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.1002089 | T. brucei Infection Reduces B Lymphopoiesis in Bone Marrow and Truncates Compensatory Splenic Lymphopoiesis through Transitional B-Cell Apoptosis | African trypanosomes of the Trypanosoma brucei species are extracellular protozoan parasites that cause the deadly disease African trypanosomiasis in humans and contribute to the animal counterpart, Nagana. Trypanosome clearance from the bloodstream is mediated by antibodies specific for their Variant Surface Glycoprotein (VSG) coat antigens. However, T. brucei infection induces polyclonal B cell activation, B cell clonal exhaustion, sustained depletion of mature splenic Marginal Zone B (MZB) and Follicular B (FoB) cells, and destruction of the B-cell memory compartment. To determine how trypanosome infection compromises the humoral immune defense system we used a C57BL/6 T. brucei AnTat 1.1 mouse model and multicolor flow cytometry to document B cell development and maturation during infection. Our results show a more than 95% reduction in B cell precursor numbers from the CLP, pre-pro-B, pro-B, pre-B and immature B cell stages in the bone marrow. In the spleen, T. brucei induces extramedullary B lymphopoiesis as evidenced by significant increases in HSC-LMPP, CLP, pre-pro-B, pro-B and pre-B cell populations. However, final B cell maturation is abrogated by infection-induced apoptosis of transitional B cells of both the T1 and T2 populations which is not uniquely dependent on TNF-, Fas-, or prostaglandin-dependent death pathways. Results obtained from ex vivo co-cultures of living bloodstream form trypanosomes and splenocytes demonstrate that trypanosome surface coat-dependent contact with T1/2 B cells triggers their deletion. We conclude that infection-induced and possibly parasite-contact dependent deletion of transitional B cells prevents replenishment of mature B cell compartments during infection thus contributing to a loss of the host's capacity to sustain antibody responses against recurring parasitemic waves.
| African trypanosomiasis caused by Trypanosoma brucei species is fatal in both humans and animals and cannot be combated by vaccination because of extensive parasite antigenic variation. Effective trypanosome control and clearance from the bloodstream involves the action of antibodies specific for the parasite's highly diverse variable surface glycoprotein antigens. However, experimental infections in mice have shown that trypanosomiasis elicits a rapid process of B cell exhaustion and loss of protective antibody responses. Indeed, both marginal zone B cells, the first line of defense against blood-borne pathogens like T. brucei parasites, and follicular B cells, which are the major source for developing high-affinity antibody-producing plasma cells and memory B cells, become depleted during infection. In addition, existing B-cell memory, both against parasite antigens and non related pathogens, is destroyed early on in infection. Here, we demonstrate that during infection, B cell development is decreased in the bone marrow and early B cell development is taken over by the spleen. However, full maturation of developing B cells is abrogated by the occurrence of transitional B cell apoptosis. This impairs the replenishment of the mature marginal zone and follicular B cell pools and prevents the buildup of protective immunity against successive parasitemic waves.
| Trypanosoma brucei is a highly antigenically variable uniflagellate protozoan of which the subspecies T. b. gambiense and T. b. rhodesiense cause Human African Trypanosomiasis (HAT), also called Sleeping Sickness. In addition the parasite infects domestic animals, contributing to Nagana, which is a fatal disease of livestock in sub-Saharan Africa. T. brucei is transmitted in tsetse fly saliva and lives and replicates in blood, lymph and interstitial fluids of its mammal hosts protected from lytic plasma components by a coat of variable surface glycoprotein (VSG). The surface coat of a T. brucei parasite consists of 107 identical densely packed VSG molecules which can be varied among a possibly unlimited repertoire of coat types via a mechanism called antigenic variation [1]–[5]. Clearance of T. brucei and other African trypanosomes from the host blood stream is mainly mediated by VSG specific antibodies [6]–[8]. T. brucei parasites have been shown to (i) deplete marginal zone and follicular B cells from the spleen [9] , (ii) induce non-specific, polyclonal B cell activation leading to clonal exhaustion [10]–[12], and (iii) cause a general decrease in bone marrow cells [13] consistent with a negative impact on lymphopoiesis and erythropoiesis. Infection of trypanosomiasis-susceptible hosts with African trypanosomes has been shown to compromise host humoral immune competence resulting in the loss of B cell responsiveness to new antigens and of recall responses to previously encountered antigens, including trypanosome VSGs and vaccines [9]. Hence, vaccination against trypanosomiasis has so far never been successful in a natural infection setting.
B2 B cell lineage development under normal conditions occurs via a series of bone marrow (BM) stromal cell facilitated processes that begin within the hematopoietic stem cell pool and proceed in hierarchical steps of lineage commitment [14], [15]. Hematopoietic stem cells (HSC), which can self renew, give rise to multi lineage progenitors (MLP) and lymphocyte primed multi lineage progenitors (LMPP) that no longer self renew. LMPP, in turn, give rise to common lymphoid progenitors (CLP), which have been shown to sustain both T and B lymphopoiesis, although these lineages may diverge within the CLP. CLP give rise to several types of precursor cells, including pre-pro-B cells [16], [17]. B lymphopoiesis then proceeds in the bone marrow yielding several developmental stages of pre-pro-B, pro-B, pre-B and eventually immature B cells, which show a high expression of the IgM form of the antigen receptor and low or no expression of the IgD maturation marker [18], [19]. To complete their development, immature B cells migrate through the periphery, however only 10% reaches the spleen as transitional B cells of the T1 type. Important is the fact that under inflammatory immune conditions, BM lymphopoiesis is often severely reduced, and is compensated for by a splenic cell differentiation process that involves the same B-cell differentiation steps, referred to as extramedullary lymphopoiesis [20], [21]. Once the transitional T1 stage has been reached, B cells develop further into T2 transitional B cells that in turn can mature into either Marginal Zone B (MZB) cells or Follicular B (FoB) cells [22]. T2 cells can also give rise to T3 transitional B cells, but the latter don't give rise to mature B cells, due to being hyper-responsive to stimulation through their BCR [23]. Each of these populations is distinguished by a unique set of cell surface antigens, allowing monoclonal antibody (mAb) phenotyping by multicolor flow cytometry [24]–[26]. Using this approach, Radwanska et al. reported that the splenic MZB and FoB cell populations become rapidly depleted in T. brucei-infected mice and do not recover [9]. Since these findings suggest an impaired replacement of mature B-cells during infection, we have now investigated B cell development, maturation and cell death of various B-cell populations in T. brucei-infected mice.
B-lymphocyte cell death in the context of inflammation and infection has been attributed in the past to several major mechanisms that include TNF-TNFR1 and Fas/Fas-L induced apoptosis, prostaglandin triggered cell death and BCR-cross-linking in the absence of proper T-cell help [20], [27]–[32]. With respect to African Trypanosomiasis, none of these aspects have been addressed so far, despite the crucial need for an effective B-cell compartment for parasitemia control. In contrast, their role and modulation during intracellular Trypanosoma cruzi infections is better documented, showing that: (i) in the absence of TNF-TNFR1 signaling, susceptibility to infection increases and coincides with abnormal B-cell differentiation in secondary lymphoid tissues [33]. (ii) CD95/FasL interaction between B cells can mediate the fratricide of IgG+ B lymphocytes [27], and (iii) myeloid cell-derived prostaglandins contribute to infection-associated apoptosis of immature B cells in a Fas-FasL independent manner [28].
Programmed cell death can be induced by a number of death factors, including Fas-FasL interaction [34], [35] and the TNF-TNF-R1 apoptosis pathway [36], [37]. With respect to the latter, it has been well established that (i) T. brucei infections induce severe inflammatory disease in susceptible hosts leading to the excessive production of pro-inflammatory factors including TNF and prostaglandins [38]–[42], and (ii) that excess induction of TNF negatively affects various lymphoid compartments [20]. In contrast to TNF, to date no information on the role of Fas has been reported in a T. brucei infection setting. The Fas apoptosis pathway is normally important in both the regulation of the immune response as well as T and B lymphocyte homeostasis [35]. Therefore, Fas-FasL mediated apoptosis plays a critical role in the mechanism for negative selection of B cells [43]–[45] and for the establishment of the B cell repertoire in the memory compartment [46]. Following activation, B cells can rapidly upregulate both Fas and FasL expression [47], [48], but the control of B lymphocyte expansion appears mainly to be regulated by FasL-expressing T cells [47].
Here we investigate the contributions of different mechanisms to T. brucei-induced abrogation of B cell development and infection-associated B cell apoptosis. Our results show that following T. brucei infection, B lymphopoiesis is truncated in the bone marrow and compensatory extramedullary B lymphopoiesis is induced (but not completed) in the spleen. Splenic B lymphopoiesis up to the stage of immature B cells was triggered but final development was severely limited by apoptosis of transitional B cells, thus preventing replenishment of mature B2 B cells. Despite the pro-inflammatory immune environment induced by experimental T. brucei infections, these events occurred independent from TNF-TNF-R1, Fas-Fas-L and prostaglandin-mediated pathways. Interestingly, in an ex vivo setting in which naïve or infection-derived splenocytes were co-cultured with living bloodstream form trypanosomes, transitional B-cell apoptosis was only observed when cell-cell contact between lymphocytes and parasites occurred. This observation corroborates the previous findings from Radwanska et al. that showed that trypanosomes can induce contact dependent cell death in anti-VSG hybridoma B-cells [49]. Using a Trans-well co-culture system, or a VSG specific Nanobody, i.e. a variable heavy chain fragment of a single chain camelid antibody devoid of its Fc part [50], we now show that preventing direct contact between the trypanosome surface coat and transitional B-cells results in an abrogation of infection-induced apoptosis in the ex vivo setting.
T. brucei infections in mice have been shown to compromise host humoral immune competence and to induce the loss of specific mature B cell populations in the spleen [9]. However, little is known about how B lymphopoiesis in the bone marrow is affected during T. brucei infection. Here, B lymphopoiesis has been examined using a C57BL/6 mouse T. brucei AnTat 1.1E infection model, which is characterized by successive waves of parasitemia and a median infection survival time of 35 days [51]. Bone marrow and spleen cells were isolated at different time points of infection and prepared for cellular characterization by multicolor flow cytometry as described in Table 1 and shown in Figure S1, S2 and S3. As presented in Figure 1, the number of very early progenitors, i.e., HSC-LMPP was minimally affected, showing only a transient reduction on day 20 p.i. Subsequent bone marrow B lymphopoiesis was severely affected in T. brucei infected mice, with major declines in all bone marrow B cell developmental stages starting with the CLP fraction. A drop in the number of CLP progenitors was detected on day 10 p.i. and this cell population remained severely depleted thereafter. The pre-pro-B cell population showed a 50% reduction by day 20 p.i., while the subsequent B-cell maturation stage i.e. the pro-B, pre-B and immature B cell populations reached more than 95% depletion by day 10 p.i. and failed to recover throughout the further course of infection. Combined, these results show that at the end stage of differentiation in the bone marrow, B cell numbers are severely depleted.
Inflammation in general has been described to induce mobilization of immature bone marrow lymphocytes [20]. Because T. brucei infection is characterized by a strong type 1 inflammatory immune response, mobilization of bone marrow precursors may result in the appearance of developing B cells in the spleen. Cells were harvested from the spleen at different time points of infection and a multicolor flow-cytometric analysis was performed according to Table 1. As shown in Figure 2, T. brucei infection induced an increase in HSC-LMPP fractions in the spleen by day 10 p.i. In addition, there was a significant rise in CLP, pre-pro-B, pro-B and pre-B cell numbers in the spleen on day 10 p.i., coinciding with the drastic losses of B cell precursors from the bone marrow. While pre-pro-B, pro-B and pre-B cell numbers remained significantly elevated in the spleen by day 20 p.i., these populations returned to pre-infection levels towards the end of infection (day 30 p.i.). In contrast, while there was no early stage decrease in immature B cells in the spleen, there was a significant loss of this population towards the end of infection.
In contrast to the elevation in early B cell developmental stages observed in the spleen, Radwanska et al. [9] have reported depletion of mature marginal zone and follicular B cells suggesting impaired replacement. Therefore, B cell development at the transitional B cell stage in the spleen was examined here, as these cells provide the link between the immature B cell stage and the mature marginal zone and follicular B cell stages. Flow cytometric analysis of the T1, T2 and T3 transitional B cell populations of T. brucei infected mice (Figure 3) revealed a transient increase in transitional B cells numbers that however rapidly faded towards day 10 p.i. On days 20 and 30 p.i. the number of transitional T2 and T3 B cells was significantly decreased compared to uninfected control mice.
To address whether apoptosis is contributing to the depletion of transitional B cells in the spleen, the amount of active caspases-1, −3, −4, −5, −6, −7, −8 and −9 inside the transitional T1, T2 and T3 B cells was measured at different time points of infection by flow cytometry (Figure 4A). Although infection did induce apoptosis in both the T1 and T2 transitional B cell population, the T3 population only showed a temporary increase of caspase activation after the first week of infection (Figure 4B). The elevation in levels of caspases in transitional B cells coincided with the contraction of these B cell populations between days 6 and 10 p.i. (Figure 3A), i.e., immediately following peak parasitemia (8×107 T. brucei/ml blood). In the model used here, infections are initiated with 5×103 T. brucei AnTat 1.1 and remission of the first parasitemic wave occurs between 6 and 7 days p.i. [41]. It is noteworthy that when the infection was initiated with 108 T. brucei AnTat 1.1, peak parasitemia and wave remission occurred at 4 days p.i. (data not shown) and transitional B-cell apoptosis was observed as early as 3 days p.i. (Figure 5). These observations suggest a direct link between levels of parasitemia and the kinetics of induction of splenic T1/2 B-cell apoptosis. Combined, these data indicate the induction of apoptosis in both T1 and T2 transitional B cell populations occurs at, or close to, peak parasitemia. The transitional B cells would, under normal conditions, give rise to the continuous replenishment of marginal zone and follicular B cell populations, but clearly are unable to do so in the infected mice.
Experimental T. brucei infections induce high circulating levels of TNF [38], a host cytokine that has been reported to be involved in the induction of immunopathology during T. brucei infections [38]–[42]. Furthermore, TNF is known to be a potent inducer of apoptosis through the TNF-R1 signal pathway. Therefore, to examine the possibility of TNF-mediated apoptosis of T1 and T2 transitional B cells, TNF−/− mice were infected with T. brucei and the number of T1 and T2 transitional B cells in the spleen was examined at different time points of infection. Figure 6 illustrates that on day 10 p.i. WT mice and TNF−/− mice both suffered from similar levels of transitional T1 and T2 B cell depletion. While on day 20 p.i. a temporary recovery of T1 transitional B cells occurred only in the TNF−/− mice, the same final level of 75% reduction in both T1 and T2 transitional B cells was observed in WT as well as TNF−/− mice towards the end of infection. As an additional control, apoptosis of transitional B cells was recorded using the poly-capase activation FACS analysis outlined above. Here, the percentage of T1 and T2 B cells undergoing apoptosis in T. brucei-infected TNF−/− as well as TNF-R1−/− mice equaled the results reported for WT mice (data not shown).
Next, the expression of the death receptor Fas (CD95) on T1 and T2 transitional B cells, and the potential involvement of the Fas-FasL apoptosis pathway were analyzed at different time points of infection. Control T1 and T2 transitional B cells express low, but detectable levels of Fas on their surface. During infection however, there was a strong increase in the level of Fas expression on the surface of both T1 and T2 transitional B cells, as shown here for days 8, 10, 14 and 20 p.i., (Figures 7A and B). The increase in surface Fas expression coincided with elevated caspases activity (Figure 4) and with the loss of transitional B cells from the spleen (Figure 6). Interestingly, also in TNF−/− mice as well as TNF-R1−/− mice, the increase in Fas expression on both T1 and T2 transitional B-cells preceded the rapid loss of these cells from the spleen further suggesting a correlation between the up-regulation of surface-expressed Fas and the trypanosomiasis-associated destruction of the transitional B-cell compartment (Figure S4).
If apoptosis of transitional B cells in T. brucei-infected mice is mediated through Fas, then the loss of the cells would be expected to be reduced in mice with constitutively low Fas expression. However, this proved not to be the case as the extent and kinetics of transitional B cell loss following infection of lpr mice with T. brucei AnTat 1.1 was the same as in wild type mice (data not shown). In addition, if apoptosis were to be mediated through Fas in wild type mice, it would be expected to be ameliorated by blocking the activation of Fas by its ligand, FasL, through administration of neutralizing anti-FasL antibody [27]. However, this also proved not to be the case. Indeed, mice infected with T. brucei and treated by i.v. injection with 100 µg of anti-FasL antibody on days 4, 5 and 6 p.i. did not manifest a measurable alteration in transitional B cell loss from the spleen (Figure 8a), nor did the remaining transitional B-cells in these mice exhibit a change in their caspase activity profile (Figure 8b) relative to mice receiving a control immunoglobulin treatment. Thus, while the occurrence of transitional B cell apoptosis after infection with T. brucei coincides with an increased expression of the death receptor Fas on the surface, apoptosis of transitional B cells could not be prevented by the anti-FasL antibody treatment used here.
Besides TNF and Fas, also prostaglandins have been implicated in B-cell apoptosis, in particular in a (intracellular) T. cruzi infection setting [28]. To examine the possible contributions of cyclo-oxygenase (COX) products to T. brucei-induced transitional B cell death in vivo, indomethacin (a nonsteroidal anti-inflammatory drug that inhibits COX activity), or control physiological buffer, was administered daily to infected mice by i.p. injection using a previously described protocol [52]. The indomethacin treated mice infected with T. brucei did not differ from control infected mice with respect to transitional B cell apoptosis (Figure S5). In a second set-up T. brucei-infected mice received indomethacin in the drinking water at a concentration of 14 µg/ml [53]. Here again, no difference was observed in the percentage of transitional B cells undergoing apoptosis between treated mice and untreated mice (data not shown).
In addition to the in vivo analysis of prostaglandin contribution to the induction of transitional B cell apoptosis, an in vitro assay was designed to further examine the contribution of cyclooxygenase products to transitional B cell apoptosis. Here, spleen cells from uninfected mice were co-cultured in a Trans-well system [29] with spleen cells from uninfected mice or from mice that had been infected for varying times with T. brucei Antat 1.1. Culture conditions included incubations in the presence or absence of trypanosomes and of indomethacin. However, in none of the experimental conditions was there a difference in percentage of transitional B cells undergoing apoptosis compared to the controls (representative data are presented in Figure S6), leading us to conclude that prostaglandins (either host or parasite derived) are not a major contributor to transitional B-cell apoptosis during T. brucei infections.
When performing the Trans-well experiments described above to address the potential role of prostaglandins in transitional B cell apoptosis, control conditions included co-cultures in which parasites and splenocytes (either from uninfected or day 5 T. brucei AnTat 1.1 infected mice) were not separated by a physical barrier. In line with previous results obtained in co-cultures of bloodstream form trypanosomes and B-cell hybridoma cells raised against VSG [49], here the T1/2 transitional B-cell population survival was significantly impaired, which was not observed when parasites and B-cells were separated by a 0,4 µm polycarbonate transmembrane. Indeed, Figure 9A (columns 1 and 2) shows that incubation of spleen cells from uninfected mice with freshly isolated living trypanosomes (10 trypanosome/spleen cell) resulted in 75% lower survival of T1 and T2 B cells relative to incubation in medium. In contrast, there was no difference in the numbers of viable IgM−ve B220−ve leukocytes recovered from splenocytes cultured in the presence or absence of trypanosomes (data not shown).
Similarly, when spleen cells from mice infected 5 days earlier with T. brucei were incubated with trypanosomes, there was a 50% reduction in viable T1 and T2 B cells relative to cells cultured in medium (Figure 9A columns 3 and 4) and again no reduction in the recovery of IgM−ve B220−ve leukocytes (data not shown). However, when cultured cells and trypanosomes were separated by a 0.4 µm polycarbonate transmembrane there was no decrease in the number of viable transitional B cells in the culture (Figure 9A column 5) indicating that B cell loss does not result from diffusible trypanosome products or from consumption of essential medium components by living trypanosomes, but from the direct contact between the cells and the parasites. Culturing splenocytes from 5 day-infected mice with parasite lysate, or purified soluble VSG (sVSG) from the parasites, rather than living parasites, also did not result in the loss of transitional B cells (Figure 9B, columns 1–3). To further examine whether contact with VSG on living T. brucei is required for depletion of transitional B cells in vitro, T. brucei AnTat 1.1 trypanosomes were pre-incubated with a VSG-specific Nanobody (Nb, a dromedary heavy chain antibody fragment, devoid of its Fc part, having no detrimental effect on parasite survival [50]), prior to addition to spleen cells from 5 day-infected mice. A comparison of Figure 9B column 4 (control) and column 5 (Nb) shows that the pre-treatment of trypanosomes with nanobodies prevented the killing of B cells. Thus, these results corroborate the obtained Trans-well results indicating that VSG dependent direct contact between intact living trypanosomes and transitional B-cells can trigger cell death.
Infection with African trypanosomes causes mice and other trypanosomiasis-susceptible mammals to develop non-specific hypergammaglobulinemia and polyclonal activation that in the end leads to B cell clonal exhaustion [10]–[12]. While the mechanisms underlying B cell clonal exhaustion have yet to be resolved, results by Radwanska et al. showed that the mature marginal zone and follicular B cell populations rapidly disappear during experimental trypanosome infections and that vaccine-induced B-cell memory is destroyed in a non-specific manner [9]. Additional results suggested that while trypanosome infection initially leads to rapid immune activation and buildup of trypanosome/VSG-specific immunity against trypanosomes in the initial parasitemic wave, humoral immune responsiveness is rapidly lost. Impairment in the replacement and recruitment of naïve B-cells to the spleen apparently prevents the efficient activation of specific immunity later on in infection [47], [48]. In order to gain insight into the mechanisms of infection-induced B-cell dysfunction, we have examined how T. brucei infection affects B cell development in the bone marrow and the survival and maturation of B cells in the periphery. Our studies show that BM lymphopoiesis is shut down during infection and that compensatory extramedullary B lymphopoiesis is truncated by apoptosis of transitional B cells thus preventing replenishment of mature marginal zone and follicular B cell compartments.
Mice infected with T. brucei exhibit reduced numbers of B cells at all developmental stages in the bone marrow and transitional stages in the spleen. Within the bone marrow, loss of CLP preceded that of other progenitor populations raising the possibility that downstream losses resulted from depletion of this precursor population. However, if the loss of CLP was solely responsible for the downstream disrupted B cell development, then the ratio of a subsequent B cell developmental population to its immediate upstream progeny and downstream precursors should remain unaffected, which was not the case. Hence, additional processes besides progenitor depletion must be operating to limit B cell development in T. brucei infected mice.
Although aberrant differentiation, apoptosis and expulsion from the bone marrow may singly or jointly contribute to loss of B cell precursors from the bone marrow, we favor the latter mechanisms based on the following arguments: (1) during infection we could not measure any increase in B cell precursor apoptosis in the bone marrow (Figure S7), (2) alterations in the expression of essential B cell development-specific transcription factors in the BM like Icaros, PU.1, EBF and E2A and the IL-7 growth factor have never been reported and were not observed in our laboratory either (data not shown), (3) a reduction in bone marrow CXCL12 expression was the only parameter found by us to correlate with the observed loss of developing B cells during infection (Figure S8). The importance of this result is underlined by the knowledge that inflammation-induced reductions in stromal bone marrow CXCL12 expression indeed have been reported by others to correlate with a premature lymphocyte efflux [20]; (4) there are elevated numbers of HSC-LMPP and CLP as well as pre-pro-B, pro-B, pre-B and immature B cells in the spleen, suggesting the possibility of an increased influx of these cells. In addition, during the early stage of infection small numbers of HSC-LMPP, CLP, pre-pro-B, pro-B, pre-B and immature B cells were also found in the blood, the liver, peritoneum and lungs and small numbers of pro-B, pre-B and immature B cells were found in the lymph nodes. Taking these findings and arguments together, we hypothesize that during early T. brucei infections, B cell precursors prematurely migrate out of the bone marrow as a result of the initiation of inflammation, and at least in part home to the spleen, allowing transient extramedullary B lymphopoiesis to take place.
Despite the initiation of extramedullary B lymphopoiesis, infection-induced loss of mature B-cells marks progressing trypanosomiasis in experimental infections. Various factors, in particular the induction of systemic inflammation, could contribute to this. Hence, we addressed the potential involvement of three likely participants, i.e. TNF, Fas and prostaglandins, all previously shown to be potentially involved in immune and B-cell malfunctioning. Paradoxically, despite the presence of high systemic TNF levels during infection [38] and the clear evidence shown here that transitional B cell loss through apoptosis coincides with elevation of Fas expression on both T1 and T2 B cells, our results showed through the use of knock-out mice as well as anti-FasL antibodies that neither the TNF- nor the Fas- death pathways acting alone are responsible for transitional B cell apoptosis in a T. brucei infection setting. In addition, despite the reported production of prostaglandins by trypanosomes [29] and macrophages in infected mice [54] inhibition of prostaglandin/cyclooxygenase activity in T. brucei infected mice by administration of indomethacin did not rescue transitional B cells, contrasting with results obtained in T. cruzi infections [28]. However, as each of the death pathways could be redundant in a multi-factorial complex event such as infection-induced apoptosis, their in vivo individual contribution to transitional B cell apoptosis in the T. brucei infection model should not be formally excluded.
The absence of a clear role for the Fas apoptosis pathway in T. brucei induced transitional B-cell death is particularly surprising, taken that (i) Fas upregulation on these cells is reported here to be a clear hallmark of progressing infection, and (ii) that Fas-FasL B cell killing in the context of an infectious disease has previously been described in the case of Trypansosoma cruzi infections, which induce Fas-mediated fratricide of IgG+ B lymphocytes specific for parasite antigens but not self antigens [27]. In addition, Fas-mediated cell death is known to be important in the regulation of the immune response and T and B lymphocyte homeostasis [35], where for instance Fas-FasL mediated apoptosis plays a critical role in the mechanism for negative selection of B cells [44], [45] and the establishment of the B cell repertoire in the memory compartment [46]. However, despite the reported role for Fas in parasite-induced fratricide, its contribution to the killing of virus-infected, damaged or excess cells and its implication in various immunopathological disorders [43], [44], [47], our study did not provide any evidence functionally linking Fas upregulation and B cell apoptosis. Indeed, treating infected mice three times (with daily interval) with 100 µg of anti-FasL antibody just prior to peak infection, in order to block in vivo the activation of the Fas death cascade did not alter the kinetics at which the transitional B cell population underwent apoptosis. It could be argued that the doses of anti-FasL antibody used in this experiment do not functionally block the activation of the Fas apoptosis pathway, however, when Lpr mice are infected with T. brucei, transitional B cells are lost from the spleen to the same extend as in the wild type mice (data not shown), arguing once again against a major involvement of the Fas-apoptosis pathway.
Having shown that T. brucei driven transitional B-cell apoptosis occurs in a TNF-, Fas- and prostaglandin-independent manner, our study next focused on a model system that could help to functionally unravel infection-induced B-cell apoptosis. Interestingly, since B cell apoptosis only occurs when living parasites are administered to the mice, and not when mice are treated on a daily basis by the injection of high doses of trypanosome lysate (results not shown), the presence of a toxin or a super-antigen-like activity by trypanosome molecules appears to be excluded. These results are in sharp contrast to the superantigen-mediated death of mature B cells in the case of Staphylococcus aureus, where injection with the virulence factor protein A alone is enough to mimic massive Fas and TNF-independent bacterial induced B cell death, and cause a ‘hole’ in the immune repertoire recognizing the pathogen [55], [56]. Also, it is worth recalling that the in vivo experiments presented in this study show that initiation of transitional B cell apoptosis depends on the timing at which peak levels of parasitemia are reached, which is a function of the number of living bloodstream form parasites used for initiation of the infection. This observation suggests a link between the presence of high numbers of living parasites and the induction of parasite-induced transitional B-cell apoptosis consistent with the possibility that B cell hyper-stimulation, through multiple VSG (variable surface glycoprotein - attached to the parasite surface)/BCR (B cell antigen specific receptor - attached to the B-cell surface) interactions, could be a major contributor to this process, as might exposure of B cells to short-lived parasite products which would be active only when directly delivered to the target cell. With respect to BCR signaling, cross-linking of the BCR has been well described to trigger apoptosis of T1 and T2 transitional B cells, which can be ameliorated in the case of T2 B cells by anti-apototic signaling through the BlyS receptor BR3. Transitional B cells of the T2 type can also be rescued from BCR crosslinking-induced apoptosis by T cell help [30]–[32], [57], [58]. However, T cell suppression is one of the hallmarks of T. brucei infection [54], making it likely that T2 transitional B cells in infected mice are exposed to the parasites in the absence of proper survival-stimulating T cell help. Of crucial importance is the notion that membrane-bound antigens that can extensively engage BCRs trigger rapid BCR-mediated apoptosis in a Fas-independent manner [59]. Hence, in the context of trypanosome-B cell interaction, the presence of 10 million identical VSG molecules densely packed on the surface of the parasite could be responsible for causing BCR receptor clustering on the surface of the transitional B cells, leading to hyper-stimulation of the B cell and TNF/Fas-independent cell death in the absence of proper T-cell signaling. Avidity, due to multiple VSG/BCR interactions, in this case would have a much higher impact than actual BCR affinity/specificity for a given VSG.
Using an in vitro co-culture system of splenocytes and live bloodstream form trypanosomes, we showed that direct contact between the living parasites and host B cells can indeed trigger the deletion of the latter. Depletion of T1 and T2 transitional B cells was maximally induced with 10 trypanosomes per spleen cell in culture but was observed with as few as 0.1 trypanosomes/spleen cell in culture, with 50% depletion occurring at a ratio of between 0.1 and 0.5 trypanosome/spleen cell (Figure S9). This falls within the physiologic range in vivo for infections with T. brucei AnTat 1.1, where the ratio of trypanosomes to viable nucleated cells in the spleen is 0.25:1 at peak parasitemia. In co-cultures of trypanosomes and splenocytes in which contact between the two is prevented using a Trans-well system or a pre-incubation of the parasite with a VSG-specific Nanobody, abrogation of transitional B cell deletion is observed. This result mirrors a previous observation by Radwanska et al. where cell death was triggered in IgM expressing hybridoma cells when cultured in the presence of living trypanosomes [49]. Again, in the co-culture system neither lysate, nor purified parasite VSG could mimic the apoptosis-inducing effect of living parasites (nor could anti-FasL or prostaglandin inhibition block the detrimental effect of living trypanosomes). Together, these observations strengthen the hypothesis that the direct interaction of B cells with epitopes on T. brucei causes transitional B cell death. Unfortunately we cannot evaluate Nanobody blockage of transitional B cell apoptosis in vivo as these antibody fragments are very rapidly cleared from the circulation by the kidneys [60]. Thus, although the mechanism of T. brucei-induced transitional B cell depletion in vivo remains to be fully elucidated, we did observe that living trypanosomes induce cell death in transitional B cells in vitro through a contact-dependent mechanism. Micro-array analysis of material from both hybridoma co-culture assays and spleen-derived B-cell co-cultures with trypanosomes are now underway in order to shed light on the signal cascades involved in this contact triggered apoptosis.
Combined, our study provides evidence for trypanosomiasis-induced apoptosis of transitional B cells in the spleen and it proposes a mechanism for T. brucei-induced B cell clonal exhaustion and loss of humoral immune competence in trypanosomaisis-susceptible hosts. Under normal conditions the production of high-affinity, antigen-specific, class-switched, antibodies takes up to 10 days after immunization [61], [62]. Here, in our infection model, about 90% of the transitional B cells are undergoing apoptosis by day 8 of infection, making the replenishment of the mature marginal zone and follicular B cell populations impossible and therefore obstructing efficient germinal center reaction and the renewal of the plasma B cell pool. Since parasite-specific antibodies are essential for parasite control, inhibition of B cell maturation at the transitional stage is an efficient evasive mechanism developed by the parasite to interfere with the protective antibody responses of the host and establish a sustained infection. It is important to stress here that our studies are based on a mouse model system for African trypanosomiasis, which manifests certain limitations. However, the results obtained in this model provide guidelines for analysis of B cell pathology in more relevant hosts including susceptible livestock species such as cattle and goat, or the retention of immune function in natural trypanotolerant animals such as the Cape Buffalo, which have a sustained capacity to generate effective protective antibody responses against T. brucei and other African trypanosomes during chronic infection, thus, maintaining cryptic parasitemia with few or no signs of disease [7]. Based on our findings and the earlier data reported by Radwanska et al. [9], it would also be crucial to investigate whether a similar B cell pathology occurs in T. brucei infected humans. In this context, it would be interesting to compare B cell pathology between T. brucei gambiense (manifesting a chronic infection in humans with prolonged parasitemia control) and T. brucei rhodesiense (manifesting an acute infection whereby parasitemia control is lost very early) infected patients, where in the latter it could be that rapid destruction of the B-cell compartment is the underlying reason for failure of parasitemia control, the rapid induction of systemic inflammation, and the subsequent passage of the parasite through the blood-brain barrier.
The 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 and Guidelines for the Use of Laboratory Animals in Research, Teaching and Testing of the International Council for Laboratory Animal Science. All animal work was approved by the appropriate committee at the University of Massachusetts (IACUC protocol #s 26-09-09/27-09-09 and 2010-0028) and at the Vrije Universiteit Brussel (ethics committee protocol # 09-220-1).
All mice were housed under barrier conditions. Male C57BL/6 wild type (Taconic, Germantown, NY), Lpr and TNF−/− C57BL/6 mice (provided by VUB, Belgium) (7–9 week old) were infected by intraperitoneal (i.p.) injection of 5000 exponentially growing pleomorphic Trypanosoma brucei Antat 1.1 (EATRO 1125 stock) [51]. Parasitemia was assessed in blood collected from the tail vein during infection. Blood was diluted in RPMI (Gibco, Grand Island, NY, USA) and the number of trypanosomes present in the blood was estimated using a hemocytometer and a light microscope.
B cell populations were analyzed by flowcytometry. Both spleen and bone marrow from femur and tibia were harvested from non-infected control and infected mice at different time points of infection. Cell suspensions were prepared in FACS buffer (1.0% BSA [Sigma, St. Louis, MO] in DPBS) and red blood cells were lysed using ACK lysis buffer (0.15M NH4Cl, 1.0 mM KHCO3, 0.1 mM Na2-EDTA). Non-specific binding sites were blocked using Fc block (CD16/CD32 Fcγ III/II, BD biosciences, San Jose, CA) for 30 minutes at 4°C. Cells were washed twice with FACS buffer and stained with biotin- or fluorochrome-conjugated primary antibodies (see following section) for 30 minutes at 4°C. After washing twice, cell suspensions stained with biotin-conjugated antibodies were incubated with streptavidin-conjugated fluorochromes (listed in the text), which detects cell bound biotinylated antibodies, and incubated for an additional 30 minutes at 4°C. Finally, cells were resuspended in FACS buffer with 1 µg of 7-amino-actinomycin D (7AAD), a fluorescent DNA dye that binds to membrane permeable dead or dying cells, (BD biosciences). Analyses were performed using a FACS Canto II flow cytometer (BD Biosciences) and data were processed using FLOWJO software (Tree Star Inc., Ashland, OR). The total number of cells in each population was determined by multiplying the percentages of subsets within a series of marker negative or positive gates by the total cell number determined for each tissue.
The following antibodies were added to 100 µl aliquots of 106 Fc-blocked leukocytes prepared as described above: 0.5 µg anti-IL7rα-FITC (clone A7R34), 0.2 µg anti-IgM-PE (clone II/41), 0.2 µg anti-IgM PE-Cy7 (clone II/41), 0,25 µg hamster IgG2 κ isotype control (clone B81-3), 0.5 µg anti-CD11b-FITC (clone M1/70; 0.5 mg/ml), 0.5 µg anti-CD23-FITC (clone B3B4), 0.5 µg anti-CD45R (B220)-FITC (clone RA3-6B2), 0.2 µg of anti-CD45R (B220)-PE-Cy7 (clone RA3-6B2), 0.2 µg anti-CD93–PE (clone AA4.1), 0.2 µg of anti-CD93-APC (clone AA4.1), 0.5 µg anti-CD95 –FITC (clone Jo2), 0.2 µg of anti-CD117 (ckit)-APC (clone 2B8), purchased from eBioscience (San Diego, CA); 0.2 µg anti-CD1d-PE (clone 1b1), 0.2 µg of anti-CD19-APC-Cy7 (clone 1D3), 0.2 µg of anti-CD43-PE (clone 1B11), 0.2 µg of anti-CD45R (B220)-APC-Cy7 (clone RA3-6B2), 0.2 µg of streptavidin-PerCP, 0.2 µg of streptavidin-PE-Texas Red, purchased from BD Biosciences (Erembodegem, Belgium); 2 µg of each of the following antibodies: CD3ε, CD11b (Mac-1), Gr-1 (Ly-6G and Ly-6C) and Ter-119 (Ly-76) from the Biotin-conjugated Mouse Lineage Panel (BD Biosciences, San Jose, CA).
Cells were stained as described in the previous section with antibodies. For the polycaspases-based apoptosis assay, labeled cells were further reacted with the FLICATM fluorescent inhibitor of caspase-1, −3, −4, −5, −6, −7, −8 and −9, using the FAM Poly Caspases Assay Kit for flow-cytometric analysis (Molecular probes, Invitrogen, Leiden, the Netherlands).
Mice infected with T. brucei where treated i.v. with 100 µg of purified anti-FasL antibody (clone MFL3, purchased from BD Biosciences) on days 4, 5 and 6 of infection and on day 7 of infection total spleen cells were isolated from treated mice and untreated controls and prepared for flowcytometric analysis as described above.
On day 5 of infection mice were sacrificed to collect blood. In addition, total spleen cells were isolated from the infected mice and uninfected control mice and prepared for cell culture as described above. Trypanosomes were purified from the blood by anion exchange chromatography according to the method of Lanham & Godfrey [63]. Then, 106 total spleen cells were put in co-culture with or without 107 (or fewer) live bloodstream form parasites at 37°C, 5% CO2 and 95% humidity in RPMI1640 medium containing 10% FBS, 2 mM pyruvate, 0.2 mM 2-mercaptoethanol, and penicillin/streptomycin under different experimental conditions and 20 h later, the cells were stained for flowcytometric analysis. Parasite lysate was prepared by 3 repeated cycles of freezing at −80°C and thawing and used at indicated concentration. Soluble VSG (sVSG) was prepared through heat-shock treatment of a purified trypanosome suspension, which forces them to release their VSG molecules, followed by anion exchange chromatography, and used at indicated concentrations. Nanobodies against soluble trypanosome VSG were prepared as described in [50], starting from an immune library of VHH fragments of heavy chain dromedary antibodies, obtained after multiple vaccination with T. brucei AnTat 1.1 VSG. Trypanosomes were pre-incubated with Nanobody BankIt1413802 Nb_An05–04 HQ680968 (monovalent and devoid of an Fc chain; 15 µg/ml medium ) at the indicated concentrations 1 hour prior to addition to the splenocytes, and the antibody was included in culture medium during subsequent incubations. Under these assay conditions, binding of the Nanobody onto the trypanosome surface did not result in altered parasite survival.
Statistical comparisons were performed by ANOVA and means were compared using Tukey and Dunnett's post test when p≤0.05 (GraphPad Prism v.4.0, GraphPad Software Inc. San Diego, CA).
|
10.1371/journal.ppat.1005462 | ER Adaptor SCAP Translocates and Recruits IRF3 to Perinuclear Microsome Induced by Cytosolic Microbial DNAs | Stimulator of interferon genes (STING, also known as MITA, ERIS or MPYS) induces the activation of TBK1 kinase and IRF3 transcription factor, upon sensing of microbial DNAs. How IRF3 is recruited onto the STING signalosome remains unknown. We report here that silencing of the ER adaptor SCAP markedly impairs the IRF3-responsive gene expression induced by STING. Scap knockdown mice are more susceptible to HSV-1 infection. Interestingly, SCAP translocates from ER, via Golgi, to perinuclear microsome in a STING-dependent manner. Mechanistically, the N-terminal transmembrane domain of SCAP interacts with STING, and the C-terminal cytosolic domain of SCAP binds to IRF3, thus recruiting IRF3 onto STING signalosome. Mis-localization of SCAP abolishes its antiviral function. Collectively, this study characterizes SCAP as an essential adaptor in the STING signaling pathway, uncovering a critical missing link in DNAs-triggered host antiviral responses.
| The stimulator of interferon genes (STING/MITA/ERIS/MPYS) is characterized as the converging point of the cytosolic DNAs-triggered innate immune signaling, and its function has been well documented in mediating the production of type I interferon and other pro-inflammatory cytokines. It remains intriguing to address how IRF3 is recruited onto the STING signalosome. In this study, we have further identified and characterized the SREBP cleavage-activating protein (SCAP) as the long-sought-after adaptor of the STING signaling. Upon microbial DNA challenge, SCAP translocates from ER, via Golgi, to perinuclear microsome in a STING-dependent manner. SCAP thus serves as a scaffold adaptor to recruit IRF3 and facilitate its integration into the perinuclear microsomes. Our study reveals an important missing link in innate immunity, further highlighting the physical and/or functional links between innate immunity and metabolism.
| Microbial infections represent an ever-present threat to host homeostasis and survival. The extracellular and intracellular microbes are dynamically and rapidly sensed by specific Pattern Recognition Receptors (PRRs), including TLRs, NLRs and RLRs [1–3]. Upon recognition of the conserved Pathogen Associated Molecular Patterns (PAMPs), PRRs initiate a myriad of signal transduction pathways, triggering innate and adaptive immune responses to eliminate the microbial pathogens [4,5].
DNAs derived from DNA viruses, bacteria or damaged host cells could activate the IRF3 and/or NF-κB signaling pathways, thus inducing the production of type I interferons (IFNs) and other pro-inflammatory cytokines [6,7]. How cells sense and respond to RNA virus infection is well characterized in the past decade [8–10]. Our understanding of the DNA-triggered signaling is relatively limited. TLR9 detects CpG DNA from endolysosome in the immune cells [11]. Multiple cytosolic sensors are proposed to detect viral or microbial DNAs in cytosol, including cGAS, RNA polymerase III, Mre11, DNA-PKcs, IFI16 and DDX41 [12–18]. Further studies are needed to clarify the physiological relevance of some of the putative DNA sensors, and to address the biochemical and functional interactions among these sensors.
Stimulator of interferon genes (STING, also known as MITA, ERIS or MPYS) is characterized as the converging point of the recently identified DNA sensors. STING is an Endoplasmic Reticulum (ER)-associated membrane protein, indispensable for inducing the antiviral innate responses triggered by microbial DNAs [19–22]. For examples, STING-deficient cells fail to induce type I IFN production after stimulation of dsDNA or infection with herpes simplex virus 1 (HSV-1) or Listeria monocytogenes [23]. STING knockout mice are highly susceptible to lethal infection by HSV-1 [23]. STING can also bind directly to cyclic dinucleotide (CDNs), including cGAMP, c-di-GMP and c-di-AMP [24,25].
CDNs and/or upstream DNA sensors could induce STING dimerization, causing its translocation from the ER, via Golgi, to perinuclear microsome [21,23,26]. Recently, we have identified the unexpected function of the autocrine motility factor receptor (AMFR, a.k.a GP78) and the insulin induced gene 1 (INSIG1) in innate immunity [27]. AMFR and INSIG1 are ER-resident ubiquitin E3 ligase, responsible for catalyzing the K48-linked poly-ubiquitination of the ER misfolded proteins, a process essential for the ER Associated Degradation (ERAD) [28]. We characterize AMFR/INSIG1 to interact specifically with STING, and to catalyze the K27-linked poly-ubiquitination of STING. The K27-linked polyubiquitin chain on STING serves as an anchoring platform for recruiting and activating TBK1, which then phosphorylates the transcription factor IRF3 [27]. Notably, IRF3 could not bind to the K27- or K63- linked polyubiquitin chain. How IRF3 is recruited onto the STING signalosome remains largely unknown.
SREBP cleavage-activating protein (SCAP) is a polytopic membrane protein on ER, harboring an N-terminal domain with eight transmembrane helices, and a C-terminal domain with five WD-repeat [29]. It is well established that SCAP interacts with INSIG1 and modulates the lipid homeostasis [30]. In this study, we demonstrate that SCAP interacts with STING independent of INSIG1, and SCAP is indispensable for DNAs-triggered host antiviral responses. Upon HSV-1 infection, SCAP translocates from ER, via Golgi, to perinuclear microsome in a STING-dependent manner. SCAP thus serves as a scaffold adaptor to recruit IRF3 onto the STING signalosome, which reveals a critical missing link in innate immunity.
Our recent study [27] identified INSIG1 to specifically interact with STING (Fig 1A). Given that INSIG1 interacts with SCAP in lipid metabolism [30], we wondered whether SCAP was also a component in the STING signalosome. We confirmed the association between INSIG1 and SCAP via the co-immunoprecipitation assay (Fig 1B). We observed that SCAP associated with STING exogenously and endogenously (Fig 1C and 1D). This association was marginally enhanced upon HSV-1 (Fig 1E left), Listeria monocytogenes (Fig 1E middle) or ISD (Fig 1E right) stimulation. Unexpectedly, silencing of INSIG1 did not affect the association between STING and SCAP (Fig 1C and 1D), indicating that STING associates with SCAP in an INSIG1-independent manner. This suggests that the STING signalosome is physically and functionally distinct from the Lipid Regulatory Complex.
The transmembrane region of STING (1–175 aa) was important for its interaction with SCAP (Fig 1F). Likewise, the transmembrane region of SCAP (1–735 aa) was mapped to mediate the same interaction (Fig 1G). Confocal microscope imaging confirmed that SCAP co-localized with STING exogenously and endogenously (Fig 1H). HSV-1 and Listeria monocytogenes infections marginally induced the expression of SCAP (Fig 1I). Taken together, these data suggest that SCAP is a new component in the STING protein complex.
To probe the potential function of SCAP in innate immunity, we screened out the specific and effective siRNAs against Scap (Scap siRNA 3060, Scap siRNA 3465 for mouse and SCAP siRNA 1302 for human), all of which could markedly diminish the expression of exogenous and endogenous SCAP (Fig 2A and 2B). It was observed that knockdown of endogenous Scap inhibited the DNA mimics poly(dA:dT)-triggered induction of Ifnb mRNA (Fig 2C). In contrast, poly(I:C)- or SeV-triggered RIG-I signaling was marginally affected in Scap knockdown cells (Fig 2C). Knockdown of Scap had no inhibitory effects on the TLRs-mediated activation of the Ifnb mRNA triggered by LPS or poly(I:C) added (Fig 2C). These data indicate that SCAP specifically regulates the cytosolic DNA-triggered expression of IFN-β.
To substantiate, we explored the effect of Scap knockdown on the expression of IRF3-responsive genes induced by cytosolic DNA challenge, using qPCR (quantitative PCR). As expected, silencing of Scap markedly attenuated the induction of the IRF3-responsive genes (Ifnb, Ifna4 and Cxcl10) in MEF cells, stimulated by the DNA mimics [poly(dA:dT) or ISD] (Fig 2D) treatment or the DNA pathogens (Listeria monocytogenes or HSV-1) (Fig 2E). However, silencing of SCAP apparently displayed no effect on the Thapsigargin-induced ER stress (S1 Fig).
To make it more physiologically relevant, we further investigated the function of SCAP in primary cells. BMDMs (bone marrow derived macrophage) were transfected with siRNAs against Scap, followed by HSV-1 or Listeria monocytogenes stimulation. Consistently, silencing of Scap markedly attenuated the expression of IRF3-responsive genes in BMDMs (Fig 2F and 2G). Collectively, these data indicate that SCAP is a positive modulator of the cytosolic DNA-triggered STING signaling pathway.
We further delineated the topology of SCAP in the STING signaling pathway. Exogenous expression of cGAS or STING could respectively activate the IFN-β-luciferase reporters, and these activations were obviously impaired when knocking down Scap (Fig 3A and 3B). In contrast, knockdown of Scap marginally affected the expression of IFN-β-luciferase reporter when ectopically expressing TBK1 (Fig 3C). Likewise, Scap knockdown had no effect on the activation of the IFN-β-luciferase reporter, when cells were stimulated with the exogenous IRF3-5D (Fig 3D). Furthermore, ectopic expression of SCAP only or both SCAP and INSIG1 could not activate the IFN-β-luciferase reporter (S2A and S2B Fig). Given the hierarchical relationships among these signaling molecules, we reasoned that SCAP modulates the STING signaling downstream of STING and upstream of IRF3.
Consistently, the expressions of PRDIII-I-luciferase reporters stimulated by cGAS, STING or TBK1 were attenuated in Scap knockdown cells, whereas the expressions of PRDIII-I-luciferase reporters stimulated by IRF3-5D remained intact in Scap knockdown cells (S3A Fig). The ubiquitination of STING or the recruitment of TBK1 was not affected by endogenous SCAP depletion (S3D to S3G Fig). Notably, knockdown of Scap led to an apparent decrease in the phosphorylation of IRF3, but not that of TBK1, when stimulating cells with either poly(dA:dT) (Fig 3E and S3B Fig) or ISD (Fig 3F and S3C Fig). Consistently, ISD-triggered dimerization and nuclear translocation of IRF3 were markedly impaired when silencing SCAP. In contrast, knockdown of Tom20 could not influence the dimerization and nuclear translocation of IRF3 (S4A and S4B Fig).
We confirmed via confocal microscopy that STING traffics from the ER to perinuclear/Golgi foci (also called perinuclear microsome) upon HSV-1 infection (Fig 4A). We wondered whether this translocation was modulated by SCAP. This possibility was ruled out by the observation that the STING translocation was intact when silencing Scap (Fig 4C). Unexpectedly, HSV-1 infection also triggered SCAP to translocate from ER to the perinuclear microsome (Fig 4B). The Cell fractionation assay further revealed that HSV-1 infection induced both STING and SCAP to be predominantly in the microsome fraction (S5C Fig). Notably, SCAP was co-localized with STING both before and after the HSV-1 infection (Fig 1H and S5A Fig). We reasoned that the SCAP translocation is instead dependent on STING.
To test this hypothesis, we monitored the SCAP aggregation in wild-type and Sting-/- MEFs upon HSV-1 infection. As expected, SCAP congregated to perinuclear microsome in wild-type MEFs upon HSV-1 infection (Fig 4D). Interestingly, STING deficiency dramatically reduced the trafficking of SCAP to the perinuclear foci (Fig 4D). This translocation of SCAP was obviously rescued in the Sting-/- MEFs reconstituted with Flag-tagged STING (Fig 4D). In Mavs-/- or Tbk1-/- MEFs, SCAP could also congregate to perinuclear microsome (Fig 4D). Taken together, these data indicate that STING specifically facilitates the translocation of SCAP to the perinuclear microsome.
To explore the action of SCAP, we noticed that SCAP interacted strongly with both STING and IRF3, but not with TBK1 and p65 (Fig 5A). The C-terminal cytosolic domain of SCAP (736–1280 aa) was mapped to bind to IRF3 (201–357 aa) (S6A and S6B Fig), whereas the transmembrane region of SCAP (1–735 aa) was mediated to interact with STING (Fig 1G). So we speculated that SCAP might serve as an adaptor for recruiting IRF3 onto STING. The association of STING and IRF3 was enhanced in the presence of SCAP, whereas silencing of Scap almost abolished this association. In contrast, SCAP did not affect the interaction between STING and TBK1 (Fig 5B and 5C). In addition, ectopic-expression of SCAP promoted the endogenous association of STING and IRF3 in response to the HSV-1 stimulation (Fig 5D). Notably, knockdown of SCAP impaired the endogenous association of STING and IRF3 (S6C Fig). The endogenous interaction between SCAP and IRF3 was also confirmed, and this interaction was enhanced upon HSV-1 stimulation (Fig 5E). In sting-/- MEFs, the interaction of SCAP and IRF3 is markedly attenuated (S6D Fig), which indicates that SCAP recruits IRF3 after its STING-dependent translocation to the microsomes.
Notably, IRF3 also congregated to the perinuclear microsome and co-localized with SCAP upon HSV-1 infection (Fig 5F and S5B Fig). Importantly, silencing of Scap blocked the congregation of IRF3 (Fig 5F), whereas TBK1 deficiency did not affect the trafficking of IRF3 to the perinuclear foci (Fig 5F). Taken together, these data indicate that SCAP bridges STING to IRF3 in the perinuclear microsome.
To determine the importance of ER localization for SCAP function, we generated three mis-localization mutants of SCAP. SCAP-ΔTM was constructed by deleting the N-terminal transmembrane domain of SCAP. SCAP-Mito was constructed by replacing the transmembrane domain with the mitochondria targeting sequence from Tom70 (translocase of outer membrane 70, a mitochondria membrane protein). SCAP-NLS was constructed by replacing the transmembrane domain with a nuclear localization sequence (Fig 6A). Confocal microscopy analysis confirmed that SCAP-ΔTM, SCAP-Mito and SCAP-NLS were targeted to whole cell, mitochondria and nucleus, respectively (Fig 6B). As expected, SCAP-ΔTM, SCAP-Mito and SCAP-NLS failed to interact with STING (Fig 6C). SCAP-NLS also failed to interact with IRF3 (S7 Fig). Notably, SCAP mis-localization mutants could not enhance the association of STING and IRF3 (Fig 6D).
We further performed rescue experiments to corroborate the functional consequences. MEF cells were first transfected with control or Scap siRNAs, followed by transfection of the control or the RNAi-resistant rSCAP plasmids, respectively. The induction of Ifnb or Ifna4 mRNA was measured after ISD stimulation. Consistently, the induction of Ifnb or Ifna4 mRNA was restored by wild type rSCAP, but not rescued by SCAP mis-localization mutants rSCAP-ΔTM, rSCAP-Mito or rSCAP-NLS (Fig 6E). Taken together, these data indicate that the ER localization of SCAP is essential for its action in the STING signaling.
We went on to explore the antiviral function of SCAP in innate immunity. The induction of IFN-β is a hallmark of host antiviral responses. Scap siRNA was transfected into MEF cells, followed by HSV-1 (Fig 7A) or Listeria monocytogenes (Fig 7B) infection. The supernatants were quantified by ELISA (enzyme-linked immunosorbent assay). As expected, knockdown of endogenous Scap drastically impaired the IFN-β protein production (Fig 7A and 7B).
Since IFN-β protects host cells against virus infection, we assessed whether SCAP could restrict HSV-1 infection. MEF cells were respectively pretreated with culture supernatants from ISD-stimulated Scap knockdown cells or control cells, followed by HSV-1 (Fig 7C) or Listeria monocytogenes (Fig 7D) infection. It was observed that, fresh cells pretreated with culture supernatants from Scap knockdown MEFs were more permissive to HSV-1 or Listeria monocytogenes infection (Fig 7C and 7D).
In addition, we investigated whether SCAP attenuated microbial replications by challenging cells with HSV-1-GFP or Listeria monocytogenes. Consistently, knockdown of Scap augmented the levels of HSV-1-GFP-positive cells (Fig 7E), and also markedly enhanced the replication of Listeria monocytogenes (Fig 7F). In contrast, the replication of NDV-GFP was unaffected by depletion of Scap (S8 Fig).
Finally, we investigated the in vivo role of SCAP. We delivered into mice, via tail vein injection, the Scap specific or control shRNAs coated with polyethyleneimine (PEI). The efficiency of in vivo ‘knockdown’ was confirmed (Fig 7G and 7H). Next, mice were infected intravenously with HSV-1, and their survival rates were monitored. As expected, Scap-knockdown mice were more susceptible to HSV-1 infection than control mice. All the Scap-knockdown mice died within 3 days, whereas 50% of the control mice remained alive until 7 days after HSV-1 infection (Fig 7I). Notably, Scap knockdown mice displayed a severer defect in the production of sera IFN-β upon HSV-1 invasion, as compared with the infected control mice (Fig 7J). These data indicate that SCAP is indispensable for protecting mice against HSV-1 infection.
Recent breakthroughs have characterized multiple cytosolic sensors that potentially monitor the cytosolic DNAs (cGAS, IFI16, DDX41, Mre11 and DNA-PKcs). STING is established unambiguously as the converging point of the DNA sensors, to further relay the activation signals on ER [31,32]. Notably, STING is induced to dimerize and traffic from the ER, via Golgi, to perinuclear microsome [6,23]. TBK1 is simultaneously recruited to the same compartment in a STING dependent manner, which then activates the transcription factor IRF3 [23]. It is intriguing to dissect the molecular mechanisms of the DNA-driven assembly of the STING signalosome on either ER or the perinuclear microsome.
Our recent study [27] has characterized the AMFR/INSIG1 protein complex as a novel component of the STING signalosome on ER. The E3 ubiquitin ligase AMFR, bridged by INSIG1, catalyzes the K27-linked polyubiquitination of STING upon microbial DNA challenge. This unique polyubiquitin chain specifically recruits TBK1 and ferries the latter to the perinuclear microsome, along with STING. Notably, IRF3 could bind neither the K27- nor K63- linked polyubiquitin chain. This suggests that IRF3 is recruited onto the STING signalosome via another uncharacterized mechanism.
So we speculated that there might be unknown adaptor protein(s) on ER to perform this function. An analogy drew our attention regarding the translocations of the STING and SREBP. SREBP is a master transcription factor of the lipid and glucose metabolism on ER, which also translocates from ER to Golgi in response to metabolic stimuli [33]. Notably, SCAP is indispensable for chaperoning SREBP to Golgi, and INSIG1 specifically interacts with SCAP to prevent this translocation [30,34]. This inspires us to explore the potential function of SCAP in innate immunity.
Both STING and SCAP reside on ER via their N-terminal transmembrane domains. We confirmed the interaction between SCAP and STING, and demonstrated the transmembrane domains of both proteins to mediate this interaction. Unexpectedly, this interaction is independent of INSIG1, suggesting that the STING signalosome might be physically and functionally distinct from the Lipid Regulatory Complex. To substantiate, there is no obvious difference of IRF3 activation upon HSV-1stimulation, with or without FBS in the cell culture medium (S9F Fig). SREBP1 knockdown did not influence the exogenous DNA-induced IRF3 activation (S9B and S9C Fig). Furthermore, HSV-1 infection could not trigger SREBP1 translocation (S9D Fig). It was well established that the SREBP signaling was dramatically impaired respectively by three SCAP mutants (SCAP-Y234A, SCAP-Y298C, SCAP-D443N) [35,36]. However, the RNAi-resistant rSCAP-Y234A, rSCAP-Y298C, rSCAP-D443N could rescue the induction of IRF3-responsive genes in SCAP knockdown cells, like that of the wild type rSCAP (S9E Fig). Taken together, these observations indicate that the STING signaling is functionally and physically uncoupled from the SREBP signaling.
During revising this manuscript, York et al reported that the cell metabolic reprogramming could also induce the expression of IFNβ and ISGs in a STING-dependent manner [37]. Interestingly, some of the data in this paper suggested that SCAP was also important for this induction. However, they did not address how SCAP could potentially modulate the function of STING.
Notably, there are some discrepancies between the two studies concerning the specific effects of SCAP in regulating the induction of IFNβ and ISGs. York’s study was performed mainly in the context of metabolic reprogramming, whereas our study only focused on the microbe-induced activation of the STING signaling. In our experimental setting, we have supplied enough FBS medium to ensure that the cells will not experience metabolic reprogramming even when knocking down SCAP. In contrast, the indicated paper mainly addressed the metabolic reprogramming-induced expression of IFN-β and ISGs, which broadly affects the overall cell metabolism and dramatically influences the basal expression of IFN-β. Although the observation is interesting, we wondered if this cell model directly or indirectly activates the STING signaling, which needs further exploration in future experiments. For example, it is intriguing to address whether perturbing metabolism will lead to the dimerization and translocation of STING to the perinuclear microsome, whether this activation is dependent on cGAS, whether SCAP and IRF3 will congregate in this scenery. We speculate that different stimulating models could potentially lead to the observed discrepancies.
In addition, we extensively employed HSV-1 and Listeria Monocytogenes as effective stimuli to address the activation of STING signaling, which are well characterized microbes in the relevant fields. As far as we know, MHV-68 is scarcely employed in elucidating the STING signaling pathway. We do not know why the indicated study did not use HSV-1 or Listeria Monocytogenes. However, MHV-68 could potentially trigger the cGAS-STING signaling [38], and it was recently reported that some RNA viruses could also activate IRF3 in a STING-dependent manner [32]. We speculate that the different species of viruses might employ subtly different mechanisms to engage the activation of the STING signaling.
Functional analyses firmly established the essential role of SCAP in mediating the STING signaling. The phosphorylation of IRF3 stimulated by cytosolic DNAs is markedly impaired when knocking down the endogenous SCAP. Silencing the endogenous SCAP resulted in the impairment of the STING-mediated induction of IFNs and ISGs, and this effect is reversed by exogenously expressing siRNA-resistant SCAP. Silencing of Scap also impairs IFN-β protein production upon microbe infection, thus crippling the host antimicrobial responses against HSV-1 and Listeria monocytogenes. In vivo ‘knockdown’ of Scap induces less interferons and accelerates the death rate of the mice upon the HSV-1 infection.
We observed that HSV-1 infection triggers both SCAP and STING to traffic from the ER, via Golgi, to perinuclear microsome. Given that SCAP chaperons SREBP from ER to Golgi via the COP-II vesicle machinery, we had supposed that the translocation of STING is dependent on SCAP. However, knockdown of Scap did not affect the translocation of STING. Instead, the SCAP translocation is dependent on STING, but not on MAVS or TBK1, as evidenced in the corresponding knockout MEFs. The autophagy-related proteins (Atg9a and LC3) were implicated to modulate STING translocation [26]. It remains to address whether the autophagy proteins could modulate the action of SCAP.
Our study further provides a possible mechanism of regulating the STING signaling pathway by SCAP, which York et al. did not address. SCAP interacts individually with either STING or IRF3, via its N-terminal trans-membrane domains or C-terminal cytosolic domain respectively. The association between STING and IRF3 was respectively enhanced or impaired in response to HSV-1 stimulation, in the presence or absence of SCAP. Mis-targeting of SCAP, to whole cell, mitochondria or nucleus, resulted in its failure to bridge STING to IRF3. Notably, IRF3 also congregates to the perinuclear microsome after HSV-1 infection; and this congregation is dependent on SCAP, but not on TBK1. However, it is technologically challenging to demonstrate whether SCAP recruits IRF3 on ER or on perinuclear microsome. Recently, Dobbs et al. [39] have suggested that STING is retained on ER by some unknown inhibitor(s) in unstimulated cells. We think that such inhibitor(s) could not only block the translocation of STING/SCAP but also mask the IRF3 binding site on SCAP. Microbial infection releases this inhibition and triggers the translocation of STING and SCAP. Consequently, SCAP could recruit IRF3 to the STING signalosome. It is conceptually possible that IRF3 is blocked access to SCAP on ER by some “safe mechanism”, and SCAP on the perinuclear microsome is exposed to IRF3.
We had proposed an evolutionary necessity for assembling the STING signalosome on ER [27]. Arguably, the current characterization of SCAP further substantiates this perspective. We speculate that SCAP was dedicated primarily to mediate the activation of SREBP. Host innate immunity adapted it, along with AMFR and INSIG1, to integrate into the later-evolved STING signalosome. We predict that future studies will uncover more proteins essential for innate immunity that modulate the ERAD and/or Lipid/Glucose metabolism on ER. Deeper functional links between innate immunity and metabolism are expected to be displayed on the interface of ER.
Taken together, this study identified SCAP as the long-sought-after adaptor for recruiting IRF3 onto the STING signalosome. All the evidence favors the STING as an assembly platform, and the translocation of STING is the major cause of the other accompanying congregations. It remains to address what drives the STING translocation and whether the COP-II vesicle is required. The microenvironment of Golgi or microsome is probably favorable to the activation of TBK1 and IRF3. The underlying mechanisms remain elusive.
C57BL/6 mice 6–8 weeks old were purchased from the Shanghai SLAC Laboratory Animal Company. The mice were maintained under specific pathogen-free (SPF) conditions at the Shanghai Institute of Biochemistry and Cell Biology. Animal experiments were carried out in strict accordance with the regulations in the Guide for the Care and Use of Laboratory Animals issued by the Ministry of Science and Technology of the People’s Republic of China. The protocol and the procedures for mice study were approved by the Institutional Animal Care and Use Committee of the Shanghai Institute of Biochemistry and Cell Biology, Chinese Academy of Sciences (Permit Number: IBCB0027 Rev2).
SCAP, STING, TBK1, IRF3, INSIG1, cGAS, p65 were obtained by PCR from the thymus cDNA library and subsequently inserted into indicated mammalian expression vectors. The reporter plasmids (IFN-β-luciferase, pTK-Renilla) have been described previously [40]. The SCAP siRNA-resistant form was generated with silent mutations introduced into the siRNA target sequence. All point mutations were introduced by using a QuickChange XL site-directed mutagenesis method (Stratagene). All constructs were confirmed by sequencing.
The polyclonal antibody against STING was generated by immunizing rabbit with recombinant human STING (221–379 aa). The goat polyclonal antibody against SCAP was from Santa Cruz Biotechnology. The rabbit polyclonal antibody against SCAP was a gift from Dr. Xiongzhong Ruan (Chongqing medical university). hemagglutinin (HA), Myc, Ub, SREBP1/2 and IRF3 antibody were purchased from Santa Cruz Biotechnology. TBK1 antibody was from abcam. Tom20 antibody was from BD Biosciences. Flag and β-actin antibodies were obtained from Sigma-Aldrich. Phospho-IRF3 and Phospho-TBK1 antibody was from Cell Signaling Technology.
Poly(dA:dT) and lipopolysaccharide (LPS) was obtained from Sigma-Aldrich. Poly(I:C) was purchased from Invitrogen. Wild type HSV-1 and HSV-1-GFP were kindly provided by Dr. Wentao Qiao (Nankai University) and Dr. Chunfu Zheng (Suzhou University), respectively. Listeria monocytogenes (10403 serotype) was a gift from Dr. Youcun Qian (Institute of Health Sciences). TBK1 kinase inhibitor BX795 was purchased from InvivoGen. ISD (Interferon stimulatory DNA) was prepared by annealing equimolar amounts of sense and antisense DNA oligonucleotides at 95°C for 10 min before cooling to room temperature. Oligonucleotides used are as follows:
HEK293, HEK293T and MEF cells were obtained from the American Type Culture Collection (ATCC). The procedure for generating BMDMs (bone marrow-derived macrophage) has been described previously [41]. HEK293, HEK293T, Sting-/- MEF, Tbk1-/- MEF and Mavs-/- MEF cells were cultured in DMEM medium (Invitrogen) plus 10% FBS and 1% penicillin-streptomycin (Invitrogen). Transfection was performed with Lipofectamine 2000 (Invitrogen) according to the manufacturer’s instructions.
The siRNAs duplexes were synthesized from GenePharma. The sequences of siRNAs are shown as follows:
MEF cells were grown on coverslips in 12-well plate. After treatment with or without HSV-1, coverslips with the cells were fixed for 15 minutes with 4% formaldehyde in PBS and permeabilized in 0.25% Triton X-100 in PBS for another 15 minutes, following by using 5% BSA in PBS for 1 hour. Then, cells were stained with indicated primary antibodies followed by incubation with fluorescent-conjugated secondary antibodies. The nuclei were counterstained with DAPI (Sigma-Aldrich). For mitochondria staining, living cells were incubated with 300 nM Mito Tracker Red (Invitrogen) for 30 min at 37°C. Slides were mounted with fluorescent mounting medium (Dako). Images were captured using a confocal microscope (TCS SP2 ACBS; Leica) with a ×63 (numerical aperture 1.4) oil objective.
For immunoprecipitation assay, cells extracts were prepared by using lysis buffer (50 mM Tris-HCl pH 7.4, 150 mM NaCl, 0.5% Triton X-100, 1mM EDTA) supplemented with a protease inhibitor cocktail (Roche). Lysate were incubated with appropriate antibodies for 4 hours to overnight at 4°C before adding protein A/G agarose beads for another 2 hours. The beads were washed three times with the lysis buffer and eluted with SDS-loading buffer by boiling for 5 minutes.
For immunoblot analysis, the immunoprecipitates samples were subjected to SDS-PAGE. The separated proteins were then electrically transferred to a PVDF membrane (Millipore). Immunoblotting was probed with indicated primary and secondary antibodies. The protein bands were visualized by using a SuperSignal West Pico chemiluminescence ECL kit (Pierce).
Luciferase reporter assays were performed as described previously [42].
Isolation of Total RNA from indicated cells was performed by using TRIzol reagent (Invitrogen) according to the manufacturer’s instructions. Reverse transcription of purified RNA was performed using oligo (dT) primers. The quantification of indicated gene transcripts were performed by real-time PCR with using FastStart Universal SYBR GREEN MASTER MIX (Roche), and Gapdh served as an internal control. PCR primers of indicated target genes are shown as below:
Cell fraction was performed as described previously [23]. Cells were harvested by centrifugation at 1,000g for 10 min. The pellet was resuspended in sucrose homogenization buffer (0.25 M sucrose, 10 mM HEPES, pH 7.4), and cells were lysed by using a dounce homogenizer. Lysed cells were centrifuged at 500g for 10 min, and the supernatant was collected. The supernatant was then centrifuged at 10,300g for 10 min. The supernatant was crude microsomal (microsome and cytosol), and the pellet was crude mitochondrial (MAM and mitochondria). The crude microsomal fraction was subjected to ultracentrifugation at 100,000g for 60 min. The pellet was microsome fraction.
Concentrations of the cytokines in culture supernatants were measured by ELISA kit (PBL Biomedical Laboratories) according to the manufacturer’s instructions.
The shRNA was delivered into C57BL/6 mice with JetPEI transfection reagent (PolyPlus Transfection, San Marcos, CA) according to the manufacturer’s instructions [43,44]. The shRNA plasmid and JetPEI was each diluted into 100ml of 5% glucose, then mixed and incubated for fifteen minutes at room temperature at a final N/P ratio of 8. Finally, the mixture (200ml) was injected into each mouse via tail vein.
For analysis of in vivo ‘knockdown’ efficiency, mice were euthanized after forty-eight hours of in vivo shRNA transfection. The mRNA or protein of SCAP was respectively checked in leukocyte from blood or extracts from livers.
The control or Scap ‘knockdown’ mice were infected intravenously with HSV-1. The viability of the infected mice was monitored for 7 days. The mouse serum was collected at six hours after infection to measure cytokine IFNβ production by ELISA.
The shRNA was designed targeting mouse Scap 5′- GCT TAG AGC TGC AAG GCA A -3′ sequence.
Student’s t-test was used for statistical analysis of two independent treatments. Mouse survival curve and statistics were analyzed with log-rank (Mantel-Cox) test. P values of less than 0.05 were considered to be statistically significant.
The GenBank (http://www.ncbi.nlm.nih.gov/Genbank) accession numbers for the genes and gene products discussed in this paper are:
SCAP (NM_012235.2, NP_036367.2); STING (NM_198282.3, NP_938023.1); TBK1 (NM_013254.3, NP_037386.1); IRF3 (NM_001197122.1, NP_001184051.1); INSIG1 (NM_005542.4, NP_005533.2); cGAS (NM_138441.2, NP_612450.2); p65 (NM_021975.3, NP_068810.3).
|
10.1371/journal.pgen.1007909 | Dual role of DMXL2 in olfactory information transmission and the first wave of spermatogenesis | Gonad differentiation is a crucial step conditioning the future fertility of individuals and most of the master genes involved in this process have been investigated in detail. However, transcriptomic analyses of developing gonads from different animal models have revealed that hundreds of genes present sexually dimorphic expression patterns. DMXL2 was one of these genes and its function in mammalian gonads was unknown. We therefore investigated the phenotypes of total and gonad-specific Dmxl2 knockout mouse lines. The total loss-of-function of Dmxl2 was lethal in neonates, with death occurring within 12 hours of birth. Dmxl2-knockout neonates were weak and did not feed. They also presented defects of olfactory information transmission and severe hypoglycemia, suggesting that their premature death might be due to global neuronal and/or metabolic deficiencies. Dmxl2 expression in the gonads increased after birth, during follicle formation in females and spermatogenesis in males. DMXL2 was detected in both the supporting and germinal cells of both sexes. As Dmxl2 loss-of-function was lethal, only limited investigations of the gonads of Dmxl2 KO pups were possible. They revealed no major defects at birth. The gonadal function of Dmxl2 was then assessed by conditional deletions of the gene in gonadal supporting cells, germinal cells, or both. Conditional Dmxl2 ablation in the gonads did not impair fertility in males or females. By contrast, male mice with Dmxl2 deletions, either throughout the testes or exclusively in germ cells, presented a subtle testicular phenotype during the first wave of spermatogenesis that was clearly detectable at puberty. Indeed, Dmxl2 loss-of-function throughout the testes or in germ cells only, led to sperm counts more than 60% lower than normal and defective seminiferous tubule architecture. Transcriptomic and immunohistochemichal analyses on these abnormal testes revealed a deregulation of Sertoli cell phagocytic activity related to germ cell apoptosis augmentation. In conclusion, we show that Dmxl2 exerts its principal function in the testes at the onset of puberty, although its absence does not compromise male fertility in mice.
| DMXL2 gene dysfunction underlies various human diseases, including breast cancer, non-syndromic hearing loss, and polyendocrine-polyneuropathy syndrome, demonstrating the large range of potential actions of DMXL2. We show here that Dmxl2 expression is crucial for survival in mice, as neonates die within hours of birth when this gene is inactivated. The transmission of olfactory information is affected, leading to an absence of suckling and impaired feeding. Severe hypoglycemia is also observed in male neonates. We observed Dmxl2 expression in several organs, including the brain, heart and adrenal glands, potentially corresponding to some of the phenotypes observed in Dmxl2-deficient pups. We also described Dmxl2 expression in the reproductive tracts and gonads and showed that Dmxl2 inactivation specifically in the testes has a significant effect on the initial waves of spermatogenesis, resulting in very low levels of sperm production at puberty. Our results suggest that DMXL2 deficits should be considered in men with impaired fertility, as pathogenic variants of this gene may be associated with male infertility in humans.
| The DMX-Like 2 (DMXL2) gene encodes a protein with multiple WD40 domains, which mediate protein-protein and protein-DNA interactions [1], [2], [3]. These WD40 domains correspond to a series of 44- to 60-amino acid sequence repeats with a characteristic sequence terminating in a tryptophan–aspartic acid (WD) dipeptide [4]. WD40 proteins function as platforms, recruiting multiple partners for a wide range of cellular functions, such as signal transduction, vesicular trafficking, cell-cycle control, chromatin dynamics and the regulation of transcription [2], [3]. Little is known about DMXL2 and its various functions in different species. This protein was first described as rabconnectin-3α (or Rbcn-3α) in the rat brain, where it forms the stoichiometric rabconnectin-3 complex with WDR7 (rabconnectin-3β or Rbcn-3β) [5]. In the brain, DMXL2 has been implicated in the regulation of neurotransmitter exocytosis via its interactions with the small G-proteins Rab3-GEP and Rab3-GAP [1], [6]. The second main function attributed to DMXL2 is modulating vacuolar-ATPase proton pump (V-ATPase) assembly and activity, through interactions with several of its subunits [7], [8]. This function is highly conserved and has been described in numerous species, including yeast [9], [10], Drosophila [7], zebrafish [11], mice [12] and humans [13], [14]. V-ATPases acidify intracellular compartments (e.g. endosomes, lysosomes, and secretory vesicles) and they play a well-established role in protein sorting, trafficking and turnover in a wide range of signaling pathways, including the Notch pathway [15], [16]. DMXL2 has been shown to be involved in the Notch signaling pathway [7], [13], [14], [17] during diverse morphogenetic processes, such as the formation of ovarian follicles in the female gonads in Drosophila [7]. A role for this protein in reproductive function has also been suggested in humans, in which DMXL2 has been implicated in a complex syndrome of congenital hypogonadotropic hypogonadism (CHH) associated with polyneuropathy and glucose metabolism disorders [18]. Three brothers in the family concerned were affected and presented incomplete puberty with a low testicular volume. Their testosterone and gonadotropin (LH, FSH) levels were low, due to a dysfunction of gonadotropin-releasing hormone (GnRH) neurons. In mice, decreasing the level of Dmxl2 expression in neurons (Dmxl2wt/loxP ; nes-Cre) resulted in 30% fewer GnRH neurons [18] and an impairment of their activation and maturation [19]. The fertility of these mice also seemed to be impaired, consistent with a hypothalamic dysfunction.
Several clues to the functions of DMXL2 have emerged from studies of these models, but Dmxl2 knockout studies have never been reported. Furthermore, DMXL2 has never been directly implicated in the functions of reproductive organs in mammals. We show here that a complete loss-of-function of Dmxl2 results in neonatal lethality within a few hours of birth. Dmxl2-knockout pups present defects of olfactory information transmission associated with an absence of feeding. A glucose metabolism disorder was also detected, with male Dmxl2 KO pups displaying severe hypoglycemia. A description of the expression profile of Dmxl2/DMXL2 revealed a possible role in ovary and testis function. We therefore studied the specific effects of Dmxl2 knockout in the germ cells and supporting cells of the gonads of both sexes. Long-term fertility appeared to be unaffected in both sexes, but young male knockout mice produced less sperm than wild-type controls at the beginning of their reproductive life.
Heterozygous Dmxl2tm1a(EUCOMM)Wtsi mice were purchased from the IMPC. These mice carry a recombinant allele (tm1a) containing an IRES:LacZ trapping cassette and a promoter-driven neo cassette in intron 6 flanked by two FRT sites. Two loxP sites are also present, one at either end of exon 7 of Dmxl2. Transcription of the tm1a allele generates a truncated/abnormal Dmxl2 mRNA and leads to LacZ gene expression (Fig 1A) [20], [21], [22]. Crosses of Dmxl2tm1a(EUCOMM)Wtsi mice generated Dmxl2wt/wt (wild-type, WT), Dmxl2wt/tm1a (heterozygous, HTZ) and Dmxl2tm1a/tm1a (knockout, KO) offspring in Mendelian proportions. The pups were genotyped with a combination of primers provided by the IMPC (Fig 1B and S1 Table). Neither a functional Dmxl2 transcript (Fig 1C), nor DMXL2 protein (Fig 1D) was produced in the organs of newborn KO mice such as the brain, demonstrating the efficiency of the knockout. KO newborns died within 12 hours of birth. KO pups were of normal size and skin color at birth (Fig 1E), but they looked weaker than control littermates, with poor motility, and their stomachs contained no milk, suggesting that they did not feed.
We investigated the tissue expression profile of Dmxl2, by assessing LacZ reporter gene expression in mutants (whole-mount X-gal staining at E14.5 and P0) and by western blotting (at P0 and in adults). At E14.5, specific X-gal staining was restricted to the olfactory mucosa (Fig 2A). At birth (P0), prominent staining persisted in the olfactory mucosa (arrowhead) (Fig 2A) but specific X-gal staining was observed in other tissues (Fig 2B). Indeed, detailed dissection and fixation of the brain revealed faint, diffuse and scattered staining of both hemispheres of the cerebral cortex, confirming the expression of Dmxl2 in this organ [18]. The heart and kidneys were also strongly stained, and the gonads of both sexes displayed punctate staining, with enrichment in the cortical region of the ovaries, and in the seminiferous cords of the testes (Fig 2B). We also tested various organs, including the liver, pancreas, digestive tract, adrenal glands, and skeletal muscles. These tissues displayed no staining or only nonspecific staining, like the WT control tissues. Western blots comparing WT and KO organs confirmed that DMXL2 was expressed specifically in the brain, heart, kidney and gonads of both sexes at P0, whereas no expression was observed in the liver and pancreas (Fig 3A). In addition, DMXL2 expression was detected specifically in the adrenal glands at P0, but these glands displayed no X-gal staining.
In adults, the DMXL2 protein was detected in large amounts in the brain, as previously reported in rats [1], but also in the pancreas and adrenal glands (Fig 3B). In addition to the larger described isoforms that predominate (341 kD and/or 338 kD), two or three smaller isoforms (180 to 250 kD) were detected, depending on the tissue (and the level of DMXL2 expression in the organ concerned). Weaker, but clearly detectable expression was also observed in the testes, epididymides, seminal vesicles, ovaries, uterus horns and mammary glands (Fig 3C). Intriguingly, X-gal staining and western blotting demonstrated the presence of DMXL2 in the heart and kidneys at P0 (Figs 2C and 3A), but this protein was no longer detectable in these organs in adult animals (Fig 3B). We performed morphometric analyses of E18.5 Dmxl2 KO embryos to investigate possible developmental defects of the heart and kidneys that might explain neonatal death. However, these studies revealed no cardiac or renal malformations or any other organ defects capable of accounting for the premature death of these mice (observations made by Prof. Manuel Mark, IGBMC, Illkirch).
Dmxl2/DMXL2 has been implicated in glucose metabolism [18]. We therefore assessed the blood glucose and plasma insulin concentrations of newborn pups of the different genotypes (S1 Fig). These concentrations were normal for female KO pups, but male KO pups were hypoglycemic (pValue ≤ 0.001) relative to their WT and HTZ littermates, despite normal insulinemia. As hypoglycemia affected the male pups only and neonatal lethality displayed no sex bias, it appears unlikely that this feature is responsible for the lethality of Dmxl2 knockout.
Strong Dmxl2 expression was detected in the olfactory mucosa as early as E14.5, and the newborn KO pups did not feed. We therefore investigated the possible effects of the loss of function of this gene on the olfactory system.
The olfactory system is known to regulate feeding behavior [23]. We therefore performed electro-olfactography (EOG) to investigate the functionality of the olfactory mucosa in KO pups. EOG signals result from the activation of the olfactory transduction cascade in a population of neurons located close to the recording electrode. This transduction cascade occurs in the neuronal cilia in contact with their environment. We stimulated the olfactory mucosa with various odorants, at several concentrations (Fig 4A). The maximum amplitude of the response to odorants was measured and did not differ significantly between KO and HTZ (control) pups (Fig 4A and 4B). The olfactory mucosa was, therefore, functional, and the peripheral olfactory sensory neuron cilia of KO pups were as capable of odorant detection as those of control newborn mice.
DMXL2 is associated with synaptic vesicles in rat brain, potentially regulating their exocytosis and signal transmission [1], [11]. We therefore investigated whether the olfactory information generated in the olfactory mucosa was efficiently transmitted to the neurons of the olfactory bulb. Neuronal activation in the olfactory bulb after odorant stimulation was assessed by c-Fos immunodetection in HTZ and KO pups (Fig 4C and 4D). We observed significantly fewer c-Fos-positive neurons in the glomerular and external plexiform layers of KO olfactory bulbs than in those of HTZ bulbs, whereas neuron density was similar (Fig 4D). The synaptic transmission of the olfactory signal from the olfactory mucosa to the olfactory bulb was, therefore, significantly altered in the absence of Dmxl2.
LacZ staining and western-blotting experiments showed that Dmxl2/DMXL2 was expressed in the gonads of both sexes (Figs 2 and 3). Studies of its transcription during gonad differentiation revealed a dynamic profile, with increases at sex-specific stages (Fig 5A). Male and female gonads displayed similar levels of Dmxl2 transcripts at early stages of differentiation (E12.5), but Dmxl2 levels increased in the ovary a couple of days before birth (between E16.5 and E18.5: pValue = 0.019), reaching maximum values on P0, before decreasing slightly. Dmxl2 transcript levels remained low in the adult ovary. The first few days after birth correspond to the breakdown of germ cell nests and the formation of the first follicles in mice [24].
In the testes, Dmxl2 transcript levels remained constant during fetal and early postnatal development, but increased markedly with the initiation of spermatogenesis after P5 (Fig 5A). Indeed, Dmxl2 transcript levels had already increased by P15, when early-stage pachytene spermatocytes are observed [25], and they peaked at P28, when the first elongating spermatids are detected [25]. Dmxl2 transcript levels remained constant thereafter in the adult testes.
Having demonstrated the specificity of the DMXL2 antibody (S2 Fig), we used immunodetection methods to detect the DMXL2 protein at various stages of gonadal differentiation (P5 and P28). A strong signal was detected in the germ cell cytoplasm in both male and female gonads (Fig 5B). A faint signal was also observed in the supporting cells of both sexes (i.e. granulosa and Sertoli cells).
As KO pups died shortly after birth, gonad phenotype could be analyzed only on P0. The morphological appearance of the gonads of female and male KO mice was assessed by classical histology methods (S3 Fig) and by the use of several markers of germ cells and supporting cells. Gonad size, organization and general appearance were similar in KO and control gonads. In particular, KO ovaries had numerous germ-cell nests at the cortex and a few primordial follicles were starting to form, as in P0 control ovary (S4 Fig).
Transcriptomic analyses were performed at P0. Microarray analyses comparing KO and WT ovaries and testes highlighted only a few genes differentially expressed between KO and WT gonads: 51 genes for KO ovaries (S2 Table), and 12 for KO testes (S3 Table) (adjusted pValue <0.1). Four of these genes were differentially expressed in the KO gonads of both sexes, as confirmed by RT-qPCR analyses for Aph1b and Fez1 (S5 Fig).
Despite the small number of differentially expressed genes in KO ovaries, two gene clusters with significant enrichment scores were identified (DAVID analysis tool; enrichment score ≥1.3) [26]. One of these gene clusters related to stress responses (enrichment score = 1.42), whereas the other concerned WD40 proteins (enrichment score = 1.68). Indeed, in Dmxl2 KO ovaries, transcript levels for three other WD40 protein-encoding genes were affected according to the microarray data, which were confirmed by RT-qPCR for Coro2b, and Fbxw8 (S5 Fig). In addition, Coro2b transcript levels were found to be upregulated in KO testes. This upregulation was not detected in global analyses.
In conclusion, Dmxl2 loss-of-function at P0 induced the dysregulation of a larger number of genes in female than in male gonads. Nevertheless, the morphology of KO ovaries was unaffected, with primordial follicle formation occurring as in the control. We evaluated the effect of Dmxl2 loss-of-function at later stages, including spermatogenesis in the male gonad in particular, by generating mice with conditional knockouts of Dmxl2 in germ cells and/or in supporting cells of both sexes.
We obtained conditional knockouts of Dmxl2 by first generating Dmxl2loxP/loxP mice (exon 7 floxed) by crossing Dmxl2wt/tm1a mice with FlpO (FLP) recombinase-expressing mice (Rosa26-FlpO) (Fig 1A) [27]. We then used several lines of Cre-expressing mice: a Vasa-Cre line, to generate a conditional Dmxl2 KO in germ cells (Dmxl2loxP/-; Vasa-Cre: germ cell cKO) [28], Amh-Cre [29], to produce a conditional Dmxl2 KO in Sertoli cells (Dmxl2loxP/loxP; Amh-Cre: Sertoli cell cKO) and Amhr2-Cre [30] for conditional Dmxl2 KO in granulosa cells (Dmxl2loxP/loxP; Amhr2wt/Cre:: granulosa cell cKO). Double conditional knockouts (dcKO) were also generated, resulting in Sertoli and germ cell-specific Dmxl2 dcKO for males (Dmxl2loxP/-; Vasa-Cre; Amh-Cre) or granulosa and germ cell-specific Dmxl2 dcKO for females (Dmxl2loxP/-; Vasa-Cre; Amhr2wt/Cre). These single and double conditional knockouts were studied at various postnatal stages.
Females of the different genotypes were fertile. The histological features of the ovary were also similar between females of the different genotypes (S6 Fig).
In males, fertility tests performed until the age of six months showed no significant differences between Sertoli cell cKO (9.1 ± 2.6 pups per litter), germ cell cKO (9.4 ± 3.4 pups per litter), Sertoli and germ cell dcKO (9.6 ± 2.9 pups per litter) and control Dmxl2loxP/loxP (9.2 ±3.4 pups per litter) mice. In addition, histological analyses of testis sections from six-month-old mice of the various genotypes revealed no specific phenotype (S7A Fig), and sperm parameters were similar in Sertoli and germ cell dcKO and Dmxl2loxP/loxP control mice (S7B Fig). Dmxl2 expression increases greatly with the onset of spermatogenesis at puberty (Fig 5A). We therefore studied sperm parameters and testis differentiation at the end of the first wave of spermatogenesis.
We first assessed Dmxl2/DMXL2 transcript and protein levels in the testes of mice of the different genotypes. We found that the germ cells were the major site of Dmxl2/DMXL2 expression in adults (Fig 6A and 6B). Nevertheless, Dmxl2 transcript detection was completely abolished only in Dmxl2 dcKO testes (Fig 6C), which were therefore considered to display a testis-specific Dmxl2 KO. The sperm parameters of control (Dmxl2loxP/loxP) (n = 12), Sertoli cell cKO (Dmxl2loxP/loxP; Amh-Cre) (n = 11), germ cell cKO (Dmxl2loxP/-; Vasa-Cre) (n = 4) and Sertoli and germ cell dcKO (Dmxl2loxP/-; Vasa-Cre; Amh-Cre) (n = 4) mice were analyzed seven weeks after birth (Fig 7A). Sperm concentration was more than 60% lower in mice with no Dmxl2 expression anywhere in the testes (dcKO), and in mice lacking Dmxl2 only in the germ line (germ cell cKO), despite a normal testis/body weight ratio (S4 Table). The percentage of motile sperm was similar in the four mouse lines, demonstrating that only the total number of spermatozoa was affected. Stereological analyses of germ cell cKO and dcKO testis sections revealed a significantly larger fraction occupied by the Sertoli cell cytoplasm in the center of the seminiferous tubules than in WT sections, whereas the lumen area was significantly smaller (Fig 7B and 7C; S8 Fig). In addition, the seminiferous epithelium occupied a smaller area in dcKO testes than in control (pValue = 0.001) or germ cell cKO testes (pValue = 0.05), suggesting that the number of germ cells was smaller.
We then used an RNA sequencing approach to characterize the molecular consequences of the absence of DMXL2 expression in testes from seven-week-old animals, comparing dcKO (Dmxl2loxP/-; Vasa-Cre; Amh-Cre) and Dmxl2loxP/loxP control testes. RNA-sequencing identified 363 genes as differentially expressed in dcKO testes relative to control gonads: 161 genes were upregulated and 202 were downregulated (S1 File). We identified the cell types affected by Dmxl2 loss–of-function in the testes, by assessing the cellular expression profiles of the 363 differentially expressed genes based on RNA-sequencing data obtained from the Gene Expression Omnibus http://www.ncbi.nlm.nih.gov/geo/; accession number GSE43717; [31]) (S1 File). According to these data, Dmxl2 is weakly expressed in Sertoli cells and much more strongly expressed in the germ line, mostly in spermatogonia and spermatocytes, consistent with our findings (S1 File). The heat maps of the 161 upregulated genes (S9A Fig) and of the 202 downregulated genes (S9B Fig) were mirror images. Indeed, the genes upregulated in dcKO mice were mostly genes expressed by Sertoli cells and spermatogonia, whereas those downregulated were mostly genes expressed by spermatocytes and spermatids. Accordingly, the leading functional annotation for downregulated genes was “spermatogenesis” (Benjamini-Hochberg adjusted pValue = 5 x 10−4, DAVID6.8; S2 File). These observations, together with our previous stereological data, suggest a defect affecting the first wave of spermatogenesis, with smaller numbers of spermatocytes and spermatids produced in dcKO gonads. As the number of Sertoli cells conditions the number of spermatogenic cells, we determined the numbers of SOX9-positive cells (i.e. Sertoli cells) in dcKO and control testes (S10 Fig). No difference was found between the two genotypes. Additional in silico analyses on upregulated genes highlighted apoptosis (apoptosis signaling, pValue = 0.0035) and endocytosis (phagosome maturation, pValue = 0.0036; clathrin-mediated endocytosis signaling, pValue = 0.0123; macropinocytosis signaling, pValue = 0.0186) processes as canonical pathways significantly upregulated in dcKO gonads, which could be related to the decrease in the number of spermatogenic cells at this stage (Ingenuity Pathway Analysis software, Qiagen; S3 File). Furthermore, the ERK1/2 pathway, which has been implicated in apoptosis and phagocytosis [32], was the leading functional network detected, encompassing 30 of the 161 upregulated genes (almost 20% of the upregulated genes). Three of the five members of the TAM (Tyro3 Axl Merkt) regulation pathway were highlighted in this network [33]: one receptor Tyro3 and two ligands: Gas6 (growth-arrest-specific 6) and Pros1 (Protein S) (Fig 8A). This pathway has been reported to be involved in the phagocytic activity of macrophages, but also in that of Sertoli cells, in which it plays a crucial role in ensuring fertility [34], [35]. Transcript levels for the five members of the TAM family (the three receptors, Tyro3, Mertk, Axl, and their two ligands, Gas6, Pros1) were analyzed in the testes of seven-week-old mice of the four genotypes (control, Sertoli cell cKO, germ cell cKO and dcKO) (Fig 8B). Interestingly, the transcript levels of all these genes were significantly higher in dcKO than in control testes (pValue <0.05), highlighting a general enhancement of the TAM regulatory pathway and, potentially, of Sertoli cell phagocytosis in the absence of DMXL2 expression in the testes. This increase in phagocytic activity may reflect higher levels of germ cell apoptosis in dcKO testes. Cleaved caspase-3 (cCasp-3) expression was detected by immunohistochemistry and cCasp-3-positive cells were counted in dcKO and control testes (Fig 9A and 9B respectively). The proportion of apoptotic cells (cCasp-3-postive cells/mm2) was significantly higher in dcKO testes (pValue = 0.02), despite variations within animals.
In conclusion, in the absence of Dmxl2 expression in the testes, spermatogenesis appears to be less efficient, with higher levels of germ-cell apoptosis, whereas the phagocytic activity of Sertoli cells seems to be enhanced during the first wave of spermatogenesis.
We describe here, for the first time, the neonatal lethality of Dmxl2 KO (Dmxl2tm1a/tm1a) in mice, with a maximum survival of 10 to 12 hours after birth. This observation contrasts with that of a previous study in which Dmxl2tm1a/tm1a was reported to be lethal during fetal development [18]. We therefore introduced the Dmxl2 tm1a mutation into two different genetic backgrounds, the C57Bl/6N and FVB/NRj strains, in which neonatal lethality kinetics were similar. Our first observations of Dmxl2 KO pups revealed abnormal feeding behavior, with an absence of milk in the stomach. A complete inability to feed inevitably leaves to neonatal death, due to the absence of nourishment, and because the liquid derived from milk is essential for homeostatic processes in newborn [36]. Nevertheless, death linked to milk deprivation occurs within 12 to 24 h and therefore seems unlikely to be the sole cause of premature death in Dmxl2 KO mice. Morphometric studies revealed no major defect in organogenesis, but signs of neurological and metabolic problems were observed.
The tissue-specific expression of Dmxl2 has been analyzed during fetal and postnatal development. Dmxl2 expression is strongly detected in the olfactory mucosa of the fetus, an area closely associated with migrating GnRH neurons [37]. DMXL2 has been reported to be crucial for the number, activation and maturation of GnRH neurons [18], [19]. It therefore seems likely that the GnRH neuron defects observed in adult nes-Cre; Dmxl2 wt/loxP mice result from early fetal dysfunction. The transmission of olfactory information was impaired in Dmxl2 KO neonates, which displayed very poor olfactory bulb neuron activation. As olfaction is required for suckling behavior in rodents [23], this neurological transmission defect in KO pups may contribute to the absence of feeding. However, this observation is insufficient to account for the premature death of these pups, and this specific example of neurological failure probably reflects more severe impairments, as observed in human patients, in whom DMXL2 haploinsufficiency leads to polyneuropathy [18].
During gestation, fetal homeostasis is essentially managed by the placenta, whereas, after birth, the neonate must rapidly adapt to a stressful situation with new metabolic needs. A failure to establish energy homeostasis rapidly leads to the death of the neonate, over a period of time similar to that observed for Dmxl2 KO mice. Indeed, death may occur within eight hours of birth in some cases of glucose homeostasis problems [38], or between 10 and 14 hours after birth if autophagy mechanisms are disrupted [39], [40]. Interestingly, neonatal Dmxl2 KO pups present signs of metabolic/homeostasis problems, such as hypoglycemia in particular. During the first few hours after birth, the pup experiences a period of starvation during which gluconeogenesis is not fully active [41]. Glycogen is stored in the liver during fetal development, to prevent hypoglycemia in the neonate. Glycogen is an important source of glucose, which is released via glycogenolysis [42]. This adaptive mechanism is managed by a hormonal network, in which insulin levels decrease and the secretion of glucagon and glucocorticoids increases. DMXL2 was not detected in the pancreas or liver at birth, but was found in the adrenal glands, focusing attention on glucocorticoids and the possible effect of DMXL2 on their secretion. Intriguingly, hypoglycemia was observed only in male KO pups, highlighting the sex-specific nature of DMXL2 function in glucose homeostasis and, possibly, in adrenal sex-specific functions. However, as the timing of death was similar for both male and female KO pups, hypoglycemia cannot be the main event causing premature death. Autophagy is the first source of energy after birth, before efficient glycogenolysis is established. Autophagy levels are low during embryogenesis, but this process is upregulated in various tissues at birth (including the heart in particular) and is maintained at high levels from 3 to 12 hours after birth [39]. Autophagy eliminates aberrant or obsolete cellular structures/organelles. It is the primary means of degrading cytoplasmic constituents within lysosomes. Autophagy is also important for the cellular response to starvation, as the amino acids it generates can be used directly as a source of energy, or converted into glucose by the liver. Mice with deficiencies of ATG5 or ATG7 (autophagy-related proteins 5 and 7, respectively), which are involved in autophagosome formation, die within the first 12 hours of birth [39], [40], a timing very similar to that observed for Dmxl2 KO pups. Many WD40 proteins have been implicated in autophagy: Atg18 (autophagy-related protein 18) [43], EPG-6 (ectopic PGL granules 6) [44], AMBRA1 (autophagy/beclin-1 regulator 1) [45], ALFY (autophagy‐linked FYVE protein) [46] and WDR47 (WD40-repeat 47) [47]. A possible new function of the DMXL2 protein in autophagy should, therefore, be investigated, as it might explain the premature death of Dmxl2 knockout mice.
In addition to its role in neurological or homeostatic processes, DMXL2 was also thought to be associated with reproductive functions. Indeed, Dmxl2 is expressed in the germ cells and supporting cells of the gonads of both sexes, with a timing during development suggesting involvement in the major events of ovary and testis differentiation.
In the ovary, Dmxl2 expression increases after the onset of meiosis (E14.5) to reach a peak at birth. This period corresponds to germ-cell nest formation, breakdown and primordial follicle formation. However, the histological features of the ovaries were identical in Dmxl2 KO mice and controls at birth, as some primordial follicles were already visible in the gonads of both genotypes. Furthermore, cell-specific Dmxl2 knockout did not result in any fertility problems or ovarian abnormalities in adult females. Analyses of gene expression in Dmxl2 KO gonads at P0 revealed molecular disorders in the ovaries, which seemed to be under stress and trying to adapt to the loss-of-function of Dmxl2, possibly by increasing the expression of other WD40 protein-encoding genes (i.e. Coro2b and Fbxw8). More interestingly, the Aph1b gene was found to be upregulated in the gonads of Dmxl2 KO mice of both sexes. Aph1b encodes one of the four subunits of the γ-secretase complex, which plays a key role in the Notch signaling pathway, and this subunit is involved in the stability of the complex [48], [49], [50]. DMXL2 has been implicated in Notch signaling, in which it controls the V-ATPase pumps responsible for regulating the pH of the endocytic vesicles in which γ-secretase acts [51], [52], [13], [14]. In Drosophila ovaries, DMXL2/Rbcn-3α has even been shown to be involved in follicle formation, through its control of the Notch pathway [7]. The Notch pathway plays an important role in folliculogenesis that has been conserved during evolution, from flies to mammals [7], [53], [54]. Nevertheless, we show here that DMXL2 is not crucial for Notch signaling in mouse ovaries. The potential decrease in γ-secretase activity in ovaries lacking Dmxl2 is probably counterbalanced by an increase in the stability of the complex (via Aph1b upregulation). The important role of DMXL2/Rbcn-3α in folliculogenesis and female fertility is, therefore, not conserved in mice.
In testes, only a few other genes in addition to Aph1b were deregulated at P0 in Dmxl2 KO gonads. Nevertheless, Dmxl2 expression in the testes began to increase between P5 and P15, coinciding with spermatogenesis, suggesting a role in postnatal gametogenesis rather than early testis differentiation. Consistent with this hypothesis, histological/stereological observations and analyses of sperm parameters in the gonads of mice with a testis-specific Dmxl2 KO (Sertoli and germ cell dcKO) revealed a disruption of the first spermatogenic wave, resulting in a sperm concentration 60% lower than that in the controls. The seminiferous tubules presented an expended Sertoli cell cytoplasm, with a shorter lumen, suggesting higher levels of phagocytosis by the supporting cells. Transcriptomic and immunohistochemical analyses confirmed these observations, highlighting a decrease in the spermatocyte/spermatid fraction and an increase in apoptosis, accompanied by an increase in the levels of phagocytosis regulators. In particular, the TAM pathway was found to be upregulated in the absence of Dmxl2 expression. Three of the TAM proteins belong to the receptor protein tyrosine kinase (RPTK): TYRO 3, AXL and MERTK. Two related proteins, GAS 6 and Protein S (Pros1), act as their ligands. These five TAM proteins are expressed by Sertoli cells in the testes [34]. Males lacking the three TAM receptors (TAM-/-) are sterile due to an impairment of the phagocytic function of Sertoli cells, which is essential for the elimination of apoptotic germ cells [34], [35], [55]. TAM receptor dimers bind their two ligands, which in turn bind to the phosphatidylserine exposed at the surface of apoptotic cells [56]. In this study, germ cell apoptosis rates were significantly higher in the absence of DMXL2 (dcKO testes), suggesting that the TAM regulation (and thus, probably, the phagocytic activity of Sertoli cells) is enhanced in mutants due to germ cell dysfunction. Mice without DMXL2 expression in the germ line had a phenotype similar to that of dcKO mutants, with low sperm concentrations at puberty (Fig 7A). Together, these results suggest that DMXL2 exerts its principal function in germ cells, during the meiotic process occurring at the onset of spermatogenesis. Changes in its expression may affect germ cell differentiation, with higher rates of apoptosis and phagocytosis by Sertoli cells clearing abnormal spermatocytes/spermatids from the seminiferous tubules and resulting in a lower sperm concentration. Nevertheless, sperm production normalized at later stages of testis development, indicating that the functions of DMXL2 are essentially limited to the first wave of spermatogenesis or that compensatory processes occur after puberty. As suggested by Busada et al. for Rhox13 [57], Dmxl2 expression in spermatogenic cells may be advantageous in mice, supporting early fertility by providing additional germ cells at the start of the animal’s reproductive life.
Animals were handled in accordance with the guidelines on the Care and Use of Agricultural Animals in Agricultural Research and Teaching (Authorization no. 91–649 for the Principal Investigator, and national authorizations for all investigators). The protocol was approved by the Ethics Committee for Animal Experiments of the Jouy-en-Josas Institute and AgroParisTech (Permit Number: 12/184). Dmxl2tm1a(EUCOMM)Wtsi (Dmxl2wt/tm1a) mice with a C57Bl/6N genetic background were provided by the Wellcome Trust Sanger Institute (International Mouse Phenotype Consortium (IMPC): https://www.mousephenotype.org/data/genes/MGI:2444630). The mutation corresponded to a knock-in of the targeting vector between Dmxl2 exons 6 and 10 (see Fig 1) [20], [21], [22]. We backcrossed Dmxl2 mutant mice onto the FVB/NRj strain (JANVIER Laboratories) for 10 generations. Dmxl2wt/tm1a mice with this genetic background were crossed to generate Dmxl2wt/wt (wild-type, WT), Dmxl2wt/tm1a (heterozygous, HTZ) and Dmxl2tm1a/tm1a (knocked-out, KO) mice.
Conditional knockouts of Dmxl2 were obtained by crossing Dmxl2wt/tm1a mice with FlpO (FLP) recombinase-expressing mice (Rosa26-FlpO) [27] to remove the β-galactosidase cassette and the neomycin resistance gene and to create Dmxl2loxP/loxP mice (in which the Dmxl2 exon 7 is floxed, see Fig 1A). Dmxl2 conditional knockout in germ cells was achieved by crossing Dmxl2loxP/loxP mice with Vasa-Cre mice (FVB-Tg(Ddx4-cre)1Dcas/J) [28] to obtain Dmxl2wt/- ; Vasa-Cre mice in the F1 generation. Due to the mode of Vasa-Cre transmission, only F1 males (Dmxl2wt/- ; Vasa-Cre) of less than eight weeks of age were then used to generate Dmxl2loxP/- ; Vasa-Cre F2 mice (germ cell cKO).
The conditional knockout of Dmxl2 in granulosa cells was achieved by crossing Dmxl2loxP/loxP mice with Amhr2-Cre mice (Amhr2wt/Cre) [30] to obtain Dmxl2wt/loxP; Amhr2wt/Cre F1 mice. We then crossed Dmxl2loxP/loxP mice with Dmxl2wt/loxP ; Amhr2wt/Cre mice to obtain Dmxl2loxP/loxP ; Amhr2wt/Cre F2 mice (granulosa cell cKO).
Dmxl2 was knocked out specifically in Sertoli cells by crossing Dmxl2loxP/loxP mice with Amh-Cre mice [29]. The Dmxl2wt/loxP; Amh-Cre F1 mice were then crossed with each other to generate Dmxl2loxP/loxP ; Amh-Cre F2 mice (Sertoli cell cKO). We also generated double conditional mutant mice (dcKO) by crossing young Dmxl2loxP/-; Vasa-Cre males (six to eight weeks of age) with either Dmxl2loxP/loxP ; Amh-Cre or Dmxl2loxP/loxP; Amhr2wt/Cre females to produce Dmxl2loxP/-; Vasa-Cre; Amh-Cre males (Sertoli and germ cell Dmxl2 dcKO) or Dmxl2loxP/-; Vasa-Cre; Amhr2wt/Cre females (granulosa and germ cell Dmxl2 dcKO), respectively.
Mice were housed under a 12 h light/12 h dark cycle at the UE0907 (INRA, Jouy-en-Josas, France), with ad libitum access to food.
Genomic DNA was obtained from tail biopsy specimens with the Kapa Express Extract kit (Kapa Biosystems), according to the manufacturer’s instructions. Total-knockout animals were genotyped by PCR amplification of the Dmxl2 wild-type and tm1a LacZ alleles. Conditional-knockout animals were genotyped by the PCR amplification of Dmxl2 exon 7, and the Vasa-Cre, Amhr2-Cre and Amh-Cre alleles (see S1 Table for primer sequences and Fig 1 for the location of Dmxl2 primers). PCR was performed with the KAPA2G Fast Genotyping Mix, according to the manufacturer’s instructions (Kapa Biosystems).
For E14.5 embryos, maternal uterine horns were dissected out and transferred to cold 1 X PBS (Eurobio) for storage. The embryonic sacs were removed and used for genotyping. Embryos were rapidly rinsed in PBS, and fixed by incubation for 2.5 hours in a fixative solution containing 2% formaldehyde and 0.2% glutaraldehyde in PBS. The embryos were washed twice, for 30 minutes each, in PBS, and stained by overnight incubation at room temperature, in the dark, in 5 mM ferrocyanide, 5 mM ferricyanide, 20 mM MgCl2, 1 mg/ml X-gal (GX12836, Genaxis), 0.02% NP-40, 0.01% sodium deoxycholate, 20 mM Tris HCl pH 7.4. The following day, they were briefly rinsed and incubated in PBS for 30 minutes, before final fixation by incubation overnight at 4°C in 4% formaldehyde. The fixed embryos were rinsed in PBS and cleared as described by Schatz and coworkers [58].
For whole-mount staining, freshly dissected tissues (skinned heads, heart and lungs, digestive system (from the stomach to the large intestine), urogenital tracts, and skeletal muscle) from P0 animals were washed for 10 min in PBS supplemented with 0.01% Tween-20 (PBS-T), fixed by incubation for 10 min in 4% paraformaldehyde (PFA), and then subjected to two more washes in PBS-T, for 10 minutes each. For the brain and kidneys, pups were perfused with 4% PFA for 10 minutes and washed by incubation in PBS overnight at 4°C. All tissues were then stained by overnight incubation in 5 mM ferrocyanide, 5 mM ferricyanide, 4 mM MgCl2, 0.1% Triton X-100, 1 mg/ml X-gal, at 32°C, in a water bath. Tissues were washed for 10 min in PBS-T and then fixed again by incubation with 4% PFA overnight at 4°C. They were stored in 100% ethanol until imaging with a Leica M80 dissecting microscope fitted with a Leica DFC420 digital camera.
Tissues from newborn WT or Dmxl2tm1a/tm1a (Dmxl2 KO) mice (brain, heart, liver, pancreas, kidneys, adrenal glands, testes and ovaries), or from adult WT mice (brain, heart, pancreas, kidneys, adrenal glands, lungs, liver, spleen, skeletal muscle, testes, epididymides, seminal vesicles, ovaries, uterus and mammary glands) were collected and snap-frozen in liquid nitrogen. For protein extraction, tissues were ground on dry ice, and transferred to a Dounce homogenizer, in which they were lysed in radioimmunoprecipitation assay (RIPA) buffer supplemented with protease inhibitors (Roche). Lysates were centrifuged for 20 min at 4°C and 16,000xg, supernatants were collected and the amount of protein present was determined by the Bradford method.
For each tissue, we subjected 25 μg of protein diluted in Laemmli buffer to electrophoresis in 4–15% Mini-PROTEAN TGX gels (Cat. 456–108310, Bio-Rad). Stained proteins of known molecular weight (range: 31–460 kD, Cat. LC5699, Invitrogen) were run simultaneously as standards. The bands obtained on electrophoresis were transferred onto a polyvinylidene difluoride membrane (Hybond-P PVDF; Amersham). The membrane was blocked by incubation in phosphate-buffered saline containing 1/1000 Tween-20 (PBS-T; Prolabo, France) supplemented with 4% (w/v) nonfat dried milk, and was incubated overnight at 4°C with primary antibody (anti-DMXL2 or anti-GAPDH; refer to S5 Table for a list of the antibodies used and the conditions in which they were used) diluted in PBS-T supplemented with 4% (w/v) nonfat dried milk. The blot was then washed three times with PBS-T, incubated for 45 min in PBS-T supplemented with 4% (w/v) nonfat dried milk plus the peroxidase-conjugated secondary antibodies, and washed thoroughly in PBS-T. Peroxidase activity was detected with the ECL-Plus detection system for western blots, according to the manufacturer’s instructions (Amersham). Immunoreaction signals were analyzed with an image analysis system (Advanced Image Data Analyzer software, LAS 1000 camera, Fujifilm).
Electro-olfactogram (EOG) recordings were made on the olfactory mucosa in an opened nasal cavity configuration, on hemi-heads of newborn HTZ and KO mice, as previously described [59]. The hemi-head was kept under a constant flow of humidified filtered air (1000 ml/min) delivered close to the septum through a 9 mm glass tube. This tube was positioned 2 cm from the epithelial surface. The olfactory system was stimulated by blowing air puffs (200 ms, 200 ml/min) through an exchangeable Pasteur pipette containing a filter paper impregnated with 20 μl of the odorant, enclosed in the glass tube. The odorants used were diluted in mineral oil (hexanal from 1:10000 to 1:10; limonene from 1:1000 to 1:10 and carvone at 1:100). EOGs were recorded at two separate centrally located positions on turbinates IIb and IIa. EOG signals were analyzed and peak amplitudes were measured with a Clampfit 9.2 (Molecular Devices). Values were averaged for each set of conditions. Means ± SEM were plotted with GraphPad, and statistical analyses were performed with a Fisher-Pitman two-sample exact permutation test (R software using the Rcmdr.Plugin.Coin package (pValue<0.05)) to compare the response between HTZ and KO animals for a given concentration of odorant.
Newborn pups were isolated from their dams and placed in a new cage for 30 min in a quiet room. They were then exposed, for 10 minutes, to odorants in a tea ball containing filter paper impregnated with 20 μl of a mixture of 12 odorants (equimolar mixture of anisole, citral, heptanal, isoamyl acetate, lyral, lilial, octanol,1-4-cineol, isomenthone, limonene, carvone, and pyridine diluted to a final concentration of 10-3M). Pups were killed 60 min after the end of the exposure period. Heads were skinned, and prepared as described in the “Immunohistochemistry” paragraph for c-Fos immunodetection. For all coronal olfactory bulb sections, we took four dorsal and ventral images. Images were acquired blind to treatment, at a magnification of x100. They were analyzed with ImageJ (Rasband, W.S., ImageJ, U.S. National Institutes of Health, Bethesda, Maryland, USA, http://imagej.nih.gov/ij/) for the thresholding of specific c-Fos staining, as previously described [60]. The area of the olfactory bulb was measured after DAPI staining, for quantification of the proportion of the plexiform extern and the glomerular layer area displaying c-Fos staining. The same threshold was applied to all images from the same experiment on the same litter. DAPI staining in each area was also quantified, for the evaluation of neuron density. Median values were plotted with GraphPad, and statistical analyses were performed with Fisher-Pitman two-sample exact permutation tests (R software, using the Rcmdr.Plugin.Coin package (pValue<0.05)) to compare HTZ and KO animals for c-fos activation and neuron density.
Glucose concentrations were determined in blood samples from newborn pups on P0, with FreeStyle Optium Xceed Blood Glucose meters (Abbott). Glucose concentrations were determined after three hours of starvation (separation of the pup from its dam).
Insulin plasma concentrations were determined for each pup, with the Mouse Ultrasensitive Insulin ELISA kit, according to the manufacturer’s instructions (Alpco).
Median values were plotted with GraphPad, and statistical analyses were performed with Fisher-Pitman two-sample exact permutation tests (R software, using the Rcmdr.Plugin.Coin package (pValue<0.05)) to compare male and female pups of each genotype.
We studied Dmxl2 expression during gonad development, by extracting total RNA from pools of ovaries or testes at different developmental stages, with the RNeasy Mini or Micro kit (Qiagen), depending on the amount of tissue. Three biological replicates were prepared for each stage and sex. The Maxima First-Strand cDNA Synthesis Kit (Thermo Scientific) was used to synthesize cDNA for RT-qPCR from 200 ng of RNA. RT-qPCR was performed in triplicate for all genes with the Absolute Blue SYBR Green ROX mix (Thermo Scientific), in the StepOnePlus Real-Time PCR System (Applied Biosystems). Based on the output of the GeNorm program, we used ActB, and Ywhaz as the reference genes for this study (S1 Table). The results were analyzed with qBase Software [61].
Dissected tissues were fixed in by incubation in 4% PFA in PBS at 4°C for 2 hours (P0 and P5 gonads), overnight (P28 gonads) or for 24 hours (skinned heads of neonates). They were then cryoprotected with various concentrations of sucrose in PBS (0, 12%, 15%, and 18% for gonads, or 0, and 30% for heads). Tissues were finally embedded in Tissue-Tek O.C.T. Compound (Sakura Finetek Japan) and frozen at -80°C. Cryosections (7 μm for gonads, 20 μm for coronal head specimens) were cut and stored at -80°C. Sections were air-dried, rehydrated in PBS and permeabilized by incubation with 0.5% Triton, 1% BSA in PBS for 30 minutes. The tissue sections were then incubated with the primary antibodies (listed in S5 Table) overnight at 4°C (2 days for c-Fos). Sections were washed several times in PBS and then incubated with secondary antibodies for 45 min at room temperature (overnight for c-Fos). Finally, slides were rinsed in PBS, mounted in Vectashield mounting medium with DAPI (Vector) and images were acquired with a DP50 CCD camera (Olympus).
Freshly dissected P0 ovaries and testes were snap-frozen in liquid nitrogen. Three independent total RNA extractions were performed on pools of WT and KO ovaries and testes, with the RNeasy Mini (for testes) or Micro kit (for ovaries) (Qiagen). RNA quality was checked with an Agilent Bioanalyzer and 200 ng of total RNA for each set of conditions was hybridized with a Mouse WG-6 v2.0 Expression BeadChip (Illumina) (Pitié-Salpêtrière Postgenomics Platform–P3S, http://www.p3s.chups.jussieu.fr, Paris, France). Raw data were corrected for background by the “normexp” method, and quantile-normalized with the Limma package, through Bioconductor in the R statistical environment (version 2.15.0). Raw pValues were adjusted by the Benjamini-Hochberg method (false discovery rate) [62]. The quality of the expression data was checked by generating boxplots for raw expression data, density plots for normalized data, and by producing scatter plots and calculating Pearson’s coefficient for the correlation between arrays, with the Ringo package. The microarray data were assigned Gene Expression Omnibus number GSE115194 and are publicly available (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE115194). RT-qPCR validations were performed as previously described, on three independent pools of gonads per genotype (XX and XY, WT versus KO) and the results were normalized against two housekeeping genes (ActB and Ywhaz, according to GeNorm analyses) with qBase Software. Means ± SD were plotted with Excel, and statistical analyses were performed by ANOVA followed by Fisher’s LSD test in InVivoStat software [63].
Six-week-old male and female conditional mutant mice were paired with wild-type FVBN mice for a period of six months. Breeding cages were monitored daily and gestations, birth dates and litter sizes were recorded. At the end of the breeding trial, the gonads and epididymides were harvested and either snap frozen for molecular analysis or fixed for histological analysis.
The evaluation of male fertility was completed by the use of the IVOS I CASA system (Computer Assisted Sperm Analysis, at Hamilton Thorne Inc., Beverly, MA, USA) to assess semen motility at the ages of seven weeks and six months. The cauda epididymis was plunged into 100 μl of TCF buffer (Tris, citrate and fructose buffer) and swimming spermatozoa were collected after incubation for 30 minutes at 37°C. A 4 μl aliquot was placed in a standardized four-chamber Leja counting slide (Leja Products B.V., Nieuw-Vennep, the Netherlands). Ten microscope fields were analyzed on an automated stage, using the predetermined starting position within each chamber. Statistical analyses were performed with the mean values for these ten fields, for at least 500 cells. Each sample was analyzed twice (two different Leja wells). In total, 30 frames were captured at 60 frames/s, with software settings as follows: cell detection with a minimum contrast of 50, a minimum cell size of 4 pixels, and a cell intensity of 80; the cutoff value for progressive cells was 50 μm/s for VAP and 80.0% for STR. Slow cells were considered to be static and had a VAP cutoff of 7.4 μm/s and a VSL cutoff of 6.6 μm/s. Median values were plotted with GraphPad, and statistical analyses were performed with Fisher-Pitman two-sample exact permutation tests (R software, using the Rcmdr.Plugin.Coin package (pValue<0.05)) to compare the sperm parameters of Dmxl2loxP/loxP males and the various conditional mutants.
The dissected gonads were fixed in Bouin’s solution for 2 hours at room temperature for P0 gonads and overnight at 4°C for adult gonads. They were washed several times in 70% ethanol and then dehydrated in a series of solutions of increasing concentrations of ethanol (90%, 100%) and butanol (50%, 100%). The tissues were then embedded in paraffin and 5 μm-thick sections were cut. Hematoxylin and eosin staining (HE staining) was performed by standard protocols for studies of tissue morphology. Images were captured with a Pannoramic Scan II (3DHISTECH) digital slide scanner.
For males, adult testis sections from control (n = 6), Sertoli cell cKO (n = 6), germ cell cKO (n = 6) and double conditional KO animals (n = 4) were analyzed in more detail. The volume fractions of the lumen, the residual Sertoli cell cytoplasm and the seminiferous epithelium were estimated on 200 seminiferous tubules per sample with the P2 grid of Appendix B of the chapter 4 of “Unbiased Stereology” [64, 65]. For each experiment, medians values were plotted with GraphPad, and statistical analyses were performed with Fisher-Pitman two-sample exact permutation tests in R software, with the Rcmdr.Plugin.Coin package (pValue<0.05).
For immunohistochemical analysis, Bouin’s solution-fixed testis sections (5 μm) from seven-week-old mice (control and dmxl2 dcKO) were deparaffinized and subjected to antigen retrieval by heating in 0.01 M citrate buffer, pH 6.0 in a pressure cooker for 5 minutes. The sections were then incubated for 10 min in H2O2 (0.3%) and then for 30 min in a blocking and permeabilization buffer (PBS/1% BSA/0.5% Triton). The sections were incubated overnight at 4°C with primary antibodies (anti-cCasp3 and anti-SOX9 antibodies [66]; see S5 Table for the list of antibodies and dilutions used). The slides were then washed in PBS and incubated with a biotinylated anti-rabbit IgG for 45 minutes at room temperature. The primary antibody was omitted as a negative control. Antibody binding was detected with a Vectastain ABC kit (Vector Laboratories, PK-6100), and sections were counterstained with hematoxylin. Images were captured with a Pannoramic Scan II (3DHISTECH) digital slide scanner. Cleaved caspase-3-positive cells were manually counted on virtual slides obtained with the Pannoramic viewer (3DHISTECH software). Seminiferous tubules were outlined manually and their surface area was obtained by the Pannoramic viewer. The total number of cCasp-3-positive or SOX9-positive cells was divided by total seminiferous tubule surface area (μm2) and multiplied by 1,000,000 to obtain a number of positive cells/mm2. Cell counting was performed on 10 (for SOX9) or 30 (for cCasp-3) round seminiferous tubules (transverse sections) on testes from 4 different animals of the control and Dmxl2 dcKO genotypes. Median values were plotted with GraphPad, and statistical analyses were performed with Fisher-Pitman two-sample exact permutation tests (R software, using the Rcmdr.Plugin.Coin package (pValue<0.05)).
Total RNA was extracted from the testes of seven-week-old Dmxl2loxP/loxP (WT mice) (n = 3) and Dmxl2loxP/-; Amh-Cre; Vasa-Cre (n = 3) mice with the RNeasy Mini kit (Qiagen). RNA quality was checked with an Agilent Bioanalyzer and 1 μg of total RNA from each sample was sent to the High-throughput Sequencing Platform of I2BC (Gif-sur-Yvette, Université Paris-Saclay, France) for oriented library preparation and sequencing. At least, 50 million 75 nt reads were generated per sample (SRA accession: SRP149657). Sequence libraries were aligned with the Ensembl 89 genome, with STAR [67], and gene table counts were obtained by applying RSEM to these alignments [68]. Statistical analyses of differential transcript accumulation were performed with R version 3.0.0 (R Development Core Team, 2013) with the Bioconductor package DESeq2 version 1.0.19 [69]. Fold-changes in expression were estimated by an empirical Bayes shrinkage procedure, which attenuated the broad spread of fold-change values for genes with low counts with negligible effects on genes with high counts [69]. The pValues were adjusted for multiple testing by the Benjamini and Hochberg method [62], and those with an adjusted pValue ≤0.05 were considered to be significant (S1 File).
RNA-sequencing data providing information about the gene expression profiles of different testis cell types were obtained from the Gene Expression Omnibus (accession number GSE43717; [31]). FPKM files containing normalized RNA‐Seq data for purified Sertoli cells (GSM1069639), spermatogonia (GSM1069640), spermatocytes (GSM1069641), spermatids (GSM1069642) and spermatozoa (GSM1069643) were compiled and data concerning the genes differentially expressed in tDmx2 KO testes were extracted (S1 File). FPKM values were log2-transformed to produce heat maps (pheatmap: Pretty Heatmaps. R package version 1.0.8; Raivo Kolde (2015); https://CRAN.R-project.org/package=pheatmap).
RT-qPCR validations were performed as previously described, with total testis RNA extracted from five animals per genotype (seven weeks of age), and results were normalized against three housekeeping genes (ActB, Ywhaz and H2afz (S1 Table), selected on the basis of GeNorm analyses) with qBase Software. For each experiment, median values were plotted with GraphPad, and statistical analyses were performed with Fisher-Pitman two-sample exact permutation tests in R software (Rcmdr.Plugin.Coin package (pValue<0.05)).
|
10.1371/journal.ppat.1007890 | Life-long control of cytomegalovirus (CMV) by T resident memory cells in the adipose tissue results in inflammation and hyperglycemia | Cytomegalovirus (CMV) is a ubiquitous herpesvirus infecting most of the world’s population. CMV has been rigorously investigated for its impact on lifelong immunity and potential complications arising from lifelong infection. A rigorous adaptive immune response mounts during progression of CMV infection from acute to latent states. CD8 T cells, in large part, drive this response and have very clearly been demonstrated to take up residence in the salivary gland and lungs of infected mice during latency. However, the role of tissue resident CD8 T cells as an ongoing defense mechanism against CMV has not been studied in other anatomical locations. Therefore, we sought to identify additional locations of anti-CMV T cell residency and the physiological consequences of such a response. Through RT-qPCR we found that mouse CMV (mCMV) infected the visceral adipose tissue and that this resulted in an expansion of leukocytes in situ. We further found, through flow cytometry, that adipose tissue became enriched in cytotoxic CD8 T cells that are specific for mCMV antigens from day 7 post infection through the lifespan of an infected animal (> 450 days post infection) and that carry markers of tissue residence. Furthermore, we found that inflammatory cytokines are elevated alongside the expansion of CD8 T cells. Finally, we show a correlation between the inflammatory state of adipose tissue in response to mCMV infection and the development of hyperglycemia in mice. Overall, this study identifies adipose tissue as a location of viral infection leading to a sustained and lifelong adaptive immune response mediated by CD8 T cells that correlates with hyperglycemia. These data potentially provide a mechanistic link between metabolic syndrome and chronic infection.
| Mouse cytomegalovirus (mCMV) infection results in initial systemic viremia that is thereafter controlled by the adaptive immune system. Control is mediated in part by T cells that render the virus undetectable systemically, and latent in specific organs, including the lungs and salivary glands. It remains unclear how latent virus is controlled across tissues given the large pool of systemic mCMV-specific T cells. We explored mCMV control in the adipose tissue, whose cellular constituents are potentially susceptible to infection. We found that mCMV infects the adipose tissue during the acute phase, causing local inflammation and a lifelong mCMV-specific CD8 T cell immune response. The response consisted largely from non-recirculating, tissue-resident T cells. The infected adipose tissue showed signs of metabolic changes, that may potentially predispose the infected host to metabolic dysregulation as evidenced by hyperglycemia. Accumulation and persistence of mCMV specific non-circulating resident CD8 T cells (Trm) in adipose tissue reveal a likely generalized mechanism of mCMV tissue reservoir control by Trm cells and identify the adipose tissue as a persistent mCMV reservoir, with potential implications for metabolic health.
| Cytomegalovirus (CMV) is a ubiquitous beta-herpesvirus that infects most of the worldwide population, with largest prevalence being observed in older adults [1–4]. Acute CMV infection is characterized by system-wide viremia after which latency and lifelong persistence is established in select cells such as CD34+ monocytes and hematopoietic progenitor cells in humans [5–7]. Although CMV infections are generally asymptomatic, untreated infections in utero or amongst the immunocompromised individuals can result in substantial developmental defects, pathology, and death [5,8–10]. However, in immunocompetent patients the substantial and varied (NK, CD8, CD4, and B cells) resources are mobilized to successfully control viral spread and reactivation. One well described arm of anti-CMV immunity, the CD8 T cell compartment, is heavily involved in viral control with up to 5–10% of total CD8s in the blood and secondary lymphoid tissues reactive to CMV antigens during a primary immune response [11–14]. Moreover, in the course of lifelong infection, cycles of latency and reactivation drive an expansion of CD8 T cells, termed memory inflation (MI), that in some cases reaches up to 30–50% of the total memory compartment in mice and men [15,16]. The magnitude of this memory response is largely unparalleled in any other infection and for this reason CMV has been used as a model to understand the effects of MI during immune aging [3,17–20].
Studies of CMV-driven MI, viral dissemination, and persistence throughout the host have largely focused on the spleen, liver, blood, lung, and salivary glands [12,21–24]. It has become clear that the blood contains a major pool of CMV-reactive T effector memory (Tem) cells that presumably scan the vasculature as a bulwark against systemic CMV reactivation and that accumulate with age [3,25]. Tissue control of CMV has been narrowly studied in the context of the lungs and the salivary gland as these sites were shown to harbor mCMV-specific resident non-circulating T cell populations in response to latent virus. However, it remains unclear at this point whether tissue resident memory T cells are universally used to control a persistent pathogen such as CMV in situ and to what extent other tissue locations contribute to such a host defense and potential viral latency.
We therefore considered potential locations for where CMV could establish itself and hypothesized that adipose tissue, given several tissue specific properties, could offer a plausible site for infection. Adipose tissue is found at a variety of anatomical locations and consists of multiple cell types including adipocyte progenitors, leukocytes, and stromal cells, some of which have shown susceptibility to mCMV infection [26,27]. Furthermore, the adipose tissue of mouse and man is home to a large proportion of both innate and adaptive immune cells, suggesting that adipose tissue contributes to the mounting of an effective immune response [28–32]. The immune system represented within visceral adipose tissue has been clearly linked to development of diseases of the metabolic syndrome in the context of obesity [28,30,31,33–37]. Given the reported linkage between lifelong CMV infection and metabolic dysfunction in humans we reasoned that mCMV could potentially drive inflammation within adipose tissue that contributes to these phenotypes [20,38]. In further support of this hypothesis, adipocytes have been demonstrated to be susceptible to adenovirus and parasitic infection, raising the possibility that infection could drive insulin resistance and glucose intolerance [39–42]. Furthermore, adipose tissue of HIV/SIV infected humans/monkeys harbored latent virus even after patients were declared virus free following retroviral treatment, suggesting that adipose tissue can provide a safe haven to viruses [43].
Lifelong herpesvirus infections have been studied for their contribution to global host inflammation in the context of frailty and immune aging, but the consequences of these infections on, and the control of infection within adipose tissue have not been investigated. Given the susceptibility of individual cells within adipose tissue to CMV infection we hypothesized that mCMV could infect a cellular constituent of adipose tissue and therefore trigger in an inflammatory immune response in situ. We show here that mCMV infects adipose tissue during early peak viremia, followed by infiltration by mCMV-specific CD8 T cells during the acute phase post infection (p.i.). The adipose T cells were specific for mCMV epitopes, with a large fraction of them possessing a Tem phenotype. The presence of inflammatory monocytes was not necessary for the mCMV-specific immune response in adipose tissue, suggesting direct viral spread to adipose tissue during initial infection. mCMV infection and the resulting anti-mCMV CD8 T cells were associated with persistent inflammation within the adipose tissue from very early points during the immune response (days three to five) through the lifespan of the infected host (greater than 450 days). Moreover, infected fat tissue exhibited a decreased production of adiponectin. Finally, an analysis of lifelong-infected animals revealed that the mCMV specific CD8 T cells were bona fide Trm (CD69+) cells that exhibit limited exchange with the vasculature based on intravenous staining. The presence of both mCMV and mCMV-specific non-circulating resident CD8 T cells in the adipose tissue for life suggests that Trm cells may be the primary mechanism by which the host controls mCMV tissue reservoirs and their association with inflammation suggest that this interaction may alter metabolic health in infected animals.
mCMV was reported to infect cells in the adipose tissue [44–46], but the local consequences of this infection in vivo have not been well characterized. We first determined if adipose tissue was susceptible to mCMV infection by comparing both RNA (S1A Fig) and DNA (S1B Fig) extracted from total adipose 3 days (d) post infection (p.i.) to uninfected counterparts, we also analyzed the presence of viral genomes in several other tissues and visceral adipose tissue had a two log higher burden at day 3, when compared to subcutaneous adipose, liver, and spleen, likely due to route of infection (S1B Fig). Analysis of RNA transcripts revealed the presence of immediate early (IE) viral gene products, IE1, in infected but not uninfected animals (Fig 1A). To determine if the presence of mCMV transcript resulted in an immune response, we measured leukocyte infiltration and found at 3d p.i. a significant increase in the global leukocyte population, which was driven by lymphocytes, neutrophils, monocytes, and eosinophils (as normalized per gram of adipose tissue, Fig 1B). We further characterized the cells participating in this early, 3d p.i., adipose response by flow cytometry and found a significant expansion of NK cells (NK1.1+CD3-; S2A Fig) and an expansion, although not significant, in the macrophage population (F4/80+CD11b+; S2B Fig). Polarization of macrophages to an inflammatory, classically activated M1 phenotype (CD11c+) also trended higher (S2C Fig). We next examined the cytokine content of total adipose tissue homogenate using a flow cytometry-based LegendPlex platform and found significant increases in IFNγ (p = 0.004), TNFα (p = 0.001), IL-1α (p = 0.0007), IL-6 (p<0.0001), and CCL2 (p<0.0001) (Fig 1C–1G) with infection in the adipose tissue on days 3 and 5 p.i. (days 3 and 5 are pooled in the figure). However, at this time point we did not observe any significant changes in the amounts of GM-CSF, IFNβ, IL-1β, IL-12, IL-17A, IL-23, or IL-27. Therefore, during the early acute timepoints of infection, we found evidence of virus transcription, cellular and cytokine immune responses in the total adipose tissue.
Following infection, mCMV replicates systemically, leading to detectable host viremia that is resolved within days of infection. It is possible that the inflammation observed in the adipose at this early timepoint could resolve during reduction of the viral load. On 7d p.i. we were able to detect mCMV viral products (S1A Fig). Given this detection we wondered if there was a continued immune response within adipose tissue. We therefore quantified inflammation during the peak adaptive immune response at the same time, again by characterizing the cellular content of the stromal vascular fraction (SVF). The global leukocyte population of infected animals (Fig 2A, left) remained elevated, driven by lymphocytes, neutrophils, monocytes, eosinophils, and basophils (Fig 2A, right). Contributing to this expansion were NK cells which were still significantly elevated (S2A Fig). Total macrophage population became significantly expanded in infected adipose tissue at this time as well (S2B Fig). Approximately half of all leukocytes detected in adipose tissue at this time were lymphocytes, and a majority of these were CD3 T cells, with significant increases in both the CD4 and CD8 (p = 0.0004) populations (Fig 2B). Both the central memory (CD44+CD62L+; p = 0.012) and effector memory (CD44+CD62L-; p = 0.004) populations were significantly increased in infected animals, whereas, as expected, there was no change in the total numbers of naïve (CD44- CD62L+) T cells (Fig 2C). Amongst the memory subsets we found roughly equivalent numbers of memory-precursor effector cells (CD127+KLRG1-; MPECs; p = 0.0028) and short-lived effector cells (CD127-KLRG1+; SLECs; p = 0.0004), with both populations significantly increased in infected animals. We next asked whether these T cells were specific for mCMV antigen or were accumulating due to perhaps inflammatory stimulation in a non-specific manner. Peptide:MHC (pMHC) tetramer staining revealed a significant expansion of T cells specific for the acute, non-inflationary immunodominant epitope M45 (p = 0.0004), with smaller, but also highly significant expansion of CD8 T cells specific for inflationary epitopes m38 (p = 0.0004) and m139 (p = 0.0004) (Fig 2E). Given an influx of antigen specific T cells we also tested whether the cytokine milieu was still altered, and found, by LegendPlex, significant increases in the protein levels of IFNγ (S3A Fig; p = 0.004) and CCL2 (S3B Fig; p = 0.0005), but not of GM-CSF, IFNβ, IL-1α, IL-1β, IL-6, IL-10, IL-12, IL-17A, IL-23, IL-27, or TNFα. In addition to these protein analyses we found several inflammatory transcripts that were significantly upregulated at this time point, including Cd3e, Ifng, Cxcr3, Ccr5, Casp1, and Adgre1, indicative of a myeloid and T lymphocyte infiltration (S4 Fig).
Adipose tissue inflammation has been demonstrated to alter adipocyte derived cytokines, adipokines, during the onset and establishment of obesity [47,48]. We therefore examined the production of two well described adipokines, leptin and adiponectin. We found that adiponectin, which is decreased during inflammation-induced by obesity [49], was significantly decreased in infected animals (S5A Fig; p = 0.0186), whereas, we found no change in the amount of leptin, which is positively correlated with body weight (S5B Fig) [50]. No significant change in total leptin, however, comes as no surprise as we saw no significant change in total epididymal fat pad or body weight during the lifespan of infection (S6 Fig). These data suggest that inflammation driven by infection and influx of immune cells can trigger a secretory response by adipocytes. Taken all together, these data indicate that mCMV infection results in a CD8 T cell response detectable in adipose tissue at 7d p.i.
CCR2+ cells are believed to be the major carriers of mCMV, involved in virus dissemination [51,52]. Moreover, intraperitoneal injection, used in our experiments, may result in an indiscriminate and non-physiological distribution of the virus, including to the epididymal adipose tissue. To assess whether the virus exhibited true tropism for adipose tissue or perhaps infected the immunological constituents of adipose tissue regardless of the route of infection, we infected animals using 105 pfu of mCMV via the footpad (f.p.) route of infection. We considered two possibilities in which this receptor could be required 1) that CCR2 is required for homeostatic seeding of CCR2+ infected cells from the periphery into the adipose tissue, resulting in “reinfection” of adipose; or 2) that CCR2+ patrolling cells are not required to maintain antigen and therefore the presence of T cells in situ. We also wondered whether CCR2+ cells were needed to spread the virus during normal homeostatic seeding of adipose. To that effect, we infected both C57BL/6 and CCR2-/- mice [53] via the f.p. route. Surprisingly, when we analyzed adipose tissue at days 3 and 7 p.i. of C57BL/6 and CCR2-/- mice we were unable to detect viral product in either strain of mouse, perhaps indicative of virus being below the limit of detection or unable to spread to adipose tissue through this route. However, at 7d p.i. when we analyzed adipose tissue by flow cytometry, we found a significant increase in CD3 T cell numbers driven by a significant expansion of CD8 T cell in adipose of both wildtype and CCR2-/- mice (Fig 3A). Of interest, we saw a very limited expansion of central memory CD8 T cells (CD62L+CD44+) in the adipose when infected through this route in both wildtype and CCR2-/- mice, while the significant expansion of effector memory cells (CD62L-CD44+) was not diminished (Fig 3B). When we assayed tetramer specificity, we found a significant expansion in both acute, M45, and inflationary, m139, epitopes (Fig 3C). Finally, we analyzed the MPEC and SLEC populations and found that MPECs did not significantly expand whereas SLECs did (Fig 3D). These results suggest that mCMV infection, perhaps even at extremely low levels of viral burden, leads to the accumulation of mCMV-specific T cells in the adipose tissue regardless of the route of infection or in the presence of CCR2 (and, presumably, of CCR2+ cells). However, it should be noted that memory precursors and central memory cells appear to be insensitive to this route of infection (Fig 3B and Fig 3D).
We then wished to determine if there was a requirement for CCR2 to maintain viral-specific T cells within adipose tissue given the lack of central memory and memory precursors through f.p. infection. We therefore infected as before, via the footpad route, CCR2-/- mice and quantified the mCMV-specific T cell in adipose at an early memory timepoint post infection, greater than 30d p.i.. We found an overall diminished T cell population within the adipose of infected animals with no single T cell subset being significantly increased during infection (Fig 3E). Just as in the 7d p.i. timepoint, we saw no significant expansion of central memory CD8 T cells and found a trend of increase in the effector memory pool (Fig 3F). However, even in the absence of this expansion of global T cell populations, there was a significant increase in the mCMV M45- and m139-specific subsets (Fig 3G) as well as an expansion of both MPECs and SLECs at this time (Fig 3H).
Taken together, we conclude that the f.p. route of infection leads to undetectable viral load in adipose tissue, which results in no expansion of central memory and memory precursor CD8 T cells. Nonetheless, effector and short-lived effector populations do expand, albeit to a lesser degree than i.p.. Finally, tetramer specific T cells still arise and persist in adipose tissue regardless of the absence of CCR2 and route of infection.
To assess the impact of mCMV infection in the adipose tissue over the lifespan, we analyzed the fat pads of lifelong infected animals. During the lifelong time points of infection (>450d p.i.) we were unable to detect mCMV RNA. We therefore looked at the maintenance of mCMV genomes in the adipose tissue of infected animals via qPCR. To get a better resolution on viral genome loads, we initially compared CD45- non-hematopoietic and CD11b+CD45+ myeloid cell pools after FACs sorting (S7A Fig). Since the CD45- fraction showed a trend towards a higher mCMV genome burden (S7B Fig), we focused on these cells in subsequent kinetic analysis. mCMV genomes persisted in the adipose CD45- tissue at comparable levels from 90 to approximately 300+ days post infection, suggesting a lifelong presence of the latent and/or reactivating virus in the adipose tissue of infected animals (S7C Fig). We next went on to characterize the immunological response at these late time points. We found that total leukocyte counts in the adipose tissue were no longer significantly elevated, showing just a trend (Fig 4A). However, there remained a significant increase in CD3 T cells in infected animals (p = 0.0407), which was entirely driven by a robust expansion of CD8 T cells (Fig 4B; p = 0.0011), with a dominant and significant increase in Tem CD8 cells (Fig 4C; p<0.0001), and a stronger skewing towards SLECs (p<0.0001) over MPECs (p = 0.0069) compared to the acute (7d p.i) infection (Fig 4D). At this point NK cells and macrophage populations in infected animals mirrored that of their aged-matched counterparts (S2 Fig), suggesting that perhaps viral control of mCMV in adipose at these later time points is more reliant upon T cells.
To examine whether CD8 T cells in the adipose tissue were recirculating from the systemic pool, we performed two experiments. First, we analyzed expression of CD69 on CD8 T cells. This molecule, often used as a marker of immediate activation, is an antagonist of the S1P1 receptor, leading to the retention of T cells in their specific tissue [54]. We found that 75% of all CD8 T cells expressed CD69 in infected animals, a significant increase compared to that of their uninfected counterparts (Fig 4E; p = 0.0015). We also analyzed the dual expression of CD103e, which has been used to define tissue resident cells in other tissues [54,55] and found no significant differences between its expression on CD8 T cells in infected and uninfected animals (S8 Fig). Second, to independently test whether and how many CD8 T cells in the adipose tissue may be of resident memory type, we assessed their accessibility to a systemic anti-CD45 antibody injected into the vasculature in vivo, as a measure of their vascular vs. tissue-resident location. We injected an Alexa Fluor 700 labeled anti-CD45 antibody intravenously (i.v.) into lifelong infected animals, harvested the adipose tissue 5 min later, as previously described, to determine the extent of T cell tissue residency [56]. We found that approximately 95% of all T cells in adipose tissue of infected (as well as uninfected) animals stained only with the ex vivo antibody and therefore could be classified as resident to adipose tissue (Fig 4F).
To ascertain that T cells in lifelong infected animals are specific for mCMV antigens, we repeated the tetramer staining as performed in earlier timepoints. We found that a majority of CD8 T cells within adipose tissue remained specific for mCMV tetramers, with an expected and significant expansion of T cells specific for the inflationary T cells epitopes, m38 (p<0.0001) and m139 (p<0.0001). A much smaller, but also significantly expanded population was specific for the acute M45 epitope (p = 0.0018), possibly indicative of recent viral reactivation (Fig 4G). Taken together, these data demonstrate that mCMV-specific CD8 T cells are maintained within the adipose tissue for the lifespan of infection, as bona fide Trm cells.
The presence of phenotypically active mCMV specific T cells in adipose tissue provides evidence of a continued surveillance against mCMV. Next, we investigated whether this significant presence of mCMV-specific Trm cells within adipose tissue over the lifespan may be associated with persistent inflammation. We found that IL-23 (p = 0.0201), IL-1α (p = 0.0071), IFNγ (p = 0.0113), TNFα (p = 0.0258), CCL2 (p = 0.0083), IL-6 (p = 0.0083), IL-27 (p = 0.0109), and GM-CSF (p = 0.0175) were all significantly elevated in lifelong infected adipose tissue when compared to uninfected age matched controls (Fig 5A–5H). By contrast, IFNβ, IL-1β, IL-10, IL-17A, and IL-12 did not exhibit significant changes when compared to uninfected animals. These data indicated that adipose tissue is a site of lifelong accumulation, or maintenance, of mCMV-specific Trm cells that exhibit phenotypic evidence of recent antigenic stimulation, and that this correlates with an inflammatory cytokine response over the entire lifespan of the host.
Adipose tissue inflammation associated with obesity has been clearly linked with multiple phenotypes of the metabolic syndrome, including glucose intolerance and insulin resistance [44,57,58]. Based on the observed increase in inflammatory cytokines and cytotoxic T cells in lifelong infected animals we hypothesized that infected mice could exhibit an altered metabolic profile. Indeed, we found that between ten- and twelve-weeks post infection there was an elevation of fasted blood glucose in infected animals (Fig 6A) with no significant differences between infected and uninfected animals in plasma insulin levels (Fig 6B). To determine if the hyperglycemia was correlated with increased adiposity of infected animals, we longitudinally followed mice and analyzed the weight of their fat pads and found no significant change in fat pad weights between infected and uninfected mice (S6A Fig). When we calculated the homeostasis model assessment insulin resistance (HOMA-IR) index and found that infected animals exhibited significant elevation of this index compared to uninfected animals (Fig 6C; p = 0.0155). Conversely, the inverse of HOMA-IR, the insulin sensitivity index, expectedly suggested that infected animals were less sensitive to insulin than uninfected controls (Fig 6D; p = 0.0155). At >450 days post infection, significantly elevated levels of blood glucose were still observed in infected animals (Fig 6E; p = 0.0006) and this occurred in the absence of a significant increase in body weight (S6B Fig). This elevation in fasted blood glucose appeared to be dependent on mature CD8 T cells as we found no significant differences between the fasted blood glucose of chronically infected and uninfected mice lacking beta-2-microglobulin (B2m KO) (S9 Fig).
Based on the results of the HOMA-IR analysis of mice between ten- and twelve-weeks post infection we analyzed more broadly the metabolic system of lifelong infected animals. First, we determined the extent to which infected and uninfected animals clear a bolus of glucose by i.p. challenge. We found that mCMV infected animals did not clear glucose from the blood as quickly as uninfected counterparts (Fig 6F). We next tested whether the HOMA-IR was an accurate representation of insulin resistance in our model. Therefore, we i.p. challenged with fast acting insulin to determine insulin sensitivity and found no significant differences between infected and uninfected animals (S10A Fig). Finally, we wondered if elevated fasted blood glucose indicated a hyperactive gluconeogenesis driven by the liver. We therefore challenged mice with sodium pyruvate i.p. to determine liver sensitivity to infection and found no difference in gluconeogenesis in infected and uninfected animals (S10B Fig).
Taken together these data are consistent with recently published work that suggests alterations in glucose tolerance and insulin sensitivity in mice acutely infected with mCMV and influenza infection of mice being fed a high fat diet [59]. When we measured adiponectin expression in adipose tissue homogenate, we expected decreased amounts of total protein as we saw in the acute time point post infection, however we found no significant change in the amount of adiponectin protein in uninfected and lifelong infected animals (S11 Fig), perhaps indicating an age related decrease in adiponectin expression that may mask changes induced by infection. Overall, we found a clear initial alteration in the glycemic profile of mCMV-infected mice following infection that appears to be driven by delayed glucose clearance in infected animals and is possibly dependent upon mature T cells. Additional studies will be required to mechanistically extend these data, one is tempted to speculate that mCMV may make animals susceptible to clinical metabolic changes pending action of other environmental stressors, including diet, as previously published, and aging in our model.
CD8 T cell immunity against mCMV infection has been extensively studied in the context of inflationary memory T cell expansion in the blood, as well as in the lungs (as the port of CMV entry) and salivary gland (as the site of intense primary CMV replication). Results of these studies have suggested that the blood contains a large pool of CMV-specific circulating Tem cells, guarding against potential systemic reactivation, whereas both the site of primary entry (lungs) and extensive initial replication and shedding (salivary gland) contain Trm cells standing guard against potential reinfection (lung) and/or local reactivation (lung and salivary gland). CMV is believed to infect many cells, but to establish latency only in very few [7,60,61]. In that context and in the context of an early and lifelong CMV infection and immunity, we know very little about CMV-specific CD8 T cell immunity and control in other tissues. For example, one fundamental question remains on whether the large systemic circulating CD8 T cell pool is responsible for the control of other potential sites of latency and reactivation.
Recent studies show that white adipose tissue is enriched in leukocytes, including a significant population of memory T cell populations even in mice housed under specific-pathogen free conditions [31,62,63]. Furthermore, it has been demonstrated that pathogen-specific T cells can arise in both mesenteric and epididymal adipose tissue following bacterial and parasitic infection [43,62]. It has also been demonstrated that murine adipose tissue can harbor infectious mCMV as demonstrated by plaque assay and microscopy during early time points post infection [45,59,64]. We show here that adipose tissue is an early site of infection which leads to generalized inflammation, maintains viral genomes for the lifetime, and possesses a sustained antigen-specific adaptive immune response. We found that mCMV-specific Tem CD8 T cells dominated the immune response early, and this response was maintained for life. Moreover, both phenotypic and functional (vascular accessibility) data were consistent with the Trm nature of fat-residing CD8 T cells.
Several groups have demonstrated that mucosal tissues, such as the salivary gland and lungs, are home to non-recirculating T cells that respond to mCMV [65–67]. This is largely believed to be in response to mCMV utilizing mucosa as a means for spread through saliva, breast milk, urine, and vaginal fluids [68]. Authors have suggested two potential methods that result in T cell accumulation in the lung and the salivary gland: 1) T cells primed in the periphery traffic to these locations; and/or 2) viral antigens, (even in the absence of full replication as demonstrated by experiments conducted with replication-incompetent mCMV) are presented in situ, evoking cytokine and chemokine cues that maintain memory T cells after original antigenic stimulation. We interpret our data as indicative of continual maintenance of memory T cells in situ. However, this raises two questions. First, why would adipose tissue be evolutionarily advantageous for mCMV infection? HCMV alters the lipid metabolism of infected cells [69,70] and given the high density of lipids within adipocytes it is possible that adipocytes or their progenitors and even fibroblasts could provide significant sources of lipids and therefore become prime targets for infection. Furthermore, different cells of the adipose tissue, including adipose tissue-derived stem cells [71], have been shown to be susceptible to CMV infection. Second, if viral antigens are not being presented within adipose tissue, why would the immune system divert a lifelong T cell population to this site? Other groups have suggested that the presence of memory T cells in the fat would be expected given the anatomical location of adipose tissue with respect to lymphatic organs, the gut, and the vasculature to provide clean-up for any antigenic leakage from these tissues. We show that fat-residing CMV-specific T cells are phenotypically activated, suggesting recent antigenic stimulation. That would support the hypothesis of antigenic presentation in situ, which may be supported by the PCR detection of viral gene products during the 10 months post infection period. While, at present, we cannot formally exclude that CMV antigens may indeed leak from these proximal tissues, we consider such a possibility less likely, given the tight temporal regulation of mCMV antigen expression. An alternative possibility would be that trafficking cells harboring CMV, such as inflammatory monocytes, could potentially be continually seeding the adipose tissue and that this would help maintain the mCMV-specific Trm cells. Against that possibility, we found that CCR2-/- mice, infected via the footpad route, as an attempt to isolate initial replication as much as possible, also exhibited significant accumulation of mCMV-specific T cell population in the fat, suggesting either cell-free spread or a non-monocyte cell-associated virus, below our limit of detection at this time, as drivers of T cell accumulation in the adipose tissue. Based on the preponderance of evidence, we favor the scenario whereby a persistent, bona fide latency established by mCMV within adipose tissue drives the accumulation of CD8 Trm cells.
Inflammation within adipose tissue has been widely investigated for its role in the development of metabolic syndromes [58]. In our experiments, mCMV infection resulted in inflammation within adipose tissue in the absence of obesity. The influx into the adipose tissue by leukocytes and specifically CD8 Tem/rm cells could potentially alter the metabolic profile of infected mice. When we measured glucose and insulin changes in infected animals, we did not observe any changes in fasted blood glucose until animals were ten to twelve weeks post infection, by which time we did not detect any difference in systemic insulin levels. This difference was maintained in lifelong infected animals, showing a significant elevation in the fasted blood glucose of infected animals, but with no statistical difference in the total weight, at end of life, or longitudinal differences in fat pad hypertrophy or atrophy. These observations are consistent with recent data demonstrating that the production of IFNγ in response to mCMV and influenza infection in a model of dietary-induced obesity was the “tipping” point in the manifestation of insulin resistance [59]. Furthermore, we find that neither gluconeogenesis nor reduced insulin sensitivity were responsible for elevated fasted blood glucose. Rather, we believe that infection potentially alters systemic glucose control, an issue that will require further experimentation. Thus, our finding could provide one potential mechanism to link epidemiological data in humans showing that HCMV infection increases the risk of developing atherosclerosis, insulin resistance, and other metabolic diseases [72–76]. In that scenario, we speculate that CMV infection alone could increase one’s risk for developing metabolic disorders, but that additional environmental factors are required, such as diet, other infections, and aging; to what extent this interplay is dependent upon adipose tissue remains to be established.
Our data identify adipose tissue as a potential reservoir for mCMV genomic persistence, through our detection of viral products at 10 months post infection. mCMV infection clearly leads to the continuous stimulation of antigen-specific CD8 T cells that take up residency within adipose tissue, based upon phenotypic data. Trm cells are maintained for the lifetime of infection and likely contribute to an inflammatory environment within adipose tissue. These data reveal a strategy by which the adaptive immune system controls mCMV in tissues and provide insights that could mechanistically link mCMV infection of the adipose tissue to metabolic dysfunction, that may depend on additional metabolic and environmental stressors, such as aging and diet.
Mouse studies were carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. Protocols were approved by the Institutional Animal Care and Use Committee at the University of Arizona (IACUC #08–102, PHS Assurance Number: A3248-01). Footpad injections were performed under isoflurane anesthesia. Euthanasia was performed by isoflurane overdose. Animal experiments performed at Helmholtz Centre for Infection Research (Braunschweig, Germany) were approved by Lower Saxony State Office of Consumer Protection and Food Safety under the license number 33.9–42502-04-14/1712.
Ten-week-old adult C57BL/6J and congenic CD45.1 (B6, H-2b), B2m KO, and CCR2-/- male mice were purchased from The Jackson Laboratory. At 12 weeks of age, adult mice were infected with 105 pfu of mCMV intraperitoneally or via footpad, both routes produce overlapping data, (Smith strain, originally obtained from M. Jarvis and J. Nelson, Oregon Health and Science University, Portland, OR, passage 3 on M210B4 cells. Mice were maintained under specific pathogen-free conditions in the animal facilities at the University of Arizona and at Helmholtz Centre for Infection Research (Braunschweig, Germany).
Animals were sacrificed by isoflurane overdose. White adipose tissue from the epididymal fat pad was excised, weighed, and cut in small pieces using forceps and scissors. Cut pieces were resuspended in DMEM containing 2 mg/ml Collagenase D (1mL solution per 0.5 grams of adipose tissue) and incubated for 30 minutes at 37°C with shaking. Digestion suspensions were thoroughly vortexed and centrifuged at (800g for 5 min). Adipocyte fraction and liquid interphase was sterile vacuumed away from pellet, which was resuspended in DMEM containing 5% BSA and pushed through a 70 μm nylon mesh filter to remove remaining cell debris. Cells were centrifuged and resuspended in 250 ul PBS containing 1% BSA. 50 ul of resuspension was used to calculate cell counts per gram of adipose tissue and the remaining used for flow cytometry. Numerical quantification of single cell suspensions was carried out using Drew Scientific HemaVet 950.
Adipose tissue was collected into a microcentrifuge tube filled with 1 mL of Qiazol and autoclaved glass beads and then snap frozen in liquid nitrogen. Samples were thawed and homogenized using a bead beater for two 2-minute cycles. RNA was extracted using Qiagen RNeasy Lipid Tissue Mini Kit per the manufacturer’s protocol. Reverse transcription was carried out using Sensiscript RT Kit per manufacturer instructions. Amplification of cDNA was performed using SYBR Green Master Mix on an ABI 7300. Standard curve was generated using plasmid gifted from Wayne Yokoyama, MD, Washington University in St. Louis. Primer set were gifted by Chris Benedict, PhD, La Jolla Institute of Immunology [72]. For applications utilizing the RT2 Miniarray Profiler samples were treated as above, analyzed for RNA Integrity Number (RIN) by the University of Arizona Genetics Core Facility on an Agilent Bioanalyzer 2100. Following manufacturer protocol only samples with a RIN greater than 7 were used. Analysis was carried out through Qiagen’s Data Analysis Center.
The liver, spleen, subcutaneous fat, and perigonadal adipose tissue were harvested from mCMV infected and uninfected age matched controls. Each tissue was collected into a microcentrifuge tube containing 1 mL of Qiazol and autoclaved glass beads, and snap frozen in liquid nitrogen. Following thaw, samples were homogenized by bead beating with two 2-minute cycles. DNA was extracted from each sample per the Qiazol manufacturer’s protocol. qPCR was performed using PowerUP SYBR Green Master Mix on an Applied Biosciences Step One real-time PCR system using the following cycle protocol: an initial step at 2 min 50°C followed by 95° for 10 min, followed by 40 cycles of 95° for 15 sec, 60° for 1 min. Recombinant plasmids containing IE1 and C57/BL6 β-actin were used as template to establish standard curves for quantification. The primer sequences were as follows: IE1-fw (5’- CCC TCT CCT AAC TCT CCC TTT-3’) and IEI-rv (5’-TGG TGC TCT TTT CCC GTG-3’), β-actin-fw (5’-AGC TCA TTG TAG AAG GTG TGG-3’) and β-actin-rv (5’-GGT GGG AAT GGG TCA GAA G-3’). Cycle 32 was set as a negative cut-off based on uninfected controls. Primer sets and recombinant plasmids were gifted by Wayne Yokoyama, MD, Washington University in St. Louis.
8-week-old C57BL/6J female mice were i.p. injected with 106 pfu of bacterial artificial chromosome–derived mCMV (pSM3fr-MCK-2 full-length [73]) and sacrificed at 90d or at greater than 240d p.i. Perigonadal adipose tissue stromal vascular fractions were isolated as described previously [74], stained with antibodies and FACS-sorted into CD45- and CD45+CD11b+ subsets. DNA was extracted using QIAamp DNA Micro Kit (QIAGEN) according to manufacturer’s protocol. Real-time PCR quantification of viral genome load was performed as described previously [75] with modifications. Briefly, equivalent volumes of each DNA sample were analyzed in qPCR reactions with primer pairs specific for either the viral gene M55/gB or the mouse gene Pthrp. Reactions were set up using Fast EvaGreen qPCR master mix (Biotium, Fremont, CA) and run in a LightCycler480 (Roche, Mannheim, Germany) using the following cycling protocol: an initial step of 2 min at 95°C followed by 50 cycles of 10 s at 95°C, 20 s at 56°C, and 30 s at 72°C. Specificity of the amplicons was confirmed through melting curve analysis and by electrophoresis on agarose gels. Absence of cross-contamination was ascertained by parallel assessment of negative water controls and of DNA samples from non-infected animals. A recombinant plasmid standard containing sequences of both gB and Pthrp genes [75] was used as a template to establish standard curves for quantification. The dynamic range of the assay stretched from 101 to106 mCMV genome copies per reaction. The following primer sequences were used: gB-fw (5′- GCAGTCTAGTCGCTTTCTGC-3′) and gB-rev (5′-AAGGCGTGGACTAGCGATAA-3′); Pthrp-fw (5′- GGTATCTGCCCTCATCGTCTG-3′) and Pthrp-rev (5′-CGTTTCTTCCTCCACCATCTG-3′).
Isolated cells were stained using flow cytometry reagents as indicated in Table 1. Dead cells (identified as 7-Amino-Actinomycin D+ or using LIVE/DEAD Fixable Dead Cell Staining Kits) and cell aggregates (identified on FSC-A versus FSC-W scatter plots) were excluded from all analyses. Cells were plated into 96-well round bottom plates (Costar). Cells were treated with FcBlock (anti-CD16/32) in PBS supplemented with 2% BSA (FACs buffer) for 10 minutes at 4 C and then surface staining antibodies, also in FACs buffer, added for an additional 45 minutes at 4 C. In experiments requiring intravascular staining animals were injected with 3 ug of anti-CD45 antibody in 50 ul of PBS retro-orbitally and waiting 5 minutes prior to sacrificing animals. After initial staining steps, cells were washed three times FACs buffer and then stained using LIVE/DEAD viability dye in PBS alone for 30 minutes at 4 C. Finally, cells were washed once with PBS and three times with FACs. Cells were fixed in BD Cytofix following manufacturers protocol and then washed three times prior to analysis. Data acquisition was performed on a custom-made, four-laser BD Fortessa flow cytometer (Becton Dickinson), and was analyzed using FlowJo software (Tree Star). Cell sorting was performed on a FACSAriaII (BD Biosciences). Gating was informed by using fluorescence minus one (FMO) controls.
Total epididymal adipose tissue was excised from infected and control animals and weighed to normalize downstream analysis per gram. Adipose was collected into 0.5% NP-40 buffer in PBS plus 1/100 protease inhibitor cocktail and homogenized using Qiagen TissueRuptor. Samples were incubated at room temperature for 30 minutes and then centrifuged at 4700 RPM for 30 minutes at 4 C. Liquid interphase was taken for downstream analysis. Adiponectin and Leptin ELISAs and BioLegend LegendPlex 13-plex Inflammation Panel analyses were carried out following manufacturer protocols.
Prior to collection of fasted blood glucose and plasma for insulin measurements mice were fasted for at least seven hours. Blood glucose was measured by tail nick and using Bayer Counter Next EZ Glucose Meter. After glucose measurements blood was obtained via retro-orbital bleeding into EDTA treated tubes (ThermoFisher) by centrifugation for 15 minutes in 4 C at 2,000 x g. Insulin was measured using Insulin Mouse ELISA kit (ThermoFisher). HOMA-IR was calculated by multiplying fasted plasma insulin and fasted blood glucose and dividing the product by 22.5, the inverse of this result was taken to represent Insulin Sensitivity [76]. For GTT, ITT, or PTT, following fasting mice were i.p. challenged with 1 mg/kg glucose, or 1 U/kg Humalog insulin (Eli Lilly), or 2.5 g/kg sodium pyruvate in PBS respectively and blood glucose measured as described above.
Statistics were performed in Prism 7.0 (GraphPad Software, La Jolla, CA, USA). Two-tailed Mann Whitney U tests with equal SD were carried out on all analyses unless otherwise noted. Significance is noted as follows throughout: ns = not significant, ****P < 0.0001, ***P < 0.001, **P < 0.01, *P < 0.05. All error bars shown are SEM. In all cases, a bar overlies groups compared for significance and the stars as described above denote significance.
|
10.1371/journal.pgen.1007908 | Estimating recent migration and population-size surfaces | In many species a fundamental feature of genetic diversity is that genetic similarity decays with geographic distance; however, this relationship is often complex, and may vary across space and time. Methods to uncover and visualize such relationships have widespread use for analyses in molecular ecology, conservation genetics, evolutionary genetics, and human genetics. While several frameworks exist, a promising approach is to infer maps of how migration rates vary across geographic space. Such maps could, in principle, be estimated across time to reveal the full complexity of population histories. Here, we take a step in this direction: we present a method to infer maps of population sizes and migration rates associated with different time periods from a matrix of genetic similarity between every pair of individuals. Specifically, genetic similarity is measured by counting the number of long segments of haplotype sharing (also known as identity-by-descent tracts). By varying the length of these segments we obtain parameter estimates associated with different time periods. Using simulations, we show that the method can reveal time-varying migration rates and population sizes, including changes that are not detectable when using a similar method that ignores haplotypic structure. We apply the method to a dataset of contemporary European individuals (POPRES), and provide an integrated analysis of recent population structure and growth over the last ∼3,000 years in Europe.
| We introduce a novel statistical method to infer migration rates and population sizes across space in recent time periods. Our approach builds upon the previously developed EEMS method, which infers effective migration rates under a dense lattice. Similarly, we infer demographic parameters under a lattice and use a (Voronoi) prior to regularize parameters of the model. However, our method differs from EEMS in a few key respects. First, we use the coalescent model parameterized by migration rates and population sizes while EEMS uses a resistance model. As another key difference, our method uses haplotype data while EEMS uses the average genetic distance. A consequence of using haplotype data is that our method can separately estimate migration rates and population sizes, which in essence is done by using a recombination rate map to calibrate the decay of haplotypes over time. An additional useful feature of haplotype data is that, by varying the lengths analyzed, we can infer demography associated with different recent time periods. We call our method MAPS for estimating Migration And Population-size Surfaces. To illustrate MAPS on real data, we analyze a genome-wide SNP dataset on 2224 individuals of European ancestry.
| Populations exist on a physical landscape and often have limited dispersal. As a result, most genetic data exhibit a pattern of isolation by distance [1], which is simply to say that populations closer to each other geographically are more similar genetically. Furthermore, the degree of isolation by distance can vary across space and time [2]. For instance, in a mountainous area of a terrestrial species’ range, a pair of individuals may be more divergent from each other than a pair of individuals separated by the same distance in a flat and open area of the habitat. Additionally, the degree of isolation by distance can change over time—for example, if dispersal patterns are changing over time. Such spatial and temporal heterogeneity is an important aspect of population biology, and understanding it is crucial to solving problems in ecology [3], conservation genetics [4], evolution [5], and human genetics [6].
Several methods have been developed to reveal spatial heterogeneity in patterns of isolation by distance [7–14]. Some methods are based on explicitly modeling the spatial structure in the data [9, 10, 12–14]; others take non-parametric approaches [7, 8]; while other methods ignore the spatial configuration of the samples and rely on researchers to make a post hoc geographic interpretation of the results [15, 16]. However, none of these methods can be flexibly applied to address temporal heterogeneity in isolation by distance patterns, and new methods are needed.
One source of information for inferring changes in demography across time is the density of mutations observed in pairwise sequence comparisons [17, 18]. For example, when individuals are similar along a long segment of their chromosomes, it suggests that these segments share a recent common ancestor [19]. These segments are often called “identity-by-descent” tracts, although here we prefer the term “long pairwise shared coalescence” (lPSC) segments (as identity by descent traditionally required a definition of a founder generation, which is not clear in most data applications). A key feature of these segments is that filtering them by length provides a means to interrogate different periods of population history. The longest segments reflect the most recent population history, whereas shorter segments reflect longer periods of time. Recent analyses using lPSC segments suggest that they can reveal fine-scale spatial and temporal patterns of population structure that are not evident with genotype-based methods such as principal components analysis [20–22].
Here we develop a new method to infer spatial and temporal heterogeneity in population sizes and migration rates. The method takes as input geographic coordinates for a set of individuals sampled across a spatial landscape, and a matrix of their genetic similarities as measured by sharing of lPSC segments. It then infers two maps, one representing dispersal rates across the landscape, and another representing population density. Importantly, building these maps using different lengths of lPSC segments can help reveal changes in dispersal rates and population sizes loosely associated with different recent time periods.
Our method is based on a stepping-stone model where randomly-mating subpopulations are connected to neighboring subpopulations in a grid. Such models are parameterized by a vector of population sizes (N →) and a sparse migration rate matrix (M). Stepping-stone models with a large number of demes can approximate spatially continuous population models [23, 24], and this can be exploited to produce maps of approximate dispersal rates and population density across continuous space.
Our method builds upon a method developed for estimating effective migration surfaces (EEMS) [12]. While EEMS infers local rates of effective migration relative to a global average, here we can explicitly infer absolute parameter values by leveraging lPSC segments and modeling the recombination process [N → and M values in the stepping-stone model, and effective spatial density function D e ( x → ) and dispersal rate function σ ( x → ) in the continuous limit]. We call this method MAPS, for inferring Migration And Population-size Surfaces.
We test MAPS on coalescent simulations and apply it to a European subset of 2,224 individuals from the POPRES data [25]. In simulations, we show that MAPS can infer both time-resolved migration barriers and population sizes across the habitat. In empirical data, we infer dispersal rates σ ( x → ) and population densities D e ( x → ) loosely associated with different time periods in Europe.
MAPS estimates demography using the number of Pairwise Shared Coalescence (PSC) segments of different lengths shared between individuals. We define a PSC segment between (haploid) individuals to be a genomic segment with a single coalescent time across its length (Fig 1A). Long PSC (lPSC) segments tend to have a recent coalescent time, and so manifest themselves in genotype data as unusually long regions of high pairwise similarity, which can be detected by various software packages [26–29]. Because lPSC segments typically reflect recent coalescent events, counts of lPSC segments are especially informative for recent population structure [19, 24, 30]. And partitioning lPSC segments into different lengths bins (e.g. 2-8cM, ≥8cM) can help focus inference on different (recent) temporal scales. However, we caution that the historical signal that gives rise to the number of segments of in a certain length bin (e.g. 2-8cM) to strongly overlap with that has given rise to a numbers of segments subsequent length bin (e.g. ≥8).
The MAPS model involves two components: i) a likelihood function (Eq (7)), which relates the observed data (genetic similarities, as measured by sharing of lPSC segments) to the underlying demographic parameters (migration rates and population sizes); and ii) a prior distribution on the demographic parameters, which captures the idea that nearby locations will often have similar demographic parameters. The likelihood function comes from a coalescent-based “stepping-stone” model in which discrete populations (demes) arranged on a spatial grid exchange migrants with their neighbors (Fig 1b). The parameters of this model are the migration rates between neighboring demes (Mα,β) and the population sizes within each deme (Nα). The prior distribution is similar to that from [12], and is based on partitioning the habitat into cells using Voronoi tesselations (one for migration and one for population size), and assuming that migration rates (or population sizes) are constant in each cell. We use an MCMC scheme to sample from the posterior distribution on the model parameters (migration rates, population sizes, and Voronoi cell configurations). We can summarize these results by surfaces showing the posterior means of demographic parameters across the habitat.
The inferred migration rates and population sizes will depend on the density of the grid used. However, using ideas from [23] and [24] we convert them to corresponding parameters in continuous space, whose interpretation is independent of the grid for suitably dense grids. Specifically, we convert the migration rates to an effective spatial diffusion parameter σ ( x → ), often referred to as the “root mean square dispersal distance”, which can be interpreted roughly as the expected distance an individual disperses in one generation (Eq (18)); and we convert the population sizes (N →) to an “effective population density” D e ( x → ), which can roughly be interpreted as the number of individuals per square kilometer (Eq (17)). These are deemed “effective” parameters because the spatial re-scaling assumes a simple approximation to the two dimensional coalescent process, see [23]. Similar to the original grid-based demographic parameters, we can summarize MAPS results by surfaces showing the posterior means of σ ( x → ) and D e ( x → ) across the habitat.
Our MAPS approach is closely related to the EEMS method [12], but there are some important differences. First, the MAPS likelihood is based on lPSC sharing, rather than a simple average genetic distance across markers. This was primarily motivated by the fact that, by considering lPSC segments in different length bins, MAPS can interrogate demographic parameters in recent time periods. However, this change also allows MAPS, in principle, to estimate absolute values for the parameters M and N →, whereas EEMS can estimate only “effective” parameters which represent the combined effects of M and N →. This ability of MAPS to estimate absolute values stems from its use of a known recombination map, which acts as an independent clock to calibrate the decay of PSC segments. Finally, MAPS uses a coalescent model, whereas EEMS uses a resistance distance approximation [12, 31].
We assess the performance of MAPS with several simulations, and compare and contrast the results with EEMS. We used the program MACS [32] to simulate data under a coalescent stepping stone model and refinedIBD [27, 28] to identify lPSC segments. Alternatively, we could of inferred lPSC segments exactly using [32] or [33], however we found the error from refinedIBD to be negligible in our simulations. All simulations involved twenty demes, each containing 10,000 diploid individuals, and each exchanging migrants with their neighbor with a per lineage migration rate equal to 0.01 per generation. We analyzed each simulated data set using PSC segments of length 2-6cM and ≥6cM, which correspond to time-scales of approximately 50 generations and 12.5 generations respectively (see Lemma 5.3 in S1 Appendix), however these are only the mean coalescent times and considerable variation exists in distribution of coalescent times. Results for other length bins also reflect the change in migration due to barrier (S1 & S2 Figs).
To illustrate MAPS on real data, we analyze a genome-wide SNP dataset on individuals of European ancestry [25]. Previous analyses of these data have shown the strong influence of geography on patterns of genetic similarity [20, 34, 35]. In particular [20] analyzed spatial patterns in the sharing of PSC segments across Europe. To facilitate comparison with their results, we use their PSC segment calls, focusing on a subset of 2224 individuals after filtering (see Methods).
We applied MAPS to these data using three different PSC segment length bins: 1−5cM, 5−10cM, and > 10cM. The longer bins correspond to more recent demography because as PSC lengths increase, the average coalescent times decrease. Indeed, the average coalescent times for each of these three length bins is inferred to be 90, 23 and 7.5 generations respectively, which roughly correspond to 2700 years, 675 years and 225 years if we assume 30 years per generation and a sufficiently large effective population size (see S1 Appendix). Here, we caution that these are only the mean coalescent times: other analyses have shown that distribution on coalescent times can have a very wide distribution and are strongly affected by the demographic history, especially in expanding populations [20].
We note that the accuracy of called PSC segments will vary across these bins: based on simulations in [20] PSC segment calls in the smallest bin (1-5cM) will likely suffer from both false positives and false negatives, whereas for the longer bins PSC calls should be very reliable. Nonetheless, even in the smallest bin, closely-related individuals will still tend to show higher PSC sharing, and so the estimated MAPS surfaces should provide a useful qualitative summary of spatial patterns of variation even if quantitative estimates may be less reliable.
We developed a method (MAPS) for inferring migration rates and population sizes across space and time periods from geo-referenced samples. Our method builds upon a previous method developed for estimating effective migration surfaces (EEMS) [12]. However there are several differences between MAPS and EEMS. Most fundamentally, MAPS draws inferences from observed levels of PSC sharing between samples, whereas EEMS draws inferences from the genetic distance. These two data summaries capture different information about the coalescent distributions: in essence, PSC sharing captures the frequency of recent coalescent events, whereas genetic distance captures the mean coalescent time. Consequently MAPS inferences largely reflect the recent past (mean coalescent time ⪅ 2,250 years for PSC segments > 2cM), whereas EEMS inferences reflect demographic history on a longer timescale across which pairwise coalescence occurs (99% of events > 6000 years old, assuming diploid Ne of 10,000 for humans, exponential coalescent time distribution).
Another consequence of modelling PSC sharing, rather than genetic distance, is that MAPS can separately estimate demographic parameters related to migration rates (M) and population sizes (N →), as in Fig 3 for example. In essence MAPS does this by using the known recombination map as an additional piece of information to help calibrate inferences. In contrast, EEMS makes no use of recombination maps and cannot separate M and N →. Instead EEMS infers a compound parameter referred to as the “effective migration rate”, which is influenced by changes in both M and N →; see Fig 3. In principle, if applied to sequence data instead of genotype data at ascertained SNPs, the genetic distances used by EEMS could perhaps also separately estimate M and N → by exploiting known mutation rates to calibrate inferences. However, this would require non-trivial additional changes to the current EEMS likelihood, which was designed to be applicable to ascertained SNPs and does not explicitly model variation in population sizes. (The EEMS likelihood instead uses a “diversity rate” eq, which reflects within-deme heterozygosity but is not explicitly a population size parameter.)
An additional useful feature of PSC segments is that, by varying the lengths analyzed, one can infer parameter values across associated with different time periods. For example, our simulations show how by contrasting shorter and longer PSC segments, the method can reveal different gene flow patterns in scenarios with recent changes (see Figs 2 and 3). Further support comes from our empirical analysis of the POPRES data-set, where we found population sizes inferred from longer PSC segments to be more correlated with census sizes than sizes inferred from shorter segments (e.g. Spearman’s ρ = 0.71 for 1−5cM and ρ = 0.84 for > 10cM; see Fig 5 and S5 Fig). Also, not surprisingly, PSC segments greatly outperform using heterozygosity as an indicator of census population size (the Spearman’s correlation coefficient between heterozygosity and census size was insignificant, p-value = 0.25).
Our estimates of dispersal distances and population density from the POPRES data are among the first such estimates using a spatial model for Europe (though see [30]). The features observed in the dispersal and population density surfaces are in principle discernible by careful inspection of the numbers of shared PSC segments between pairs of countries (e.g. using average pairwise numbers of shared segments, S4b Fig, as in [20]). For example, high connectivity across the North Sea is reflected in the raw PSC calls: samples from the British Isles share a relatively high number of PSC segments with those from Sweden (S4b Fig). Also the low estimated dispersal between Switzerland and Italy is consistent with Swiss samples sharing relatively few PSC segments with Italians given their close proximity (S4b Fig). However, identifying interesting patterns directly from the PSC segment sharing data is not straightforward, and one goal of MAPS (and EEMS) is to produce visualizations that point to patterns in the data that suggest deviations from simple isolation by distance.
The inferred population size surfaces for the POPRES data show a general increase in sizes through time, with small fluctuations across geography; In our results, the smallest inferred population sizes are in the Balkans and Eastern Europe more generally. This is in agreement with the signal seen previously [20]; however, taken at face value, our results suggest that high PSC sharing in these regions may be due more to consistently low population densities than to historical expansions (such as the Slavic or Hunnic expansions).
Although consistent with previous results, our estimates of dispersal and population sizes do not exactly agree with empirical estimates. For example, our estimates of population sizes are consistently lower than the census sizes (S5 Fig). This is to be expected for several reasons. First, census sizes include non-breeding individuals (juvenile and post-reproductive age) that do not impact the formation of PSC segments. Second, MAPS is fitting a single population size per location, and in a growing population the best fit population size will be an under-estimate of contemporary size. Third, in a wide class of population genetic models, the effective size, even among reproductive age individuals, is lower than the census size because of factors that inflate the variance in offspring number. Fourth, some discrepancy is expected simply because the stepping-stone population genetic model used here is only a coarse approximation to the complex spatial dynamics of human populations. Finally, there is probably cryptic relatedness in the POPRES samples which can decrease population size estimates.
Here, as in EEMS, we use a discrete stepping-stone model to approximate a process that might be more naturally modelled as continuously varying in space [12]. Recent work exploits continuous models to estimate dispersal and population density parameters from sharing of lPSC segments [24, 30]. However, these methods assume that dispersal and population density are constant across space: extending them to allow these parameters to vary across space could be an interesting avenue for future work.
Here, we infer demography given a PSC length bin. These PSC length bins correspond to very approximate time periods, and we report the mean age of the segment in the specified time period to give an idea of the approximate time period under an assumption of a large effective population size (see Lemma 5.3 in S1 Appendix). However, as mentioned previously, the variance in the distribution of ages can be very large. A major advancement would be to infer demography explicitly as a function of time. In principle, our method allows for inference of demography across time by treating PSC segments as roughly approximating independent across length bins conditional on the demography, see S1 Appendix. However, this requires fitting multiple migration/population surfaces and is computationally unfeasible with our current MCMC routine. Other computational techniques (e.g. Variational Bayes or fast optimization of the likelihood) might make this goal possible.
For the empirical data analysis, we ran MAPS with 200 demes. The MAPS output was obtained by averaging over 20 independent replicates (the number of MCMC iterations in each replicate was to set 5e6, number of burn-in iterations set to 2e6, and we thinned every 2000 iterations). We provide the the MAPS here: https://github.com/halasadi/MAPS, and the plotting scripts here: https://github.com/halasadi/plotmaps.
Our pipeline to call PSC segments for simulations can be found here: https://github.com/halasadi/ibd_data_pipeline. We follow the recommendations of [27, 28] and [20] by running BEAGLE multiple times and merging shorter segments.
For the empirical data analysis, we use the PSC segments (“IBD”) calls from [20], which can be found here: https://github.com/petrelharp/euroibd. The calls from [20] were obtained by running fastIBD (implemetned in BEAGLE [27]) and applying custom post-processing steps derived by simulation. We further applied a filter to retain countries with at least 5 sampled individuals, and removed Russian and Greek individuals to restrict the geographic region to a smaller spatial scale.
MAPS assumes a population genetic model consisting of triangular grid of d demes (or populations) with symmetric migration. The density of the grid is pre-specified by the user with the consideration that the computational complexity is O(d3). We use Bayesian inference to estimate the MAPS parameters: the migration rates and coalescent rates M and q_ respectively. Its key components are the likelihood, which measures how well the parameters explain the observed data, and the prior, which captures the expectation that M and q_ have some spatial structure (in particular, the idea that nearby edges will tend to have similar migration rates and nearby demes have similar coalescent rates).
MAPS estimates the posterior distribution of Θ=M,q_ given the data. The data used for MAPS consists of a similarity matrix X R = { X i , j R } which denotes the number of PSC segments in a range R = [μ, ν] base-pairs shared between pairs of haploid individuals (i, j) ∈ {1, ⋯, n} × {1, ⋯, n} where n is the number of (haploid) individuals. Furthermore, a recombination rate map is required as input for MAPS. The likelihood is a function of the expected value of X i , j R (E [ X i , j R ]). Below we describe the computation of E [ X i , j R ] and the other key components of the likelihood. Finally, we briefly describe the prior used and an MCMC scheme to sample from the posterior distribution of Θ.
|
10.1371/journal.pntd.0004283 | Etiology and Epidemiology of Diarrhea in Hospitalized Children from Low Income Country: A Matched Case-Control Study in Central African Republic | In Sub-Saharan Africa, infectious diarrhea is a major cause of morbidity and mortality. A case-control study was conducted to identify the etiology of diarrhea and to describe its main epidemiologic risk factors among hospitalized children under five years old in Bangui, Central African Republic.
All consecutive children under five years old hospitalized for diarrhea in the Pediatric Complex of Bangui for whom a parent’s written consent was provided were included. Controls matched by age, sex and neighborhood of residence of each case were included. For both cases and controls, demographic, socio-economic and anthropometric data were recorded. Stool samples were collected to identify enteropathogens at enrollment. Clinical examination data and blood samples were collected only for cases.
A total of 333 cases and 333 controls was recruited between December 2011 and November 2013. The mean age of cases was 12.9 months, and 56% were male. The mean delay between the onset of first symptoms and hospital admission was 3.7 days. Blood was detected in 5% of stool samples from cases. Cases were significantly more severely or moderately malnourished than controls. One of the sought-for pathogens was identified in 78% and 40% of cases and controls, respectively. Most attributable cases of hospitalized diarrhea were due to rotavirus, with an attributable fraction of 39%. Four other pathogens were associated with hospitalized diarrhea: Shigella/EIEC, Cryptosporidium parvum/hominis, astrovirus and norovirus with attributable fraction of 9%, 10%, 7% and 7% respectively. Giardia intestinalis was found in more controls than cases, with a protective fraction of 6%.
Rotavirus, norovirus, astrovirus, Shigella/EIEC, Cryptosporidium parvum/hominis were found to be positively associated with severe diarrhea: while Giardia intestinalis was found negatively associated. Most attributable episodes of severe diarrhea were associated with rotavirus, highlighting the urgent need to introduce the rotavirus vaccine within the CAR’s Expanded Program on Immunization. The development of new medicines, vaccines and rapid diagnostic tests that can be conducted at the bedside should be high priority for low-resource countries.
| Infectious diarrhea is a major cause of illness and death among children under five years from low-income country. In order to identify infectious agents associated with diarrhea, we conducted a case-control study in the Pediatric Complex of Bangui, the sole public pediatric hospital from Central African Republic (CAR). A total of 333 hospitalized children with diarrhea and 333 controls were included, controls being pair matched to the cases according to age, sex and neighborhood. At least one of the sought-for pathogens was identified in 80% of hospitalized children, and approximately one in ten cases presented mixed bacterial/viral co-infections. Five pathogens were positively associated with hospitalized diarrhea, namely rotavirus, norovirus, astrovirus, Shigella/EIEC and Cryptosporidium hominis/parvum. Giardia intestinalis was negatively associated with hospitalized diarrhea. A seasonality effect—viruses during the dry season, bacteria and parasites during the rainy season—but also an age effect, were observed, which may guide clinicians in the management of diarrhea. As rotavirus was the leading cause of severe diarrhea, the introduction of the rotavirus vaccine in CAR will certainly provide considerable direct health benefits in terms of reduced illness and deaths. New medicines, vaccines and rapid diagnostic tests that can be conducted bedside should be urgently developed for low-resource countries.
| In 2013, 6.3 million children under the age of five years died, 578,000 of them from diarrheal diseases. Nearly half of these diarrhea-related deaths were reported from Sub-Saharan Africa [1]. The fourth Millennium Development Goal, established after the United Nations Millennium Summit in 2000, seeks to decrease the mortality of children under five by two-thirds before 2015 [2]. Since 2000, childhood mortality due to diarrhea has diminished by 6.5% every year, but this trend requires an acceleration to reach the 2030 objectives. In order to achieve this decline in childhood diarrheal mortality, the World Health Organization (WHO) published guidelines for the clinical management of childhood diarrhea [3]. These guidelines recommend using antibiotics only for bloody diarrhea, suspected cholera, or associated sepsis. They also encourage zinc supplementation and use of oral rehydration solution (ORS) to treat and prevent diarrhea. However, in practice, antibiotic treatments are overused, resulting in the emergence of antibiotic resistance: only 40% of the children with diarrhea receive the recommended zinc supplementation and ORS [4].
Nowhere are these problems of childhood diarrhea and shortfalls in its management more evident than in the poorest and most unstable countries of Sub-Saharan Africa. The Central African Republic (CAR) is a resource-limited country in equatorial Africa (ranked 180/187 according to the Human Development Index in 2013). Mortality among under five year-old children was 179/1000 in 2010 [5]. No high-quality epidemiological and biological data on severe childhood diarrhea, however, exist for CAR. Indeed, the consequences of decades of poverty, civil war, and economic and political crisis have complicated the management of severe childhood diarrhea. Moreover, a recent qualitative investigation in CAR revealed the complex home management of childhood diarrhea. Parents’ beliefs that diarrheal illness needs to be stopped immediately, that it requires medication, that they should avoid consulting primary health centers and minimize expenses were the most important reasons hampering effective home management of diarrhea [6]. The CAR therefore provides a strong case study in which to understand better the epidemiology and etiology of severe childhood diarrhea in the poorest and most unstable countries without functioning health care systems. Investigating childhood diarrhea in such a context can assist in managing this treatable pathology, by highlighting the most appropriate and adaptive public health interventions.
The Global Enteric Multicenter Study (GEMS) study is a matched case-control study of moderate-to-severe diarrhea in children aged 0–59 months which aimed to estimate the pathogen-specific disease burden in populations from four sites in Africa and three in Asia. The GEMS showed that preventive strategies targeting five pathogens (rotavirus, Shigella, ST-ETEC, Cryptosporidium, typical enteropathogenic E coli) are likely to substantially reduce the burden of moderate-to-severe diarrhea. However, the public health interventions in very low income countries like the CAR, suffering from long-term instability, are different from relatively stable countries with higher income and better medical infrastructure.
Our study, a matched case-control study of diarrhea among hospitalized children under five years was conducted at the Pediatric Complex (PCB) in the CAR’s capital city, Bangui. It was performed in collaboration with the Institut Pasteur de Paris (IPP), the Institut Pasteur de Bangui (IPB) and the PCB.
The study’s primary objective was to identify pathogens associated with diarrhea in hospitalized children under five years of age. Secondary objectives were i) to describe the clinical symptoms of severe diarrhea among hospitalized children, ii) to identify the risk factors associated with severe diarrhea (anthropometric, socio-economic, environmental characteristics), iii) to describe the management of diarrhea before and during hospitalization, and iv) to describe the vital status of children with severe diarrhea during hospitalization and two months after discharge.
Our study was a matched case-control study conducted in Bangui, CAR from December 2011 to November 2013 at the PCB, the country’s sole public pediatric hospital. Cases were children under 60 months of age, hospitalized for diarrhea. Other inclusion criteria were residence in one of Bangui’s eight districts and a general health condition that would support blood and stool sampling. Exclusion criterion was being positive for human immunodeficiency virus (HIV). Controls, identified from the community, were pair matched to the cases according to age (±2 months for infants (0–11 months), ±3 months for toddlers (12–23 months) and ±6 months for children (24–59 months)), sex and neighborhood. To be eligible, controls had to be in good general health, with no history of diarrhea or antibiotic use during the seven days before sampling. HIV status was not systematically tested in controls, but if parents spontaneously declared a child to be seropositive, that child was not included. Cases and controls could not be included more than once.
The research protocol was approved by the Scientist Committee of the Sciences and Health University of Bangui, the CoRC (Clinical Research Committee of Institut Pasteur), the CCTIRS (Comité Consultatif sur le Traitement de l’Information en matière de Recherche dans le domaine de la santé) and the CNIL (Commission Nationale de l’Information et des Libertés) in France. Written informed consent was obtained from all children’s parents or legal guardians for both cases and controls. This study was conducted according to the protocol and ethical principles with their origins in the Declaration of Helsinki. The project provided treatment and laboratory testing free of charge. The CN/CNLS (Coordination Nationale du Comité National de Lutte contre le Sida), with financial support from the Global Fund, covered all costs for HIV treatment.
To show an odds ratio of 2 characterizing the association between a given pathogen and severe diarrhea, with a pathogen prevalence of 5%, a power of 80% and a two-sided α = 0.05, a sample size of 600 cases and 600 controls was necessary.
The variables collected are defined in Table 1. Data were double entered and analyzed using STATA: SE 13.1 (Stata Corp Station, TX, USA). Continuous variables were expressed as mean (±SD) or median [interquartile range] and discrete variable as percentage and 95% CI. Univariate analyses for continuous variables were performed using Student t-test or Mann-Whitney test when appropriate. For discrete variables, univariate analyses were performed using Chi-2 test or Fisher’s exact test. Tests were two-sided and a p-value<0.05 was considered significant.
Comparison between cases and controls were made by univariate conditional logistic regression to take into account the matching of cases and controls. Pathogens potentially associated with severe diarrheas in univariate analysis with a p-value<0.25 were included in a backward conditional logistic regression, adjusted on the presence of other pathogens. Results are reported as adjusted OR (aOR) with 95% CI. The attributable fraction (AF) was calculated for pathogens with significant aOR with the following formula: aAF = P (pathogens) among cases* (aOR-1)/aOR, that is, the proportion of severe diarrhea attributable to this specific pathogen. When the association was significantly negative with an aOR<1, the protective fraction was calculated with the following formula: aPF = (1- aOR) x p (events) among controls, that is, the proportion of severe diarrhea avoided by the presence of this specific pathogen.
For cases, associations between socio-economic, anthropometric and clinical data with the use of ORS and antibiotics before and during hospitalization, and association with vital status were determined by univariate analysis. All variables associated with a p value<0.25 were included in a backward logistic regression. The final model includes only variables with a p-value<0.05. Interactions were tested and the goodness-of-fit of the model was studied using the Hosmer-Lemeshow statistic.
General characteristics of cases are described in Table 2. During the 24 months of the study, 428 consecutive cases were hospitalized for diarrhea. Twenty-two were tested positive for HIV and consequently excluded on that basis. No mothers refused their children’s participation in the study. Nine cases did not match any control. Sixty-four case-control pairs were wrongly matched: 3 for sex and 61 for age. Finally, 333 cases (78%) and 333 (79%) controls were analyzed (Fig 1). Fifty-six percent were male. The mean age at inclusion was 12.9 months (±9.8). The distribution of age was as follows: 195 (59%) between 0–11 months; 103 (31%) between 12–23 months and 35 (10%) between 24–59 months.
The mean time since onset of diarrhea before cases were presented at the hospital was 3.7 (±1.8) days. Cases presented the following clinical symptoms: severe dehydration (N = 239, 72%); hemodynamic shock (N = 216, 65%); serious neurological injury (N = 239, 72%); fever (body temperature ≥38°) (N = 264, 79%); vomiting (N = 267, 80%); and extra digestive signs (N = 94, 28%), including 69 cases of upper respiratory tract infection, 5 cases of pulmonary signs, 8 cases of digestive signs and 12 cases of other signs. 26 cases (8%) were positive for malaria, all of them due to Plasmodium falciparum. The mean episodes of vomiting in the last 24 hours were 4.2 (±2), and the mean number of stools during the last 24 hours was 6.1 (±2). In 16 cases (5%) and 56 cases (17%), macroscopic blood and mucous were found in the stools, respectively.
Demographic characteristics of cases and controls are summarized in Table 2. Cases and controls were well balanced for age and sex at baseline. Cases were significantly more severely or moderately malnourished than controls, 40% versus 14%, respectively, p<0.001. Cases belonged to a higher socio-economic class, had more mothers who completed primary school and had more access to improved water than controls. No difference in diet was found between the two groups.
The results are summarized in Table 3. At least one of the sought-for pathogens was identified in 78% and 40% of cases and controls, respectively, p<0.001. Mixed bacterial/viral infections were detected in 10% of cases and 4% of controls, p = 0.001. Viruses were the most prevalent pathogens, detected in 55% of cases and 15% of controls, p<0.001. Conversely, Giardia intestinalis was more frequent in controls (7.8%) than in cases (0.9%), p<0.001). The ipaH gene was detected in all 14 Shigella-positive culture specimens except one and in 67/651 (10%) culture-negative specimens. In all, 50/333 (15%) and 30/333 (9%) samples were considered positive for the entity Shigella/EIEC in cases and controls, respectively. Presence of Shigella/EIEC was significantly associated with blood in the stool (p<0.001). Diarrhoeagenic E.coli were found in as many cases as controls. Of the diarrhoeagenic E.coli pathotypes detected, the most frequent was EAEC (5.7% of cases and 4.5% of controls) followed by ETEC (4.8% of cases and 4.8% of controls). No significant difference between cases and controls for LT, ST, and LT/ST toxin-producing ETEC was found.
Other pathotypes (EPEC, ATEC, EIEC and STEC) were found in less than 3% of the children. Cryptosporidium parvum/hominis was identified more frequently in cases than in controls (12.6% of cases and 2.7% of controls, p<0.001).
In multivariate analyses, when adjusted on the presence of other pathogens, five pathogens were positively associated with diarrhea: rotavirus, norovirus, astrovirus, Shigella/EIEC, Cryptosporidium parvum/hominis and one appeared negatively associated: Giardia intestinalis. The adjusted AFs (aAF) were 39% for rotavirus, 7% for both norovirus and astrovirus, 9% for Shigella/EIEC and 10% for Cryptosporidium parvum/hominis. The protective fraction for Giardia intestinalis was 6%, (Table 3).
In cases, the prevalence of pathogens varied according to age categories (Fig 2). Globally, the prevalence of viruses and Cryptosporidium parvum/hominis in cases decreased with age, whereas they increased for Shigella/EIEC and Giardia intestinalis. In cases, viruses were found in 63% (122/195) of infants, 50% (52/103) of toddlers and 20% (7/35) of children; (p<0.001). Cryptosporidium parvum/hominis were found in 16% (32/195) of infants, 7% (7/103) of toddlers and 9% (3/35) of children, (p = 0.04). Shigella/EIEC were found in 9% (17/195) of infants, 22% (23/103) of toddlers and 29% (10/35) of children, (p<0.001). In controls, the percentage of virus was stable with age, and the proportion of Cryptosporidium parvum/hominis remained low. In contrast, the percentage of Giardia intestinalis was higher among controls than in cases and increased with age.
Viruses were more frequently identified during the dry season, with 65% (115/179) of cases compared to 43% (66/154) in rainy season, p<0.001. In contrast, Shigella/EIEC and Cryptosporidium parvum/hominis were mainly identified during the rainy season: for Shigella, 19% (30/154) cases in the rainy season and 11% (20/179) in the dry season (p = 0.03); and for Cryptosporidium parvum/hominis, 16% (25/154) cases in the rainy season and 9% (17/179) in the dry season (p = 0.06) (Fig 3).
Before hospitalization, ORS and zinc supplementation were prescribed to 38% and 0.9% children, respectively. Antiparasitic treatments and antibiotics were administered to 44% and 34% of children, respectively. One on four children (25%) received traditional treatments that consisted mainly in infusions or herbal decoctions (72%), herbal enemas (16%) or fruit porridge (9%). During hospitalization, 99% of children received ORS, 87% intravenous rehydration, 70% antibiotics, 66% antiparasitic treatments, 55% zinc, and none received traditional treatments (Fig 4). The only factor identified as independently associated with the prescription of ORS before hospitalization was bloody stools. No factors were identified to be associated with the use of antibiotics before hospitalization or the use of any treatment during hospitalization.
The mean duration of hospitalization was 4.8 (±2.8) days. Four percent of children (12/333) died during hospitalization and 1% (3/271; 50 missing values) died during the two months after discharge. Among the 12 deaths, 5 were infants (< 11 months), 5 were toddlers (12 to 23 months), and 2 were children (24–59 months). Three cases were positive for Plasmodium falciparum and probably died of severe malaria, 2 from severe anemia, 3 from septic shock, and 1 from small bowel obstruction. Finally, 3 died from metabolic acidosis secondary to hemodynamic or septic shock. In the multivariate analysis, the only factor found to be significantly associated with the vital status was the nutritional status at the time of hospitalization: 11% (4/38) of children with SAM (WHZ < -3DS or MUAC < 115mm) died during hospitalization, compared to 3% (8/293) for MAM or normally well-nourished children, (p = 0.026).
This study is the first case-control study conducted in CAR that provides the etiology and clinical outcome of children of less than five years that were hospitalized for diarrhea in Bangui. A large prospective study, the GEMS study, used a similar approach with matched case-control in children between 0 to 59 months suffering from moderate-to-severe diarrhea in seven countries, including four in Africa: The Gambia and Mali (West Africa), Mozambique (South Africa) and Kenya (East Africa) [11]. Our study differs from GEMS by its recruitment. No country from Central Africa was included in the GEMS study. In addition, our study was conducted in a very low-income country, suffering from long-term instability, substantial poverty, and resource-poor primary health centers. Finally, the study also differs from the GEMS study in its inclusion criteria. In our study, children were included because they were hospitalized for diarrhea, whether or not they met the WHO criteria for severe dehydration [12]. This inclusion criterion, easier to apply in resource-poor settings, fits well with the WHO criteria for severe diarrhea. Indeed, an overwhelming majority of cases had severe dehydration (72%), more than 60% had signs of hemodynamic shock and almost 90% had to be intravenously rehydrated. In this way, our study complements the GEMS study.
HIV positive children were excluded from the present study. Indeed, HIV (co-) infection increases the risk of severe diarrhea by impairing the immune system, making problematic the interpretation of results in the context of case/control comparison. Moreover, it was not possible to test the community controls for HIV without returning to provide individual announcement and counseling, which was not possible in our study. Even if the HIV status of the controls was unknown, we supposed that it was low because only 22 hospitalized cases were tested positive for HIV and the children selected as controls were in good health. Among the 406 HIV-negative cases, only 333 were included in our analysis because they were well-matched with the controls as shown in the flow chart. The main criterion for impaired matching was age. The majority of these cases were borderline, with only few days or weeks of difference between cases and controls. The addition of the wrongly matched children for age and sex did not modify the results. In the current analysis, we prefer to maintain the criteria that were defined in the protocol.
At least one of the sought-for pathogens was identified in around 80% of children hospitalized for diarrhea, and approximately one in ten cases presented mixed bacterial/viral co-infections, consistent with findings in other African countries [13, 14]. Five pathogens were significantly associated with severe hospitalized diarrhea, namely rotavirus, norovirus, astrovirus, Shigella/EIEC, Cryptosporidium hominis/parvum. Seasonality and age effects were also observed, consistent with other studies [9, 15–17]. These results may assist clinicians diagnosing the causes of diarrhea. The frequency of Giardia intestinalis detection was lower in children with diarrhea (0.9%) than in asymptomatic carriers (7.8%), a finding also in keeping with previous data [18, 19] and supporting claims that this parasite is not a major cause of severe diarrhea. Further studies are needed to determine whether Giardia intestinalis protects against or is a consequence of diarrhea. As in GEMS [11], most attributable episodes of diarrhea were associated with rotavirus (40% of the cases versus 3.3% in controls) in concurrence with previous data from CAR [20, 21] and other Sub-Saharan countries [15]. Most rotavirus infections occurred in the youngest children, consistent with the limited protection conferred by maternal antibodies during the first months of life and the effective immunity granted by repeated infections [22]. The incidence and severity of rotavirus infections has declined significantly in countries that have integrated the rotavirus vaccine into their routine childhood immunization policies [23], highlighting the urgent need to introduce it within the CAR’s Expanded Program on Immunization [24]. The proportion of norovirus infections among cases and controls was slightly lower than the prevalence reported from high-mortality developing countries (about 14% in cases versus 7% in controls) [16].
Cryptosporidium, a major cause of chronic diarrhea in malnourished patients or those with positive HIV status [25], was a significant pathogen in our study. Our findings on Cryptosporidium are consistent with those from GEMS [11] and indicate the high global burden of cryptosporidiosis among children in Africa, regardless of their HIV-status. These elements support the need to inform healthcare professionals about this pathogen and to develop practical, inexpensive kits for its detection in resource-poor settings.
Shigella spp. are consistently reported as highly associated with diarrhea in case-control studies [11, 26, 27]. In our study, Shigella spp are the third most important pathogen to be associated with diarrhea, after rotavirus and Cryptosporidium. The proportion of Shigella/EIEC infections among cases and controls was 15% and 9%, respectively, indicating that the prevalence of asymptomatic shedders is higher than expected. However, comparing our findings with those of other studies is difficult, because the pathogen was largely detected using PCR. Traditionally, the current gold standard for Shigella species detection is culture, which is highly selective, but poorly sensitive due to inconsistent bacterial load, loss of bacterial viability during specimen transport and frequent antibiotic treatment before culture. In addition, the gene ipaH used for the diagnosis is also carried by EIEC, implying that PCR cannot differentiate between Shigellosis and EIEC. Nevertheless, it is well established that Shigella is much more prevalent and thus, probably represents most of the ipaH-associated organisms detected. Usually, PCR-based methods have a higher sensitivity compared to conventional culture methods, which improves the ability to detect pathogens in a stool sample. As individuals with diarrhea tend to have higher quantities of bacteria isolated from their stool than do those without diarrhea [28–30], the Shigella-specific disease burden might be underestimated using qualitative PCR. A recent study showed that a cutpoint threshold of approximately 1.4 × 104 ipaH copies could be the new reference standard for the detection and diagnosis of shigellosis in children in low-income countries [31].
In case-control studies, diarrhoeagenic E. coli pathotypes show inconsistent association with diarrhea patients, whereas Salmonella enterica, Campylobacter and adenovirus are often found in similar proportions in patients with or without diarrhea, as we observed in our data [10, 11, 19, 26]. This finding suggests that comparing prevalence between cases and controls may have little value for pathogens with frequent asymptomatic excretion [32, 33].
As reported in previous studies [34, 35], acute malnutrition was associated with severe diarrhea and was shown to be a risk factor for death during hospitalization. This finding is of major concern, because in 2012 in CAR, 8% of the children under five years suffered from acute malnutrition and 39% of chronic malnutrition [36]. This situation has likely worsened because the country has experienced civil war since March 2013.
Our findings also shed light on the management of severe childhood diarrhea. Although the WHO recommends the use of antibiotics in children with bloody diarrhea (5% of cases in our study), suspected cholera, or associated sepsis, we found that 40% of children before their arrival and 70% during hospitalization received an antibiotic treatment. Combined with the widespread use of antiparasitic treatments before consultation (44%), this finding could partially account for the relatively low number of bacteria or parasites found among cases. However, the bias was minimized for Shigella species and Cryptosporidium parvum/hominis as they were detected using PCR.
Furthermore, the uncontrolled consumption of antimicrobial agents is cause for concern in countries like the CAR with inadequate healthcare systems, because it favors the spread of antimicrobial resistance [37–39]. Antibiotic treatments were primarily pills obtained without prescription from street vendors, who offer also diagnosis and limited diagnostic and medical services [6]. These medicines are much less expensive than in pharmacies, but there are no oversights on the safety, appropriateness or duration of such treatments [6].
Despite guidelines that recommend the use of ORS and zinc supplementation for all children, only 40% of children received ORS and less than 1% took zinc before their hospitalization. Although the pre-hospitalization ORS finding is higher than that found in other studies (ORS before hospitalization in 16% of cases according to UNICEF in CAR between 2008 and 2012, 20% in Senegal), it is essential to continue education of mothers on the importance of rehydration and zinc in home management of diarrhea.
The WHO also recommends exclusive breastfeeding for the first six months of life. Only 16% of children under six months old were exclusively breastfed, whereas others were exposed to putative pathogens in weaning foods or inadequate diversity of complementary foods.
Our study has limitations. We could not complete the planned inclusion of 600 cases and 600 controls, mainly due to security problems in Bangui. Only 2 or 3 children were recruited per day during the study-period. This low recruitment can be explained by the cost associated with transport to the hospital, hospitalization, treatment and care. As previously described [6], children afflicted with severe diarrhea have a complex therapeutic itinerary. On average, parents bring their diarrheic children to the hospital after three days of symptom, and after frequently receiving various street medicines or home remedies. These findings are confirmed by the socio-economic differences observed between cases and controls, the cases being “less poor” than the controls. It is likely that a certain number of children never go to the hospital because their parents cannot pay. Moreover, at a time where ORS was widely adopted in the community, it is not surprising that children with moderate diarrhea that predominates in Bangui were not hospitalized, and therefore were not cases included in our study. The high childhood mortality rate reported in our study (4%) may be even higher for children with poor access to healthcare services.
The choice of immunological methods to detect rotavirus, astrovirus, adenovirus and norovirus is questionable. Indeed, some more sensitive molecular methods have been used in other epidemiological studies on diarrhea [40, 41]. However, the impact of using lower sensitivity methods on the interpretation of the results is likely to be low. The sporadic low level viral shedding is not necessarily clinically relevant and if detected using highly sensitive methods can complicate the interpretation of the results. As described above for Shigella species, high sensitivity of PCR-based methods could lead to an underestimation of virus-specific disease burden when comparing cases with controls. Nevertheless, quantitative methods would have been very useful to assess the relationship between virus shedding and clinical severity in our study [42].
The data reported here are particularly important, given the significance of childhood diarrhea in countries with inadequate healthcare systems and long-term instability, as well as the lack of high-quality data and the difficulty of carrying out such studies in these contexts. Rotavirus, norovirus, astrovirus, Shigella/EIEC, Cryptosporidium hominis/parvum and Giardia intestinalis were significantly associated with severe hospitalized diarrhea. Because rotavirus was the most common cause of severe diarrhea, the introduction of the rotavirus vaccine in CAR will certainly have a major impact on childhood diarrhea. Severe diarrhea requiring antibiotics was extremely rare among CAR children under five years old. The observed overuse of antibiotics poses a major risk for the emergence of resistance, particularly when new discoveries of antibiotics are almost non-existent and the risk of therapeutic impasse real. The development of new medicines, vaccines and new rapid diagnostic tests that can be conducted bed-side should be high priorities for low-resource countries, particularly those suffering from instability and poorly-functioning health systems, as well as for global health structures. Future studies measuring the impact of antibiotic overuse on the intestinal microbiome of children will help to shed light on the complex linkages between malnutrition, diarrhea and immunity.
|
10.1371/journal.ppat.1006327 | Analysis of host microRNA function uncovers a role for miR-29b-2-5p in Shigella capture by filopodia | MicroRNAs play an important role in the interplay between bacterial pathogens and host cells, participating as host defense mechanisms, as well as exploited by bacteria to subvert host cellular functions. Here, we show that microRNAs modulate infection by Shigella flexneri, a major causative agent of bacillary dysentery in humans. Specifically, we characterize the dual regulatory role of miR-29b-2-5p during infection, showing that this microRNA strongly favors Shigella infection by promoting both bacterial binding to host cells and intracellular replication. Using a combination of transcriptome analysis and targeted high-content RNAi screening, we identify UNC5C as a direct target of miR-29b-2-5p and show its pivotal role in the modulation of Shigella binding to host cells. MiR-29b-2-5p, through repression of UNC5C, strongly enhances filopodia formation thus increasing Shigella capture and promoting bacterial invasion. The increase of filopodia formation mediated by miR-29b-2-5p is dependent on RhoF and Cdc42 Rho-GTPases. Interestingly, the levels of miR-29b-2-5p, but not of other mature microRNAs from the same precursor, are decreased upon Shigella replication at late times post-infection, through degradation of the mature microRNA by the exonuclease PNPT1. While the relatively high basal levels of miR-29b-2-5p at the start of infection ensure efficient Shigella capture by host cell filopodia, dampening of miR-29b-2-5p levels later during infection may constitute a bacterial strategy to favor a balanced intracellular replication to avoid premature cell death and favor dissemination to neighboring cells, or alternatively, part of the host response to counteract Shigella infection. Overall, these findings reveal a previously unappreciated role of microRNAs, and in particular miR-29b-2-5p, in the interaction of Shigella with host cells.
| MicroRNAs are genome-encoded small non-coding RNAs that play a pivotal role in post-transcriptional regulation of gene expression. In addition to other biological functions, microRNAs are increasingly recognized as important players in the complex interaction between host and bacterial pathogens. In this work we show that microRNAs modulate infection by Shigella flexneri, a major causative agent of bacillary dysentery in humans. Specifically, we characterize the dual regulatory role of miR-29b-2-5p during infection, showing that this microRNA strongly favors Shigella infection by promoting both bacterial binding to host cells and intracellular bacterial replication. Moreover, we show that miR-29b-2-5p, through the repression of its direct target UNC5C, induces a dramatic increase of host cell filopodia, which are essential to promote bacterial capture and internalization. Importantly, the expression of miR-29b-2-5p is decreased during Shigella replication, as a strategy to promote balanced bacterial replication or as part of the host response to counteract infection. This study constitutes the first demonstration that host microRNAs play an instrumental role in the regulation of infection by Shigella.
| Shigella flexneri (Shigella), a facultative Gram-negative bacterium, is a major causative agent of bacillary dysentery in humans [1, 2]. A key feature in the pathogenesis of Shigella concerns its capability to efficiently colonize the colonic mucosa, in particular the epithelial cells therein, which constitute the primary targets of infection. During the early steps of epithelial cell infection by Shigella, bacteria are captured by pre-existing filopodia, which subsequently retract to bring the bacteria in close proximity to the cell body [3]. Although Shigella does not express classical adhesins, it was recently shown that the surface exposed autotransporter protein IcsA (also referred to as VirG) exhibits inducible adhesin-like properties upon exposure to bile salts [4]. To enter cells, Shigella uses a trigger mechanism, characterized by the formation of actin-rich membrane ruffles that mediate bacteria internalization [5]. Following invasion, Shigella escapes from the vacuole and replicates within the host cell cytosol; replication is accompanied by the spreading of Shigella to neighboring cells, using actin-based motility [6, 7]. Bacterial effector proteins, injected into the host cytosol through a virulence plasmid encoded type-3 secretion system (T3SS), are essential for all the steps of the interaction of Shigella with the host [8–10].
Notwithstanding the ability of Shigella, and other bacterial pathogens, to manipulate a vast range of host cellular functions and the multiple mechanisms used by host cells to counteract infection, a still unexplored question regards the interplay between Shigella and host cell microRNAs. MicroRNAs (miRNAs) are small RNA molecules, of ca. 22 nt. in length, that play crucial roles in post-transcriptional regulation of gene expression [11]. In mammals, miRNAs act mainly by binding to partially complementary sequences in the 3’ untranslated regions (3’UTRs) or coding sequences of target mRNAs, leading to translation inhibition and/or degradation of target transcripts [12]. The observations that each miRNA targets multiple transcripts, that multiple miRNAs can simultaneously regulate individual transcripts, and the incomplete knowledge of the principles ruling miRNA-target interaction, render the identification of miRNA targets and the assignment of specific phenotypes a complex process [13, 14].
Considering the pervasive role of miRNAs in the control of gene expression [15] and their involvement in numerous pathophysiological processes [11, 16], the premise that miRNAs are relevant players in the interaction between bacterial pathogens and host cells is of particular interest. Indeed, miRNAs have been extensively described as part of the immune response to various bacterial pathogens, from plants to vertebrates, and are increasingly recognized as a novel molecular strategy exploited by bacteria to subvert host cell pathways and thus promote infection [17–19]. Along this line, we have recently shown, through a systematic genome-wide screening approach, that host miRNAs strongly determine the outcome of infection by Salmonella Typhimurium [20]. Importantly, functional characterization of selected miRNAs identified through this approach, and of their targets, revealed novel pathways and molecular players crucial to Salmonella infection.
Here, we demonstrate that miRNAs modulate infection by Shigella, and characterize the dual regulatory role of miR-29b-2-5p during infection, which results from the concomitant regulation of bacterial capture by host cells and of intracellular bacterial replication. Moreover, we show that miR-29b-2-5p through the repression of its direct target UNC5C, a member of the UNC5 netrin-1 receptor family, induces a dramatic increase in the formation of host cell filopodia that are essential for increased bacterial capture and internalization.
To systematically identify miRNAs that modulate infection by Shigella, we performed a microscopy-based high-throughput screening of a genome-wide library of miRNA mimics (Fig 1A). Representative images of the 10 highest ranking candidates increasing Shigella infection, as well as the percentages of Shigella infected cells upon treatment with these 10 miRNAs, are shown in S1A and S1B Fig and S1 Table. Among the miRNAs favoring Shigella infection more efficiently, we focused on miR-29b-2-5p, which strikingly increased both the percentage of infected cells and the number of intracellular bacteria per cell (Figs 1A, S1A and S1B).
To validate and further characterize the effect of miR-29b-2-5p on Shigella-host interaction, infection was assessed at early, intermediate and late times post-infection (0.5, 3 and 6 hpi, respectively) using three complementary approaches: i) fluorescence microscopy analysis, ii) colony-forming unit (cfu) assays, and iii) qRT-PCR. The results obtained consistently revealed an increase of Shigella infection elicited by transfection of miR-29b-2-5p mimic, at the three time points tested (Figs 1B, 1C and S1C). In all experiments, a C. elegans miRNA (cel-miR-231) was used as negative control; this miRNA has no sequence homology to any known human miRNA. Time-lapse microscopy documenting the dramatic increase of Shigella infection by miR-29b-2-5p is included as supplementary material (S1 Video); images extracted at different time points are shown in S2A Fig. It should be noted that for all the experiments (except bacterial binding to host cells), cells were treated with the antibiotic gentamicin to kill the non-internalized bacteria. To confirm the efficacy of the gentamicin treatment, we performed immunofluorescence labelling of Shigella before and after a permeabilization step. In contrast to what was observed in permeabilized cells, no labelling with the Shigella antibody was detected in non-permeabilized cells, either at early (0.5 hpi) or late (6 hpi) time-points of Shigella infection (S1D Fig), confirming that the gentamicin treatment was effective in killing extracellular bacteria.
A significant increase of the fraction of cells with bound bacteria was observed for cells transfected with miR-29b-2-5p mimic (wild-type, WT; Figs 1D, 1E and S2B). This was frequently accompanied by an increase in the formation of actin rich membrane ruffles (Fig 1D), clearly indicating that miR-29b-2-5p promotes a productive interaction of Shigella with host cells. To uncouple Shigella binding to host cells from bacterial invasion, we performed parallel experiments with the Shigella ΔIpaB mutant strain, which binds efficiently to host cells but is impaired in the subsequent steps of invasion [21]. Binding of Shigella ΔIpaB to host cells was also increased by miR-29b-2-5p (ΔIpaB; Figs 1D, 1E and S2B), to a similar extent as the wild-type strain, demonstrating that miR-29b-2-5p promotes Shigella binding with host cells, prior to the invasion step.
To determine if the increase in Shigella infection upon miR-29b-2-5p treatment observed at early times post-infection (2.5-fold at 0.5 hpi) could explain the strong increase observed at later times post-infection (4.1- and 9.9-fold at 3 and 6 hpi, respectively) we performed a time-course experiment using different Shigella MOIs (S2C Fig). We observed that infection with MOI 50 results in a 2.6-fold increase of bacterial invasion compared to MOI 10 (0.5 hpi). This difference is comparable to the increase in invasion observed in cells transfected with miR-29b-2-5p mimic in relation to control miRNA (compare Figs 1C and S2C). Interestingly, this difference was maintained at later times of infection (2.3-fold at 3 hpi and 2-fold at 6 hpi, S2C Fig). These results demonstrate that an increase in invasion per se cannot explain the striking increase in bacterial load observed in cells treated with miR-29b-2-5p at later time points, which should thus be ascribed to a cumulative effect of the miRNA on bacterial invasion and intracellular replication. Accordingly, normalization of the results of the cfu experiments at 3 and 6 hpi to the amount of bacteria internalized at the early time point (0.5 hpi) revealed a further increase of bacterial infection at later time points, in cells treated with miR-29b-2-5p (S2D Fig). Moreover, a detailed single-cell analysis of Shigella infected cells performed at different time points showed that the number of bacteria bound per infected cell was similar between cells transfected with miR-29b-2-5p and control miRNA (Fig 1F; binding), while at intermediate and late times post-infection the number of bacteria per cell was significantly increased in cells transfected with miR-29b-2-5p (Fig 1F; 3 and 6 hpi). Overall, these results show that miR-29b-2-5p favors two independent steps of Shigella interaction with host cells: it promotes binding of bacteria to host cells, and increases Shigella intracellular replication.
We observed that miR-29b-2-5p increases infection of the Shigella ΔIcsA mutant strain, which is unable to spread to adjacent cells [22, 23], similarly to the wild-type Shigella (S3A–S3C Fig), suggesting that the effect of miR-29b-2-5p is not related to spreading. Quantification of the area of Shigella infection foci, using automated image analysis, revealed indeed that the average foci size was comparable in miR-29b-2-5p treated cells and control cells (S3D and S3E Fig). Shigella ΔIcsA mutant strain was used as control. Of note, infection of cells transfected with miR-29b-2-5p mimic was performed with a lower MOI than that used for control cells, to achieve a comparable level of bacterial invasion. Together, these results exclude the involvement of miR-29b-2-5p in the actin based spreading of Shigella to neighboring cells.
When analyzing the fraction of cells containing internalized bacteria (Shigella +), we observed that, at late times post-infection, cells transfected with miR-29b-2-5p exhibited significantly higher cell death than cells transfected with a control miRNA (6 hpi; Figs 1G and S2E). These results indicate that the increase of intracellular replication prompted by miR-29b-2-5p decreases viability of the infected cells.
Interestingly, miR-29b-2-5p had no effect on the early or late stages of infection by Salmonella Typhimurium, a closely related bacterial pathogen (4 and 20 hpi; S3F–S3H Fig). Binding experiments revealed a mild but significant increase of Salmonella binding to host cells treated with miR-29b-2-5p mimic compared to control miRNA (S3I–S3K Fig). Importantly, this effect was clearly less pronounced than that observed in Shigella binding experiments (1.3-fold for Salmonella, S3I–S3K Fig vs. 3.2-fold for Shigella, Figs 1D, 1E and S2B). The binding experiments were performed with Salmonella WT and a mutant strain (Salmonella Δ4), which binds efficiently to host cells, but it is incapable of triggering membrane ruffling and invasion due to the lack of four key effector proteins (SipA, SopE, SopE2 and SopB) [24–26]. These findings are in full agreement with the results from a systematic analysis of miRNAs modulating Salmonella infection that we have performed recently [20]. It should be noted that from the 10 miRNAs that increased Shigella infection more efficiently in the screening, only 6 were present in the study by Maudet et al. [20], which was performed with an earlier version of the library of miRNA mimics (988 mimics, miRBase Release 13.0, 2009). Importantly, only one of these miRNAs slightly increased Salmonella infection (miR-365, 1.8-fold increase; S1 Table).
Taking into account the strong dependence of Shigella infection on the levels of host miR-29b-2-5p, we hypothesized that the levels of this miRNA could be altered during infection. MiR-29b-2-5p originates from the processing of pri-miR-29b-2/c [27], which also results in the production of three other mature miRNAs, specifically miR-29b-3p, miR-29c-5p and miR-29c-3p (Fig 2A). Importantly, miR-29b-2-5p has a unique seed sequence (highlighted in red in Fig 2A), not shared by any of the ca. 2,500 miRNAs currently annotated in the human genome; miR-29b-3p, which has a seed sequence identical to miR-29a-3p and miR-29c-3p, can also originate from a different pri-miRNA (pri-miR-29b-1/a [27]).
qRT-PCR specific for the mature miR-29b-2-5p revealed a decrease in the levels of this miRNA during the course of infection (Fig 2B). The levels of miR-29b-2-5p were significantly decreased at late times post-infection (6 hpi), but remained unchanged at 0.5 and 3 hpi, even at high multiplicity of infection (MOI 100; Fig 2B). Infection with the Shigella ΔIcsA mutant strain, invasive and replication competent albeit defective in intercellular spreading [22, 23], reduced miR-29b-2-5p levels similarly to wild-type Shigella (Fig 2C), whereas infection with invasion-deficient strains (ΔIpaB and BS176) and incubation with heat-killed Shigella did not change miR-29b-2-5p levels (Fig 2C). Moreover, analysis of miR-29b-2-5p levels in sorted populations of cells with internalized Shigella (Shigella +) and bystander cells (Shigella -) at 6 hpi showed that the decrease in miR-29b-2-5p occurs exclusively in the Shigella + cell population (Fig 2D). Collectively, these data demonstrate that Shigella infection decreases miR-29b-2-5p and that this effect is dependent on bacterial replication and is restricted to cells with internalized bacteria.
To investigate whether Shigella-induced regulation of miR-29b-2-5p occurs at the level of miRNA biogenesis or stability, we quantified the levels of the corresponding primary miRNA transcript (pri-miR-29b-2/c; Fig 2A), which remained unchanged upon Shigella infection (Fig 2E). Apart from miR-29b-2-5p, the levels of the other mature miRNAs resulting from the processing of pri-miR-29b-2/c remained unchanged (Fig 2D). These results indicate a selective regulation of miR-29b-2-5p at the post-transcriptional level, most likely through modulation of mature miRNA stability.
Although various regulatory mechanisms have been shown to affect miRNA biogenesis and activity, less is known about the regulation of mature miRNA levels through degradation. To evaluate if known miRNA-degrading enzymes are involved in the turnover of miR-29b-2-5p during Shigella infection, we tested whether the knockdown of RRP41 (ribosomal RNA-processing protein 41; core component of the exosome complex), PNPT1 (polyribonucleotide nucleotidyltransferase 1, aka PNPaseold-35) and XRN1 [28], could affect the regulation of miR-29b-2-5p in the context of infection. Interestingly, knockdown of PNPT1 significantly impaired the decrease of miR-29b-2-5p levels in infected cells (Fig 2F), indicating that PNPT1 is involved in the degradation of miR-29b-2-5p during Shigella infection. Time course experiments of Shigella infection revealed that both bacterial binding and intracellular replication are increased upon PNPT1 knockdown compared to cells treated with a control siRNA (S4A–S4F and S4J Fig). The increase at late times post-infection (6 hpi) can, at least in part, result from a decreased degradation of miR-29b-2-5p in infected cells. However, the increase of infection observed in PNPT1 knockdown cells at the early times post-infection (binding and 0.5 hpi) cannot be explained by changes in the miR-29b-2-5p expression. Moreover, we have analyzed the expression of PNPT1 during the course of infection (0.5, 3 and 6 hpi) and in sorted populations of cells with internalized Shigella (Shigella +) and bystander cells (Shigella -). Neither the PNPT1 mRNA (S4G Fig) nor protein levels (S4H and S4I Fig) were noticeably changed upon Shigella infection, indicating that Shigella infection is not affecting the expression of PNPT1, but presumably enhancing its exonuclease activity. Overall, these results show that PNPT1 plays an important role in Shigella infection, in part through the regulation of miR-29b-2-5p levels.
To determine if decreasing miR-29b-2-5p expression levels affects Shigella interactions with host cells, we have applied two experimental approaches: i) transfect cells with a miR-29b-2-5p inhibitor (LNA-based antisense oligonucleotide against mature miR-29b-2-5p); and ii) generate miR-29b-2 knockdown cell lines using CRISPR/Cas9 mediated genome editing [29]. Using these two strategies, we achieved a decrease in miR-29b-2-5p expression of more than two-fold, compared to the control (S5A Fig), which correspond to the levels observed in Shigella infected cells at late times of infection (6 hpi; Fig 2B). It should be noted that in the case of miR-29b-2 knockdown cell lines, the expression of the whole precursor miRNA is affected and therefore in addition to miR-29b-2-5p, miR-29b-3p levels were also reduced. The results of both approaches revealed that the interaction of Shigella with host cells was impaired upon miR-29b-2-5p knockdown. Specifically, we observed a decrease in binding (S5B–S5D Fig), as well as of intracellular bacteria at 0.5 and 6 hpi (S5E–S5G Fig). The decrease in Shigella infection was similar at the binding, invasion (0.5 hpi) and intracellular replication (6 hpi) steps. The lack of a more pronounced effect at 6 hpi (compared to binding or 0.5 hpi) can most likely explained by Shigella infection inhibiting miR-29b-2-5p expression (6 hpi, Fig 2B), resulting in a comparable inhibition of replication in control and miR-29b-2-5p knockdown cells at this time point. Together with the data obtained with the miR-29b-2-5p mimic, these results corroborate the concept that miR-29b-2-5p levels are a strong determinant of the extent of Shigella infection.
Considering that Shigella infection leads to a decrease in miR-29b-2-5p levels, we reasoned that a subset of the genes up-regulated upon infection correspond to derepressed miR-29b-2-5p targets. Therefore, to identify potential direct or indirect miR-29b-2-5p targets that are relevant in the context of Shigella infection we compared the genes down-regulated upon miR-29b-2-5p overexpression with those up-regulated at late times of Shigella infection (6 hpi; Fig 3A). Transcriptome analysis of HeLa cells transfected with miR-29b-2-5p and control miRNA, performed by deep-sequencing, identified 1,915 putative targets of miR-29b-2-5p (≥1.5 down-regulation; reads ≥25 for cells transfected with control miRNA; Fig 3A). Analysis of Shigella infected cells revealed 173 up-regulated genes, compared to mock treated cells (6 hpi; ≥2-fold up-regulation Shigella + population relative to mock; ≤ 2-fold change Shigella - population relative to mock; reads ≥25 for mock; Fig 3A). Overall, we identified 52 genes that are both repressed by miR-29b-2-5p and have increased expression in Shigella infected cells (S2 Table).
To pinpoint, amongst these genes, those relevant to Shigella infection, we performed a targeted RNAi screening by individually knocking down 46 genes, for which siRNAs were available (Fig 3A, S3 Table). Knockdown of 6 genes (UNC5C, ZCCHC24, MAFB, SH3RF2, VAV3 and PALLD) increased the percentage of Shigella infected cells by at least 2.5-fold at 6 hpi, compared to a control siRNA (Fig 3B and 3C, S3 Table). Interestingly, despite the comparable increase in the percentage of infected cells, knockdown of these transcripts did not increase Shigella intracellular load as efficiently as miR-29b-2-5p (Fig 3C, S3 Table), suggesting that the phenotype induced by the miRNA on infection ensues from its cumulative effect on multiple targets, presumably acting at different stages of infection.
Among the putative miR-29b-2-5p targets with a marked effect on Shigella infection, we focused our attention on the top hit of the siRNA screening—UNC5C, a member of the UNC5 netrin-1 receptors. This receptor family has been implicated in various cellular processes [30], including neuronal migration, embryonic angiogenesis and control of cell survival, but has not been previously associated to host-pathogen interactions. Consistent with UNC5C being a target of miR-29b-2-5p, we confirmed that the expression of UNC5C was increased at late times of Shigella infection, inversely correlating with miR-29b-2-5p (compare Figs 3D and 2B) and that UNC5C expression was decreased upon transfection of miR-29b-2-5p (Fig 4E). To provide evidence of direct binding of miR-29b-2-5p to UNC5C transcript, we first generated a reporter construct containing the full-length UNC5C 3’UTR cloned downstream of the Renilla luciferase coding sequence (psiCHECK-2-UNC5C 3’UTR). Firefly luciferase expressed from the same vector was used for normalization. As shown in Fig 3F and 3G, miR-29b-2-5p repressed Renilla luciferase activity significantly, in comparison with the control miRNA, denoting direct targeting of the UNC5C 3’UTR by miR-29b-2-5p. Given the length of the UNC5C 3’UTR (>6,700 nt) and the presence of multiple predicted miR-29b-2-5p binding sites (using TargetScan [15] and RNAhybrid [31]), we subsequently generated reporter constructs containing different fragments of the UNC5C 3’UTR, as well as mutant constructs to finely map the miR-29b-2-5p binding sites. This strategy identified of two binding sites in the 5’-end of the UNC5C 3’UTR (Fig 3F and 3G).
Both the knockdown and knockout of UNC5C significantly increased Shigella infection, when compared with control siRNA or parental cells, respectively (Figs 4A, 4B and S6A). A comparable increase of infection was observed at 0.5 hpi upon transfection of siRNAs targeting UNC5C, in UNC5C KO cells and upon transfection of miR-29b-2-5p mimic (ca. 2.5-fold; compare Figs 4A, 4B, S6A, 1B, 1C and S1C). At late times post-infection (6 hpi), miR-29b-2-5p treatment led to an increase of Shigella intracellular load significantly stronger than that obtained upon UNC5C knockdown or knockout (ca. 2.2- vs 9.9-fold; compare Figs 4A, 4B, S6A, 1B, 1C and S1C), whereas the number of infected cells remained comparable (Figs 1B and 4B). Taken together, these observations support the involvement of UNC5C in the early interaction between Shigella and host cells. Accordingly, binding of Shigella wild-type, as well as ΔIpaB strain, to host cells was significantly enhanced by UNC5C depletion or knockout when compared with control siRNA transfected or parental cells, respectively (Figs 4C, 4D and S6B). The comparable number of bacteria present within infected cells at intermediate and late stages of infection (3 and 6 hpi) between UNC5C knockdown cells and control cells (S6C Fig) further demonstrates that UNC5C does not affect Shigella replication; this is in contrast with the strong increase in Shigella intracellular load in cells treated with miR-29b-2-5p (S6C Fig). Moreover, these results further demonstrate that increased invasion cannot per se explain the striking increase in bacterial load observed in cells treated with miR-29b-2-5p at late time points post-infection. Of note, knockdown of UNC5C by RNAi led to a decrease in UNC5C expression similar to that achieved by miR-29b-2-5p mimic (siRNA UNC5C; Fig 4E), and the UNC5C knockout cell line generated through CRISPR/Cas9 mediated genome editing [29] showed a complete absence of UNC5C expression (UNC5C KO; Fig 4E).
We observed a slight increase in cell death upon UNC5C knockdown compared to cells transfected with a control siRNA, at late times post-infection (6 hpi; Figs 4F and S6D). Despite this increase, cell death remained significantly lower than that observed in cells transfected with miR-29b-2-5p (compare Figs 4F, S6D, 1G and S2E). This is most likely related to the fact that miR-29b-2-5p dramatically increases Shigella intracellular replication, whereas UNC5C knockdown/knockout enhances Shigella binding/invasion to host cells but does not affect intracellular replication. These results indicate that the role of UNC5C in the control of cell survival is not preponderant to its function during Shigella infection. We have also confirmed that similarly to miR-29b-2-5p, UNC5C knockdown does not affect Shigella intercellular spreading (S3D and S3E Fig). In addition, expression analysis of a panel of pro-inflammatory cytokines and downstream proteins (IL-8, IL-6, TNF-α, TNFAIP3, IL-1α and IL-1β) revealed that miR-29b-2-5p and UNC5C knockdown either did not affect the expression of these genes or slightly increased their expression in mock-treated cells (Figs 4G and S6I). Their expression was sustained or increased in infected cells transfected with miR-29b-2-5p mimic or UNC5C siRNA compared to control cells, at both 3 and 6 hpi (Figs 4G and S6I). These results show that miR-29b-2-5p and UNC5C knockdown are not dampening the pro-inflammatory response to Shigella infection.
Consistent with previous reports showing that UNC5C is not expressed in colon cancer cell lines and patient samples due to promoter methylation [32, 33], transfection of miR-29b-2-5p mimic in HCT-8 colon cancer cells did not affect Shigella binding to host cells (Figs 4H, 4J and S6E). Of note, we confirmed the lack of UNC5C expression in HCT-8 cells (Fig 4L). A slight, but significant, reduction in Shigella binding to cells overexpressing GFP-UNC5C compared to control cells transfected with GFP alone was observed (1.3-fold; S6G Fig). This mild effect is most likely explained by the low transfection efficiency of HCT-8 cells (15–20% for GFP-UNC5C). In agreement, microscopy analysis revealed a significant reduction in Shigella bound to cells expressing GFP-UNC5C compared to GFP expressing cells (S6H Fig). Importantly, in CCD 841 CoN cells, a human normal colon cell line that expresses UNC5C at levels significantly higher than HeLa cells (Fig 4L), treatment with miR-29b-2-5p mimic or UNC5C knockdown resulted in an increase of Shigella binding to host cells (Figs 4I, 4K and S6F).
Overall, these results demonstrate that the modulation of UNC5C is determinant to the effect of miR-29b-2-5p on Shigella binding to host cells, including in colon epithelial cells, the biological target of Shigella infection.
Detailed examination of cell morphology, following F-actin staining, revealed that cells transfected with the miR-29b-2-5p mimic or UNC5C siRNA, as well as UNC5C KO cells, display a remarkable increase in the number of cell extensions when compared with control cells (Fig 5A). It has recently been shown that, during the early steps of interaction between Shigella and host cells, bacteria establish contacts with pre-existing filopodia extensions on the cell surface [3]. We thus hypothesized that miR-29b-2-5p, through the repression of its target UNC5C, could be increasing Shigella interaction with host cells through induction of filopodia formation.
To address this hypothesis, we performed scanning electron microscopy (SEM) experiments, which confirmed a dramatic increase of filopodia, both in periphery and in the dorsal surface of HeLa and CCD 841 CoN colon cells transfected with miR-29b-2-5p mimic or UNC5C siRNA, as well as in UNC5C KO cells (Figs 5B and S7E). Moreover, we observed an increase of host cells with Shigella associated with peripheral filopodia, as well as of cells with Shigella entrapped by dorsal filopodia (Fig 5C). Since Shigella is rapidly internalized following contact with the host cells, these experiments were performed with the Shigella ΔIpaB strain, which is able to bind but does not invade host cells, thus allowing easier analysis of the capture step. When using the wild-type strain, an increase of filopodia emanating from the ruffles formed at Shigella invasion sites was also observed in miR-29b-2-5p treated cells and UNC5C knockdown and knockout cells (S7A Fig). In agreement with the pivotal role of actin in filopodia dynamics [34, 35], treatment of cells with the inhibitor of actin polymerization cytochalasin D reverted the potentiating effects of miR-29b-2-5p mimic and UNC5C knockdown/knockout on Shigella interaction with host cells (S7B–S7D Fig).
Analysis of the effect of UNC5C knockdown on infection by other bacterial pathogens revealed a slight increase of Salmonella binding to cells treated with UNC5C siRNA (ca. 1.3-fold, S8A–S8C Fig), which was reflected in a comparable increase of intracellular bacteria at 4 and 20 hpi (S8D–S8F Fig). For Listeria monocytogenes, a food-borne gram-positive pathogen [36], a mild increase of binding to UNC5C knockdown cells was also detected (ca. 1.3-fold, S8G–S8I Fig), with a stronger increase detected at 20 hpi (2.4-fold, S8J–S8L Fig). Despite these slight changes, the UNC5C knockdown and consequent filopodia formation did not affect Salmonella and Listeria binding and invasion of host cells to the same extent as Shigella. Concordantly, scanning electron microscopy showed that the degree of interaction of Salmonella with the filopodia present in host cells is significantly lower than that observed for Shigella infection (compare Fig 5C and S7F Fig). Specifically, we did not detect long-distance Salmonella interactions with the longer cell periphery filopodia, as observed for Shigella. We did detect interaction of Salmonella with shorter dorsal surface filopodia, however this contact was not noticeably increased in cells treated with miR-29b-2-5p mimic or UNC5C siRNA (S7F Fig), as observed for Shigella (Fig 5C).
GTPases of the Rho family are small proteins involved in regulating the organization of the actin cytoskeleton, and thereby cell shape, adhesion and movement. Both RhoF (aka Rif) and Cdc42 have been shown to induce filopodia, albeit via distinct mechanisms and with different characteristics [37–39]. Interestingly, the knockdown of either RhoF or Cdc42 abolished the potentiating effect of miR-29b-2-5p mimic and UNC5C knockdown on filopodia formation (Figs 5D, 5E and S9A–S9C) and consequently on Shigella binding to host cells (S9D–S9F Fig). Moreover, increased Cdc42 activation, reflected by more GTP-bound Cdc42, was observed in extracts from UNC5C knockdown and knockout cells as well as of miR-29b-2-5p treated cells, when compared to control (Fig 5F).
These results demonstrate that the major event underlying the effect of miR-29b-2-5p, and consequent repression of its target UNC5C, on Shigella early interaction with host cells derives from the increased filopodia formation, which promotes bacterial capture.
MicroRNAs are emerging as essential players in the complex interaction between bacterial pathogens and host cells, participating as part of the host defense mechanisms against pathogens, or exploited by the microorganisms to manipulate host cell pathways to promote invasion, survival, replication and dissemination. The identification and characterization of miRNAs modulating infection by specific bacterial pathogens (e.g. Salmonella Typhimurium, Mycobacterium tuberculosis) is under way, however for most bacterial pathogens, including Shigella, this remains essentially unexplored.
In this work, we focus on miR-29b-2-5p, a miRNA with a seed sequence unique in the human genome, which favors Shigella infection very efficiently. Detailed characterization of the effect of this miRNA revealed that it plays a dual regulatory role by modulating two distinct steps of Shigella infection (Fig 5G): miR-29b-2-5p promotes bacterial capture by dramatically increasing filopodia formation and it increases Shigella intracellular replication leading to increased bacterial loads. Interestingly, this miRNA had a mild effect on Salmonella binding to host cells and no effect on intracellular replication. Indeed, we show that Salmonella binding to host cells is not markedly affected by the increase of filopodia elicited by miR-29b-2-5p. Moreover, we speculate that the differential effect of this miRNA on Shigella and Salmonella intracellular replication results from their different intracellular lifestyles: it is likely that miR-29b-2-5p is affecting a pathway(s) relevant for Shigella replication in the host cytoplasm that is not relevant for Salmonella, which mostly replicates in a vacuole.
Interestingly, we observed that miR-29b-2-5p levels are decreased at late times of Shigella infection. Regulation of mature miR-29b-2-5p occurs independently of the primary miRNA transcript (pri-miR-29b-2/c) and of the three other mature miRNAs that result from its processing (miR-29b-3p, miR-29c-5p and miR-29c-3p), the levels of which remain unchanged. Importantly, we show that this regulation is dependent on PNPT1, a human polynucleotide phosphorylase with 3’-to-5’ exoribonuclease activity [40], which has been implicated in the degradation of several mature miRNAs (e.g. miR-221 and miR-222) in melanoma cells [41]. To our knowledge, this is the first report of the involvement of an exoribonuclease in the degradation of mature miRNAs during infection by a bacterial pathogen. Moreover, knockdown of PNPT1 led to increased Shigella infection both at early and late time points, revealing its relevance in the context of infection, at least in part dependent on miR-29b-2-5p regulation. Future work will focus on characterizing how PNPT1 activity is modulated during Shigella infection that, as demonstrated in this study, does not occur at the transcriptional level and is not related to protein steady state levels. Our data showing that increased Shigella intracellular replication sustained by high levels of miR-29b-2-5p induce host cell death, suggest that dampening of miR-29b-2-5p levels can constitute a bacterial strategy to favor balanced intracellular replication, thus avoiding premature host cell death and allowing bacteria spreading to neighboring cells. Alternatively, it is conceivable that regulation of miR-29b-2-5p levels constitutes part of the host response, in an attempt to counteract Shigella infection. Importantly, the relatively high basal levels of miR-29b-2-5p in epithelial cells at the start of the infection, and the fact that reduction of miR-29b-2-5p levels occurs only at later stages of infection, ensure efficient capture of Shigella by host cell filopodia and consequent invasion.
A major challenge in the analysis of miRNA function concerns the identification of the key miRNA targets underlying a given phenotype [13]. Here, we have used a combination of transcriptome analysis followed by a targeted RNAi screening approach to pinpoint, among the putative targets of miR-29b-2-5p, those relevant in the context of Shigella infection. Treatment with siRNAs targeting 6 of these genes (UNC5C, ZCCHC24, MAFB, SH3RF2, VAV3 and PALLD) significantly increased the percentage of infected cells. In this manuscript we focused on the role of UNC5C, although the other hits clearly deserve further analysis. VAV3 is a guanine nucleotide exchange factor preferentially for RhoG, RhoA, and to a lesser extent Rac1 [42], which has been involved in Shigella invasion. Surprisingly, VAV3 depletion increases Shigella infection, and therefore it is likely that this effect is unrelated to invasion. Another interesting candidate is PALLD, a protein involved in regulation of actin dynamics, namely binding, bundling and polymerization [43]. MAFB is a transcription factor with important functions in development and differentiation, which has been shown to act as regulator of type I interferon transcription and of actin organization in macrophages [44, 45]. SH3RF2 and ZCCHC24 are significantly less studied and currently there is no relevant information that may relate their function to Shigella infection. Importantly, knockdown of none of the 46 tested genes increased Shigella intracellular load to the levels achieved by miR-29b-2-5p. Given that miR-29b-2-5p acts in two distinct steps of the bacterial infection cycle, it is likely that multiple independent targets are responsible for the two phenotypes. Accordingly, we found that UNC5C is the miRNA direct target responsible for increased bacterial capture by host cells, and that its knockdown fully recapitulates the miRNA phenotype observed at early stages of Shigella infection. Moreover, we could uncouple the effect of miR-29b-2-5p on bacterial binding to host cells from its effect on intracellular bacterial replication, by generating a UNC5C knockout cell line. This cell line constitutes an interesting resource to study not only the early stages of Shigella-host cell interaction, but also the intracellular steps of Shigella infection, overcoming current limitations associated with the low adhesion activity of Shigella in vitro and the use of bacterial strains expressing ectopic adhesins (e.g. E. coli AfaE adhesin [46]) or coated with poly-lysine [47].
Interestingly, we demonstrate that miR-29b-2-5p and its target UNC5C are involved in the regulation of filopodia formation in epithelial cells. Consistent with our results showing that UNC5C negatively regulates filopodia formation in human cells, the ortholog of UNC5C in C. elegans, UNC5, has been shown to inhibit filopodia protrusion in neuronal growth cones [48, 49]. Filopodia are thin actin rich cell protrusions, which extend beyond the cell body, and are present in multiple cell types. By sensing mechanical and chemical cues in the extracellular milieu, filopodia are involved in several processes, including cell motility, cell-cell communication, cell-matrix adhesion and exploration of the environment [50, 51]. Viruses, such as murine leukemia virus, human immunodeficiency virus and human papilloma virus, have been shown to adhere to and glide along the side of filopodia to entry sites located at the cell body [52]. Likewise, several bacteria, including Yersinia pseudotuberculosis and Campylobacter jejuni, have also been shown to adhere to filopodia prior to internalization [53, 54]. Filopodia have also been shown to play an important role in the capture of Shigella [3]. Our results demonstrate that miR-29b-2-5p, through repression of UNC5C, induces filopodia formation, leading to a dramatic increase in Shigella capture and consequent bacterial invasion. Interestingly, we observed Shigella interaction with filopodia present both in the cell periphery, as well as in the dorsal cell surface: in general, we found bacteria at a distance of the cell surface associated with the longer cell periphery filopodia, whereas bacteria closer to the cell body are mostly associated with dorsal surface structures. Based on these observations, we hypothesize that the cell periphery extensions are involved in long-distance capture of Shigella, whereas the shorter dorsal surface filopodia function in maintaining the captured bacteria closer to the cell body, prior to invasion. It has been previously suggested that capture of Shigella mediated by filopodia increases the amount of bacteria invading per entry site [3]. Our results showing a similar number of Shigella associated per cell, despite the increase in filopodia, rather indicate that filopodia increase the probability of bacterial capture, but not the efficiency of invasion at a given site. Mechanistically, our results demonstrate that the formation of filopodia sustained by miR-29b-2-5p and knockdown/knockout of UNC5C is dependent on Cdc42 and RhoF Rho-GTPases. It has been shown that Cdc42 induced filopodia are generally short and project from the cell periphery, whereas the structures induced by RhoF are usually longer and arise from both the cell periphery and dorsal surface [37–39]. Interestingly, the cells treated with miR-29b-2-5p or knockdown/knockout of UNC5C present a combination of both phenotypes, displaying both short and long filopodia projections.
This study constitutes the first demonstration that host miRNAs modulate infection by Shigella. In this context, it is particularly interesting that miR-29b-2-5p acts on two independent steps of the infection process. Further studies will reveal if the concomitant regulation of multiple steps of the interaction between bacterial pathogens and host cells by miRNAs is a recurrent feature, which may explain the very strong phenotypes observed for some miRNAs.
Human epithelial HeLa-229 cells (HeLa CCL-2.1, ATCC) were cultured in DMEM GlutaMAX containing 1.0 g l-1 glucose (Life Technologies, 21885), human colon cancer HCT-8 cells (ATCC CCL-244) in RPMI 1640 GlutaMAX (Life Technologies, 72400), human colon normal CCD 841 CoN cells (ATCC CRL-1790) in DMEM-F12 GlutaMAX (Life Technologies, 31331), supplemented with 10% fetal bovine serum (Biochrom, S0615-1047D). Cells were maintained at 37°C in a 5% CO2 humidified atmosphere.
Shigella flexneri serotype 5 strain M90T,Salmonella enterica serovar Typhimurium strain SL1344 expressing GFP constitutively from a chromosomal locus [55], and Listeria monocytogenes serovar 1/2a EGD-e are referred to as wild-type (WT) in this study. Shigella flexneri M90T, its plasmid cured derivative BS176, and Listeria monocytogenes serovar 1/2a EGD-e were kindly provided by Prof. J. Vogel (IMIB, University Würzburg, Germany). The isogenic ΔIcsA and ΔIpaB mutants were generated by deleting the coding sequence of IcsA and IpaB, respectively, and inserting the kanamycin cassette via phage lambda Red recombinase-mediated homologous recombination. The following primer pairs were used to generate the deletion mutants: ΔIcsA 5'-TTTTCAGGGGTTTATCAACCACTTACTGATAATATAGTGCGTGTAGGCTGGAGCTGCTTC-3' and 5'-AGAGAAATGCAGGACATCAACACGCCCTGCATTTTTATTAGGTCCATATGAATATCCTCCTTAG-3', ΔIpaB 5'-CTGGTTTTCCTCTTGCCAAAATATTGACTTCCACTGAGCTGTGTAGGCTGGAGCTGCTTC-3' and 5'-GGTATAAGGTCTGTGAGGGTTTTACCTATTATTTTGCCAAGGTCCATATGAATATCCTCCTTAG-3'; the following primers were used to verify the deletion: ΔIcsA 5'-AACACAGCTCTCATGTTTTGG-3' and 5'-AGGCATACCATCATGTGCAC-3', ΔIpaB 5'-AGTGCTTCGAACTCGTAATTC-3' and 5'-TATGCGCTGCAATCTGCTG-3'. Shigella flexneri strains expressing GFP and DsRed were obtained by transforming the pXG-1 [56] and pCIG (kindly provided by Dr. A. Zumsteg, MPI-IB, Berlin, Germany) plasmids, respectively.
Shigella and Salmonella were grown aerobically in Luria broth (LB), and Listeria was grown in Brain Heart Infusion (BHI) medium. When appropriate, media were supplemented with antibiotics (Shigella GFP: 20 μg ml-1 chloramphenicol, Shigella DsRed: 20 μg ml-1 ampicillin, Salmonella GFP: 20 μg ml-1 chloramphenicol).
For Shigella and Salmonella infections, overnight bacterial cultures were diluted 1:100 in LB and grown aerobically at 37°C with shaking until OD600 0.4 (Shigella) or OD600 2.0 (Salmonella). Bacteria were harvested by centrifugation (12,000 g, 2 min) and resuspended in complete medium. Unless otherwise specified, infections were performed at multiplicity of infection (MOI) of 10 (Shigella) or 25 (Salmonella). After addition of bacteria, cells were centrifuged at RT at 2,000 g for 15 min (Shigella) or 250 g for 10 min (Salmonella), followed by incubation at 37°C in a 5% CO2 humidified atmosphere for 15 min or 20 min, respectively. The medium was then replaced with fresh medium supplemented with 50 μg ml-1 gentamicin (defined as time point 0 of infection) and incubated for 30 min to kill extracellular bacteria. After this incubation, the medium was replaced and cells were maintained in medium supplemented with 10 μg ml-1 gentamicin, until analysis. For Listeria infections, overnight cultures were diluted 1:50 in BHI medium and grown at 37°C with shaking until OD600 0.7. Dilution of bacterial cultures and infections were performed as described above for Salmonella.
For the quantification of Shigella spreading, infections were performed as described above; cells were infected at MOI 5 (control miRNA) or 1.25 (miR-29b-2-5p mimic and UNC5C siRNA) to ensure comparable levels of infection between the different treatments at 0.5 hpi.
For treatments with heat-killed (HK) Shigella, after reaching OD600 0.4, bacteria were harvested by centrifugation as described above, washed twice in PBS, incubated for 3 h at 100°C and then resuspended in complete medium. Incubation of cells with heat-killed Shigella was performed as described above for live bacteria.
For binding experiments, cells were exposed to Shigella WT or ΔIpaB mutant strain (MOI 50), centrifuged at 2,000 g for 15 min and incubated for 10 min at 37°C, 5% CO2. Cells were then extensively washed with PBS to remove non-bound bacteria and immediately processed for microscopy, cfu assay or RNA extraction. For Salmonella and Listeria a MOI 50 was used for binding experiments, and cells were processed as described above immediately after the centrifugation step (250 g, 10 min). To block actin polymerization, when indicated HeLa-229 cells were treated with 0.5 μg ml-1 cytochalasin D (CytD; Sigma, C8273), or DMSO (control), for 90 min prior to infection.
To quantify intracellular bacteria replication by colony forming units (cfu) assay, at the indicated time points cells were washed three times in PBS and lysed in PBS containing 0.1% Triton-X-100. The cell lysates were then serially diluted in PBS and plated on LB agar plates.
Transfection of miRNAs mimics and siRNAs into HeLa-229 and HCT-8 cells was performed with Lipofectamine RNAiMAX (Life Technologies), using a standard reverse transfection protocol at a final miRNA/siRNA concentration of 50 nM, as described previously [20]. Co-transfections experiments were performed with 25 nM of each miRNA/siRNA at a final concentration of 50 nM. For CCD 841 CoN cells, forward transfection at a final miRNA/siRNA concentration of 50 nM using Lipofectamine RNAiMAX was performed, according to the manufacturer’s instructions.
The library of miRNA mimics corresponding to all the human mature miRNAs annotated in miRBase 19 (2,042 mature miRNA sequences) was obtained from Dharmacon, GE Healthcare; for the targeted siRNA screening, selected siRNAs were ‘cherry-picked’ from the human genome-wide siRNA library stock plates (siGENOME SMARTPools, 4 siRNAs per gene, Dharmacon, GE Healthcare); gene IDs and gene symbols are included in S3 Table. MiRIDIAN hsa-miR-29b-2-5p mimic (C-301149-01), siGENOME SMARTpool UNC5C siRNA (M-010628-01), siGENOME SMARTpool Cdc42 siRNA (M-005057-01), siGENOME SMARTpool RhoF siRNA (M-008316-00-01), siGENOME non-targeting siRNA #5 (D-001210-05) and miRIDIAN microRNA mimic negative control #4 (cel-miR-231; CN-004000-01) were purchased from Dharmacon, GE Healthcare. PNPT1 (SASI_Hs01_00228542), RRP41 (SASI_Hs01_00239912) and XRN1 (SASI_Hs01_00212296) pre-designed siRNAs were purchased from Sigma. The hsa-miR-29b-2-5p miRCURY LNA Power inhibitor (4103894–101) and the negative control A miRCURY LNA Power inhibitor (199006–101) were purchased from Exiqon.
The miRNA and siRNA screenings were performed in 384-wellplates, essentially as described previously [20]. Briefly, miRNAs and siRNAs were transferred robotically from stock library plates and arrayed in 384-well plates (Viewplate-384 black, clear bottom, PerkinElmer). Transfection of miRNA mimics and siRNAs was performed using a standard reverse transfection protocol with Lipofectamine RNAiMAX, at a final miRNA/siRNA concentration of 50 nM; 1.5×103 HeLa cells were seeded per well. Forty-eight hours after transfection, cells were infected with Shigella (MOI 25), as described above. Cells were fixed at 6 hpi and counterstained with HCS CellMask Deep Red stain (1:15,000; Life Technologies) and Hoechst 33342 (1:5,000; Life Technologies), according to the manufacturer’s instructions. Transfection was optimized using a toxic siRNA targeting ubiquitin C. The screening was performed in triplicate.
For cfu assay, RNA isolation and microscopy, 5.0×104 HeLa and CCD 841 CoN cells or 8.0×104 HCT-8 cells were seeded per well in 24-well plates. Infections with Shigella and Salmonella were performed as described above, 60 h (Shigella) or 48 h (Salmonella) after transfection with miRNA mimics or siRNAs.
For HCT-8 cells transfection with plasmid DNA, 7.0×104 cells were seeded per well in 24-well plates, and grown for 48 h to reach ca. 70% confluency. Cells were then transfected with Lipofectamine 3000 (Life Technologies), with 1 μg of pEGFP-C1 or pEGFP-C1-UNC5C, according to the manufacturer’s instructions. Transfected cells were infected as described above, 48 h after plasmid transfection.
Cells were fixed with 4% paraformaldehyde (PFA) for 15 min at RT, permeabilized with 0.5% Triton-X-100 in PBS for 10 min. Blocking was performed for 30 min in 1% Bovine Serum Albumin (BSA) in PBS. Cells were then stained with a primary antibody directed against Shigella LPS (1:150, 2 h RT; Gentaur, YCC310-130) or Listeria (1:750, 2 h RT; antibodies-online, ABIN 237765) diluted in blocking buffer. Cells were further washed and incubated with the corresponding secondary antibody conjugated with Alexa Fluor 488 (1:400, 1 h RT; Life Technologies, A21441).
For F-actin staining, following the blocking step cells were incubated with Alexa Fluor 594 Phalloidin (1:150, 1 h RT; Life Technologies, A12381).
When indicated, cells were stained with HCS CellMask Deep Red stain (1:10,000, 1 h RT; Life Technologies, H32721) and nuclei were counterstained with Hoechst 33342 (1:5,000, 15 min RT; Life Technologies, 62249). Slides were mounted in Vectashield (VectorLabs).
To confirm the efficacy of the gentamicin treatment, a control experiment based on immunofluorescence labelling of Shigella before and after a permeabilization step was performed. Briefly, HeLa-229 cells were infected with a Shigella strain constitutively expressing mCherry, as described above. At 0.5 and 6 hours after infection, cells were washed in PBS three times and immediately incubated with a rabbit polyclonal anti-Shigella antibody (1:50, 1 h RT; Gentaur, YCC310-130). Cells were then further washed with PBS, fixed with 4% PFA and incubated with Alexa Fluor 488 conjugated anti-rabbit secondary antibody (1:250, 1 h RT; Life Technologies, A21441). Of note, incubation with the primary antibody was performed in live cells, prior to fixation, since fixation can permeabilize the cell membrane. For each time point, a parallel set of samples was fixed and permeabilized prior to incubation with the primary antibody to ensure complete staining of intracellular bacterial, as described above.
For quantification of bacterial infection, image acquisition was performed using an ImageXpress Micro (Molecular Devices) or an Operetta (PerkinElmer) automated high-content screening fluorescence microscope at a 20× magnification; a total of 9–16 images were acquired per coverslip/well, corresponding to approximately 2,500 cells analyzed per experimental condition and replicate. Image analysis was performed using the ‘Multi-Wavelength Cell Scoring’ application module implemented in MetaXpress software (Molecular Devices), or Columbus image analysis software (PerkinElmer). Briefly, nuclei and cells were segmented based on the Hoechst and CellMask stainings, respectively; cells were then classified as positive or negative for Shigella depending on the total area of bacterial staining. For the miRNA and siRNA screenings, miRNA mimics or siRNAs that decreased the total number of host cells to less than 65% of the control were excluded from further analysis.
Time-lapse Shigella infection experiments were performed in an Operetta automated high-content screening fluorescence microscope equipped with the live cell chamber option for temperature and CO2 control. Prior to infection, cells were treated with Hoechst (1:100,000, 1 h RT) and then infected as described above, in 384-well plates. Images were acquired at 15 min intervals from 0.5 to 6 hpi, at 37°C, 5% CO2.
Confocal microscopy images were obtained using a laser scanning Leica SP5 confocal microscope (Leica Microsystems). For the analysis of the number of bacteria per cell, at least 50 infected cells per condition and independent experiment were counted manually from maximum-projected Z-stack confocal images.
For quantification of Shigella spreading, image acquisition was performed using an Operetta (PerkinElmer) automated high-content screening fluorescence microscope at a 10× magnification; a single image was acquired per well, corresponding to an area of ca. 1.4 mm2, approximately 2,000 cells were analyzed per experimental condition and replicate. Image analysis was performed using Columbus image analysis software (PerkinElmer). Briefly, Shigella infection foci were identified based on texture and intensity features, using the ‘Find Texture Regions’ building block implemented in Columbus image analysis software, and the size of each individual foci was determined using the ‘Calculate Morphology Properties’ building block. Infection foci of very small dimension (below 400 μm2; average HeLa-229 area 500 μm2) and containing low bacterial load were excluded since they represent non-productive infections; infection foci touching the borders of the images were also excluded, to improve the overall accuracy of the measurements. At least 500 individual foci from a total of 5 independent experiments were plotted per experimental treatment. Shigella ΔIcsA mutant strain, defective in intercellular spreading was used as a control.
Total RNA, including the small RNA fraction, was isolated in TRIzol (Life Technologies) and extracted by phenol-chloroform followed by isopropanol precipitation.
For quantification of gene expression, total RNA was reverse-transcribed using hexameric random primers (Life Technologies) and M-MLV reverse transcriptase (Life Technologies), according to the manufacturer’s instructions. Real-time quantitative analysis was performed using SsoAdvanced Universal SYBR Green Supermix (BioRad) according to the manufacturer’s instructions, using a CFX96 Touch Real-Time PCR detection system (BioRad). The following primer pairs were used: GFP 5'-ATGCTTTTCCCGTTATCCGG-3' and 5’-GCGTCTTGTAGTTCCCGTCATC-3'; DsRed 5'-TCATCTACAAGGTGAAGTTC-3' and 5’-AGCCCATAGTCTTCTTCT-3'; UNC5C 5'-GATTGTGATAGCAGTGAT-3' and 5'-CGATGATTCTTCCGATAC-3'; pri-miR-29b-2/c 5'-TTGCCACTTGAGTCTGTT-3' and 5'-CTTCTCTACTGTCACCTCTC-3'; Cdc42 5’-GCCAAGAACAAACAGAAGCCT-3’ and 5’- ACTTGACAGCCTTCAGGTCA-3’; RhoF 5’-AGCAAGGAGGTGACCCTGAAC-3’ and 5’- CCGCAGCCGGTCATAGTC-3’; β-actin 5'-CCTGTACGCCAACACAGTGC-3' and 5'-ATACTCCTGCTTGCTGATCC-3'. Expression was normalized to β-actin.
For miRNA quantification, reverse transcription was performed using the miRCURY LNA Universal cDNA synthesis kit (Exiqon) followed by qRT-PCR using miRCURY LNA SYBR Green master mix (Exiqon) and predesigned mercury LNA PCR primer sets (Exiqon), according to the manufacturer’s instructions. The following primer sets were used: hsa-miR-29b-2-5p (204208), hsa-miR-29b-3p (204679), hsa-miR-29c-5p (204132), hsa-miR-29c-3p (204729) and U6 (203907). MiRNA expression was normalized to U6.
The 2-ΔΔCt method was used to calculate fold changes.
Analysis of cell viability following infection with Shigella was performed using 7-Amino-Actinomycin (7-AAD; BD Biosciences, 51-68981E), a viability dye excluded from cells with intact membranes (viable cells), as described previously [20]. Flow cytometry was performed on a BD Accuri C6 flow cytometer (BD Biosciences); a minimum of 10,000 cells was analyzed for each subpopulation (Shigella - and Shigella +). FlowJo software (Tree Star Inc) was used for data analysis.
Cell sorting to separate the subpopulation of cells infected with Shigella (Shigella +) from bystander cells (Shigella -) was performed as described previously [20]. HeLa-229 cells infected with Shigella expressing GFP or DsRed at MOI 10, were collected at 6 hpi and sorted based on the intensity of the GFP or DsRed signal, using a FACSAria III flow cytometer (BD Biosciences). Control, mock-infected cells were processed in parallel; RNA isolated from the sorted cells was processed as described above.
After washing thoroughly with PBS five times, cells were fixed overnight at 4°C in 2.5% glutaraldehyde in 0.1 M cacodylate buffer (pH 7.2). Samples were then washed five times in 0.1 M cacodylate buffer (pH 7.2), dehydrated through a graded series of 30%, 50%, 70%, 90% acetone solution in 10 min incubations, and incubated six times in 100% acetone for 10 min, followed by critical point drying with CO2. Dried specimens were sputter coated with gold/palladium. Samples were analyzed with a JEOL JSM-7500F (JEOL GmbH) or a Zeiss Crossbeam 340 (Zeiss) field emission scanning electron microscope operating at 5Kv. Representative images were selected from 3 independent experiments.
The UNC5C knockout HeLa-229 cell line (UNC5C KO) and miR-29b-2 knockdown HeLa-229 cell lines (miR-29b-2 KD#1 and miR-29b-2 KD#2) were generated essentially as described previously [29]. Briefly, sgRNA sequences targeting the first exon of UNC5C or the miR-29b-2 precursor miRNA were designed using the online CRISPR Design Tool (http://tools.genome-engineering.org) from F. Zhang laboratory (MIT, USA). Selected sgRNAs were ordered as single strand DNA oligonucleotides (IDT), with the following sequences: UNC5C_sg1 5’-CACCGCACACCTCTGTCTACGATG-3’ and 5’-AAACCATCGTAGACAGAGGTGTGC-3’; UNC5C_sg2 5’-CACCGACAGCGGCCCGCTGCGGACT-3’ and 5’-AAACAGTCCGCAGCGGGCCGCTGTC-3’; miR-29b-2_sg1 5’-CACCGTTCTGGAAGCTGGTTTCACA-3’ and 5’-AAACTGTGAAACCAGCTTCCAGAAC-3’; miR-29b-2_sg2 5’-CACCGAATGGTGCTAGATACAAAGA-3’ and 5’-AAACTCTTTGTATCTAGCACCATTC-3’; miR-29b-2_sg3 5’-CACCGCCTAAAACACTGATTTCAAA-3’ and 5’-AAACTTTGAAATCAGTGTTTTAGGC-3’. The oligos were annealed and cloned directionally into the px330 vector using the BbsI enzyme; the px330 plasmid encodes a chimeric guide RNA and a human codon-optimized S. pyogenes Cas9 [29]. Equimolar amounts of the constructs encoding the guide RNAs targeting UNC5C or the miR-29b-2 genomic regions were transfected into HeLa-229 cells seeded on 24-well plates, using FuGENE HD (Promega, E2311) or Turbofect (Thermo Scientific, R0531) transfection reagent, according to the manufacturer’s instructions. To isolate clonal populations of edited cells, four days after transfection the cells were plated into 96-well plates, at a limiting dilution. Colonies originated from single cells were expanded and cell fractions were collected for genomic analysis. The genomic region encompassing the target sites of the designed sgRNAs was amplified using the following primers: UNC5C 5’-GTCGTTATTTCTTCGGACTGCTTC-3’ and 5’-AAGGAGGGAAGGAAGAAGCTAAG-3’; miR-29b-2 5’-GTGCTTGTGTCCTGATGAAGTA-3’ and 5’-AAATCGGTCAGCCTGTGTAAG-3’; genome editing was evaluated by Sanger sequencing. Knockout of UNC5C and knockdown of miR-29b-2-5p was validated by qRT-PCR, as described above.
The full length and fragments of the UNC5C 3′UTR were amplified by PCR and cloned into the dual luciferase reporter vector psiCHECK-2 (Promega) using PmeI or XhoI and NotI restriction sites. Mutants of miR-29b-2-5p putative binding sites were generated by site directed mutagenesis. The following oligonucleotides were used for cloning of the UNC5C 3’UTR, fragments and mutagenesis: full-length 5’- CGTGATGTTTAAACCACCATGCTGGAAGGGGAAA-3’ and 5’-ACGAGCGGCCGCTGGCTGACAGAATAAATGCAGG-3’; (1–2372) 5’- CGTGATGTTTAAACCACCATGCTGGAAGGGGAAA-3’ and 5’- GTACGAGCGGCCGCCTCATTGCTACAAACATCCACTG-3’; (2252–4681) 5’- CGTGATGTTTAAACATAAAATCTCTCTCAGCCCACC-3’ and 5’- GTACGAGCGGCCGCTATGAGGCACTCGAGAGGTA-3’; (4556–6730) 5’- CGTGATGTTTAAACTTGTGAAGCAGTACATCATTGC-3’ and 5’- ACGAGCGGCCGCTGGCTGACAGAATAAATGCAGG-3’; (1–612) 5’- CGATCGCTCGAGCCACCATGCTGGAAGGGGAAA-3’ and 5’- GTACGAGCGGCCGCGATTCTCCACTTCCCTGATAC-3’; (580–1680) 5’-CGATCGCTCGAGGTGTTACTGTTTGTATCAGGG-3’ and 5’- GTACGAGCGGCCGCGTCTAGATATCAAGATGCCAAG-3’; (1619–2372) 5’- CGATCGCTCGAGGCCTTCCTGACTGGTTGCAAAT-3’ and 5’- ACGAGCGGCCGCTGGCTGACAGAATAAATGCAGG-3’; mutant site1 5’- AGCTGAGGAGGAAATCCAGAATGCACTTCACAGGCAAGAC-3’ and 5’- GTCTTGCCTGTGAAGTGCATTCTGGATTTCCTCCTCAGCT-3’; mutant site 2 5’- GACTCTAAACTGGAGATCAATTCTCATTGTGGGTCTGTCAC-3’ and 5’- GTGACAGACCCACAATGAGAATTGATCTCCAGTTTAGAGTC-3’.
For luciferase reporter assay, HeLa cells were seeded in 96-well plates (8×103 cells/well) one day before transfection and were transfected with 200 ng of the constructs containing the wild-type or mutated sequences of the 3′UTR-UNC5C using FuGENE HD Transfection reagent. Twenty-four hours after plasmid transfection, cells were transfected with cel-miR-231 or miR-29b-2-5p mimics (final concentration 50 nM) and the experiment was stopped after an additional 48 h. Firefly and Renilla luciferase activities were measured using the Dual-Luciferase Reporter Assay Kit (Promega, E2920) according to the manufacturer’s protocol. Results were expressed as normalized Renilla/Firefly luciferase activity ratios.
Cells were washed in PBS, collected in Laemmli’s sample buffer and separated in 10% or 12% SDS-PAGE, followed by Western-blotting. The following antibodies were used: β-actin (1:3,000; Sigma, A2228) and PNPT-1 (1:750, Santa Cruz, sc-365049). Anti-mouse secondary antibody coupled to horseradish peroxidase were used (1:10,000; GE Healthcare, NA931). Signals were detected using SuperSignal West Dura Extended Duration Substrate (Pierce, 34075) and an ImageQuant LAS 4000 CCD camera (GE Healthcare).
Activation of Cdc42 was determined using the Cdc42 Pull-down Activation Assay Biochem Kit (Cytoskeleton, BK034), according to the manufacturer’s instructions. Briefly, prior to analysis HeLa cells were transfected in a 6-well plate (10 wells per condition, 1.5×105 cells/well) using a standard reverse transfection protocol with cel-miR-231, miR-29b-2-5p mimic or UNC5C siRNA, at a final concentration of 50 nM, and incubated during 72 h. For UNC5C KO and parental cells, cells were grown for 48 h (10 wells per condition, 2.4×105 cells/well). All cells were serum-starved for 24 h before collection. After starvation, cells were washed with PBS and scraped in lysis buffer. Equal amounts of total protein (600 μg) of each lysate were then incubated with the PAK-PBD beads, according to the manufacturer’s instructions. Total and activated GTP-bound Cdc42 (pulldown) were analyzed by SDS-PAGE and Western-blot as described above. Relative Cdc42 activation was calculated by densitometric analysis of the intensities of the Cdc42 bands from the pulldown (GTP-Cdc42) and total Cdc42 present in the sample, which was used for normalization.
Library preparation and deep-sequencing was performed by Vertis Biotechnology AG, as described previously [20]. RNA-seq analysis was performed using the READemption pipeline version 0.3.0 [57], with Segemehl version 0.1.7 [58]. Reads were mapped against the human (build GRCh37) and Shigella flexneri (NCBI Reference Sequence NC_008258) genomes. Analysis of differential gene expression was performed with DESeq 1.18.0 [59]. The demultiplexed FASTQ files and gene-wise quantifications have been deposited in NCBI’s Gene Expression Omnibus and are accessible through GEO Series accession numbers GSE75746 and GSE75747.
No statistical methods were used to predetermine sample size. The investigators were not blinded to allocation during experiments and outcome assessment. The data are presented as mean ± standard error of the mean (s.e.m.), of at least three independent experiments. The exact number of experiments performed for each panel is indicated in Figure legends. Statistical analysis was performed using Prism Software (GraphPad). For statistical comparison of datasets from two conditions, two-tailed Student’s t-test was used; for data from three or more conditions, ANOVA with Tukey’s or Dunnet’s post-hoc test was used. A P-value of 0.05 or lower was considered significant.
|
10.1371/journal.pcbi.1002522 | Emergence of a Small-World Functional Network in Cultured Neurons | The functional networks of cultured neurons exhibit complex network properties similar to those found in vivo. Starting from random seeding, cultures undergo significant reorganization during the initial period in vitro, yet despite providing an ideal platform for observing developmental changes in neuronal connectivity, little is known about how a complex functional network evolves from isolated neurons. In the present study, evolution of functional connectivity was estimated from correlations of spontaneous activity. Network properties were quantified using complex measures from graph theory and used to compare cultures at different stages of development during the first 5 weeks in vitro. Networks obtained from young cultures (14 days in vitro) exhibited a random topology, which evolved to a small-world topology during maturation. The topology change was accompanied by an increased presence of highly connected areas (hubs) and network efficiency increased with age. The small-world topology balances integration of network areas with segregation of specialized processing units. The emergence of such network structure in cultured neurons, despite a lack of external input, points to complex intrinsic biological mechanisms. Moreover, the functional network of cultures at mature ages is efficient and highly suited to complex processing tasks.
| Many social, technological and biological networks exhibit properties that are neither completely random, nor fully regular. They are known as complex networks and statistics exist to characterize their structure. Until recently, such networks have primarily been analyzed as fixed structures, which enable interaction between their components (nodes). The present work is one of the first empirical studies investigating the adaptation of complex networks [1]. Network evolution is particularly important for applying complex network analysis to biological systems, where the evolution of the network reflects the biological processes that drive it. Here, we characterize the functional networks obtained from neurons grown in vitro. Network properties are described at seven day intervals during the neurons' maturation period. Initially, neurons formed random networks, which spontaneously reorganized to a ‘small-world’ architecture. The ‘small-world’ concept derives from the study of social networks, where it is referred to as ‘six-degrees of separation’: the connection of any two individuals by as few as six acquaintances. In brain networks, this translates to rapid interaction between neurons, mediated by a few links between locally connected clusters (cliques) of neurons. This architecture is considered optimal for efficient information processing and its spontaneous emergence in cultured neurons is remarkable.
| The organizational properties of biological, technological and social systems are increasingly being characterized by representing them as abstract networks of interacting components and quantifying non-random features of their structure [2], [3], [4], [5], [6]. Many real-world networks have an organization (topology) that is neither completely random, nor completely regular. Termed complex networks, these typically afford excellent integration between their constituent parts yet they also provide tightly interconnected subnetworks that segregate efficient within-group interaction. An example is social networks, for which a seminal study [7] revealed that any two individuals in the world could communicate via only a small number (∼6) of mutual acquaintances. Such networks are sparse – only a tiny proportion of the world's population are associated, yet they are incredibly well-connected. The phenomenon has been termed ‘small-world’ – hence the concept of a small-world network.
For neuronal connectivity, the abstract network (graph-theoretic) approach to analysis has allowed common organizational principles to be identified at both the macroscale level of whole brain imaging [8], [9], [10], [11], and the microscale level of connections between individual neurons [12], [13]. Importantly, this form of analysis enables the relationship between neuronal network organization and (whole or partial) brain function to be investigated. There are numerous complex network statistics for assessing the non-random properties of these abstract networks (for review see [14]), each of these statistics enable direct comparison of results from diverse experiment modalities and over a range of species and scales [2], [5]. Moreover, properties may also be compared with those of networks from other domains [15]. Two important metrics are the level of integration and segregation; high levels of which are found in random and lattice networks, respectively. Since small-world networks have high levels of both properties, the extent to which a given network approximates or deviates from small-worldness may be evaluated by considering the balance between integration and segregation [16], [17]. This balance has become an important benchmark for the assessment of neuronal networks and the small-world topology has been found at multiple scales over a range of species in both structural [16], [18] and functional [9], networks. Moreover, its influence on network efficiency [8] and robustness [9] has been demonstrated, and deviation from the small-world topology has been associated with abnormal or decreased brain function [8], [19], [20], [22], [23].
The focus of the present study is the development of complex network properties within cultures of neurons, grown in vitro. Unlike in-vivo brain networks, where the range of experimental conditions is typically constrained by the availability of subjects with a given condition, or strict regulation regarding experimental manipulation, cultures of dissociated neurons grown on multi-electrode arrays (MEAs) provide an experimental platform for the long-term investigation and manipulation of neuronal cells. Such neurons spontaneously form connections [24], [25] and non-random properties have been found in the resulting structural network [13]. Moreover, cultures share several important characteristics with their in vivo counterparts [26], [27], [28], for review see [25]. Consequently, these preparations are increasingly being used in investigations of cellular and network processes that underlie complex cognitive functions [29], [30], [31], [32], [33] and as models of pathophysiological states (e.g. epilepsy and stroke [34]). Importantly, since the cultures have no pre-built infrastructure, they allow the network formation to be observed - making them well-suited to investigating neuronal network development in a living biological system.
Two aspects of the cultures that are of particular interest are their structural (anatomical) circuitry and the interactions which take place over this circuitry, both determining the computational capacity of the underlying network. Whilst cultures are typically too dense for accurate observation of their structural connectivity, analysis of functional connectivity provides a probabilistic estimation of the relationship between distributed neuronal units [2], thereby enabling spatio-temporal interactions between areas of the network to be measured throughout experiments. This provides a useful means to investigate the network properties of cultures, particularly since functional connectivity estimated over certain timescales may contain information about the underlying structural network [35].
Existing literature indicates that the functional network properties of cortical cultures change during maturation [36] and following stimulation [29], [37], [38]. However, such studies have focused on changes in the expected link-level properties such as the mean strength and metric distance of connections [36], or the proportion of links which are strengthened or weakened following stimulation [37], [38]. These aggregate measures capture gross changes in global connectivity, but they do not reflect the organizational features of the network, e.g. the distribution of properties amongst the neural units, or whether there are groups of neural units that are more densely connected than others. Analysis of such organizational features would reveal the architecture of the network, enabling investigation into which interactions the network could support and how the network organization changes under different experimental conditions. Importantly, by assessing the complex network properties, the relevance of results from cultures to investigations of whole-brain networks would be increased.
Reports that rigorously compare culture's complex network properties under different experimental conditions are very sparse. Mature cultures were assessed in [39] and networks from cultures subject to an in vitro glutamate injury model of epileptiform activity were assessed in [34]. The utility of cultures for investigating changes in cognitive function, characterizing drug effects and modeling disease states, could be greatly extended by applying complex network statistics to quantify the influence of experimental manipulation on the network architecture. Moreover, comparison with results from in vivo networks may reveal basic organizational principles common to both.
Experiments utilizing cultures can be undertaken across a range of ages, yet little is known about whether developmental changes occur in culture's complex network properties. Questions such as when and which non-random properties are present, their stability over time and the variability between cultures and their ages remain largely unanswered. The nature of such spontaneously occurring changes in a culture's functional network are important a priori knowledge for assessing experimental outcomes using complex network statistics. Moreover, by analyzing these ‘known’ conditions, a framework can be established for evaluating a variety of experimental conditions, including those resulting from embodying a culture in a closed loop system. [40], [41], [42], [43].
The density at which cultures are seeded exerts an important influence on the rate of maturation. Dense cultures mature faster than their sparse equivalents, and they demonstrate bursting activity earlier in development [44]. For the purpose of the present paper, dense cultures were deemed preferable, since their use enabled network properties to be measured earlier in development than would have been possible on much sparser cultures. Additionally, to investigate changes in the functional network properties during culture maturation, maintaining consistency in plating parameters was important to minimize differences in cultures structural properties. Such differences would have complicated the analysis and interpretation of results. Therefore, cultures at a fixed density were used (those described in [44] as ‘dense’). At ∼1,500 to 6,500 cells within the ∼1.6 mm2 recording area of the MEA, the cells in such cultures form a monolayer. Moreover, they can be maintained for many months [45] and the density is comparable to that used by other groups (typically ∼2,500–3,000 cells per mm2 [24], [30], [36], [42], [46], [47]).
The present study establishes the baseline network statistics for cultures at specified stages of development and uses them to characterize culture maturation. The topological, spatial and performance properties of functional networks captured every 7 days (7 to 35 days in vitro [DIV]) were compared using a population of 10 cultures. The study is one of the first to investigate functional connectivity in an evolving complex system. Here, the evolution of network properties is a counterpart of biological processes shaping the culture's development.
Since the graph-theoretic approach and use of complex network statistics is a relatively novel method for investigating functional connectivity in cultures, the key methodological decisions are described next.
Results are split into two sections. The first presents topological, then spatial network statistics from persistent networks. The second presents statistics on the propagation of activity over the network (from the transient networks). Network statistics were obtained for each culture at each age (DIV 7, 14, 21, 28, 35). NOTE, at DIV 7 only one culture was found to have a persistent network, therefore this age was not considered for the significance testing.
The number of nodes and links for a given culture was used to calculate the edge density of its network. Figure 2 shows the expected values for each property. The mean number of nodes was relatively constant and independent of age (P = 0.272). In contrast, the mean number of links measured at DIVs 14 and 21 was lower than at DIVs 28 and 35, with a strong trend towards a significant increase between the younger and older ages (P = 0.074). Edge density increased significantly between DIVs 14 and 21 (P = 0.012) and showed no significant change thereafter. Statistics quoted are for the n = 5–8 cultures valid for complex network analysis (see Materials and Methods). However, results were comparable when all cultures were used. Numbers of nodes and links followed a comparable trend for two different persistence thresholds (see Figure S1), indicating their robustness to threshold selection. Edge density followed different trends for the different link persistence thresholds; this was due to small differences in the numbers of nodes at each age, resulting in larger differences in edge density (Figure S1).
Results presented thus far have focused on identifying changes in the network infrastructure (via the persistent interactions between different areas [nodes] in the cultures). Here, the results focus upon the activity that takes place over this infrastructure. Each transient network is considered as a ‘snapshot’ of network activity, measured over a short time-scale (duration of a network-wide burst) and reflects interactions between different areas of the culture in this period.
As per the persistent networks, the basic properties relating to network size were compared. Additionally, since there were multiple transient networks for each culture, the coefficient of variation was also analyzed (see Materials and Methods). Figure 8 panel A shows the expected number of transient links as a function of culture age, panel B shows the equivalent data for number of nodes. There was a strong trend towards an increase in the mean number of transient network links (P = 0.087), and a strong trend towards an increase in the number of nodes (P = 0.089). Figure 8 panel C shows the expected coefficient of variation for the number of transient links. This was largest at DIV 21 and there was a significant increase in coefficient of variation between DIV 14 and DIV 21 (P = 0.021). This demonstrated that transient networks at DIV 21 varied considerably in their numbers of links, more so than at any other age. Panel D shows the equivalent data for number of nodes (no significant difference).
The present study characterizes the evolution of functional networks observed in cortical cultures and extends previous work where network properties of cultures were investigated at a single developmental stage [34], [39]. Analysis of activity from multiple bursts allowed the identification of frequently activated links - the persistent network, which was robust to inter-burst fluctuations in activity and suitable for analysis of complex network statistics. Results demonstrated that cortical cultures exhibit developmentally-dependent structured interactions, which are reflected in their persistent patterns of activity. These data suggest the evolution of a complex network of links that supports increasingly efficient information flow and specialized processing. Given the absence of external chemical or electrical stimulation applied to the cultures, these findings support the assertion that such complex network evolution is an intrinsic property of neuronal maturation. Moreover, the characterization of age-dependent network properties enables appropriate selection of culture development stages for specific experiments [24], [37], [38], [42].
Immature cultures (DIV 14) exhibited limited interactions between neuronal units, resulting in a network of few nodes and links. The observation that at DIV 14 activity could spread rapidly between any two neuronal units (short mean path length in Figure 3, reflects high integration), but was slow to propagate network-wide (Figure 9) indicated an absence of functional organization. The homogeneous node degree distribution and low clustering coefficient exemplified the poor functional differentiation between nodes, with no evidence of densely interconnected areas that could support segregation of neural processing. Together, these network properties implied a disordered spread of activity, across a random network topology, whilst the long burst propagation time indicated an inefficient structure for widespread information transfer. Since dissociated neurons were seeded randomly onto the MEA and received no external stimulation, it could be expected that their initial connectivity resulted in a random topology. Moreover, since neuron-synapse maturation is incomplete at DIV 14 [24], [53], it is unsurprising that the complex network properties found in mature cultures [39] were not present at this age. However, the prevalence of long-distance connections at DIV 14 (Figure 5 and [36]) is counter to the economy of wiring principle [54] and suggests that units are not simply making spatially convenient connections. In in vivo and ex vivo preparations the cell type and neurochemical identity have been proposed as guiding influences for connectivity [55] and there is evidence that the variety and proportions of neuron types in cortical cultures are similar to those found in vivo [25], [27], [56], therefore connectivity in cultures could be similarly guided by these influences.
Whilst interactions at DIV 14 were clearly unstructured, the subsequent 14 days of development represented a critical window, during which functional complexity increased (Figure 3), leading to the emergence of the small-world topology at DIVs 28 and 35. Figures S2, S3 and S4 demonstrate the robustness of the small-world result.
We consider the possible driving forces behind this topology change to include the level of synchronization, the ratio of excitation-inhibition and the mechanism of Hebbian learning.
Synchronization of culture activity can be defined over a range of timescales – from ‘synchronous busting’ [57], where areas of the network are synchronously active (usually to within ∼100 milliseconds), to precise synchronization between the spike times of two or more neural units [36] (usually to within ∼10 ms or less). For the present study, the network links were derived from firing-pattern correlations and thus represent synchronization levels between neural units (nodes); the low number of nodes and links at DIV 14 reflects a low level of synchronization (i.e. between only a few units), compared to a high level of synchronization (i.e. between many units) at DIVs 28 and 35. Literature indicates that a low level of synchronization at DIV 14 may be due to an excitatory-inhibitory imbalance [53]. Conversely, evidence suggests that a high level of network-synchronization found in older cultures (whereby many neural units are activated within a short time-window [24]) is supported by a balanced excitatory-inhibitory subsystem [53], with tight synchronization between pairs of neural units (as observed in [36]) arising from the activity-based refinement of synaptic connection strengths [24], [58].
In a previous study of functional connectivity during development [36] culture properties at DIV 14 and DIVs 28–35 are in accordance with those of the present study. However, at DIV 21 [36] reported an increased level of synchronization and a dramatic change in burst properties (compared to those at DIV 14). In contrast, the present study revealed no such increase in synchronization at DIV 21, yet burst properties were highly variable - as reflected by a highly variable number of transient links (Figures 2,8), and there was a highly variable burst propagation time (Figure 9). Results herein suggest a network with an uneven balance between highly and poorly interconnected areas, whereby bursts initiated from different sites (as reported in [59]) propagate at different rates, with little link activation regularity (as reflected by the low link persistence at this age). We posit that the highly variable burst properties reported herein and in [24], [36] point to itinerant rather than persistent synchronization at DIV 21. Such transient synchronization effects may be averaged out by requiring multiple occurrences of correlated activity over long time-scales [58]. Therefore, our persistent networks at this age may not reflect the increased synchronization found in [36] (where links required only a period of correlated activity during the entire recording).
Crucially, the combination of varied burst properties and transient synchronization at DIV 21 indicates a mixture of regular and irregular activity. Modeling studies have suggested that such mixed activity constitutes optimal conditions for the emergence of a small-world topology via Hebbian learning rules and activity driven plasticity [60]. Thus, a change in the culture's spontaneous activity patterns could drive the topology transformation. Results herein and in [61] suggest that once the topology of the network has emerged, equilibrium states may exist at different time scales - from transient synchronization between subgroups of neural units at the short time-scale to regular occurrence of such transiently activated subgroups over longer time-scales. Modeling studies may provide further insight into the role of synchronization and the evolution of such equilibrium states [62], whilst pharmacological manipulation of specific neuron sub-types could verify biological mechanisms behind activity modulation.
Networks at DIVs 28 and 35 were classified as small-world, exhibiting several highly connected areas (clusters of highly inter-connected neural units), alongside the ability for any two areas to interact via few intermediary connections (short mean path length). Interestingly, when the network properties at DIVs 28 and 35 were compared, smaller differences were found than between earlier ages, suggesting a state of maturity [24], [32], [36], [63]. The non-trivial network structure demonstrated at DIVs 28 and 35 corresponds well with previous work [39], which concluded that mature cultures had complex network properties similar to those found in vivo.
Small-world networks have an architecture which supports efficient information transfer [8], [64]. Accordingly, our results showed a developmental reduction in burst propagation time that accompanied the emergence of cultures' small-world properties (Figure 9). Furthermore, variability of burst propagation time was lower at DIVs 28 and 35 than at younger ages. Since burst events are typically initiated from a number of sites [59], this reduced variability suggests that burst propagation times in mature cultures are not influenced by burst source; information propagates efficiently from all parts of the network. Interestingly, a small-proportion of links at DIVs 28 and 35 were activated extremely frequently (Figure 10), suggesting that they facilitate many of the interactions; it is possible that they represent activation of the small-world ‘short cuts’ between clusters.
The increasing prevalence of highly connected nodes in older cultures suggests that such hubs play a greater role in network activity as the cultures mature, perhaps indicating sources [65], sinks, or bridges [18], [33] for network activity. Interestingly, structural and functional hubs have recently been identified in the developing hippocampus where GABAergic interneuron hubs were found to orchestrate network synchrony [52], firing immediately prior to network bursts. Similarities between connectivity of GABAergic interneurons in the hippocampus and neocortex [66] and suggestions that cortical cultures develop subsystems akin to those found in vivo [25], [53], [56], imply that similar functional hubs may be present in the primary cortical cultures employed herein.
The present study has demonstrated that networks derived from the spontaneous activity of cultures develop non-random properties despite a lack of external input. Based on these results, we draw four main conclusions. Firstly, to mitigate fluctuations in spontaneous activity, multiple network bursts should be assessed to obtain the persistent network. Secondly, the functional network of a cortical culture evolves from an initial random topology to a small-world topology; we propose this is due to a change in the culture's spontaneous activity patterns that is driven by the maturing excitatory-inhibitory balance and an increase in network-wide synchronization. Thirdly, the reduction in burst propagation time with culture maturation that accompanies the evolution of a small-world topology supports the efficient network-wide flow of information afforded by a small-world network. Lastly, the presence of hubs and increasing contribution of links with high persistence suggests a proportion of highly influential nodes and links.
To the authors' knowledge, this is the first demonstration of small-world properties evolving in the functional networks of cortical neurons grown in vitro. This further supports work suggesting maturation of in vitro networks around the age of DIV 28 to 35; importantly, our results indicate that experiments which require complex network features should be undertaken from DIV 28 onwards, whilst those aiming to shape network maturation should be undertaken before DIV 28. Moreover, the work herein further supports the use of complex network statistics to quantify network level changes resulting from different experimental conditions, and importantly it provides a benchmark against which to assess the influence of closed loop stimulation on shaping cultures network properties - a fundamental question for the work on closed loop culture embodiment.
An important area for future work is to investigate the role of frequently activated nodes (hubs) in cultured neurons; including whether the presence of network-synchrony controlling hubs in the underlying substrate could mediate the timing and extent of functional interactions between otherwise segregated clusters, perhaps coordinating synchronous network-wide bursting. Additionally, the use of staining to identify the location and proportion of the different neuron types and sub-types, and the use of pharmacological manipulation to verify their effect on activity may help elucidate mechanisms behind the different network properties.
Data used for the present study was collected for [44], from cultures of pre-natal (E18) rat dissociated cortical neurons and glia cells, seeded onto multi-electrode arrays (MEAs, Multi Channel Systems, Reutlingen, Germany). Cultures were maintained in Teflon sealed MEAs in an incubator at 5% CO2, 9% O2, 35°C and 65% relative humidity [44]. For the present study, ‘dense’ cultures (estimated cell density of 2,500±1,500 per mm2) were used.
Culture's electrical activity was recorded daily during their first 5 weeks of development. For the present study, a sample population was selected from the large number of cultures recorded, specifically, 10 cultures from 4 preparations (plating batches). Cultures were arbitrarily selected from those that had recordings every 7 DIV, i.e. those which survived for the full 5 weeks and for whom none of the weekly recordings were missed. The use of multiple preparations is important as bursting patterns across preparations vary considerably [44]. Additionally, since the variation in burst properties measured at the same age (DIV) from different cultures (of the same plating), can exceed day-to-day differences in their properties (and inter-plating differences are significantly larger) [44], network properties were compared at weekly intervals. This also allows easy comparison with results from other studies [32], [36].
Data were recorded from cultures for 30 minutes daily in the incubator used for culture's maintenance. Unit and multi-unit spontaneous spike firing was recorded from the MEA (8×8 array of 59 planar electrodes, each 30 µm diameter with 200 µm inter-electrode spacing [centre to centre]). The pre-amplifier was from Multi Channel Systems (MCS), excess heat was removed using a custom Peltier-cooled platform. Data acquisition and online spike detection was performed using MEABench [67]. According to the MEA user manual (MCS) spike detection is reliable up to ∼100 µm from the electrode centre, beyond which spikes become indistinguishable from the background noise. Therefore, each MEA provides a grid of 59 non-overlapping 100 µm recording horizons (once the four analogue channels and single ground electrode are removed). It should be noted that data recorded on each channel may be from multi-neuron activity, no attempt was made to spike sort the data as overlapping waveforms found during a burst can present problems [44]. Lastly, as recording began immediately after the cultures were transferred to the pre-amplifier, the first 10 minutes were discarded from the analysis in order to mitigate any movement induced changes in culture activity [44], [68].
Spikes were detected online (using MEABench), positive or negative excursions beyond a threshold of 4.5× estimated RMS noise, were classed as spikes. Their peak amplitude timestamp (µs), plus electrode number were stored. For the present study, all positive amplitude spikes were removed to avoid counting spikes on both upwards and downwards phases.
In cortical cultures, global bursts (population bursts), characterized by an increase in culture activity across the entire MEA, are typically present from DIV 4–6 onwards [44], but sometimes as late as DIV 14 onwards [36]. Such bursts provide a time window during which many culture interactions take place and thus a useful opportunity to assess network-wide connectivity. For the present study, global bursts were identified as an increase in the number of spikes detected per unit time, summed over all electrodes in the array: specifically ≥4 spikes per channel in 100 ms, on ≥4 channels within 250 ms; based on the SIMMUX algorithm, included as Matlab (The MathWorks, Natick, MA, USA) code with MEABench. Burst start was determined by the timestamp of the first spike included in the global burst, and burst end taken as the timestamp of the last spike included. To assess interactions between neural units underlying all the electrodes, global bursts in which at least 25% (15/59) electrodes registered channel bursts (≥4 spikes in 100 ms) were selected. These were termed ‘network-wide’ bursts and ensured that many neural units participated in the burst (increasing the probability that the resultant networks would have sufficient numbers of nodes for the analysis of network properties). Additionally, since there were typically 10 to 150 such bursts in the 20 minute recording segment used, it provided a good balance between having sufficient numbers of bursts for analysis, whilst avoiding the inclusion of ‘tiny’ bursts [44] since these may have biased results.
All activity occurring from the first spike in the nw-burst to the last spike in the nw-burst (including tonic activity from electrodes not included in the nw-burst) was used for assessing the relationships between channel pairs. Spike occurrences were counted in 1 ms bins, this allowed a certain amount of jitter in the spike arrival times (which could otherwise decrease the likelihood of identifying correlated activity). Bin size was selected based on experimentation with 1, 5 and 10 ms bins. The 1 ms bins provided a greater separation between correlated and un-correlated channels, data not shown.
Functional connectivity was assessed by correlating spike times recorded on pairs of electrodes during a network-wide burst (as per [34]). This linear link analysis method assesses the probability of a spike at time t on one electrode being accompanied by a spike arriving at t±k on another electrode, where k is the allowable lag time. Spike times arriving within ±13 ms of each other were considered to be related (under the assumption that a linear relationship between spike arrival times on pairs of electrodes indicates their coupling). The maximum lag time was based on speed of axonal propagation, time for synaptic transmission and the maximum distance between 2 points on the MEA. Since the firing rates recorded on each channel may be different, cross-covariance was used, this correlates deviations in firing rates from their respective means as a function of lag.
Channels that had fewer than 8 spikes recorded during the burst were excluded from the cross-covariance analysis, as results from synthetic data testing showed that performing cross-covariance on vectors with fewer than 8 spikes was poor at distinguishing related vectors from independent ones (data not shown).
The cross-covariance function calculates the covariance of two random vectors:(1)In the case where X and Y are time-series the cross-covariance may depend on the time when it is estimated and on the lag between the time series:(2)For wide-sense stationary time series, covariance is a function of the lag only:(3)Cross covariance was calculated using the built in Matlab function xcov; specifically, each pair of channels with at least 1 ms overlap in their activity were compared from the time of the first spike on either channel to the time of the last spike on either channel. The tightness of the correlation window (X or Y channel recording spikes), and requirement for overlapping activity was to mitigate the effects of long periods of quiescence and to ensure that the data were as wide-sense stationary as possible.
Calculation of the cross-covariance at each lag resulted in a cross-covariance plot for each channel pair. The maximum cross-covariance value (peak of the plot) was used to determine whether a link between nodes was present by comparing it to a threshold as detailed next.
Table S1 provides the mathematical definitions for the topological properties and complex network statistics. Basic topological properties (related to network size), and complex network statistics, were calculated from the adjacency matrices (using Matlab, with additional scripts from the Brain Connectivity Toolbox [14]). For each transient network, only basic topological properties were measured, complex network statistics were not calculated due to the highly variable network size and edge density (see verification of network size and edge density). Instead, the mean numbers of nodes and links were calculated over all transient networks in the recording. Additionally, the coefficient of variation for number nodes and for number of links was calculated over all transient networks in the recording. The expected numbers of nodes, and links and the expected coefficients of variation were calculated over all 10 cultures.
In addition to the networks' topological properties, the spatial and temporal features of the networks were also assessed; link distance was calculated as the Euclidean distance between the electrodes on the MEA, based on 200 µm centre-to-centre spacing of the electrodes. For the present study, connections between nodes up to 566 µm (2 electrodes) apart were considered as ‘nearby’ and those greater than 566 µm as ‘distant’. Link persistence was calculated using the weighted persistent network adjacency matrix (i.e. prior to thresholding), normalized so that the persistence value was the percentage of transient networks in which the link was found.
For both link length (derived from the distance between connected nodes) and link persistence, histograms were obtained over all links from all cultures at each age. Thus, for link length, a count of the number of links in each bin (bin size = 1 electrode spacing) was calculated for each network, this was normalized to the total number of links in the network. For link persistence, a count of the links at each persistence level (bin size 5%) was calculated for each network. In both cases, median bin values were obtained over all 10 cultures, therefore the histogram proportions may not always sum to 1.
To quantify the changes in link length and persistence, two further measures were assessed: for link length, the proportion of links between spatially nearby vs distant nodes was calculated for each culture, and the median of these values was used to compare results between ages; for link persistence, the contribution of persistent links was measured as the number of links in each 5% persistence category multiplied by the category persistence value (e.g. if 20% of links were found in the 10% persistence category, the contribution was 200). The link contribution counts were further binned into transient (<25%) and persistent (≥25%).
The efficiency of information broadcast was measured as burst propagation time (time to recruit all channels in a network-wide burst). This was calculated in milliseconds from the time of the first spike in the burst, until the time at which all channels participating in the burst had been recruited. Channels could be recruited to the burst whilst the burst was in progress (i.e. sufficient channels displayed the required activity) but once the number of channels bursting dropped below the threshold, channels could no longer be recruited. For each channel included in the burst, recruitment time was the timestamp of the first spike in the burst activity sequence. Burst propagation times were calculated for all bursts of a culture at each age and the median of these was calculated for each age. Outliers (values <5th and >95th percentile) were removed from the data.
All statistics were obtained using SPSS version 17.0 (SPSS Inc., Chicago, USA). Unless otherwise specified P<0.05 was set as the significance level. Statistical tests for each network property were selected based on the experiment design and form of the resultant data; Checks were performed to ensure that the assumptions of each test were met. Following test selection, statistical power was verified at the 80% level (checking that the proposed test statistic had sufficient power to detect a genuine effect [71] [typically set to a difference of 1–2 times standard deviation of the mean], given the n numbers and variability of the data). For the present study, where some of the tests were applied to data with relatively low n numbers it was important to ensure that the power of each test was sufficient [72]. It was also important to ensure that the assumptions of the statistical tests were not violated (Text S1 describes the selection and validation of statistical tests used in the present study). The selected tests were as follows:
To check for a significant increasing or decreasing linear trend of the network properties as a function of the culture age, results for each network property were compared using a one-way ANOVA. Culture age (DIV) was the factor, and the network property was the dependent variable. The following properties were assessed in this manner: number of nodes, number of links, edge density, normalized mean path length, normalized clustering coefficient, small-worldness, goodness of fit ratio. In cases where a significant trend was found, Bonferroni and Tukey post-hoc tests were performed to check for significant differences between each pair of conditions, where found, the homogeneous subsets are mentioned in the results. Homogeneity of variances was tested using the Levene test.
Normality was tested using the Shapiro-Wilk normality test. In cases where the sample means were not normally distributed, non-parametric tests were used. For the burst propagation times a Kruskal-Wallis test was performed on the median burst propagation times for each culture at each age, with culture age as the grouping factor and median burst propagation time as the dependent variable. For the proportion of links to nearby vs distant nodes at each culture age, a 2-tailed Wilcoxon signed rank sum test was used. To compare the contribution of persistent links at each age, Friedman's rank test was used. Lastly, for the skewness of the link length distributions, a z-test was calculated based on the skewness estimate taken over the standard error of the skewness estimate. The P value was then calculated using the online statistics analysis tool (http://www.quantitativeskills.com/sisa/calculations/signhlp.htm, accessed November, 2010).
|
10.1371/journal.pmed.1002448 | The US President's Malaria Initiative, Plasmodium falciparum transmission and mortality: A modelling study | Although significant progress has been made in reducing malaria transmission globally in recent years, a large number of people remain at risk and hence the gains made are fragile. Funding lags well behind amounts needed to protect all those at risk and ongoing contributions from major donors, such as the President’s Malaria Initiative (PMI), are vital to maintain progress and pursue further reductions in burden. We use a mathematical modelling approach to estimate the impact of PMI investments to date in reducing malaria burden and to explore the potential negative impact on malaria burden should a proposed 44% reduction in PMI funding occur.
We combined an established mathematical model of Plasmodium falciparum transmission dynamics with epidemiological, intervention, and PMI-financing data to estimate the contribution PMI has made to malaria control via funding for long-lasting insecticide treated nets (LLINs), indoor residual spraying (IRS), and artemisinin combination therapies (ACTs). We estimate that PMI has prevented 185 million (95% CrI: 138 million, 230 million) malaria cases and saved 940,049 (95% CrI: 545,228, 1.4 million) lives since 2005. If funding is maintained, PMI-funded interventions are estimated to avert a further 162 million (95% CrI: 116 million, 194 million) cases, saving a further 692,589 (95% CrI: 392,694, 955,653) lives between 2017 and 2020. With an estimate of US$94 (95% CrI: US$51, US$166) per Disability Adjusted Life Year (DALY) averted, PMI-funded interventions are highly cost-effective. We also demonstrate the further impact of this investment by reducing caseloads on health systems. If a 44% reduction in PMI funding were to occur, we predict that this loss of direct aid could result in an additional 67 million (95% CrI: 49 million, 82 million) cases and 290,649 (95% CrI: 167,208, 395,263) deaths between 2017 and 2020. We have not modelled indirect impacts of PMI funding (such as health systems strengthening) in this analysis.
Our model estimates that PMI has played a significant role in reducing malaria cases and deaths since its inception. Reductions in funding to PMI could lead to large increases in the number of malaria cases and deaths, damaging global goals of malaria control and elimination.
| The United States contributes a significant proportion of the global budget for malaria control in the form of foreign aid through the President’s Malaria Initiative (PMI).
Due to proposed cuts to US foreign aid and PMI funding, it is important to demonstrate the impact and cost-effectiveness of PMI.
We used an established malaria transmission model to investigate the impact of PMI funding for malaria control.
We estimated the past impact of PMI funding on malaria-related cases and deaths and the potential future impact if PMI funding were to be cut.
PMI funding is highly cost-effective, averting an estimated 185 million cases and saving 940,049 lives since it was set up in 2005.
A reduction in funding of 44% would lead to an additional 67 million cases and 290,649 deaths over the next 4 years.
Ongoing support from PMI is critical to maintain recent advances in malaria control and progress towards malaria elimination goals.
PMI has proven to be a highly cost-effective means by which US foreign aid can be invested to reduce malaria burden.
| Unprecedented effort has seen the global burden of malaria halve since the turn of the 21st century due to the widespread distribution of highly effective preventative interventions such as long-lasting insecticide treated nets (LLINs) and indoor residual spraying (IRS) and the provision of highly efficacious treatment with artemisinin combination therapies (ACTs) [1]. However, funding for malaria control has plateaued, falling well behind what is necessary to expand protection to all those in need [2,3]. The continued high level of support for foreign aid contributions in a fluid global political landscape is not guaranteed and gains in malaria control can be fragile if intervention coverage, which is largely dependent on donor funding, is not maintained [4].
The US is the world’s largest donor of foreign aid for malaria control [5] and therefore a mainstay in global malaria efforts. The President’s Malaria Initiative (PMI), established in 2005 and funded by the United States Agency for International Development (USAID), has been particularly influential in investing in malaria control over the past 12 years [6]. PMI provides support to malaria control programmes in 19 African focus countries and the Greater Mekong Subregion (GMS) and is the largest bilateral funder of malaria prevention and treatment [5,7]. In the 12 years since its inception, PMI has procured 197 million LLINs and 378 million courses of ACTs, provided over 215 million person-years of protection with IRS, and distributed 35.7 million courses of preventative therapy for pregnant women [6]. In 2015, PMI funding represented over one-fifth of the global malaria budget envelope [5,6]. In a recent statistical analysis, the influence of PMI funding has been estimated to have had significant impact on under-5 mortality in sub-Saharan Africa, with an estimated reduction of 16% [8]. The US’s commitment to overseas aid has been threatened in recent months [9], highlighting the fragility of global funding for malaria control and a reliance on global political stability. In May 2017, Congress published the Congressional Budget Justification [9], which outlined a commitment to malaria control for 2018 of US$424 million. This is equivalent to a 44% reduction relative to commitments reported for 2017 [10].
To quantify the importance of the PMI contribution to global malaria efforts, we combined data on PMI commodity contributions over time and by country [6] with a mathematical model of the impact of interventions on malaria transmission, morbidity, and mortality parameterised at the subnational level [11] and previously used to inform the Global Technical Strategy (GTS) for malaria [5]. We used this to estimate the global health impact of past PMI funding and the potential implications that a reduction in funding from a key stakeholder and donor in the near term could have on malaria globally.
We linked data on PMI financing, historical intervention coverage, and the underlying epidemiology in modelled countries with estimates of the potential effect of reduction in PMI funding on the coverage of interventions nationally. These estimates were then used as inputs for an established transmission model of P. falciparum malaria [11,12] to project the impact of reductions in funding on cases, deaths, and Disability Adjusted Life Years (DALYs) (Fig 1).
We used an established individual-based malaria transmission model that incorporates a full dynamic mosquito-vector element to allow vector-control interventions to be accurately represented [13]. We briefly describe the model structure below. Full mathematical details can be found in S1 Appendix, Text A-I, and associated references [11,12].
Modelled humans are initially susceptible and may become infected, with a given probability, via the bite of an infectious mosquito. Upon infection, following a period reflecting liver-stage infection, an individual may become symptomatic and seek treatment. Successfully treated individuals benefit from a period of drug-dependent prophylaxis before returning to the susceptible compartment. Symptomatic individuals who do not receive treatment experience a period of symptomatic disease (which has high onward infectivity) before recovering to an asymptomatic state. These individuals, along with those who experience asymptomatic infection, move from being patently asymptomatic to subpatent before natural clearance moves them back into the susceptible compartment. Superinfection can occur from all asymptomatic and subpatent states. Those who experience clinical disease are considered at risk from severe disease and its associated mortality [14].
Naturally acquired immunity is incorporated at several stages of the infection process [12]. Clinical immunity is developed earliest, protecting individuals against severe disease and then clinical disease, and is exposure driven with an age-dependent component to the severe disease pathology and associated mortality rate. Antiparasite immunity develops later, driven by both age and exposure to infection, and reduces the detectability of infections through the control of parasite density. A degree of anti-infection immunity develops later in life, reducing the probability that an infectious bite results in patent infection. The parameters determining the acquisition of immunity were estimated through fitting to severe disease incidence, clinical incidence, and parasite prevalence data stratified by age across a range of transmission settings [12,15].
All infection states are assumed to be onwardly infectious to mosquitoes, with infectivity correlated with parasite density (i.e., highest for clinical disease, intermediate for patent asymptomatic infection, and lowest for subpatent infection), with the parameters estimated by fitting to mosquito feeding studies [12,14,15].
Vectors are modelled as a stochastic compartmental formulation incorporating the larval stages of infection and adult female infection stages [10, 13].
We modelled each first administrative unit (first administrative level below national) in all countries with stable malaria transmission, totalling 1,020 administrative units. Prior scale-up of interventions (LLINs and IRS) was estimated from World Malaria Report data [16], which are based on reports from National Malaria Control Programmes (NMCPs). Demographic and Health Survey (DHS) and Multiple Indicator Cluster Surveys (MICS) for within Africa [17] and World Malaria Report [18] estimates for elsewhere were used to estimate treatment coverage. It was assumed there was no prior scale-up of seasonal malaria chemoprevention (SMC). Each administrative unit was assigned a seasonal pattern that determined the seasonal fluctuation in the carrying capacity of the environment. Seasonality was estimated using Fourier transformations of daily rainfall data from 2002–2009 from Garske et al. (2013) [19]. The carrying capacity was then fitted to 2015 estimates of prevalence (within Africa) [1] or cases (outside of Africa) [16,20] using a root-finding algorithm. Data on populations were compiled from the Gridded Population of the World dataset, adjusted for United Nations estimates of country-level populations [21]. Estimates of the spatial limits of P. falciparum transmission [20] were used to delimit populations at risk.
To estimate the impact of PMI funding, we firstly estimate the proportion of intervention coverage that is attributable to PMI funding in each location. This is then subtracted from the total intervention coverage estimated. The number of LLINs procured and distributed, the number of people protected by IRS, and the number of ACTs procured and distributed stratified by year and country were all obtained from PMI’s 10th Annual Report to Congress [6]. Absolute numbers were converted to coverage using the appropriate denominators: the estimated population at risk for LLINs and IRS and estimates of the total number of ACT treatment courses delivered [5] for ACT in each country. Examples of this process are detailed in Box 1 (and S1 Table). Throughout, we assumed that 1 LLIN covered 1.8 people (in line with WHO methodology [5]). To estimate the relationships between net delivery, coverage, and usage, we follow an approach by Bhatt et al. (2015) relating distribution data (i.e., procurement as reported by PMI) to household ownership and usage, accounting for household size [22]. The coverage estimates in the model relate to usage and also incorporate wear and tear and decay of insecticide over time. We make an optimistic assumption that ACTs delivered are efficiently used (i.e., reach the health clinics and are effectively employed to treat malaria). In Senegal and Mali, where PMI funds support SMC, we assumed that SMC coverage attributable to PMI was 20%, supporting and complementing SMC implementation by NMCPs and other nongovernmental organisations (NGOs) in these countries [6]. These estimates are then used to simulate malaria trajectories, both retrospectively and prospectively, assuming varying levels of PMI funding.
We considered 3 budget scenarios, one in which PMI funding was kept constant to 2017 levels, one in which 100% of the PMI budget was removed, and a third in which the budget was reduced by 44% (applied uniformly across PMI-supported countries) to reflect the difference in budget attributed to malaria control detailed in the 2017 financial omnibus [10] and the proposed budget for 2018 onwards [9]. The relationship between PMI’s budget and intervention coverage was assumed to be linear, whereby an assumed budget cut of 44% was associated with a proportional decrease in the PMI-attributable intervention coverage. We also ran a scenario with a less drastic reduction in funding of 20%. We assume no mitigation through alternative funding routes or reallocation of reduced budgets. Extra savings and benefits to the health system of PMI funding were also estimated. The savings to the health system of cases averted due to PMI-funded interventions were calculated as the costs of case management and drug commodity costs of the cases averted. In addition, we calculated the additional deaths that may occur if PMI-funded interventions were removed and a national health system did not have the capacity to absorb and adequately treat the additional severe cases.
All scenarios were run multiple times in a sensitivity analysis using 20 separate sets of parameters drawn from the posterior of the modelling fitting [15]. Associated outputs are presented as the median and 95% credible intervals.
To date, PMI has allocated over US$5 billion to 19 PMI focus countries in sub-Saharan Africa as well as the GMS [23] (Fig 2). We attribute increases in coverage of 8.13% for LLINs, 4.18% for IRS, and 12.9% for ACTs to PMI funding in supported countries in 2015. We estimate that in the 12 years since its inception, PMI has prevented 185 million malaria cases (95% CrI: 138 million, 230 million) (Fig 3A) and saved 940,049 lives (95% CrI: 545,228, 1.4 million) (Fig 3B), the majority of which (77%, 95% CrI: 75%, 81%) would have occurred in children under the age of 5. In sub-Saharan Africa, we estimate that PMI investment has led to an 11.6% (95% CrI: 9.5%, 13.0%) reduction in incidence and an 18.3% (95% CrI: 16.3%, 20.4%) reduction in under-5 malaria-mortality rates in 2015. We estimate the biggest impact in terms of absolute cases averted to have occurred in long-term supported countries with the highest burden. For example, Nigeria, the country with the highest burden globally [5], has received approximately US$345 million from PMI since 2010 [6], leading to an estimated 13.8 million cases (95% CrI: 8.7 million, 17.0 million) averted and 128,861 lives (95% CrI: 75,852, 200,075) saved. Angola has benefitted from continuous support since 2005, seeing investments totalling US$248 million dollars [6], leading to an estimated 8.7 million cases (95% CrI: 6.3 million, 10.4 million) averted and 43,752 lives (95% CrI: 24,946, 61,433) saved.
We estimate that a 44% cut in PMI funding would lead to an additional 67 million cases (95% CrI: 49 million, 82 million) (Fig 3B) and 290,649 deaths (95% CrI: 167,208, 395,263) (Fig 3C; S1 Appendix, Table F) from malaria compared to maintaining current levels of funding from 2017 to 2020. A 20% reduction in funding was associated with an additional 31 million cases (95% CrI: 21 million, 38 million) and 127,799 deaths (95% CrI: 73,313, 178,234) over the same period. If PMI-funded coverage of interventions can be maintained over the next 4 years, PMI support will be responsible for averting an estimated total of 162 million cases (95% CrI: 116 million, 194 million) (Fig 3B) and 692,589 deaths (95% CrI: 392,694, 955,653) (Fig 3C) in the 4-year period from 2017 to 2020, compared to no PMI support.
The impact on malaria burden will be focussed in high-burden countries receiving significant financial support in sub-Saharan Africa. Additionally, with ongoing concern surrounding the emergence and spread of ACT drug resistance [24], support for the GMS is also contributing to the malaria elimination goals in that region.
We estimate that PMI support would avert an additional US$174 million dollars (95% CrI: 121 million, 224 million) of national health system spending through averted malaria cases from 2017 to 2020 (Fig 4A). In the absence of PMI funding, a failure of health systems to absorb the extra caseload (through lack of capacity, finances, or both) would lead to an estimated 69,314 extra deaths (95% CrI: 39,102, 94,888) over this period (Fig 4B), in addition to the 692,589 deaths estimated to be directly caused by reductions in intervention coverage. These impact estimates are likely conservative, not accounting for the indirect impacts of increased transmission.
Over the period 2013–2015, when the PMI programme was fully scaled to current levels, PMI reported that spending in the 19 focus countries in sub-Saharan Africa was approximately US$1.7 billion. Translating the modelled epidemiological impact into system-wide cost-effectiveness, we estimate a cost of US$20.6 per malaria case averted (95% CrI: US$15.2, US$31.4), US$4,081 per death averted (95% CrI: US$2,084, US$7,435), and US$94 per DALY averted (95% CrI: US$51, US$166) (Table 1). This represents a range of 2%–57% as a proportion of per-capita gross domestic product (GDP) in these countries. Cost-effectiveness estimates are driven by the intervention mix and national-level differences in the cost of treating clinical and severe cases. Differences between cases and deaths averted are primarily driven by the intervention mix, especially the proportion of funding that went towards treatment (treatment contribution is positively associated with the proportion of deaths to cases averted, linear model coefficient = 0.012, p = 0.035).
Here, we have produced modelled estimates of the programme-wide effectiveness of PMI in terms of the impact it has had upon malaria morbidity and mortality since its inception in 2005. We estimate that PMI has averted 185 million cases and 940,049 deaths to date. If funding for PMI is maintained, we predict that a further 162 million cases and 692,589 deaths could be averted over the next 4 years, compared to no PMI funding. However, in comparison to continued full PMI support, a 44% cut in the PMI budget, as indicated in the May 2017 Congressional Budget Justification, could result in an additional 67 million cases and 290,649 deaths in the next 4 years.
Our results highlight the fragility of the gains in malaria control that have been made to date, particularly given the changing geopolitical landscape [25]. International funding, including that from governments, such as from PMI, the United Kingdom’s Department for International Development (DFID), the Global Fund, and others, accounts for a large proportion (approximately 68% [5]) of the funds available for malaria control worldwide. Malaria control is therefore reliant on sustained long-term investment from foreign donors. Without continued commitment to support programmes, recent gains in the control of malaria will be difficult to sustain and potential rebound epidemics likely [4].
Prudent investment of foreign aid relies on being able to effectively implement cost-effective interventions to maximise health gains. PMI has proven to be a capable mediator of this process for malaria. The estimates of cost per DALY averted here are significantly below the WHO threshold for cost-effectiveness of less than 300% of a country’s per-capita GDP [26]. Even among highly cost-effective interventions, malaria control compares favourably as a means by which to improve global health [27]. Between-country variation in cost-effectiveness is pronounced. The effect is driven by the intervention mix and underlying epidemiological variation (such as the intrinsic transmission potential). Costs are driven by the intervention mix and, specifically, the impact of PMI support on treatment costs, which varies between countries. Whilst the past and current positive health impacts of PMI-funded interventions is very apparent, there remains much debate as to the impact that foreign aid has on recipient countries [28].
In addition to its direct impact on cases and malaria-attributable mortality, investment in malaria control brings about substantial further potential health gains by alleviating the burden that malaria places on health systems in affected countries [29]. Supporting vector control interventions is expected to decrease caseloads, freeing up health system capacity and reducing costs incurred from treating clinical and severe cases of malaria. A recent PMI-supported study demonstrated reductions in malaria-related inpatient and outpatient admissions and hospital costs after the scale-up of interventions in Southern Province in Zambia [30]. Funding cuts lead to increased caseloads due to the negative impacts of reduced intervention coverage, the stress of which will be borne by the national health systems of malaria-endemic countries. Lack of health-system capacity was a critical factor in the recent Ebola epidemic in West Africa [31], the impact of which reverberated globally. Those countries worst affected are highly malaria endemic and had health systems already dealing with the challenges of a high malaria burden [32,33]. A redistribution of emergency funds earmarked for the Ebola epidemic [9] could potentially help to mitigate budget cuts for malaria control. However, this is a finite fund that would only serve as a very near-term solution to budget reductions.
Our results provide a conservative estimate of the overall impact of PMI funding, as we do not capture the impact of all PMI-associated activities, notably intermittent preventive treatment in pregnancy (IPTp), which we have not modelled but is one of the most cost-effective malaria interventions [34,35]. PMI presence in a country further catalyses and facilitates the procurement, distribution, and implementation of interventions from other funders with the initiative distributing 80 million LLINs and 34 million ACT courses procured by other donors in the period 2006–2015 [6]. Furthermore, PMI is involved with a number of capacity and health system-strengthening initiatives, such as training health workers in malaria diagnosis and treatment [6], the loss of which would compound issues of increased caseload if PMI support were reduced. Our estimates of reductions in under-5 mortality attributable to PMI funding are lower when compared with estimated reductions of a similar magnitude in all-cause mortality in a recently published difference-in-differences analysis of PMI impact [7]. Whilst our estimates of intervention coverage attributable to PMI funding are similar, the additional impact estimated by Jakubowski et al. may be ascribed to indirect impacts of PMI funding on nonmalaria outcomes (through, for example, health systems strengthening), although considerable uncertainties also impact both analyses. We also do not capture the wider societal costs of the disease, such as missed workdays by carers, reduced education, or impact on future lifetime earnings, nor the economic effects of endemic malaria on factors such as migration, trade, tourism, or foreign investment within a country [36]. It is likely that, when facing cuts, PMI and NMCPs would reallocate existing funds to cover those interventions seen as vital. However, in an already budget-restricted environment, a limit to the potentially mitigating effects of such reallocations would quickly be reached. There are a number of difficulties associated with estimating accurate coverage estimates and uptake for interventions with a wide range of definitions and methodologies adopted. We have assumed that PMI-reported contribution and interventions figures, taken from their 10th Annual Report to Congress [6] and building upon a well-established monitoring and evaluation strategy, are representative and accurate. We also are including assumptions that the PMI-delivered interventions are reaching required recipients in an efficient manner. Whilst we know inefficiencies do exist, for example in LLIN distribution [22], these are difficult to attribute to specific sources. Furthermore, due to the nonlinear impact of interventions such as LLINs, it is difficult to split contributions from different funding sources (i.e., should an X% funding contribution be linked to the first N% or last N% of observed LLIN coverage?). We do account for falloff between coverage and usage as well as deterioration of insecticide and wear and tear of LLINs in this analysis. Similarities to empirical estimates [8] indicate that we are accurately capturing broad trends in intervention coverage due to PMI funding.
As malaria transmission is brought to low levels, increased efforts are needed to target hard-to-reach populations as well as increase surveillance efforts, and hence the programmatic costs are likely to increase [7]. In such circumstances, investment decisions need to take into account the potential for permanent gains that would be accrued if an area or country can achieve elimination. However, there still remain large, extremely cost-effective gains that can be obtained by investing further to reduce the burden of malaria in areas of high endemicity. WHO GTS for malaria has set targets of achieving of at least 90% reductions in global case incidence and mortality rates by 2030 compared to levels in 2015, with vector control, chemoprevention, diagnosis and treatment, and surveillance being key pillars of the outlined strategy [37]. Based on the estimates of our model, PMI’s ongoing support of these activities in countries of high burden or strategic importance is vital in order to avoid a rapid erosion of the progress made in the last 15 years on the road towards malaria eradication.
|
10.1371/journal.pcbi.1002182 | Barriers to Diffusion in Dendrites and Estimation of Calcium Spread Following Synaptic Inputs | The motion of ions, molecules or proteins in dendrites is restricted by cytoplasmic obstacles such as organelles, microtubules and actin network. To account for molecular crowding, we study the effect of diffusion barriers on local calcium spread in a dendrite. We first present a model based on a dimension reduction approach to approximate a three dimensional diffusion in a cylindrical dendrite by a one-dimensional effective diffusion process. By comparing uncaging experiments of an inert dye in a spiny dendrite and in a thin glass tube, we quantify the change in diffusion constants due to molecular crowding as Dcyto/Dwater = 1/20. We validate our approach by reconstructing the uncaging experiments using Brownian simulations in a realistic 3D model dendrite. Finally, we construct a reduced reaction-diffusion equation to model calcium spread in a dendrite under the presence of additional buffers, pumps and synaptic input. We find that for moderate crowding, calcium dynamics is mainly regulated by the buffer concentration, but not by the cytoplasmic crowding, dendritic spines or synaptic inputs. Following high frequency stimulations, we predict that calcium spread in dendrites is limited to small microdomains of the order of a few microns (<5 μm).
| Diffusion is one of the main transport phenomena involved in signaling mechanisms of ions and molecules in living cells, such as neurons. As the cell cytoplasmic medium is highly heterogeneous and filled with many organelles, the motion of a diffusing particle is affected by many interactions with its environment. Interestingly, the functional consequences of these interactions cannot be directly quantified. Thus, in parallel with experimental methods, we have developed a computational approach to decipher the role of crowding from binding. We first study here the diffusion of a fluorescent marker in dendrites by a one-dimensional effective diffusion equation and obtained an effective diffusion constant that accounts for the presence heterogeneity in the medium. Furthermore, comparing our experimental data with simulations of diffusion in a crowded environment, we estimate the intracellular calcium spread in dendrites after injection of calcium transients. We confirm that calcium spread is mainly regulated by fixed buffer molecules, that bind temporarily to calcium, and less by the heterogeneous structure of the surrounding medium. Finally, we find that after synaptic inputs, calcium remains restricted to a domain of 2.5 µm to each side of the input location independent of the input frequency.
| Dendrites of neurons contain a complex intracellular organization made of organelles, such as mitochondria, endoplasmic reticulum, ribosomes and cytoskeletal network generated by actin and microtubules [1]–[3]. The cell cytoplasm is thus a crowded rather than diluted medium in which diffusional mobility of small molecules is restricted [3]–[7]. Molecular crowding can affect many biochemical processes such as, protein folding [8]–[10], enzymatic reactions [11]–[13] and signal transduction [14]. Although electromicroscopy images [15] reveal the complexity of dendritic organization, there are no direct methods to estimate the functional consequence on diffusion. Modeling in combination with Monte-Carlo methods [16]–[19] allowed to study diffusion in crowded media. Depending on the size of the diffusing molecule and the interactions with the heterogeneous media, crowding can lead to anomalous or normal diffusion [4], [20]–[24].
Neuronal calcium is an fundamental and ubiquitous messenger [25], [26]. It is regulated by cytoplasmic crowding, mobile and immobile calcium buffers [27]–[30], pumps and dendritic spines, which cannot be easily dissociated experimentally. It was already noticed and quantified [31] that cellular calcium buffers can determine amplitude and diffusional spread of neuronal calcium signaling. Precisely, fixed calcium buffers tend to retard the signal and to lower the apparent diffusion coefficient, whereas mobile buffers can contribute to calcium redistribution. To study calcium dynamics, we develop in the first part, a model of diffusion in a crowded three-dimensional dendrite, that we reduce to a one-dimensional effective diffusion process. The model is general and can be applied to protein diffusion in membranes or in endoplasmic reticulum-like networks [16], [32]. In a second part, we use uncaging experimental data of an inert dye (fluorescein) in a spiny dendrite and in a glass tube of similar size filled with aqueous solution to estimate the reduction of the diffusion constant in a dendrite. These experiments are repeated by Brownian simulations in a 3D model dendrite in order to validate our one-dimensional model.
In the last part, we use the previously derived effective diffusion constant and simulate a system of reaction-diffusion equations in one dimension to study calcium dynamics in a dendrite. We accounted for calcium buffers, pumps, dendritic spines and synaptic inputs. We show that for moderate organelle crowding, calcium spread is mainly restricted by the buffer and the pump concentration and not by obstacles or dendritic spines. Although crowding restricts dendritic diffusion by a factor 20, it is not responsible for the high calcium compartmentalization () in dendrites [33], [34]. We further show that following high frequency stimulations, calcium spread does not exceed . In summary, calcium microdomains are highly regulated by various active processes such as calcium buffers, pumps and stores.
Our results are divided into three sections. In the first section, we present the diffusion model for an inert dye in a crowded dendritic medium. The model is derived from a periodic compartmentalization of the dendritic domain. It is followed by an extension of the model to almost periodic compartments and the analysis of the mean time a particle takes to travel across the dendrite. In the second part, we present the outcome of the uncaging experiments of fluorescein to probe the dendritic medium and to estimate the model parameters. It is followed by a comparison to Brownian simulations, which repeat these experiments on a computer. Finally, we provide mean-field simulation results for calcium spread in a dendrite under the additional presence of stationary buffers, pumps and synaptic input.
In addition to cytoplasmic crowding, calcium dynamics is regulated by many factors such as binding to buffer molecules (e.g., calmodulin and calcineurin), dendritic spines and various types of pumps located on the dendritic surface (PMCA, NCX) and on the surface of internal organelles such as the endoplasmic reticulum (SERCA). It is usually not possible to dissect experimentally the contribution of each process, and we shall apply our previous result to study calcium spread in dendrites.
We present a reaction-diffusion equation (Materials and Methods) to simulate calcium dynamics in both spiny and aspiny dendrites. At this stage, we do not take into account the intracellular calcium stores, and thus, we exclude the generation of calcium waves through CICR, nor do we model spontaneous dendritic calcium spikes or calcium transients associated with back-propagating action-potentials. We shall focus here on the local spread of calcium transients and we ignore global calcium events. We include in our simulations the effect of buffers, pumps, spines and synaptic input. The contribution to calcium dynamics for each active component is provided in the Materials and Methods.
We first simulated calcium diffusion in an aqueous solution (contained in a glass pipette) by initiating a calcium transient and solving the one dimensional diffusion equation (41)–(45) with a diffusion constant of (Figure 4A). The effect of crowding alone on calcium diffusion in a dendrite was simulated by reducing the free diffusion constant to (Figure 4B). We assume here that the effects of crowding on motion are the same for fluorescein molecules and calcium ions attached to a dye molecules. As expected, crowding leads to a more localized and persistent calcium transient compared to free diffusion in an aqueous solution.
We next added two types of imobile buffers, calmodulin (CaM) and calcineurine (CN), as well as pumps (NCX and PCMA) to the simulation. The buffer concentration was varied between low ([CaM]: , [CN]: ), medium ([CaM]: , [CN]: ) and high ([CaM]: , [CN]: ) levels. Figure 4C and D show the effect of fast buffering on calcium dynamics in aqueous solution and in a crowded dendrite, respectively, for medium buffer concentration. The differences are small. The calcium signal in the crowded medium is more localized in space and slightly longer lasting than in aqueous solution. From these simulation results, we conclude that the spatiotemporal extent of the calcium signal is highly restricted by the stationary buffer activity. These results agree qualitatively with other uncaging experiments of calcium in glass tubes and dendrites [47].
We next analyze calcium spread originating from localized inputs such as synapses. At dendritic synapses calcium can enter through NMDA-receptors. To estimate calcium spread as a function of the synaptic input frequency, we simulated -influx in the middle of a dendritic segment (Figure 4E). Buffers and pumps were set to their default values (Table 1). We initiated calcium transients in the crowded model dendrite for different input frequencies (). The spatiotemporal extent of the calcium signal for different input frequencies is given in the intensity plots Figure 4F. Calcium spread is measured by the full width at half maximum (FWHM) of the calcium signal. Interestingly, for input frequencies larger than 20 Hz, the calcium signal in the dendrite reaches a stationary value. For high input frequencies (20 Hz) calcium spread does not exceed ( = 0.5FWHM) as measured from the input source. This is in agreement with the experimental data where calcium spread was contained within a domain of about . We conclude that buffer and pumps limit calcium spread to few micrometers.
We have shown here that dendritic crowding reduces the diffusion constant of inert Brownian molecules by a factor of 20 when compared to diffusion in an aqueous solution. We have used this result to estimate calcium spread in dendrites. We found that in the absence of regenerative mechanisms (VSCC, calcium stores), the spread of calcium largely depends on the buffer concentration and moderate molecular crowding does not play a significant role in shaping calcium dynamics. Thus, crowding has only a minor effect compared to the cumulative effect of pumps and buffers. In addition, the presence of a single (passive) spine at the location of calcium release did not influence calcium diffusion in the dendrite.
In this study, we have analyzed the effect of molecular crowding on calcium spread under the presence of stationary buffers. Assuming that the diffusion constant of calcium and fluorescein are reduced by the same factor due to the effect of molecular crowding, our results confirm previous studies that calcium spread is largely restricted by the effect of stationary buffers [31], [48]–[50]. Our analysis showed only a small effect of molecular crowding on calcium spread (Figure 4C and 4D): slightly more calcium molecules were bound to buffers in the crowded condition.
These results are qualitatively consistent with stochastic simulations in a cubic cell model under different crowding and buffer mobility conditions [19], where it has been shown that molecular crowding affects the calcium signaling system mainly through crowding-induced binding of calcium to buffer molecules and less through the direct hindrance of calcium diffusion. This study showed further that these effects are not additive. Interestingly, the reduction in diffusion constant due to molecular crowding was found to be for moderately crowded environments with excluded volume fraction. In our study, the reduction of the calcium diffusion constant was extrapolated from fluorescein uncaging experiments in the dendritic medium, which resulted in a much higher value. This difference might result from additional crowding effects such as cavities that were not modelled in the stochastic simulations.
Calcium microdomains have been observed during spontaneous and electrically evoked activation of synapses on dendritic shafts in aspiny neurons [34]. Compartmentalization into domains of about resulted from fast kinetics of calcium permeable AMPA receptors and fast local extrusion via the exchanger [34]. In general, as observed in Figure 4, calcium spread is robustly confined in a domain of less than from the input source and this seems to be independent of the synaptic firing frequency. Thus, calcium dynamics seems to be well regulated by buffers, stores and extrusion mechanisms.
It is certainly a requirement for dendrites to prevent calcium spread over large distances because it is not only the primary messenger in the induction of synaptic plasticity, such as long term potentiation (LTP) [51], but it is also involved in morphological changes and in the regulation of receptor trafficking such as AMPA [52]. While organelle localization might depend on the dendritic local needs (protein syntheses, energy supply and local calcium stores), calcium pump densities and calcium buffer concentrations might be regulated independently to maintain calcium homeostasis. It remains an unsolved question to determine how pumps and calcium buffer molecules are regulated along a dendrite.
Using our previous computations, we found that (passive) dendritic spines in this mean-field approach do not contribute much in dendritic calcium regulation (data not shown). In general, our result suggests that spines should not significantly affect the movement of diffusing particles along the dendrite. However, in the case of calcium, we have not taken into account a possible calcium propagation through the endoplasmic reticulum network, which may lead to a very different type of propagation.
Dendritic spines can be seen as the ultimate place of confinement in dendrites: indeed, calcium exchangers located on the endoplasmic reticulum surface or on the spine neck membrane can prevent calcium from diffusing into the spine head [53], [54]. In addition, large crowding observed at the spine base due to various types of organelles such as the endoplasmic reticulum or the spine apparatus [2], [15] can prevent diffusing molecules from entering the spine neck. However, it is not clear whether mRNA or transcription factors can enter dendritic spines by passive diffusion or whether active processes are required.
Cultures were prepared as detailed in [47]: we use wistar rat pups at P1. Hippocampal tissue was mechanically dissociated and plated on 12 mm glass coverslips at 3–4105 cells per well in a 24 well plate. Cells were left to grow in the incubator at C, 5% CO for 4 days, at which time the medium was changed to 10% HS in enriched MEM. The medium was changed four days later to 10% HS in enriched MEM. Cells were transfected at 1 wk in culture with DsRed plasmid to visualize the dendrites and spines using a lipofectamine 2000 (Invitrogen) method. On the day of imaging, the glass was transferred to the recording medium containing (in mM): NaCl 129, KCl 4, MgCl 1, CaCl 2, glucose 10, HEPES 10, and TTX . pH was adjusted to 7.4 with NaOH, and osmolarity to 315 mOsm with sucrose. Ten-fourteen day old cultured cells were patch clamped at the soma and recorded with a glass pipette containing (in mM): K-gluconate 140, NaCl 2, HEPES 10, EGTA 0.2, Na-GTP 0.3, Mg-ATP 2, phosphocreatine 10, and 100 of caged fluorescein (Molecular Probes) at pH 7.4 having a resistance of 6–12 M. Signals were amplified with Axopatch 200 (Axon Instruments Inc. Foster City, CA). Cells were imaged with a 63× water immersion objective (NA = 0.9). UV laser was aimed at a spot of 1 m in the center of the field of view. A line scan mode (0.7 msec/line) was used along an imaged dendrite to measure fast changes in fluorescence following flash photolysis of caged fluorescein. In the second stage of the experiment, the content of patch pipettes, containing caged fluorescein, was sucked out and introduced into additionally prepared pipettes with long and sharp tips, having tens of microns in length and about 1–2 in diameter making their geometry similar to a “typical” dendrite. Same line scan mode was used to compare changes in fluorescence in a dendrite and in a glass tube, containing similar concentrations of caged fluorescein. Data were analyzed using custom made MATLAB-based programs. Steps of 0.6 from the center of the uncaging sphere were defined through the line scans and pixels inside every step were horizontally averaged. Every line scan trial was repeated 7–14 times. Statistical comparisons were made with t-tests.
We implemented the Brownian simulations in MATLAB using a ray-tracing algorithm. To overcome the huge computational burden that Brownian simulations in complex domains impose, we made heavily use of MATLAB's object-oriented programming and vectorization features as well as of the external C/C++ interface functions capabilities (MEX-files). We first constructed a triangular mesh of the simulation domain (e.g., cylinder, cylinder with spine, see Figures 3 A,B) using a simple mesh generator based on distance function (DistMesh package, [55]). The (meshed) simulation domain was then equipped with user-defined sampling boxes, an initial distribution of particles and diffusion barriers (e.g., disks with small holes, see Figures 3 B,C). We predefined a sampling interval ( ms) at which the particle concentrations in the sampling boxes were measured.
Surface mesh elements were defined to be either reflective or absorbing. The top and the bottom of the cylindrical domain was set to be absorbing while all other surface elements were defined to be reflective. Particle rays crossing reflecting boundaries or obstacles were reflected according to the law of light reflection. To speed up the code we divided the simulation domain into partition voxels. For each partition voxel a list of contained objects (mesh elements, obstacles) was pre-computed and provided to the algorithm during execution.
The Brownian simulation was implemented using an Euler-scheme with adaptive-step size. Steps were defined by the distance to mesh elements and obstacles. The closer the particles were to objects the smaller the step size was chosen. As a rule of thumb, the minimal step size was determined by 0.3–0.5 of the smallest length scale that had to be resolved (e.g., the radius of the hole of the disks, see Figure 3 B). The (vectorized) particle rays were traced in the voxels and tested for intersections with mesh elements or objects. If intersections occurred the particles were either reflected or absorbed. It is important to note that an adaptive-step size algorithm leads for each particle to a different progress in physical time. Hence, the measurement of particle concentrations at fixed sampling times, required the implementation of a scheduler that removed particles temporarily from the simulation and stored their positions. Our simulations lasted between several hours to several days on a cluster depending on the number of particles, number of objects and the minimal step size. We have made extensive use of MATLAB's visualization tools to monitor the simulations and to generate visual outputs of the simulation results (see snapshots in Figures 3 A–C and a movie (Video S1) in the Text S1). We have included in the Text S1 a validation study of diffusion in a cylindrical domain with absorbing boundaries at the top and bottom. Different measures such as global and local particle concentrations as well as the mean first passage time to the absorbing boundaries are extracted from the simulations and compared with existing analytical results. The test-simulation is shown in Video S2. A good agreement between these results was obtained, and thus, evidence for the correctness of the implemented algorithm in the Monte-Carlo simulation tool is provided.
The spatiotemporal calcium signal in the dendrite is regulated by several active and passive components that are described next.
|
10.1371/journal.pcbi.1006229 | Dopamine transporter oligomerization involves the scaffold domain, but spares the bundle domain | The human dopamine transporter (hDAT) is located on presynaptic neurons, where it plays an essential role in limiting dopaminergic signaling by temporarily curtailing high neurotransmitter concentration through rapid re-uptake. Transport by hDAT is energized by transmembrane ionic gradients. Dysfunction of this transporter leads to disease states, such as Parkinson’s disease, bipolar disorder or depression. It has been shown that hDAT and other members of the monoamine transporter family exist in oligomeric forms at the plasma membrane. Several residues are known to be involved in oligomerization, but interaction interfaces, oligomer orientation and the quarternary arrangement in the plasma membrane remain poorly understood. Here we examine oligomeric forms of hDAT using a direct approach, by following dimerization of two randomly-oriented hDAT transporters in 512 independent simulations, each being 2 μs in length. We employed the DAFT (docking assay for transmembrane components) approach, which is an unbiased molecular dynamics simulation method to identify oligomers, their conformations and populations. The overall ensemble of a total of >1 ms simulation time revealed a limited number of symmetric and asymmetric dimers. The identified dimer interfaces include all residues known to be involved in dimerization. Importantly, we find that the surface of the bundle domain is largely excluded from engaging in dimeric interfaces. Such an interaction would typically lead to inhibition by stabilization of one conformation, while substrate transport relies on a large scale rotation between the inward-facing and the outward-facing state.
| The human dopamine transporter efficiently removes the neurotransmitter dopamine from the synaptic cleft. Alteration of dopamine transporter function is associated with several neurological diseases, including mood disorders or attention-deficit hyperactivity disorder, but is also a major player in addiction and drug abuse. Functional studies have revealed that not only is transporter oligomerization involved in surface expression and endocytosis, but, more importantly, in reverse transport (efflux) of dopamine that is triggered by amphetamine-like drugs of abuse. Structural knowledge of transporter oligomerization is largely missing. We performed a large scale comprehensive computational study on transporter oligomerization to reveal dimer geometries and the residues involved in the interfaces. The dimer conformations we find in our dataset are fully consistent with all available experimental data in the literature, but also show novel interfaces. We further verified all dimer geometries by free energy calculations. Our results identified an unpredicted—but for the mechanism of substrate transport essential—property: the bundle domain, which moves during the transport cycle, is excluded from contributing to dimer interfaces, thereby allowing for unrestrained movements necessary to translocate substrates through the membrane.
| The human dopamine transporter (hDAT) is a member of the monoamine transporter family [1], which also includes the transporters for serotonin (hSERT) and noradrenaline (hNET). Within the central nervous system, these transporters are localized to pre-synaptic neurons close to the synaptic junctions, but primarily outside the synapse [2]. Their role is to efficiently curtail neurotransmitter-mediated signaling via uptake of associated neurotransmitters into pre-synaptic neurons. These transporters utilize the sodium/chloride gradient across the plasma membrane to provide the driving force for transport [3]. Dysfunction in hDAT leads to severe neuronal disorders, such as Parkinson’s disease, attention-deficit hyperactivity disorder (ADHD), depression and schizophrenia [4,5]
The first crystal structure for this class of transporters became available for the bacterial homolog small amino acid transporter (LeuT) isolated from Aquifex aeolicus [6]. The Drosophila melanogaster dopamine transporter (dDAT) was the first eukaryotic member of the SLC6 family to be resolved by X-ray crystallography [7], followed more recently by crystallization of the human serotonin transporter (hSERT) [8]. LeuT was crystallized in three conformations [6,9,10], revealing that during the transport cycle the bundle domain (consisting of transmembrane helices TMH 1, 2, 6, 7) rotates by ~30° relative to the scaffold domain (TMH 3, 4, 8, 9.10, 12), thus anchoring the transporter to the membrane. TMH 5 and 11 are also assigned to the scaffold domain, though they bend during the conformational transition of the transport cycle. The similarity of these crystal structures confirmed that several findings can be conveyed from the bacterial LeuT to the human homologs, including transporter topology, conformational changes of the transport cycle [6,9–11] and the substrate binding site [12]. These structures also revealed important differences, as the human transporters do not share the dimeric form with LeuT, due to the conformation of transmembrane helix 12 (TMH12) [6,7]. TMH12 is central in the dimer interface of LeuT, but is remarkably different in dDAT and hSERT. Cross-linking experiments and Förster resonance energy transfer (FRET) measurements indicated the existence of higher-order oligomers of human monoamine transporters [13–20] in the plasma membrane. Functional studies revealed the importance of higher-order oligomers for their physiological function, and also revealed substrate-transport and inhibitor binding-dependent oligomerization [19,21–25]. Spectroscopic studies of hSERT showed that the transporter assumes a large distribution of oligomeric sizes [26,27]. The stability of the oligomer and protomer exchange rates were contingent on the presence of phosphatidylinositol-4,5-bisphosphate (PIP2) [28]. Monoamine transporters have mainly been simulated in their monomeric form, although LeuT-based dimers have also been examined [29,30]. Recently, protein docking predicted a single hDAT dimer conformation that included residue C306 in the interface, which was assessed using MD simulations [30] and site-directed mutagenesis studies.
In the current study, our key aim was to develop a comprehensive description of hDAT oligomerization by computational approaches. We hence characterized the interfaces of human DAT (hDAT) oligomers and identified the transmembrane helices and residues involved using a hDAT model based on the outward-facing dDAT crystal structure. Membrane properties, including lipid entropy, play a crucial role in the process of oligomerization. We have therefore used the unbiased DAFT (Docking assay for transmembrane components) [31] approach to directly follow the formation of hDAT dimers, starting from two randomly-oriented membrane inserted hDAT transporters. We observed the formation of four stable symmetric and four asymmetric dimer conformations. Importantly, we found only a very modest contribution of the bundle domain in the dimer interfaces, consistent with its role as a moving domain during the transport cycle.
Neurotransmitter transporters are known to oligomerize in the plasma membrane [32]. It was shown that the oligomeric state modulates transporter function [22], but structural knowledge is largely missing. Only a few residues have been found to most likely reside within protomer interfaces [13,14,17,21]. The aim of this study was to observe dimer conformations and to identify transmembrane helices and residues involved in hDAT dimerization. In this study we used an unbiased computational approach [31] (see Material and methods for detailed description), in which we repeatedly embedded two hDAT transporter molecules in random relative orientation in a palmitoyl-oleoyl-phosphatidyl-choline (POPC) membrane bilayer and simulated each system for 2 μs (Fig 1). This process was repeated 512 times to obtain an ensemble of structure and trajectories (total simulation time of over 1 ms) that is large enough to be representative for hDAT oligomerization.
Relative orientation of the two protomers in the dimer was assessed through the introduction of an internal coordinate system, as shown in Fig 1C. We first defined an orientation vector as the reference vector (red arrow) within the frame of each hDAT molecule. The position (β) of protomer B relative to protomer A is measured as the angle between the reference vector and the vector connecting the center of mass of the two hDAT monomers. In addition to the orientation described in the β angle, the second hDAT (protomer B) molecule can also rotate around it own axis. Therefore, we introduced a second angle (χ) that measures the orientation of the reference vector of protomer B relative to the line connecting the center of mass of the two hDAT molecules.
The orientation plot (Fig 2) summarizes across the ensemble the relative orientation of the two protomers using these two angles (β and χ). The orientation plot therefore reflects the frequency with which each orientation was observed, while not averaging over the symmetry-related orientation of the homodimers. The series of plots in Fig 2 shows the time evolution of orientations at 0.0 μs, 0.5 μs, 1.0 μs, 1.5 μs and 2.0 μs. Over time an enrichment of a few prominent clusters or orientations became apparent, visible as red-to-yellow regions, while the frequency of observing other conformations decreased, which are indicated by blue colored areas, in which the dimer probability is below average.
The orientation plot is symmetric with respect to the diagonal once converged, because we measure the orientation of homodimers. Clusters on the diagonal represent symmetric dimers, while all off-diagonal dimers are asymmetric, differing in the helices in the shared interfaces. We did not impose symmetry in the analysis, but rather used it as a measure of convergence. The inherent symmetry of the orientation plot requires that at full convergence all off-diagonal peaks must be present in both symmetry-related positions and must be sampled with the same frequency. The plot at 2.0 μs was remarkably symmetric, indicative of an almost fully converged dataset in which the appearance of additional conformations at even longer simulation times is unlikely. The time evolution of the degree of symmetry of the four off-diagonal clusters E-H showed that the ratio converges from 1:3 at 0.5 μs to 1:1.2 at 2.0 μs. The ratios were estimated by calculating the difference between the number of systems in the four clusters above the diagonal over the number of systems in the four clusters below the diagonal. The second applied measure of convergence was the time evolution of the potential energy of hDAT-hDAT interactions (S1 Fig). The plot showed that interacting dimers were present in large numbers and the profile is leveling off. Exchange between dimer conformations with the same interaction energy cannot be detected by the interaction energy plot, but would remain visible in the orientation plot. The plateau in the interaction energy plot is reached once the number of interacting dimers in the full ensemble has equilibrated and the populations of weaker and stronger interacting dimers no longer change. Experimental data for the parolog human SERT have shown that ~35% of all hSERT transporter are monomeric in the endoplasmatic reticulum (ER) membrane as well as in the plasma membrane [27,28]. The ER membrane is similar to our bilayer in that it mainly consists of phosphatidyl lipids, while being almost devoid of cholesterol and free of PIP2. Given the similarity between hSERT and hDAT we should also not expect that all hDAT transporters in the ensemble interact and form dimers. When comparing frequencies between symmetric and asymmetric dimers, the two symmetry-related off-diagonal conformations of an asymmetric cluster must be added together, while for the diagonal clusters these are already integrated, because of overlapping on the diagonal in the plot. In total, we found four symmetry-related conformations on the diagonal (cluster A-D) and four asymmetric conformations not on the diagonal (cluster E-H) were observed (See S1 Table for number cluster members).
The orientation plot showed that hDAT dimers clustered into a limited number of conformations. To set these into a structural perspective, we fitted all final structures (512 structures) to the first protomer and analyzed the position of the second protomer. The distribution of the second protomer was, as expected, not uniform throughout the ensemble (S2 Fig). In Fig 3A we show an overlay using one representative dimer per cluster (Fig 2D). All dimers were fitted to protomer A to place all 8 clusters into a single reference frame. Position and orientation of protomer B therefore shows its relative arrangement in the hDAT dimers. Importantly, this overlay revealed that possible dimer interfaces included a large part of the membrane-exposed hDAT surface, enclosing the transmembrane helices of the scaffold domain, but also highlighted a prominent exception, which was adjacent to the bundle domain. A comparison with the orientation plot (Fig 2) revealed that the frequency of finding dimers that included the bundle domain within the interface (corresponding to the region of ~105–150 °) was decreasing over time.
We then screened all trajectories to investigate if any dimer was formed in which the bundle domain was central for the dimer interface at any time-point throughout the 2 μs long simulations. In a condition of complete random encounters, we should have expected 20–40 contacts, but we found only 7 such trajectories (1.4% of all simulations). In all seven simulations, these dimers were unstable and separated within 0.5 μs (Fig 3B).
The orientation plot showed the formation of 8 distinct clusters. One representative dimer was extracted from every cluster, converted into fine-grained representation and simulated for 100 ns to test their stability at the all atom resolution of classical force fields. These systems proved to be stable, confirming that the structure obtained through the DAFT workflow at the coarse-grained representation are stable low-energy conformations (S3 Fig). Fig 4A and 4B show representative structures for every cluster, whereby all dimers are fitted to protomer A. The position of protomer B is thereby shown (and indicated in degrees) using the internal coordinate system as references, as is also used in Fig 2. Comparison between Fig 4A and 4B shows that the binding geometry of some clusters like cluster C and G have mutually exclusive conformations, while clusters A and G could in principle contemporaneously form hDAT-hDAT dimers, leading to higher oligomeric structures.
Analysis of the transmembrane helices and loops within the interfaces is summarized in Fig 4C. The axes of Fig 4C are labeled by residue numbers; transmembrane helices and (frequently observed) loops are indicated next to the axes, colored according to the colors used for all DAT structures. Fig 4C shows a contact heat map at single residue resolution, integrated over the last 100 ns of each trajectory. Contact frequencies between residues are shown in the triangle above the diagonal coded in gray scale. It shows that some interactions are very frequent, indicating that these interactions could be essential for dimer stability. We find that transmembrane helices TMH1, 2, and 7 are completely absent in any interface. TMH6 can be found in the interface, although it is limited to interactions of residues of the first helical turn.
The lower panels color residue contacts according to association to one of the clusters. Overall, most contact on or close to the diagonal are from symmetric dimers. A per residue analysis (S4 Fig) shows interactions and interaction frequency for all dimers associated with one of the clusters. The symmetric dimers include less transmembrane helices within the interface regions than the asymmetric dimers. This was expected, because in the symmetric assembly the same regions are present on both protomers. On average, the interfaces of the symmetric dimers consisted of 1–2 transmembrane helices or loops on each protomer surface. We can infer from experimental data on the hSERT [28] that showed exchange rates of protomers on a time scale of minutes, that dimer association and dissociation are slow processes. Interestingly, the exchange rate was comparable between the cholesterol-free ER membrane and the cholesterol-containing plasma membrane, if devoid of PIP2, indicating that the exchange rate is not very sensitive to the membrane composition. Sequential residues along a helix change orientation by ~100° relative to the main helix axis. The interaction of helices therefore follows an alternating pattern, which is visible in Fig 4C as a strand that is oriented parallel to the diagonal (for parallel-oriented helices) or normal to the diagonal (for antiparallel-oriented helices), while the interacting residues show the typical heptad repeat pattern (S4 Fig). The length for the pattern is representative for the extent of helix-helix interactions.
The symmetric interfaces of clusters B and D were dominated by interactions of residues outside transmembrane helices (Fig 5). Cluster B was mainly stabilized by interactions between the EL2 loops at the extracellular site, whereby the highly-conserved residues W184 and N185 were found in the core of the interface. It was shown experimentally that a mutation of W184 leads to a loss of surface expression [33]. Possible explanations might be a folding defect, which would lead to transporter degradation or a lack of transporter dimerization, which is required for ER export. Several lipids filled the space between the protomers. A similar lipid-bridged dimer was observed for cluster D, in which the protomer interactions were dominated by interaction between the C-terminal helices at the intracellular site. The dimer interface was stabilized by the salt bridge between R588 and E589 across the dimer interface, and further stabilized by additional interactions between aromatic residues on the C-terminal helix. The configurations of clusters B and D shared an additional property. The EL2 loop and the C-terminal helix were immersed in the membrane, and thus strongly interacted with the lipid headgroup region. In addition to direct interactions, the immersion perturbed the membrane bilayer, which could contribute to the attractive dimer-stabilizing forces due to membrane deformation.
In contrast, the symmetric interfaces of clusters A and C established transmembrane contacts (Fig 5). In cluster A, these comprised the entire TMH9, while TMH4 contributed mainly at its extracellular side. The geometry of the interface in cluster C was less specific, as TMH6 and TMH11 were within the interface in only a subset of the dimers. Common to both clusters was the fact that they included the cysteine residues C243 and C306, which have been shown by cross-linking experiments to be part of interfaces [13,14]. Cluster C showed dimers, which were stabilized by symmetric inter-molecular salt bridges between R304 and E307. These interactions oriented the two C306 residues to face each other at cross-linking distance. The conformation in cluster C was mainly stabilized by this pair of salt bridges, as the shared surface is otherwise small, leaving space to allow for dimer plasticity, mobility and dynamics.
The off-diagonal clusters in the orientation plot are asymmetric and therefore share different interfaces (Fig 5). The orientation plot (Fig 2) showed four main off-diagonal clusters and also revealed that the interface on one protomer was able to interact with more than one interface on the second protomer. Cluster E1/2 has several properties in common with cluster B, including the EL2 loop that is prominently present in the interface and the lack of extensive contacts in the hydrophobic core of the membrane. The EL2 loop interacted with TMH3 and 4 on the first protomer and TMH5 on the second. Cluster H1/2 shared the EL2 loop as part of the interface, but the orientation of the second protomer was remarkably different from cluster E2/3, as the contacting interface was mainly TMH12, though limited to its extracellular side. The A559V mutation is located within the extracellular region of TMH12 and is associated with Autism spectrum disorders (ASD) [34,35], caused by trafficking and functional alterations, while the P554L mutant, located in EL5, is associated with the Dopamine Transporter Deficiency Syndrome (DTDS) [36] and has been described as intracellularly retained.
The dimer clusters (F1/2 and G1/2) formed an extensive interface throughout the transmembrane region. Both clusters shared TMH11 in their interface, often including contacts to EL3. TMH11 has been reported to be located in dimerization interfaces [17] of human monoamine transporters. The main helix interfacing with TMH11 was TMH5 in cluster F1/2 and TMH4 in cluster G1/2. This change in relative orientation is achieve by a ~60° rotation of the second protomer. Cluster G1/2 is of particular interest, because residue C243 and C306 were found at cross-linking distance from each other. Cross-linking experiments showed that hDAT can form asymmetric trimer and tetramers [14] through a chemical bond established between C243 and C306.
We quantified the stability of observed dimers by determining Potential of Mean Force (PMF) profiles for dimer dissociation of two representative dimers per cluster (Fig 6) by applying the same CG system representation as used for the DAFT simulations. The starting structures of all systems are shown in S5 and S6 Figs. In the first step, the two protomers were pulled apart by steered MD (SMD) simulations in a 100 ns long trajectory, while restraining the relative orientation using the enforced rotation module. The applied pulling velocity (0.025 nm/ns) was set to be in the range of the fastest diffusion observed in the unbiased simulations. The rationale for selecting such a slow pulling rate was to use a velocity that is compatible with normal diffusion in the membrane in order to obtain SMD trajectories that minimally perturb the membrane. In the second step we extracted equidistant structures at 0.1 nm separation over 1.6 nm and carried out PMF calculations using umbrella sampling. Additional conformations with 0.025 nm increments over the first 0.4 nm were added in the initial rising part of the PMF to improve sampling and achieve sufficient histogram overlaps. Overall, the PMF profiles of dimers extracted from all clusters showed qualitatively similar profiles, with a dimer-stabilizing potential between 30 and 90 kJ·mol-1 (Fig 6). S2 Table reports the integrals of each PMF profile. The error bars reflect the variability observed in the PMF calculations and represent lower boundaries, because additional variability which might only appear over a much longer timescale is not sampled in the umbrella window calculations. The restraining forces of the enforced rotation module applied to prevent translation (S7 Fig) and rotation (S8 Fig), showed a Gaussian-like distribution, which indicates that the restraints did not mask any hidden energy gradient or forces. To provide a better estimate of the variability of the PMF profiles, we calculated two PMF profiles per cluster, using two different hDAT dimers as starting structures. The profiles of clusters D-H showed limited difference between the two profiles, suggesting that these are representative profiles. The profiles for the clusters A, B and C showed a large difference between the two profiles. The very large energy difference between the dimeric and the monomeric state of one profile of cluster A indicates that this profile might be overly attractive. The difference between the two profiles of cluster B and C are more likely a consequence of differences in the starting structures, which were selected according to the centrality of each cluster, but differed in the details of their interfaces (S5 and S6 Figs). The dimers of clusters B, D, but also F, showed weaker and broader interaction profiles of 20–50 kJ·mol-1. The dimers with a large transmembrane interface exhibited especially strong and short-range interaction profiles, showing 60–75 kJ·mol-1 of stabilizing energy. The short-range nature of these profiles indicate that they are dominated by van der Waals interaction between transmembrane helix side chains, while the broader profile indicates an important role of lipids and/or the changes on the intracellular and extracellular loops. Of particular importance were the details of loop interactions across the dimer interface as obtained from the final structure of the DAFT simulations. The broader minima were associated with the sliding of residues from the entangled opposing loops during the initial phase of separation.
The PMF profiles were strikingly different for the 7 dimers that formed transient interactions involving the bundle domain (Figs 3B and 6I). From every unbiased trajectory we extracted the frame with the smallest dimer-dimer distance as a starting point for the PMF calculations. All seven profiles showed a flat energy landscape, revealing that these structures were not held together by any stabilizing interactions, therefore transient and short-lived.
Binding of the negatively-charged lipid PIP2 was shown for hSERT to affect oligomerization [28,37,38]. The protomers in each oligomer exchanged with a timescale of minutes in both the plasma membrane (if depleted of PIP2) and the ER. The presence of PIP2 in the plasma membrane blocks this exchange process and kinetically traps the oligomer in its state at the plasma membrane. PIP2 carries a charge of -5. We thus assumed that binding is dominated by electrostatic interactions and calculated the electrostatic potential for each cluster (Fig 7) using apbs [39]. A representative structure for each cluster was converted from the coarse-grained representation to a fine-grained all-atom structure using the back-mapping module [40] by applying two steps of energy minimization and four steps of system relaxation. Visualization of the potential isosurfaces at 2 eV revealed potential PIP2 binding sites as areas of a positive electrostatic potential extending into the membrane. A large positive electrostatic field reaching into the membrane was generated by three clusters: the symmetric cluster C, which includes C306 in its interface at the extracellular site, the asymmetric clusters F1/2 and G1/2, which both showed extensive contacts across the lipid bilayer. The respective fields were generated by the positively-charged lysine and arginine residues of IL5 for the symmetric cluster C, by residues in IL3, IL5 and the N-terminus for cluster F1/2 and by residues in IL2, IL4 and IL5 in cluster G1/2. These fields extended into the headgroup region of the lipid bilayer, therefore pre-disposing these conformations for interactions with negatively-charged lipid such as phosphatidylserine or PIP2.
The physiological role of the hDAT is to rapidly remove the previously released dopamine by uptake in the pre-synaptic neuron, which leads to a fast drop of dopamine concentration. Oligomerization of human monoamine transporters is involved in transporter function at multiple levels: surface expression requires transporter dimerization to pass quality control at the ER exit sites for loading into COPII vesicles [41]. Transporter function seems to be modulated by oligomerization, including substrate transport activity and amphetamine-elicited neurotransmitter efflux by reverse transport through the transporter [19,23,24] as well as endocytosis from the plasma membrane [42]. Also, some evidence of cooperativity between transporter protomers has been reported in the literature [22,25,43]. Early studies on transporter oligomerization have identified C243 and C306 [13,14] in hDAT to be involved in oligomer formation. These two residues are positioned at opposite sites on the transporter surface, i.e. C243 is located within TMH4, while C306 is found towards the end of EL3. It was therefore conceivable that large oligomer structures of theoretically infinite size could form, if both elements are involved. Unexpectedly, only dimer, trimer and tetramers of hDAT were identified after chemical cross-linking using Cu2+ or Cu2+-phenanthroline as oxidizing agent. Interestingly, the C243 cross-link was formed only to a limited extent, in contrast to the C306 cross-link. These data on human DAT are fully consistent with our results: We found a symmetric cluster of conformations, in which two C306 residues were pre-disposed for cross-link formation, but we were not able to observe such clusters with residue C243 in the interface. Notably, chemical cross-link formation is not necessarily representative of a conformational equilibrium situation, because it reports on reactivity, rather than on equilibrium distribution. Thus, a dimer has to exist only for a time period that is long enough for the chemical reaction to take place. hDAT is not chemically cross-linked at the plasma membrane under normal conditions, while under oxidizing conditions [13,14] the cross-linked dimers were a minor component. This is important because we can infer that the C243 dimer should be very rare, while the C306 dimer should be of low abundance, albeit higher than the C243-C306 heterodimer species. Such low frequency is precisely what we observed in our dataset.
TMH11 and TMH12 have been observed to be within a dimer interface in hSERT [17]. The similarity of hSERT and hDAT in many functional and biophysical aspects suggests that such interfaces may exist in hDAT as well. We identified TMH11 in the interface of heterodimers, whereas the involvement of TMH12 appears to be more complex. Of all the transmembrane helices among the SLC6 transporter family, only the conformation of TMH12 is not conserved between the bacterial homolog LeuT and human transporters [7–9]. This is attributable to a prominent kink in the center of the membrane. The dimeric LeuT crystal structure showed TMH9/12 at its interface. However, the kink in TMH12 makes the same dimer geometry virtually impossible. We still find TMH12 in the dimer interface, nonetheless the interface is limited to residues located predominantly towards the extracellular and possibly intracellular ends.
Sequence analyses carried out before solving the first LeuT structure uncovered a putative leucine heptad repeat in TMH2, which was suggested to be involved in transporter oligomerization of rGAT1 [44] and hDAT [21]. Surface targeting of both transporters was largely reduced and FRET analysis established a loss of oligomer formation in rGAT1 following mutagenesis of the respective leucine side chains to alanine. However, the crystal structures of LeuT, dDAT and hSERT did not confirm the existence of the predicted leucine heptad repeat [45]. A π-helix element in the middle of TMH2 creates a frame shift in the positioning of the helix residues with respect to a canonical α-helix. Hence, the leucine residues can no longer be aligned to form a heptad repeat. Furthermore, two of the four residues are not surface-exposed and cannot be involved in protein-protein interactions.
The C-terminal helix, directly following TMH12, is absolutely essential for the surface expression of hDAT and the related transporters [32,42,46,47], since it plays a key role in protein folding and trafficking, as well as in the interaction with the transporter core [48,49]. Moreover, a motif located in the C-terminal region is necessary for recognition by the SEC24C vs. SEC24D components of the ER export (COPII) machinery [47,50]. In addition, the C-terminus harbors the FREK sequence (residues 587–590), which is the binding site for the small ras-like GTPase Rin1, that is involved in PKC-mediated endocystosis of the transporter [51,52]. Cluster D revealed dimers which interacted only through the C-terminal helices, stabilized by a salt bridge and aromatic interactions between the protomers, but devoid of contacts with the transmembrane region. The C-terminal helix was embedded in the lipid headgroup region of the membrane, thereby affecting the lipid bilayer. This is indicative of a general biophysical property, which can possibly be adopted by every member of the SLC6 transporter family, and is also consistent with the specificity of SEC24 isoforms required for ER export of individual proteins. Namely, there are four human SEC24 isoforms (SEC24A-D) that recognize and bind some ~6000 membrane proteins encoded for in the human genome.
The entire ensemble revealed that we obtained a set of structures that clustered into 8 dimer conformations, consisting of four symmetric and four asymmetric dimers, showing in total 6 partially overlapping and consequently mutually-exclusive interfaces. Mapping the entire ensemble revealed that the putative interfaces covered a large part of the hDAT surface: it consisted of the scaffold domain comprising TMH 4, 9, but also TMH 3, 8, and 12 at their extracellular or intracellular ends. In addition, TMH 5 and 11, which are diagonally oriented at the transporter surface, extensively contribute to the dimer interfaces. It is remarkable that the four helices of the bundle domain (TMH 1, 2, 6 and 7) appear to be absent from the dimer interface, with the exception of the first turn of TMH6 until residue W311, which formed contacts in some dimers. Analysis of all trajectories revealed that on rare occasions (a total of 7 trajectories), transient dimers formed with the bundle domain as the core of the dimer interface. However, they were always unstable and separated within 0.5 μs. The lack of such dimers was not due to limited sampling, because from our ensemble of 512 independent simulations, we would have expected 20–40 dimers forming through random encounters. Importantly, the PMF profiles of these seven dimers were flat (Fig 6I), verifying the lack of dimer-stabilizing interactions. Collectively, these data suggest that hDAT establishes a force that opposes dimer formation at its bundle domain surface. Our data indicate that the bias is, at least, two-fold: i) hDAT seems to perturb the membrane to make encounters less likely. ii) Interactions of residues in the bundle domain with a second hDAT transporter are generally weak, because EL3 forms a rim-like structure at the transporter-lipid interface and is thereby floating in the membrane headgroup region, strongly limiting the extent of the transporter surface in direct contact with the dimer interface.
Biological membranes are a complex mixture of many types of lipids, asymmetrically and non-homogeneously distributed over the two leaflets. A full description of this complexity would be desirable, but is not yet achievable. The cholesterol content in the ER membrane is below 5%, and it is devoid of PIP2. Oligomerization of the homolog hSERT was shown to be indistinguishable between the ER membrane and a PIP2 free plasma membrane, suggesting that transporter oligomerization is largely insensitive to the membrane composition. The use of a POPC lipid membrane is therefore a good compromise, as it consists of the dominant phospholipid species in the plasma membrane as well as the ER membrane, though details might differ.
Not all membrane properties have the same importance for oligomerization. Experimental data showed that unexpectedly the cholesterol content does not play a major role [27]. Lipids form annular structures around proteins. The analysis of lipid order [53] revealed that the first layers of lipids showed increased ordering and therefore changed dynamics that differed from unperturbed lipids. The ordering of lipids does not differ much between the clusters A-H (S9 Fig) and the transient dimers which include the bundle domain in the interface, suggesting that the annular lipids do not play a major role in preventing dimer formation at the bundle domain. The hydrophobic region of membrane proteins does not always perfectly match with the thickness of the membrane. The hDAT is no exception as it shows regions which induce membrane thinning, while other parts of the hDAT lead to a thickening of the membrane (S10 Fig). The analysis of membrane thickness [53] indicated a pattern. The unstable transient dimers that included the bundle domain frequently showed areas of strong membrane deformation with opposing sign in close proximity originating from dimerization. In contrast, strong changes in membrane thickness across the interface were largely absent in the stable dimer, suggesting that mismatching membrane thickness might play a role in preventing dimer formation at the bundle domain by inducing strong membrane deformations.
Our studies of LeuT and hDAT simulations [54,55] showed that the scaffold domain anchors and stabilizes the transporter in the membrane. In addition, crystal structures showed that the bundle domain, which moves during the transport cycle (Fig 8), comprises only a small fraction of the membrane-exposed surface of hDAT. The work required for pushing against the membrane to allow for bundle domain movements is therefore limited. Any larger interaction surface including TMH 1, 2, 6 or 7 would block transport by locking the bundle domain in one single conformation, thereby impeding the switch between the inward- and outward-facing conformations, a movement necessary for the transport cycle (Fig 8A and 8B). We also identified residue C306 within the interface; this residue is located in EL3, and in close proximity to TMH6 (as shown in Fig 5C). However, the frequency of formation of this particular dimer is relatively low. The dimer was stabilized by a double inter-molecular salt bridge between R304 and E307, though the same residues could also form an intra-molecular salt bridge, thereby weakening the interaction across the dimer interface. Mutation of either of the two charged residues (R304 and E307) resulted in a strong reduction of cross-linked hDAT, following oxidative treatment [30]. This distinct symmetric dimer interface, including EL3, could still allow for transport, since the interacting side chains of R304 and E307 are long and flexible and can permit a large range of adjustments whilst sliding relative to one another. The direction of conformational changes during the transport cycle is such that this last segment of EL3 would slide in parallel, and thus limit the impact of this dimer geometry on the conformational changes in the transport cycle. The R304A and E307A mutants showed reduced expression but a gain in transporter function [30], providing further evidence in support of our data. The salt bridge may be able to switch between the inter- and intra-molecular arrangements. This in turn allows for conformational changes in the transport cycle, which may not be necessary for the R304A and E307A mutants.
As discussed above, several oligomeric contact points identified in this study have already been observed experimentally. Our results suggest the existence of a range of symmetric and asymmetric DAT dimers rather than one well-defined oligomeric structure. The coarse-grained force field, which is necessary to reach the time scales required in this study, comes with the limitation that approximations in the short-range interactions needed to be introduced in the description of the system. Application of the DAFT approach to study GPCRs dimerization showed reproduced experimental data [31,56–58], indicating that the coarse-grained system does nevertheless allow for direct study of protein dimerization. It was recently indicated that protein-protein interactions might be too strong [59] in the Martini Force Field. This leads to an overstabilization of dimers, which most affects the weak and transient unstable dimers. The hDAT dimer geometries observed in the 8 clusters are least affected, because the timescale for dimer separation is on the timescale of minutes, as experimentally measured for hSERT [26–28]. Overstabilization would therefore not significantly affect the geometry of dimer, but the PMF profiles might be exceedingly attractive. It will be important to experimentally confirm the predicted dimer interfaces and establish whether all interfaces would equally contribute to larger oligomer structures. The observation of a set of possible hDAT-hDAT interface and oligomeric structures has additional functional implications that warrants further experiments. It will be rather unlikely that one defined positive or negative allosteric interaction would exist between protomers. In this light, it is surprising that a very recent study [25] reported that a DAT dimer likely consists of one active and one non-active protomer as has been suggested for SERT [43]. Such a binary impact (active or inactive) is difficult to reconcile with the range of oligomers and interfaces as reported in this study or observed for SERT in single molecule studies [28].
The lipid species PIP2 is well known for its interaction with—and regulation of—membrane proteins. PIP2 is a negatively-charged phospholipid that resides in the intracellular membrane leaflet of the plasma membrane and accounts for 1% of the total membrane lipid content. We have previously reported that: (i) PIP2 interacts with monoamine transporters, (ii) the size and stability of transporter oligomers depends on PIP2 [28] and (iii) that transporter efflux is PIP2-dependent as well [37], thereby affecting behavioral response to psychostimulants [38]. The hSERT residues SERT-K352 and SERT-K460 are exposed on the cytosolic membrane and were found to play a role in the PIP2-mediated effects [37]. Here we analyzed the electrostatic field generated by hDAT dimers. A subset of dimer conformations showed a positive electrostatic field that connected the two protomers, whilst bridging over the lipid bilayer. It is conceivable that these extended areas of positive electrostatic fields could act as the regions that attract PIP2 lipids, since PIP2 carries a total charge of -5. The size of the PIP2 headgroup would than allow for interactions with both hDAT protomers. These areas included three regions on hDAT: a) IL1 and IL4, b) IL5 and c) IL2, TMH1 and IL3. The loops carry a large number of positively-charged lysine and arginine residues, including the sites, which were found to interact with PIP2 in the homolog SERT (SERT-K352 in IL3 and SERT-K460 in IL4). Removal of these charges would reduce the positive field, hence reducing the interaction with PIP2. It is worth mentioning that three of these clusters were asymmetric; the exception being the sparsely populated and symmetric cluster C. In hSERT, depletion of PIP2 led to a dramatic reduction of higher oligomers, while the dimer population remained mostly unchanged. Assuming that hSERT and hDAT behave comparably, our results indicate that the highly populated symmetric dimers would not be affected by PIP2 binding, while the highly populated asymmetric dimers would be. We can therefore surmise that the assembly of oligomeric structures larger than dimers ought to involve asymmetric interfaces that are PIP2-sensitive.
The oligomeric arrangement of the monoamine transporters differs from ion channels and ionotropic receptor in that the latter form well-defined trimeric, tetrameric or pentameric structures, while the former show an oligomeric size distribution that decreases mono-exponentially, indicating dynamic exchange, thereby excluding the possibility of a single well-defined oligomeric arrangement. Structures larger than dimers require at least two non-overlapping interfaces. Our simulations elucidate that hDAT forms stable dimers through 6 different interfaces, including the already identified interfaces comprised of residues C243 and C306, TMH11 and TMH12. Importantly, the bundle domain appears to be excluded from these interfaces, thus allowing for efficient motions during the transport cycle. The number of transporters must be high to support fast neurotransmitter clearance, as neurotransmitters need to be removed faster than they are released to prevent large concentration buildup and systemic spillover. The observed interface distributions support a high transporter density via several interfaces, while maintaining maximum transporter function, by avoiding transport impeding interactions at the bundle domain, that must be free to move during the transport cycle.
Homology models of the human dopamine transporter (hDAT) were created from residues 44 to 602 based upon the outward-open crystal structure of the Drosophila melanogaster dopamine transporter (dDAT) [7] (PDB ID: 4XP1) using modeller 9.15 [60] which share 72% sequence identity of their transmembrane region (TMH1 to TMH12). The sequence alignment is given in the S11 Fig. The models contained 2 Na+ and 1 Cl- ions in their respective binding sites, and a single dopamine molecule bound in the central binding site as observed in the crystal structure of dDAT. A previously modeled extracellular loop 2 (EL2) [54] was introduced by fitting to the structurally shared connecting secondary structure elements of helix TMH3 and the helix of EL2, because the EL2 of dDAT was truncated for crystallization and showed extensive crystal contacts with the co-crystallized antibody. The 250 models, produced by applying the automodel procedure using the refinement protocol “normal”, were scored and sorted by their discrete optimized protein energy (DOPE) score [61]. The best 20 of these models were re-ranked by their root mean-squared deviation (RMSD) of the Cα atoms from the template. The model with the lowest RMSD has a DOPE score of -78938, and shows a Ramachandran plot (S12 Fig), where 94.0% of residues are in the most favourable region, 5.2% in the additional allowed region, 0.4% in the generously allowed region and 0.4% in the disallowed region. The quality of the model was also assessed using the QMEAN score [62], which shows that the local quality of the model is especially high for the transmembrane region (S13 Fig). This model was inserted as apo protein into the DAFT work flow. The Cl- and two Na+ ions need to be removed from the structure, because these ions cannot be described correctly by the Martini coarse-grained force field. Ions can only be represented by including their first hydration shell as one single particle, while these three ion are stripped of their hydration shell when bound to hDAT. We therefore also removed dopamine, because its binding depends on the present of bound sodium.
The DAFT (Docking Assay For Transmembrane components) approach [31] allows to identify protein-protein interactions and binding orientations. It uses molecular dynamics simulations for exploring the conformational search space, therefore explicitly including the entropic component associated with the dimerization event. Moreover, it uses a coarse grain representation to allow for a microsecond to millisecond time scale. The DAFT workflow consists of several different modules coupled together. Conversion of the all atom fine-grained (FG) homology model of apo hDAT to the coarse-grained (CG) representation of the ElNeDyn CG implementation of the Martini force field was done by the martinize module [63]. Then, 512 hDAT dimer systems were created with random relative orientations in the membrane plane and a center of mass distance between the two protein molecules of 8.4 nm, which resulted in a typical minimal distance between protomer of > 2.5 nm. The only exception was the orientation in which the C-terminal helices of both protomers face each other, in which case the minimal distance was close to 2.0 nm. The time evolution of trajectories starting from these conformation showed that the closer distance was without bias towards a specific dimer formation during the DAFT simulations as simulations drifted away from this initial orientation. The membrane was built around the proteins using POPC lipids by the insane module [64], adding CG water and ions (Fig 1). Each system contained 172 to 190 lipids per leaflet, the salt (NaCl) concentration was set to 150 mM. The systems were first energy minimized, than equilibrated with NVT ensemble simulations while restraining the protein. Production runs (2.0 μs each) with NPT ensemble were carried out using a timestep of 20 fs, the v–rescale thermostat [65] was used to maintain the temperature at 310 K, the weak coupling barostat applied to keep the pressure at 1 bar [66]. The electrostatic interactions were defined according to the Martini force field by a coulomb type shift and the values are switched between 0 to 1.2. The Van der Waals interactions were represented by Lennard-Jones potentials using the shift type to 0 between 0.9 to 1.2 nm.
Selected final structures of the CG systems were back-mapped into FG all atom representation using the backward module [40] using two steps of energy minimization and four steps of system relaxation. The resultant structures were simulated using the amber99sb-ildn force field [67] for the transporter and applying the Berger parameters for the lipids [68], as this combination of force fields showed the best performance [69]. For system relaxation, the protein was restrained by applying 1000, 100, 10 and 1 kJ·mol-1 restraints on the protein, each for one nanosecond. Further, the production run was carried out for 100 ns. The temperature was maintained at 310 K using the v-rescale temperature coupling [65] and the pressure was maintained semi-isotropically at 1 bar using the weak coupling barostat [66]. The pressure coupling time constant was set to 1 ps and the compressibility to 4.5×10−5 bar-1. Long range electrostatics interactions were represented by the particle mesh Ewald method [70] with a cutoff of 1.0 nm. The Van der Waals interactions were imposed using Lennard Jones potential using the 1.0 nm cutoff. All the bonds are constrained using LINCS [71].
To quantify the energies involved in dimer stabilization we carried out PMF calculations using the CG system representation. We selected two dimers from each cluster and also included all seven system of the DAFT dataset which showed dimer contacts that included the bundle domain at any timepoint of the 2.0 μs long trajectories. The center of mass distance was used as the reaction coordinate. The starting conformations for the PMF calculations were develop by a series of steered MD (SMD) simulations [72], in which we increased the center of mass distance between the protomers. The velocity for the movement of the reference point was set to 0.025 nm/ns, which is in the range of the fastest diffusion as measured by protein diffusion in the unbiased simulations of individual protomers before dimerisation. Large membrane deformation by a pulling velocity, which is much faster than normal diffusion can thereby be avoided. To allow for sufficient protomer separation, we extend the membrane in the direction of the reaction coordinate by 5 nm and filled the additional space with lipids, water and ions. While restraining the hDAT protomers, these extended system were equilibrated using first a time step of 2 fs for 10 ns to relax initial structural strain, followed by a 100 ns long equilibration using a 20 fs time step.
During the SMD simulations we restrained the relative orientation of the protomers using the enforced rotation module [73] implemented in Gromacs [74] to limit the conformational search space, thereby making sampling of the reaction coordinate affordable. This procedure is formally correct and allows to set the bound and the unbound state into an energetic relationship. The restraints allow to circumvent the sampling problem at intermediate distances by limiting the exploration of phase space to the reaction coordinate. Quantification of the complete free energy hypersurface would require extensive sampling of the hDAT binding and unbinding events, for which the single molecule data of the hSERT indicate that the process requires minutes to equilibrate [28].
From these SMD trajectories we extracted hDAT dimer conformations along the reaction coordinate from the first 1.6 nm at 0.1 nm intervals. Each structure was used as the starting point for an umbrella simulation of 100 ns, applying the enforced rotation module to maintain relative hDAT orientation. Moreover a harmonic potential of 1000 kJ/mol/nm was applied between the center of mass of two protomers, acting on their backbone. Additional umbrella windows with a 0.025 nm spacing and a restraining force of 5000 kJ/mol/nm were added within the first 0.4 nm to enhance sampling at shorter distances, resulting in a total of 32 umbrellas per PMF profile. Block analysis showed that all systems relaxed within 50 ns; we therefore use the second half of each umbrella trajectories for the PMF analysis using WHAM [75] and applying the bootstrap method [76].
|
10.1371/journal.pgen.1001332 | A Novel Unstable Duplication Upstream of HAS2 Predisposes to a Breed-Defining Skin Phenotype and a Periodic Fever Syndrome in Chinese Shar-Pei Dogs | Hereditary periodic fever syndromes are characterized by recurrent episodes of fever and inflammation with no known pathogenic or autoimmune cause. In humans, several genes have been implicated in this group of diseases, but the majority of cases remain unexplained. A similar periodic fever syndrome is relatively frequent in the Chinese Shar-Pei breed of dogs. In the western world, Shar-Pei have been strongly selected for a distinctive thick and heavily folded skin. In this study, a mutation affecting both these traits was identified. Using genome-wide SNP analysis of Shar-Pei and other breeds, the strongest signal of a breed-specific selective sweep was located on chromosome 13. The same region also harbored the strongest genome-wide association (GWA) signal for susceptibility to the periodic fever syndrome (praw = 2.3×10−6, pgenome = 0.01). Dense targeted resequencing revealed two partially overlapping duplications, 14.3 Kb and 16.1 Kb in size, unique to Shar-Pei and upstream of the Hyaluronic Acid Synthase 2 (HAS2) gene. HAS2 encodes the rate-limiting enzyme synthesizing hyaluronan (HA), a major component of the skin. HA is up-regulated and accumulates in the thickened skin of Shar-Pei. A high copy number of the 16.1 Kb duplication was associated with an increased expression of HAS2 as well as the periodic fever syndrome (p<0.0001). When fragmented, HA can act as a trigger of the innate immune system and stimulate sterile fever and inflammation. The strong selection for the skin phenotype therefore appears to enrich for a pleiotropic mutation predisposing these dogs to a periodic fever syndrome. The identification of HA as a major risk factor for this canine disease raises the potential of this glycosaminoglycan as a risk factor for human periodic fevers and as an important driver of chronic inflammation.
| Shar-Pei dogs have two unique features: a breed defining “wrinkled” skin phenotype and a genetic disorder called Familial Shar-Pei Fever (FSF). The wrinkled phenotype is strongly selected for and is the result of excessive hyaluronan (HA) deposited in the skin. HA is a molecule that may behave in a pro-inflammatory manner and create a “danger signal” by being analogous to molecules on the surface of pathogens. FSF is characterized by unprovoked episodes of fever and/or inflammation and resembles several human autoinflammatory syndromes. Here we show that the two features are connected and have the same genetic origin, a regulatory mutation located close to a HA synthesizing gene (HAS2). The mutation is a 16.1 Kb duplication, the copy number of which correlates with HAS2 expression and disease. We suggest that the large amount of HA responsible for the skin condition predisposes to sterile fever and inflammation. HAS2 was previously not known to associate with autoinflammatory disease, and this finding is of wide interest since approximately 60% of human patients with periodic fever syndrome remain genetically unexplained. This investigation also demonstrates how strong artificial selection may affect not only desired and selected phenotypes, but also the health of domestic animals.
| Shar-Pei dogs have been companion animals for centuries within China where they were commissioned to guard and hunt, and to sometimes serve as fighting animals. At the beginning of the communist era dog ownership was highly taxed and the breed was brought close to extinction. A few Chinese Shar-Pei dogs were exported to the United States in the early 1970's and Shar-Pei descending from this limited number of animals have undergone strong selection for a wrinkled skin phenotype and heavily padded muzzle and are called the “meatmouth” type (Figure 1A–1C) and have now found global popularity. The ancestral Shar-Pei, referred to as the “traditional” type Shar-Pei, still occurs and it presents with a less accentuated skin condition (Figure 1D). The major constituent of the deposit in the thickened skin is hyaluronan or hyaluronic acid (HA). HA is a large, multifunctional, linear, negatively charged, non-sulfated glycosaminoglycan of the extracellular and pericellular matrices. It is composed of repeating disaccharides and is widely spread throughout epithelial, connective and neural tissues [1], [2]. The biological role of HA depends on its size, location and equilibrium between synthesis and degradation [1]–[3]. Meatmouth Shar-Pei show two- to five-fold higher serum levels of HA compared to other breeds [4], allowing us to propose the term hyaluronanosis, a definition also used for a comparable human condition [5]. HA is synthesized at the plasma membrane by three HA synthases, HAS1, HAS2 and HAS3, with HAS2 being the rate limiting-enzyme [6]. HAS2 is overexpressed in dermal fibroblasts of Shar-Pei compared with other canine breeds [7] suggesting a regulatory mutation as causative for hyaluronanosis. HA is deposited throughout the skin of Shar-Pei, often in microscopic lakes and grossly evident vesicles, leading to the formation of thickened skin folds around the head and tibiotarsal (hock) joints (Figure 1E). Almost all Shar-Pei seem to be affected by hyaluronanosis, however the extent varies among individuals and adults exhibit less skin folds and hyaluronanosis than puppies. Strong selection by breeders for dogs who retained their skin folds into adulthood has altered the phenotype of the breed to the more commonly heavily wrinkled meatmouth type.
Meatmouth Shar-Pei also suffer a strong predisposition to an autoinflammatory disease, Familial Shar-Pei Fever (FSF), which clinically resembles some human hereditary periodic fever syndromes, such as Familial Mediterranean Fever (FMF) [8]. Both diseases are characterized by seemingly unprovoked episodes of fever and inflammation and both FMF and FSF present as short (12–48 hour) recurrent bouts of high fever, accompanied by localized inflammation usually involving major joints (especially the tibiotarsal joints). Patients with FMF or Shar-Pei with FSF can suffer episodes as often as every few weeks, but in the interim seem symptom free. However, since acute phase reactants may endure between episodes, a subclinical state and chronic autoinflammation may persist (Linda Tintle unpublished data). As a secondary complication, the chronic state puts human patients, as well as affected Shar-Pei dogs, at risk of developing reactive systemic AA amyloidosis and subsequent kidney or liver failure [8], [9]. In Shar-Pei, the fever episodes are typically more frequent during the first years of life and the percentage of affected dogs is very high, estimated to be 23% in the US in 1992 [9].
In order to find candidate loci for the breed-specific phenotype (hyaluronanosis), known to be under selective pressure, we screened the genome for signatures of selective sweeps. These sweeps can be recognized as long chromosomal segments with a low degree of heterozygosity within populations [10]. Using 50,000 single nucleotide polymorphisms (SNPs) distributed throughout the dog genome, the level of heterozygosity in windows of ten consecutive SNPs was compared between a set of Shar-Pei (n = 50, all from the US, Table S1) and the average of 24 other canine breeds (n = 230). On four chromosomes (Cfa 5, 6, 13 and X) the reduction in heterozygosity in Shar-Pei was greater than 4-fold the average of control breeds (Figure 2A). The strongest signal of reduced heterozygosity appeared within a 3.7 Mb stretch on chromosome 13 (CanFam 2.0 Chr13: 23,487,992–27,227,623) (http://genome.ucsc.edu/) near the HAS2 gene, where almost complete homozygosity was observed in Shar Pei (Figure 2C). Here the reduction in heterozygosity was greater than 10-fold in Shar-Pei and several smaller regions showed complete homozygosity. The same region was confirmed to show high levels of homozygosity when the analysis was repeated in 37 additional Shar-Pei dogs sampled from Spain (Table S1) and was overlapping a sweep region reported by others for this breed [11]. The strong signal, together with the known function of HAS2 and its aberrant expression pattern in Shar-Pei, made this region an obvious candidate for the mutation causing the wrinkled skin phenotype (hyaluronanosis).
In parallel, we performed a genome-wide association study to map the susceptibility locus for FSF, using Shar-Pei strictly classified as FSF affected (n = 24, classification code FSF+A and FSF+, described in Materials and Methods) and unaffected (n = 17, classification code H+, described in Materials and Methods). Five SNPs were significantly associated (best SNP praw = 7.0×10−7, pgenome = 0.005 based on 100,000 permutations; software package PLINK http://pngu.mgh.harvard.edu/~purcell/plink [12]), all on chromosome 13 (CanFam 2.0 Chr13: 22.4–30.7 Mb, Figure 2B). After correcting for putative stratification, two outlier cases were removed (Figure S1) and the same SNPs, forming the same signal of association remained (best SNP praw = 2.3×10−6, pgenome = 0.01; Table S2) with a genomic inflation factor of 1.2. When the association signal and the sweep signal were compared they appeared interspersed, so that individual SNPs were either part of homozygous regions or showed association with FSF (Figure 2C). It was therefore difficult to determine exactly where the strongest association fell, as variation is required to detect association.
Targeted sequence capture technology was used to further investigate the sweep signal and to search for the hyaluronanosis causative mutation. We resequenced 1.5 Mb around and upstream of our candidate gene, HAS2 (CanFam 2.0 Chr13: 22,937,592–24,414,650) in four Shar-Pei (two meatmouth type with high serum HA levels and two traditional type) and three control dogs from other breeds. The obtained sequences were mapped to the boxer reference sequence providing at least 5X coverage for 96–98% of the resequenced region in each individual. The targeted region also included the large intergenic noncoding RNA, HAS2 antisense (HAS2as; Table S3) which has been proposed as a negative post-transcriptional regulator of HAS2 mRNA [13]. After masking repetitive sequences we identified ∼670 indels and ∼1,500 SNP in each dog (Table S4) as well as two overlapping duplications in the Shar-Pei (Figure 3A). Nine mutations (eight SNPs and one indel) located in conserved elements as well as two SNPs possibly regulating transcription, were selected for further investigation due to their unique pattern in the sequenced Shar-Pei dogs. Additional genotyping in Shar-Pei and dogs from other breeds (Tables S1, S5) showed these mutations were not specific to Shar-Pei and the variants were subsequently excluded as causative.
The two duplications were named after the Shar-Pei type in which they were first identified. The “meatmouth” duplication was the larger fragment, 16.1 Kb (CanFam 2.0 Chr13: 23,746,089–23,762,189) with breakpoints located in repeats (a SINE at the centromeric end and a LINE at the telomeric end) and individual copies separated by seven base pairs (Figure 3B). The “traditional” duplication was 14.3 Kb (CanFam 2.0 Chr13: 23,743,906–23,758,214) and was identified in the two Shar-Pei with a less accentuated skin phenotype (Figure 3B). We first examined the duplications via Southern blot with control breeds (n = 2), traditional (n = 2) and meatmouth Shar-Pei (n = 6) (Figure 3C). As the digest cut outside and within both duplications, we were able to observe the absence of the variants from control breeds and separate restriction patterns in traditional and meatmouth type Shar-Pei. Interestingly, one meatmouth dog contained both duplication types (Figure 3C lane 6 and confirmed by PCR across break points, data not shown). Two copy number assays were developed to quantify these elements. The first (CNV-E) measured only the meatmouth duplication whilst the second (CNV-748), detected both the traditional and meatmouth duplications. Copy number analysis was estimated as the relative fold enrichment (ΔΔCt) between an amplicon within the duplication and one outside the duplication in a housekeeping gene. Assay CNV-E was run on 90 Shar-Pei and 73 dogs from 24 other breeds (Table S1) and assay CNV-748 on a subset of 44 Shar-Pei and 14 dogs from other breeds. Assay CNV-748 demonstrated that both the traditional and meatmouth duplications are unique to the Shar-Pei breed (Figure 4 and Figure S2).
We used the results of both assays to search for a relationship between Familial Shar-Pei Fever (FSF) and either meatmouth copy number (Assay CNV-E), traditional copy number (the normalized difference between CNV-748 and CNV-E) or total traditional+meatmouth copy number (Assay CNV-748). Shar-Pei dogs were strictly classified as affected by FSF (n = 28, FSF+A and FSF+) or unaffected by FSF (n = 16, H+). The most significant association was found when only the meatmouth copy number was considered (p<0.0001, Figure 4) although a weaker association with total copy number (p<0.01) was also seen. The observed association between fever and meatmouth copy number, despite the very high homozygosity in this region, strongly suggests that a high copy number is not just a genetic marker for FSF but is causally related to the development of disease.
Of the 153 dogs analyzed with the meatmouth copy number assay, 31 Shar-Pei and 18 control animals also had serum measures of HA available. No clear association was detected between HA levels and copy number (Figure S3), however the mean HA level in Shar-Pei with ≥ six copies was 905±403 ug/L (n = 21), whilst Shar-Pei with fewer copies had a mean concentration of 770±494 ug/L (n = 12) and control breeds had HA serum levels of 206±145 ug/L (n = 19). Interestingly, the three traditional Shar-Pei dogs had serum HA levels between 73 and 266 ug/L, which fell within the normal range [4].
The link between copy number and the expression of HAS2 and HAS2as was examined on a smaller scale using dermal fibroblasts cultured from six separate meatmouth Shar-Pei. The expression of both genes was calibrated against the Shar-Pei with lowest copy number (CNV estimate = 5) and both genes showed an increasing trend of expression with copy number (Figure 5). These data suggest that a regulatory element for HAS2 is located in the duplicated region, however the interpretation of the HAS2as result is less clear. A single study of a human osteosarcoma cell line demonstrated that the expression of two isoforms of HAS2as were able to reduce HAS2 expression, and so these mRNAs may act as regulators of HA production [13]. Our data could indicate that HAS2as expression is also influenced by a regulator element in the duplication, or that HAS2as is up-regulated in response to HAS2 levels. If either of these scenarios were true, it is possible that if RNA expression were measured at multiple time points we would see temporal HAS2 repression. It could also be that the interaction between canine fibroblast HAS2 and HAS2as does not mirror the human system and that the canine antisense mRNA is non-functional. At present our results must be considered as preliminary and it is clear that further exploration of the interaction between canine HAS2 and HAS2as is required.
Here we have identified a 16.1 Kb duplication located approximately 350 Kb upstream of HAS2. This is clearly a derived mutation since it occurs as a single copy sequence in other dog breeds. We postulate that this is a causative mutation associated with both hyaluronanosis and Shar-Pei fever, as the observed correlation between copy number and susceptibility to Shar-Pei fever was not expected if this was a linked, neutral polymorphism. We suggest that the unique region of the meatmouth type duplication identified in Shar-Pei contains one or more regulatory elements that alter the expression of HAS2. It appears possible that as the duplication copy number increases, so does the copy number of potential enhancer elements within the duplication, likely leading to a higher expression of HAS2 and elevated HA levels, and resulting in the development of hyaluronanosis in this breed. We propose a scenario whereby the traditional duplication arose de novo in the traditional type of Shar-Pei causing a milder skin phenotype. This event made the region unstable and allowed the second meatmouth duplication to occur. Breeders subsequently selected the meatmouth duplication as a higher copy number enhanced the phenotypic effect in appearance. However, it is not yet possible to say whether the meatmouth duplication first occurred at low frequency in the Chinese Shar-Pei population and quickly rose during breeding in America, or if the mutation occurred spontaneously during breed expansion in the West.
Tandem duplications are notoriously unstable and may show copy number variation due to unequal crossing-over, as is clearly illustrated by the copy number variation of a 450 Kb duplication associated with dominant white colour in pigs [14]. The meatmouth Shar-Pei duplication adds to the list of copy number variants (CNVs), which affect phenotypic traits in domestic animals (e.g. dominant white in pigs [14], gray color in horses [15], the hair-ridge in Rhodesian ridgeback dogs [16], and pea-comb in chicken [17]), several of which are linked not only to the desirable trait but also to disease. Interestingly, all of these except pea-comb, represent novel duplications derived from single copy sequences. This is in contrast to most reported CNVs in humans, which are mainly benign and represent expansions or contractions of duplicated sequences [18].
Although we failed to find a significant correlation between serum HA levels and copy number, this does not exclude our proposed hyaluronanosis scenario. Difficulties in correlating fluctuating serum levels of HA with other clinical and biomedical parameters have also been reported in many human studies, where no or only weak correlations were observed [19], [20]. We have shown that the 16.1 Kb duplication appears only in meatmouth Shar-Pei, a breed type that has elevated levels of HA compared to both traditional Shar-Pei and other breeds, and that copy number correlates with a breed-specific syndrome associated with excessive HA deposition and the over expression of a HA synthesizing gene. Because HA is primarily a component of the extracellular matrix, serum measurements may only broadly reflect total body HA.
Hyaluronan can bind to several cellular receptors (e.g. CD44, RHAMM and layilin), however it is the interaction between CD44 and HA which acts as a biological regulator, differentially modulating the cellular microenvironment in response to homeostatic versus inflammatory conditions [21]. Alterations in the balance between native high molecular weight HA versus fragmented HA may result in activation of innate immunity. HA has been linked to sterile inflammation as an endogenous response molecule to sterile tissue injury [21]. Shorter fragments of HA can be generated by environmental insults such as sterile trauma [22], reactive oxidative species (ROS) [23], or pathogenic hyaluronidases, and it is these low molecular weight fractions which can become pro-inflammatory danger associated molecular pattern (DAMP) molecules [22], [24] mimicking microbial surface molecules.
Using a mouse model, Yamasaki and colleagues [25] showed that HA can interact with the cell through two separate pathways that culminate in the release of IL-1β, which together with IL-6, is one of the main promoters of fever. In the first route, CD44 bound HA is degraded at the plasma membrane by hyaluronidase-2 (HYAL2) prior to endocytosis and further cleavage by lysosomal hyaluronidase-1 (HYAL1). The resultant small intracellular oligosaccharides of HA activate the NLRP3 inflammasome, a multiprotein complex consisting of the NLRP3 scaffold, the ASC adaptor and caspase-1 [26]. In the second arm, the CD44-HA complex activates toll like receptors 2 and 4 (TLR2 and 4), leading to intracellular IL-1β mRNA transcription and the formation of pro-IL-1β. Activation of the NLRP3 inflammasome by HA oligosaccharides allows cleavage of this pro-IL-1β by caspase-1 and subsequent release of IL-1β. The NLRP3 inflammasome is present in the cytosol of many cells including monocytes, macrophages and mast cells, and has been implicated in the pathogenesis of numerous autoinflammatory diseases in humans including the cryopyrin-associated periodic syndromes which result from mutations in NLRP3/CIAS1 [26].
The actual role of excessive HA in Shar-Pei needs to be investigated further. Shar-Pei may experience exogenous fragmentation of their over-abundant HA from sterile or pathogenic trauma. This, plus endogenous degradation of excessive native HA, may contribute to induction of recurrent episodes of fever and inflammation. Acute fever events in Shar-Pei respond rapidly to dipyrone, a potent antipyretic and analgesic pyrazolone, which has been demonstrated to inhibit IL-1β induced fever [27]–[29 and Linda Tintle unpublished data]. It is therefore not surprising that the strong selection on the hyaluronanosis phenotype, with increased levels of cutaneous HA, may predispose Shar-Pei to autoinflammation, potentially contributing to other pathologies seen in this breed. One such example is renal medullary amyloidosis. Histopathologically, kidneys of Shar-Pei in renal failure have multifocal non-suppurative tubulointerstitial nephritis with fibrosis. Medullary amyloidosis predominates and glomerular deposition, although consistent, is highly variable in its extent [8], [30]. The renal medulla is naturally HA rich and enhanced renal interstitial HA accumulation can be coupled to inflammatory responses, such as ischemia-reperfusion injury, transplant-rejection, tubulointerstitial inflammation and diabetes [31]. In addition, Shar-Pei are prone to mast cell disease including mast cell tumors [32], [33]. The binding of HA to CD44 has been shown to play a critical role in regulation of murine cutaneous and connective tissue mast cell proliferation [34]. As the CD44-HA interaction may modulate local immune responses through regulation of mast cell functions [35], excessive HA and its subsequent damage and degradation may play a role also in the Shar-Pei breed’s predilection for allergic skin disease and other mast cell driven inflammation.
This study suggests that HAS2 dysregulation can trigger a periodic fever syndrome in dogs and therefore it will be relevant to examine the approximately 60% of human fever patients who currently have unexplained disease. Previously, the role of hyaluronan in sterile inflammation has focused on HA signaling and degradation; for example a deficiency of hyaluronidase causing mucopolysaccharidosis type IX in humans has some autoinflammatory features [36]. However by directly implicating HAS2 in inflammation, we suggest that a reexamination of genes further up the biosynthetic pathway, such as those involved in HA synthesis and polymerization is called for. In addition, the canine mutation appears regulatory in nature and therefore regulators of HA should be also be included in a broader scope pathway analysis of human patients with unexplained autoinflammatory disease.
Finally, this study illustrates how copy number variations can shape phenotypic traits and how strong artificial selection for certain phenotypic traits may not only affect the desired trait but also the health of the animal.
All dog samples were collected from pet dogs after owner consent following the ethical approval protocols (SLU, Dnr: C103/10, MIT 0910-074-13). DNA was extracted from blood samples using QIAamp DNA Blood Midi Kit (QIAGEN) or PureLink Genomic DNA kit (Invitrogen). All dogs, their breed type, geographic origin, health status and experiment in which they were utilized are listed in Table S1.
Classification of Shar-Pei fever: Purebred Shar-Pei individuals were divided into the following six groups based on their medical records and evidence by owner and/or veterinarian:
1. FSF+A, the individual had experienced recurrent episodes of high fever accompanied by inflammation of joints from an early age (less than one year old). Additionally, post-mortem examination detected depositions of amyloid in kidneys and/or liver (amyloidosis).
2. FSF+, the individual had experienced recurrent episodes of high fever accompanied with inflammation of joints from an early age (less than one year old).
3. Atypical FSF, the individual had experienced occasional unexplained fever episodes or recurrent episodes with a late onset (greater than three years old).
4. H+, the individual had never experienced unexplained fever and/or inflammation, was older than five years old at the time of sampling and also lacked first-degree relatives that could be classified into the groups FSF+A, FSF+ or Atypical FSF.
5. H-, the individual had never experienced unexplained fever and/or inflammation but was younger than 5 years at the time of sampling and/or had first-degree relatives that could be classified into the groups FSF+A, FSF+ or Atypical FSF.
6. Unknown, the individual’s medical record was not available.
Hyaluronanosis: Serum Hyaluronic Acid (HA) concentration was used as a proxy for hyaluronanosis but no distinct cut-off value was established. However, dogs with normal and abnormal concentrations of serum HA were interpreted as before [4]. HA measurements were performed using the Hyaluronan ELISA kit (Echelon Biosciences INC) according to the manufacturer’s instructions. The absorbance was read at 405 nm, and a semi-log standard curve was used to calculate hyaluronic acid concentrations.
A whole genome scan was performed with two array types, the 27K (v1) and 50K (v2) canine Affymetrix SNP chips. Results were called using Affymetrix’s snp5-geno-qc software. The 50K array was used when the rate of heterozygosity was calculated for US Shar-Pei separately and for a reference group of 24 other breeds. The ratio of heterozygosity in 10 SNP (≈1 Mb) sliding windows between the two groups was used as a measure of relative heterozygosity. To look for regions of homozygosity within the Shar-Pei genome only, the software package PLINK [12] was used. This was performed both for the 50 K array with 50 US Shar-Pei and replicated for 37 Spanish Shar-Pei using 22,362 SNPs genotyped with the Illumina CanineSNP20 BeadChip. These data were collected with an Illumina BeadStation scanner and genotypes were scored using GenomeStudio. Regions of homozygosity were defined if shared across all Shar-Pei samples.
A case-control association analysis using 17,227 SNP common to both the 27K and 50K arrays (MAF>0.05, call rate >75%) was performed in Shar-Pei classified as affected (FSF+A and FSF+, n = 39) or unaffected (H+, n = 17) by Shar-Pei fever. The software package PLINK [12] was used for the analyses and to ensure genome-wide significance, p-values were corrected for multiple testing. Values used are the max (T) empirical p-values obtained after 100,000 permutations. To assess whether signals from the two genome scans overlapped, the 39 Shar-Pei with unambiguous phenotypes were analyzed with the 17,227 SNPs common to both SNP platforms.
Targeted capture of the 1.5 Mb candidate region (CanFam 2.0 Chr13: 22,937,592–24,414,650) was performed using a 385K custom-designed sequence capture array from Roche NimbleGen. Hybridization library preparation was performed as following: Genomic DNA (15–20 µg) was fragmented using sonication; blunting of DNA fragments using T4 DNA Polymerase, Klenow Fragment and T4 Polynucleotide Kinase; adding A-overhangs using Klenow Fragment exo− and ligation of adaptors using T4 DNA Ligase with Single-read Genomic Adapter Oligo Mix (Illumina). All enzymes were purchased from Fermentas and used following manufacturers instructions. Purification steps were performed using QIAquick PCR Purification Kit (QIAGEN). Hybridization was performed following the manufacturer’s instructions without amplification of the fragment library prior to hybridization. Eluted captured DNA and uncaptured libraries were amplified using Phusion High Fidelity PCR Master Mix (Finnzymes) and the SYBR Green PCR Master Mix (Applied Biosystems) was used to estimate the relative fold-enrichment. Capture libraries with the estimated enrichment-factor of >200 were sequenced using Genome Analyzer (Illumina) and obtained sequences were aligned to CanFam 2.0 [37] and to the targeted region using Maq assembly (http://maq.sourceforge.net/) [38]. For each individual, sequence coverage was calibrated by dividing the coverage in 100 bp windows by the average coverage for the total region. Three control breeds (Pug, Neapolitan Mastiff, Standard Poodle) and two of each type of Shar-Pei (meatmouth type and traditional type) were sequenced. The two traditional type Shar-Pei were sequenced at different read lengths but were aligned using the same strict criteria (allowing two mismatches per read) and therefore vary in the percentage of mapped reads as well as coverage when compared to the other individuals. Individual 7 (Table S4) was sequenced from whole genome amplified material and this may have impacted the ability to map reads and detect SNPs. This individual was not plotted in Figure 3A, but was used in downstream analyses.
All primers used were designed using Primer3 (http://frodo.wi.mit.edu/primer3/) [39] and are listed in Table S6. PCR and Sanger Sequencing was performed to investigate putative mutations (ten SNPs and one indel) and were carried out with 20 ng genomic DNA using AmpliTaq Gold DNA Polymerase (Applied Biosystems) following the manufacturer’s instructions. The amplification of the copy number variant (CNV) breakpoints was performed with 400 ng of DNA and a Long-range PCR with Expand Long Template PCR System Mix 1 (Roche), cloned using Zero Blunt TOPO Cloning Kit (Invitrogen) and plasmid DNA prepared using QIAprep Spin Miniprep Kit (QIAGEN). PCR products and plasmids were sequenced using capillary electrophoresis 3730xl (Applied Biosystems), aligned and analyzed using CodonCode Aligner version 2.0.6 (CodonCode).
Four micrograms of genomic DNA from each sample was digested with BsrGI (New England BioLabs) and separated on a 0.7% agarose gel. A 910 bp probe (targeting CanFam 2.0 Chr13: 23,746,12–23,747,522) was used to detect the duplicated region.
Estimation of copy number was performed using the comparative CT (ΔΔCT) relative quantification method and a calibrator animal (German Shepherd 95). The duplex reaction contained a primer limited copy number assay (CNV-E: 300 nM each of forward and reverse primers, 250 nM FAM labeled MGB probe; CNV-748: 50 nM of forward and 300 nM reverse primers, 250 nM FAM labeled MGB probe, Applied Biosystems) and a reference assay designed to C7orf28B (900 nM of forward and reverse primers, 250 nM VIC and TAMRA labeled probe, Applied Biosystems). Real Time PCR was performed in quadruplet using 10 ng of gDNA, Genotyping Master Mix (Applied Biosystems) and a 7900 HT Real Time PCR machine (Applied Biosystems). The PCR primers used and dogs evaluated can be found in Tables S4 and S1 respectively.
Cultures of dermal fibroblasts were established from skin samples of Shar-Pei dogs as described previously [40]. Skin samples were well shaved and cleaned with 70% EtOH/Betadine before biopsy and cell isolation. Fat tissue and blood vessels were removed from the skin and then samples were washed with PBS, cut into small fragments (0.5 cm2) and digested with dispase II solution (Boehringer Mannheim) for 16 h at 4°C. The next day, after incubation for 30 min at 37°C in the same solution, the dermis was separated from the epidermis. Washed dermal samples were chopped into 1 mm3 fragments and incubated for 140 min in 15 ml of DMEM per gram of skin containing 30 mg bacterial collagenase (Gibco), 18 mg hyaluronidase, 12 mg pronase, 1.5 mg DNAse, supplemented with bovine albumin (all from Sigma) and antibiotics. After digestion, cutaneous cells were washed with PBS and grown in a humidified atmosphere at 37°C with 5% CO2 for two days. Medium was changed twice a week and cells were used at passages two-five.
RNA extraction from fibroblast cultures was performed as described elsewhere [41]. 500 ng of RNA was reverse transcribed using the High-Capacity cDNA Archive Kit (Applied Biosystems) with random primers and following the manufacturer’s instructions. Two assays were designed to target HAS2 and HAS2as cDNA, respectively. Real Time PCR in a volume of 20 ul was performed in duplicate using SYBR Green PCR Master Mix (Applied Biosystems) and primers at 300 nM in a 7900 HT Real-Time PCR system (Applied Biosystems) with standard cycling. PCR specificity assessment was performed by adding a dissociation curve analysis at the end of the run. Each amplification run contained negative controls. Relative fold-enrichment was performed using the comparative ΔCT-method with Glucose-6-phosphate dehydrogenase (G6PD) for normalization.
http://pngu.mgh.harvard.edu/~purcell/plink/
http://www.codoncode.com/
http://genome.ucsc.edu/
http://maq.sourceforge.net/
http://frodo.wi.mit.edu/primer3/
|
10.1371/journal.pgen.1006857 | Genomic introgression mapping of field-derived multiple-anthelmintic resistance in Teladorsagia circumcincta | Preventive chemotherapy has long been practiced against nematode parasites of livestock, leading to widespread drug resistance, and is increasingly being adopted for eradication of human parasitic nematodes even though it is similarly likely to lead to drug resistance. Given that the genetic architecture of resistance is poorly understood for any nematode, we have analyzed multidrug resistant Teladorsagia circumcincta, a major parasite of sheep, as a model for analysis of resistance selection. We introgressed a field-derived multiresistant genotype into a partially inbred susceptible genetic background (through repeated backcrossing and drug selection) and performed genome-wide scans in the backcross progeny and drug-selected F2 populations to identify the major genes responsible for the multidrug resistance. We identified variation linking candidate resistance genes to each drug class. Putative mechanisms included target site polymorphism, changes in likely regulatory regions and copy number variation in efflux transporters. This work elucidates the genetic architecture of multiple anthelmintic resistance in a parasitic nematode for the first time and establishes a framework for future studies of anthelmintic resistance in nematode parasites of humans.
| Teladorsagia circumcincta is an economically significant nematode (roundworm) pathogen affecting sheep and goats in temperate regions of the world. The widespread use of prophylactic treatment has resulted in rapid selection for anthelmintic (anti-worm drug) resistance in this and other species of livestock parasites. The mechanism of resistance is not well understood because most studies have focused on the role of candidate genes using simplistic models of single gene selection, despite evidence that the evolution of resistance is more complex. Here, we report on a comprehensive whole-genome analysis that elucidated resistance-associated genes, which was facilitated by developing a pair of T. circumcincta strains sharing a largely common genetic background but differing markedly in their susceptibility to anthelmintic drugs. The results show that multiple genetic factors contribute to anthelmintic resistance in a variety of ways, including possible reduction/modulation in target site sensitivity, reduced target site expression, and increased drug efflux, to name a few. This suggests that drug resistance in these parasites is a multifactorial quantitative trait rather than a simple discrete Mendelian character. With this study, we established a genomics-based experimental paradigm for investigating anthelmintic resistance, at a time when its medical importance is rapidly increasing.
| Anthelmintic resistance is already a global problem for agriculture and a growing concern in relation to human pathogens [1, 2]. In the absence of effective vaccines, treatment and prophylaxis of helminthiases rely on a limited number of chemotherapeutic agents whose efficacy is increasingly undermined by the selection and spread of resistant parasites. Although fundamental to our ability to conserve sensitivity to existing drugs and to design improved interventions, the molecular and population genetic bases of anthelmintic resistance remain inadequately understood [3, 4]. To date, most studies have focused on the role of individual candidate genes such as drug targets or transporters. However, while such studies have been instrumental in identifying some causal genetic variants associated with drug resistance, frequently they have accounted for only a proportion of the drug resistant phenotypes present in the population, suggesting that the trait probably has a complex multi-genic nature [5–7]. Efforts to comprehensively map functional polymorphisms and to clarify the genotype-phenotype relationships in anthelmintic resistance have been challenging, given the relatively poor genetic and experimental tractability of helminth systems, which has impeded genome-wide studies beyond targeted analysis of particular candidate genes [8].
The problem of anthelmintic resistance is most severe in the trichostrongylid nematodes of livestock and particularly those infecting small ruminants such as sheep and goats [2]. The troubling propensity of these parasites to develop drug resistance has been attributed to their enormous effective population size and the resulting genetic diversity upon which selection is able to act [9, 10] and although variation is a prerequisite for selection, the extreme genetic heterogeneity in parasite populations often confounds the identification or association of genetic components contributing toward anthelmintic resistance. This difficulty is further hampered by factors such as the degree of parasite population connectivity due to parasite and/or host movement [11], the influence of population size and life history traits on genetic drift within parasite subpopulations, and the variation in local parasite management strategies, all of which likely influence the ability to detect and correctly interpret genetic differentiation between anthelmintic resistant and susceptible parasites [10]. Our approach towards identifying drug resistance associated genes involved the controlled crossing of a multidrug-resistant parasite strain with a characterized susceptible strain followed by repeated backcrossing and drug selection, which resulted in the introgression of resistance associated alleles into a largely susceptible, partially inbred genetic background. By identifying the alleles derived from the original resistant parent in the resulting backcrossed progeny, it was possible to generate a genetic map of resistance loci within the genome. Similar approaches have been used in mapping drug resistance associated loci in Haemonchus contortus [12, 13] and elucidating the genetic basis of drug resistance in some trematode parasite species [14].
In this study, we extended our previous work [15] by combining a genetic introgression approach with whole-genome sequencing to further elucidate the genetic basis of field-derived multiple-anthelmintic resistance in Teladorsagia circumcincta, the most economically important nematode pathogen affecting sheep and goats in temperate regions of the world. T. circumcincta is a monoxenous, obligately sexual species that infects the fourth stomach (abomasum) of small ruminants, leading to reduced wool, milk and meat production, and in severe cases, death. Widespread anthelmintic resistance has arisen in this trichostrongylid parasite, including multiple-anthelmintic resistance to all major broad-spectrum drug classes available prior to 2008 (i.e., benzimidazoles, imidazothiazoles and macrocyclic lactones, which target microtubule polymerization, nicotinic acetylcholine receptors and glutamate-gated chloride channels respectively) [16] and also to the more recently released amino-acetonitrile derivatives [17]. Through controlled genetic crosses set up by surgical transplantation, we undertook a serial backcrossing experiment that aimed to introgress the resistance-related genes from a field isolate into the genomic background of a partially inbred susceptible recurrent parental strain. Using this partially inbred susceptible parental strain, we generated a draft reference genome of T. circumcincta by de novo assembly, which was subsequently used to conduct comparative genome-wide single nucleotide and copy number variant analyses of the resistant strain. In addition, using a combination of pooled (Pool-seq) and individual (ddRAD-seq) genome sequencing and RNA-seq, we identified genes with differential patterns of diversity associated with multiple-anthelmintic resistance.
We generated a draft genome of T. circumcincta using the partially inbred anthelmintic susceptible strain (Sinbred) which was used as the recurrent parent in the backcrossing program undertaken to introgress anthelmintic resistance-associated genes/alleles into a susceptible genetic background (Fig 1A and 1B). The draft nuclear genome of ~701 Mb (93.4% CEGMA completeness [18]) comprises 81,730 supercontigs (S1 Table), with 35.0% (28,621) of the supercontigs accounting for 90% of the genome. The GC content was 44.8%. The amino acid composition was comparable to that of other phylogenetically close parasitic species such as Necator americanus or non-parasitic Caenorhabditis elegans (S2 Table). In total, 1,583 repeat families were predicted and annotated, spanning 38.5% of the genome (S3 Table). We predicted a total of 25,532 protein-encoding genes, representing 2.3% of the genome at an average density of 36.4 genes per Mb with an average GC content of 47.8%. Compared to C. elegans, the gene density in T. circumcincta is lower and the average size of gene loci is larger (Mann–Whitney U test, P < 2.2 × 10−16) with longer introns (S1 Table). The majority of predicted genes (80.6%) were supported by transcriptional evidence from mixed-sex adult worm samples with RNA-seq coverage of at least 50% of the length of the annotated coding exons. We predicted secreted proteins (1,603 classical and 9,642 non-classical secretion) and putative membrane-bound proteins (3,749), representing 44% and 15% of the proteome respectively. Functional annotation of deduced proteins on the basis of primary sequence similarity comparisons identified 4,456 unique InterPro domains, 1,563 Gene Ontology terms, and 7,458 KEGG Orthology groups, for 66%, 51%, and 64% of the T. circumcincta genes respectively. When considered together, 78% of all T. circumcincta genes had some form of putative functional annotation. In spite of our inbreeding (two generations of sibling mating) efforts to reduce heterozygosity in preparation for genome sequencing, the quality of the final assembly still suffered from residual heterozygosity, which is consistent with previously reported genome assemblies of obligate outcrossing nematode species with a large effective population size [19, 20]. Although polymorphic haplotypes can be collapsed into consensus sequences during assembly, high genetic diversity tends to result in a fragmented, larger-than-expected assembly [21]. Reliably discriminating uncollapsed alleles from truly paralogous loci remains a significant challenge, and this caveat calls for a careful interpretation of our reference-alignment-based variant analysis.
To analyze genetic variation between the RS3 and Sinbred strains, whole genome re-sequencing analysis was conducted using DNA obtained from pools of 300–500 mixed-sex worms. Approximately 92-fold coverage of the genome was obtained in total across the populations (43.6× and 48.3× from RS3 and Sinbred populations, respectively). Based on the depth of coverage, mapping quality, and gapped regions across all loci, 68.8% of the genome (482.3/700.6 Mb) and 90.6% of the coding sequences (14.9/16.4 Mb) were estimated to have at least the minimum sequence coverage for variant detection in both populations (S4 Table). A set of 17.6 million SNPs was obtained, of which 17.2 million (97.8%) were bi-allelic. The number of segregating (polymorphic) sites was overall ~2-fold lower for the Sinbred strain than for the RS3 strain (7,354,798 vs 16,489,377) (see S5 Table) indicating that the partial inbreeding strategy we adopted to reduce heterozygosity in the susceptible reference genome was successful. While a relatively small proportion of SNPs were differentially fixed in the two populations (Sinbred: 4,094; RS3: 147,114 / 17,176,467), there was a notable excess of private SNPs in the RS3 population (9,617,901 + 19,932 cf 422,182 + 81,072) which were most likely introgressed from the resistant parent strain, Rpar. In addition to private Rpar derived SNPs, the majority of SNPs observed in the Sinbred population were also segregating in the RS3 population (6,851,544 / 7,354,798) (S5 Table), as expected from the introgression strategy. Two-dimensional allele frequency spectra based on the bi-allelic sites illustrates this asymmetric distribution of private alleles with concentration of counts in cells along the vertical axis representing the RS3 population (Fig 1C). The observed pattern is consistent with the expected outcome of our experimental design which relied on a high level of genetic divergence between the two parental isolates (Sinbred and Rpar) and a unidirectional gene flow driven by the repeated use of Sinbred in backcrossing. In both populations, low-frequency SNPs were in deficit relative to neutral expectations (genome-wide Tajima’s D: 2.08 and 2.10 for Sinbred and RS3 respectively), likely due to a combination of ascertainment bias resulting from the limited sampling depth, the exclusion of singleton polymorphisms and the random loss of rare alleles following the inbreeding and introgression strategies employed in the construction of these strains. In addition, the observed level of genetic variability, particularly in the RS3 population, may be an underestimation considering a possible mapping bias against non-reference alleles. While these biases have the potential to increase uncertainty in population genetic parameter estimation, they likely have limited impact on our ability to detect outlier genomic regions showing the most extreme levels of divergence between the RS3 and the Sinbred populations.
In the RS3 parasites we expected the introgressed alleles associated with anthelmintic resistance to be contained in the divergent genomic regions originating from the Rpar field isolate, which were maintained by drug selection in the face of repeated gene flow from the Sinbred reference strain. We further expected that independent meiotic recombination events would lead to variation in the introgression break points among the haplotypes segregating in the RS3 population such that, at the population level, a gradient of allelic divergence would be created peaking around the directly selected loci. A genome-wide scan of FST following kernel smoothing resulted in demarcation of contiguous regions of the genome with high levels of population differentiation, representing putative introgression blocks. Outlier regions were determined on the basis of the empirical distribution of the smoothed FST values (Fig 2A) with the goal of prioritizing candidate variants under anthelmintic selection as specific targets for future functional studies. Using 4.5 standard deviations above the mean FST as a cutoff (i.e., z-score > 4.5; see Methods), genomic regions of ~0.86 Mb were identified across 34 contigs, of which 25 overlapped with a total of 58 protein-coding genes (Fig 2B). Considering the fragmented nature of our draft reference genome, these regions may not all represent independent unlinked loci, particularly when outlier windows are located near the ends of the contigs and thereby miss flanking regions of low FST. One important consequence of this relatively fragmented assembly is that we cannot be certain of exactly how many high FST outlier regions (or QTL) differentiate Sinbred and RS3, so we have focused on high FST SNPs that fall within these outlier regions, especially where those SNPs fall within or close to predicted genes for which a plausible case can be made for variation in or around that gene to contribute to variation in drug response. Although the size of the individual outlier regions and the number of genes annotated in each were heterogeneous, the majority of the identified regions spanned less than 100 kb (median: 35.5 kb; interquartile range: 27 kb) and harbored less than 4 genes (Fig 2A; S6 Table). Of notable exception was the 290 kb region located on Contig53, which contained 16 outlier genes. While this region may harbor multiple, spatially separated causal variants collectively resisting the gene flow from the Sinbred population, it is more likely that recombination has not yet substantially eliminated hitchhiking loci due to a reduced local recombination rate and/or an overall insufficient number of serial backcrosses and drug screening. Although regions containing resistance loci are expected to display higher population differentiation relative to the genomic background, non-uniform distribution of shared ancestral polymorphisms and within-population allelic diversity has likely added a layer of noise to our FST-based introgression mapping approach. At the most fundamental level, however, mapping resolution is limited by the extent to which causal variants are decoupled from neutral hitchhiking loci, and therefore, additional rounds of backcrosses and drug screening would be expected to have helped more fully resolve causal variants from those that are closely linked. Notwithstanding these caveats, this analysis reveals an architecture of resistance genetics that is characterized by multiple regions of elevated [outlier] FST between Sinbred and RS3. This observation leads to the conclusion that multidrug anthelmintic resistance is likely a polygenic trait and to the hypothesis that these outlier regions of elevated FST represent quantitative trait loci (QTL) that are the products of selection for resistance.
To test this hypothesis, and to help prioritize the candidate genes located in the outlier regions (Fig 2; S7 Table) for more detailed analysis, we critically evaluated whether any of these genes have known/predicted functions that can be plausibly connected to anthelmintic resistance in light of our current understanding of the mechanisms of drug action [7]. Among the most notable candidates were a β-tubulin gene (TELCIR_01271), a major target of benzimidazole anthelmintics, and putative orthologs of Cel-unc-29 nicotinic acetylcholine receptor (nAChR) subunit (TELCIR_06180) that may constitute a component of a levamisole-sensitive receptor in T. circumcincta. Interestingly, additional members of the Cys-loop ligand-gated ion channel families (LGICs) were represented including putative orthologs of the Cel-acr-11 nAChR (TELCIR_03607) and the Cel-lgc-54 LGIC (TELCIR_00170). Overall, GO terms attributable to these LGICs, such as acetylcholine-activated cation-selective channel activity (P = 2.6 × 10−3), extracellular ligand-gated ion channel activity (P = 6.3 × 10−3), and postsynaptic membrane (P = 6.7 × 10−3) were significantly overrepresented among the outlier genes (Table 1). These results underscore the potential contribution of target gene variation in the development of anthelmintic resistance in T. circumcincta under field conditions, although our data do not rule out a more complex genetic architecture involving additional, presently uncharacterized genes. To further contextualize our findings in relation to previously reported anthelmintic resistance-associated genes and gene families, we examined gene-wise FST values and highlighted non-synonymous SNPs that showed not only significant (P < 1.0 × 10−5) but indeed substantial differentiation between the susceptible and the resistant strains (Fig 3; S7 and S8 Tables).
In the present genome assembly, we identified two paralogs of β-tubulin (isotype-1 and -2) that are co-orthologous to Cel-ben-1, the locus which confers benzimidazole (BZ) sensitivity in C. elegans [22]. This finding is in line with the model of a lineage-specific duplication in trichostrongylid species [23]. Both isotypes have been implicated in BZ resistance [24, 25], with the hypothesis that selection occurs in two stages [26]: an initial reduction in diversity at isotype-1 followed by the loss of isotype-2. We observed only isotype-1 (TELCIR_01271) variation associated with the outlier loci with high FST (z-score > 4.5) in our genome-wide survey of the resistant backcross progeny (RS3). Furthermore, no evidence of selection was detected in any of the remaining members of the T. circumcincta β-tubulin gene family (Fig 3A). In the case of β-tubulin isotype-1, two non-synonymous coding variants, E198L (GAa/TTa) and F200Y (tTc/tAc), were exclusively found in the RS3 population and present at allele frequencies of 28.1% and 72.8%, respectively. The F200Y variant confers BZ resistance in H. contortus [27], and has been widely recognized in many species of parasitic nematode as a major resistance determinant. Although amino acid substitutions at position 198 (e.g., E198A) are less common in nematodes, variants at this position have been linked to BZ resistance phenotypes in H. contortus [28] and, more recently, in T. circumcincta [29], and molecular modeling suggests that the associated loss of hydrogen bonding interactions may play a role in the resistance mechanism [30]. We reconstructed three segregating haplotypes of β-tubulin isotype-1 in the RS3 population over the exonic region harboring E198L and F200Y variants (Fig 3A; S1 Fig). The inferred haplotype structure indicates that (i) F200Y occurs on at least two distinct and diverse haplotype backgrounds, suggesting multiple independent origins of the variant allele, and (ii) E198L and F200Y variants occur in trans on separate haplotypes. In agreement with this haplotype reconstruction, genotyping of individual male worms from the Sinbred (n = 94) and RS3 (n = 79) populations failed to detect any individuals homozygous for both resistance alleles (R198R198/R200R200) although worms homozygous resistant at one locus only (S198S198/R200R200 and R198R198/S200S200) were observed, as were double heterozygotes (S198R198/S200R200) (S2 Fig; S9 Table). Considering that S198R200 haplotype was segregating in the RS3 population at a minimum inferred frequency of 79.7% (S9 Table), the absence of single heterozygotes (especially worms heterozygous for only P198) was consistent with (and also supported) the conclusion of our Pool-seq analysis that R198R200 haplotype was not present in the RS3 population. Sequencing of isotype-1 from the same individuals confirmed the existence of multiple haplotypes, supporting the conclusion that BZ resistance conferring alleles arose several times in the OR parent of RS3 (S3 Fig), as has been reported for BZ resistance conferring alleles in the UK [29]. Although further work will be necessary to fully determine the extent to which E198L contributes to the overall resistance phenotype, the absence of haplotype(s) simultaneously harboring both variants suggests that, under field conditions, E198L can confer BZ resistance (and hence was selected) independently of F200Y, and that β-tubulin carrying both variants either is detrimental to organismal fitness (i.e., negative intramolecular epistasis) or the double mutation event (or recombination over an interval of <6bp) required to give rise to a cis haplotype is sufficiently unlikely that it is not observed.
Cholinergic anthelmintics, such as LEV and pyrantel, induce spastic neuromuscular paralysis by selectively opening nAChRs, a family of pentameric ion channels belonging to the Cys-loop LGIC superfamily that convert neurotransmitter binding into membrane electrical depolarization. Each receptor subunit has an N-terminal extracellular ligand-binding domain (ECD) followed by four transmembrane helices (TMD) that form the ion channel [31]. Nematode genomes encode a large number of nAChR subunits (~30) [32] and different subunit combinations result in pharmacologically distinct receptors [33]. Although the precise subunit composition of LEV-sensitive nAChR in T. circumcincta has not yet been determined, a putative model for trichostrongylid species suggests a likely involvement of parasite orthologs of Cel-unc-29, Cel-unc-38, Cel-unc-64 and Cel-acr-8 in receptor formation [33], and mutation, truncation and decreased expression of these subunits have been observed in field-selected LEV-resistant trichostrongylids [34–36]. Tci-unc-29.4 and Tci-acr-11 nAChR genes were identified in our FST outlier analysis and, in addition, the genomic regions encoding Tci-unc-29.2 (TELCIR_06181) and Tci-acr-8 (TELCIR_04316), displayed relatively high FST values (z-score of 4.32 and 4.18, respectively) (Fig 3B). Within these subunit genes, 16 non-synonymous coding variants were found. When considering the potential consequences of the substitutions based on the amino acid properties [37] and their locations relative to the ligand-binding or the transmembrane domains, the probability that any of these variants have drastic detrimental effects on protein structure and function appears to be relatively low. This observation is consistent with the view that, under field conditions, loss of function variants are likely to experience negative selection due to reduced fitness. Several of the variants (e.g., T388I and *462W (stop loss) in Tci-acr-11 and P24Q in Tci-acr-8) appear to have a greater potential to alter protein function, although the allele frequency of these variants in the resistant population is not substantially different from that in the susceptible population, suggesting that they are unlikely to play a direct role in LEV resistance (S8 Table). Further biochemical, pharmacological and structural modeling work will be necessary to fully assess and understand the impact of these alleles on drug-target interactions. These results also raise the possibility that drug selection may have acted primarily on the noncoding regulatory variants of the FST outlier nAChR genes. A preliminary assessment of the transcript abundance levels (see Methods) suggested that Tci-acr-11 transcript was substantially less abundant in the resistant strain (RS3) relative to the susceptible strain (Sinbred) (RS3/Sinbred log2 ratio: -5.02; see S7 Table), in a manner similar to previous reports that showed decreased expression of nAChR subunit genes in various LEV-resistant trichostrongylid populations [38–40]. However, because of the limited mapping resolution of the present study and our generally poor understanding of the functional consequences of noncoding variants, we are unable to identify specific candidate noncoding mutation(s) that could contribute to the resistance phenotype. Furthermore, although mutant screens in C. elegans indicate that genetic variations in calcium-mediated muscle contraction signaling pathway and ancillary proteins involved in nAChR assembly/maintenance may influence LEV susceptibility [41, 42], we did not observe any significant evidence of genetic differentiation among the genes implicated in the LEV excitation-contraction pathway in the RS3 population (S7 Table).
The predicted gene TELCIR_00170 (Tci-lgc-54) is one of the top FST outliers in our analysis (S7 Table). It belongs to the Cys-loop ligand gated chloride channel branch of the LGIC superfamily and is distinct from the nAChRs associated with LEV resistance (Fig 3C) and from GluCl and GABA receptor family members. The likely C. elegans ortholog, Cel-lgc-54, is described as a predicted “ligand unknown” biogenic amine-gated chloride channel [43] and as a GABA-receptor [44] but has not been implicated previously in relation to IVM resistance. Although the ligand for nematode LGC-54 is not yet identified, the predicted protein contains a tryptophan in ligand-binding loop C (amino acid position 231), which has been hypothesized to be a key residue for binding amines [45], and it is known that other family members (including Cel-lgc-55, the most closely related paralog in C. elegans) are activated by serotonin, dopamine and tyramine [46–48]. Furthermore, a gene encoding a putative dopamine receptor, Hco-ggr-3, has been implicated in IVM resistance in H. contortus [48]. In the RS3 population, we identified 5 non-synonymous variants in Tci-lgc-54: T9S, K11E, S18fs, A20P, and S336N. The former four are located upstream of the N-terminal ECD, and the latter is located at the beginning of the cytosolic loop between the third (M3) and the fourth (M4) TMD alpha helices. Notably, the frameshift variant at position 18 introduces a premature stop codon and the A20P variant located at the predicted signal peptide cleavage site (position 20–21) has a potential to interfere with the proper signal peptide processing (S4 Fig). Failure of signal peptide cleavage is likely to result in mislocation and/or degradation of the protein and thus behave as a loss-of-function mutation. It is of interest in this context that (a) large deletion alleles of Cel-lgc-54 and Cel-ggr-3 in C. elegans are viable suggesting that these channel subunits are not essential (although loss-of-function mutations in Cel-lgc-55 confer subtle behavioral phenotypes [47]), (b) reduced IVM sensitivity in H. contortus is associated with reduction in the transcript abundance of Hco-ggr-3 [48] and (c) a four-amino-acid deletion in the N-terminal region of Cel-glc-1 has been linked to IVM resistance in C. elegans [49].
Although glutamate-gated chloride channels (GluCls) are considered its main targets, IVM may also interact directly with other anionic Cys-loop LGICs, including GABAA and glycine receptors [50, 51] and irreversible activation of these inhibitory chloride channels by IVM results in flaccid paralysis and eventual expulsion of the parasite [52, 53]. Several glutamate and GABA gated chloride channel genes have been implicated in IVM resistance in H. contortus (e.g. Beech et al. 2013) but none of these genes showed significant values for FST in our analysis of T. circumcincta. In addition to genes involved directly or indirectly in neurotransmitter functions, genes putatively responsible for amphid neuron defects in C. elegans and H. contortus, such as Cel-che-3, Cel-dyf-7 and Hco-dyf-7, have been implicated in IVM resistance in these species [54–56]. Again, none of the likely che-3 and dyf-7 orthologs in T. circumcincta displayed notably high FST values in our analysis (S7 Table). Although it is possible that different IVM resistance mechanisms are involved in different nematode species, in a recent study of UK field populations of H. contortus, no evidence of selection by IVM was detected for Hco-lgc-37, Hco-glc-5, Hco-avr-14 or Hco-dyf-7 [57]. Furthermore, a recent analysis of IVM resistant backcross populations in H. contortus has also suggested that these candidate genes were not associated with resistance [58]. It is thus conceivable that some of the putative candidate genes from earlier single-locus studies may represent false-positive associations. If, as we conclude on the basis of the data reported here, multidrug anthelmintic resistance is a polygenic quantitative trait, one explanation for this apparent discrepancy is that in the resistant population that we analyzed these genes are not under selection. The observation that >1 genotype can result in the same phenotype is expected for quantitative traits. This may also, for example, explain why macrocyclic lactone resistance in particular seems so genetically heterogeneous, with many different candidates apparently under selection in different resistant populations.
We also examined copy number variations (CNVs) between the RS3 and Sinbred strains of T. circumcincta (Fig 2C). Since our reference assembly was generated using the Sinbred strain, we focused our analysis on genomic regions displaying increased copy number in the RS3 population relative to the Sinbred population, because any copy number decrease in the RS3 strain was likely to have been confounded with a potential mapping bias against highly divergent reads containing non-reference alleles (especially in the intronic and intergenic regions). The top 10 protein-coding CNV regions showing the most extreme inter-strain variation (log2 read count ratio > 2.9) contained 13 genes, 4 of which are likely orthologs of C. elegans P-glycoprotein 9 (Tci-pgp-9; TELCIR_08885, TELCIR_13813, TELCIR_19247 and TELCIR_19884) (Table 2; Fig 3D). These sequences (one complete and three partial genes) appear to represent the individual haplotypes of Tci-pgp-9 that are segregating in the Sinbred population. While our pooled sequencing data provide strong evidence of an increase in Tci-pgp-9 copy number on average in the resistant population relative to the susceptible population, it remains challenging to reliably resolve the full haplotype sequences and their respective within-population copy number variability for each population. Nevertheless, confirmation that RS3 strain parasites carry additional copies of this gene compared to Sinbred parasites was provided by a separate investigation of Tci-pgp-9 copy number using single worm genomic DNA quantitative PCR (S5 Fig). Furthermore, these single worm data showed that certain Tci-pgp-9 haplotypes (i.e., allelic variants defined on the basis of the first inter-nucleotide binding domain (IBDA) sequence polymorphisms) (S6 Fig) occurred only in worms exhibiting increased Tci-pgp-9 copy number (IBDA haplotypes 3, 6 and 10) (S10 Table), suggesting (a) that these haplotypes arose as a result of the gene duplication event(s) that gave rise to the increase in copy number, (b) that the duplication(s) occurred long enough ago that the duplicated copies have started to diverge and/or (c) selection for or against functional differences in IVM-affinity conferred by specific haplotypes. One further haplotype that appeared to be enriched in the RS3 population, haplotype 2, does not occur in worms that carry additional copies of Tci-pgp-9 (S10 Table), suggesting that selection for this haplotype in resistant worms is not related to increased copy number.
P-glycoprotein (P-gp) is an ATP-binding cassette (ABC) transporter with two homologous halves, each containing a TMD and a cytoplasmic nucleotide-binding domain (NBD). Using the energy from ATP hydrolysis, P-gp actively transports many lipophilic compounds (both endogenous metabolites and xenobiotics) out of the cell from the inner leaflet of the membrane, providing a mechanism by which anthelmintic concentration at the receptor site may be reduced. Sequence polymorphism and constitutive or inducible overexpression of P-gp’s have been reported in IVM-resistant populations of several nematode species, including T. circumcincta, in support of the hypothesis that an increased drug efflux due to changes in expression, activity and/or substrate specificity of ABC transporters can contribute to IVM resistance [59, 60]. Our preliminary assessment (by RNA-seq) indicates that Tci-pgp-9 transcripts are more abundant in the RS3 population relative to the Sinbred population (log2 ratio > 3) (S7 Table), suggesting that the Tci-pgp-9 copy number increase facilitates (constitutive or inducible) increased expression of the transporter in the resistant population. Cloning and sequencing of individual cDNA transcripts corresponding to the N- and C-terminal transmembrane domains (with their extracellular loops) showed that the predominant transcripts in the RS3 population carry either a splice variant that results in a deletion of 45 aa from the first predicted extracellular loop between TM1 and TM2 (a region of the protein hypothesized to play a role in substrate binding), or a full length variant that contains 3 non-synonymous amino acid substitutions in the same predicted loop (S7–S9 Figs). Thus, it appears likely that a combination of increased expression (via increased copy number) and sequence polymorphism may contribute to the association between Tci-pgp-9 and IVM resistance in the RS3 strain.
The design of our introgression strategy, which aimed to identify genomic regions concurrently selected in response to BZ, LEV and IVM, did not allow us to directly assess the relative contribution of candidate loci to resistance against each of the individual anthelmintic classes. We therefore undertook a complementary mapping approach with a specific focus on IVM resistance using F2 populations derived from a cross between the Sinbred susceptible and the Rpar multiple-anthelmintic resistant isolates, i.e., the parental strains of the RS3 backcross progeny population on which our whole-genome introgression study was based (Fig 1A). Individuals from IVM-screened and drug-naïve F2 mapping populations (n = 24 male worms from each group) were genotyped by ddRAD-seq, a reduced-representation genome sequencing method, yielding a total of 0.59 million variant calls. Using the FST estimates of segregating SNPs that satisfied a minimum sampling depth of 10 individuals in both populations (n = 2,628) (S11 Table), we identified outlier loci and linked genes (i.e., contigs) most strongly differentiated in the IVM-survivor group relative to the drug-naïve control group, and compared the outcome against the outlier genes identified from the introgression mapping experiment. Even though there was a high rate of allele dropout (most likely due to sequence polymorphism within the restriction sites in our mapping population leading to the loss of affected restriction fragments from our RAD-seq libraries), we were able to survey part of the genome (349 contigs; combined length = 57.6 Mb or ~8% of the genome) for outlier loci. It has been shown that, in RAD-seq experiments, missing data can inflate FST values and rates of false-positive outliers increase as the chromosome sampling depth cutoff decreases [61]. We indeed observed a strong dependency between the variance of FST estimates and the allele dropout (S10 Fig) and therefore determined outliers in each individual sampling depth category separately under the assumption that outliers were evenly distributed across loci irrespective of missing data (Fig 2D).
Within the part of the genome that was subjected to FST outlier analysis, we identified 18 genes across 5 contigs that displayed some evidence of genetic differentiation in both the introgression and F2 mapping experiments (i.e., minimum FST z-score of 2.5 in both datasets) (S12 Table). Included in this combined list is Tci-lgc-54, one of the top outlier genes from the genomic introgression analysis. Although the evidence of selection is not as strong in the IVM-survivor F2 mapping population as it was in the introgressed multiple-anthelmintic resistant population (RS3), the amine-gated chloride channel Tci-lgc-54 is the only candidate LGIC that is supported strongly by both of the mapping approaches (S13 Table) and thus merits further analysis as a potential IVM resistance gene in T. circumcincta. It is important to note that the individual F2 generation IVM treatment survivors analyzed here were not significantly more resistant to either BZ or LEV treatment as a result of the IVM treatment (see S14 and S15 Tables). Thus the resistance phenotypes appeared to have segregated independently, indicating that Tci-lgc-54 was an FST outlier in IVM-resistant F2 segregants that remained susceptible to BZ and LEV treatment. Consequently, despite strong evidence for selection of ABC transporters, these data do not support a single “multidrug resistance mechanism” able to confer resistance simultaneously to all 3 drug classes.
We were unable to assess whether the Tci-pgp-9 revealed by Introgression analysis as a strong candidate IVM-resistance locus co-segregates with IVM resistance in the F2 mapping population because of the absence in the RAD-seq data of linked SNP makers with adequate sampling depth for FST outlier analysis. A different ABC transporter, Tci-mrp-6 (TELCIR_03131) (Fig 3D), is however present on the combined introgression/segregation list. Our results suggest the interesting possibility that multiple ABC transporters may be involved in IVM efflux, either within the same worm or in different worms, in the RS3 multiple-anthelmintic resistant population. In support of this conclusion, several reports in other parasitic nematodes implicate a range of ABC transporter family members [62–64], implying that there may be many combinations of ABC transporters able to contribute to IVM-resistance.
A novel, strongly-supported, candidate region in our combined outlier list (S12 Table) is on Contig209, which contains two putative triacylglycerol lipase genes (TELCIR_02985 and TELCIR_02988; FST z-score > 3.3). Intriguingly, a triacylglycerol lipase/cholesterol esterase gene (F54F3.3) has been shown in C. elegans to respond transcriptionally to IVM exposure [65], although it remains to be determined whether the lipase activity plays a role in a drug-induced starvation-related stress response that facilitates tolerance of or recovery from ivermectin toxicity, or whether lipid metabolism may play a more direct role in IVM metabolism or detoxification. It is of interest to note that a similar pool sequencing analysis of ivermectin response in Onchocerca volvulus [66] points to the involvement of likely orthologues of these genes in a distantly related nematode parasite of significant medical importance.
The work reported here is the first to combine classical genetic methods such as introgression and segregation analysis with new genomic tools such as RAD-seq and whole genome re-sequencing to analyze multiple-anthelmintic resistance in a parasitic nematode. The nematode species examined, T. circumcincta, is an economically significant, globally distributed gastrointestinal parasite of small ruminants. More importantly however, with the rapid proliferation of mass drug administration programs globally for treatment of helminth infections of humans, there is an urgent need to better understand the genetic basis of resistance to the drugs that form the basis of those programs. We show clearly that resistance to each of the three drug classes segregates independently of the others and that for LEV and IVM resistance in particular, multiple loci likely contribute to the resistance in a variety of ways (possible reduction/modulation in target site sensitivity, reduced target site expression, increased drug efflux, etc.), so that drug resistance in these parasites should best be thought of as a multifactorial quantitative trait rather than a simple, discrete Mendelian character. The polygenic genetic architecture of resistance provides an explanation for the apparent discrepancies between the many single, candidate gene studies and, since many genes can contribute to resistance, it seems likely therefore that the combination selected in any given circumstance is as likely to be a product of genetic drift as of selection per se. Furthermore, it is also clear from these data that alleles that contribute to resistance for each drug class have arisen many times on different genetic backgrounds, giving rise to a heterogeneous mix of “resistance haplotypes” that implies soft rather than hard selective sweeps. This is most obvious at the β-tubulin isotype-1 locus. Although selection at this locus appears to be necessary and sufficient for BZ-resistance, we observed extensive polymorphism surrounding the amino acid 198 and 200 determinants of resistance, suggesting soft selection from multiple pre-existing variants at P198 or P200 rather than hard selection of a single allele at a single position has occurred for this resistance. Similar genetic heterogeneity at this locus has been observed in other BZ-resistant isolates of T. circumcincta and H. contortus [29].
In conclusion, this work elucidates the polygenic, quantitative trait genetic architecture of multiple anthelmintic resistance in a parasitic nematode for the first time and establishes a framework on which future studies of the inevitable evolution of anthelmintic resistance in nematode parasites of humans can be based. In this context, it is significant that a similar study of ivermectin response in O. volvulus [66] points to a similar genetic architecture, with hits either to likely orthologues of genes identified here or to similar neuronal functions, thus demonstrating the utility of studies in more tractable parasite species such as T. circumcincta.
All experimental procedures used in generating the parasite material for this study were approved by AgResearch’s Wallaceville Animal Research Centre Animal Ethics Committee under the Animal Welfare Act 1999 in New Zealand [AEC application numbers 516, 562 & 636].
The multiple-anthelmintic resistant field strain of T. circumcincta (OR strain) used in this study was isolated in New Zealand in 1996 from lambs which had been grazing a property previously occupied by Angora goats (Leathwick DM, personal communication). Fecal nematode egg count reduction tests undertaken on the lambs revealed that none of the three broad-spectrum anthelmintic families available at that time, i.e., oxfendazole (BZ), levamisole (LEV) and ivermectin (IVM), were fully effective against this isolate. Prior to inter-strain crosses being set up, a population of OR was maintained for five generations in pen-raised goat kids, and screened at each generation with selected representatives of BZ (Systamex; Schering Plough, Kenilworth, NJ; 4.5 mg/kg), LEV (Levicare; Ancare New Zealand, Auckland, New Zealand; 7.5 mg/kg) and IVM (Ivomec liquid for sheep and goats, Merial New Zealand; 0.2 mg/kg) to maximize the proportion of worms homozygous for resistance to each of them. Anthelmintics were administered to the goat kids at the manufacturer’s recommended dose rate unless no specific goat dose rate was provided, in which case a standard sheep dose rate was used (as was common practice on goat farms in New Zealand before the widespread emergence of anthelmintic resistance in this host species). The efficacy of all these drugs, used either individually or in combination, was very low against the resulting resistant parental (Rpar) strain (Fig 1B). The anthelmintic susceptible S strain of T. circumcincta was originally isolated from field-grazed lambs in New Zealand during the 1950’s prior to the widespread use of broad-spectrum anthelmintics (Elliott DC, personal communication). This isolate had subsequently been maintained at Wallaceville Animal Research Centre (AgResearch, New Zealand) by annual passage through pen-raised drug-naïve lambs. Given that significant genetic diversity can be maintained even in laboratory-passaged nematode populations of limited size, the S isolate was subjected to two generations of half-sib mating in an attempt to reduce the background genetic variance in preparation for introgression mapping. Briefly, mature eggs were collected from the oviduct of a single gravid adult female recovered from the host’s abomasum, and cultured to collect infective larvae. Thirteen of these sibling larvae were used to orally infect a pen-raised parasite-free goat kid, and half of the resulting progeny were used to re-infect the same host to supplement the existing infection. A second goat kid was subsequently infected using larvae cultured from the first. Sibling mating was repeated as before by isolating and culturing eggs from a single adult female worm isolated from the abomasum of the second kid. Twenty sibling larvae from this culture were used to infect a third worm-free goat kid whose fecal output was cultured for subsequent infections. Anthelmintic efficacy testing on this partially inbred S strain (Sinbred) revealed that representatives of all three broad-spectrum anthelmintic classes were highly effective (efficacy >99.0% for BZ, LEV, and IVM) [15].
A schematic of the backcrossing and selection experiment is outlined in Fig 1A. Crosses between Rpar and Sinbred strains of T. circumcincta were performed by surgical transfer of worms from separate donor goat kids (containing either Rpar or Sinbred worms) into the abomasum of a recipient goat kid. In order to ensure that the female Sinbred worms had not yet mated, infection of the donor kids was timed so that the females would be 10 days old and thus still at the late-fourth developmental stage (L4) at the time of transfer, while male Rpar worms would be 5 weeks old and thus adults. Ten days before transfer, the Rpar worms were screened with BZ, LEV and IVM. Worms were collected from donor goat abomasa, rinsed with phosphate-buffered saline over a 45μm Endecott sieve, and inspected microscopically. Approximately 300 male adult male (Rpar) and an equivalent number of L4 female (Sinbred) worms were collected into a modified Nematode Growth Medium containing 0.4% w/v agar [67], and surgically transferred into the abomasum of a previously worm-free recipient goat kid. The [heterozygous] F1 progeny resulting from this cross were then used to infect another worm-free kid to obtain an F2 generation in which the alleles for anthelmintic resistance were expected to have segregated. The F2 infective larvae cultured from this kid were used to infect a further worm-free goat kid, which was then subjected to successive doses of IVM, BZ and LEV over a period of 24 hrs at 28 days post-infection so that the F3 generation would be derived from worms carrying the full complement of genes needed for multiple-anthelmintic resistance. In order to maximize the frequency of resistance alleles in the population, F3 worms were passaged for a further two generations of drug screening with IVM, BZ and LEV. At this point a backcross between the anthelmintic-screened F5 generation and Sinbred worms was set up using similar procedures to those described for the initial crosses. F2, F3 and F4 generations of the backcross worms were then each screened, as before, with all 3 anthelmintic classes before a final round of backcrossing and a further 4 generations of drug screening. Because each generation of backcrossing reduces the proportion of the donor parent genome present in the population by half, the resultant multiple-anthelmintic resistant worm population (RS3) was expected to have a genetic makeup largely similar (7/8) to that of the susceptible recurrent parent (Sinbred) but, at the same time, be carrying the anthelmintic-resistance genes derived from the Rpar strain. Crosses based on mass mating and multiple generations of drug selection (between and after backcrosses) were an important design feature to create variation in recombination breakpoints and divide individual introgression blocks (i.e., segments of DNA of Rpar origin untouched by recombination) into smaller fragments.
Parasite-free lambs, maintained indoors on a diet designed to avoid any unintended nematode infections, were each infected with approximately 24,000 larvae of either the RS3 strain (2 lambs) or the Sinbred strain (3 lambs) of T. circumcincta. At 28 days post-infection, lambs that had received the RS3 strain were treated successively with IVM, BZ and LEV over a period of 24 hrs. At 37 days post-infection (9 days post-treatment), the RS3 worms were collected from the abomasum, washed free of all debris in physiological saline, and then transferred in mixed sex batches of 300–500 individuals into 1.5 ml tubes in which they were snap frozen at -80°C. The Sinbred worms were similarly collected and snap frozen at 28 days post-infection without anthelmintic treatment. Genomic DNA was isolated from each of the strains using a method modified from that described by Sulston and Hodgkin [68]. Each worm sample (after thawing) was suspended in 150–200 μl of lysis solution [100 mM NaCl, 100 mM Tris-HCl (pH 8.5), 50 mM EDTA (pH 7.5), 1% SDS, 2% β-mercaptoethanol and 200 μg/ml proteinase K], placed in a glass/teflon tissue grinder and thoroughly homogenized. The homogenate was transferred into a sterile 10 ml tube and supplemented with additional lysis solution to a total volume of 3 ml. After incubation at 65°C for 16 hr with gentle mixing, 10 μl of RNase A (100 mg/ml; QIAgen, Hilden, Germany) was added and the lysate incubated for 10 min at 45°C. To remove protein/polysaccharide complexes (which can be problematic in nematode DNA preparations), 750 μl 5M NaCl and 500 μl CTAB/NaCl (10% CTAB in 0.7M NaCl) were added, and after gentle mixing the tube was incubated for 15 min at 65°C. The lysate was then extracted successively with equal volumes of phenol/chloroform/isoamyl alcohol (25/24/1) and chloroform/isoamyl alcohol (24/1). Following recovery of the aqueous layer from the final extraction, DNA was precipitated by the addition of 2 volumes of absolute ethanol (4°C), pelleted by centrifugation, washed twice in 70% ethanol (4°C), briefly air-dried, and re-suspended in 150 μl 10mM TrisCl (pH 8.5). Total RNA was also prepared from the mixed-sex batches of adult T. circumcincta from each of the Sinbred and RS3 strains. Frozen worm samples were added to 1 ml of TRIzol reagent (Invitrogen, Carlsbad, CA) and the mixture ground to a powder in liquid nitrogen in a mortar and pestle. The resulting powder was transferred into a microfuge tube, 200 μl chloroform was added and the tube was shaken vigorously before being centrifuged at 12,000g for 15 min at 4°C. Following recovery of the upper phase, the RNA was precipitated by the addition of 500 μl of isopropanol and pelleted by centrifugation at 12,000g for 10 min at 4°C. The RNA pellet was then washed in 75% ethanol and air-dried briefly before re-suspending in 40 μl UltraPure DNase/RNase Free Distilled Water (Invitrogen). The integrity and yield of the RNA was verified using the Bioanalyzer 2100 (Agilent Technologies, Cedar Creek, TX).
As a basis for introgression mapping and comprehensive variant analysis, we generated a draft genome sequence for the anthelmintic-susceptible Sinbred strain of T. circumcincta (BioProject ID: PRJNA72569). Whole genome shotgun libraries (fragments and mean insert size of 3kb and 8kb) were generated as previously described [69] and sequenced using a Genome Sequencer Titanium FLX (Roche Diagnostics, Basel, Switzerland) platform (S16 Table), and assembled using Newbler v. 2.6 [70]. To improve scaffolding, an in-house tool CIGA (Cdna tool for Improving Genome Assembly) was used to map 454 cDNA reads using BLAT [71] to the genomic assembly to link genomic contigs. Gaps were then closed using Pygap, an in-house tool, which utilizes the Pyramid assembler and uses Illumina paired-end reads to close gaps and extend contigs. The repeat library was generated using Repeatmodeler (http://repeatmasker.org), and Tandem Repeat Finder [72] was used in addition for sequence annotation. Repeats and predicted RNAs were then masked using RepeatMasker (http://repeatmasker.org). The ribosomal RNA genes were identified using RNAmmer [73] and transfer RNAs were identified with tRNAscan-SE [74]. Non-coding RNAs, such as microRNAs, were identified by sequence homology search of the Rfam database [75]. Protein-coding genes were predicted using a combination of ab initio programs Snap [76], Fgenesh [77] and Augustus [78] and the annotation pipeline tool Maker [79] which aligns mRNA, EST and protein information from the same species or cross-species to aid in gene structure determination and modifications. A consensus gene set from the above prediction algorithms was generated, using a previously described, logical, hierarchical approach [69]. In summary, the following Quality Index (QI) criteria were calculated: i) length of the 5’ UTR; ii) fraction of splice sites confirmed by an EST alignment; iii) fraction of exons that overlap an EST alignment; iv) fraction of exons that overlap EST or protein alignments; v) fraction of splice sites confirmed by a SNAP prediction; vi) fraction of exons that overlap a SNAP prediction; vii) number of exons in the mRNA; viii) length of the 3’ UTR and; ix) length of the protein sequence produced by the mRNA, and then the following decision making steps were followed: a) genes were screened for overlaps (<10% overlap was allowed); b) if QI[ii] and QI[iii] were greater than 0, or QI[iv] was greater than 0, then the gene was kept; c) the gene was BLASTed against Swissprot [80] (E < 1e-6). If there was similarity to a Swissprot entry, then the gene was kept; d) RPSBLAST was ran against Pfam [81] (E < 1e-3). If there was similarity to a Pfam entry, then the gene was kept; e) RPSBLAST was run against CDD [82] (E < 1e-3 and coverage > 40%). Genes that met both cut-offs were kept and f) if no hit was recorded, then a sequence similarity-based search was run against GenesDB from KEGG [83], and genes with at least a 55% identity and a bit score of 35 or higher were kept. Genes of interest (discussed in the paper) underwent manual evaluation and improvements. Gene product naming was determined by BER (http://ber.sourceforge.net). Functional domains and Gene Ontology (GO) terms were assigned using InterProScan (v. 4.8) [84]. Genes were mapped to KEGG Orthology (KO) groups using wu-blastp (E < 1e-5). Proteins with signal peptides and transmembrane topology were identified using the Phobius web server [85], and non-classical secretion was predicted using SecretomeP 1.0 [86]. CEGMA (v.2.4) [18] was used to assess the completeness of the genome without excluding partial matches. MEGA6.06 [87] was used to estimate maximum likelihood phylogenies (LG+G model).
The same RNA samples were used to generate both Roche/454 and Illumina cDNA libraries. Both data types were used for genome annotation and the Illumina reads were used for differential expression analysis. Non-normalized oligo dT libraries for Roche/454 were generated as previously described [88]. The Roche/454 library was sequenced using a Genome Sequencer Titanium FLX (Roche Diagnostics) and the ‘sffinfo’ program was used to extract information from the SFF files. Adaptor sequences were trimmed from the sequenced reads using the 'seqclean’ software and host and bacterial contamination was removed using Newbler’s ‘gsmapper’. For the Illumina RNA-seq library construction, total RNA was treated with Ambion Turbo DNase (Ambion/Applied Biosystems, Austin, TX) and 1μg of the DNAse treated total RNA was poly(A) selected using the MicroPoly(A) Purist Kit according to the manufacturer's recommendations (Ambion/Applied Biosystems). One ng of the mRNA isolated was used as the template for cDNA library construction using the Ovation RNA-seq (v.2) kit according to the manufacturer's recommendations (NuGEN Technologies, San Carlos, CA). Non-normalized cDNA was used to construct multiplexed Illumina paired-end small fragment libraries as previously described [69], according to the manufacturer's recommendations (Illumina, San Diego, CA) with the following exceptions. In summary, 500 ng of cDNA was sheared using a Covaris S220 DNA Sonicator (Covaris, Woburn, MA) to a size range of 200-400bp. Four PCR reactions were amplified to enrich for adaptor ligated fragments and index the libraries. The final size selection of the library was achieved by an AMPure paramagnetic bead (Agencourt, Beckman Coulter Genomics, Beverly, MA) cleanup, targeting 300-500bp. To produce cluster counts appropriate for the Illumina sequencing, the concentration of the library was determined by qPCR according to the manufacturer's protocol (Kapa Biosystems, Woburn, MA). The Illumina HiSeq 2000 platform was used for generation of sequences of 100bp from samples of pooled individuals of 300–500 RS3 or Sinbred adult worms (S16 Table). RNA-seq reads were aligned to the Sinbred reference assembly of T. circumcincta using STAR aligner (v.2.3.0) [89] with default parameters, following the 2-pass method [90]. Transcript abundance levels were expressed in FPKM (fragments per kilobase of exon per million fragments mapped). To infer rankings of differentially expressed genes according to their effect size, GFOLD (generalized fold change) algorithm was used with default parameters [91]. Although GFOLD provides a more consistent and biologically meaningful approach to ranking differentially expressed genes than other methods for single replicate RNA-seq experiments, an analysis of variation between replicate samples is necessary to draw sound conclusions, especially on the individual gene level. The differential expression patterns observed must be considered as only preliminary because of the study design limitations and confounding factors that called for cautious interpretation of the results. For example, the RS3 worms were collected for RNA extraction at 9 days post drug treatment, resulting in a temporal separation between the stage of drug effect (and selection) and the stage at which gene expression was measured. In addition, potential sex-ratio heterogeneity among samples (that consisted of 300–500 mixed-sex adult worms) while expected to be close to 50/50 was not fully controlled for, which could result in expression variation in gender associated genes.
Genomic DNA samples from pooled individuals of 300–500 RS3 or Sinbred worms were subjected to Illumina GAII, GAIIx, HiSeq 2000, HiSeq 2500 and MiSeq paired-end sequencing (S16 Table). Sequencing adapters were removed using trimmomatic (v.0.33) [92], and the resulting reads were aligned to the Sinbred reference assembly of T. circumcincta using BWA-MEM (v.0.7.12) [93]. Picard (v.1.95) was used to remove duplicate reads and local re-alignments were performed around indels using GATK (v. 3.4–46) [94]. Variants were called from the RS3 and Sinbred population pools using GATK HaplotypeCaller (v. 3.4–46) [94]. In line with the developers’ recommendations for analyzing pooled DNA samples (as opposed to diploid individuals),—sample_ploidy parameter was set to 10. In addition to requiring a minimum mapping quality score of 20 and a minimum base quality score of 20 in the reference alignment, the following set of quality filters were applied to SNP calls using GATK VariantFiltration (v. 3.4–46) [94]: DP (maximum depth) > median depth+(median absolute deviation×1.4826)×3; QD (variant confidence divided by the unfiltered depth of non-reference samples) < 2.0; FS (Phred-scaled p-value using Fisher’s Exact Test to detect strand bias in the reads) > 60.0; MQ (Root Mean Square of the mapping quality of the reads across all samples) < 40.0; MQRankSum (Mann-Whitney Rank Sum Test for mapping qualities) < -12.5; ReadPosRankSum (Mann-Whitney Rank Sum Test for the distance from the end of the read for reads with the alternative allele) < -8.0. Population allele frequency was estimated based on the relative abundance of reads supporting each allele (i.e., allelic depth), and Fisher’s exact test was used to assess the statistical significance of allele frequency differences between the populations. To delineate the introgressed loci in the RS3 strain, we conducted a genome-wide scan of fixation index (FST) using nucleotide frequencies at polymorphic sites, and identified genomic regions that were most divergent relative to the parental Sinbred genetic background. Our mapping approach was based on the expectation that, after serial backcrossing and drug screening, causal and closely linked SNPs in the RS3 strain would retain the allelic profile of the anthelmintic resistant parental Rpar isolate, whereas the genomic regions not associated with drug resistance would be represented by either the Sinbred genotype or a mixture of the two parental genotypes (depending on the variation in the recombination breakpoints among individuals). A SNP site was included in the FST analysis if it was supported by at least two alternative reads, did not overlap with indels, and had a minimum depth of 25× coverage in both populations. Following methods described by Kofler et al. [95] and assuming a pool size of 300 individuals in each population, FST values were estimated per site and averaged over non-overlapping 1-kb sliding windows. In addition, the number of loci meeting the depth of coverage threshold (25×) was examined for each window, and those windows with covered fraction > 0.5 (total combined length = 325.7 Mb) were included in the analysis. Position-sorted mean FST values (for each 1-kb window) were scanned for peaks after applying a kernel smoothing algorithm with adaptive bandwidth selection using the lokern package in R [96, 97] to identify blocks of genomic regions with extended linkage disequilibrium and elevated FST, while reducing the effects of sequencing error, mapping artifacts, and base-to-base variation in coverage. Outlier windows were identified based on the empirical distribution of the smoothed FST values. A z-score of 4.5 was chosen as the cutoff threshold, guided by the z-score exhibited by the β-tubulin gene (TELCIR_01271; z-score = 4.66), a widely recognized BZ-resistance conferring locus in trichostrongylid nematodes. We reasoned that β-tubulin isotype 1 could effectively serve as a “positive control” and that loci with FST values similar to or more extreme than that for β-tubulin gene represented candidate loci worthy of further investigation. Statistical enrichment of GO terms among the genes overlapping outlier regions was assessed using a conditional hypergeometric test implemented in GOstat [98]. Gene-wise FST values per protein-coding loci were calculated as the maximum smoothed FST value among 1-kb windows that overlap the gene footprint (exon + intron). Using SnpEff (v.3.5), variants were annotated on the basis of their genomic location (e.g., exon, intron, intergenic, upstream/downstream (5 kb flanking regions) or splice site donor/acceptor), and their mutational effects were predicted (e.g., missense, nonsense or silent). Segregating haplotypes were reconstructed using either a phasing approach supported by sequencing reads [99, 100], or a de novo assembly-based method [101]. DNA copy number variation (CNV) was examined using CNV-seq [102] with p-value parameter set to 0.00001. Two-dimensional allele frequency spectra for RS3 and Sinbred populations were produced after alleles were subsampled without replacement to a uniform coverage of 10×. Only bi-allelic sites with a minimum coverage of 10× in both populations were included in the analysis, while SNPs with high depth of coverage (within the top 5% of the empirical distribution) were excluded. In total, 14,562,483 SNPs were used to obtain frequency spectra. Based on these variants, Tajima's D statistic was computed to assess whether allele frequency distributions deviated from neutral expectations [103].
A segregating F2 mapping population of T. circumcincta was generated as part of the work undertaken to introgress anthelmintic resistance genes into an anthelmintic susceptible genetic background. A cross was initiated by surgically transferring ~300 anthelmintic resistant adult males (Rpar) and an equivalent number of susceptible late L4 stage females (Sinbred) into the abomasum of a previously worm-free kid goat. F1 eggs collected from the host feces were then cultured and ~10,000 of the resulting infective larvae were used to infect a second worm-free kid goat to produce an F2 generation. Two experiments were carried out to investigate the segregation of resistance to each class of anthelmintic, one using kid goats as the recipients and a second using lambs, with essentially the same results. We describe the lamb experiment here (S14 Table). Two groups of worm-free lambs (n = 27 per group) were administered an oral dose of ~8,000 (Group 1) or 16,000 (Group 2) of the F2 generation infective larvae each. On Day 27 post-infection, the lambs in Group 2 were treated with IVM. Those in Group 1 remained untreated. Three days later (Day 30 post-infection) individual fecal nematode egg counts (FECs) were undertaken on all animals and, using a restricted randomization procedure based on weight and FEC, each of the infection groups was subdivided into 3 equal-sized anthelmintic treatment groups (1a-1c and 2a-2c). On Day 31 post-infection, anthelmintic treatments were administered as follows: Groups 1b and 2b received BZ while Groups 1c and 2c received LEV. Groups 1a and 2a remained untreated as controls. Anthelmintic doses were calculated on the basis of individual live-weights and each dose was administered orally with a disposable syringe. For ddRAD-seq analysis, a total of 24 adult male F2 IVM-treatment survivors from the parallel, kid goat experiment (Group 2a) and 24 drug-naïve adult male (Group 1a) F2 worms also from the kid goat experiment were individually transferred into 100 μl of DirectPCR Lysis Reagent (Mouse Tail; Viagen Biotech, Los Angeles, CA), supplemented with 3% proteinase K (10 mg/ml; Roche) and incubated at 55°C for 16 h followed by 90°C for 1 h to denature the proteinase K. A 20 μl volume from each worm lysate (~10 ng DNA) was digested with EcoRI and MspI overnight, after which sequencing adapters (P1-EcoRI-inline-barcode and P2-MspI) were ligated to the fragment termini. The reaction was purified using a 0.5X and 0.7X double size selection (modified from Lennon et al. [104]) using Agencourt AmpureXP beads (Beckman Coulter, Brea, CA), and PCR amplified to incorporate index sequences for multiplexing using KAPA HiFi Real Time master mix using the following protocol; 98°C for 2 min, followed by 14 cycles of 98°C for 15 sec, 60°C for 30 sec and 72°C for 30 sec. PCR reactions were purified using AmpureXP beads, after which the DNA concentrations were standardized using a Qubit 3.0 Fluorometer (Life Technologies, Carlsbad, CA) and samples pooled at equimolar concentration. Adapter-ligated and PCR amplified fragments approximately 500-600bp in length were obtained by gel size selection and purification. The ddRADseq library, supplemented with a 10% PhiX spike-in control, was sequenced using an Illumina MiSeq (reagent kit v3), resulting in 150bp single-end sequencing reads. Sequencing data were demultiplexed using process_radtags [105], and were mapped to the T. circumcincta reference assembly using BWA-MEM (v0.7.10) [93]. Local realignments were performed around indels using the GATK (v3.3–0), after which variants were called by HaplotypeCaller under default parameters. FST estimation [106] was carried out using VCFtools (v.0.1.12b) [107].
An allele-specific multiplex PCR strategy based on that developed by Humbert and Elard [108] was used in a 96-reaction format to assess the presence of a Phe (TTC)/Tyr (TAC) substitution at codon 200 in the β-tubulin isotype-1 gene in individual worms from the anthelmintic susceptible (Sinbred) and multiple-anthelmintic resistant (RS3) strains of T. circumcincta. Only adult male worms were used for these allele-specific reactions to avoid the possibility of DNA from sperm and/or fertilised eggs present in female worms interfering with the genotype identifications. The strategy involved the use of four primers per reaction, two of which–one forward and one reverse–were generic (allele-nonspecific), while the remaining two–again one forward and one reverse–were allele-specific (S2 Fig). Primer designs, which differed slightly from those of Humbert and Elard [108] to account for minor DNA sequence differences between the strains studied by them and those used in the present study, were as follows: generic forward [“TubGF”] 5′ CTTAGATGTTGTTCGTAAAGAGG 3′; generic reverse [“TubGR”] 5′ CATGTTCACAGCCAACTTGC 3′; Phe-specific [“TubSASRev”] 5′ AGAGCTTCATTATCGATGCAGA 3′; Tyr-specific [“TubRASFwd”] 5′ TGGTWGAAAAYACCGATGAAACRTA 3′. Note that TubRASFwd was degenerate at three nucleotide positions (5, 11 and 23) in order to accommodate the presence of SNPs in those positions in some haplotypes containing a Tyr at codon 200 (see S3 Fig). Two further primers–[“TubRASH3Rev”] [5′ CTTCATTATCGATGCAGAATGTTAA 3′] and [“TubSASH1Fwd”] [5′ CAGTTGGTTGAAAATACCGATGA 3′]–were designed to detect the presence or absence of a Glu198Leu substitution.
Adult worms from each of the above T. circumcincta strains were isolated from experimentally infected goats (approximately one week after successive treatments with all three anthelmintic families in the case of RS3), washed free of all debris in physiological saline and then transferred, in batches of either 100 male worms or ~300 mixed-sex worms, into cryovials where they were frozen in liquid nitrogen until required. Crude genomic DNA template was prepared from individual adult male T. circumcincta from the Sinbred and RS3 strains (96/strain) by overnight incubation in lysis solution [Viagen DirectPCR (MouseTail), 50 μl per worm with 3 mg/ml ProteinaseK] without further purification. SYBR®Green real-time PCR assays were performed in a GeneAmp 5700 sequence detection system to compare Tci-pgp-9 gene copy number in individual male worms from the Sinbred and RS3 worm populations. Primers constructed for this purpose corresponded to genomic DNA sequence within the first putative inter-nucleotide binding domain of Tci-pgp-9 (i.e., Tci-pgp-9-IBDA) and were designed to amplify an equivalent 99bp product from each of the seven Tci-pgp-9-IBDA haplotypes identified from the Sinbred and/or RS3 strain worms. Primer sequences were as follows: forward [“IBD77RTGF”] 5′ CGHTATGGACGTGAAAAAGTCACAGA 3′ and reverse [“IBD77RTGR”] 5′ CCAACTCACGTCRGGGAAYGACTG 3′. Although designs of both these primers were based on well conserved Tci-pgp-9 IBDA haplotypes (see S6 Fig) it was necessary to incorporate some degeneracy to ensure perfect matches in all cases. To take variation in the concentration of genomic DNA between single worm DNA preparations, Tci-pgp-9 copy number was calculated in reference to a single copy gene. The single copy reference in this case was T. circumcincta β-tubulin isotype-1 using primers forward [“TUBRTGF2”] 5′ GGGCTTCCAACTGACGCATTCTTTG 3′ and reverse [“TUBRTGR2’] 5′ GGGCTTCCAACTGACGCATTCTTTG 3′ which amplified a 122bp product from an exon in the central region of the gene. The Tm for both primers was within the same range as those of IBD77RTGF and IBD77RTGR. All reactions were performed in duplicate in 96-well optical reaction plates (Applied Biosystems) using 25 μl reaction volumes which contained SYBR®Green PCR mastermix (Applied Biosystems), 0.2 μM of each gene-specific primer and 1 μl of 10-fold diluted crude genomic template. For both the target and reference genes “no-template controls” were included on each plate. Amplification conditions for the above reactions were as follows: initial incubation at 50°C for 2 min, followed by 95°C for 10 min to denature the template, followed by 40 cycles of 95°C for 15 sec and 60°C for 1 min. Following the reactions a melting curve and cycle threshold (CT) value were generated for each sample. The CT value indicates the fractional cycle number at which the amount of amplified DNA reaches a fixed threshold. Mean CT values of duplicate samples were used in subsequent quantification analyses. No product was amplified in the “no-template control” reactions. As indicated above the amount of target measured in each case was standardised in relation to an endogenous reference gene to account for any between-worm variation in the total amount of gDNA template available. This was done by calculating ΔCT values for each sample–the ΔCT value indicates the difference in cycle number required to reach the fixed threshold for the target and reference genes. Two-sample t-tests were used to compare ΔCT values for worms from the Sinbred and RS3 populations, as well as specific genotype groups within the RS3 population.
Tci-pgp-9-IBDA haplotypes were identified based on PCR clones (n = 66) amplified from gDNA preparations (see S6 Fig). “Allele-specific” primers were designed to differentiate between each of the haplotypes using a nested PCR strategy to allow genotyping of individual male worms from the Sinbred and RS3 strains of T. circumcincta. Primary reactions were performed using the degenerate primers based on the following amino acid sequences: VEIDKINIE (sense) [“IBD77GF3”] and GTQMSGGQ (antisense) [“IBD77GR2”]. Reactions were carried out in a Mastercycler thermal cycler (96 well block) using 20 μl reaction volumes containing 0.5 unit Platinum Taq polymerase (Invitrogen), 2 μl 10x Taq buffer, 2.5 mM MgCl2, 200 μM each dNTP, 20 pmol of each primer and 2 μl template. A touchdown protocol was used as follows: 95°C for 8 min to denature the template and activate the enzyme, followed by 12 cycles of 94°C for 30 sec, 58°C (–0.5°C/cycle) for 30 sec and 72°C for 1 min, followed by 28 cycles of 94°C for 30 sec, 52°C for 30 sec and 72°C for 1 min, and finishing with a final elongation step of 72°C for 7 min. A single generic (“allele-nonspecific”) forward primer–“IBD77GF5” [5′ GAGTAGTKTCACARGARCCNATGCT 3′]–was used for all subsequent nested allele-specific reactions. This primer was degenerate at four nucleotide positions (8, 14, 17 and 20) to accommodate the presence of SNPs at those positions in some haplotypes. The allele-specific reverse primers used in combination with IBD77GF5 to assess worm genotypes are shown in S17 Table. Allele-specific reactions were similarly performed in a Mastercycler using 20μl reaction volumes but unlike the first round reactions they contained 1.0 mM MgCl2 and 10 pmol of each primer, and amplification conditions used were more stringent than for the first round reactions, i.e., 95°C for 8 min, followed by 35 cycles of 94°C for 30 sec, 60–61°C for 30 sec and 72°C for 1 min, finishing with a final elongation step of 72°C for 7 min. Five microliters of each reaction were run on a 2% agarose gel in the presence of ethidium bromide to assess the incidence of each of the sequence variants in each of the worm populations. Tci-pgp-9- IBDA genotype information derived from the allele-specific reactions was checked and verified in selected worms from each population by sequencing PCR fragments amplified from these worms using a nested protocol similar to that used for the allele-specific reactions.
Total RNA was isolated from mixed-sex batches of adult worms from each of the Sinbred and RS3 strains using TRI REAGENT LS (Molecular Research Center, Cincinnati, OH). Synthesis of first-strand complementary DNA (cDNA) was carried out using SuperScript II Reverse Transcriptase (Invitrogen) and poly(A) oligo(dT)12-18 primer (Invitrogen) as per the manufacturer’s instructions. The resulting cDNA solution was diluted with DEPC-treated water to equate to an initial RNA concentration of 20 ng/μl before being stored at -20°C until required for subsequent PCR. Overlapping fragments, encoding the complete transmembrane region from each half of the T. circumcincta PGP-9 protein molecule, were amplified from first-strand cDNA derived from Sinbred (two separate pools) and RS3 worms using degenerate primers in nested or partially nested PCRs. Primer designs were based on the deduced amino acid sequence of Tci-PGP-9, and corresponded to: N-terminal transmembrane region, fragment 1, first round reactions–DAILVCFQ (sense) [“PGP9AF”]/ MIICGAFI (antisense) [“PGP9AR”]; nested reactions–VCFQFRYT (sense) [“PGP9AFnest”]/ APFMIICG (antisense) [“PGP9ARnest”]; N-terminal transmembrane region, fragment 2, first round reactions–GGFIVAFT (sense) [“PGP9BF”]/ YNPADGKI (antisense) [“PGP9BR”]; nested reactions–IVAFTYDW (sense) [“PGP9BFnest”]/ GCGKSTII (antisense) [“PGP9BRnest”]; C-terminal transmembrane region, fragment 1, first round reactions–VTEDTGVA (sense) [“PGP9CF”]/ QAIQMKFM (antisense) [“PGP9CR”]; nested reactions–ATAQNDP (sense) [“PGP9CFnest”]/ PGP9CR; C-terminal transmembrane region, fragment 2, first round reactions–IALYFGW (sense) [“PGP9DF”]/ GCGKSTVI (antisense) [“PGP9DR”]; nested reactions–LYFGWQMA (sense) [“PGP9DFnest”]/ VGPSGCG (antisense) [“PGP9DRnest”]. Approximate locations of these primer sites in relation to each other and to the expected transmembrane structures and the ATP sites of the PGP-9 protein molecule are depicted in S7 Fig. Although appropriate products were amplified in each case using the above primer combinations, it subsequently became apparent that there were errors in the design of the sense primers PGP9AF and PGP9AFnest. These in fact should have corresponded to VPKASIGQ and IGQLFRYT respectively as indicated in S9 Fig. All PCR amplifications were performed in an MJ Research PTC-200 thermal cycler using final volumes of 20 μl containing 0.5 unit Platinum Taq polymerase (Invitrogen), 2 μl Taq buffer, 2.5 mM MgCl2, 200 μM each dNTP, 20 pmol of each primer and 2 μl cDNA template. Both first round and nested reactions were undertaken using a touchdown PCR procedure as follows: 95°C for 5 min to denature the template, followed by 12 cycles of 94°C for 15 sec, 58°C (–0.5°C/cycle) for 30 sec and 72°C for 1 min, followed by 28 cycles of 94°C for 15 sec, 52°C for 30 sec and 72°C for 1 min, and finishing with a final elongation step of 72°C for 7 min. PCR products (see S8 Fig) were ligated into a TOPO TA Cloning vector (Invitrogen) and multiple clones sequenced for each product.
|
10.1371/journal.pgen.1000453 | The Impact of Local Genome Sequence on Defining Heterochromatin Domains | Characterizing how genomic sequence interacts with trans-acting regulatory factors to implement a program of gene expression in eukaryotic organisms is critical to understanding genome function. One means by which patterns of gene expression are achieved is through the differential packaging of DNA into distinct types of chromatin. While chromatin state exerts a major influence on gene expression, the extent to which cis-acting DNA sequences contribute to the specification of chromatin state remains incompletely understood. To address this, we have used a fission yeast sequence element (L5), known to be sufficient to nucleate heterochromatin, to establish de novo heterochromatin domains in the Schizosaccharomyces pombe genome. The resulting heterochromatin domains were queried for the presence of H3K9 di-methylation and Swi6p, both hallmarks of heterochromatin, and for levels of gene expression. We describe a major effect of genomic sequences in determining the size and extent of such de novo heterochromatin domains. Heterochromatin spreading is antagonized by the presence of genes, in a manner that can occur independent of strength of transcription. Increasing the dosage of Swi6p results in increased heterochromatin proximal to the L5 element, but does not result in an expansion of the heterochromatin domain, suggesting that in this context genomic effects are dominant over trans effects. Finally, we show that the ratio of Swi6p to H3K9 di-methylation is sequence-dependent and correlates with the extent of gene repression. Taken together, these data demonstrate that the sequence content of a genomic region plays a significant role in shaping its response to encroaching heterochromatin and suggest a role of DNA sequence in specifying chromatin state.
| Epigenetic packaging of DNA sequence into chromatin is a major force in shaping the function of complex genomes. Different types of chromatin have distinct effects on gene expression, and thus chromatin state imparts distinct features on the associated genomic DNA. Our study focuses on the transition between two opposing chromatin states: euchromatin, which generally correlates with gene expression, and heterochromatin, which is typically refractive to gene expression. While heterochromatin is capable of spreading into euchromatic domains, the parameters that influence such spreading are unknown. We established heterochromatin at ectopic sites in the genome and evaluated whether specific DNA sequences affected the extent of heterochromatin spreading and the transition between heterochromatin and euchromatin. We found that the nature of the genomic DNA neighboring the heterochromatic sequence dramatically affected the extent of heterochromatin spreading. In particular, the presence of genes antagonized the spread of heterochromatin, whereas neutral sequence elements were incorporated into the domain. This study demonstrates that genome sequence and chromatin identity are inextricably linked; features of both interact to determine the structural and functional fate of underlying DNA sequences.
| Correct patterns of gene expression are established by orchestrated interactions among cis-regulatory elements, trans-acting factors and the surrounding chromatin environment. How these interactions are coordinated and to what extent genomic sequence serves as a blueprint, directing these interactions towards normal growth and development, remain major questions in genome biology.
Chromatin has classically been divided into two functionally distinct types: heterochromatin and euchromatin. Genes inserted within, or proximal to, major heterochromatin domains can exhibit either variegated or complete silencing [1]–[7]. This repression, referred to as position effect variegation (PEV), results from the propagation of heterochromatin marks along the chromosome, placing the euchromatic gene into a chromatin context that is incompatible with normal gene expression [2],[8],[9]. While PEV and the factors that contribute to it have been most thoroughly elucidated in yeast and flies, position-dependent gene silencing has been observed in a range of organisms including both mice and humans [1]–[3],[10]. Indeed, there are examples of human disease that can be attributed to gene silencing associated with aberrant formation of heterochromatin [11]–[14]. Together, these studies highlight the important relationship between chromatin context and gene expression and suggest that eukaryotes have developed mechanisms to counter the spread of repressive heterochromatin [2],[8],[15],[16]. However, the nature of these mechanisms and the extent to which they utilize specific DNA sequences remains incompletely understood.
Several studies have pointed towards the importance of genome sequence in shaping epigenetic states. For example, insulators are specific DNA sequences that protect genes from the regulatory effects of neighboring domains, thus enforcing domain boundaries [17]. As presently defined, insulator activity has two components: the ability to prevent cross-talk between an enhancer and promoter (enhancer blockers) and the ability to stop the spread of repressive heterochromatin (heterochromatin barriers) [17]–[20]. First identified and characterized in flies [18],[21], insulators have since been identified in vertebrates [22]–[24].
Elucidating the role of genome sequence in shaping chromatin domains requires an experimental system in which heterochromatin nucleation can be initiated in a controlled manner. To this end, we have examined heterochromatin spreading from a de novo nucleation site in the fission yeast, Schizosaccharomyces pombe. The unique advantage of this system, in addition to its genetic tractability, is the presence of well-defined DNA sequences, referred to here as heterochromatin-nucleating sequences, that are sufficient to induce heterochromatin formation de novo [25]–[27]. Moreover, introduction of a de novo heterochromatin domain at a euchromatic locus permits a simplified view of this process, in contrast to native domains of heterochromatin that result from the complex interplay of multiple sites of nucleation and heterochromatin barriers [27]–[29]. Analysis of the resulting de novo heterochromatin domains clearly implicates primary DNA sequence in defining both the magnitude and extent of the heterochromatin domain. The conceptual framework that emerges from this study provides a basis for exploring the nature of complex genomes and the impact of genome sequence on the establishment and maintenance of chromatin domains, in organisms ranging from yeast to mammals.
Previous studies in S. pombe have demonstrated that a fragment of pericentromeric DNA, called L5, is capable of nucleating heterochromatin, marked by di-methylation at H3K9 (H3K9me2) and the presence of the HP1 homologue, Swi6p, at an ectopic site through an RNAi-dependent mechanism [25],[30]. Integration of the L5 element leads to the repression of an adjacent reporter gene in a manner that appears largely similar to that observed at the endogenous centromere [25],[27],[30]. What is unknown, however, is the extent to which L5-nucleated heterochromatin is capable of extending past the reporter construct into endogenous genomic sequences.
To address this question, we created a construct containing the 1.6 kb L5 element upstream of an ade6+ reporter gene. This construct was then integrated at the euchromatic ura4+ locus in order to create ura4::L5-ade6+ strains. In addition to the L5-containing construct, a control construct bearing only the ade6+ gene was also integrated at the ura4+ locus (ura4::ade6+). The effect of L5-integration on the chromatin environment of sequences within the ura4 locus was characterized by quantifying H3K9me2 and Swi6p levels throughout the region using chromatin immunoprecipitation (ChIP).
In the presence of the L5 element, H3K9me2 was enriched ∼2- to >10-fold over both the reporter gene and the surrounding genomic neighborhood (Figure 1A and Figure S1A), extending 4 kb proximal and 10 kb distal to L5. The pattern of Swi6p enrichment is remarkably similar to the level of H3K9me2, consistent with previous reports demonstrating that H3K9me2 and Swi6p have tightly overlapping distributions within heterochromatin domains (Figure 1A and Figure S1A) [28]. These data demonstrate that heterochromatin assembly is not limited to the L5-element and the reporter gene; instead, heterochromatin spreads bi-directionally into adjacent, formerly euchromatic, sequences, resulting in a de novo heterochromatin domain that spans approximately 15 kb. Throughout, we will describe the properties of a heterochromatin domain by its extent, the distance over which heterochromatin is enriched, and its magnitude, the level of heterochromatin enrichment at a given location.
Because the chromatin state of genes near ura4 changes upon insertion of L5, we sought to determine whether gene expression at the ectopic locus was also altered. Quantitative RT- PCR (qRT–PCR) was used to quantify the levels of mRNA in the presence of L5 relative to control strains lacking L5. As expected from earlier studies [25], we observed an L5-dependent decrease in ade6+ expression; however, the reduction in expression was moderate (34±5%), indicating that silencing is incomplete in these strains (Table 1). In addition to ade6+, two genes within the de novo heterochromatin domain, located 2.7 kb proximal and 4.9 kb distal from the L5 element, also exhibited a decrease in expression in the presence of L5, 43±10% and 52±8%, respectively. Gene expression outside of the de novo heterochromatin domain was also analyzed (Figure 1A). As predicted, three genes (A, B, H) had no significant difference in transcript abundance in the presence of L5 (Table 1). The remaining gene, G, as well as gene F that lies within the de novo domain, exhibit a discordant relationship between the enrichment of heterochromatin marks and the level of gene expression. Together, these results suggest that gene-specific features may have a greater influence on the level of gene repression, as compared to the centromere, where gene repression is more complete [31].
To further explore the extent of ade6+silencing, we utilized a phenotypic assay for ade6+expression. This assay allows the extent of silencing to be resolved on a sub-colony level, as opposed to the population level queried by qRT–PCR. Under conditions of limiting adenine, yeast that are mutant, or silenced [1], for ade6+ accumulate a metabolic intermediate that results in red pigmentation. In contrast, cells in which ade6+is expressed at wild type levels remain white. Results from the phenotypic assay indicate that there is significant heterogeneity among colonies in the ura4::L5-ade6+ strains (Figure 1B). Similar to classic PEV, the colony phenotypes ranged from white to red [8]. However, distinct from PEV in Drosophila, we also observed intermediate phenotypes of pink and red with white sectors, consistent with PEV as observed in yeast [1],[32],[33].
We next wanted to determine whether the magnitude and extent of a de novo heterochromatin domain depends upon its location in the S. pombe genome or whether domain properties are intrinsic to the L5 element itself. To explore this, we identified a second integrant of the ura4::L5-ade6+construct on chromosome 2 (spbc2f12.03::ura4::L5-ade6+). Comparison of H3K9me2 and Swi6p enrichment at L5 and ade6+ between the two sites of integration reveals similar patterns of enrichment, suggesting that the nucleation of heterochromatin and local spreading are not sensitive to the changes in genomic location from ura4 to spbc2f12.03 (Figure 1C).
We next compared the magnitude and extent of the de novo heterochromatin domains formed at these two genomic locations. Distal to L5, the patterns of heterochromatin enrichment are markedly similar between the ura4 and spbc2f12.03 loci. In contrast, heterochromatin is observed 9 kb proximal to L5 at the spbc2f12.03 locus as compared to only 4.9 kb at the ura4 locus (compare Figure 1A and 1C). While this expansion at the spbc2f12.03 locus is only modestly enriched in H3K9me2, it is also marked by the presence of Swi6p (Figure 1C), suggesting there is a 4.1 kb expansion of the heterochromatin domain relative to the ura4 locus. Thus, the proximal boundary of the de novo heterochromatin domain at the ura4 locus is influenced by genomic location as opposed to reflecting an intrinsic limitation of the L5-element.
To determine whether the de novo heterochromatin domain at spbc2f12.03 alters gene expression, qRT–PCR was used to analyze mRNA levels at the spbc2f12.03 ectopic locus. Repression is observed at ade6+ and the nearby rpl1701+ gene, but not genes 2, 3 and 5 (Figure 1C, gene 4; Table 1). Thus, analogous to gene F at the ura4 locus, the recruitment of heterochromatic marks to an ectopic locus is not always associated with significant gene silencing.
The experiments described above demonstrate that the extent of an L5-dependent de novo heterochromatin domain can vary between different locations in the genome. To explore whether these differences are attributable to cis-acting factors, we engineered constructs in which different DNA sequences were placed adjacent to L5. These constructs were then inserted at the ura4 locus to examine the role of sequence in defining the heterochromatin domain, without altering its location in the genome.
First, 5 kb of S. pombe DNA taken from a region between two divergently transcribed genes (spcc320.02+ and spcc320.03+) was positioned between L5 and ade6+. This region was selected because it is one of the larger regions in the S. pombe genome in which known protein-coding genes are absent and because it normally lacks heterochromatic modifications [28]. Having established that this region maintains the absence of H3K9me2 when moved to the ura4 locus in strains lacking L5 (Figure S2A), H3K9me2 levels were queried over this DNA in the presence of the L5 element. Heterochromatin was robustly enriched over the 5 kb insert DNA (Figure 2A). Strikingly, the magnitude of H3K9me2 enrichment is comparable to the level observed at S. pombe centromeres (Figure 2A), suggesting that H3K9me2 can reach and sustain high levels of occupancy over the entire region. This is in contrast to the pattern of heterochromatin spreading over the gene-rich ura4 and spbc2f12.03 neighborhoods (Figure 1A and 1C). The difference between the magnitudes of heterochromatin enrichment between these DNA sequences supports the role of cis-acting DNA sequences, potentially the genes themselves, in shaping the characteristics of heterochromatin domains.
Consistent with this hypothesis, there is a significant reduction in H3K9me2 enrichment coincident with the start of the ade6+ gene, and the level of enrichment at this location is similar to the level observed when ade6+ is adjacent to L5 (compare Figure 1A and 2A). One interpretation of these data is that only low levels of heterochromatin can exist in transcriptionally active regions. Thus, when heterochromatin spreads from the spacer DNA into the ade6+ gene, the ade6+ gene behaves as a heterochromatin attenuator. Alternatively, this may indicate that the extent of spreading is constrained such that heterochromatin cannot spread, with high levels of enrichment, farther than 5.6 kb from L5.
To address this latter possibility, a longer spacer sequence was selected and inserted between L5 and ade6+. Sequences from lambda phage were chosen, as they have been used in previous epigenetic studies as spacer DNA [34]. No significant enrichment was observed over the length of the 7 kb insert in the absence of L5, suggesting that these sequences do not nucleate heterochromatin on their own (Figure S2B). In contrast, when the L5 element is present, robust enrichment in H3K9me2 was observed over the length of the lambda fragment, at levels similar to that of the centromeres and the 5 kb S. pombe spacer fragment (Figure 2B). Moreover, when we extended our analysis to include the levels of H3K9me2 enrichment at endogenous sequences in ura4::L5-7kb-ade6+ strains we found that it was remarkably similar to the levels observed in strains lacking spacer DNA (Figure S3). Thus, the addition of spacer DNA (up to 7 kb) does not constrain the extent of a de novo heterochromatin domain. Instead, our data are consistent with a model in which both the extent and magnitude of a heterochromatin domain are dictated by features of the underlying DNA sequence.
Because endogenous sequences can influence the extent of a heterochromatin domain, we next sought to determine the factors that mediate the interaction of DNA sequence and chromatin. Based on the observation that some barriers require formation of a transcription complex [22],[27],[29],[35],[36], we investigated the relationship between domain size and transcriptional activity.
The level of transcriptional activity within both the ura4 and the spbc2f12.03 regions could be assessed using previously reported data sets [29],[37]. Transcriptional activity was inferred from both the steady state levels of mRNA and the level of RNA Polymerase II (Pol II) and RNA Polymerase III (Pol III) enrichment at the promoter. Between the two regions, there were five loci that were transcriptionally exceptional: one gene that was transcribed by Pol III and four genes with unusually high levels of Pol II transcriptional activity (Table 2) [29],[37].
The ura4 genomic neighborhood includes a Pol III-transcribed tDNAGly (gene D in Figure 1A), which is coincident with an H3K9me2 gap. This gap could be attributed to general nucleosome depletion or, alternatively, to nucleosomes shielded from H3K9me2 modification by the Pol III transcription complex. Supporting the former hypothesis, tDNA genes are generally depleted of nucleosomes when compared to the genome average [38]. To distinguish between these two possibilities, an antibody to the C-terminus of histone H3 was used to characterize nucleosome occupancy surrounding to the tDNAGly gene. Indeed, the level of H3 enrichment at the tDNAGly was reduced relative to sequences in the surrounding neighborhood (Figure S1B), indicating that the observed H3K9me2 gap is due to decreased nucleosome occupancy surrounding the tDNAGly.
The ura4 genomic neighborhood includes the gene spcc330.06+(gene C in Figure 1A), which is highly expressed and enriched in Pol II (at the 94th percentile genome-wide) at its promoter (Table 2) [37]. This gene is located within a striking transition in H3K9me2 enrichment from 14-fold to <2-fold enrichment over a distance of only 2.7 kb (Figure 1A). In contrast to the nucleosome gap discussed above, this transition marks a boundary of heterochromatin enrichment and cannot be explained by reduced nucleosome occupancy (Figure S1B). We hypothesize that this gene may behave as a heterochromatin barrier and, more broadly, that highly expressed genes in general may be effective heterochromatin barriers.
Within the spbc2f12.03 genomic neighborhood, three genes are highly transcribed (genes 1,4 and 6 in Figure 1C). One of these genes (gene 6) is distal to the boundary of the de novo heterochromatin domain, and as such is uninformative. However, genes 1 and 4 (Figure 1B) are located at boundaries of the de novo heterochromatin domain, consistent with the hypothesis that highly expressed genes weaken and/or stop the spread of de novo heterochromatin.
To directly test whether the presence of transcribed genes can influence the extent of a de novo heterochromatin domain, we constructed a chimeric reporter gene composed of the strong, repressible, nmt1+ promoter driving expression of the his3+ open reading frame (Pnmt1-his3+) [39]. This construct was then inserted within the 7 kb spacer fragment, and heterochromatin spreading was monitored over the spacer sequences and the inserted gene. As expected, H3K9me2 was highly enriched over the spacer DNA proximal to Pnmt1-his3+, consistent with the levels observed in uninterrupted spacer strains (Figure 3). However, the magnitude of H3K9me2 enrichment decreases over the Pnmt1-his3+sequences and remains reduced over the distal portion of the spacer DNA (Figure 3). These data demonstrate that the insertion of genic sequences within the spacer DNA attenuates the spread of heterochromatin and further support the hypothesis that the presence of genes within the ura4 and spbc2f12.03 neighborhoods limits heterochromatin spreading. It is interesting, however, that the Pnmt1-his3+ construct, despite being more highly transcribed than spcc330.06+ (Figure S5A), does not exhibit complete barrier activity (Figure 3), suggesting that factors other than high levels of transcriptional activity are required for complete barrier activity.
Because the presence of genes antagonizes heterochromatin spreading, we sought to determine whether a high level of transcriptional activity is required for attenuator activity. To test this we took advantage of an engineered allele of the nmt1+ promoter that results in reduced transcription efficiency [40], and cultured these strains in medium containing thiamine, which results in further repression of the nmt1+ promoter [39]. Despite a ∼570 fold decrease in expression the weakened Pnmt1-his3+ gene still exhibited significant attenuation ability, indistinguishable from the strongest allele (Figure S5B). Thus, other features of the nmt1+ promoter may serve to attenuate the spread of heterochromatin. Indeed, the region of the promoter that is required for thiamine repression binds a protein complex independent of thiamine conditions [41]. This protein complex, or other complexes that localize to the promoter independent of thiamine and transcription efficiency, may serve to attenuate the spread of heterochromatin.
Having demonstrated the impact of genome sequence on the extent of spreading from a heterochromatin-nucleating sequence, we wanted to determine whether changes to the sequence content, in terms of L5 copy number, would alter the properties of a de novo heterochromatin domain. Thus, two copies of L5 were inserted in tandem at the ura4 locus. The magnitude and extent of heterochromatin enrichment in these strains was markedly similar to strains bearing one copy of L5 (Figure 4A), suggesting that the copy number of L5 does not notably enhance heterochromatin enrichment or spreading within a de novo domain.
We also wanted to address the possible role of trans-acting factors in regulating the extent of the heterochromatin domain, either by competition with other heterochromatic regions for limiting heterochromatin components [42] or by competition between heterochromatic and euchromatic factors for the same nucleosome substrate [43]. We hypothesized that increasing the dosage of heterochromatin proteins (or reducing the amount of competing factors) should result in the expansion of a heterochromatin domain [9], [42]–[45].
Swi6p is a dosage-dependent modifier of heterochromatin levels at the S. pombe mating-type locus as well as a limiting factor in heterochromatin formation [46],[47]. Thus, we analyzed the magnitude and extent of the ura4 de novo heterochromatin domain in strains bearing three copies of swi6+ [47]. We confirmed that the level of swi6+ mRNA is increased by 2.7-fold in these strains (data not shown). While the local magnitude of H3K9me2 proximal to L5 was increased in these strains (Figure 4B), the increased dosage of swi6+ did not result in the expansion of the heterochromatin domain. Consistent with the increased level of H3K9me2 enrichment, ade6+ expression was further reduced in these strains, resulting in an increased proportion of red colonies (Figure 4C and Figure S4). In contrast, increased swi6+ dosage did not significantly affect expression of other genes within the ura4+ neighborhood (Figure S4). This suggests that, while the level of Swi6p influences both the local concentration of H3K9me2 and the level of gene expression, the extent of the de novo heterochromatin domain is not sensitive to increased dosage of swi6+.
In the absence of known transcribed elements, H3K9me2 spreads unattenuated over distances at least up to 7 kb (Figure 2B), resulting in a consistent level of H3K9me2-enrichment at ade6+ independent of the presence of spacer DNA (Figure 5C). What remains to be addressed is whether the level and the stability of gene silencing differ between strains containing spacer DNA versus ura4::L5-ade6+.
To explore this question, we compared the levels of ade6+expression by qRT–PCR and found that when ade6+was located 7 kb away from the L5 element, distal to the lambda spacer, silencing was no longer observed despite the presence of H3K9me2 (Figure 5A and 5C). This finding was confirmed using the phenotypic ade6+ assay, which revealed a much lower level of silencing in ura4::L5-7kb-ade6+ strains (Figure 5B). These data suggest that, even when the levels of H3K9me2-enrichment are similar (Figure 5C), heterochromatin formed proximal to L5 and heterochromatin formed over spacer DNA can have different effects on gene expression.
We suspected that the differences in silencing could be attributed to the levels of Swi6p at the ade6+ gene in ura4::L5-ade6+ as compared to ura4::L5-7kb-ade6+ strains. To address this hypothesis, we assessed the level of Swi6p enrichment over the lambda spacer DNA and distal ade6+ gene. We observed a significant decrease in Swi6p enrichment relative to the level of H3K9me2 across spacer and ade6+ DNA when compared to ura4::L5-ade6+ and spbc2f12.03::ura4::L5-ade6+ strains, as well as to other heterochromatic loci (Figure S6 and Figure 5C). This reduction is consistent with the decreased levels of silencing and could be a function of long distance spreading or a sequence-dependent affect of spacer DNA.
In addition to the total level of gene silencing, another manner in which the reduced levels of Swi6p in spacer strains could affect gene expression is by altering the stability of gene repression. When transgenes are placed within the centromere, or at locations throughout the mating type locus, their phenotypic stability (silenced or expressed) can vary with location and Swi6p dosage [31],[32],[46],[47]. To address whether the silenced and expressed states are stable to equivalent degrees in cases of local (high levels of Swi6p) versus spreading over spacer DNA (reduced levels of Swi6p) we chose colonies that were either silenced or expressed, as determined by their ade6+ expression phenotype (ie. entirely red or entirely white, respectively). The stability of the silenced state was determined by the proportion of progeny that exhibited silencing after a period of overnight growth. We examined the phenotypic stability of ade6+ and ade6− controls, and as expected, the progeny maintained the appropriate phenotype (Figure 6A). However, when expressed colonies were chosen from ura4::L5-ade6+ strains, only ∼62% maintained the completely expressed phenotype, while the remaining colonies switched to a partially or completely silenced phenotype. This is in stark contrast to the ura4::L5-7kb-ade6+ strains, in which 95% of the progeny maintained the expressed state (Figure 6B), suggesting that the establishment of silencing (i.e., switching from an ade6+ expressed state to a silenced state) occurs less often when ade6+ is distal to 7 kb of spacer DNA and less enriched in Swi6p, despite comparable levels of the epigenetic mark H3K9me2.
When silenced colonies were selected from ura4::L5-ade6+ strains, ∼18% of the progeny exhibited phenotypes indicative of complete silencing and 90% exhibited at least partial silencing. In contrast, silenced colonies from ura4::L5-7kb-ade6+ strains were less likely to give rise to progeny that exhibited complete or partial silencing (∼1% and 69%, respectively), suggesting that maintenance of silencing is also less frequent when ade6+ is separated from L5 by spacer DNA and reduced in Swi6p localization (Figure 6C). These data provide evidence that the level of Swi6 impacts the establishment and maintenance of silencing, despite consistent levels of H3K9me2.
Ectopic gene silencing and/or heterochromatin formation has previously been studied in mammalian systems [48]–[51]. Ectopic X inactivation, for example, has been shown to affect gene expression on a large scale [52]. Typically, however, the complex nature of the mammalian genome restricts the focus of these studies to local heterochromatin formation and single gene repression. In this study, the compact nature of the S. pombe genome and our ability to robustly query for the presence of heterochromatin allowed us to rigorously test the response of multiple DNA sequences to encroaching heterochromatin. Our data demonstrate a clear effect of genomic sequence in shaping both the extent and magnitude of a heterochromatin domain and demonstrate that, while the eukaryotic genome is permissive to the negative transcriptional effects of heterochromatin, euchromatic sequences can counteract encroaching heterochromatin.
The relationship between the size of a heterochromatin domain and the presence of specific heterochromatin barriers has been previously established in a number of eukaryotic organisms [53]. Our study extends this conclusion, demonstrating that DNA sequences exert a range of effects on heterochromatin domains. For example, the ade6+ gene dampens heterochromatin enrichment independent of both genomic location and distance from L5, but is insufficient to completely stop heterochromatin spreading (Figures 1A, 1C, 2A, and 2B). In contrast, intergenic and spacer DNA sequences promote the assembly of robust H3K9me2. We propose that there is a spectrum of effects, ranging from antagonistic to cooperative, that genomic sequence can exert on heterochromatin (Figure 7). This model incorporates the complexity and context dependence of genomic sequence and its relationship to heterochromatin and is applicable to sequences in yeast, as seen here, or in more complex genomes, as will be discussed below.
For this model, we have subdivided the discrete extremes of DNA sequences noted previously (that is, heterochromatin-nucleating sequences and heterochromatin barriers) into subclasses that include attenuators, neutral elements and protosilencers/boosters. While this is helpful for purposes of discussion, we do not wish to impose strict definitions, especially in light of data from this study, suggesting that particular sequences may be placed at multiple points along the continuum, depending on their context.
Our data confirm that heterochromatin can spread from L5 in both directions over euchromatic DNA, resulting in a de novo heterochromatin domain encompassing multiple endogenous genes and altering gene expression (Figure 1A and 1C, Table 1). However, gene repression within the de novo domain is moderate, at most about 50%. The incomplete silencing observed within de novo heterochromatin domains, as well as the boundaries of these domains, may be a consequence of the factors present within euchromatic domains that antagonize the propagation of heterochromatin.
The boundaries of de novo heterochromatin domains are marked by three highly transcribed genes, implicating Pol II transcription in barrier activity (Figure 1A and 1C, Table 2) [37]. The ade6+ gene is also transcribed and enriched in Pol II, albeit to a lesser extent than the three putative barriers. These four sequences may rely on transcription to counteract the spread of heterochromatin from the L5 heterochromatin-nucleating sequences. However, high levels of transcription are insufficient for complete barrier activity (Figure 3). Furthermore, we find that, in the case of Pnmt1-his3+, the presence of genes can attenuate the spread of heterochromatin independent of the level of transcription (Figure S5B). We conclude that DNA sequences modify heterochromatin spreading through the sequence-dependent recruitment of other mediating factors, such as transcription complexes, and dictate whether a sequence behaves as a true barrier or falls in the range of heterochromatin attenuators (Figure 7). These findings are consistent with previous results implicating transcription factors and promoters with barrier activity [35], [54]–[58]. Sequence could also influence heterochromatin directly, as is the case with some examples of nucleosome positioning [59] or could reflect selective pressure to maintain domain boundaries (Table 1).
In addition to protein-coding genes, the ura4 de novo heterochromatin domain includes a tDNAGly gene (discussed below) and non-coding RNAs (Figure 1A). The ura4 locus is not unique in its transcriptional makeup, as recent studies have provided insight into the vast amount of transcription occurring in the S. pombe genome outside of canonical protein coding genes [60],[61]. Additionally, the ura4 neighborhood also includes solo long terminal repeats (LTRs) [62]. How these features interact with heterochromatin spreading, and whether they shape the formation of de novo heterochromatin domains warrants further genome-wide studies.
Transcription by Pol III complexes has an established relationship with barrier activity in yeast genomes [27],[29],[35],[36]. A tDNAAla within the S. pombe centromere 1 prevents the spread of heterochromatin into the abutting domain of centromeric chromatin. In contrast, the tDNAGly gene is not coincident with the domain boundary of the ectopic heterochromatin domain formed at ura4+; however it is deficient in H3K9me2 enrichment, due to the absence of a nucleosome(s) (Figure 1A and Figure S1B). Nucleosome depletion has been shown previously to restrict heterochromatin spreading [63]; in this context, the nucleosome gap may weaken the spread of heterochromatin, resulting in the gradual attenuation observed distal to the tDNAGly. We suggest that, like the ade6+ gene, the tDNAGly behaves as heterochromatin attenuator in our experimental system. It is interesting to note that, while other tDNAs substituted at the centromere recapitulate barrier activity, re-positioning of tDNAAla at a euchromatic locus resulted in an attenuation of heterochromatin spreading, but not complete barrier activity [27]. Together, these data establish a mechanistic link between heterochromatin barriers and attenuators, and implicate genomic context as an additional factor in determining where a sequence falls along the continuum of effects on heterochromatin (Figure 7).
Whether the effect of DNA sequence could be abrogated by increased dosage of heterochromatin proteins was also examined. Increased swi6+ resulted in increased levels of H3K9me2 over sequences adjacent to L5, as well as enhanced repression of ade6+, consistent with an increase in local heterochromatin (Figure 4B). However, this change in heterochromatin enrichment is not accompanied by an expansion of the domain. Further expansion of the domain is likely prevented by the barrier and attenuator activity of adjacent sequences, indicating that these sequences are robust to the increasing magnitude of heterochromatin in these strains. This is the also the case with models of PEV in mouse where enhancing heterochromatin formation is insufficient to cause PEV when a transgene is flanked by chromatin insulators [64]. Alternatively, enhanced propagation of heterochromatin could be limited by selection against increased silencing of genes within the de novo heterochroamtin domain.
As heterochromatin antagonists are characterized by different strengths, we propose that DNA sequences also differ in their ability to initiate or promote heterochromatin spreading. The identification of protosilencers, sequences that can actively contribute to gene silencing but only in specific “silencing-conducive” environments [65], supports this hypothesis. DNA sequences that are permissive to heterochromatin spreading can be conceptually subdivided into those that rely on active mechanisms, like those above, and those that passively allow heterochromatin but do not actively propagate the heterochromatic state (Figure 7). The spacer and S. pombe intergenic fragments may fall into this class of sequence elements. Both sequences allow formation of large heterochromatin domains with levels of H3K9me2 enrichment similar to that observed at the centromere (Figure 2A and 2B). Alternatively, these sequences may contain elements that enhance heterochromatin spreading, and thus would belong in the former class of sequences that actively promote heterochromatin spreading.
Interestingly, while high levels of H3K9me2 are sustained over the length of the lambda spacer DNA, the ratio of Swi6p/H3K9me2 is reduced, relative to both genome-wide data and data from the ura4 and spbc2f12.03 de novo heterochromatin domains (Figure 5C and Figure S6) [28]. The reduced levels of Swi6p correlate with reduced ability to establish and maintain silencing at ade6+ when compared to ura4::L5-ade6+ strains (Figure 6). These data suggest that lambda spacer DNA exerts a sequence-specific effect on the associated heterochromatin domain that results in reduced levels of gene repression.
The spreading of heterochromatin from L5 shares at least conceptual similarities with the spreading of gene silencing and, presumably, facultative heterochromatin from an ectopic X inactivation center in mammalian systems [52],[66]. Furthemore, as we demonstrate in fission yeast, genome sequence is also implicated in the organization of chromatin on the mammalian X chromosome [67]. The inactive X chromosome is organized in alternating domains of genes that are subject to inactivation (silenced) and domains of genes that escape from X inactivation (expressed) [68],[69], as well as by domains of different types of heterochromatin [70],[71]. A CTCF site on the mouse inactive X chromosome, located within such a transition region, exhibits insulator activity in transgene assays [23], thus implicating DNA sequence in maintaining the boundaries of expression domains. Moreover, the presence of specific DNA features on the X chromosome can be used to accurately predict whether a gene will be subject to, or escape from, X inactivation [72],[73]. However, as with the intergenic and spacer fragments in this study, it is unknown whether the sequences correlated with gene silencing passively permit the silent state, or whether they actively promote the propagation of gene silencing. Finally, LINE-1 elements have been proposed to behave as protosilencers, or booster elements, relaying transcriptional inactivation from sites of nucleation [72],[74]. While such evidence points to the importance of DNA sequence in regulating domains of gene expression on the X chromosome, the presence of barriers and other sequences in mammals has yet to be addressed fully.
The genotypes for strains used in this study are as listed (Table S1). Fission yeast media were prepared using standard procedures [75]. For repression of the nmt1 promoter 15 uM thiaimine was added [76]. A strain bearing the ade6DN/N allele (a loss of function mutation created by a 153 bp deletion of the ade6+ open reading frame [77]) was generated (Kfy539) and was transformed via electroporation (1.5 kV, 200Ω, 25uF) on a BioRad Gene Pulser II. Transformed cells were selected on PMG media lacking adenine [75]. Colonies derived from strain Kfy539 were then patched onto media containing 2 g/L of 5-fluoro-orotic acid (FOA) (MP Biomedicals) to select for disruption of ura4+. The resulting strains were screened, using Southern analysis, for appropriate integration of ade6+. Additionally, BW17 transformants were screened by Southern blot for the maintenance of the 7 kb lambda DNA fragment. At least three independent transformants of each genotype were maintained (with the exception of the random integrant, Kfy812) and used for further analysis. All transformants were then crossed into a swi6+ strain and the ura4::L5-ade6+ allele was selected for on the basis of FOA resistance. To create swi6+333 strains, ura4::L5-ade6+ strains were crossed into SPG1232 (Shiv Grewal) [47]. To create ura4::L5-7kb::(Pnmt1-his3+)-ade6+ the Pnmt1-his3+ construct was transformed into Kfy589, colonies were selected for on the basis of growth on media lacking histidine. After integration within lambda was confirmed by Southern analysis, these strains were crossed into a swi6+ strain.
To create plasmid BW5, ade6+ was amplified from S. pombe genomic DNA using primers BWP34F and BWP34Rb (Table S2) to add StuI, SpeI, ClaI, and BglII sites to the 5′ end of the product and Sac1, Sma1 and Stu1 sites to the 3′ end. The PCR product was then digested with StuI and inserted into the StuI site of ura4+ in pUC13/18. The ade6+ open reading frame and upstream region were sequenced to ensure no mutations had been introduced during cloning.
Plasmid BW7 was constructed through digestion of YL317 with SpeI and ClaI and subsequent purification of the L5-containing fragment [27]. L5 was then inserted into the SpeI/ClaI site of BW5. Plasmids BW32 and BW34 contain 4.9 kb of S. pombe intergenic DNA taken from between SPCC320.02 and SPCC320.03 inserted into the BglII site of BW5 and BW7, respectively. The intergenic fragment was digested from the cosmid SPCC320 using XbaI, subcloned into pUC13/18, and then digested with BamHI before inserting into the appropriate plasmid. Plasmids BW30 and BW17 were created by digesting the lambda phage genome (NEB) with BamHI and purifying the 7.2 kb fragment, which was then ligated into the BglII sites of BW5 and BW7, respectively. To create BW20 an additional copy of L5, as a BamHI–BglII fragment, was inserted into the BglII site of BW7.
Plasmids BWP40 and BWP41 were created by replacing the GFP ORF with his3+ within the plasmids pFA6a-kanMX6-P3nmt1-GFP and pFA6a-kanMX6-P41nmt1-GFP, respectively (A gift from Jian-Qiu Wu) [78]. A subfragment of the lambda spacer DNA was liberated from BW17 by digestion with BglII and cloned into pUC1318. The Pnmt1-his3+ containing fragment was then inserted within the PstI site in the lambda fragment.
To identify random integrants that did not disrupt the ura4 locus, we selected transformants on the basis of growth on PMG–adenine and death on FOA. These strains were then confirmed via Southern blot to have a single ade6+ insertion and the site of integration was mapped using an inverse PCR protocol modified from [79]. Genomic DNA (2 µL) was digested with MboI or Nde1 and incubated for 3.5 hours at 37°C. The digest was heat inactivated at 65°C for 20 minutes. 2 µL of the digest was added to a standard ligation reaction (T4 ligase, NEB) and incubated overnight at room temperature. Inverse PCR was performed using primers E367/BWP89F for the Nde1 digest and BWP37F/BWP32F for the Mbo1 digest. The PCR products were purified and sequenced using the PCR primers listed above.
Strains were grown overnight with shaking in YES media at 32°C and diluted to a concentration of 1e6 cells/mL. Cultures were diluted serially (1∶9) and plated on YES media lacking adenine.
To assess the stability of silencing, colonies that were scored as either completely white or completely red were identified using a Leica MZ7.5 microscope and grown for 24 hours in YES media before plating on YE plates lacking adenine.
For both protocols, plates were grown for three nights at 32°C and shifted to 4°C for 24 hours before photographing or counting.
Total nucleic acid was isolated from logarithmically growing cells in YES media at 32°C, and was then subjected to DNAse treatment and RT–PCR using oligodT as a primer. Expression was analyzed by quantitative PCR using SYBR Green on a Bio-rad myCycler, using primers specific to the wild type copy of ade6+ (BWP85F/R). Levels of mRNA from ade6+, and other genes queried, were expressed relative to act1+(BWP74F/R). The standard curve was generated using genomic DNA isolated from strain Kfy1. In order to be included in this study a PCR experiment had to have a PCR efficiency between 90–110% and a correlation coefficient >0.99.
The H3K9me2 ChIP protocol was adapted from [80]. Logarithmically growing cells from control and experimental strains were treated with 1% paraformaldehyde for 15 minutes. The cell wall was then destroyed through bead beating twice for two minutes in buffer containing protease inhibitors. The resulting material was then sheared to an average fragment size of 600 bp using sonication. Chromatin preps were then subdivided into three tubes: an input sample that was used to check shearing, an IP sample to which protein beads and antibodies to H3K9me2 (from Takeshi Urano) were added, and a mock sample to which only protein beads were added. The mock and IP samples were incubated overnight, and the beads were isolated and subjected to a series of washes. Finally, DNA was purified from all three samples (IP, mock, and input) with phenol-chloroform extraction and ethanol precipitation using glycogen as a carrier.
H3 ChIPs were preformed as above using an antibody to H3 (abcam 1791).
Swi6p ChIPs were performed using the above protocol modified from [81]. 2.5e8 cells were shifted to room temperature for two hours prior to fixation. Cells were fixed with 3% paraformaldehyde for 30 minutes at room temperature. 1 µL of antibody (from Shiv Grewal) was incubated with the IP sample overnight, prior to incubation with protein beads for two hours.
Quantitative PCR was used to assay levels of query/act1+ in IP reactions relative to a no-antibody control.
|
10.1371/journal.pbio.1002557 | Mitochondrial 16S rRNA Is Methylated by tRNA Methyltransferase TRMT61B in All Vertebrates | The mitochondrial ribosome, which translates all mitochondrial DNA (mtDNA)-encoded proteins, should be tightly regulated pre- and post-transcriptionally. Recently, we found RNA-DNA differences (RDDs) at human mitochondrial 16S (large) rRNA position 947 that were indicative of post-transcriptional modification. Here, we show that these 16S rRNA RDDs result from a 1-methyladenosine (m1A) modification introduced by TRMT61B, thus being the first vertebrate methyltransferase that modifies both tRNA and rRNAs. m1A947 is conserved in humans and all vertebrates having adenine at the corresponding mtDNA position (90% of vertebrates). However, this mtDNA base is a thymine in 10% of the vertebrates and a guanine in the 23S rRNA of 95% of bacteria, suggesting alternative evolutionary solutions. m1A, uridine, or guanine may stabilize the local structure of mitochondrial and bacterial ribosomes. Experimental assessment of genome-edited Escherichia coli showed that unmodified adenine caused impaired protein synthesis and growth. Our findings revealed a conserved mechanism of rRNA modification that has been selected instead of DNA mutations to enable proper mitochondrial ribosome function.
| RNA modifications constitute an important layer of information, with functional implications that are not written in the underlying DNA sequence. Recently, we observed an apparent RNA-DNA difference (RDD) at position 947 of the human mitochondrial 16S ribosomal RNA (rRNA), but its nature and mechanism were unclear. Here we show that this disparity reflects an m1A modification (methylation at position 1 of the adenine moiety), and demonstrated by a combination of knock-down experiments in cells and in vitro methylation assays that the tRNA methyltransferase TRMT61B is the best candidate enzyme to introduce this modification. We also show that this modification is present in most of the 16S rRNA molecules in isolated mitochondrial ribosomes, and that it occurs in all vertebrates with an adenine (90% of the vertebrates), but not in those with a thymidine at this 16S rRNA position. Finally, as the first step towards understanding the functional importance of this rRNA modification, we used a genome-edited bacterial system to demonstrate that an unmodified adenine reduced the growth and translation rates of the bacteria as compared to both wild-type bacteria and mutant bacteria with a thymidine in the relevant position. Hence, three solutions were selected during evolution to allow proper function of the mitochondrial 16S rRNA—either RNA modification or two alternative ancient evolutionary DNA mutations.
| Most RNAs are enzymatically modified after transcription. To date, more than 100 different species of chemical modifications have been identified in various RNA molecules from all domains of life [1]. Historically, analyses of RNA modifications have been limited to abundant RNA molecules such as tRNA, rRNA, or UsnRNA. Recently, transcriptome-wide analyses using deep-sequencing combined with biochemical approaches have facilitated the identification of several modified bases in mRNAs and non-coding RNAs [2–5]. Additionally, transcriptomes were compared to their corresponding genomic sequences as a practical approach for the detection of RNA-DNA differences (RDDs) and identification of the canonical A-to-I and C-to-U RNA editing sites in diverse eukaryotes [6]. Accumulating evidence suggests the presence of non-canonical RDD sites (i.e., edits other than A-to-I [G] or C-to-U) [7–9]. However, their very existence and general importance has been questioned [10], and neither the mechanisms generating non-canonical RDDs nor conclusive experimental evidence for their functional role have been described in vertebrates.
It was previously suggested that many non-canonical RDDs are not editing events per-se but rather reflect RNA modifications [8]. RNA modifications, such as methylation of specific bases, occur in mitochondrial RNAs in many vertebrates and invertebrates [11–14]. These modifications play crucial roles in mitochondrial activity, and their absence leads to pathological consequences [15]. While RNA modifications in mitochondrial tRNAs of mammals were thoroughly mapped [16], the mitochondrial ribosomal (r)RNAs have been far less investigated.
By sequencing both the genome and corresponding transcriptome of the human mitochondria in lymphoblastoid cells, we recently identified three new RDD sites [8] that were also corroborated in a larger human sample size [13]. The most prevalent RDD occurred at adenine 947 of human mitochondrial 16S rRNA (mitochondrial DNA [mtDNA] position 2617). Specifically, we showed that this position was a mixture of A, T, or G in RNA-seq reads, suggesting the presence of an unidentified modified base [8].
Recently, cryo-EM structures of mammalian mitochondrial ribosomes (mitoribosomes) have been solved at high resolution, sufficient to map residues of rRNAs [17,18]. A947 is located in helix 71 (H71) of 16S rRNA in close proximity to the inter-subunit bridge B3, suggesting that the modified base at this position may play a functional role in mitochondrial translation. In this study, we investigated the function of A947 and its RDD. We found that the non-canonical mitochondrial 16S rRNA RDDs resulted from a 1-methyladenosine (m1A) RNA modification and identified the candidate modifying enzyme. We demonstrate that the modification occurred in most vertebrates and is enriched in the mature mammalian mitoribosome. Finally, we used a bacterial model to directly link the nucleotide identity of 16S rRNA position 947 to cellular growth and protein translation. Indeed, mutations in the equivalent structural position of A947 in Escherichia coli imply functional importance for modifying this base during evolution.
We previously hypothesized that the observed 16S rRNA RDDs in RNA-seq represent an RNA modification that was manifested as a mixture of reads with thymine, adenine, and lower occurrences of guanine. However, the nature of this putative modification and its underlying mechanism remained to be elucidated. Previous RNA-seq analysis of templates harboring a m1A modification [19] resulted in nucleotide distribution of sequencing reads similar to our observed reads in the mitochondrial 16S rRNA transcript at position 947 [8].
To elucidate the chemical identity of 16S rRNA position 947, we isolated 16S rRNA from HeLa cells and analyzed its modifications by capillary liquid chromatography and nano electrospray mass spectrometry [20–22]. By assigning RNase T1-digested fragments of 16S rRNA, we observed known 2’-O methylations including Gm1145, Um1369 and Gm1370 (S1 Fig). In addition, we clearly detected the mono-methylated RNA fragment (positions 939–950, MW 3833.5) containing the RDD site at position 947 (Fig 1A). Further probing of the RNA fragment by collision-induced dissociation revealed that the methylation occurs at the adenosine residue occupying position 947 (Fig 1B).
It is known that two isozymes introduce the m1A modification in human mitochondrial tRNAs [16]. TRMT10C is a subunit of the mitochondrial RNase P complex that additionally acts as a methyltransferase generating m1A as well as m1G at position 9 in mitochondrial tRNAs [23]. TRMT61B, another methyltransferase, is responsible for m1A at position 58 in some mitochondrial tRNAs [24]. Notably, a recent Genome Wide Association Study revealed association between the levels of our observed 16S rRNA RDDs with SNPs in TRMT61B [13]. We thus hypothesized that the RDD at 16S rRNA position 947 echoes an m1A modification, likely introduced by TRMT61B.
To examine whether TRMT61B or TRMT10C catalyzes the formation of methyladenosine at position 947 in 16S rRNA, we knocked down each of them by siRNAs in HeLa cells, followed by total RNA extraction to obtain templates for primer extension. In control experiments of mock or luciferase knockdown (Fig 2A), the cDNA extended from the primer was strongly arrested at position 948, and partially extended to C942 by inserting dideoxy guanosine. This indicated that A947 is partially methylated. In addition, we also observed a clear band due to m1A58 in tRNALeu(UUR) from HeLa cells treated by luciferase siRNA (Fig 2B). Upon knockdown of TRMT61B (Fig 2A), nearly half of the cDNA extended past position 947 and stopped at position 942. This demonstrated that the methyladenosine at position 947 was m1A, introduced by TRMT61B. In contrast, knock down of TRMT10C did not lead to altered cDNA extension. As a positive control, hypomodification of m1A58 in tRNALeu(UUR) was observed when TRMT61B was repressed (Fig 2B). To confirm this observation by RDD, total RNAs from HeLa cells treated by siRNAs targeting for TRMT61B or luciferase as a control were subjected to RNA-seq analyses, followed by mapping to the mtDNA sequence (Fig 2C, S1 Data). Mixed nucleotide frequencies at position 947 observed in the cells treated with the control siRNA were dramatically altered and converged into adenosine in the cells treated with siTRMT61B, supporting the result from the primer extension experiment. Moreover, we observed a slight decrease in the read coverage around position 947. This observed pattern is in agreement with a recently published report describing the signature of m1A in RNA-seq data [25]. Notably, the observed levels of reads with either a T or a G decreased (as well as the reduction in coverage) in the RNA extracted from the siTRMT61B cells (Fig 2C).
To demonstrate that TRMT61B is a methyltransferase directly responsible for m1A947 formation in mitochondrial 16S rRNA, we carried out an in vitro reconstitution of m1A using recombinant TRMT61B. Total RNA extracted from HeLa cells treated by siRNA targeting TRMT61B was incubated with recombinant TRMT61B in the presence of AdoMet. m1A947 formation was specifically detected by primer extension assay (Fig 2D). In the presence of both recombinant TRMT61B and AdoMet, the cDNA band that extended up to position 942 decreased, and the cDNA band that arrested at position 948 clearly increased, indicating that m1A947 was reconstituted in vitro. In the negative control experiment, m1A947 was not introduced without AdoMet. Moreover, using in vitro transcription, we prepared a 114-nucleotide-long RNA segment (16S rRNA positions 866–979) and performed an in vitro methylation assay by TRMT61B. As a negative control, we prepared an active-site mutant of TRMT61B (D335A mutation), according to the biochemical study on TrmI [26], a bacterial ortholog of TRMT61B. Then, we carried out an in vitro methylation assay of the 114-mer RNA segment with either the wild-type TRMT61B or the D335A mutant in the presence or absence of AdoMet, followed by RNase A digestion, and subjected them to capillary LC/nano-ESI-MS analysis. The results indicate that TRMT61B clearly introduced the m1A947 in the 114-mer RNA segment in the presence of AdoMet (Fig 2E). Moreover, the methylated tetramer (AAm1AUp) produced by RNase A digestion was probed by CID, and its sequence was confirmed by assignment of the product ions (S2 Fig). As expected, the D335A TRMT61B mutant failed to introduce m1A947 in the segment (Fig 2E). Taken together, these results clearly demonstrate that the RDD site at 16S rRNA position 947 is m1A introduced by mitochondrial methyltransferase TRMT61B.
Our previous phylogenetic analysis revealed high conservation of mtDNA position 2617, i.e., 16S rRNA position 947 [8]. Specifically, this position was an adenine in nearly 90% of all tested vertebrate species, while the remaining 10% had a thymine. The above findings are consistent with the idea that the presence of the m1A modification results in a mixture of reads with mostly adenine or thymine and less with a guanine in RNA-seq data, as previously observed [19]. To assess the extent to which position A947 is methylated across vertebrate phylogeny, we analyzed RNA-seq data from nine species representing major vertebrate taxa (placental mammals, marsupials, monotremes, birds, reptiles, amphibians, and bony fish) (Fig 3, S1 Data). Notably, special care was employed to assure proper RDD identification while excluding sequencing and mapping errors (see Materials and Methods) [8].
Our analysis revealed that RDDs (A-to-U and A-to-G) occurred in the 16S rRNA of all tested species in which an adenine occupied mtDNA position 2617, indicating the presence of m1A947 in these 16S rRNAs. In contrast, species with a thymine in their mtDNA maintained a uridine in their RNAs (Fig 3, S1 Data). Notably, the RDD levels varied among the tested species, possibly due to physiological differences between different vertebrates. These observations further support our interpretation that position 947 of 16S rRNA is modified and that this modification is highly conserved across vertebrates.
The m1A947 modification was identified in RNA-seq data generated from total RNA samples, harboring both mature and premature mitochondrial transcripts. As a first step to elucidate the functional importance of this modification, we assessed the extent to which 16S rRNA position 947 is present in the modified version in the mature ribosome. To this end, we purified whole mitochondria and isolated mitoribosomes from a single Sus scrofa liver specimen and sequenced mtDNA and RNA from the isolated mitochondria as well as RNA from the purified mitoribosome. As expected, 100% of mtDNA reads from isolated mitochondria showed an adenine at the S. scrofa orthologous position of human mtDNA position 2617 (Fig 3B, S1 Data). Remarkably, while RDDs appeared in ~75% of the mitochondrial total RNA sample, their prevalence increased to ~90% in the purified mitoribosome (p < 10E-10, χ2 test, Fig 3B, S1 Data). This RDD enrichment in the mammalian mitoribosome supports the interpretation that the mature mitoribosome likely almost entirely contains m1A947 16S rRNA, thus further supporting the functional importance of this modification.
Structural analysis of position 947 of 16S rRNA revealed that it is likely involved in anchoring H71 by forming interactions with H64 and H92 of the 39S subunit (Fig 4).
Notably, this position is structurally conserved from bacteria to mammals in both mitochondrial and cytoplasmic ribosomal structures [27–30]. Our RNA-seq analysis of diverse vertebrate species showed an RDD in species with an adenine, but not with a thymine, at 16S rRNA position 947. Furthermore, this position is occupied by a thymine in the human cytoplasmic ribosome and by a guanine in 95% of all tested bacterial species. We therefore hypothesized that three functional alternatives arose during evolution, all of which are capable of maintaining full ribosomal activity: (A) an unmodified thymine (human cytoplasmic and 10% of the vertebrate mitochondrial ribosomes), (B) an unmodified guanine (in most bacteria), and (C) an m1A modification in 90% of the mitochondrial ribosomes. According to this hypothesis, forcing an unmodified adenine into the mature ribosome would interfere with its activity. Because there is no available technology to modify specific mtDNA nucleotides in cells, we chose E. coli as our model system to test this hypothesis. Using the genome engineering technology MAGE [31], we successfully replaced the endogenous bacterial nucleotide at position 1954, which corresponds to human mitochondrial 16S rRNA position 947, in all seven 23S rRNA gene copies. We independently replaced the wild-type guanine for either a thymine or an adenine (S3 Fig) and tested the resulting strains in terms of growth rate and translation efficiency. Remarkably, strains harboring an adenine grew significantly slower as compared to strains harboring either the WT base (guanine) or a thymine (Fig 5A, S1 Data). This result is consistent with our hypothesis that the presence of either thymine or guanine is important for cell growth, and that an unmodified adenine has a negative effect at this position on the mitochondrial and bacterial translational machineries.
Next, we aimed at assessing the effect of the mutations on ribosomal activity by examining the impact of each base at the modified position on protein synthesis. Indeed, in vivo examination of yellow fluorescent protein (YFP) production revealed that strains harboring an adenine at position 1,954 showed decreased maximal protein synthesis rate (64.4% and 76.2% in the mutant and WT, respectively) as well as a decrease in total protein production (75% and 83.6% in the mutant and WT, respectively; Fig 5B, S1 Data). This observation suggests that the major growth defect of the bacterial strain harboring an adenine at position 1,954 is caused by impaired protein synthesis. Interestingly, the strain with a thymine demonstrated a comparable, and even higher, maximal protein production rate (109% in the mutant as compared to WT) as well as total protein production (110.1% in the mutant, as compared to WT; Fig 5B, S1 Data). These results were further validated by a bacterial in vitro translation assay using cell-free synthesis of green fluorescent protein (GFP) as a reporter (Fig 5C, S1 Data). Taken together, our experiments strongly support our hypothesis that mitochondrial 16S rRNA position 947 should either be a modified adenine or harbor a thymine or guanine to maintain proper mitoribosomal activity.
To interpret the structural basis of our functional observations, we modeled adenine, m1A, or uridine at 16S rRNA position 947 within the recently solved cryo-EM structures of the human and porcine 55S mitoribosome [17,18]. Interestingly, m1A947 resides within helix 71 (H71) in domain IV of the 39S large ribosomal subunit, which is located at the subunit interface near intersubunit bridge B3, where interactions with the 28S small ribosomal subunit are formed (Fig 4). The loop region of H71 interacts with H92 in 16S rRNA domain V to form an interdomain interaction (S4 Fig), which is likely stabilized by the methylation of U2552 of H92 in E. coli (human mitochondrial U1369, porcine mitochondrial U1373) [32]. Accordingly, m1A947 may also be involved in stabilizing the tertiary rRNA interactions in this region of the ribosome. m1A may do so by forming stabilizing electrostatic interactions between the positive charge induced by N1-methylation and the negatively charged phosphate rRNA backbone of H64 (Fig 4). The unmodified adenine, in contrast, lacks this positive charge and may not be able to provide these stabilizing interactions. These observations raise the possibility that m1A enrichment in mitochondrial ribosomes is due to its contribution to the formation of more stable ribosomal subunits. In the case of the E. coli ribosome, which contains a guanine at the corresponding position, the exocyclic amino group of this guanine residue may act as a hydrogen bond donor to the backbone of the 23S rRNA.
Examination of the model (and the superimposed E. coli ribosomal structure) shows that in addition to m1A, uridine may also form some stabilizing interactions that an unmodified adenine is unable to provide. This possibly explains why 10% of vertebrates harbor only thymine in their mtDNA and uridine in the corresponding 16S rRNA. More specifically, due to the size difference between adenine and uridine, the uridine may interact with the rRNA backbone indirectly via a water molecular bridge, reducing the effect of the partial negative charge of its exocyclic oxygen atoms. Hence, stabilization by the uridine may occur via a different mechanism from the fully-charged m1A modification.
Here we report that the non-canonical RDDs in mitochondrial 16S rRNA position 947 result from an m1A modification generated by the mitochondrial RNA methyltransferase TRMT61B. Thereby, our finding sets the basis for assessing the prevalence of the underlying m1A modification echoed by A-to-U/G RDDs throughout the genome [33]. TRMT61B introduces m1A modifications in position 58 of the T-loops of six tRNA species in bovine mitochondria (S5A Fig) [16,24]. Thus, similar to the bacterial RlmN methyltransferase [34], TRMT61B modifies both tRNA and rRNA and, hence, constitutes the first vertebrate methyltransferase that modifies both tRNA and rRNAs.
Alignment of the T-loops of these six tRNAs revealed a weak consensus sequence (YMRAW) surrounding m1A58 (S5B Fig), which is also present in the loop capping 16S rRNA H71 (UAAAU) (S5B Fig). Close inspection of the tRNAs and the 16S rRNA revealed similarity in the structure of the backbone loop and orientation of the methylated base (S6 Fig). In vitro reconstitution of m1A947 using recombinant TRMT61B in combination with total RNA extracted from siRNA-treated HeLa cells revealed the capability of TRMT61B to recognize deproteinized naked 16S rRNA as a substrate to introduce m1A947 in Helix 71. This finding indicates that m1A947 is likely introduced at the early assembly stage of the mitoribosomal 39S subunit. Therefore, TRMT61B likely recognizes its tRNA and rRNA targets by a similar molecular mechanism.
Notably, we observed partial reduction in the methylation state of 16S rRNA position 947 in human cells (Fig 2). Such a partial phenotype is likely due to the stability of tRNA and rRNA molecules, which have relatively long half-lives as compared to mRNA molecules. Thus, pools of tRNA and rRNA molecules are not completely replaced with newly-synthesized hypomodified ones during cultivation after knockdown, even if the knockdown efficiency is quite high. Alternatively, although TRMT61B is the first methyltransferase that introduces an m1A modification in both mitochondrial tRNA and 16S rRNA, we cannot exclude the possible involvement of other additional enzymes. Therefore, our results underline the importance of a future detailed analysis of the mitochondrial m1A methylation mechanism and possible screen for its underlying components.
Our study revealed three changes that have occurred independently at the loop capping rRNA H71, which served as convergent evolutionary solutions for the ribosomal large subunit to allow formation of fully active ribosomes: one solution is generated post-transcriptionally (the 16S rRNA 947 m1A modification), and the two other solutions occurred at the DNA level (thymine in 10% of the vertebrate mitoribosomes and guanine in most bacteria). Our structural analysis suggests that in contrast to the unmodified adenine, the presence of m1A947 or an unmodified uridine in this position of the mitoribosomal 16S rRNA or a guanine in the corresponding position of the 23S bacterial rRNA may create stabilizing interactions within the ribosome, thus likely explaining their importance for protein translation in mitochondria and bacteria. Once a reconstituted in vitro human mitochondrial translation system is available, one will be able to assess the functional impact of mitochondrial 16S rRNA mutants, wild-type, and RNA-modified molecules and study their importance for mitochondrial translation.
It is intriguing that most vertebrate mitochondrial ribosomes rely on an rRNA modification, while a nucleotide compatible with a fully active ribosome is already encoded by the bacterial gene of the large rRNA subunit, the cytosolic ribosome in eukaryotes, and in a subset of vertebrate mitochondria. Intuitively, mutations at the DNA level (the vertebrate thymine and bacterial guanine) seem like more elegant solutions, eliminating the need for RNA modification at 16S rRNA position 947. Since the adenine (which is modified at the RNA level) was retained in 90% of the vertebrates, there might be selective pressure in favor of this base at the DNA level. Three possible explanations emerge: (A) The mitochondrial 16S rRNA transcript has a second role in addition to its role in the mitoribosome, which requires the 947A. (B) Random mutagenesis led to an adenine, which, in turn, has been retained due to co-occurrence with flanking sequences that together completed the recognition motif of TRMT61B. (C) The adenine is maintained at this mtDNA position because it strengthened a yet-to-be defined regulatory element. Interestingly, recent findings may favor the third hypothesis: Recently, ChIP-seq and DNase-seq analyses enabled us to identify candidate regulatory elements even within coding mtDNA sequences, thus suggesting a dual role for such sequences [35]. Hence, the assessment of the putative regulatory impact of this region (with or without the mutations) merits further investigation.
In summary, our findings unveiled that the previously reported RDDs at 16S rRNA position 947 marked an m1A RNA modification introduced by the mitochondrial methyltransferase TRMT61B. As this modification is present in most 16S rRNA in the mature mammalian mitoribosome and occurs throughout vertebrate evolution, modification of position 947 is most likely important for mitoribosomal structure and function. In agreement with this idea, our bacterial model experiments indicate that in the absence of methylation, the adenine at position 947 had to be mutated (A-to-T) at the DNA level to enable translation and cell growth, as also observed in 10% of vertebrate species. As G in this position is likely the ancestral state, and is compatible with efficient protein translation, m1A or U likely evolved later at this position to meet the specific requirements of mitochondrial function. Finally, it is intriguing that the expression of TRMT61B is altered in Alzheimer’s disease, thus suggesting altered levels of mitochondrial tRNA and rRNA modifications in this disease [36]. In summary, three alternative evolutionary solutions (i.e., RNA modification and either of two DNA bases) were selected to maintain ribosomal function in bacteria and mitochondria.
HeLa cells were grown in Dulbecco’s modified Eagle medium (DMEM) supplemented with 10% fetal bovine serum at 37°C, under a humidified atmosphere with 5% CO2.
The C-terminal Flag-tagged MRPL44 (MRPL44-Flag) expression vector was generated by LR reaction of Human Gateway Entry Clone FLJ12701AAAF with pDEST 12.2 Flag [37]. Approximately 2 × 108 HeLa cells were transfected with MRPL44-Flag expression vector using FuGENE HD (Roche). At 40 h post-transfection, the cells were harvested, and 39S subunit of mitoribosome was immune-precipitated by anti-Flag M2 agarose (Sigma) as previously described [38]. Co-precipitated RNA was extracted using TRI Pure (Roche), and mitochondrial 16S rRNA was resolved further purified by denaturing PAGE.
Isolated mitochondrial 16S rRNA was digested by RNase T1 at 37°C for 60 min in an 8-μl reaction mixture containing 20 mM ammonium acetate (pH 5.3) and 10 units/μl RNase T1 (Epicentre). Three quarters of the digested RNA fragments were analyzed by capillary liquid chromatography coupled with nanoelectrospray ionization linear ion trap-orbitrap hybrid mass spectrometer (LTQ Orbitrap XL, Thermo Fisher Scientific) [21,39].
Primer extension was conducted essentially as described previously [24]. The sequences of oligonucleotides are listed in S1 Table. The 5’ 32P-labeled primer (0.1 pmol) was mixed with 2.25 μg of total RNA in a 5-μL solution containing 10 mM Tris-HCl (pH 8.0) and 1 mM EDTA and incubated at 80°C for 2 min, followed by cooling down to room temperature for annealing. Then, the mixture was mixed with a 4.5-μL solution containing 2 μL of 5× FS buffer (Invitrogen), 0.25 μL of 1.5 mM dATP, dTTP, dCTP, and ddGTP mix, 0.75 μL of ddH2O, and 1.5 μL of 25 mM MgCl2. Upon addition of 0.5 μL of SuperScript III (Invitrogen), the reverse transcription was carried out for 1 h at 55°C for 16S rRNA and 47°C for tRNALeu(UUR). To terminate the reaction and digest the template RNA, the mixture was added with 0.5 μL of 4M NaOH and boiled at 95°C for 5 min, then neutralized by adding 4.5 μL of 1 M Tris-HCl (pH 5.0). The cDNAs were analyzed by 20% PAGE with 7M urea. The gel was exposed to an imaging plate, and the radiolabeled bands were visualized by FLA-7000 (FujiFilm).
Knockdown of target genes using RNAi was basically preformed as described previously [24]. The siRNAs used here are listed in S2 Table. The RNAi efficiency was checked by real-time RT-qPCR with a set of primers listed in S2 Table. Total RNA (2 μg) extracted from the knockdown cells was treated with 2 U of RQ1 DNase (Promega) to remove genomic DNA in a 20 μl 1×Reaction Buffer (Promega) at 37°C for 30 min, followed by adding RQ1 DNase stop solution (Promega) and incubated at 65°C for 15 min. The DNase-treated total RNA (2 μg) was incubated at 65°C for 5 min in a 10 μL solution containing 2.5 μM oligo(dT18) primer, 60 μM random N6 primer, and 1 mM dNTPs, then cooled on ice. Subsequently, 10-μL mixture containing 2×Transcriptor RT reaction buffer (Roche), 10 U RNase inhibitor (Roche), and 5 U Transcriptor RTase (Roche) was added to the solution. The cDNAs were synthesized in the mixture by sequential incubation at 25°C for 10 min, at 55°C for 30 min, and at 85°C for 5 min. The PCR was performed in a 20 μL mixture containing a 1 μl aliquot of the cDNA solution, 0.2 μM of each PCR primers, and 1×KAPA SYBR FAST Master Mix optimized for LightCycler480(Kapa biosystems). The thermal cycling conditions included 45 cycles of 95°C for 10 s, 58°C for 20 s, and 72°C for 1 s. Amplification of cDNA was monitored by LightCycler 480 (Roche).
His-tagged human recombinant TRMT61B was expressed in E. coli and purified by Ni-NTA chromatography as described [24]. A fraction containing the recombinant TRMT61B was dialyzed overnight against a buffer composed of 50 mM Hepes-KOH (pH 7.5), 50 mM KCl, 5 mM MgCl2, 10% glycerol, and 7 mM 2-mercaptoethanol. Recombinant TRMT61B was further purified by anion exchange chromatography using Mono-Q (GE healthcare) at pH 7.5 and 50–1000 mM KCl gradient. The concentration of the purified protein was determined by the Bradford protein assay (Bio-Rad) using bovine serum albumin as a standard.
In vitro reconstitution of m1A947 was carried out essentially as described previously [24].
The 114-mer RNA segment including Helix 71 (G866-U979) of human mitochondrial 16S rRNA was transcribed in vitro using T7 RNA polymerase. Template DNA with T7 class III promoter for the 114-mer RNA segment was prepared by assembling the following DNA sequences:
5’-gctaatacgactcactataggcaccgcctgcccagtgacacatgtttaacggc-3’,
5’-gtgacacatgtttaacggccgcggtaccctaaccgtgcaaaggtagcataatcac-3’,
5’-gtgcaaaggtagcataatcacttgttccttaaatagggacctgtatgaatggctccacgagggtt-3’,
and 5’-aaccctcgtggagccattc-3’. In vitro transcription by T7 RNA polymerase was performed as described [40]. The transcript was purified by denaturing PAGE and quantified by measuring the optical density at 260 nm. The reaction mix (50 μL), consisting of 25 mM Hepes-KOH (7.5), 100 mM KCl, 2.5 mM MgCl2, 1 mM DTT, 1 mM Ado-Met, 1.5 μg total RNA, and 1 μM His-tagged TRMT61B, was incubated for 2 h at 37°C, followed by adding phenol-chloroform isoamylalcohol (Nacalai) to terminate the reaction. Total RNA was recovered by ethanol precipitation and subjected to primer extension as described above.
RNA-Seq data (publicly available from Sequence Read Archive [SRA]) from seven vertebrate species was analyzed: (M. musculus [SRR579545], M. domestica [SRR306744], O. anatinus [SRR306726], A. carolinensis [SRR579556], G. gallus [SRR579551], X. tropicalis [SRR579560], and T. nigroviridis [SRR579565]) [41,42]. Additionally, we sequenced the following samples: RNA from HeLa cells in which TRMT61B was silenced (SRR3964513) and control cells (SRR3964514), purified total RNA (SRR3963545) or DNA (SRR3963556) from isolated S. scrofa heart mitochondria, isolated mitoribosome from the same S. scrofa sample (SRR3963521), and C. chamaeleon RNA extracted from whole blood (SRR2962875).
Porcine mitochondria were prepared from a liver sample extracted from a single freshly slaughtered pig (S. scrofa). The preparation was done as previously described by Greber et al. [43]. To avoid contamination from other liver samples, only one liver was processed at a time. The volumes of buffers used at different steps during the preparation were reduced according to the protocol that indicates the volumes for a preparation of five livers.
The preparation of the 55S mitoribosome follows the procedure previously described by Greber et al. [43]. Sixty-six grams of frozen mitoplasts were thawed in 150 ml lysis buffer (20 mM HEPES-KOH, pH 7.6, 100 mM KCl, 20 mM MgCl2, 1 mM dithiothreitol [DTT], 125 μM spermine, 125 μM spermidine) and brought to a total volume of 225 ml with monosome buffer (20 mM HEPES-KOH, pH 7.6, 100 mM KCl, 20 mM MgCl2, 1 mM DTT). Twenty-five milliliters Triton X-100 buffer (monosome buffer with 16% [v/v] Triton X-100) were added and the solution was stirred for 15 min at 4°C before homogenization using a Dounce homogenizer. The suspension was centrifuged (SLA-1500, 13,000 rpm, 20 min, 4°C), and the supernatant was PEG precipitated in 5% (w/v) PEG 10,000 for 15 min. The precipitate was collected by centrifugation (Sorvall SLA-3000 [Thermo Fisher Scientific], 2,500 g, 7 min, 4°C). Each pellet was re-suspended in 35 ml monosome buffer (2 h shaking), and the suspension was homogenized using a Dounce homogenizer before centrifugation using a Beckman (Beckman-Coulter) Type 45Ti rotor (28,000 rpm, 17 min, 4°C). The supernatant was loaded onto 50% (w/v) sucrose cushions (15 ml) and centrifuged (Beckman Type 70Ti [Beckman-Coulter], 50,000 rpm, 24 h, 4°C). Pellets were dissolved in 500 μl monosome buffer (shaking 230 rpm, 1 h) and cleared (tabletop centrifuge, 16,000 rpm, 20 min, 4°C). The sample was distributed onto 10%–40% (w/v) sucrose gradients (1 ml per gradient) and centrifuged (Beckmann SW-32 Ti [Beckman-Coulter], 26,000 rpm, 12 h, 4°C). The gradients were fractionated, and fractions corresponding to the 55S mitoribosome were collected and pooled. 55S mitoribosomes were pelleted using a Beckman (Beckman-Coulter) Type TLA-55 rotor (50,000 rpm, 6 h, 4°C). The supernatant was immediately discarded, and the pellets were flash-frozen in liquid nitrogen.
DNA was extracted using the Genomics DNA Extraction Mini Kit (RBC Bioscience), and RNA was extracted using the Perfect Pure RNA Cell and Tissue Kit (5 PRIME), following the manufacturers’ protocol. RNA and DNA were extracted from isolated mitoplasts, and RNA was purified from 55S mitoribosomes and blood of a single C. chamaeleon. The C. chamaeleon sample was collected as part of a different study in our lab [44]. The sample was collected and returned to its capturing site (UTM coordinates: 681839.21E/3597036.47N) using permits from the Israel Nature and Parks authority, number 2013/40003, and was approved by the animal experiments board at Ben-Gurion University number IL-18-03-2012.
One microgram of total RNA was subjected to cDNA synthesis using the iScript cDNA Synthesis Kit (Bio-Rad), following the manufacturer’s protocol.
DNA libraries were prepared using the Nextera XT DNA Sample Preparation Kit (Illumina). RNA libraries were prepared using the TruSeq RNA Kit (Illumina) according to the manufacturer’s protocol. TRMT61B-si and control-si RNA samples were sequenced using the Illumina HiSeq 2500 platform (Technion Genome Center, Israel) with 50-nt single-end reads. Both S. scrofa DNA and RNA libraries were sequenced using the MiSeq platform (Illumina). DNA and RNA libraries were sequenced using 151-nt or 300-nt, paired-end reads. C. chamaeleon libraries were sequenced on Hi-Seq 2000 platform (Illumina) using 101-nt paired end reads.
Sequencing reads were aligned against the publicly available mtDNA sequence of each species (M. musculus: NC_005089.1, M. domestica: NC_006299.1, O. anatinus: NC_000891.1, A. carolinensis: NC_010972.2, G. gallus: NC_007236.1, X. tropicalis: NC_006839.1, T. nigroviridis: NC_007176.1, S. scrofa: NC_000845.1, and C. chamaeleon: JF317641.1). For multiple sequence alignment, we utilized BWA [45] following the default protocol of the 1,000 Genome Sequence Analysis (ftp.1000genomes.ebi.ac.uk/vol1/ftp/). Only reads that were aligned to the corresponding mtDNA were used for further analyses. SAMtools [46] was used to convert the SAM to the BAM sequence format. MitoBam Annotator [47] was used to identify secondary read changes in the corresponding RNA sample. The orthologous sequences of the human mtDNA sequence position 2617 in each of the tested species were identified and analyzed. Secondary read changes were considered high quality only if identified in at least 1,000 high-quality sequence reads (filter A, except M. domestica), if their minimal read fraction was at least 1.6% (i.e., 0.8% from the reads of each of the strands [filter B] [48], and after manual inspection using the Integrative Genomics Viewer [49] to exclude mutations at the edges of the reads (S7 Fig).
In order to mutate the 947 orthologous position in all seven rRNA (23S) genes (position 1954), we used the E. coli strain EcM2.1 (a specially designed strain for high MAGE efficiency) to carry out three successive MAGE cycles as previously described [31]. We used two 90bp single-strand oligonucleotides to target the lagging strand of all seven genes. The first oligo was used to replace the endogenous G with an A: G*T*GGAGACAGCCTGGCCATCATTACGCCATTCGTGCAGGTCGGAATTTACCCGACAAGGAATTTCGCTACCTTAGGACCGTTATAGTT*A*C, and the second to replace the endogenous G with a T: G*T*GGAGACAGCCTGGCCATCATTACGCCATTCGTGCAGGTCGGAAATTACCCGACAAGGAATTTCGCTACCTTAGGACCGTTATAGTT*A*C. The mutated base is underlined. Asterisks represent phosphorothioate bonds. Briefly, cells were grown overnight at 34°C. Then, 30 μl of the saturated culture was transferred into fresh 3 ml of LBL medium until reaching OD = 0.4 and then moved to a shaking water bath (350 RPM) at 42°C for 15 min, after which it was moved immediately to ice. Next, 1 ml was transferred to an Eppendorf tube, and cells were washed twice with ddW at centrifuge speed of 13,000 g for 30 s. Next, the bacterial pellet was dissolved in 50 μl of DDW containing 2 μM of SS-DNA oligo and transferred into a cuvette. Electroporation was performed in 1.78 kV, 200 ohms, 25 μF. After electroporation, the bacteria were transferred into 2 ml of fresh LBL and incubated in 34°C until again reaching OD = 0.4 for an additional MAGE cycle.
To identify positive MAGE colonies (referred to as bacterial strains throughout the text), we PCR amplified two fragments encompassing the bacterial genomic regions (E. coli) orthologous to position 2,617 in all seven large rRNA (23S) genes. The amplified fragments correspond to E. coli genome (gi|556503834:4168641–4171544 [rrlB] positions 1,929–2,043 [fragment one] and to positions 1,929–2,333 [fragment two]). Restriction fragment length polymorphism (RFLP) was conducted on fragment one using MlucI (New England Biolabs—#R0538S) to identify the G-to-A or G-to-T mutations (both changes created a MlucI restriction site). Thus, complete restriction digestion of fragment one products implied genome editing (i.e., from wild type G to either T or A) in all seven copies of the 23S gene. To verify this interpretation, we amplified fragment two in samples showing complete digestion of fragment one. We then purified and sequenced those samples using primer 3 (S3 Table 3 and S4 Table). These sequences were aligned and visualized using Sequencher 4.10 (GeneCodes). Furthermore, after the initial screen, we PCR amplified each of the seven 23S rRNA genes by a set of specific primer combinations (S3 and S4 Tables) to ensure that the resulting modified strains harbor the desired mutation in all seven genes. To this end, the gene-specific templates were created by 50X dilution of the original PCR product of fragment 1, from which 1 μl was used as template for a new PCR amplification using primer pairs specific to each of the seven 23S genes. The resulting gene-specific amplification products were subjected to restriction digestion by MlucI, as mentioned above. All primers, PCR, and RFLP reactions and conditions are described in S3 Table 3 and S4 Tables. PCR and RFLP products were visualized by an EtBr-stained 1% agarose gel. PCR fragments were purified using Wizard SV Gel and PCR Clean-up system (Promega), following manufacturer’s protocol, and sequenced at the BGU sequencing core facility.
Cultures were grown for 48 h in LB medium, back diluted in a 1:100, ratio and dispensed on 96-well plates. Wells were measured for optical density at OD600, and measurements were taken during the growth at 30min intervals until reaching stationary phase. Qualitative growth comparisons were performed using 96-well plates (Thermo Scientific). For each strain, a growth curve was obtained by averaging over 48 wells.
Strains were transformed with the plasmid pZS*11-YFP-CGC-Kan harboring a YFP gene and Kan resistance cassette. YFP was measured as described in the section “Liquid Growth Measurements.” YFP production rate was measured by subtracting the YFP value at time t by time t-1 and dividing this value by the OD value at time t. Maximal production rate was defined as the highest value of this graph, and total production is the area under it.
Each of the MAGE-treated E.coli strains has been subjected to a 30S cell extract protocol [50]. All strains were grown to O.D600 2.0 ± 0.05, then lysed according to the 30S cell extract protocol while carefully maintaining all the strains under the same exact conditions throughout all processes. Next, cell extracts were used for a cell-free protein synthesis assay using EGFP fluorescence as a reporter. The assay was conducted in a Nunc 384 (120 μL) well plates (Thermo Fisher Scientific, Waltham, MA) and was monitored using time-dependent florescence measurements using a plate reader (Excitation 485 nm, Emission 525 nm). A typical cell-free reaction assay consists of 10 μL reaction mixture containing 33% (by volume) E. coli cell extract, and 66% of the reaction volume is composed of the reaction buffer containing nutrients, metabolites, and crowding agents. The reporter plasmid (pBEST-OR2-OR1-Pr-UTR1-deGFP-T500 [Addgene #40019]) is finally added to final concentration of 2nM. For detailed methodology, please see [51].
|
10.1371/journal.pcbi.1002240 | Most Random Gene Expression Signatures Are Significantly Associated with Breast Cancer Outcome | Bridging the gap between animal or in vitro models and human disease is essential in medical research. Researchers often suggest that a biological mechanism is relevant to human cancer from the statistical association of a gene expression marker (a signature) of this mechanism, that was discovered in an experimental system, with disease outcome in humans. We examined this argument for breast cancer. Surprisingly, we found that gene expression signatures—unrelated to cancer—of the effect of postprandial laughter, of mice social defeat and of skin fibroblast localization were all significantly associated with breast cancer outcome. We next compared 47 published breast cancer outcome signatures to signatures made of random genes. Twenty-eight of them (60%) were not significantly better outcome predictors than random signatures of identical size and 11 (23%) were worst predictors than the median random signature. More than 90% of random signatures >100 genes were significant outcome predictors. We next derived a metagene, called meta-PCNA, by selecting the 1% genes most positively correlated with proliferation marker PCNA in a compendium of normal tissues expression. Adjusting breast cancer expression data for meta-PCNA abrogated almost entirely the outcome association of published and random signatures. We also found that, in the absence of adjustment, the hazard ratio of outcome association of a signature strongly correlated with meta-PCNA (R2 = 0.9). This relation also applied to single-gene expression markers. Moreover, >50% of the breast cancer transcriptome was correlated with meta-PCNA. A corollary was that purging cell cycle genes out of a signature failed to rule out the confounding effect of proliferation. Hence, it is questionable to suggest that a mechanism is relevant to human breast cancer from the finding that a gene expression marker for this mechanism predicts human breast cancer outcome, because most markers do. The methods we present help to overcome this problem.
| Proving that research findings from in vitro or animal models are relevant to human diseases is a major bottleneck in medical science. Hundreds of researchers have suggested the human relevance of oncogenic mechanisms from the statistical association between gene expression markers of these mechanisms and disease outcome. Such evidence has become easier to obtain recently with the advent of microarray screens and of large public-domain genome-wide expression datasets with patient follow-up. We demonstrated that in breast cancer any set of 100 genes or more selected at random has a 90% chance to be significantly associated with outcome. Thus, investigators are bound to find an association however whimsical their marker is. For example, we could establish outcome associations for a signature of postprandial laughter and a signature of social defeat in mice. Association was not stronger than expected at random for 28 (60%) of 47 published breast cancer signatures. The odds of association are 5–17% with random single gene markers—a finding relevant to older breast cancer studies. We explained these results by showing that much of the breast cancer transcriptome is correlated with proliferation, which integrates most prognostic information in this disease.
| Ethics limits experimental investigation on human subjects. Hence, most experimental biomedical research is performed on animal and/or in vitro models. Proving that findings from model systems are relevant to human health is a major bottleneck.
Hundreds of studies in oncology have suggested the biological relevance to human of putative cancer-driving mechanisms with the following three steps: 1) characterize the mechanism in a model system, 2) derive from the model system a marker whose expression changes when the mechanism is altered, and 3) show that marker expression correlates with disease outcome in patients—the last figure of such paper is typically a Kaplan-Meier plot illustrating this correlation.
Breast cancer has been a test bed in oncogenomics. Several landmark studies (reviewed in ref. [1]) uncovered multi-gene mRNA markers of disease recurrence, which are independent of classical clinical markers and may provide useful information to guide treatment. These clinically motivated multi-genes markers, also called signatures, were derived from compendia of genome-wide mRNA tumoral profiles by selecting genes whose expression correlated with outcome [2]–[5], or with known aggressiveness markers such as proliferation [6]–[9] or grade [10]–[12].
Beyond clinical utility, many signatures were derived as markers of specific mechanisms and/or biological states and their association with outcome was evaluated in the context of studies structured along the 3-steps outlined above. These include signatures of stem cells [13]–[15], aneuploidy [16], wound healing [17], [18], hypoxia [19], [20], stromal component [21], epithelial-mesenchymal transition [22]–[24]; of mutations in TP53 [25], ALK5 [26]; of loss of PTEN [27]; of perturbations of E2F1 [28], bromodomain 4 [29], mir31 targets [30], p18ink4c [31], retinoic acid receptor [32]; of anchorage-independent growth [33], activation of modules related to the proteasome and mitochondrions [34], etc. Contrasting with this diversity, meta-analyses of several outcome signatures have shown that they have essentially equivalent prognostic performances [35], [36], and are highly correlated with proliferation [7]–[8], [37], a predictor of breast cancer outcome that has been used for decades [38]–[40].
This raises a question: are all these mechanisms major independent drivers of breast cancer progression, or is step #3 inconclusive because of a basic confounding variable problem? To take an example of complex system outside oncology, let us suppose we are trying to discover which socio-economical variables drive people's health. We may find that the number of TV sets per household is positively correlated with longer life expectancy. This, of course, does not imply that TV sets improve health. Life expectancy and TV sets per household are both correlated with the gross national product per capita of nations, as are many other causes or byproducts of wealth such as energy consumption or education. So, is the significant association of say, a stem cell signature, with human breast cancer outcome informative about the relevance of stem cells to human breast cancer?
Resolving this issue has become more pressing recently. Several large cohorts with genome-wide tumoral expression profiles and patient follow-ups are available in the public domain. Servers resting on these data [41], [42] make step #3 accessible to anyone with an Internet connection. Genome-wide expression profiling has also considerably lowered the barrier to step #2. The search for markers is reduced to a nearly automated screen by comparing microarray profiles in situations where the putative cancer-driving mechanism is active or inactive. The end result is an increasing number of signatures.
Few studies using the outcome-association argument present negative controls to check whether their signature of interest is indeed more strongly related to outcome than signatures with no underlying oncological rationale. In statistical terms, these studies typically rest on H0 assuming a background of no association with outcome. The negative controls we present here prove this assumption wrong: a random signature is more likely to be correlated with breast cancer outcome than not. The statistical explanation for this phenomenon lies in the correlation of a large fraction of the breast transcriptome with one variable, we call it meta-PCNA, which integrates most of the prognostic information available in current breast cancer gene expression data.
In order to assess whether association with outcome was specific, we tested the association with breast cancer outcome of three signatures whose rationale does not suggest any connection with cancer: a signature of the effect of postprandial laughter on peripheral blood mononuclear cells [43], a signature of skin fibroblast localization [44] and a signature of social defeat obtained from mice brains [45]. For the sake of simplicity, and because this is the most commonly used setup in the field, we focused on the 295 patients of the Netherlands Cancer Institute (a.k.a. NKI) cohort [2] and the overall survival end-point. Details on the procedure used to estimate association with outcome are provided in Supporting Information (Text S1). Surprisingly, the three control signatures were significantly associated with outcome (Figure 1, panels A–C).
To check that these were not anecdotal observations, we downloaded all signatures from MSigDB database [46] belonging to the c2 category and assessed their association with outcome. MSigDB c2 signatures are manually curated from the literature on gene expression and also include gene sets from curated pathways databases such as KEGG. Trivial single-gene signatures were removed. The 1890 signatures examined in MSigDB c2 encompass all the fields of biomedical sciences, nevertheless we discovered that 67% of them were associated with breast cancer outcome at p<0.05, 23% at p<10−5 (Figure 1D).
Cancer is a major subject matter of biomedical research, thus MSigDB c2 may be enriched for cancer-related signatures. To rule out the potential effect of a cancer bias, we generated for each signature in MSigDB c2 a signature of identical size but selected its genes randomly in the human genome. Although they are completely devoid of any biological rationale, 77% of these signatures were associated with outcome at p<0.05, and 30% at p<10−5 (Figure 1D).
Thus, nominal p-values should not be used directly because a signature associated with outcome with a significance of 10−5 and even more so, 0.05, is not more related to outcome than a random set of genes.
Although most random signatures are significantly associated with breast cancer outcome, the association could be much stronger for published breast cancer signatures and provide valid statistical support for their relevance.
We compiled 47 signatures from the literature. Association with outcome has been reported for most of them (Supporting Information, Text S1), either for the purpose of finding better prognostic tools, or, in most cases, to suggest biological relevance. We compared the outcome association of each signature to that of 1000 random signatures of identical size (Figure 2). We confirmed the outcome association of 42 in these 47 signatures. Yet, 11 of them (23%) showed a weaker association than the median of random signatures. Abiding to statistical standard, one may consider a signature biologically relevant if its association with outcome is stronger than the association of the best 5% random signatures. Only 18 signatures in 47 (40%) met this criterion.
Figure 2 reveals that larger signatures are more significant. More than 90% of the signatures >100 genes we generated were significant at p<0.05. For the two largest ones, 714 and 1345 genes respectively, all 1000 random signatures tested were significant.
At the other end of the size spectrum, we found that 26% of individual genes printed on the NKI arrays were associated with outcome at p<0.05. Thus, a single gene study has 26 chances in 100 to yield a significant association. When we applied a q-value correction [47]—relevant to genome-wide studies—17% of all genes were associated with outcome at q<0.05. A comparable calculation was presented by Ein-Dor et al. [48]: 1234 genes among 5852 that passed their initial filter were associated with outcome with a false discovery rate <10%.
Proliferation is a well-known breast cancer prognostic marker [38]–[40]. Cycling cells express thousands of specific genes [49], thus genome-wide expression profiles are likely to capture the fraction of cycling cells within a tissue. A proliferation cluster was noticed in early breast cancer microarray studies [50]–[52], and proliferation is the major variable behind gene expression-based breast cancer prognosis [7]–[9]. We devised a new metagene, meta-PCNA, in order to investigate further the role of proliferation.
The proliferating cell nuclear antigen, PCNA, is a ring-shaped protein that encircles DNA and regulates several processes leading to DNA replication [53]. As suggested by its name, this is one of the most widely used antigen target for immunohistochemical measures of the fraction of proliferating cells in tissues. Ge et al. [54] profiled with microarrays 36 tissues from normal, healthy, individuals encompassing 27 organs. We call ‘meta-PCNA’ the signature composed of the 1% genes the most positively correlated with PCNA expression across these 36 tissues (Table S1). In plain language, meta-PCNA genes are consistently expressed when PCNA is expressed in normal tissues and consistently repressed when PCNA is repressed. We define the meta-PCNA index as the median expression of meta-PCNA genes. Beside PCNA itself, meta-PCNA includes other canonical proliferation markers such as MKI67, TOP2A, MCM2, etc.
We next compared for each one of the 47 published signatures its association with outcome in the original NKI data set and after adjustment of expression levels for the meta-PCNA index (Figure 3, Kaplan-Meier plots in Supporting Information, Text S1). Their association with outcome dropped dramatically after adjustment, although a few signatures remained strongly outcome associated. Any transformation damaging expression data will trivially decrease the association between outcome and expression. To control that was not the case with our adjustment procedure we reran the same analysis, except that meta-PCNA values were permuted randomly among patients prior to adjustment. In contrast with the adjustment of the actual non-permuted index, outcome association was not affected (Supporting Information, Text S1).
We plotted the hazard ratios of the 47 signatures against the absolute correlation of their first principal component with the meta-PCNA index. The more a signature was correlated with meta-PCNA, the higher its hazard ratio (R2 = 0.9, Figure 4A, details for each data point in Supporting Information, Text S1).
Since only a limited set of genes is included in the 47 signatures, we plotted the distribution of correlations with the meta-PCNA index of all genes significantly associated with outcome and, as a negative control, of all genes printed on the microarrays (Figure 4B). Among the 17% of genes associated with outcome at q<0.05, 91% were significantly correlated with meta-PCNA. Thus, any predictor resting on a linear combination of genes associated with outcome has a high probability to be confounded by proliferation.
The potential confounding effect of proliferation has been recognized by a number of authors who attempted to rule it out by removing known proliferation genes from expression data [17], [14], [15]. These genes have been defined in various ways, including the Gene Ontology ‘cell cycle’ category, the genes periodically regulated in a cell-cycle time course [49], or genes of the breast cancer ‘proliferation cluster’ [55].
Following Ben Porath et al. [14], we defined as cell-cycle genes any gene present in at least one of these three categories. We calculated the distributions of correlations between the meta-PCNA index and genes of the Embryonic Stem Cell Module (ESCM) of Wong et al. [15], with and without the cell cycle genes (Figure 5). Purging these genes out of the ESCM did eliminate signals in the highest correlation range, but the ESCM remained vastly more correlated with meta-PCNA than the bulk of genes printed on the arrays (p = 10−25).
Moreover, 58% of the genes printed on the array were significantly correlated with the meta-PCNA index in the NKI cohort. Thus, the correlations with meta-PCNA extend far beyond cell cycle genes. Removing these genes fails to rule out the confounding effect of proliferation. Similarly, a signature does not have to be enriched with known cell cycle genes to convey a strong cell proliferation signal.
Previous sections rested on the NKI data set and the overall survival end-point. Are our observations specific of this popular, but not universal, setup? We reran the analyses using recurrence-free survival, and on another cohort [56] using both overall survival and relapse-free survival.
We calculated hazard ratios for the 47 published signatures using all combinations of end-points and cohorts. Correlation between hazard ratios among the different cohorts/end-points was ≥0.97 (Figure 6). Thus, the ranking of the signatures with respect to association with outcome was highly reproducible. However, the combination of NKI data and overall survival gave hazard ratios ∼1.3 units higher (median HR = 3.8 in NKI and OS vs. <2.5 in other setups). Accordingly, p-values were ∼4 orders of magnitude smaller than when association with outcome was estimated from the overall survival in the cohort of Loi et al. [56], although it included ∼30% more patients. This difference between the 2 cohorts is less marked with relapse-free survival. Nevertheless, our analysis (summarized Table 1) reveals that, irrespective of the specific setup, at least 40% of MSigDB c2 signatures and 5% of all genes are associated with outcome, and at most 40% of the 47 published signatures are better than the 5% best same-size random signatures.
There are many ways to estimate association between the expression of a multi-gene marker and disease outcome, and different studies have taken different routes. Our goal to compare signatures and assess them against negative controls, however, implied a uniform statistical framework. We present a comparison of a number of such methods in the Supporting Information (Text S1). A popular approach used in the studies we reviewed consists in stratifying the patients by hierarchical clustering in the signature subspace [57], [21], [29], [24], [28], [15], [58]. In most cases, our method of choice (using the first component of a Principal Components Analysis of a signature as a prognostic score) reveals stronger outcome associations than this approach. Our method is validated by the fact that we could reproduce the outcome association of most published signatures, which, conversely, validates the prognostic value of those signatures. The choice of association method is of course important, as there is a possibility that it misses some signals captured by specific combinations of signatures and models. However, most papers use similarly simple methods as ours. Furthermore, the strength of such association might be doubted if it depended on an elaborate algorithm, as it is likely to be caused by spurious signals arising from model selection biases.
The main message of this paper is that, if the purpose of a study is to assert the biological relevance to human cancer of a signature, the association between this signature and outcome cannot rest on the nominal p-values, as obtained on breast cancer by the Cox analysis. This follows from elevated likelihood that random sets of genes are related to the outcome. Thus, an investigator finding that her/his signature is associated with outcome with a significance of 10−5, and even more so, 0.05, gains no specific information because sets of random genes would likely yield similar, or better, results. Nominal p-values do not answer the appropriate statistical question: the question is not whether a given set of genes is related to survival, but whether it is more related to survival than random sets of genes.
This problem extends to single-gene markers and therefore to many studies published in the pre-genomic era. Claims similar to those concerning signatures have been made, that single genes, important in a model system, are relevant for human cancer progression based on differential expression between short- and long-survival groups. As 26% of the genes are related to survival at p<0.05 (17% at q<0.05), much tighter p-values than commonly used should be imposed to demonstrate such a relation.
Several studies in the panel of 47 we investigated developed arguments independent of outcome association. For example, Hu et al. [59] used outcome association not as a validation argument, but as an exploratory tool to discover driver DNA copy number aberrations, which were then directly investigated. However, most of these studies, and many more not reviewed here, extrapolated the results from animal or highly artificial in vitro models to human in vivo cancer on the basis of questionable association statistics alone.
The present study addresses purely correlative association between gene-expression and disease outcome. We have shown that proliferation integrates most of the prognostic information contained in the breast cancer transcriptome. Yet—we cannot stress this enough—we have not shown that proliferation is a core driving force behind breast cancer progression. Disentangling the role of a biological process in cancer progression in vivo from the role of proliferation and from the role of the other processes associated with it is a crucial issue. The adjustment methodology we propose may be useful in assessing whether markers of biological processes do or do not rest on association with proliferation. Our results also imply that such markers should be evaluated against the outcome association of comparable negative control markers.
Our study questions the biological interpretation of the prognostic value of published breast cancer signatures, but has no bearing on their usefulness in the clinic: a marker may be accurate without yielding interesting biological insight regarding the mechanism of disease progression. Nevertheless, the prominence of proliferation should be taken into account in future clinical research. Are there transcriptional signals in breast cancer that are prognostic, but independent of proliferation? Is there any hope to perform better than the 70 genes NKI signature [2]? The studies we reviewed assessed outcome prediction from gene expression measured in bulk tumors sampled from a relatively wide spectrum of patients, thus prognostic transcriptional signals detectable in specific tumor cells and/or specific patient groups were out of scope. Yet, proliferation-related signals are prognostic mostly in ER+ tumors [1]. Immunological genes convey prognostic information in ER- tumors and in tumors with HER2 amplification [8], [60]–[64]. This information is unquestionably independent of proliferation since it improves prognostic accuracy beyond the abilities of proliferation-driven signatures and classical clinical markers [65]. Larger cohorts allowing the analysis of patients sub-groups and expression profiling of specific tumor cells/tumor areas may lead to better prognostic tools in the future.
In conclusion, we have shown that 1) random single- and multiple-genes expression markers have a high probability to be associated with breast cancer outcome; 2) most published signatures are not significantly more associated with outcome than random predictors; 3) the meta-PCNA metagene integrates most of the outcome-related information contained in the breast cancer transcriptome; 4) this information is present in over 50% of the transcriptome and cannot be removed by purging known cell-cycle genes from a signature.
All analyses were run with R 2.9.0 [66] with packages specified in the following sections. Functions were run with default parameters unless specified otherwise.
The code and data underlying the results and figures of this study are available as a Bzip2-compressed tar bundle from the PLoS Computational Biology web site (Dataset S1, size is 87 MB). The scripts assume a UNIX/LINUX environment.
All the data were available from public sources:
Probes mapping to the same genes were averaged in each one of the three datasets.
Whenever possible, the signatures were compiled from the publications online supplementary tables. When not available, the gene symbols were automatically read with an optical character recognition system from the papers tables and figures. In rare instances, signatures were encoded manually and double-checked. Because gene names and symbols are changing over time, the gene symbols of all signature genes were updated to match the HUGO nomenclature and therefore maximize the match with microarray gene annotations. HUGO gene symbols and their older aliases were obtained from the file gene_info as available on May 9th 2007 from the NCBI ftp server.
MSigBD 2.0 c2 signatures were downloaded as a *.gmt file from the Broad Institute page www.broadinstitute.org/gsea/msigdb/index.jsp.
We computed the Pearson correlation between PCNA and all the genes in the Ge et al. [54] dataset and selected the 1% most positively correlated, i.e., 131 genes out of 13,077, to form the meta-PCNA signature (Table S1). The meta-PCNA index of a tissue was computed from its expression profile by taking the median expression of these genes.
The expression of each gene was fitted with R's ‘lm’ function and each expression measurement was substituted by the sum of its residual and its mean expression across the cohort.
In order to systematically compare the published signatures to random signatures and evaluate the relation between outcome association and meta-PCNA, we needed an outcome association estimation procedure that is robust and fully automated. We systematically compared three procedures and selected among them the most sensitive and stable one. This is described in Supporting Information (Text S3.), only the selected method is described here. It consists in computing the first principal component (PC1) of the signature (with R's prcomp) and then split the cohort according to the median of PC1. Probes mapping to the same gene were averaged and, following Ramaswamy et al. [57], data were median polished (R's medpolish) before the dimension reduction step.
|
10.1371/journal.pcbi.1006558 | Suboptimal community growth mediated through metabolite crossfeeding promotes species diversity in the gut microbiota | The gut microbiota represent a highly complex ecosystem comprised of approximately 1000 species that forms a mutualistic relationship with the human host. A critical attribute of the microbiota is high species diversity, which provides system robustness through overlapping and redundant metabolic capabilities. The gradual loss of bacterial diversity has been associated with a broad array of gut pathologies and diseases including malnutrition, obesity, diabetes and inflammatory bowel disease. We formulated an in silico community model of the gut microbiota by combining genome-scale metabolic reconstructions of 28 representative species to explore the relationship between species diversity and community growth. While the individual species offered a broad range of metabolic capabilities, communities optimized for maximal growth on simulated Western and high-fiber diets had low diversities and imbalances in short-chain fatty acid (SCFA) synthesis characterized by acetate overproduction. Community flux variability analysis performed with the 28-species model and a reduced 20-species model suggested that enhanced species diversity and more balanced SCFA production were achievable at suboptimal growth rates. We developed a simple method for constraining species abundances to sample the growth-diversity tradeoff and used the 20-species model to show that tradeoff curves for Western and high-fiber diets resembled Pareto-optimal surfaces. Compared to maximal growth solutions, suboptimal growth solutions were characterized by higher species diversity, more balanced SCFA synthesis and lower exchange rates of crossfed metabolites between more species. We hypothesized that modulation of crossfeeding relationships through host-microbiota interactions could be an important means for maintaining species diversity and suggest that community metabolic modeling approaches that allow multiobjective optimization of growth and diversity are needed for more realistic simulation of complex communities.
| The gut microbiota serve a critical role in maintaining a healthy state in the human host. The gut contains approximately 1,000 bacterial species that provide a wide range of metabolic capabilities including the breakdown of dietary compounds and the synthesis of useful metabolites. The robust function of the gut community is intimately connected to its diversity, both with respect to the number of species and the relative abundance of those species. The gradual loss of diversity is a key element of microbiota dysbiosis, which has been correlated with a wide range of health problems including inflammatory bowel disease. To investigate species diversity in the gut microbiota, we developed an in silico community model by combining genome-scale metabolic reconstructions of 28 representative species from the most abundant genera in the human gut. Our model predicted that maximal community growth produced low species diversity and no synthesis of the health-promoting metabolite butyrate. After reducing the community model to 20 species, we showed that suboptimal community growth allowed much higher species diversity and butyrate synthesis more consistent with in vivo studies. The model predicted that increased diversity could be achieved through modulation of metabolite crossfeeding relationships between species, an experimentally testable hypothesis.
| The gut microbiota comprise a highly complex ecosystem that has been characterized as an additional organ within the human host [1, 2]. The microbiota form a mutualistic relationship with the host, with saccharolytic species enzymatically degrading complex carbohydrates into fermentable sugars and fermentative species converting sugars and other available nutrients into a variety of absorbable metabolites [3, 4]. A particularly important function of the microbiota is to ferment dietary fiber into the short-chain fatty acids (SCFAs) acetate, butyrate and propionate [5, 6]. While significant variations are possible depending on diet, the molar ratio of these three SCFAs is approximately 60:20:20 [7]. SCFAs are consumed by host colonocytes as a primary energy source, with butyrate being the preferred SCFA but acetate probably supplying more energy to its higher concentration in vivo.
The gut microbiota consist of approximately 1,000 species [8] and 7,000 unique strains [2] in a typical human host. The two dominant phyla in healthy humans are Firmicutes and Bacteroidetes, which comprise more than 90% of the community [9, 10]. Other important but much less abundant phyla are Proteobacteria, Actinobacteria, Euryarchaeota and Verrucomicrobia as well as Eukaryota such as fungi [11, 12]. Metagenomic studies have shown wide variations in bacterial composition in healthy humans [13, 14], demonstrating that microbiota composition is an individual characteristic and an inadequate measure for assessing gut health across patient populations.
A hallmark of healthy gut communities is high diversity [15, 16], both in terms of the species present and the relative abundance of these species [17, 18]. An integrated gene catalog developed from 1,267 sequenced samples and comprising almost 10 million genes provides detailed information on the number of genes from over 700 gut genera [19]. The high diversity of the gut microbiota is demonstrated by 75 genera having relative gene counts (genes counts in that genus divided by total gene counts) of at least 0.1%. Numerous studies have shown a strong correlation between bacterial diversity and health/disease states, with long-term loss of diversity a key characteristic of dysbiosis [20–22]. The loss of bacterial diversity has been implicated in broad range of diseases including Clostridium difficile infections [23], inflammatory bowel and Crohn’s diseases [24, 25], obesity [26], diabetes [27], cardiovascular disease [28], rheumatoid arthritis [29], colorectal cancer [30], cystic fibrosis [31] and depression [32].
Genome-scale metabolic modeling has emerged as an important tool for computationally interrogating the metabolism of microbial communities. A number of alternative methods for combining metabolic reconstructions of single species into community models are now available. These methods are invariably based on growth rate maximization, either with regard to the species individually [33, 34] or the community as a whole [35–37]. Compared to other techniques, the recently developed SteadyCom method represents an important advance by performing community flux balance analysis (FBA) to determine the relative abundance of each species for maximal community growth while ensuring that all metabolites are properly balanced within each species and the community [38]. While the usual FBA objective of maximal growth has been experimentally demonstrated for individual bacteria such as Escherichia coli [39, 40], limited data is available to support the adoption of maximal growth as a community objective [41–43]. Indeed this objective has the potential to favor the fastest growing species and produce communities with low diversity that are inconsistent with healthy gut communities observed in vivo. In this study, we utilized the SteadyCom method to investigate the tradeoff between community growth and species diversity for in silico communities comprised of 20 and 28 representative gut species.
Semi-curated genome-scale metabolic reconstructions for representative species within the 28 most abundant genera [19] in the human gut were obtained from the Virtual Metabolic Human database (vmh.uni.lu) [44]. These models represented the five major bacterial phyla (Actinobacteria, 2 species; Bacteroidetes, 4 species; Firmicutes, 15 species; Fusobacteria, 1 species; Proteobacteria, 6 species), including 10 species from the highly prevalent Firmicutes order Clostridia (Table 1). The 28 genera covered almost 85% of reference genes by occurrence frequency according to a recent integrated catalog of the human gut microbiota [19]. Each species was constrained according to either a Western or high-fiber diet [44] and assigned a non-growth ATP maintenance (ATPM) value of 10 mmol/gDW/h, which is within the range reported for curated bacterial reconstructions. Because the species biomass equations were not curated [44], all species models used the same growth-dependent maintenance energy. The community metabolic model was constructed from the single-species reconstructions using the createCommModel function provided within the SteadyCom suite of MATLAB tools [38]. The community model accounted for 22,203 genes, 26,867 metabolites and 35,031 reactions within the 28 species as well as 354 uptake and 354 secretion reactions for the extracellular space.
The uptake reactions in the extracellular space of the community model were constrained with the chosen diet as the union of all the maximum nutrient uptake rate constraints from the 28 species models. An Excel file with community uptake constraints for the four models (20 and 28 species; Western and high-fiber diets) is available in the Supporting Information (S1 File). With the exceptions noted below, the maximum nutrient uptake rates of each species were set equal to values defined by the chosen diet for that species. This approach ensured that each species would produce the same single-species growth rate with SteadyCom as obtained with standard flux balance analysis (FBA). Crossfeeding of all 21 amino acids and eight common metabolic byproducts (acetate, CO2, ethanol, formate, H2, D-lactate, L-lactate, succinate) was promoted by increasing the maximum nutrient uptake rates of these nutrients to 10 mmol/gDW/h. A single value was used due to the lack of metabolite- and species-dependent data for byproduct uptakes in the literature. While these constraints had no effect on single-species metabolism due to the extracellular constraints, they allowed species with different single-species growth rates to coexist through crossfeeding. The SCFAs butyrate and propionate were not allowed to be consumed under the assumption that crossfeeding of these two SCFAs was negligible compared to utilization by the host [45].
The community metabolic models were solved with the SteadyComCplex function within SteadyCom [38]. The IBM ILOG Cplex solver was used for linear program (LP) solution. SteadyCom performs community FBA by computing the relative abundance of each species for maximal community growth while ensuring that all metabolites are properly balanced within each species and the community. SteadyCom provides the capability to constrain the species abundances to explore various features of community behavior. Most simulations were performed with the default abundance constraints (lower bound zero, upper bound unity), which is referred to as the unconstrained case. Community flux variability analysis (FVA) was performed with respect to the species abundances using the SteadyComFVACplex function within SteadyCom. FVA showed that all maximal growth communities had unique species abundances. Using community FVA results for the 28-species community, species which could only coexist at 70% or less of the maximal community growth rate were eliminated to yield a 20-species community (Table 1). The 20-species model was used to explore the tradeoffs between community growth and diversity by constraining the species abundances with upper bounds computed from community FVA results obtained at different percentages of the maximal growth rate (see below).
Simulation results were analyzed with respect to the community growth rate, species abundances and diversity in the community, and the number and type of crossfeeding relationships. The growth rate and species abundances were direct outputs of SteadyCom. Species diversity was quantified using the Inverse Simpson equitability index [46–48] that accounted both for the number of participating species (e.g. richness) and the abundance of each species (e.g. evenness),
D c o m = 1 N 1 ∑ i = 1 N p i 2 (1)
where N is the total number of species and pi ∈ [0, 1] is the relative abundance of species i, another direct output of SteadyCom. The diversity measure Dcom varied from 1 N if the community has a single participating member to unity if the all species participated and had the equal abundances. While the community models were built from metabolic reconstructions of particular species, the species name and associated genus were used interchangeably since the modeling goal was to achieve diversity in the genera.
To investigate the tradeoff between community growth and species diversity, maximum abundances in the 20-species community model were constrained using FVA results as follows,
p j U = 2 r p j m a x ∑ i = 1 N p i m a x (2)
where p j m a x is the maximum abundance of species j calculated at a specified percentage of the maximal growth rate using FVA, r is a uniform random number in the closed interval [0 1] and p j U is the upper abundance bound of species j. The maximum abundance of species j was divided by the sum of the maximum abundances of all species such that the scaled maximum abundances summed to unity. The bounds were randomized to more completely sample the growth-diversity tradeoff and were generated subject to the constraint ∑ j = 1 N p j U ≥ 1 to avoid producing structurally infeasible LP problems. The multiplier 2 was introduced because r was uniformly distributed with expected value 0.5. As compared to completely random bounds, Eq (2) allowed species with higher FVA abundances to have higher upper bounds on average. These bounds tended to constrain the solution such that higher diversity than the optimal solution was obtained, especially when FVA solutions at lower growth rates (e.g. 60% of maximal) were used. Eq (2) is heuristic in the sense that the calculated upper bounds do not ensure Pareto optimality [49, 50] of the resulting solutions in the growth-diversity space. More sophisticated methods that allow the calculation of the maximal diversity at a given growth rate would be required for this purpose. For each diet, a total of 900 cases were performed at FVA growth rates 60–99% of the maximal value, with more cases run at lower growth rates (150 cases at 60% and 65%, 125 cases at 70% and 75%, 100 cases at 80%, 85% and 90%, 25 cases at 95% and 99%) to adequately sample higher diversities. When ∑ j = 1 N p j U was close to unity, SteadyCom often returned a solution where ∑ i = 1 N p i m a x was outside the default tolerance of 10−4 specified within SteadyCom. These solutions were discarded to maintain the same accuracy of all suboptimal growth solution, substantially reducing the total number of cases (e.g. 900 to 595 for the Western diet).
First the community model was constrained to investigate the metabolic capabilities of each species individually on the in silico Western diet. These single-species simulations were performed within SteadyCom by constraining the abundance of all other species to zero. The growth rate and the secretion rates of ten primary metabolic byproducts were determined for each species (Fig 1). The 28 species exhibited a wide range of growth rates, including three species (Escherichia, Enterobacter, Citrobacter) with growth rates exceeding 0.4 h−1 and four species (Bifidobacterium, Pseudoflavonifractor, Phascolarctobacterium, Megasphaera) with growth rates of zero for the ATP maintenance value of 10 mmol/gDW/h. While species with high individual growth rates were expected to have a competitive advantage in the in silico community, slower growing species had the possibility of coexisting by increasing their growth rates through metabolite crossfeeding. With regard to SCFA synthesis, 24 species secreted acetate, 7 species secreted butyrate including 3 major butyrate producers (Faecalibacterium, Eubacterium, Fusobacterium), and 13 species secreted propionate including 3 major propionate producers (Bacteroides, Veillonella, Parabacteroides). While the SCFA synthesis capabilities of the genera Bacteroides, Faecalibacterium and Eubacterium are well documented [51, 52], the other SCFA predictions also appear to be consistent with experimental studies [53, 54].
SteadyCom was used to determine the optimal growth rate and species abundances of the 28-member community on Western and high-fiber diets. The growth rate on each diet was similar (0.69 h−1 Western diet, 0.65 h−1 high-fiber diet) (Fig 2A) and appeared to be consistent with limited data available for in vivo gut community growth rates [55, 56]. Each community consisted of a small number of species, with only five species for the Western diet and six species for the high-fiber diet having non-zero abundances. The abundances of the 28 species correlated strongly with their single-species growth rates (Fig 2B; P < 10−4 for either diet), as would be expected from a community modeling methodology based on growth maximization. The dominant species included generally beneficial commensals from the genera Clostridium [57], Collinsella [58] and Coprococcus [59] but also represented several genera associated with inflammatory bowel disease (IBD) pathogenesis, including Escherichia [60], Enterobacter [61] and Citrobacter [62]. As a result of dominance by a few species, both communities exhibited low diversity, which has been correlated with a wide variety of gut pathologies [20–22]. Only slightly higher diversity was achieved for the Western diet when the ATP maintenance value of each species was tuned to the extent possible to achieve a uniform single-species growth rate of 0.2 h−1 (S1 Fig), demonstrating that the domination of particular species was partially attributable to their ability to more effectively exploit crossfed metabolites for growth.
The community model was formulated to allow crossfeeding of all 21 amino acids and eight metabolic byproducts. Ethanol, D-lactate, L-lactate and succinate were crossfed to the extent that their net secretion rates (difference between the sum of the species synthesis and uptake rates) on either diet were zero (not shown). Both diets produced relatively high formate and acetate net secretion rates (Fig 2C), while the rates of CO2, H2, propionate were comparatively low and no butyrate was produced. Predicted ratios of the acetate:butyrate:propionate rates of 91:0:9 and 94:0:6 for the Western and high-fiber diets, respectively, were inconsistent with reported in vivo SFCA levels, which commonly are in the range of 60:20:20 [7]. The SCFA imbalance predicted for both diets was attributable to acetate production by all participating species, a lack of propionate producers (only Clostridium and Collinsella), and the absence of acetate consumers and butyrate producers. The five most significantly cross-fed metabolites were predicted to be the amino acids aspartate and serine and the byproducts D-lactate, L-lactate and CO2 (Fig 2D). Clostridium, Escherichia and Enterobacter formed a mutualistic three-species subcommunity with both large uptake and secretion rates of the five metabolites. By contrast, Collinsella and Coprococcus exhibited commensal interactions by solely consuming the secretion products while Citrobacter did not participate in the crossfeeding of these metabolites.
In silico communities optimized for maximal growth exhibited a lack of species diversity and SCFA imbalance characterized by low butyrate levels, both of which are strongly correlated to gut disease [5, 6, 20, 53]. To explore species diversity and SCFA synthesis at suboptimal growth rates, SteadyCom was used to perform community flux variability analysis (FVA) with respect to the species abundances. For growth rates between 10% and 99.99% of the maximal value, the number of species abundances that could be maximized to exceed 1% of the community abundances (possible species) or could be minimized to be exceed 1% of the community abundances (essential species) were determined. Uniqueness of the maximal growth communities was indicated by the convergence of the number of possible and essential species to a single value at the maximal growth rate (Fig 3A). FVA produced unique species abundances for all maximal growth communities S1 Folder. While the Western diet could only support five species at the maximal growth rate, the possible community size increased to 13 species at 99% and to 18 species at 80% of the maximal growth rate. Nine species were predicted to be incapable of coexistence at growth rates greater than 70% of the maximal value (Fig 3B), including saccharolytic Alistipes and the common probiotics Lactobacillus and Bifidobacterium (Fig 3C). No species were essential until 96% of the maximal growth rate and all five species that comprised the optimized community were not essential until 99.92% (Fig 3D). Similar results were obtained with the high-fiber diet. These results suggest that substantially enhanced diversity may be achievable at even marginally suboptimal growth rates.
Based on the results in Fig 3B, the 28-species community was reduced to a 20-species community by removing eight species that could only coexist at growth rates less than 70% of the maximal value on both diets (Table 1). These species belonged to the genera Lactobacillus, Alistipes, Bifidobacterium, Parabacteroides, Pseudoflavonifractor, Phascolarctobacterium, Megasphaera and Acidaminococcus. This reduction in community size allowed a more efficient exploration of species diversity at growth rates above 70% of the maximal value, where the eliminated species were ensured not to coexist. Community FVA performed with the 20-species community suggested that 18 species could coexist at 80% and 15 species could coexist at 95% of the maximal growth rate on the Western diet (Fig 4A). Similar results were obtained for the high-fiber diet. By construction, all species could coexist at 70% of the maximal growth rate for at least one diet (Fig 4B).
We hypothesized that the tradeoff between community growth and species diversity could be described by a Pareto optimal surface [49, 50, 63–65] with increased diversity achieved only at the expense of reduced growth. However, community FVA did not necessarily provide a means to sample this tradeoff surface since diversity is not considered as part of the analysis. To investigate this issue, the 40 FVA solutions (20 solutions each for species abundance minimization and maximization) generated at each growth rate were used to compute the equitability measure Dcom. The average Dcom value for all growth rates and both diets was in the small range 0.11–0.22 (Fig 4C). Of the 2,720 FVA solutions tested, the largest Dcom was 0.43 and only nine cases produced Dcom > 0.35. In other words, the FVA solutions proved inadequate for generating high diversity, which was not surprising given that FVA solutions were computed by minimizing or maximizing a particular species abundance. This lack of diversity again translated to SCFA imbalance, with the average butyrate and propionate fractions over all 1,360 cases for the Western diet being 2% and 3%, respectively (Fig 4D).
While useful for generating bounds on achievable species diversity and SCFA production, community FVA did not provide a direct means to investigate the tradeoff between community growth and diversity. However, we found that this tradeoff could be explored effectively by using FVA solutions at a particular growth rate as inputs to Eq 2 for calculation of an upper bound on each species abundance. When FVA solutions at low growth rates (e.g. 60% of maximal) were used for constraint calculation, the resulting solutions tended to have relatively high diversity and low growth. Using 575 simulation cases performed with the 20-species community and the Western diet, the fraction of cases in which the abundance of each species exceeded 1% was calculated as a measure of species fitness over a wide range of growth rates (Fig 5A). The five species that comprised the community at the optimal growth rate were present in all 575 communities. Citrobacter and the high butyrate producer Fusobacterium were present in over 98% of communities, while Klebsiella and the high propionate producer Bacteroides were present in at least 80% of communities. By contrast, the high butyrate producer Faecalibacterium and and high propionate producer Veillonella were present in no more than 6% of communities. Similar results were obtained with the high-fiber diet (S2A Fig).
Results of the 575 Western diet cases were collected into 15 bins in the growth rate space. While the maximal growth solution consisted of only five species with non-zero abundances, the richness of suboptimal solutions was found to routinely exceed ten species (Fig 5B) with a maximum richness of 19 species achieved at 70.3% of the maximal growth rate. The binned richnesses showed small variabilities, partially due to the use of FVA solutions with zero maximum abundance values for some species. Similar richness trends were observed for the high-fiber diet, with 61% and 18% of the 568 simulation cases having richnesses of at least 10 and 15 species, respectively (S2B Fig).
The 575 Western diet cases produced a remarkably simple tradeoff between community growth and species diversity represented by a line (R2 = 0.996) for growth rates less than 97% of the optimal value (Fig 5C). We hesitate to refer this curve as “Pareto optimal” because our computational procedure does not ensure Pareto optimality of the calculated points. Regardless, this curve clearly showed that species diversity could only be achieved at the expense of community growth and visa versa. A very similar growth-equitability curve was generated with the high-fiber diet (S2C Fig).
Enhanced species diversity at suboptimal growth rates tended to produce more favorable ratios of SCFA net synthesis rates (Fig 5D). For the Western diet, the bin centered at 0.79 produced average butyrate and propionate fractions of 18% and 17%, respectively, which appeared to be more consistent than the maximal growth solution with the 20% values commonly reported for in vivo levels of these two SCFAs. The binned SCFA fractions showed small variabilities, consistent with predicted richness variations. Less favorable SCFA synthesis rates were predicted for the high-fiber diet (S2D Fig). While the propionate fractional rate averaged 18% over the range of 70–95% of the maximal growth rate, the butyrate fractional rate never averaged more 12% in any bin and reached a single-case maximum of 13%. Lower butyrate synthesis compared to the Western diet was attributed to reduced participation of the high butyrate producers Fusobacterium, Eubacterium and Faecalibacterium in the simulated communities.
While our method of imposing calculated bounds on the species abundances proved effective for investigating the growth-diversity tradeoff, such a direct mechanism for modulating species abundances is not biologically plausible. Several mechanisms for tuning community composition have been widely studied, including spatial structuring of the participating species in multispecies biofilms [66, 67] and the modulation of metabolite crossfeeding between species [68–70]. Because SteadyCom is based on the assumption of a homogeneous environment, the only mechanism available to tune community composition is modulation of nutrient uptake rates, including the uptake rates of crossfed metabolites. To further investigate how crossfeeding was modulated to achieve high diversity at suboptimal growth rates, we compared the optimal solution for the Western diet to the results in Fig 5 for the simulation cases binned at 79% of the optimal growth rate.
A heatmap of uptake/secretion rates of the 29 crossfed metabolites for each of the 20 species shows that the optimal solution was characterized by relatively high crossfeeding rates between a small number (five) of participating species (Fig 6A). Crossfed metabolites with the largest uptake/secretion rates were aspartate, serine, CO2, D-lactate, L-lactate (metabolites 4, 17, 23, 27, 28; see also Fig 2D). Comparable results for suboptimal growth at 79% of the optimal growth rate were generated by averaging the metabolite uptake/secretion rates across the 34 simulation cases within this bin. By contrast to the optimal case, suboptimal growth was characterized by relatively low crossfeeding rates between a large number (15.4 on average) of participating species (Fig 6B).
To better understand the modulation of crossfeeding rates between optimal and suboptimal growth, we averaged the absolute values of the exchange (uptake and secretion) rates of each crossfed metabolite across the 20 species and the 34 simulation cases (for suboptimal growth). While individual crossfed metabolites differed with respect to their average exchange rates, the overall utilization of the 29 metabolites was similar (Fig 6C) with the average exchange rate across all metabolites being 0.25 mmol/h for maximal growth and 0.24 mmol/h for the suboptimal growth cases. By contrast, the standard deviation of the exchange rate across all metabolites was 0.64 mmol/h for maximal growth and 0.32 mmol/h for the suboptimal growth cases. Similar results were obtained for the high-fiber diet (S3 Fig). This analysis reinforced a central theme of this in silico study: optimal growth resulted in large metabolite exchange rates between few species while suboptimal growth was characterized by reduced metabolite exchange rates between many species.
Metabolic modeling of the human gut microbiota has emerged as an important in silico tool for investigating community growth and composition as well as species interactions. While yielding useful insights into community behavior, previous gut microbiota models [38, 71–75] have been limited with respect to the number of species included and the metabolic interactions allowed. To our knowledge, the gut community models developed in this study represent the most complete descriptions to date both with respect to the number of species (e.g. 28 bacteria) and model size (e.g. 26,867 metabolites, 35,031 reactions). Our ability to generate and solve such large community models is directly attributable to the availability of gut bacteria reconstructions in the Virtual Metabolic Human database (https://vmh.uni.lu) [44] and the computational efficiency of the SteadyCom method [38]. The model database allowed the species to be chosen based on the most abundant genera in the gut [19] rather than by the availability of curated reconstructions, which remains limited to a few dozen species. The tradeoff for this unprecedented diversity of modeled species was that the reconstructions were only semi-curated and could not be expected to have the fidelity of fully curated models. Given the qualitative nature of our study focused on examining community growth and diversity, this limitation was deemed acceptable. Rather than requiring a priori specification of particular interactions between species, SteadyCom allowed arbitrary crossfeeding of secreted metabolites between all species. We limited crossfeeding to the 21 amino acids and 8 common byproducts, yielding 21,924 possible crossfeeding relationships for the 28-species community.
The 28 species included in the first gut community offered a wide diversity of metabolic capabilities, including the synthesis of short-chain fatty acids (SCFAs) used by host colonocytes as a primary energy source. As observed in healthy gut communities [5, 6, 76], SCFA synthesis was diversified across the community, with 24 species, 7 species and 13 species secreting acetate, butyrate and propionate, respectively. However, the maximal growth communities determined with SteadyCom consisted of only five species for the Western diet and six species for the high-fiber diet. The optimized communities were enriched in genera known to be overrepresented in inflammatory bowel disease, namely Escherichia [60], Enterobacter [61] and Citrobacter [62]. The communities also exhibited large imbalances in SCFA production, with over 90% of SCFA synthesis yielding acetate and no butyrate secreted, another hallmark of IBD [20–22]. Our simulation results suggest that maximal community growth unchecked by the host may evolve to disease states such as IBD.
We used SteadyCom to perform community flux variability analysis (FVA) as a means to determine limits on achievable species diversity and SCFA synthesis. FVA suggested that suboptimal community growth rates offered the potential for substantially enhanced diversity with 20 of the 28 species capable of coexisting at 70% of the optimal growth rate for at least one diet. Based on these results, we generated a reduced 20-species community by eliminating the eight species capable of coexisting only at growth rates less than 70% of the maximal values. The eight species eliminated represented several genera known to be beneficial for gut health, most notably Lactobacillus [77], Alistipes and [78], Bifidobacterium [79]. Due to their low predicted growth rates compared to other community members, these species may need to establish favorable metabolic niches along the intestine to robustly coexist [80–82]. While outside the scope this study due to the homogeneous assumption underlying SteadyCom, the effect of such spatial gradients would be interesting topic for future research.
FVA performed with the reduced 20-species community further demonstrated the potential for achieving high species richness (defined as the number of species with abundances of at least 1%) and equitability (see Eq 1) as well as balanced SCFA production at suboptimal growth rates. We developed a simple randomized method for using the FVA results to constrain species abundances in SteadyCom to achieve suboptimal growth rates and sample the growth-diversity tradeoff surface. A remarkably simple linear relationship between the community growth rate and species equitability was predicted, with high levels of diversity (richness ≥ 12 species, equitability ≥ 0.55) achievable at growth rates below 85% of the maximal value. Increased species diversity resulted in more balanced SCFA synthesis, with butyrate comprising 14–18% of total production. These predictions are consistent with known characteristics of healthy gut communities, suggesting that simulated suboptimal growth represents a “healthy” state and simulated maximal growth represents a “dysbiosis” state such as IBD.
We further analyzed our suboptimal growth solutions to determine how SteadyCom achieved enhanced diversity and how the host might promote diversity in vivo. In addition to modulating the uptake of available nutrient across species, SteadyCom tuned the secretion/uptake rates of crossfed amino acids and metabolic byproducts to enhance the growth rates of otherwise slower growing species. Compared to maximal growth, suboptimal solutions were characterized by lower secretion and uptake rates of crossfed metabolites between a larger number of species. These results suggest modulation of crossfeeding relationships is one possible mechanism available to the host for promoting diversity at the expense of growth. From a theoretical perspective, host-microbiota metabolic interactions might be viewed as a type of bilevel optimization problem with the microbiota attempting to achieve maximal community growth and the host modulating the gut environment to maximize species diversity. Gut diseases such as IBD might result from the host “losing the battle” due to inflexibilities resulting from poor diet and/or sudden loss of diversity due to antibiotic treatment. While a few metabolic modeling methods that address diversity have been proposed [83–85], additional modeling tools that directly address the growth-diversity tradeoff in microbial communities are needed.
As has been reported in numerous in vivo studies [86–89], we expected the in silico high-fiber diet to promote species diversity and enhance butyrate synthesis as compared to the Western diet. Instead our model predicted that diet had little effect on the growth-diversity tradeoff and high fiber actually resulted in reduced butyrate levels. These discrepancies could reflect insufficient numbers of fiber-degrading and butyrate-producing species in our simulated communities. Indeed only four members (Bacteroides, Prevotella, Alistipes, Desulfovibrio) of the 28-species community exhibited faster single-species growth rates on the high-fiber diet, and two of these species were removed to generate the 20-species community. The communities contained only three major butyrate producers, with one species (Faecalibacterium) being among the least competitive members of the 20-species community. The computational efficiency of SteadyCom allows the construction of larger community models with more representation of fiber-degrading and butyrate-producing species.
|
10.1371/journal.pbio.2003502 | “Pomacytosis”—Semi-extracellular phagocytosis of cyanobacteria by the smallest marine algae | The smallest algae, less than 3 μm in diameter, are the most abundant eukaryotes of the World Ocean. Their feeding on planktonic bacteria of similar size is globally important but physically enigmatic. Tiny algal cells tightly packed with the voluminous chloroplasts, nucleus, and mitochondria appear to have insufficient organelle-free space for prey internalization. Here, we present the first direct observations of how the 1.3-μm algae, which are only 1.6 times bigger in diameter than their prey, hold individual Prochlorococcus cells in their open hemispheric cytostomes. We explain this semi-extracellular phagocytosis by the cell size limitation of the predatory alga, identified as the Braarudosphaera haptophyte with a nitrogen (N2)–fixing endosymbiont. Because the observed semi-extracellular phagocytosis differs from all other types of protistan phagocytosis, we propose to name it “pomacytosis” (from the Greek πώμα for “plug”).
| The global significance of microorganisms is a consequence of their astronomical numbers. This is certainly true for the smallest planktonic algae on earth, which are less than 3 μm in diameter and the most numerous eukaryotic organisms of the oceans. Contrary to the general belief that algae use only sunlight and dissolved mineral nutrients to grow, these microscopic plants consume large numbers of bacteria. Their acting as mini-predators on bacteria of nearly their own size is hard to imagine. A tiny algal cell is cramped with organelles—such as nucleus, mitochondria, chloroplasts—and there is simply no space inside this cell to engulf a large bacterium in the usual manner. To find out how the 1.3-μm–diameter haptophyte algae feed, we studied them using high-resolution electron microscopy. We found that prey handling by the alga differs from all other types of cell feeding. We showed that this alga holds the 0.8-μm–diameter prey in the open cytostome (cell mouth) and that, from among planktonic bacteria, the alga apparently selects a ball-shaped Prochlorococcus (an abundant cyanobacteria responsible for most of global photosynthesis) that tightly fits into the open cytostome like a plug. Instead of full prey digestion, we observed that the alga leaves behind the doughnut-shaped carcass of the prey. We conclude that such unusual feeding, which we call “pomacytosis”, of this tiny predatory alga is caused primarily by the space limitation inside its cell.
| In conventional phagocytosis, the caught prey is internalized, i.e., enclosed by a phagocytic membrane inside the predator cell to form a food vacuole, within which prey is digested and its contents are absorbed through the vacuole membrane [1]. Apart from secure isolation of the prey from the environment, full closure of the food vacuole benefits the predator in a number of ways. The fully closed vacuole allows the predator to pump excess water to reduce the vacuole volume, to adjust pH inside the vacuole to facilitate prey digestion by lytic enzymes, and to contain lysed prey for efficient nutrient assimilation. Only refractory prey material, e.g., moieties of cell wall, is egested when the closed food vacuole finally fuses back with the plasma membrane [2]. Thus, conventional phagocytosis of internalized prey requires enzymes, microfilament, microtubule, and membrane investments and can be limited by the predator size [3].
Phagocytosis of prey of similar size or bigger is difficult but achievable for protists. For example, some dinoflagellates use a feeding tube to inject lytic enzymes into prey and to extract digested prey contents [4,5]. Other dinoflagellates and several haptophytes form extracellular, yet closed, food vacuoles [6–8]. Such extensive extracellular vacuoles can only be completed by large predatory cells that can produce and stock sufficient amounts of the required investments. Compared to extracellular phagocytosis, internalization of similar-sized prey requires from the predator fewer investments but sufficient intracellular space free from organelles. In protists, the nucleus, mitochondria, and chloroplasts (the latter in algae) can vary in size, but these organelles cannot be smaller than a certain minimal volume. Owing to the presence of such “nonscalable” organelles [9], the intracellular volume available for investment storage and prey internalization shrinks as a power function of the predator cell size. Consequently, small protists may be unable to internalize (conventionally phagocytose) similar-sized prey. To test that, we focused on feeding of the smallest algae (<3 μm in diameter), whose chloroplast-packed cells in addition to the nucleus and mitochondria should have the minimal organelle-free space among free-living protists.
According to our morphometric estimates, organelles occupy approximately 70% of a haptophyte alga with a cell volume of 2.8 ± 0.8 μm3 (n = 10; S1 Fig). Even after taking into account scalable but vital cell components, e.g., endoplasmic reticulum rich in ribosomes and enzymes, both the haptophyte alga as well as the smallest known prasinophyte alga with a cell volume of 1.1 to 5.7 μm3 [10] are still capable of internalizing a bacterial cell of 0.1 to 0.3 μm3 [11]. This is in agreement with the substantial indirect experimental evidence that, despite their diminutive size, the smallest (1- to 3-μm diameter) algae are the main predators of bacterioplankton in the open ocean [12,13]. However, because of insufficient resolution of optical microscopy, phagocytosis by these algae could only be inferred [14].
In order to find out how algae less than 3 μm in size phagocytose similar-sized bacteria, we chose to study the smallest oceanic picoeukaryotic algae, plastidic eukaryote small (PES), separated from other protists and bacteria living in seawater by flow cytometry. Using high-resolution electron microscopy to observe fine cellular details of the sorted algae, we found that their semi-extracellular bacterial phagocytosis—“pomacytosis”—differs from all other types of phagocytosis.
Low concentrations of bacterioplankton and PES (6 × 105 cells ml-1 and 4 × 102 cells ml-1, respectively; S2 Fig) in the studied region of the Eastern subtropical North Atlantic Ocean were typical for open ocean waters [11,13].
The main PES population was well defined by flow cytometry and selected for sorting (S3 Fig). High-throughput barcoding analysis of flow-sorted PES cells (S3 Fig) yielded 10,416 high-quality (≥300 nt) 16S rRNA gene reads and identified the dominant taxa: 51% of the amplicons were sequences of cyanobacteria, composed of Prochlorococcus (26%) and unicellular diazotrophic cyanobacteria group A (UCYN-A; 25%), and 38% of the amplicons were chloroplast sequences, the majority of which (58%) belonged to the Braarudosphaeraceae, a coccolithophore family of the Haptophyta (Fig 1). The remaining chloroplast sequences belonged to 10 other types of small algae, each of which represented only a minor fraction of the PES cells (Fig 1). The negligible number of sequences of SAR11 alphaproteobacteria (Rickettsidae)—the most abundant bacteria in the samples (thus the most probable by-sorted cells)—validated the high purity of PES sorting.
Analyses of nearly full-length ribosomal gene sequences confirmed the phylogenetic affiliation obtained with shorter amplicons. Full-length sequences of the 16S rRNA gene of Prochlorococcus and UCYN-A were 99% identical to high light–adapted Prochlorococcus marinus strain MIT9301 and 100% identical to the Candidatus Atelocyanobacterium thalassa isolate a long-term oligotrophic habitat assessment (ALOHA) [15], respectively. The 18S rRNA gene sequence was 99% identical to a calcifying Braarudosphaera bigelowii isolate TMRscBb7 [16] (S4 Fig) and to a small noncalcifying alga collected from oligotrophic waters of the South East Pacific Ocean (S4 Fig) [17], confirming the chloroplast 16S rRNA gene-based identification.
Scanning and transmission electron microscopy (SEM and TEM, respectively) showed no curved rod-shaped cells of the most abundant SAR11 bacteria (S5A Fig) among flow-sorted PES cells. The absence of by-sorted SAR11 bacteria reaffirmed the high sorting purity. The majority (95%) of the imaged PES cells (185 out of 195) were ball-shaped small cells with an estimated diameter of 1.3 ± 0.22 μm (n = 33, size corrected for 30% linear cell shrinkage during sample dehydration [18]). Some of them bore organic, noncalcified scales (S6 Fig). These morphotypes represent coccolythophore life cycle stages found in nutrient-poor waters [19–21]. Among the sorted PES cells, there were no cells with external mineral investments, i.e., pentagonal-shape liths characteristic of Braarudosphaera species found in nutrient-replete waters [22]. A few morphologically different cells (10 out of 195 examined cells) had one or two well-preserved flagella (S7 Fig) that ruled out the artificial loss of external investments by the dominant alga.
Out of 185 cells of the dominant alga, 155 (84%) were associated with smaller coccoid cells 0.81 ± 0.08 μm (n = 10, size corrected for 30% linear cell shrinkage during cell dehydration) in diameter (Fig 2). An additional intracellular body of the dominant algal cells was observed using TEM (Fig 2E). The 0.47 ± 0.05 μm (n = 4) diameter body occupied a particular location at the cell periphery next to one of the two chloroplasts. When the body was absent, a rupture in the algal cell wall was observed (Fig 2F, thick arrow; S5B Fig), confirming that the body was intracellular but could be lost under mechanical stress caused by sorting PES cells directly on TEM grids. A similar intracellular “spheroid body” in B. bigelowii isolate TMRscBb7 was identified as an obligate N2-fixing UCYN-A endosymbiont [16]—cyanobiont. Contrary to the cyanobiont, the molecularly identified Prochlorococcus associated with PES is a free-living planktonic cyanobacterium that was numerous in the studied seawater (1.7 × 105 cells ml-1).
Synthesising the above evidence, we concluded that the UCYN-A amplicon derived from the “intracellular body” and the Prochlorococcus amplicons represented the extracellular cocci attached to the PES cells. We interpreted the latter association as phagocytosis of Prochlorococcus by the naked haptophyte (hereafter referred to as B. bigelowii JC142). We assigned the observed B. bigelowii cells to two major groups and one minor group, as follows: (a) alga with an associated Prochlorococcus, of which less than 50% cell surface is inside the cytostome (49%); (b) alga with an associated Prochlorococcus, of which more than 50% cell surface is inside the cytostome (35%); and (c) alga with a cytostome but without prey (16%) (Fig 2).
The cytostome is most likely used for shape- (and possibly surface-)selective prey recognition and capture. In support of the notion of selection, neither molecular nor microscopic evidence suggested that B. bigelowii JC142 fed on SAR11 alphaproteobacteria—the most abundant free-living bacteria in the studied seawater (2.8 × 105 cells ml-1; S2 Fig). The algae preferred to feed on less abundant Prochlorococcus (1.7 × 105 cells ml-1), which made up only 27% of total bacterioplankton in the seawater (6 × 105 cells ml-1; S2 Fig), i.e., the haptophyte selected on average one out of four encountered free-living bacterial cells.
Because of high-purity PES sorting, the individual Prochlorococcus observed by SEM were not by-sorted cells but were in fact cells detached from the haptophytes during sorting (e.g., Fig 2D). Both intact and doughnut-shaped, deformed Prochlorococcus cells were observed (S5 Fig). The intact, spherical Prochlorococcus (14 observed cells) were probably at the start of pomacytosis, while the doughnut-shaped Prochlorococcus with a central small spot of depressed surface area (23 cells) were at the end of pomacytosis (S5 Fig). Similarly, deformed Prochlorococcus cells were observed by SEM and TEM (Fig 2L and 2M), affirming that the deformation was a result of pomacytosis rather than an artefact of SEM sample preparation.
High-power TEM revealed that in the groups (a) and (b), the prey Prochlorococcus cell is fitted into a semicircular cytostomic depression, which—according to SEM—is in reality hemispherical and is anchored in the cytostome between the two algal chloroplasts (Fig 2C and 2E). In all 155 specimens observed with the Prochlorococcus cell, the latter remains at least partially free of the algal cytostome membrane (Fig 2, S8 Fig). To our knowledge, this is the first observation of semi-extracellular phagocytosis of prey by a protist using a partially opened cytostome.
The B. bigelowii JC142 is the smallest haptophyte that was directly observed to phagocytose free-living bacteria. However, B. bigelowii ability to internalize the selected bacterium is evolutionary evidenced—its cyanobiont is of a phagocytic origin. The intracellular UCYN-A symbiont cell in B. bigelowii isolate TMRscBb7 is surrounded by a food vacuole membrane [16]. The presence of the UCYN-A cyanobiont further reduces the intracellular space of B. bigelowii available for prey internalization. The size of the UCYN-A symbiont of B. bigelowii JC142 is at the lowest end of the reported UCYN-A size range [15,16,23,24]. The cyanobiont occupies less than 5% of the estimated volume of the B. bigelowii JC142 cell, while the Prochlorococcus prey measures more than 20% of the algal volume. Perhaps the choice between conventional phagocytosis and pomacytosis depends on the size ratio between the alga and its prey. In order to overcome its space limitation, 1.3-μm B. bigelowii JC142 cell, pomacytoses 0.8-μm Prochlorococcus instead of using whole-cell phagocytosis.
Selective feeding of B. bilgelowii JC142 on Prochlorococcus implies that, despite the internal supply of fixed nitrogen by the UCYN-A cyanobiont—or perhaps owing to this supply as well as to metabolic demands of the symbiont—the haptophyte could be limited in other main inorganic nutrients [25], e.g., phosphorus and iron. However, this limitation is unlikely because B. bigelowii JC142 was collected in the Eastern subtropical North Atlantic Ocean fertilised by aeolian dust from the Saharan desert. Consequently, the surface waters in the studied area are enriched in phosphate and iron [26] but are poor in nitrogen salts [27]—the environment that facilitates the growth of N2-fixing photoautotrophs. Instead of photoautotrophy, B. bigelowii JC142 cells unconstrained by inorganic nutrients, including nitrogen (fixed by its cyanobiont), pomacytose Prochlorococcus. Therefore, the main nutrient the haptophytes gain from Prochlorococcus prey is, perhaps, fixed carbon.
B. bigelowii may require fixed carbon because it has the cyanobiont. The UCYN-A cyanobiont lost its photosystem II complex (PSII) but retained its photosystem I (PSI) [28] to use light energy to fix N2. In return for the shared fixed nitrogen, the Braarudosphaera host should share its fixed carbon with the cyanobiont [15,29]. Furthermore, to minimize inhibition of the cyanobiont N2 fixation, a B. bigelowii cell needs to keep its intracellularly dissolved O2 concentration low. Large host cells, e.g., Rhizosolenia and Rhopalodia diatoms, do that by spatially segregating their chloroplasts from N2-fixing cyanobionts within their cells [30]. In the 1.3-μm B. bigelowii JC142 cell (Fig 2E), O2 produced by the adjacent chloroplast could directly inhibit N2 fixation by the cyanobiont, and the haptophyte needs to reduce [29] if not to halt photosynthesis by its own chloroplasts. Consequently, both the host and cyanobiont become starved of fixed carbon and require its alternative, external source. In order to acquire that fixed carbon, B. bigelowii JC142 selectively pomacytose free-living Prochlorococcus cyanobacteria.
Based on our observations (Figs 1 and 2), we suggest interpreting the reported association between the “unknown structure” and UCYN-A–bearing haptophyte (Fig 6 in [29]) as Prochlorococcus cell pomacytosed by the haptophyte. Low CO2 fixation by the haptophyte chloroplasts compared with high CO2 fixation by the “unknown structure”—Prochlorococcus (Fig 6 in [29])—supports our suggestion that the Braarudosphaera could acquire fixed carbon from its prey rather than from its own chloroplasts. Perhaps, because a CO2-fixing Prochlorococcus cell also produces O2, the B. bigelowii JC142 cell does not internalize it. Instead, live Prochlorococcus is kept segregated from the O2-sensitive cyanobiont, and the haptophyte keeps the cytostome semi-open to allow O2 dissipation. Therefore, pomacytosed Prochlorococcus could be viewed as a temporary chloroplast substitute.
Conventional phagocytosis is a relatively quick process that usually takes seconds (e.g., [8]), and one seldom observes a protist predator in the process of internalizing prey. Because the majority of the B. bigelowii JC142 collected during six-hour sampling was in a process of feeding (84% held prey), pomacytosis should be a slow process that takes hours. The absence of internalized Prochlorococcus cells and nearly 1:1 ratio between pomacytosed Prochlorococcus with more than half cell surface exposed (group [a]) and with less than half cell surface exposed (group [b]) suggest that the haptophyte controls exposure of the prey cell to seawater. During slow pomacytosis, the predator could gain extra benefit from the prey that fixes CO2 and takes up nutrients through the cell wall exposed to seawater (Fig 2E–2K). Unlike conventionally phagocyting cells, the pomacyting B. bigelowii JC142 detained Prochlorococcus in their cytostome without full internalization, perhaps; harvested fixed carbon released by prey; and egested the deformed, spent prey without full digestion (S5 and S8 Figs).
Therefore, a combination of intracellular space limitation (primarily) and physiological requirements of the tiny predatory alga (secondarily) leads to semi-extracellular phagocytosis of selected prey.
This is an oceanographic study carried out in the international waters. This research does not require special permission.
The study was carried out in the Eastern subtropical North Atlantic Ocean (23° 37′ N 20° 43′ W) on board the Royal Research Ship “James Cook” during the research cruise JC142 from November to December 2016. Seawater samples from 25 m (a representative depth of the surface mixed layer) were collected using a rosette of 20-l Niskin bottles mounted on a conductivity-temperature-depth (CTD) profiler. All plastic- and glass-ware for handling seawater was prewashed with 10% HCl and rinsed with sampled seawater.
Concentrations of total bacterioplankton, including Prochlorococcus and SAR11, the latter as a population of cells with low nucleic acid content [31], were determined by flow cytometry. Routinely, samples were fixed with 1% (w/v) paraformaldehyde (PFA) final concentration, stained with SYBR Green I DNA dye [11,32], and analysed with the custom-modified FACSort instrument (Becton Dickinson, Oxford, UK) equipped with the blue diode laser (488 nm, 50 mW; Quantum Analysis, Munster, Germany) using the CellQuest software.
For determining concentrations of PES and Synechococcus and for cross-referencing microbial populations in the concentrated samples (used for flow sorting), seawater samples were fixed with 2% PFA, stained with 0.1 μg ml-1 Hoechst 33342 (final concentration), and analysed with the custom-built MoFlo XDP instrument (Beckman-Coulter, High Wycombe, UK) (S2 Fig) using the Summit 5.4 software. The first UV diode laser (355 nm, 100 mW; JDSU, CY355-100, Thailand) and the second blue diode laser (488 nm, 240 mW; Cobolt, Solna, Sweden) were aligned through the first and third pinhole, respectively. Shallow angle light scatter (forward scatter [FSC]) of the UV light was detected using the 351 ± 5–nm optical filter and the H957-18 photomultiplier (Hamamatsu, Japan). More sensitive H957-27 photomultipliers (Hamamatsu) were used for detecting particle fluorescence at four wavelengths (457 ± 25 nm, 530 ± 20 nm, 580 ± 15 nm, >643 nm) and the three wavelengths (505–550 nm, 580 ± 15 nm, 670 ± 15 nm) excited by the first and second laser, respectively.
A reference mixture of yellow-green (505/515 nm) 0.5-μm beads (Life Technologies, Eugene, Oregon, US) and multifluorescence 1.0-μm beads (Fluoresbrite Microparticles, Polysciences, Warrington, Pennsylvania, US) were used as an internal standard for both fluorescence and flow rates. The absolute concentration of beads in the stock solution was determined using syringe pump flow cytometry [33].
For flow sorting, microbes were gravity concentrated approximately 103-fold using sterile 0.2-μm pore size Sterivex filter units (Millipore, Watford, UK) attached directly to Niskin bottles. For molecular identification, concentrated microbial samples were fixed with Lugol iodine solution [34] and stored at +4°C before being flow sorted within 48 hours. Samples were discoloured with thiosulfate [34] and stained with Hoechst 33342 prior to sorting. For electron microscopy analyses concentrated samples were fixed with 2% PFA and stained with Hoechst 33342 prior to sorting. The same dominant distinct population of the smallest picoeukaryotic algae—PES—was flow sorted with the MoFlo XDP instrument (S3 Fig) using the Summit 5.4 software. The instrument was optically aligned, and its sorting purity and recovery were optimised using blue (350/440 nm) 1.0-μm beads (Life Technologies). Only PES cells gated by both gates (S3B, S3C, S3E and S3F Fig) were sorted. Purity of sorted PES cells was validated by the molecular and electron microscopy analyses.
For TEM analyses, 1 × 103 to 2 × 103 target PES cells were flow sorted directly on formvar/carbon–covered 200 mesh copper grids (Agar Scientific, Stansted, UK) stained with 2% w/v Gadolinium (aqueous solution), rinsed with pure deionized water, and stored in a desiccator for analysis ashore. The grids were examined at 200 keV with the Jeol 2011 LaB6 TEM instrument fitted with a Gatan UltraScan 1000 camera at the University of Warwick’s Research Technology Platform in Advanced Bioimaging in the United Kingdom.
For SEM analyses, 20 × 103 target cells were flow sorted into sterile 1.5-ml microcentrifuge tubes containing aqueous solution of 1% glutaraldehyde (Electron Microscopy Sciences). The tubes were stored at 4°C and brought ashore. The sorted cells were collected onto 0.2-μm pore size, 13-mm polycarbonate filters under low vacuum, dehydrated in the ethanol series, and critical point dried using 99.9% hexamethyldisilazane (Sigma-Aldrich). The dehydrated filters were stored in a desiccator at room temperature. Prior to SEM analyses, the filters were sputtered with Au/Pd (3:2) to a thickness of 10 nm using the High-Resolution (208hr) Sputter Coater coupled with the MTM20 film thickness controller (Cressington). The filters were examined with the high-resolution SEM UltraPlus instrument (Zeiss Gemini) at 5 keV using the secondary electron detector at the Imaging and Analysis Centre of the Natural History Museum in London, UK.
Cell dimensions were measured on both TEM and SEM micrographs using the ImageJ software [35]. The values obtained from the SEM micrographs were corrected to account for approximately 30% cell shrinkage [18]. Average cell volumes were calculated assuming a ball or spheroid shape of algal cells (4/3πa2b), a spherical segment for chloroplasts (πh2[b-1/3h]), an ellipsoid for a nucleus (4/3π[a-h]2b), and half of this ellipsoid for a mitochondrion (S1 Fig).
For molecular analyses, 20 × 103 to 50 × 103 PES cells were flow sorted into sterile 1.5-ml microcentrifuge tubes. An aliquot of 2 μl containing approximately 2 × 103 cells was added into a 0.2-ml PCR tube containing 30 μl of Q5 High Fidelity Master Mix (New England BioLabs) complemented with primers and nuclease-free water (Ambion). For full-length 16S or 18S rRNA gene amplification, we used 27f/1492r [36] or 63f/1818r [37] primers with annealing temperature of 59°C. The amplicons were added with A-tails (OneTaq DNA polymerase, New England BioLabs), ligated to the pGEM T-Easy vector (Promega), and transformed into the NEB 5-alpha competent Escherichia coli cells (New England BioLabs). Plasmids from the positive colonies were sequenced with T7 and SP6 primers to cover the full amplicon length. The 18S rRNA gene sequences were aligned with 18 reference sequences of haptophytes (1,400 positions), and phylogenetic relationships for the dataset were calculated with MrBayes software [38].
For a massively parallel sequencing, hyper variable regions V3–V4 (490 bp) were amplified by PCR using S-D-Bact-0341-b-S-17 and S-D-Bact-0785-a-A-21 primers [39]. The forward primer included the PGM barcode adapter (Ion Xpres Barcode Adapters 1–96 Kit, ThermoFisher Scientific), and both primers were tailed with the Ion Torrent sequencing adapters to allow direct downstream multiplexed sequencing. Following amplification, PCR products of approximately 490 bp were gel purified with NucleoSpin Gel and PCR Cleanup kit (Macherey-Nagel), and 1.5 ng of the product was used for template preparation with the Ion Torrent OneTouch System (ThermoFisher Scientific). The templates were sequenced on an Ion Torrent PGM sequencer (ThermoFisher Scientific) using the Hi-Q sequencing chemistry.
After sequencing, the individual sequence reads were first quality trimmed using the Ion Torrent software suite and then further processed using the bioinformatics pipeline of the Silva NGS project [40]. This involved quality controls for sequence length (≥300 bp) and the presence of ambiguities (<2%) and homopolymers (<2%). The remaining reads were split into individual sample FASTA files using mothur [41] and aligned against the SSU rRNA seed of the SILVA database release 119. The classification was done by a local BLAST search against the SILVA SSU Ref 115 nonredundant (NR) database using BLAST 2.2.22+ with standard settings. The analysis gave (semi)quantitative information (number of individual reads representing in a taxonomic pool) on the composition of the original PCR amplicon pool [39]. The classification of plastidic SSU rRNA sequence reads was done by nucleotide BLAST search against the NR database at the National Center for Biotechnology Information (NCBI; www.ncbi.nlm.nih.gov).
|
10.1371/journal.pntd.0002737 | Serologic Prevalence of Toxoplasma gondii in Indian Women of Child Bearing Age and Effects of Social and Environmental Factors | Seroprevalence and incidence of toxoplasmosis in women of child bearing age has remained a contentious issue in the Indian subcontinent. Different laboratories have used different patient recruitment criteria, methods and variable results, making these data difficult to compare.
To map the point-prevalence and incidence of toxoplasmosis in India.
In this cross-sectional study, a total of 1464 women of fertile age were recruited from 4 regions using similar recruitment plans. This included women from northern (203), southern (512), eastern (250) and western (501) regions of India. All samples were transported to a central laboratory in Delhi and tested using VIDAS technology. Their age, parity, eating habits and other demographic and clinical details were noted.
Most women were in the 18–25 years age group (48.3%), followed by 26–30 years (28.2%) and 31–35 years (13.66). Few (45) women older than 35 yr. were included. Overall prevalence of anti-Toxoplasma IgG antibodies was seen in 22.40%, with significantly more in married women (25.8%) as compared to single women (4.3%). Prevalence increased steadily with age: 18.1% in the 18–25 yr. age group to 40.5% in women older than 40 yr. The prevalence was high (66%) in those who resided in mud houses. Region-wise, the highest prevalence was observed in South India (37.3%) and the lowest (8.8%) in West Indian women. This difference was highly significant (P<0.001). Prevalence was 21.2% in East India and 19.7% in North India. The IgM positivity rate ranged from 0.4% to 2.9% in four study centers.
This pan-India study shows a prevalence rate of 22.4% with a wide variation in four geographical regions ranging from as low as 8.8% to as high as 37.3%. The overall IgM positivity rate was 1.43%, indicating that an estimated 56,737–176,882 children per year are born in India with a possible risk of congenital toxoplasmosis.
| Toxoplasmosis is a protozoan parasitic disease commonly transmitted and propagated by cats as family pets. Infection acquired during pregnancy can lead to congenital abnormalities in the fetus, still birth or intrauterine death. Seroprevalence and incidence of toxoplasmosis in Indian women of child bearing age has remained a contentious issue. Different laboratories have used different patient recruitment criteria, methods and variable results, making these data unreliable. There is no published pan-India seroprevalence study. Hence, a seroprevalence study was undertaken comprising 1464 women of reproductive age representing four distinct geographical regions of India. This resulted in an estimated prevalence of 22.4% (328); the highest prevalence being in South India (37.3%) followed by East India (21.2%) and North India (19.7%). West Indian women had the lowest seroprevalence (8.8%). This difference was highly significant. In our analysis we determined the possible risk-factors of infection in these women. These included lower socioeconomic status, residing in mud plastered houses, consumption of raw salad, drinking untreated water, owning pets and advanced age. Overall, the incidence rate of toxoplasmosis was 1.43%. Extrapolating the data, we estimate that between 56,737 and 176,882 children a year may be born in India with a possible risk of congenital toxoplasmosis, which can manifest itself in-utero or several years after birth.
| Toxoplasma gondii (T. gondii) infection is a significant member of the TORCH group of diseases which cause congenital abnormalities, and even fetal loss. TORCH group infectious agents also consist of Rubella, Cytomegalovirus, Herpes viruses and Treponema pallidum. In India, awareness about these infections that cause congenital conditions is poor [1]–[4]. Most women who seek medical attention, or are referred by obstetricians, are those who have had an undesirable pregnancy outcome [5], [6].
Toxoplasmosis is caused by the protozoan parasite T. gondii. It has a wide host range, infecting most warm-blooded species but the life cycle is completed only in felids [7], [8]; Indeed only cats can shed the environmentally-resistant stage of the parasite (oocyst) in their feces [9], [10]. Humans usually become infected by ingesting food or water contaminated with cat faeces containing oocysts [11], [12] or by eating under-cooked meat containing the encysted stage of the parasite (tissue cysts) [13], [14]. Infection acquired during pregnancy can be transmitted to the fetus, sometimes with serious consequences. There are numerous serological surveys of T. gondii infection in pregnant women in India, but most of them were based on convenience sampling, and often selectively in women with bad outcome of pregnancy [15]–[18].
Here, we present the first designed survey for determining the prevalence rate of anti-T. gondii antibodies in Indian women of reproductive age from four geographic regions: East, West, North and South India.
Blood sampling was performed from October 2011 to October 2012. The Institutional Review Board of AIIMS approved this study (IEC/NP-92/2011). Informed written consent was obtained from all participating women who agreed to participate in the study. The study followed the STROBE guidelines (Supplementary file S2).
Serum samples were assayed for anti-Toxoplasma IgG and IgM antibodies by a commercially available Vitek Immuno-Diagnostic Assay System (VIDAS, BioMerieux SA, France), strictly following the manufacturer's instructions. All IgG and IgM positive samples were tested for IgG avidity using the same technology (VIDAS, BioMerieux SA, France), as published previously [3]. An avidity index of <0.200 indicates low avidity; an index of 0.200–0.299 indicates borderline avidity, and an index of >0.300 denotes high avidity for IgG. High avidity enables exclusion of a recent infection of <4 months duration. More details are provided in a related publication [3].
Data was entered in Excel sheet and imported into SPSS statistical program for analysis. For statistical evaluation of binomial data, the χ2 test with 95% confidence intervals according to Clopper and Pearson were used; P values <0.05 were considered statistically significant. Incidences (IgM positivity) and prevalence (IgG positivity) rates are expressed as percentages. To estimate the approximate number of babies born with risk of congenital T. gondii infection in India per annum, following formula was used.
Total population and live birth rates were taken from the Government of India official website [19].
For example, we had 434 women in their first trimester, of whom 7 (1.61%) were IgM positive. In the first trimester of pregnancy the rate of congenital transmission is reported to be 13% [6]. Hence, presuming that all women were in their first trimester, the approximate number of children born with a risk of congenital T. gondii infection would be
Using the same formula, out of 177 women in their second trimester, 4 (2.25%) women were IgM positive. Hence presuming a transmission rate of 29% in the second trimester [6], the approximate number of children born with a risk of congenital T. gondii infection would be
Data on 1464 women of reproductive age which ranged from 18 to 45 years (mean ± SD, 26.9±5.9) from four geographically distinct regions of India are presented here. Of these, 250 (17.1%) were from East India (Assam), 203 (13.8%) from North India (Delhi and national capital region), 499 (34.1%) from Western India (Gujarat), and 512 (34.9%) from South India (Karnataka). Region-wise, the mean age of the women in the East was 24.2±4.2 yr., 29.3±5.8 yr. in the North, 25.2±5.9 yr. in the West and 29.2±5.7 yr. in the South. The difference was insignificant. The distribution of various age groups is shown in figure 1. Out of 1464 women, 233 (15.9%) were single with a mean age of 22.4±3.8 (18–41 yr. age range) while 1231 (84.08%) were married with a mean age of 27.8±5.9 (18–45 yr. age range). Of the 1231 married women, 297 (24.1%) were nulliparous with a mean age of 30.2±6.3 and 934 (75.9%) were parous with a mean age of 27.1±5.6 (Figure 2, Flow chart 1). All single women were nulliparous and non-pregnant. Of the 934 parous women, 471 (50.4%) were single gravida and 463 (49.6%) were multigravida. Their mean age was 25.3±5.0 and 28.8±5.6, respectively. The number of gravida was as high as 7 (6 women; 1.3%). Of the 934 gravida women 356 (38.2%) had at least one live birth, while 153 (16.4%) had all adverse pregnancy outcomes. The remaining 424 (45.4%) were primigravida. Overall, 751 women were pregnant at the time of sample collection. Of these, 434 (57.8%) were in their first trimester, 176 (23.4%) in their second and 141 (18.8%) in their third trimester.
The overall seroprevalence was 22.4% (328 of 1464). The prevalence rates varied significantly across the 4 regions, with the highest (37.3%) in South India and the lowest in West India (8.8%). The difference was highly significant (Figure 3). We also observed a significant difference in the prevalence rates of anti-Toxoplasma antibodies between single women (4.29%, 95% CI; a range of 1.9% to 6.9%) and married women (25.0%, 95% CI; a range of 22.6% to 27.4%) (p<0.005). However, seroprevalence increased with age (Figure 4, trend line). Prevalence was lower than 11.7% among those under 20 yr. of age, but steadily increased to 40.4% in those who were older than 41 yr. Region-wise proportion of anti-Toxoplasma IgG antibody positive women in various age groups is shown in table 2. Most (74.2%) of the 328 seropositive pregnant women were multigravida and only a quarter (25.8%) were primigravida (Figure 5, Flow chart 2). Of the 208 seropositive pregnant women at the time of sampling, two thirds (138/208 or 66.35%) were in their first trimester while 36/208 (17.30%) were in their second trimester and 34/208 (16.35%) in their third.
Only 21 women out of 1464 (1. 43%) had anti-Toxoplasma IgM antibodies. Trimester-wise, 434 women were in their first trimester and 7 (1.61%) of them were IgM positive, while 177 women were in their second trimester and 4 (2.25%) of these had IgM antibodies. None of the 141 pregnant women in their third trimester was IgM positive. Ten of 479 non-pregnant women had IgM antibodies. IgM positivity was highest (2.9%, 15 out of 512) in South India, followed by 0.8% in Eastern (2 out of 250), 0.6% in Western (3 of 499) and 0.4%, in North (1 out of 203) India. The region-wise pattern was similar to the prevalence rate of IgG antibodies. All IgM positive women were also IgG positive and none, except for one woman, was IgG negative and IgM positive. From South India, 191 women were seropositive, of which 15 (7.8%) were IgM positive, while in North India only 1 out of 40 (2.5%) and in East India 2 out of 53 (3.7%) IgG positive women were also IgM positive. However, in West India the IgM positivity rate amongst the IgG positives was 6.8% (3 out of 44). This change was significant (<0.001) when estimated out of the total study subjects as opposed to only out of IgG positives. Out of 21 IgM positive subjects, 6 had low avidity, 8 were borderline and 7 showed high avidity. All low avidity subjects were followed up for 6 months. Of these, 3 were pregnant and 2 delivered normal babies while one newborn had congenital hydrocephalus and microphthalmia. Follow-up samples from babies could not be tested. All low avidity seropositive women were from South India.
Most women (97.1%) belonged to low or lower middle income groups. The majority of women from South and East India resided in mud-plastered houses and consumed tube well/hand pump water without using any disinfectant or filter. The difference in the living conditions was highly significant (p<0.0001) between South India and West India and also between South India and North India (p<0.001), while the difference between North and East India was not significant (Figure 3). General socio-behavioral characteristics of these women are shown in table 3. Most women (1253 of 1464; 85.6%) were involved in housekeeping. In spite of having a rural background (75.1%), our study showed that 90.6% (1328) of the women had elementary education. A history of contact with animals was found in 426 (29.1%) women but pets were significantly more common (53.5%) in South Indian households (Table 3). Consumption of raw salad was common (1360/1464; 92.89%) across the country. Of the 308 women from North India who are excluded from our final data analysis, 178 were parous and 130 were nulliparous. Their mean age was 29.1 yr. (range of 15–52 yr.). Of the parous women, 78 (43.8%) were primigravida and 45 (25.3%) were two gravida. Fifty two (29.2%) women were multigravida and the number of gravida was as many as eight.
This community based study provides a significant and much needed resource material specific to India. Our study showed that age was one major variable of a higher prevalence rate of toxoplasmosis. The prevalence was also higher in those who were married and multigravida than those who were unmarried/single. This higher prevalence was not associated with the number of gravida per se but rather to higher mean age [27.8±5.9 (range 18–45 yr.)] of those who were married than those who were unmarried/single women [22.4±3.8 (range 18–41 yr.)]. This conclusion is further validated with observations that age of women from South India was higher than other regions and prevalence was also significantly higher in women of this region. There are also anecdotal reports associating multiple sexual exposures with high prevalence and possibility of sexual transmission of toxoplasmosis [20]. Over all prevalence of toxoplasmosis in the present study was significantly lower than reported earlier from India [3]. One plausible explanation for this difference could be that in the present study we included all women of fertile age from various cohorts, as compared to an earlier study in which we included only pregnant women. However, even if we combine those with suspected TORCH infections, who were otherwise excluded from final data analysis, the overall prevalence rate in North Indian women was 17.4%, which was significantly lower than the 45% reported by us 10 years ago, in the same population using the same diagnostic techniques. Whether it was due to improved awareness and social hygiene in the last 10 years, or due to selection biases in the two studies, cannot be ascertained with certainty.
The difference in the prevalence rate between women from South and West India was highly significant. It could be mainly due to socio-cultural and climatic factors. The climatic conditions in South India favour sustenance and proliferation of Toxoplasma oocysts. Also a highly significant number of households owned cats in this region. Moreover, as a social culture, South Indians do not wear shoes and most often are barefoot or wear sleepers only. This might increase chances of transferring T. gondii oocysts from soil and water to their food [9]–[13]. Western India, on the other hand, is a dry arid climatic zone where temperature in May-June averages 46°C. Socio-culturally also, the population in West India must wear shoes due to the high temperature and sandy soil. These climatic conditions are detrimental to T. gondii to maintain its life cycle. We [21] and others [22], [23] have previously demonstrated the role of environmental conditions on the prevalence of toxoplasmosis.
Although consumption of raw/undercooked meat and exposure to soil through farming or gardening have been associated with a higher risk of infection in various studies [1], [4], [21], no such correlation was observed in the multivariate analysis done in the present study. In India, consumption of raw or undercooked meat is extremely rare, hence this route of infection is theoretically negligible. The type of food, social considerations and quality of water consumed were the most likely factors associated with high prevalence of toxoplasmosis in South India, besides higher age. Water and food-borne outbreaks of toxoplasmosis have been well documented worldwide [11], [12] and also from India [24].
Anti-toxoplasma IgM antibodies were considered indicative of a recent infection for several decades, but then it was realized that these antibodies could persist for several months, even years after the primary infection [2], [3], [5]. Hence in the 1990s a new test, based on affinity levels of IgG antibodies binding with antigen, was developed, and known as the avidity test. Determination of avidity helps in determining if the infection is of recent origin or more than 4 months old [25]–[27]. The IgM positivity and low avidity rates we observed in this study may seem very low, but cumulative figures are alarming. India has a population of more than 1.22 billion and the live birth rate is 22.22/1000 per annum [19]. Taking these IgM rates into consideration, a conservative estimate of child births with a possible risk of congenital toxoplasmosis [6] would be between 56737 and 176882. This would translate into health and rehabilitation expenses to treat and rehabilitate the congenitally-infected children, many of whom may remain asymptomatic for several years. Unfortunately, many of such congenitally-infected adolescents and adults are not diagnosed accurately, whether the infection was acquired in-utero or after birth. Our calculations, which are based only on IgM positivity rates, are not a very reliable marker of recent infection, as discussed in the previous paragraph. Therefore, these estimates need more validation studies from India for transplacental transmission rates of T. gondii using IgG avidity or molecular methods, such as PCR. Due to a high false IgM positivity rate, we also conclude that carrying out only IgM testing without IgG testing is not an advisable approach of investigating toxoplasmosis, and all patients must be tested first for IgG and, if found positive, samples would be subjected to IgM and/or avidity tests.
Our study had some limitations. We attempted to extract a maximum of information regarding pan-India seroprevalence of toxoplasmosis and correlate it with environmental and socio-cultural considerations. It would have been ideal to include more centers and more samples but this was not feasible due to time and financial constraints. Also, we could not do multiple follow-up sampling to find out true incidence rates in various seasons and in IgM positive women.
|
10.1371/journal.pcbi.1005980 | Simulation enabled search for explanatory mechanisms of the fracture healing process | A significant portion of bone fractures fail to heal properly, increasing healthcare costs. Advances in fracture management have slowed because translation barriers have limited generation of mechanism-based explanations for the healing process. When uncertainties are numerous, analogical modeling can be an effective strategy for developing plausible explanations of complex phenomena. We demonstrate the feasibility of engineering analogical models in software to facilitate discovery of biomimetic explanations for how fracture healing may progress. Concrete analogical models—Callus Analogs—were created using the MASON simulation toolkit. We designated a Target Region initial state within a characteristic tissue section of mouse tibia fracture at day-7 and posited a corresponding day-10 Target Region final state. The goal was to discover a coarse-grain analog mechanism that would enable the discretized initial state to transform itself into the corresponding Target Region final state, thereby providing an alternative way to study the healing process. One of nine quasi-autonomous Tissue Unit types is assigned to each grid space, which maps to an 80×80 μm region of the tissue section. All Tissue Units have an opportunity each time step to act based on individualized logic, probabilities, and information about adjacent neighbors. Action causes transition from one Tissue Unit type to another, and simulation through several thousand time steps generates a coarse-grain analog—a theory—of the healing process. We prespecified a minimum measure of success: simulated and actual Target Region states achieve ≥ 70% Similarity. We used an iterative refinement protocol to explore many combinations of Tissue Unit logic and action constraints. Workflows progressed through four stages of analog mechanisms. Similarities of 73–90% were achieved for Mechanisms 2–4. The range of Upper-Level similarities increased to 83–94% when we allowed for uncertainty about two Tissue Unit designations. We have demonstrated how Callus Analog experiments provide domain experts with a fresh medium and tools for thinking about and understanding the fracture healing process.
| Translation barriers have limited the generation of mechanism-based explanations of fracture healing processes. Those barriers help explain why, to date, biological therapeutics have had only a minor impact on fracture management. Alternative approaches are needed, and we present one that is intended to help develop incrementally better mechanism-based explanations of fracture healing phenomena. We created virtual Callus Analogs to simulate how the histologic appearance of a mouse fracture callus may transition from day-7 to day-10. Callus Analogs use software-based model mechanisms, and simulation experiments enable challenging and improving those model mechanisms. During execution, model mechanism operation provides a coarse-grain explanation (a theory) of a four-day portion of the healing process. Simulated day-10 callus histologic images achieved 73–94% Similarity to a corresponding day-10 fracture callus image, thus demonstrating feasibility. Simulated healing provides an alternative perspective on the actual healing process and an alternative way of thinking about plausible fracture healing mechanisms. Our working hypothesis is that the approach can be extended to cover more of the healing process while making features of simulated and actual fracture healing increasingly analogous. The methods presented are intended to be extensible to other research areas that use histologic analysis to investigate and explain tissue level phenomena.
| Annually, there are approximately 15 million fractures in the United States, and a significant portion (10–15%) fail to heal properly [1]. Both numbers and costs are predicted to increase as the population ages and as the number of osteoporosis-related fractures increases [2]. Therefore, developing intervention strategies to stimulate fracture healing is expected to positively impact health. Many of the advances made in fracture management in recent years were in mechanical stabilization and biologic bone augmentation materials such as autogenous bone graft, synthetic bone ceramics, or demineralized bone matrix [3]. The clinical impact of biological therapeutic agents, such as bone morphogenetic proteins, has fallen short of expectations for largely unknown reasons [4]. It is noteworthy that the gold standard, and most commonly used strategy for fracture nonunion treatment, autogenous bone graft, has not changed in the last 100 years [3, 5]. Introductions of new therapeutics have slowed despite expanded research [6]. Such ineffectiveness reflects significant translation barriers. The problem is not unique to fracture-healing research; it is encountered within many research domains [7, 8].
A translation barrier exists when mechanistic understanding of a particular medical process, such as fracture healing, is insufficient to posit a reliable, efficacious intervention strategy. A goal of the research described herein is to develop and demonstrate feasibility for a simulation-based approach, facilitating incremental improvement to a plausible mechanism-based understanding of fracture healing processes. We are not yet aspiring to utilize simulation methods to discover new mechanistic insights; knowledge is currently too sparse to support doing so. However, the approach that we employ does provide a novel means to explore and think more deeply about plausible virtual (implemented in software) mechanism-based fracture healing processes. Our approach is intended to be extensible to other processes that, like fracture healing, benefit from histologic analyses. We aim for our model mechanisms to follow a design such that it is straightforward to make them incrementally more biomimetic and fine-grained as new wet-lab knowledge becomes available.
Before proceeding, we need a concise definition of a mechanism. In S1 Text, we provide several definitions of a mechanism which are drawn from literature sources. In support of achieving the above research goals, we are using the more detailed definition developed by Darden [9]. Paraphrasing, a biological mechanism is concrete and can be defined as a real system of entities and activities orchestrated so that it produces the phenomenon of interest, which for this work can be a feature of the fracture healing process. Thus, a model mechanism is a system of biomimetic software entities and activities organized such that, during execution, the process produces a phenomenon that is analogous to one or more features of the fracture healing process in particular ways. A model mechanism capability essential to achieving our research goal is that it facilitates hypotheses about corresponding plausible underlying features of the biological mechanism, which produces the fracture callus attribute being simulated.
Fracture healing is described as comprising two phases and three stages that overlap temporally over several weeks: anabolic and catabolic phases; and inflammatory, endochondral, and coupled remodeling stages. The dominant cell types and subprocesses [10] change as healing progresses. Recent analyses of transcriptomes present during fracture healing have shown that most of the genes and signaling pathways that are involved in skeletal development in embryos are also expressed in cells of the fracture callus [11]. Consequently, some pathway components have become the focus of empirical research efforts to develop therapeutic interventions [10], despite the fact that there is no model of explanation—even at a coarse-grain—for stages in the fracture healing process.
Core phenomena of embryogenesis and some types of tissue regeneration include the evolving small- and large-scale patterns that are readily apparent in recorded images. There has been considerable progress in developing mechanism-oriented explanations for those phenomena [12]. However, stained tissue sections of mouse tibia fractures obtained at intervals of several days lack the hallmarks of orderly, organized evolving phenomena exhibited by embryogenesis. The strikingly less organized callus tissue obscures the ongoing order of the various subprocesses and their mechanisms. Part of the problem traces to limitations of experimentation. Healing of mouse tibia fractures typically spans four-to-five weeks. A major complication is that, within the same experiment, no two fractures are the same. Although the healing phenomenon is the same, the unfolding healing subprocesses within each callus are unique. Large observational gaps coupled with the necessary limitations of standard histological techniques means that informative subprocesses or phenomena may be missed. It is also plausible that informative phenomena—patterns and features—are being observed and recorded, but are not yet recognized as such.
Analogous circumstances have existed in non-biological domains, and significant progress has been achieved using computational and grid-based simulation methods to provide plausible model representations of the missing processes and phenomena. For example, looking for improved insight into processes occurring at the interface of ecology and geomorphology, Fonstad opined, “we have thousands of such images, but no theories in geomorphology nor ecology can fully explain the patterns in any of them” [13].
The fact that callus mechanisms have been successfully healing bone fractures for more than 150 million years [14] implies the existence of a well-orchestrated, robust process. Similarities of callus and embryonic transcriptomes support that inference [2]. If we accept the premise that fracture healing is a well-orchestrated, robust process, then we need to answer this question: how can we begin developing a theory about the healing process—even if initially coarse and somewhat abstract—so that we can begin theorizing about its orchestration? A clearly described phenomenon is a precondition for developing a theory intended to explain that phenomenon (S1 Text). However, we do not yet have a clear temporal description of the fracture healing process, or even for portions of the process. We do, however, have detailed descriptions of features of the process at different stages.
With current technology, it is not feasible to measure a callus continuously. Likewise, it is infeasible to track the changing variety of local structures and cell types. Must we plead for more data, and then come back to the problem in another decade or two? Absent a plausible explanation and theory to test, more data may not be the answer. In discussing comparable issues at the ecology-geomorphology interface, Fonstad observed that, “both of these disciplines are data-rich … it is immediately apparent that both of these disciplines are far more theory-poor” [13]. Fracture-healing research is handicapped because it is relatively data-poor and theory-poor. So, although we can draw inspiration from the explanatory, pattern-oriented simulation methods used by Fonstad and others, their models and those pattern-oriented techniques are not yet applicable in advancing fracture-healing research.
Given the growing interest in increasing the clinical relevance of modeling and simulation research, it is not surprising that the number of such reports in which authors utilize histology images to support face validation and/or guide calibrations is also increasing. The following are three recent examples. Marino et al. [15] utilized their model of lung granuloma formation to compare in silico granulomas to those of the nonhuman primate Macaca fascicularis. Gardiner et al. [16] utilized an agent-based particle system at various granularities to simulate mechanical behaviors of cells and tissues. Simulations using selected parameterizations bore a close resemblance to histological observations of an epithelial layer, cell clusters, and single cells. Ziraldo et al. described an agent-based model of ischemia/reperfusion-induced inflammation coupled with pressure ulcer formation and progression in humans with a spinal cord injuries [17]. Serial photographic images spanning several clinical stages were used to calibrate progression and healing of virtual pressure ulcers. Virtual pressure ulcers were interrogated to explore how and when a irritation might resolve or become chronic.
The prospect of pulling together a start-to-finish tissue-level mechanism-oriented description of a fracture healing process, even one that is initially coarse-grain, seems distant. Why? It is a consequence of four interrelated obstacles arising from fracture-healing research using rodent models.
Given those obstacles, current knowledge and methods are insufficient to describe, much less begin building a conventional molecular and cellular biology-based model of fracture healing. The most pressing current need is to develop strategies and methods to circumvent and eventually overcome each of the above four obstacles. We conjectured that the software-based model mechanism methods, which we have used successfully in other contexts (e.g., see [22–27]), could provide the foundation for such strategies, even though, for those earlier applications, considerably more mechanism related fine-grain knowledge was available. Briefly stated, the cited software-based model mechanism approach begins with a target phenomenon. We build an extant (actually existing, observable), working mechanism in software that is parsimonious and, based on similarity criteria, exhibits essentially the same phenomenon. Doing so requires making no assumptions about the biology.
However, even when the mechanism is kept coarse-grain, the space of possible software mechanisms capable of generating essentially the same phenomenon can be huge. So, biologically inspired requirements and constraints along with mechanism granularity limits are imposed incrementally to shrink and constrain possible model mechanism space. That process shrinks a large set of possible coarse-grain mechanisms into a much smaller set of plausible, incrementally more likely and increasingly biomimetic, model mechanisms.
For fracture healing, we envision simulations generating plausible scenarios for how discretized features of a callus tissue section on one day might transform progressively into the tissue section features—target features—observed several days later. Wet-lab experiments can target differences in two model mechanisms, where the resulting new evidence is expected to support one mechanism and falsify the other (as in [28]), further shrinking plausible mechanism space. At that stage, the surviving software mechanism can stand as a coarse-grain theory for how a portion of mouse tibia fracture healing occurs.
Eroding the four obstacles in meaningful ways requires coupling the preceding methods with an important new capability: use of image interpolation strategies to build plausible sequential image models of the same fracture at different stages of the healing process. A prerequisite for an interpolation strategy is having and aligning discretized coarse-grain models of tissue section images of tibia fractures from different mice at different times.
We report results of a focused demonstration that meets the above requirements. We present results of workflows that support the feasibility of the approach, while also bringing its weaknesses into focus. For this demonstration, we limited attention to the critical interval from day-7 to day-10 during healing of a mouse tibia fracture and focused on discretized models of specific tissue sections on both days. From the latter, we obtained the initial state and final target state for our simulations. Biomimetic software mechanisms involving actions of quasi-autonomous tissue units spanning, typically, 5,000–6,000 time steps are responsible for simulated healing. Similarities (defined in Methods) between simulated and referent final states ranged from > 73% to > 93%, depending on the nature and stringency of the Similarity criterion. Despite the narrow focus, it is clear that a major benefit of the approach demonstrates that simulation experiments can enable discovering, challenging, and improving theories of healing subprocesses.
Because our approach and methods are unconventional, somewhat new, and still evolving, we present that information next under Methods to provide the context needed to present and discuss results. There are weaknesses and limitations associated with every aspect of our approach. Some are identified in Methods, and others are addressed under Discussion. We undertook this demonstration with the expectation that the more successful methods could be repurposed to begin lowering similar barriers faced within some domains of disease progression research.
We begin with a synopsis of our approach from a workflow perspective, as diagrammed in Fig 1. We then provide details on methods used during each of the six stages. In several places, we also provide essential background information that influenced decisions for a particular stage. Words, such as tissue, mechanism, healing, and process, are used in discussing actual mouse tibia fracture healing and corresponding simulations. To reduce confusion, we capitalize those words hereafter when discussing Callus Analogs.
We focus on the day-7 to day-10 interval of mouse tibia fracture healing because histomorphological evidence indicates that the relative contributions of chondrogenesis and osteogenesis may undergo important changes during that interval. The goal is to develop a concrete, quantitative (and thus challengeable) but partially coarse-grain theory that may explain how characteristic tissue level features on day-7 are being transformed into corresponding features observed on day-10. The discovery effort would be greatly simplified if we could obtain collocated day-7 and day-10 tissue sections from the same callus (mouse 1), but that is infeasible. Instead, we used an evidence-based illustration (created by coauthor M.M.) of an envisioned tissue section of the mouse 1 callus on day-10 at the same callus location as the day-7 tissue section. Three domain experts (see Acknowledgments) judged it plausible and acceptable. Hereafter, we refer to the illustration as the day-10i tissue section. A square grid was used to discretize the day-7 and day-10i images. The area of tissue at each grid location was labeled as one of nine tissue types, based on staining and preponderance of cell types within that space. The result (stage 2) was a discretized coarse-grain model of the day-7 and day-10i tissue sections.
Because we are at the beginning of this explanatory discovery process, we needed to select a target region on which to focus (discussed further under Target Region). From a simulation perspective, the target region has an initial state, which maps to the day-7 tissue section, and a corresponding final state, which maps to day-10i tissue section. During stage 4 we used the MASON simulation toolkit [29] to create a 2D 25×25 Target Region initial state, in which objects representing tissue units are assigned to each grid space. We start with a 2D Target Region to limit uncertainty in tissue type identification and to adhere to our parsimony guideline.
Stage 5 efforts focused on answering the following question: how do we enable the Target Region initial state to transform itself so that the arrangement of tissue types mimics the Target Region final state? The steps followed to answer that question involved iterative refinements (discussed below) and had two objectives. 1) Explore logic to be used by simulated tissue units that enable them to successfully transition into biomimetic final states. 2) In doing so, keep the logic simple and avoid process features that may appear non-biomimetic. Once we had evidence that reasonably biomimetic final states were achievable, we shifted attention to improving the simulated healing process sufficiently to achieve the following quantitative target Similarity value (stage 6): compositional and organizational similarity between simulated Target Region and day-10i final state is ≥ 70%. So doing would support the feasibility of achieving the long-term Fig 1 goals. A simulation that uses concrete objects (simulated tissue units) to generate a process that is analogous to callus healing in several ways is a software analog of the healing process. We call the parameterized software a Callus Subregion Analog. Hereafter, for convenience, we refer to the software as Callus Analog and, in some places, simply Analog.
Histologic slides of sagittal sections through mouse tibia calluses at various stages of healing were available from a previous study. The sections were stained using Hall-Brunt Quadruple to highlight tissue, bone, and cartilage. Shown in Fig 2 are the tissue sections from mouse 1 on day-7 and from mouse 2 on day-10. Coauthor R.M. selected them because they have similar fracture features and exhibit all characteristic callus features.
The following nine distinct microscopic tissue types are common to all normal mouse tibia calluses, beginning before day-7 and extending beyond day-10. We assigned a different color to each tissue type, which was used to colorize a discretized version of Fig 2A.
A microscopic area of callus containing about 20 or more cells can be distinguished as being either new marrow (4), new bone (5), hypertrophic cartilage (6), mature cartilage (7), or young cartilage (8) based on the characteristic heterogeneous mix of cell types, the dominant cell type, and extracellular matrix. As healing progresses the mix of cell types within a microscopic area changes. Some areas may undergo multiple tissue type transitions. A working hypothesis is that each of the microscopic tissue types is engaged in somewhat different activities, which are integral to the overall healing process.
The first stage 2 task was to select a square grid mesh size and overlay it on Fig 2A. Choice of mesh size was somewhat arbitrary. If it is too fine, there are fewer cells within the microscopic area and so the uncertainty in specifying the dominant cell type increases. If too coarse, the fraction of microscopic areas containing clearly distinguishable tissue types 4–8 decreases, rendering a single tissue assignment inadequate (and actions of the analog counterpart would likely require unique logic). A guideline for selecting grid size was that the cellular heterogeneity observed within the larger local callus area be reasonably preserved in the discretized image. For example, for a macroscopic region characterized by a heterogeneous mix of predominately ~ 60% new marrow (gray) and ~ 40% osteoblasts (burgundy; new bone), the discretized image counterpart should be a mix of ~ 60% gray and ~ 40% burgundy tissue units. We selected a mesh size that corresponds to an 80×80 μm area in Fig 2A, which typically contained roughly 40 cells, and overlaid that grid on the day-7 and day-10i tissue sections. We then designated each microscopic area as being one of nine concrete, quasi-autonomous, Tissue Unit types, where the behavior of each Tissue Unit type was controlled by a software agent. Fig 3 contains the resulting discretized, colorized images.
Although physically correct image interpolation (e.g., between day-7 and day-10i) is infeasible, sophisticated image interpolation methods, as demonstrated by Stich et al. [30], are available to create high-quality, convincing model images that represent unobserved transitions between recorded images of the same object. The criterion for an acceptable interpolation used by Stich et al., is qualitative: the interpolated images are perceived as visually correct by human observers. During stage 1, we faced the more daunting problem identified in Fig 4: we needed an image that plausibly anticipates the appearance of the mouse 1 fracture if it had been sectioned on day-10 rather than day-7. Starting with the features evident in Fig 2A, and drawing on the tissue features in Fig 2B, coauthor M.M. created an illustration of the envisioned mouse 1, day-10i section. It was judged plausible and acceptable by coauthor R.M. and, separately, by three independent domain experts (see Acknowledgments), thus concluding stage 1. Clearly, a different medical illustrator, one knowledgeable about callus progression, would create a somewhat different illustration. However, we suggest that variability introduced by such illustrations will not add measurably to the considerable variability and uncertainties already present, as illustrated by Fig 4.
To demonstrate feasibility, we needed to designate a Target Region, but first, we needed to select a portion of Fig 2A in which to locate the Target Region. For the latter, we selected the yellow-boxed area in Fig 2A. It is bordered on one side by bone and marrow cavity, which means that transitions in that area will be focused rightward, rather than occurring in two or more directions. Because that area, and the corresponding region in Fig 2B, exhibit similarities, we conjectured that the variety of feature changes occurring during transition from day-7 to day-10i might be representative of key healing features occurring elsewhere in the callus during that 4-day interval. There is no indication that unique healing features may be occurring within this area but not elsewhere during that 4-day interval.
Specifying the size of the target region is subject to opposing constraints. If the region is too large, with a large variety of tissue transition types, we run the risk that the process of discovering plausible and parsimonious logic to direct transitions will become unwieldy, possibly even problematic. If the region is too small, the variety of transition types may be too few to enable adequately simulating Target Region final state. We selected the 25×25 grid region designated by the white box in Fig 3A. Fig 3B shows the corresponding Target Region final state.
Limiting attention to just one target region can be viewed as a weakness. On the contrary, it is an essential part of a recognized, long-term mechanism-discovery strategy that can build on methodological lessons learned while using the Iterative Refinement Protocol in other contexts [23, 25, 27, 28, 31]. That strategy employes variations of the forward/backward chaining (described in S1 Text) and requires selecting a Target Region (stage 3, Fig 1).
After we achieve stage 6 for the day-10i Target Region (described below), we envision expanding the temporal reach of Callus Analog Mechanisms along the dotted line illustrated in Fig 4 to include an earlier stage within that same Target Region, such as day-4i, and a later stage, such as day-14i, and doing so all while continuing to simulate the original day-10i Target Region. Those objectives are illustrated by two unshaded bars labeled a and b in Fig 4. Further, the histological evidence suggests that, on the same day, different subregions within a callus can be at somewhat different stages of repair and may progress at different rates. Given that, a parsimonious strategy is to select separated target regions within the same callus and develop simulations for each in sequence. They could be treated as independent modules. Future work based on simulations of independent target regions will help bring regional issues into focus prior to engineering their merger. The process of merging initially independent modules into a unified model of a tissue healing process would occur further downstream. Given that this work strives to establish the feasibility of the Fig 1 approach, it is efficient to focus first on one Target Region.
Simulation requirements—and thus software requirements—flow directly from desired use cases [32]. In the Introduction, we stated that a primary use case is exploratory simulations capable of the following: aiding image interpolation and providing plausible explanations for how callus features are progressively transformed, all while shrinking the space of possible explanatory transformation scenarios. The last two requirements involve generation of plausible mechanism-oriented explanations, illustrated in Fig 1.
To realize use cases, we employ the virtual experiment approach described by Kirschner et al. [33], along with enhancements drawn from Smith et al. [28] and Petersen et al. [31]. In doing so, the methods employed must meet the following three requirements, which are based on broader sets of requirements discussed by Hunt et al. [32].
To achieve requirement 2, Callus Analogs are written in Java, utilizing the MASON multi-agent simulation toolkit [29]. The data presented herein along with Callus Analog code are available [36].
We customized the established Iterative Refinement (IR) Protocol [22, 27, 28, 31, 32] to meet the challenges evident in Fig 4. Given a software Mechanism that may explain a specified attribute and a virtual experiment design, the goal of an IR Protocol cycle is to test this hypothesis: upon execution, simulation features will mimic the target attribute within a prespecified tolerance. A concrete software mechanism can be falsified—shown to be inadequate—when, during the course of many Monte Carlo trials, it too often fails quantitatively to achieve its objective and/or exhibits non-biomimetic features. It is from encountering and overcoming such failures that explanatory insight improves. Each falsification improves credibility incrementally and shrinks plausible Mechanism space. Our customized IR Protocol follows:
Well-organized processes are responsible for the callus remodeling that occurs between day-7 and day-10. Our operating hypothesis is that information available in day-7 and day-10 tissue section images can be used to draw simplified inferences about unobserved transitions that occur during intervening days, analogous to the approach used by Stich et al. in simulating unobserved transitions between recorded images of the same object [30].
A Callus Analog Mechanism is a system of biomimetic software entities and activities organized such that, during execution, it produces a representation that is measurably similar to the day-10i Target Region. Feature changes within the Target Region explain how the Phenomenon is generated. Stained tissue sections provide snapshots of the healing process. To be explanatory, a biological mechanism will exhibit the fourteen features identified in S1 Text. Because Callus Analog Mechanisms exhibit those same features (also identified in S1 Text), the two processes may be analogous.
Simulations are discrete time; time advances in steps. Fig 5 shows the Target Region initial and final state. The Process responsible for transitioning from initial to final state is the top-level Mechanism. Changes within local subregions from one time step to the next are lower level Phenomena. The lower level Mechanisms responsible for those changes are characterized by individual TU changes, which are controlled by the logic that governs TU agent actions during each time step, discussed below. During each time step, each TU agent, selected randomly, has one opportunity to update and act, based on changes that have occurred within its Moore neighborhood since it last updated. An action may change a TU type or one of its Moore neighbors. Coauthor R.M. identified the following as allowed but not required biomimetic transitions.
At each time step, the current simulated Target Region is compared to the Target Region final state and percent Similarity is calculated as follows:
%Similarity=100(∑(i=2−8)(CSi/CFi)n(CSi/428))t,
where t designates the time step, and i specifies the TU type, 2–8. CSi is the count of TU type i in the simulated Target Region; CFi is the count of TU type i in Target Region final state; 428 is the number of active TUs in Target Region; n = 1 if CFi > CSi, and n = –1 if CSi > CFi.
A case can be made that, if there is strong similarity between gray and burgundy TUs in the simulated and actual Target Region final states, then the similarity score should not be penalized because there are too many simulated gray TUs and too few simulated burgundy TUs, or vice versa. New marrow (gray) and new bone (burgundy) are always formed together. Thus, in some cases, the decision to designate an 80×80 μm area of stained tissue section as either gray or burgundy can be arbitrary; two experts may make different assignments. There are several ways to address that issue but they involve adding at least one new TU type. Given our strong parsimony guideline and the fact that we are at a very early stage in developing the Fig 1 approach, we elected to also calculate a Similarity value when gray are burgundy treated as the same during the calculation. The resulting value is designated Upper-Limit Similarity, simply UL-Similarity hereafter. A more realistic value may be between Similarity and UL-Similarity.
When working to discover plausible model mechanisms, there is a risk that the modeler, subconsciously or otherwise, will favor Mechanism features and logic that ensure that outcomes of generated behaviors are as the modeler thinks that they should be. We strove to eliminate that risk by adhering to the guidelines in steps 4 and 6 of the IR Protocol.
In developing model mechanisms, we did not aim to include established biological features. Instead, we worked to develop model mechanisms that did not contradict known biology. Along the same lines, our model mechanisms were developed not to specifically describe characteristics of the fracture healing process, but instead to allow for a new, coarse-grained manner in which to think about the process. Additional information about the disadvantages of absolute grounding can be found in [40]. By adhering to a strong parsimony guideline (IR Protocol step 2), we avoided adding unnecessary details; doing so enabled us to avoid inscription error. More details on overfitting and analog-to-referent mappings can be found in the Discussion section of Kim et al. [26]. In the subsections that follow, we describe four workflow stages, designated Mechanisms 1–4.
What further improvements in similarity values, as calculated above, might reasonably be achieved? We answered that question with Mechanism 4 internal control calculations that draw on the fact that the closest Similarities that can be achieved for a simulated day-10i Target Region will be those achieved by independent executions of the Mechanism that generated it. The analog Healing Process from day-7 initial state to a simulated day-10 final state is unique for each Monte Carlo execution of Mechanism 4. We selected one Mechanism 4 Monte Carlo execution from 25 and recorded its Target Region configuration at the time step for which simulated Target Region final state maximum Similarity was achieved. We designated it to be the internal control simulated day-10i target state. We then measured maximum Similarity of each of the other 24 Monte Carlo Healing Processes to that simulated day-10i target state. Data are available and labeled as “Internal Control” in S1 Dataset.
Callus Analogs are a form of data, using both the implicit schema of MASON/Java and the explicit schema of configurations. Analog and configuration data are maintained, archived, and released using the Subversion version control tool in two repositories: one public for collaboration and one private (Assembla) for rapid and prototyping development with project partners [36, 41]. Input-output (I/O) data is handled separately. Smaller data sets (tissue data) are stored in CSV format.
The entire Callus Analog toolchain is open-source, thereby enabling repeatability. Similarly, all generated and released data from the project is licensed and available as open data. Callus Analogs are built and maintained for a cloud environment (e.g., Google Compute Engine) to ensure platform and infrastructure repeatability across future experiments, project team members, partners, and the wider community.
The protocols, procedures, and methods that we employ to ensure that results of Callus Analog experiments are reproducible and to establish credibility that they are scientifically useful are described in detail in [28]. Except as noted, we followed those best practices during the workflows described above. They include 1) quality assurance and control protocols, 2) face validation, 3) verification procedures for model mechanisms, 4) repeatability, 5) methods for generating narrowly focused predictions, and 6) Callus Analog validation methods used during IR Protocol cycles.
However, it is too early for systematic sensitivity analyses or efforts to quantify uncertainties associated with particular Mechanism 4 configurations. That is because Callus Analog is still at a very early stage; the focus is on acquiring new insights. Mechanism life cycles are expected to remain relatively short. As soon as we target additional attributes, it is likely that Mechanism 4 will be falsified (because it cannot achieve those new targets). Thus, it will be necessary to alter Mechanism 4 during additional cycles through the IR Protocol. We acquire evidence that we are adhering to our strong parsimony guideline at IR Protocol step 2 in part through documentation of sources of uncertainty and focused assessments of simulated final state sensitivities to modest changes in the logic followed by each TU.
Each execution of our Mechanisms is a unique, simulated healing process, which may (or not) mimic an interval of actual fracture healing. Because the focus is on similarity to the day-10i Target Region, we refer to the simulated state having maximum Similarity value for a given Monte Carlo execution as the simulated day-10i Target Region final state. However, we currently have no data to guide mapping time steps to wet-lab clock time.
Summarized maximum Similarity results for Mechanisms 2–4 are included in Table 1 and the complete results are included in S1 Dataset. Examples of simulated Target Region final states for Mechanisms 2–4 are provided in Fig 9. S1 Video is an example of the complete simulated healing process for Mechanisms 2. It includes the final state having the largest Similarity value. Mechanism 2 improved upon Mechanism 1, generated reasonable simulated states, but failed to meet the biomimesis requirement. A feature of a simulated healing process that has no known real counterpart is designated non-biomimetic. Fig 9A shows the simulated Target Region having the largest maximum Similarity value from a set of 25 executions. Although maximum Similarity values over 70% were achieved for Mechanism 2, at least two non-biomimetic features were observed. 1) In simulated final states, teal TUs were absent from the southeast region of Target Region and that reduced Similarity values significantly. The small islands of teal and blue TUs within a gray/burgundy region in Fig 9A may also be non-biomimetic features. 2) Temporal profiles of Similarity values reached a plateau prior to time step 5000, which persisted beyond time step 20,000. Consequently, the time step at which maximum Similarity occurred appeared somewhat random.
Fig 9B shows the simulated Target Region having the largest Mechanisms 3 maximum Similarity value from a set of 25 executions. S2 Video includes Fig 9B. Table 1 data shows that Mechanism 3 improved on Mechanism 2. However, the absence of blue TUs adjacent to teal in the lower right limited maximum Similarity values and may be a non-biomimetic feature.
Fig 9C and 9D, which are included in S3 and S4 Videos, are examples of simulated Target Regions having the largest Mechanisms 4 maximum Similarity value from the set of 25 executions summarized in Table 1. No non-biomimetic features were observed. Based on simple qualitative visual comparisons, we judged the similarity of Fig 9E–9H to the day-10i Target Region to be comparable to that of Fig 9C and 9D, even though their Target Regions had smaller maximum Similarity values. That observation indicates that, moving forward, improved measurements of Similarity will be needed. S5 and S6 Videos includes Fig 9E and 9F.
Each Mechanism 4 is a unique, simulated healing process, which is intended to mimic the interindividual variability of actual fracture healing. To observe and measure that uniqueness, single Mechanism 4 executions were selected from those used to provide the summary results for Mechanism 4 in Table 1. A Similarity value was calculated and plotted each time step in Fig 10.
Looking for such features and for wet-lab evidence that pushes the decision either way should be part of future research Callus Analog research. We saved the images corresponding to the maximum Similarity for both sets of 25 Mechanism 4 executions summarized in Table 1. Similarity value is just one way to judge good overall biomimicry.
Assessments of simulated final state sensitivities to modest changes in TU logic help identify sources of uncertainty. They also provide evidence for how tightly we are adhering to our strong parsimony guideline. For example, changing action and event probabilities in Mechanism 4 by ± 10–15% for a particular TU produces changes in Similarity value for simulated final state and temporal profiles that are well within the range produced by 25 Monte Carlo executions; the behavior space of Mechanism 4 is not significantly altered. The following is a specific illustration. We changed the Moore neighborhood probability values for blue (randomly chosen) in Fig 8A from [0.2 north/south and 0.6 east] to [0.165 north/south and 0.67 east] and repeated the 25 executions tabulated Table 1 using the same seeds. The average maximum UL-Similarity was 93.6% vs. 93.3% in Table 1; and for Similarity it was 77.8% vs. 77.6% in Table 1. For mean time step at which those value occurred, the new (vs. Table 1) time step was 5317 (vs. 5274) for UL-Similarity and 5299 (vs. 5244) for Similarity. However, changing how and when invasion of TUs from the north is triggered is an example of a change that can have a more significant influence: for such changes, the behavior space of Mechanism 4 can be significantly altered. Data labeled “Blue Probability 2” in S1 Dataset reflect this experiment.
Given the reality illustrated in Fig 4, the prospect of discovering plausible mechanism-based explanations of fracture healing by relying solely on results of wet-lab experiments is problematic. We sought computational methods that could be facilitative, and we ruled out pattern-based methods because the generative methods cannot be made biomimetic [42]. We also ruled out established biomedical multiscale modeling and simulation methods, as recently reviewed by Walpole et al. [43]. A requisite for those methods is sufficient information and knowledge to provide an adequately detailed mechanism-based explanation that is believed to account for the phenomenon of interest. That requisite can be met when the focus of the research can be characterized as being right-of-center on the four Fig 11 spectra. Each spectrum represents a broad attribute of the research. Location on the spectrum characterizes the current reality. While the conventional inductive methods of multiscale modeling and simulation are located right-of-center, fracture-healing research cannot yet meet that requisite because it is characterized by locations on all four spectra that are considerably center-left. For center-left locations, combining analogical models (e.g., electrical, mechanical, chemical, systems engineering, etc.) and analogical reasoning is a proven alternative strategy for developing plausible explanations of phenomena [34, 35]. But because there are no material systems that can serve as analogical models, we focus on developing analogical model mechanisms in software.
Fracture-healing research focus issues are located center-left on the four spectra. Consequently, there are many equally possible explanations for the healing phenomenon. That reality is illustrated in Fig 4 and by the green curve in Fig 11. The conventional strategy is to perform new experiments that generate new data and knowledge. In doing so, one’s location shifts rightward and that shrinks the number and variety of equally possible and plausible explanations. Given the Fig 4 reality, that approach is not available. As explained in [22] and demonstrated in [23, 24, 26, 28], an advantage of utilizing a software-based model mechanism approach is that, by keeping model mechanism entities parsimoniously coarse-grain and concrete, we limit the variety and space of model mechanisms capable of generating the targeted phenomenon.
By showing that Mechanism 2 was inadequate, we eliminated it from further consideration along with all finer-grain variants of Mechanism 2. By so doing, we shrank the space of possible model mechanisms. In moving from Mechanism 2 to 3, the granularity of model mechanism entities is unchanged, but their activities are changed. Thus, we are working within a marginally smaller space of possible model mechanisms. As demonstrated in [28], we can reduce that space significantly by increasing the number and variety of targeted phenomena—validation targets—that must be generated by the analog system, and that is our plan moving forward.
During our effort to keep the Callus Analog simple, we encountered conflicting demands. To be scientifically useful and facilitate discovery, a Callus Analog must be sufficiently biomimetic in both model mechanism and generated phenomena. Increasing biomimicry requires that we make Callus Analog features finer-grain. However, making a Callus Analog finer-grain absent an evidence-based reason for doing so (as stipulated by the IR Protocol) risks dramatically increasing the space of equally possible explanatory mechanisms. Adhering to a strong parsimony guideline at step 2 of the IR Protocol helps resist that pressure.
Each Callus Analog execution provides a record of an analog Healing Process, which is the top-level analog phenomenon. Executions generate the succession of changes by which earlier states of the Target Region gradually become a simulated Target Region final state. Each video explains how the initial state is transformed into a final state that is measurably similar to Target Region final state. It is too early to claim that strong analogies exist between features of simulated and actual fracture healing for Mechanism 4. Nevertheless, we can hypothesize that, at comparable levels of granularity, the simulated healing processes seen in the S3–S6 Videos have actual fracture healing counterparts. Taken together, each video is a low-resolution (coarse-grain) model of explanation—a theory—that maps to a 4-day portion—day-7 to day-10—of the tibia fracture healing process in a mouse. There are currently no comparably detailed competing theories of fracture healing. Because Callus Analog mechanisms are concrete, they are easy targets for scientific challenge, and it is through that use that we anticipate Callus Analogs will provide scientific value moving forward.
The changes occurring within the Target Region during Mechanism 4 executions are intermediate level phenomena; they have histomorphometric counterparts in callus tissue sections. Two examples are the eastward progression of teal TUs replacing blue TUs and the eastward expansion of the mixture of gray and burgundy TUs. The mechanisms responsible for those intermediate level phenomena are mediated by the individual activities of the participating TUs. The logic dictating TU actions at each time step, as illustrated in Fig 7C, provides Mechanism orchestration. A change in TU type at a particular grid location is the lowest level (finest-grain) model mechanism phenomenon.
Mechanisms 1–4 were developed sequentially. Derived results were most useful when a particular configuration (logic, utilization of neighborhood information, probability values, etc.) failed to meet expectations. Some failures were marked by poor maximum Similarity values. Others were marked by a feature or features within Target Region that was unexpected or judged non-biomimetic. In such cases, we would hypothesize a plausible explanation for the problem and a possible solution, and then conduct experiments to challenge that hypothesis and solution. When successful, we achieved an incremental Analog improvement. Failures provided new knowledge by enabling us to marginally shrink the space of plausible Callus Analog healing processes. Mechanism 1 was unsuccessful because it was flawed in several ways. Nevertheless, observations made during IR Protocol cycles stimulated ideas for other logic that might be explored, including the ideas that drove development of Mechanism 2 and, later, Mechanism 3.
From observations made during explorations leading to Mechanisms 1 and 2, we inferred that, in order to make simulated healing processes more biomimetic, it would be necessary to include two features. 1) Allow for multiple TU changes at any grid location during simulated healing. 2) Have sustained directional influences, spanning many TUs, guiding or constraining the direction and type of TU transitions. The latter may map to the combined net effects of multiple factors such as angiogenic impairment [44], relative abundance and activity of immune cells [45], signaling influencing osteogenic and chondrogenic transcription networks [46], O2 gradients [47], and mechanical influences [48]. Callus Analog has achieved its current objectives without needing to bring any of those influences into focus, consistent with our strong parsimony guideline. As the list of targeted phenomena expands, it will become necessary to make model mechanisms finer-grain. It is during such refinements that a newly added Callus Analog feature may map to one or more of those influences.
The logic used by Mechanisms 3 and 4 limits the direction in which a TU can affect the transition of a neighbor, and it imposes preconditions on number and type of Moore neighbors that must be present before a TU transition can occur. A consequence of those constraints was the emergence of apparent cohesion of TUs within three areas that are clearly evident in Fig 9E–9H: one area dominated by blue TUs, another dominated by teal TUs, and a third dominated by gray and burgundy TUs. That apparent cohesion is clearly evident during S2–S6 Videos.
From the simulation engineering perspective, Callus Analog could be simplified if those three areas were represented as large, quasi-autonomous, sub-Callus organized units, within which TUs are simply parts under the control of each organized unit. However, there is, as yet, no direct wet-lab evidence that would support or require such simplification. Interfaces between those areas map to well-documented transition zones (e.g., see [39]). Moving forward, features of transition zones will be added to an expanding list of targeted phenomena to further shrink the space of plausible explanatory model mechanisms.
Special attention was given to understanding why, during an IR Protocol cycle, a model mechanism failed. As Petersen states, "having that information is essential to the scientific process because it is falsification that provides new knowledge: specifically, the current (falsified) mechanisms are flawed—they are not a good analogy of the referent biological mechanisms" [31]. Building upon and revising flawed hypotheses offers a new perspective and new way of thinking about plausible networked callus healing processes, and that alternative way of thinking may well become the primary value of the Callus Analog approach.
Fracture healing occurs primarily through the process of endochondral ossification, a process in which cartilage matrix is replaced by bone. This is the same process by which many bones are formed and grow. During endochondral ossification at the fracture site, chondrocytes express vascular endothelial growth factor, which induces vascular invasion of the cartilage matrix [49–51]. Along with the invading vasculature, osteoclasts that have entered the callus degrade the cartilage matrix. Previously, the chondrocytes were thought to undergo programmed cell death [52], and concurrently, osteoblasts, which are delivered by the vasculature [19], replace cartilage matrix with bone. In this two-stage theory, chondrogenesis—cartilage development—serves chiefly as a means for producing hypertrophic chondrocytes, and they, in turn, initiate bone formation (carried out by other cells). However, a competing theory has emerged. Chondrocytes enter a transient stem cell-like state from which they transform into the osteoblasts and osteocytes that form the new bone [37, 38]. The earlier theory envisioned chondrogenesis and osteogenesis as separate processes, whereas the more recent theory is mechanistically simpler: it envisions chondrogenesis and osteogenesis as characteristic sequential features of the same process.
All of the wet-lab observations on which those two competing theories are based are below the resolution of Mechanism 4. They are subsumed by the Fig 7 logic. So, at this stage, the model does not provide evidence for or against either theory. As we add new target attributes downstream, we envision replacing the Fig 7 logic with model mechanism details using the tuneable resolution process of Kirschner et al. [33], which is a systems biology approach for discretized multi-level, multi-compartment computational models. The process involves fine- or coarse-graining of entities and activities. Such an approach allows for the adjustment of the level of resolution specific to a question, an experiment, or a level of interest. At that stage we should be able to challenge those competing theories.
Because we are still early stage, the Mechanism 4-based simulated healing process comes with an ample supply of weaknesses and limitations. Both a weakness and limitation is that there are no (fine-grain) 1:1 counterparts to the cellular entities and molecular level events that are the focus of the majority of wet-lab experiments. As Callus Analog credibility improves and finer-grain features are included as validation targets, it will become feasible to increase model mechanism resolution further utilizing the tuneable resolution process in Kirschner et al. [33].
The day-10i illustration is a stage 1 requirement. It is an important source of information but also a source of uncertainty. A requisite for building an explanation for fracture healing is having staged representations of the same fracture (e.g., on days 4, 7, 10, 14, 20, etc.) that are, within reasonable tolerances, reliable, semi-quantitative, and scientific. There are currently no protocols to achieve that requisite. However, many histologists, pathologists, and biologists are trained in accurately illustrating representations of specimens, including tissue sections. A logical next step would be to acquire two (or more) independently generated day-10i illustrations of the same section and then document where and why they are similar and different. Thereafter, we envision protocols and methods for developing credible staged illustration representations becoming increasingly standardized, and, where feasible, automated. To increase standardization, we can draw from the robust best practices developed over decades for standardization of pathologic and histologic evaluations and reporting. We can also draw on the medical image registration methods [53] that enable rapid advances in computed tomography and magnetic resonance imaging.
Feature discretization and simplification at stage 2 helps manage uncertainties, but the process itself is also a source of new types of uncertainty. There is a risk that increasing or decreasing grid mesh density can alter analog-to-mouse healing process mappings in scientifically meaningful ways. A good future time to assess that risk will be when Callus Analog insights have advanced sufficiently to begin exploring the first testable theory of fracture nonunion.
The cellular components of all 80×80 μm areas within the day-7 tissue section Target Region are heterogeneous, but discretization requires that the corresponding analog grid space be occupied by one of nine TU types. To limit unintended bias, we can, as above, continue to draw from the mature best practices of pathologists to develop protocols to minimize added uncertainties. Longer term, we envision discretization protocols becoming automated. Near-term, several strategies can be explored to discover and ameliorate discretization weaknesses. Here are two examples. 1) Acquire two (or more) independently generated target region discretizations, and then independently develop a plausible explanatory analog for each that achieves the same final state Similarity criteria. 2) There will always be instances where it will be difficult for a domain expert or an automated process to make some TU assignment, such as choosing between a new marrow TU (gray) or an osteoblast-dominated TU (burgundy), because those cell types are similar. In such cases, those grid spaces can be given a new designation, TU = g/b. At the start of each simulation experiment, all g/b spaces are randomly designated as either gray or burgundy. The result is a set of Monte Carlo Target Region starting states. One then cycles through the IR Protocol for each in parallel until the Target Region final state Similarity criterion is achieved. Both strategies require increased work, and automating IR Protocol tasks will help avoid reducing the overall workflow pace.
Selecting an initial target region at stage 3 was essential to demonstrate feasibility. Moving forward, the approach must be expanded in stages to cover the entire callus, possibly as follows. First, develop and improve simulated healing processes for small portions of a callus and then explore how best to merge them incrementally to simulate more of the fracture healing phenomenon. It seems likely that multiple sub-callus processes will be needed to simulate healing within the entire callus. Evidence suggests that different callus subregions can be at somewhat different stages of repair and may progress at different paces. Separate simulations of independent subregions will help bring these issues into focus. A plausible next step would be to select a new day-7 target region (possibly larger than 25×25) and determine if Mechanism 4 is able to achieve a corresponding day-10i final state with Similarity values ≥ 70%. Following that, we envision extending those two analog healing processes forward to day-14 and backward to day-4.
Each Mechanism 4 video is a sample from the circumscribed space of model healing processes, and each is biomimetic. Are all other Mechanisms in that space also biomimetic? It is too early to answer, but it seems likely that the answer is no. Each video (the record of one Monte Carlo trial) provides a means to search for and address the emergence of non-biomimetic features. Observing more videos provides one with a better overall impression of the space of simulated healing. Domain experts observing videos can identify features that may be non-biomimetic, such as the small islands of teal and blue TUs within a gray/burgundy region in Fig 9A. Features that appear in one video may not appear in another. Assume that domain experts identify a likely non-biomimetic feature in several Mechanism 4 videos. Mechanism 4 would be falsified. We would then seek a marginally different—yet still parsimonious—model mechanism in which the logic used by each agent type has been revised to avoid exhibiting the non-biomimetic feature, while still meeting all similarity criteria. The revised Mechanism would circumscribe a smaller set of analog healing processes. In the preceding scenario, the videos provide domain experts with an entirely new means of thinking about callus healing. More broadly, simulated healing provides a new perspective on the actual healing process, and it is from that perspective that we encourage the use of simulations to enhance mechanism discovery. Doing so can help overcome translation barriers through the development of coarse-grained mechanism-based explanations. Later, as additional validation targets are met, incrementally better explanations will shrink the space of possible mechanism-based theories, and putative mechanisms will become finer-grained, which we anticipate will enable novel intervention strategies to be brought into focus.
|
10.1371/journal.pgen.1007133 | Genome wide comparison of Ethiopian Leishmania donovani strains reveals differences potentially related to parasite survival | Leishmania donovani is the main cause of visceral leishmaniasis (VL) in East Africa. Differences between northern Ethiopia/Sudan (NE) and southern Ethiopia (SE) in ecology, vectors, and patient sensitivity to drug treatment have been described, however the relationship between differences in parasite genotype between these two foci and phenotype is unknown. Whole genomic sequencing (WGS) was carried out for 41 L. donovani strains and clones from VL and VL/HIV co-infected patients in NE (n = 28) and SE (n = 13). Chromosome aneuploidy was observed in all parasites examined with each isolate exhibiting a unique karyotype. Differences in chromosome ploidy or karyotype were not correlated with the geographic origin of the parasites. However, correlation between single nucleotide polymorphism (SNP) and geographic origin was seen for 38/41 isolates, separating the NE and SE parasites into two large groups. SNP restricted to NE and SE groups were associated with genes involved in viability and parasite resistance to drugs. Unique copy number variation (CNV) were also associated with NE and SE parasites, respectively. One striking example is the folate transporter (FT) family genes (LdBPK_100390, LdBPK_100400 and LdBPK_100410) on chromosome 10 that are single copy in all 13 SE isolates, but either double copy or higher in 39/41 NE isolates (copy number 2–4). High copy number (= 4) was also found for one Sudanese strain examined. This was confirmed by quantitative polymerase chain reaction for LdBPK_100400, the L. donovani FT1 transporter homolog. Good correlation (p = 0.005) between FT copy number and resistance to methotrexate (0.5 mg/ml MTX) was also observed with the haploid SE strains examined showing higher viability than the NE strains at this concentration. Our results emphasize the advantages of whole genome analysis to shed light on vital parasite processes in Leishmania.
| Approximately 200,000–400,000 new cases of visceral leishmaniasis (VL) occur annually resulting in an estimated 40,000 deaths. Almost 90% of the reported cases are caused by the Leishmania donovani occurring primarily in East Africa and the Indian subcontinent. Parasites in East Africa are more polymorphic than those isolated in other regions, and differences in vectors, biotopes and patient response to drugs are found in Ethiopia. Large-scale whole genome sequence (WGS) analysis of L. donovani strains and clones isolated from VL and HIV+-VL co-infected patients in Ethiopia was carried out. Single nucleotide polymorphism (SNP) and gene copy number variation (CNV) analysis shows genetic differences correlated with geographic regions in Ethiopia. These differences are associated with distinct biological processes and molecular functions, and may be associated with genes involved in drug resistance and parasite survival.
| Leishmania donovani, together with L. infantum, are the main causative agents of visceral leishmaniasis (VL). The World Health Organization (WHO) estimates that this disease causes an estimated 200,000 to 400,000 new VL cases worldwide, and >40,000 deaths yearly. The majority of VL cases occur on the Indian subcontinent, Brazil, and East Africa with most cases in the latter region found in Sudan, South Sudan and Ethiopia [1]. While treatment regimens for VL, including combination therapy based on existing drugs, have improved safety and prognosis, they are still suboptimal, and new drugs are urgently needed. High parasite resistance on the Indian subcontinent to the pentavalent antimonial sodium stibogluconate (SSG) has led to its discontinuation, however SSG is still part of the primary treatment regimen for VL in Ethiopia. The situation in Ethiopia is further complicated by the presence of AIDS, a leading cause of adult illness and death in this country [2, 3]. Between 20–40% of VL patients are co-infected with HIV and relapses, up to 50% at a year post-treatment, are frequently observed [2, 4]. Treatment following relapse normally utilizes alternative drugs like liposomal amphotericin B, pentamidine and paromomycin. Paromomycin is an antibiotic recently approved to treat VL in India and is in clinical trials in Africa [5]. Differences in the dose of paromomycin required to treat VL patients from northern Ethiopia (NE) and Sudan, as compared to southern Ethiopia (SE) and Kenya have been reported [5, 6]. Interestingly, other differences in the ecology, sand fly vectors and parasites have been described between NE and SE. Endemic regions of NE are typically semi-arid, with commercial monoculture fields and scattered Acacia–Balanite forests [7–9]. Phlebotomus orientalis is the primary vector in this region. On the other hand, transmission in SE occurs in areas of savannah and forest where termite mounds abound; and Phlebotomus martini and Phlebotomus celiae have been implicated as vectors [7, 8, 10, 11]. Molecular characterization including multilocus enzyme electrophoresis (MLEE), multilocus microsatellite typing (MLMT), protein and DNA sequence analysis of individual genes, and k26—PCR targeting the hydrophilic acylated surface protein B (HASPB) repeat region have been used to characterize the L. donovani complex in East Africa [7, 12–15]. MLMT analysis indicated that different genetic populations and subpopulations are present in NE and SE [13].
Recent developments in whole genome sequencing (WGS) and computational analysis allow the in-depth exploration and comparison of leishmanial genomes at a high level of resolution and accuracy [16–18]. Genomes of individual Leishmania species [18]; of L. infantum or L. donovani strains in limited geographic regions of Turkey or India/Nepal respectively [16, 19, 20], and strains showing differences in drug resistance and tropism [16, 20–24] have been analyzed by WGS. In this report 41 patient strains and clones from NE and SE are analyzed and compared by WGS. This study provides insights into the population structure, and genetic differences of parasites circulating in the distinct ecologies of Ethiopia. The evidence of genomic variation between the two L. donovani populations (NE and SE) may provide an additional insight about parasite virulence, development of drug resistance and give new directions for new treatment strategies.
WGS of 41 L. donovani parasites, both clones and isolates, from 15 Ethiopian patients was carried out and analyzed (S1 Table). All 18 patient strains were isolated in the years 2009 and 2010 in NE (n = 9) or SE (n = 10). The parasites analyzed were isolated from spleens, bone marrows and skin lesions. Five of the patients had AIDS at the time the parasites were isolated; and in three cases, two strains obtained from different organs (either spleen and skin, or spleen and bone marrow), were analyzed. None of the patients received treatment for VL prior to parasite isolation. A Sudanese strain, isolated in 1998, was also included in some analysis (S1 Table).
Chromosome copy number was predicted based on whole chromosome median read coverage as described in Material and Methods. The predicted values for each chromosome in the patient isolates and their clones are given in Fig 1 and S2 Table. Normalized read depth, showed that 68.5% of the chromosomes had a predicted chromosome (PCHR) copy number of 2 ± 0.5, i.e. disomic, with smaller percentages showing higher copy numbers; i.e., trisomic (3 ± 0.5) in 23.4%, tetrasomic (4 ± 0.5) in 6.7%, pentasomic (5 ± 0.5) in 1% and hexasomic (6 ± 0.5) in 0.4%, similar to what was previously reported for Nepalese L. donovani strains [16]. Several strains, mostly patient isolates and clones belonging to the SE population; show intermediate ploidy (mixoploidy) for specific chromosomes, e.g. chromosomes 1, 6 and 23 (S2 Table), suggestive of parasite populations with mixed polysomic diversity, perhaps due to variation in chromosomal amplification between individual cells in culture. Differences in chromosome aneuploidy between SE patient strain and its clones was not significant (unpaired two samples student's t-test, p = 0.2), suggesting that aneuploidy in the clonal population is not due to selection of clones exhibiting different ploidy for identical chromosomes. On the other hand, there is a significant difference in chromosome aneuploidy (unpaired two samples student's t-test, p = 0.028) when SE parasites, patient and clones, were compared to NE parasites, patient and clones.
Cluster analysis based on chromosome copy number provides insight on three levels: first, that similarities and differences exist between all parasites examined; second, validation that clones derived from the same patient strain are highly similar; and third, that differences in chromosome copy are not correlated with the parasite geographic origin. Similar to earlier studies [16, 19, 20, 25] where aneuploidy was examined in clinical isolates of Leishmania from India and Nepal (L. donovani), in sand flies from Turkey (L. infantum), and in laboratory strains (L. major, L. braziliensis, L. donovani, L. infantum, L. mexicana and L. panamensis); all the Ethiopian L. donovani lines examined showed aneuploidy, and each has a different karyotype. Seven chromosomes (17, 25, 27, 30, 32, 34 and 36) were disomic in all lines examined, as was chromosome 3 with the exception of one strain AM554. Four of these chromosomes (30, 32, 34 and 36) were used for normalization of average chromosome read coverage as described by Rodgers et al [18]. Interestingly, disomy was also observed for five of these chromosomes, 17, 25, 30, 34 and 36, in all the 17 Indian and Nepalese L. donovani lines originally studied [16], and in almost all 206 strains from the Indian subcontinent recently examined [20]. Several other chromosomes (18, 19, 21 and 28) that were disomic in all the Indian and Nepalese lines showed somewhat more diverse ploidy (2 to 3 copies) in the Ethiopian lines. Chromosome 31 was polysomic in all the lines used in this study. Even GR363sk/cl.I and GR363sk/cl.II, which showed the least aneuploidy of all the lines examined, were still trisomic for chromosome 31. Chromosome 31 is polysomic in every Old World Leishmania strain examined to date [16, 19, 25].
Chromosome copy number was compared using clones from several strains (AM560, n = 4; GR356, n = 9; GR363sp, n = 3; GR363sk, n = 3; GR364sp, n = 3; GR383, n = 5). In one case (GR364sk, n = 3) two clones and the original patient strain were examined. Validation of aneuploidy similarities was carried out using clustering based on ploidy patterns taking into account all 36 chromosomes. What is readily apparent, from both the heat map and dendrogram (Fig 1), that in most cases karyotypes of clones isolated from the same strain are highly similar. For instance, the karyotypes of the five GR383 clones (I, II, X, XII and XIII) are almost identical (unpaired student's t-test, p > 0.73), and fall in a tight cluster. This is also seen for all the clones of GR363sp, GR363sk, GR364sk and AM560, and 8/9 clones of GR356. Only two clones, GR356/cl.I and GR364sk/cl.II, show karyotypes significantly different from their sister lines or the patient strain from which they were derived, and fail to cluster with the former (Fig 1). Interestingly, clone GR356/cl.1 groups with its sister lines based on SNP analysis (Figs 2 and 3), and exhibits a k26-PCR amplicon identical to the other NE strains (290 bp). Of note, clone GR364sk/cl.II had a k26-PCR amplicon (450 bp) similar in size to SE strains, even though it was derived from GR364sk, a NE strain with a 290 bp amplicon typical of the NE region [7]. This strain was isolated from a HIV-VL patient, and the sister clone, GR364sk/cl.I which is very similar to the patient strain also has a k26—PCR product of 290 bp. Clone GR364sk/cl.II also groups with the SE strains by SNP analysis (Figs 2 and 3), but is distinct from them suggesting that this patient might have had a mixed infection.
Cluster analysis of the karyotype data doesn’t separate the SE and NE strains by geographic region instead they are mixed together, interspersed among each other. Despite this, some differences in chromosome copy number between the two populations are apparent. Chromosomes 13, 15, 16, 23, 28, and 31 show significantly higher average ploidy for SE strains compared to NE strains, while chromosome 4 is the only chromosome where the NE strains show a significantly higher average ploidy than the SE population (Table 1 and S2 Table). In addition, we found that SE strains show a significantly higher average total chromosome somy (2.52 ± 0.09) than the NE strains, (2.35 ± 0.06; paired t-test, p = 0.0016), perhaps indicating that the SE strains tend towards higher polyploidy than the NE strains (S2 Table).
Finally, parasites from different organs of three HIV-VL co-infected patients were examined. Parasites isolated from the skin or spleen of the same patient form separate groups, and the karyotypes of parasites isolated from the respective sites show greater differences, in most cases, than clones generated from the same site, either skin or spleen. GR364 spleen and skin strains, with the exception of GR364sk/cl.II, form separate clusters (Fig 1). Overall these five lines (skin–original patient strain and clone I, versus spleen—all three clones) show no significant difference in chromosome ploidy distribution, i.e. karyotype. However, when each GR364 spleen and skin chromosome was compared on an individual basis, 6 out of 36 chromosomes (Ld4, 5, 8, 14, 20 and 31) show significant differences in ploidy (p = 0.0004, 0.04, 0.01, 0.05, 0.01 and 0.001, respectively). This probably accounts for the fact that the skin and spleen strains form separate subgroups (Fig 1). Likewise, GR363sk and GR363sp taken from the skin and spleen, respectively, of the same patient cluster in separate branches of the dendrogram (Fig 1). Significant overall differences in ploidy between the GR363 skin and the spleen clones (two tailed paired t-test, p = 0.0067) was observed. Interestingly, the spleen clones from this patient demonstrate on average higher ploidy than the skin clones. Finally, strains isolated from the spleen and bone marrow of one patient, LDS373sp and LDS373bm respectively, also show different karyotypes (7/36 chromosomes differ), even though they cluster together on the dendrogram. These parasites also show different k26—PCR fragment sizes, gene CNV and SNP profiles. An overall difference in ploidy was found in chromosomes 1, 4, 6, 20 and 35 by the comparison of all spleen against skin clones. The differences in chromosome ploidy of strains isolated from different organs may result from clonal selection of the parasites in the host due to specific selective pressures at the infection site.
SNP calling, compared to the L. donovani reference strain (BPK282A1), for each of the strains and clones was carried out as described in material and methods. The two parasite populations, SE and NE, show significantly different number of SNPs on average, ~153K and ~168K (p < 0.043) respectively, compared to the Indian reference strain (R). In addition, the percentage of homozygous and heterozygous SNPs within each geographic population (S3 Table), represented by a single alternate allele (A), show significant differences i.e., for the SE (mean homozygous AA = 87.2% and heterozygous RA = 12.7%, respectively, p<0.05) and the NE populations (mean homozygous AA = 83.3% and heterozygous RA = 16.6%, respectively, p<0.05). The relationship between SE and NE strains and clones based on whole genome SNP pair-wise analysis is shown in Fig 2. Each colored square in the matrix indicates the percent SNP similarity for a strain/clone listed on the left compared to strain/clone listed along the bottom of the matrix. Unlike the chromosomal aneuploidy profiles (Fig 1), SNP population analysis divides the L. donovani population into large clades or groups based on geographical distribution (Fig 2). This is easily seen both in the hierarchical cluster tree where SE and NE parasites each form separate groups with clones from each strain showing highest similarity to each other (Fig 2A) and the heat map (Fig 2B). The only exceptions are three atypical strains/clones (AM422, AM553, and LDS373bm) which fall outside the main clades, and clone GR364sk/cl.II, as mentioned above, which groups with the SE rather than the NE parasites. The two atypical SE strains; AM422 and AM553 fall closest to the NE clade, yet are distinct from the NE strains. Interestingly, LDS373bm does not cluster with the spleen strain, LDS373sp, isolated from the same patient, even though the karyotypes are similar. AM422, AM553 and LDS373bm seem to have SNP profiles falling between the NE and SE populations.
Principal component analysis on the SNPs was also used to examine the population structure. SNPs showing high linkage disequilibrium were removed prior to analysis by SNP pruning [26] (S4 Table). As can be seen in Fig 3A, two clusters or near-clades containing most the NE (black circles) or SE (red circles) strains and clones are observed. The two atypical SE strains, AM422 and AM553, are outliers falling far outside both clusters (EV1 = 0.229, EV2 = -0.954 and EV1 = 0.945 EV2 = 0.265 respectively), while the two atypical NE strains, GR364sk/cl.II and LDS373bm, group together with the rest of the SE population confirming the results shown in Fig 2. PCA was repeated excluding the two atypical SE strains to allow better separation of the remaining 39 isolates (Fig 3B). All the SE strains still group together in one dense cluster that includes GR364sk/cl.II suggesting that these strains are more homogenous, however the NE strains show more diverse distribution with each strain and its clones forming a separate cluster. LDS373bm no longer groups with the SE strains.
SNPs in protein coding regions were examined, and SNPs unique to the SE and/or NE populations identified. No significant difference in the percentage of synonymous, nonsynonymous or nonsense mutations was found between SE and NE parasite populations: 47%, 52.6% and 0.4% versus 50.0%, 49.8% and 0.2%, respectively. Altogether 683 common genes containing at least one SNP causing either a nonsynonymous or a nonsense mutation are present in both NE and SE parasites. The remaining SNPs resulting in nonsynonymous or nonsense mutations are only found in either SE (412 genes) or NE (595 genes) parasite populations (S5 Table). As such, SNPs in these genes are unique markers for parasites in each geographic region.
Gene Ontology enrichment analysis of the “unique” SNP containing genes found in each population indicates that the proteins are involved in different biological processes (S6 Table). A web based semantic cluster algorithm, REVIGO, was used to remove redundant GO terms [27]. After removal of redundant GO terms the remaining terms were graphed as scatterplots in two-dimensional space according to semantic similarity. Semantically similar GO terms should remain close together in the plot, and size of the circle indicates the frequency of GO term. Unique SNPs in NE parasites (Fig 4A and S7 Table) are associated with biological processes such as tRNA aminoacylation for protein translation, glutamine family metabolism, regulation of transferase activity e.g. protein kinases, and phosphate ion transport, while those in SE parasites are primarily associated with cation transmembrane transport, purine nucleoside triphosphate and nucleobase metabolism and DNA conformation (Fig 4B). Similar differences are also noted when the molecular functions of the genes with unique SNPs are analyzed. Unique SNPs in the NE population are concentrated mainly in genes involved in glutamine family biosynthesis and metabolism, tRNA aminoacetylation, pyrimidine metabolism, and cyclins involved in protein kinase regulation during cell division. On the other hand, unique SNPs associated with the SE population are found in genes such as glutathione metabolism, protein translation initiation and elongation factors, transport and oxidoreductase activity (S5–S7 Tables).
Several genes associated with the development of leishmanial drug resistance also contain nonsynonymous SNPs and/or nonsense mutations. A unique SNP, only present in the NE population, was identified in the aquaglyceroporin (LdBPK_310030) gene, a protein that plays a role in trivalent antimony (SbIII) uptake, located on chromosome 31 [20]. This unique heterozygote nonsynonymous mutation g.7444A>T causes an amino acid exchange (Ser251Thr) in TML-5 of aquaglyceroporin, and is only found in the NE population (S5 Table). The MRPA gene (PGPA) encodes an ABC-thiol transporter (LdBPK_230290.1) that sequesters thiol-Sb conjugates and is also involved in antimony resistance [28]. This gene contains several unique nonsynonymous SNPs unique to the NE (four homozygote and one heterozygote), and SE (two homozygote) populations (S5 Table). It is not clear how these unique SNPs affect transporter function, as no difference in response to antimonial chemotherapy between L. donovani isolates from NE and SE has been reported.
Comparative read coverage was examined for 41 SE and NE isolates using a sliding window (5000 bp) in order to detect genomic copy number variation (CNV) as described in Material and Methods. Chromosome somy was not taken into account at this stage. While both increases, and decreases in CN were observed (S8 Table), increases in CN (>2) were more prevalent occurring 83% of the time. In addition, a significant overall difference (p < 0.00001) in average CNV between the NE and SE populations was noted (Table 2). Sixty-two different genes showed significant differences in CN between the SE and NE populations (S9 Table and S1 Fig). In the SE strains, genes with an average CN < 1.5 are primarily found on chromosomes 10, 11, 12, 22, 27, 31, 34, and 36; and include the folate-biopterin transporters, ABC transporters, ATP-binding cassette protein, D-lactate dehydrogenase, branched-chain amino acid aminotransferase, amastin-like proteins, phosphoglycerate mutase, tartrate-sensitive acid phosphatase, mitogen activated protein kinase homolog, as well as numerous uncharacterized proteins. Genomic CNV of the atypical NE strains was very similar to that observed for the SE strains. On the other hand, only one gene, an amastin-like protein (LdBPK_341700) on chromosome 34, shows an average CN <1.5 in a majority of NE strains. Interestingly, low copy number genes were more prevalent in the SE population with 84% of all strains and clones exhibiting an average CN < 1.5 over all genes.
One striking difference between the NE and SE strains and clones is in the number of folate/biopterin transporter (FBT) gene(s) on chromosome 10 (Table 3, S10 Table and S2 Fig). Leishmania are auxotrophs for folic acid, and 14 different members of FBT family have been identified in L. infantum [29]. Several of these genes are known to play roles in parasite drug resistance and viability. Eight of the 13 L. donovani FBT homologs are located on chromosome 10, of which 7/8 are present in a tandem array. This chromosome is disomic in most leishmanial strains examined to date [16, 18–20, 25]. Interestingly, three of the FBT genes present in this tandem array on chromosome 10 (LdBPK_100390, LdBPK_100400 and LdBPK_100410) show gene amplification in several NE (GR364sk, GR364sp and GR356) and Sudanese parasites (S10 Table). In addition, LdBPK_355160 on chromosome 35, an ortholog of L. infantum biopterin transporter 1, also a member of the FBT family, is amplified in the NE strain GR383 (CN = 5), but not the other NE isolates.
SE parasites exhibit the opposite trend for these three genes (LdBPK_100390, LdBPK_100400 and LdBPK_100410) on chromosome 10 showing loss of heterozygosity in eight, ten, and ten out of eleven SE strains, respectively (S10 Table). Loss of heterozygosity was also observed for one additional FBT gene on chromosome 10 (LdBPK_100380) in the SE strain AM553. Interestingly, two atypical NE strains, LDS373bm and GR364sk/cl.II, which group with the SE strains by SNP analysis also show loss of heterozygosity for the same three FBT genes on chromosome 10. The average haploid CN taking into account chromosome somy for these three genes is 2.75 in the NE strains versus 1.05, 1.25 and 1.25 respectively in the SE strains (Table 3).
Genomic CN for LdBPK_100400 (L. infantum FT1 homologue) in the NE and SE leishmanial strains was also determined by qPCR (Fig 5) in 20 strains and clones using a novel dual priming oligonucleotide system [30]. qPCR tended to give higher CNs for this gene than found by computational analysis (cn.mops), however there was a good correlation overall between FT1 CN based on computational estimation with cn.mops [31] and qPCR (ρ = 0.91). Strain LDS373bm also showed loss of heterozygosity by qPCR, similar to what was found above by computational analysis.
In addition, both methods show significant difference in FT1 CN between the SE and NE populations. The mean FT1 CN for the two different methods and populations is as follows: qPCR; SEmean = 0.79, NEmean = 2.44, p = 0.00034; cn.mops; SEmean = 1, NEmean = 2.6, p = 0.00009 confirming the trend toward loss of heterozygosity in the SE and amplification in the NE strains / clones examined.
FT1 is the main transporter for folate. Resistance to methotrexate (MTX) is correlated with reduced folate uptake [32, 33], and CN for this gene was reduced in some resistant parasites [29, 34]. Therefore, the effect of MTX on the viability of eight SE and NE strains/clones that vary in FT1 CN was examined (Fig 6). Several of the SE and NE strains examined also show CNV for other FBT genes on chromosome 10 flanking FT1 (S10 Table). All of the SE strains/clones tested are single copy for FT1 and are significantly less sensitive (6–30% growth inhibition) to MTX (0.5 mg/ml) than the NE strains/clones (42–78% growth inhibition), p = 0.005; and a linear correlation (r2 = 0.937) between CN, for the genes demonstrating CNV on chromosomes 10 (LdBPK_100380, LdBPK_100390, LdBPK_100400 and LdBPK_100410) and 35 (LdBPK_355160), and sensitivity to MTX was observed (Fig 6 and S11 Table). NE parasites are significantly more sensitive (p = 0.02 to 0.0009) to inhibition by MTX over a wide range of concentrations (33 to 900 μg/ml) when grown at limiting folate concentrations (S3 Fig). A high correlation between FT1 CN and sensitivity to MTX was found (Pearson's correlation coefficient ρ = 0.85, p = 0.007).
Plasticity in gene organization has been reported for several Leishmania species with the number of gene copies varying between isolates from the same species [35–37] and changes in gene dosage may be correlated with differences in protein expression [18]. For instance, the region on chromosome 10 containing LdBPK_100480, LdBPK_100510, LdBPK_100520 and LdBPK_100521 encodes a Zn-binding protein whose function is unknown, two tandem copies of gp63 and an uncharacterized protein, respectively (S8 and S9 Tables). This region is amplified (CN = 3) in 3/10 SE strains and one SE-like NE clone, GR364sk/cl.II that clusters by SNP analysis with the SE strains. All other strains are diploid for this region. Gp63 is a protease involved in parasite virulence and survival [38, 39], and is frequently present on chromosome 10 in other species as a multicopy gene family e.g., L. infantum (LinJ.10.0490, 10.0500, 10.0510, 10.0520 and 10.0530) or L. major (LmjF. 10.0460, 10.0465, 10.0470 and 10.0480). Likewise, on chromosome 19 there are two glycerol uptake proteins (LdBPK_191340 and LdBPK_191350) that have an additional gene copy (CN = 3) in 9/10 SE strains and the SE-like clone (GR364sk/cl.II). Interestingly, in other species these genes are part of a tandem multicopy gene family (L. infantum 7 genes, L. major 6 genes, L. braziliensis 8 genes) that may be involved in the remodeling of lipids on glycerol phosphoinositol lipid anchors. Amplification of the 48 kb H-region on chromosome 23 has been associated with drug resistance in vitro [40, 41]. Part of this region is also amplified in some wild-type strains [40, 41]. Deletions (CN = 1) or duplications (CN = 4) of part of the H-region (9 kb) were seen in several SE and NE parasites, respectively. The deleted region was found in 3/10 SE strains and contains the genes coding for the ABC-thiol transporter (MRPA) (LdBPK_230290), involved in resistance to antimony [42], and argininosuccinate synthase (LdBPK_230300), involved in arginine synthesis [43]. Interestingly a similar region is amplified in 8 clones from 2 different NE patient strains, and contains the genes coding for argininosuccinate synthase (LdBPK_230300), a putative uncharacterized protein (LdBPK_230270), the Terbinafine resistance locus protein (Yip1) (LdBPK_230280), and the PTR1 gene (LdBPK_230310). These genes are present in the H-region and frequently amplified in some drug resistant cell lines [44], however they were unchanged, diploid, in all the other parasites belonging to NE and SE populations. This is similar to CN analysis of antimony resistant and sensitive L. donovani strains from Nepal where amplification was not observed for the H-region genes [16], even though MRPA, and in some cases the PTR1 gene, were shown to be amplified in naturally resistant parasites examined by other techniques [45, 46]. Finally, CNV was also found in part of a 15.8 kb region located on chromosome 36 known as the MAPK locus (LdBPK_366740, LdBPK_366750, LdBPK_366760 and LdBPK_366770). Amplification of this region was found in antimony resistant L. donovani from Nepal and associated with higher gene dosage in drug resistant lines [16] [47]. Interestingly, CN of 3/4 genes found in this locus were significantly lower (p ≤ 0.5 x 10−7) in the SE and SE-like NE strains/clones than the NE strains/clones (S8 and S12 Tables). No significant different in CNV between the NE and SE parasites, CN = 2 in 40/41 strains and clones, was observed in the case of the histidine secretory acid phosphatase (LDBPK_366770) which is considered part of the MAPK-locus and amplified in antimony resistant parasites [16]. The only exception was seen with the SE-like NE strain (LDS373bm) that show a complete deletion of LDBPK_366770, as well as LDBPK_366780 (CN = 0).
When haploid gene CN is calculated, taking into account both gene CNV and chromosome ploidy, most of the differences between the NE and SE parasite populations (Table 3 and S13 Table) still remain even though chromosome polyploidy is statistically more common in SE parasites.
Genome wide sequencing (WGS) and analysis of pathogens has proven widely useful for investigations on molecular epidemiology and evolution; genotype—phenotype associations; identification of genes involved in various biological processes such as drug resistance and virulence; as well as new targets for drug and vaccine development [16–18, 21, 48, 49]. Developments in next generation sequencing over the last two decades have provided a relatively low cost, fast pipeline for the exploration and comparison of Leishmania genomes. This study focused on comparison and analysis of WGS data from a large number of L. donovani strains and clones (n = 41) originating from fifteen VL patients in southern and northern Ethiopia. Previous studies on population genetics using multilocus microsatellite typing (MLMT) or individual gene sequences suggest that L. donovani is comprised of distinct populations associated with specific geographic regions in East Africa [7, 15, 50], and that African parasites are in large part distinct from those found on the Indian subcontinent [13, 17]. Differences have also been documented in the parasites, sand fly vectors and host responses between these geographic regions e.g., sensitivity of antigen based serodiagnostic assays [51], clinical response to paromomycin [6], incidence of PKDL [52, 53]. Our whole-genome sequence data confirms the presence of two very different populations of L. donovani in the region, exemplified by an absolute difference of ~15,000 SNPs between the NE and SE populations. These parasite populations likely arose due to unique evolutionary pressures associated with local sand fly species, hosts, reservoirs, ecology, and other factors. A unique advantage of whole-genome data is that it gives us a comprehensive catalog of genetic variation that could underpin these adaptations.
When chromosome aneuploidy is analyzed, a picture appears suggesting great diversity among the Ethiopian strains within each population. This picture is unlike that shown for the Ethiopian reference strain LV9 where all 35 chromosomes, except for chromosome 31, were disomic [18]. This reference strain was extensively passaged in numerous laboratories since first isolated from a VL patient in 1967. Unlike the reference strain, all the isolates used in this study were rapidly cryopreserved and only briefly cultured under identical conditions prior to DNA purification for WGS, yet they still show unique, highly variable karyotypes compared to strains from the same geographic population or strains isolated from different organs of an identical patient. This indicates that chromosome aneuploidy, unlike SNPs, cannot be used to map leishmanial population genetics. Interestingly, MLMT analysis of L. donovani strains from Libo Kemkem, a previously non-endemic region in NE where an outbreak of VL occurred in 2004–2005, also demonstrated high genetic diversity among parasites isolated from patients, including a unique genetic group that shared several alleles with strains from SE [50].
Leishmania chromosome ploidy in individual cells can change rapidly in response to environmental conditions, even routine culture, resulting in mosaic aneuploidy [54], however in this study individual clones generated from patient strains generally showed similar karyotypes clustering together, and when examined, were highly similar to the parental patient isolate. These results suggest that the average karyotype for each strain is relatively stable i.e., minimally affected by culturing and/or cloning prior to DNA extraction. Downing et al [16] also reported that chromosome aneuploidy was stable in culture for 17 Nepalese L. donovani strains even though they also showed diverse karyotypes. This suggests that the karyotypes observed for the parasites in this study are probably very similar to the original patient isolate, assuming changes don’t take place upon differentiation from intracellular amastigote to extracellular promastigote. However, this can only be confirmed by direct measurement of aneuploidy in parasites taken directly from patients without prior culturing, something not currently possible.
Similar to previous reports, chromosome 31 was supernumerary in all 41 Ethiopian strains and clones studied. This appears to be a defining characteristic in all Leishmania species examined so far [16, 18, 19, 25, 55]. As genes involved in iron metabolism and related functions are highly enriched on chromosome 31, it has been suggested the chromosome polyploidy arose to expedite iron uptake, and that expression of Iron–Sulfur proteins that are important in oxidation-reduction reactions, and synthesis of metabolites essential for parasite survival and growth [55].
One interesting finding is that strains concurrently isolated from different organs of identical patients, in the two cases examined, have significantly different karyotypes. Thus, while clones and/or parental strain from spleens of each patient clearly grouped together, they clustered separately from clones originating from the skin of the same patient. While changes in specific chromosome ploidy associated with parasite tropism were not identified, these results suggest that the aneuploidy patterns observed are a result of parasite origin (spleen or skin), and the differing conditions, perhaps temperature or host immune responses, to which the parasite is exposed.
SNP analysis, similar to MLMT [15], clearly shows that the Ethiopian L. donovani strains group, in large part, into two main populations, NE and SE, delineated by geography rather than clinical history (VL or HIV-VL co-infection; spleen or skin). Interestingly, the NE population appears to be more polymorphic than the SE population (Figs 2 and 3), reflecting the finding from MLMT data that inbreeding is higher for SE strains than NE strains [15]. In most cases, clones generated from an individual strain (GR363sp/sk, GR364sp/sk, GR383, GR356, AM560, etc), show more limited genetic polymorphism than that observed between strains, generally clustering together regardless of the method used for analysis (chromosome aneuploidy, SNPs or gene CNV). This was the case even when the patient strain(s) were isolated from different organs, such as skin and spleen, of the same HIV-VL co-infected patient, showing that the genotypes present in visceral organs can spread systemically in immunosuppressed patients to the skin where they get transmitted to sand flies. Only in one case, GR364sk/cl.II, did a cloned line fail to cluster with other clones generated from the patient strain. Instead, this clone grouped with SE strains both by SNP and CNV analysis. While contamination during cloning can’t be ruled out, parasites isolated from HIV co-infected patients have been shown to be more polymorphic than those isolated from patients with VL, and differences following patient treatment have been noted [56–58]. The chromosome karyotype and SNPs for this clone are distinct from all the SE strains suggesting that contamination, if it occurred, did not take place during generation of the cloned line.
SNP analysis also identified three strains, AM422, AM553 and LDS373bm, that didn’t fall, as expected, together with their respective geographic genotypes, SE and NE. LDS373bm and LDS373sp were isolated in parallel from different organs, bone marrow and spleen respectively, of the same HIV co-infected patient. The latter parasite (sp) is genetically similar to other parasite strains belonging to the NE population, while the former (bm) belongs to another genotype. These parasites are also different by k26 PCR typing of the HASP B repeat region [7]. This patient was apparently infected by at least two genotypes, with the genotype present in the bone marrow perhaps less virulent and only surviving in immune suppressed hosts. While several reports using MLMT, multilocus sequence typing and kDNA RFLP show that HIV-VL patients can be sequentially infected with genetically different parasites [58–60], to our knowledge this is the first time that an HIV-VL patient was shown to be simultaneously infected with two genetically different parasites. The amplicon (290 bp) seen for LDS373sp was typical of most NE strains examined (37/41), while for LDS373bm it was larger (410 bp), and observed in 4/41 NE strains, all HIV-VL co-infected patients. WGS of other parasites exhibiting the 410 bp amplicon was not carried out. It would be interesting to analyze more of these parasites and see if they form a separate genetic group. AM422 and AM553, both isolated in SE, fell outside the main NE and SE populations when SNPs were analyzed by two methods (Figs 2 and 3). Neither of these isolates was from patients co-infected with HIV. AM422 showed a k26 PCR amplicon typical of NE strains (290 bp) and originated in the Omo valley near the border with South Sudan, while AM553 had a unique k26 amplicon (360 bp). The unique SNP profiles for these two isolates suggest that additional genotypes are circulating in SE, perhaps a result of the more varied ecology in this region.
Gene CNV analysis also identifies differences that are typical of each geographic parasite population, NE and SE. Candidate genes that can be attributed to essential biological processes like drug resistance, virulence and parasite viability demonstrate differential CN among SE and NE strains and clones. While it is not clear which environmental and host factors resulted in the selective amplification of different genes in the NE and SE populations, many of them are essential genes important for parasite survival. Amplification or deletion of specific genes may give the parasites a growth advantage in the sand fly vector or human host. In the NE parasites, there are three times more copies of folate/biopterin transporter (FBT) genes on chromosome 10. Leishmania are folic acid auxotrophs, and LDBPK_100400 is a homologue to the Leishmania infantum FT1 transporter that was defined as the main folate transporter in this Leishmania species. It is also known that FT1 transporter expression is upregulated in log phase of promastigote stage [34, 44], the stage found in sand fly midgut. Therefore, increased folate concentration in sand fly midgut may result in better parasite growth in the vector and provide the parasite with a better chance for survival and infection.
This study was conducted according to the Helsinki declaration, and was reviewed and approved by the Institutional Review Board (IRB), Medical Faculty, Addis Ababa University. Written informed consent was obtained from each adult study participant.
For this work 18 L. donovani strains isolated from 15 patients with VL in Ethiopia during the years 2009–2010 were selected for WGS (S1 Table). The selection was based on three criteria: 1) geographical origin (northern or southern Ethiopia); 2) Patient's pathology such as HIV/VL co-infection versus VL; and 3) source of parasites, skin versus spleen. Parasites were cultivated in M199 medium with supplements and rapidly frozen [7]. Additional Ethiopian (GR373 [7]) and Sudanese (LEMS3570, kindly provided by Prof. Patrick Bastein, National Reference Center of Leishmania, University Hospital Centre of Montpellier, France) strains were included in analysis of FT1 copy number, and are also listed in S1 Table. All parasites used in this study were characterized by ITS1 -, cpb—and k26—PCR ([7] and S4 Fig).
Eight patient strains were cloned prior to DNA extraction for WGS. The cloning procedure was carried out essentially as described [61].
DNA was purified from Leishmania promastigotes that were harvested in their stationary growth stage in ~20ml M199 medium. DNA extraction was carried out as described by [7]
Genomic DNA was sheared into 400–600-base pair fragments by focused ultrasonication (Covaris Adaptive Focused Acoustics technology (AFA Inc., Woburn, USA)) and standard Illumina libraries were prepared. 100 base pair paired end reads were generated on the HiSeq 2000 v3 according to the manufacturer’s standard sequencing protocol [62]. Raw sequence data was deposited in the European Nucleotide Archive with the accession number ERP016010.
Sequence reads were mapped against the reference genome Leishmania donovani_21Apr2011 [16] using SMALT (version 0.7.4 https://sourceforge.net/projects/smalt/) to produce bam files. SMALT was used to index the reference using a kmer size of 13 and a step size of 2 (-k 13 -s 2) and the reads aligned. Reads were mapped if they had an identity of at least 90% to the reference genome and mapped uniquely to the genome. Reads in pairs were mapped independently, and marked as properly paired if they mapped in the correct orientation no more than 1.5 kb apart. PCR duplicate reads were identified using Picard v1.92(1464) and flagged as duplicates in the bam files.
Aneuploidy was predicted based on whole chromosome median read coverage. For the normalization of median read coverage over all 36 chromosomes for a given strain, the average median coverage of four stable diploid chromosomes (chromosome 30, 32, 34 and 36) was calculated and taken as the mean read coverage for a diploid chromosome (DCmean). These four chromosomes have been previously shown to be diploid in almost all L. donovani isolates examined, and have been used for prediction of chromosome copy number [18, 20]. The predicted chromosome copy number is calculated as the fold change compared to DCmean. This prediction was applied to all 41 L. donovani sequences over 36 chromosomes, and saved in a 41*36 chromosome copy number matrix. We used R (version 3.1.3) [63] to evaluate of similarities and differences between strains and clones, and their chromosome aneuploidy patterns and computed a heat map over the chromosome copy number matrix with R heatmap.2() function from the gplots (version 2.17.0) R package. The numerical data for the creation of the heat map (Fig 1) is given in the supporting information (S2 Table). Median chromosome ploidy was used to compare average ploidy between NE and SE strains in Table 1. For isolates with multiple clones, average chromosome ploidy was calculated by determining the median somy of all the clones from an individual strain separately for each of the 36 chromosomes (S2 Table).
For the identification of population typical SNPs a procedure was implemented based on SNP and indel calling with the Genome Analysis Toolkit (GATK version 3.1) [64], VCFtools (version 0.1.12b) [65] and a custom R script based on Bioconductor packages (www.bioconductor.org) (Release (3.2)). The parameters for the first filtering procedure with GATK were set as follows: Emission confidence threshold > 10, Calling confidence threshold >50, read-depth > 500. A calculation of level of similarity between all 41 samples based on SNP profiles was computed with VCFtools, and processed with R based procedures into a similarity matrix. Finally, this matrix was visualized with heatmap.2() function from the gplots R package. Principal component analysis (PCA) following pruning of SNPs with high linkage disequilibrium was run using the Bioconductor package SNPRelate [26]. For the creation of unique SE or NE SNP profiles, only SNPs that were identified in >4 or >7 of the SE or NE clones and strains, respectively, were taken for final analysis. This protocol created a consensus SNP profile for each parasite population, as repetition of SNPs in more than 1/3 of the strains and clones for each population supported the accuracy of SNP calling and served as an additional quality control step. As a final step, unique SNP profiles were detected with VCFtool for each population. Unique SNPs present in coding regions that affect protein translation, namely non-synonymous or nonsense mutation(s) that change the amino acid or cause a stop codon were identified by a self-implemented R procedure based on Bioconductor packages. The mean absolute number of SNPs was compared between the SE and NE parasite populations, as well as to the Indian reference strain. Mean absolute SNPs for an individual patient isolate was calculated by averaging the SNPs from all clones of the respective strain. Variation in the number of SNPs between clones of an individual isolate was small, and % standard error varied from = 0.34–6.0 x 10−3, except for GR364sk (% s.e. = 0.064) when the atypical clone GR364sk/cl.II was included. In the absence of the atypical clone the % s.e. = 0.001.
The detection of copy number variations (CNV) and aberrations was done using the R- bioconductor [66] (www.bioconductor.org) package cn.mops (version 1.16.1) (Copy Number estimation by a Mixture Of PoissonS). cn.mops detected copy number variation as the normalized read depth variation at a certain genomic position over all 41 strains and clones. For that cn.mops calculated the read count matrices across all BAM files. For this analysis, the genomic window length was set to 5 Kbp and used as a sliding window for the prediction of genomic CNV. Therefore, a genomic position of certain strain or clone is considered as one with "CNV" if it shows a significant change in normalized read depth (> or < two i.e., diploidy) in a window length of 5 Kbp compared to other strains or clones at the same given genomic position. Further analysis genomic CNV that is localized in coding regions was carried out.
Parasites 2 x 106 / ml were added in triplicate to 96-well plates and incubated with 0.5 mg/ml MTX (Sigma catalog number M8407) and in non-drug treated medium in final volume of 200 μl for 66 hours at 26°C. AlamarBlue (25 μl/well; AbD Serotec) was added and the viability was measured after four hours (λex = 544; λem = 590; Fluroscan Ascent FL, Thermo) [67]. The percentage of killing was calculated as the fraction of fluorescence level of MTX treated wells compared to the non-drug treated wells.
The polycistronic folate/biopterin transporters (FT) on chromosome 10 have high DNA sequence similarity [29]. FT1 (LDBPK_100400) dual priming oligonucleotides (FT1-DPO) were used to specifically evaluate genomic CN for this gene by qPCR. FT1-DPO primers, forward FT1-DPO 5'-CGCCAGAACCCGAAGCCTGIIIIIGCACTGG-3' and reverse FT1-DPO 5'-GTTCATCACAGTCGCGATGAGTIIIIIAATCATTATG-3', were designed to include a polydeoxyinosine linker (IIIII) [30] that allows the specific primer annealing to the LDBPK_100400 (FT1 ortholog) gene, and not to the other LD FT homologues. Specificity of the FT1-DPO PCR product was confirmed by cloning and sequencing of the amplicon. The L. donovani housekeeping gene alpha-tubulin was used for normalization. The qPCR conditions were the same for both FT1-DPO and housekeeping genes, and was carried out as follows: DNA (~10–20 ng) or no DNA control was added to HRM PCR Kit reaction mix (10 μl, QIAGEN GmbH, Germany) containing the FT1—DPO primers (1 μM each final concentration), and ultra-pure PCR-grade water (final volume 25 μl/PCR). Amplification conditions were as follows: 3 min denaturation at 95°C, followed by 40 cycles of denaturation 1 s at 95°C; annealing 20 s at 55°C; and extension 1 s at 65°C. HRM Ramping was carried out at 0.2°C/s from 65 to 95°C. HRM PCR and analysis were performed using a Rotor-Gene 6000 real-time PCR analyzer (Corbett Life Science, Australia). All reactions were carried out in duplicate and a negative-control reaction without parasite DNA was included in each experiment. For the calculation of the FT1 relative copy number (CNrel) the threshold (Ct) for all FT1 amplified samples were compared with their corresponding alpha-tubulin amplified samples as followed: CNrel = 2Ct(alphatubulin)-Ct(FT1). The CNrel was further normalized based on the mean of the six lowest predicted relative CN. The mean value was considered as GCN = 1.
|
10.1371/journal.pntd.0007448 | Seasonal and geographic variation in insecticide resistance in Aedes aegypti in southern Ecuador | Insecticide resistance (IR) can undermine efforts to control vectors of public health importance. Aedes aegypti is the main vector of resurging diseases in the Americas such as yellow fever and dengue, and recently emerging chikungunya and Zika fever, which have caused unprecedented epidemics in the region. Vector control remains the primary intervention to prevent outbreaks of Aedes-transmitted diseases. In many high-risk regions, like southern Ecuador, we have limited information on IR. In this study, Ae. aegypti IR was measured across four cities in southern Ecuador using phenotypic assays and genetic screening for alleles associated with pyrethroid IR. Bottle bioassays showed significant inter-seasonal variation in resistance to deltamethrin, a pyrethroid commonly used by the Ministry of Health, and alpha-cypermethrin, as well as between-city differences in deltamethrin resistance. There was also a significant difference in phenotypic response to the organophosphate, Malathion, between two cities during the second sampling season. Frequencies of the resistant V1016I genotype ranged from 0.13 to 0.68. Frequencies of the resistant F1534C genotype ranged from 0.63 to 1.0, with sampled populations in Machala and Huaquillas at fixation for the resistant genotype in all sampled seasons. In Machala and Portovelo, there were statistically significant inter-seasonal variation in genotype frequencies for V1016I. Resistance levels were highest in Machala, a city with hyperendemic dengue transmission and historically intense insecticide use. Despite evidence that resistance alleles conferred phenotypic resistance to pyrethroids, there was not a precise correspondence between these indicators. For the F1534C gene, 17.6% of homozygous mutant mosquitoes and 70.8% of heterozygotes were susceptible, while for the V1016I gene, 45.6% homozygous mutants and 55.6% of heterozygotes were susceptible. This study shows spatiotemporal variability in IR in Ae. aegypti populations in southern coastal Ecuador, and provides an initial examination of IR in this region, helping to guide vector control efforts for Ae. aegypti.
| Mosquito control is the primary method of managing the spread of many diseases transmitted by the yellow fever mosquito (Aedes aegypti). Throughout much of Latin America the transmission of diseases like dengue fever and Zika fever pose a serious risk to public health. The rise of insecticide resistance (IR) is a major threat to established vector control programs, which may fail if commonly used insecticides are rendered ineffective. Public health authorities in southern coastal Ecuador, a high-risk region for diseases vectored by Ae. aegypti, previously had limited information on the status of IR in local populations of mosquitoes. Here, we present the first assessment of IR in adult Ae. aegypti to insecticides (deltamethrin, Malathion, and alpha-cypermethrin) routinely used in public health vector control in four cities along Ecuador’s southern coast. Observed patterns of IR differed between cities and seasons of mosquito sampling, suggesting that IR status may fluctuate in space and time. The highest overall resistance was detected in Machala, a city with hyperendemic dengue transmission and a long history of intense insecticide use. Monitoring for IR is an integral component of vector control services, where alternative management strategies are deployed when IR is detected.
| In Ecuador, dengue, chikungunya, and Zika viruses are present and transmitted to people by the Aedes aegypti mosquito, causing a high burden of febrile illness in susceptible populations. This species is a particularly effective vector of these aboviruses because it has evolved to live in urban environments, lay its eggs in small containers of water in and around human dwellings, and feed preferentially on humans [1]. In Ecuador in 2016, there were 13,612 reported cases of dengue fever [2], 2,025 cases of chikungunya, [3] 3,531 cases of Zika fever [4]. For all three of these diseases, actual incidence in Ecuador is likely much higher than the reported figures indicate, because many cases are asymptomatic or mild, and access to laboratory diagnostics can be limited [5]. Furthermore, the co-circulation of these viruses can lead to more complex disease outcomes, and nonspecific febrile symptoms make clinical differentiation of the three diseases difficult [6]. To reduce the burden of disease, public health organizations rely heavily on vector control methods, particularly insecticide applications [7]. Additionally, individuals purchase products such as aerosols, repellents, and nets to reduce mosquito populations and prevent disease transmission at the household level, with low-income households in Machala, Ecuador, spending as much as 10% of their discretionary income on these items [5]. While vector control is considered to be the only tool available for the control of these arthropod-borne viruses, the extent to which these interventions produce significant reductions in disease burden has been difficult to ascertain [8,9].
There are four main categories of chemical insecticides regularly used for disease vector control: organophosphates, pyrethroids, carbamates, and organochlorines, with organophosphates and pyrethroids being the most widely used in Ecuador for Ae. aegypti control [10]. For all four of these insecticide classes, regular deployment in locations around the world has been associated with the development of insecticide resistance (IR) in targeted vector populations, resulting in resurgences of mosquito-borne diseases when vector control fails [11–13]. Although public health organizations recommend monitoring and managing IR [14], these practices are often resource-intensive, meaning that many areas do not have the capacity to conduct regular IR testing. Additionally, the largely unregulated application of insecticide treatments at the household level can influence local-scale IR. In an experimental study, exposure to commercially available aerosolized insecticides, applied in a simulation of typical household use, resulted in significant increases in genotypic and phenotypic IR in Ae. aegypti [15]. These results, along with field observations [16–18], indicate that fine-scale selection pressures can contribute to IR development. When resistance remains undetected, public health organizations may spend substantial time and resources applying insecticides that are ineffective [19] and may have negative environmental and human health impacts [20].
A better understanding of the spatial and temporal variability in IR would greatly enhance the ability of vector control to predict and mitigate IR in vector populations. Research has shown that IR status can differ dramatically across cities within a country [21–23], but further work needs to be conducted to understand the spatial scale of this variability, particularly across cities with differing vector control needs and strategies. Furthermore, most available reports from within-country spatial scales determine IR status at a single point in time, but IR exhibits temporal fluctuations that should be taken into consideration [24,25]. This study addresses a substantial knowledge gap regarding IR in Ae. aegypti by investigating differences in IR across both space (with four cities included in the study area) and time (with sampling occurring in three seasons), while considering both genetic and phenotypic lines of evidence to determine IR status. Furthermore, assays were conducted with multiple commonly-used insecticides, allowing for comparisons of effectiveness that can inform public health decision makers about local-scale vector-control.
In Ecuador, vector control is conducted by field workers of the Ministry of Health (MoH) in arbovirus endemic areas, as well as focal control in and around homes with suspected arboviral infections. Interventions to control adult mosquitoes include indoor residual spraying (IRS) with deltamethrin (pyrethroid) and ultra-low volume (ULV) fumigation with malathion (organophosphate). Interventions to control immature mosquitoes include application of an organophosphate larvicide (temephos/Abate) to containers with standing water, as well as community mobilization to eliminate larval habitats.
The objectives of this study were to evaluate overall IR status and to describe seasonal and inter-city variability in IR of Ecuadorian populations of Ae. aegypti, specifically testing for susceptibility to insecticides commonly used in mosquito control campaigns by the MoH. In recent work in El Oro province, researchers demonstrated resistance to deltamethrin in phenotypic assays and detected genetic mutations associated with resistance to pyrethroids in Ae. aegypti [26]. Beyond this small-scale study, however, IR monitoring is not regularly performed and research in this area has been limited, meaning that the prevalence of IR in these Ae. aegypti populations is largely unknown. Due to the high morbidity associated with the viruses transmitted by Ae. aegypti, as well as concerns regarding the cost and efficacy of vector control efforts, MoH leaders have indicated that there is a critical need for operational research on IR in Ecuador.
This study was conducted as part of a longitudinal cohort study examining social-ecological correlates of arboviral risk in southern Ecuador. The study protocol was reviewed and approval by Institutional Review Boards (IRBs) at SUNY Upstate Medical University (IRBNET ID 4177710–25), the Luis Vernaza Hospital in Guayaquil, Ecuador, and the Ecuadorean Ministry of Health. Prior to the start of the study, all adult participants (18 years of age or older) engaged in a written informed consent process. Data collection for this project was also conducted with input from local vector control and public health organizations. Sampling coordinates were stored in a secured database and all resulting data were pooled to evaluate citywide trends in the analysis, meaning identifiable individual sites were not shared.
Field samples of Ae. aegypti were collected in four cities in El Oro, a dengue endemic province in southern coastal Ecuador (Fig 1). The cities included in this study range from low to high dengue case burden (e.g. in 2017 incidence ranged from 0.78–16.8 per 10,000 people). In Machala (3°15’09”S, 79°57’20”W; 6m elevation, 279,887 people), a port city and important center of agribusiness, 136 unique sites were visited. In Huaquillas (3°28’33”S, 80°13’33”W; 15m elevation, 57,366 people), a town on the border with Peru, 142 unique sites were visited. Both of these cities are located at sea level along the coast and have endemic transmission of dengue, typically seeing higher numbers of annual reported cases (Machala 148, Huaquillas 33, reported in 2017). In Portovelo (3°42’58”S, 79°37’08”W; 645 m elevation, 13,673 people), 52 unique sites were visited and in Zaruma (3°41’31”S, 79°36’47”W; 1,155m elevation, 25,615 people), 42 unique sites were visited. These are both mining towns located further inland and at higher elevations, with Zaruma highest, with limited autochthonous transmission (Portovelo had 20 cases on average per year from 2014–2017, and Zaruma 4). The discrepancies in case burden between the cities translate into differential mosquito control demands at the municipal level, making our selection of study sites a diverse backdrop for investigating IR.
Collection of field samples was conducted in 2017 over three sampling periods: Season 1 (February 1 –April 30), Season 2 (May 1 –June 30), and Season 3 (July 1 –August 31). Seasons were explicitly chosen to collect mosquito eggs at different phases of annual arbovirus transmission in Ecuador, sampling during the peak (Season 1), decline (Season 2), and low transmission (Season 3). These collection seasons correspond with historical trends in both dengue transmission and mosquito densities in El Oro Province [27,28], and matched the observed dengue transmission pattern during the study (see S1 Fig, weekly dengue cases for each city in 2017).
A longitudinal cohort study across the four cities included household surveys with heads of households, regarding the purchase of insecticides. Surveys were conducted from May to July 2017. We additionally collected information from the MoH in each city regarding the type, timing, and method of insecticide application over the duration of the present study.
Ae. aegypti eggs were collected from households in the four cities using ovitraps lined with oviposition paper [29]. Two or three ovitraps were placed at each trapping site; details on the number of traps by season are shown in S2 Fig. Households were selected from ongoing surveillance study sites where the MoH designated areas of the cities as having high historic risk of dengue. These houses are distributed across each of the four cities to capture geographic variability, and are part of a larger study looking at arboviral and vector dynamics in a cohort of homes over 3 years. Ovitrapped houses for this study were purposely targeted to collect the greatest number of eggs, so ovitraps were strategically placed in areas known to have a greater abundance of eggs based on previous surveillance efforts, representing a subset of the larger household cohort study as well as additional sites nearby. Ovitraps are a sensitive means of identifying the presence of Ae. aegypti, especially in areas with low vector densities [30].
Papers with eggs were collected in the field and transported to the Center for Research on Health in Latin America (CISeAL) in Quito, Ecuador, where all insect rearing and handling was performed under standard insectary conditions (28 ± 1°C; 80% +/- 10% relative humidity; 12h light, 12h dark photoperiod). Egg hatching was achieved by placing the papers in a plastic tray containing distilled water. Larvae were fed finely ground fish flakes. Upon pupation, mosquito specimens were placed in cages for adult emergence. To increase the number of adult specimens available for experimentation, F0 adults were allowed to mate and adult female mosquitoes were blood fed on mice and allowed to oviposit. Collected eggs (F1 generation) were hatched and maintained under the aforementioned conditions. Upon pupation, F1 specimens were sorted by sex and females were used for further experimentation. Males were killed by freezing and discarded.
We monitored phenotypic resistance to pesticides using the bottle bioassay method as described by Brogdon and Chan [31]. Briefly, we coated glass bottles with ethanolic-insecticide solutions at established diagnostic doses: 10 μg / bottle for both deltamethrin and alpha-cypermethrin (both pyrethroids), and 50 μg / bottle malathion (an organophosphate). Each bioassay replicate consisted of four pesticide-coated bottles and one control bottle, coated only with the ethanol diluent. After the ethanol had evaporated from the bottles, groups of 15–25 female mosquitoes were introduced into each of the four treatment bottles, resulting in a total combined count of mosquitoes in the treatment bottles for each replicate ranging from 62 to 100. Mortality was recorded at 15-minute intervals. Mosquitoes were considered knocked down when they were incapable of flying or maintaining an upright posture on the surface of the bottle. Mortality counts were recorded after 30 minutes of exposure, following the protocol for diagnostic time for Ae. aegypti given in [31]. This period was chosen to be consistent with CDC protocols used at other sites, and following the specific diagnostic time recommended for South American Ae. aegypti. We did not have a reference susceptible colony to establish specific diagnostic time; we did not record final mortality times.
The number of F1 females procured from field-collected specimens allowed us to perform one to three bioassay replicates for each city. The selection of pesticides (deltamethrin, alpha-cypermethrin and malathion) corresponds to pesticides typically used for vector control in Ecuador.
In addition to the monitoring of phenotypic resistance, we also performed genetic screening for two mutations of the voltage-gated sodium channel gene (V1016I and F1534C) that are associated with IR to pyrethroids in Ae. aegypti. Both mutations have been previously described and are reported to alter the transmembrane domain of the voltage-gated sodium channel, preventing pyrethroids from binding to these sites [28,29]. Previous genotyping work has shown that Ae. aegypti that are homozygous for the isoleucine variant at position 1016 (I1016) are resistant to pyrethroid treatments, while those that are homozygous for the valine variant (V1016) allele are susceptible, and those that are heterozygous show intermediate resistance [32]. At position 1534, the cysteine variant (C1534) confers resistance, and the phenylalanine variant (F1534) is susceptible, while the heterozygote typically shows intermediate resistance, though there is some disagreement regarding the extent to which the heterozygous genotype confers resistance, independent of additional resistance mechanisms [32,33].
Following the aforementioned bioassays, DNA was extracted from specimens of known phenotype (susceptible/resistant to each of the pesticides used for testing). Genomic DNA was obtained using the Wizard Genomic DNA Purification Kit (Promega Corporation, Madison, WI, USA), following the protocol established by the manufacturer. Quantification of the concentration (ng/μL) and purity (absorbance index at 260nm / 280 nm) of the obtained DNA was completed using a Nano-Drop 1000 V3.7 spectrophotometer (Thermo Fisher Scientific Inc., Wilmington, DE, USA).
Screening for the V1016I and F1534C mutations was performed as previously described [28,29], using a Bio-Rad real-time thermocycler CFX96 (Bio-Rad, Hercules, CA, USA). Results were visualized using CFX Manager Software (Version 3.1, Bio-Rad, Hercules, CA, USA). Determination of the genotype was possible through the visualization of PCR product melting curves. In the case of the V1016I mutation, a melting peak at 79°C corresponded to isoleucine (I/I–resistant mutant) and a melting peak at 85°C corresponded to a valine (V/V–susceptible wild type) (S3 Fig). In the case of the F1534C mutation, a melting peak at 85°C corresponded to cysteine (C/C, resistant mutant) and a melting peak at 80°C corresponded to a phenylalanine (F/F–susceptible wild type) (S4 Fig). Genotypic and allelic frequencies were calculated for both mutations.
Genotypic frequencies for all four cities were mapped using ArcMap© version 10.6 (Environmental Systems Research Institute, Redlands, California, USA). Inter-city and inter-seasonal differences in genotypic frequencies were tested for statistical significance (α = 0.05) using Fisher’s exact test of independence in the R software for statistical computing, version 3.4.3 (The Foundation for Statistical Computing, Vienna, Austria). When significant differences were detected, post-hoc pairwise Fisher’s tests with Bonferroni corrections were then performed to identify which cities and seasons differed significantly. The same statistical methods were used to examine phenotypic frequencies of resistance, as determined by bottle bioassays for deltamethrin, alpha-cypermethrin, and malathion, between cities for each collection season and in relation to corresponding genotypes. Additionally, populations for each city and season were assessed for Hardy-Weinberg equilibrium using the exact test in the HardyWeinberg R package [34].
In Machala, 27.5% (14/51) surveyed homes reported purchasing pyrethroid insecticides to use at home, while 18.9% (10/53) of surveyed homes in Huaquillas reported these purchases. Of households surveyed in Portovelo and Zaruma, 45.6% (22/48) and 36.5% (19/52) respectively, reported purchase of pyrethroid insecticides for home use. Ministry of Health insecticide applications in Machala utilized a combination of deltamethrin, alpha-cypermethrin, and malathion over the duration of the study. Both Portovelo and Zaruma applied deltamethrin and malathion during the peak transmission season, followed by deltamethrin only for the following months. Huaquillas included malathion and an additional (unidentified) product during peak transmission season, followed by deltamethrin and an additional (unidentified) product in the following months (S5 Fig).
High levels of IR to deltamethrin, alpha-cypermethrin, and malathion were detected in all four cities, as determined by bottle bioassays. Machala had the lowest combined mortality averaged across the three seasons (18.29%, SE±5.75), indicating the highest level of resistance, followed by Portovelo (31.08%, SE±6.23) and Huaquillas (44.79%, SE±10.87). The mean mortality rate for Zaruma was 23.67% (SE±17.93) in Season 1, which was the only time period in which sufficient numbers of Ae. aegypti were collected for bioassays at this location.
The proportion of mosquitoes that were resistant to deltamethrin differed significantly across cities in all three collection seasons (Fisher’s exact test P-value <0.001) (Fig 2A). This pattern was also seen in the post-hoc pairwise tests (S1 Table). Mean mosquito mortality after 30 minutes of exposure to deltamethrin in bottle bioassays ranged from 0.70% (SE±0.6) in Machala to 18.99% (SE±0) in Huaquillas in Season 1, from 7.09% (SE±4.19) in Portovelo to 16.88% (SE±7.51) in Huaquillas in Season 2, and from 0.37% (SE±0.37) in Machala to 56% (SE±2.49) in Huaquillas in Season 3, with no mortality observed in control groups (Fig 2A).
Mean mosquito mortality after 30 minutes of exposure to alpha-cypermethrin across bottle bioassay replicates ranged from 59.24% (SE±13.86) in Zaruma to 59.41 (SE±1.49) in Machala in Season 1, from 12.1% (SE±0.42) in Machala to 17.27 (SE±0.33) in Portovelo in Season 2, and from 0.38% (SE±0.38) in Machala to 11.39 (SE±1.03) in Portovelo in Season 3, with no mortality observed in control groups (Fig 2B). Significant differences in mean mortality rates between cities were detected in seasons two and three (Fisher’s exact test P value <0.001; Fig 2B), with Machala having the highest proportion of phenotypically resistant specimens in these two seasons (S2 Table). Collection counts were insufficient for alpha-cypermethrin bioassays in Portovelo for Season 1, Huaquillas for Seasons 1 and 3, and Zaruma for Seasons 2 and 3.
Mean mosquito mortality after 30 minutes of exposure to malathion in the bottle bioassays ranged from 5.07% (SE±0.35) in Machala to 3.03% (SE±1.52) in Zaruma in Season 1, from no mortality in Machala to 1.45% (SE±1.45) in Portovelo in Season 2, and from 0.78% (SE±0.78) in Machala to 2.86% (SE±0.08) in Portovelo in Season 3, with no mortality observed in control groups. Statistically significant differences in Malathion resistance were only detected in season two between Machala and Portovelo (Fisher’s exact test P value <0.001). Due to low sample sizes, bioassays could not be conducted in Portovelo for Season 1, Zaruma for Seasons 2 and 3, and Huaquillas for all three seasons, and therefore further comparisons were not feasible.
Inter-seasonal variations in phenotypic resistance for each city were also assessed for statistical significance. In the deltamethrin bottle bioassays, mean mortality in Machala increased from Season 1 to Season 2 (Post-hoc Fisher’s exact test P value <0.001), but decreased from Season 2 to Season 3 (Post-hoc Fisher’s exact test P value <0.001). Mean mortality increased from Season 1 to Season 2 (Post-hoc Fisher’s exact test P value = 0.003) and Season 2 to Season 3 (Post-hoc Fisher’s exact test P value <0.001) in Huaquillas. Mean mortality increased in Portovelo from Season 1 to Season 2 (Fisher’s exact test P value = 0.005) and from Season 2 to Season 3 (Fisher’s exact test P value = 0.02; S3 Table). In the alpha-cypermethrin bottle bioassays, significant inter-seasonal differences were only detected in Machala, where the percent mortality decreased from Season 1 to Season 2 (Post-hoc Fisher’s exact test P value < 0.001) and from Season 2 to Season 3 (Post-hoc Fisher’s exact test P value = 0.05; S4 Table). Similarly, significant inter-seasonal differences in mean mortality in the malathion treatment were only detected in Machala, where mortality decreased from Season 1 to Season 2 (Post-hoc Fisher’s exact test P value = 0.05; S5 Table).
Observed genotype frequencies varied significantly between cities only in Season 1 for both V1016I (Fisher’s exact test P value <0.001) and F1534C alleles (Fisher’s exact test P value <0.001; Fig 3). Pairwise post-hoc analysis revealed significant differences in the frequencies of genotypes I/I (mutant) and V/I (heterozygous) for the V1061I gene, with Huaquillas (n = 34) having a significantly higher frequency of heterozygotes compared to Machala (n = 22; Post-hoc Fisher’s exact test P value <0.001) and Portovelo (n = 22; Post-hoc Fisher’s exact test P value = 0.03; S6 Table). Portovelo and Zaruma (n = 40) differed in the frequency of I/I (mutant) and V/V (wild type) genotypes, with Zaruma having a significantly higher frequency of wild type mosquitoes during Season 1 (Post-hoc Fisher’s exact test P value = 0.05; S6 Table). Genotypic frequencies of C/C (mutant) and F/C (heterozygous) genotypes of the F1534C resistance gene differed across cities in Season 1, with Zaruma having a significantly higher proportion of the heterozygous genotype than Huaquillas and Machala (Post-hoc Fisher’s exact test P values = 0.02 and <0.001, respectively; S7 Table). Although the frequencies of F1534C genotypes also varied significantly in the Season 3 (Fisher’s exact test P value = 0.03), conservative post-hoc analysis did not reveal any significant pairwise relationships (S7 Table).
Significant inter-seasonal variation in V1016I genotype frequencies was detected in Huaquillas (Fisher’s exact test P value = 0.04), Machala (P value <0.001), and Portovelo (P value <0.001), though conservative post-hoc analysis did not identify significant pairwise differences in Huaquillas. In Machala, the proportion of the I/I (mutant) and V/I (heterozgote) genotypes differed significantly from Season 1 to Season 3 (Post-hoc Fisher’s exact test P value <0.001), with the I/V genotype increasing while the I/I genotype decreased (S8 Table). In Portovelo, there were significant differences in the frequencies of the I/I and V/I genotype between Seasons 1 and Season 2 and between Season 1 and Season 3 (Post-hoc Fisher’s exact test P values = 0.05 and 0.001, respectively), with the frequency of the I/I genotype decreasing while V/I increased. Similarly, the frequency of the V/V (wild type) genotype increased significantly from Season 1 to Season 3 (Post-hoc Fisher’s exact test P value = 0.005) while the I/I genotype decreased (S9 Table). No significant inter-seasonal differences in F1534C genotype frequency were detected.
There were statistically significant associations between the resistance genotypes and phenotypic resistance results for 632 Ae. aegypti that were genotyped and subjected to the pyrethroid (deltamethrin or alpha-cypermethrin) bottle bioassays. Due to low sample sizes, we pooled pyrethroid assay results (Tables 1 & 2), and present the separated analyses in supplemental information (S10 and S11 Tables). For mutation V1016I, while the majority of the resistant individuals (96.4%) presented the mutant or heterozygous genotype, 12 (3.6%) phenotypically resistant individuals presented the wild type (V/V) genotype that is typically associated with susceptibility to pyrethroids, and 40 (13.5%) susceptible individuals presented the genotype that typically confers resistance. Similarly, at the F1534C locus, 276 (93.2%) individuals with the mutant (C/C) genotype, typically associated with resistance, were susceptible to pyrethroid treatments. Due to low sample sizes of the wild type (F/F) for both pyrethroids, separate Fisher’s exact tests were not conducted (S6 Table). Because the two loci studied here contribute additively to resistance, the proportions of resistant and susceptible individuals for each combined V1016I and F1534C genotypes were also considered. Of the 227 mosquitoes with the mutant I/I and the mutant C/C genotype, 187 (82.4%) were resistant in the pyrethroid assay. There were twelve genotyped mosquitoes that were heterozygous at both loci, and of these, six individuals were resistant; likewise, only three genotyped specimen were homozygous wild type at both loci, and all three were susceptible (Tables 1 & 2). The V1016I genotype frequencies indicate that populations from Huaquillas (Fisher’s exact test P value = 0.03) and Machala during Season 1 (Fisher’s exact test P value = 0.03) and Portovelo during Season 3 (Fisher’s exact test P value = 0.02) were not in Hardy-Weinberg equilibrium (S12 Table).
In this study, exposure to diagnostic doses of deltamethrin, alpha-cypermethrin and malathion resulted in mortality rates below 80% after 30 minutes of insecticide exposure in all populations tested, regardless of collection season. Based on the World Health Organization´s (WHO) recommendations for assessing the significance of detected resistance [31], our results suggest that these Ae. aegypti populations should be considered as resistant to all the insecticides considered in our study. Furthermore, the results of the bioassays for malathion susceptibility are of particular interest, as they indicate that these populations are extremely resistant to malathion, with no population showing more than 5.07% mortality, and populations from Machala reaching values as low as 0% mortality during Season 2.
The CDC bottle bioassay used in this study is one of several methods used to detect resistance in mosquito populations. The other most commonly used method, the WHO susceptibility test, is a response-to-exposure analysis that uses insecticide-impregnated papers obtained directly from a distribution center [34]. While this pre-manufactured quality means there is greater consistency and control in the administration of the WHO susceptibility test, the CDC bottle bioassay can be conducted without specialized equipment, which often makes it the assay of choice in resource-limited settings [35]. In addition to these methods, IR can be assessed based on calculation of the lethal concentration needed to kill half of a sample of mosquitoes (LC50) after topical application of the insecticide of interest [36]. This method allows for the calculation of resistance ratios to quantify and compare resistance across populations; however, this test is more time-consuming and requires specialized equipment, so it is not as commonly conducted [37].
In this study, molecular characterization showed that the resistance-associated mutant alleles V1016I and F1534C are present at all the locations studied. In particular, allele F1534C was present at very high frequencies (close to 1) in all the locations studied and across all three seasons, similar to results from recent work in Mexico [35]. This suggests that this gene has been subjected to selective pressures in the past and is approaching or has reached fixation in these populations. By contrast, the results from the test for Hardy-Weinberg equilibrium for the V1016I gene and the significant inter-seasonal differences in genotype frequencies indicate that the populations are still responding to varying selective pressures and are not in a state of equilibrium. Both of these mutations typically confer resistance to dichlorodiphenyltrichloroethane (DDT) as well as pyrethroids [36,37], meaning selection for these alleles likely began with earlier widespread usage of DDT (which was used by the MoH until 1996 [38]) and has persisted as pyrethroids became more commonly used [39]. Significant distinctions in genotypic and phenotypic frequencies were not detected for many cities in this study during subsequent seasons.
It is worth noting that for some locations and seasons, periods of low vector densities in the field resulted in the collection of low numbers of mosquito eggs. Overall, collection counts were consistently low in Zaruma, a small city that has the highest elevation of the four study cities. This observation is consistent with other work that has documented a negative relationship between elevation and the probability of Ae. aegypti presence [40]. Similarly, collection counts during Season 2 and Season 3 were not high enough for bioassays with each of the three insecticides of interest. These seasons have historically corresponded with periods of low mosquito activity and dengue transmission in Ecuador, as observed in 2017 [27,28] This scarcity of F0 individuals translated into missing data for some cities and/or very low frequency counts in some categories, thus providing a limited basis for making quantitative comparisons.
The detection of resistance to organophosphate and pyrethroid insecticides in Ecuador is consistent with broader regional trends identified in recent years. In neighboring Peru, Ae. aegypti strains have been shown to be resistant to multiple organophosphates and pyrethroids in WHO susceptibility tests [41]. Similarly, pyrethroid resistance has been detected in Colombia in both CDC and WHO bioassays, though the levels of resistance vary throughout the country [22]. However, Colombian Ae. aegypti populations tested with both WHO and CDC bioassays were broadly susceptible to malathion, in contrast to the widespread resistance seen in the Ecuadorian Ae. aegypti in this study [42]. Throughout the rest of the Americas, pyrethroid resistance, as measured in studies using LC50 or percent mortality, appears to be broadly distributed, particularly in Brazil and French Guiana, though some populations in Costa Rica, Panama, and northern Colombia are still susceptible [10].
The temporal and spatial variability in the results from the bioassays highlight the importance of regularly conducting IR monitoring across multiple locations to understand the true extent of IR and make appropriate vector control decisions. For example, in Machala, the mortality rate of Ae. aegypti treated with alpha-cypermethrin decreased significantly from Season 1 to Season 3, indicating this population was becoming less susceptible throughout the course of the study. By contrast, the mortality rate associated with the deltamethrin assays on populations from Huaquillas increased from Season 1 to Season 3, meaning this population was becoming more susceptible over time. The impact of city-level insecticide application on IR is difficult to infer with the information we obtained from the MoH of each municipality. Portovelo did not use alpha-cypermethrin and we saw no differences in alpha-cypermethrin mortality rates across seasons. Machala used some form of deltamethrin, alpha-cypermethrin, and malathion throughout the study duration, but mortality rates differed for these insecticides by season, with mortality rates decreasing for both alpha-cypermethrin and malathion. There was some seasonal variation in method of application, strength of insecticide solution, and neighborhood coverage, which could impact the IR of our sampled mosquito populations. Continued work in this area could determine if our observed trends are due to seasonal fluctuations, differential insecticide application parameters, local-scale movements of Ae. aegypti populations with varying levels of resistance, or long-term, inter-annual trends. There were also statistically significant differences in mortality rates across the cities for the deltamethrin bioassays in all three seasons and in two of the three seasons for the alpha-cypermethrin bioassays. Considering this variability in the resistance phenotypes found within a single province of Ecuador, organizations involved in decision-making about insecticide applications should be cautioned against inferring the IR status of one Ae. aegypti population based on the status of populations in neighboring municipalities.
To better contextualize this work for appropriate vector-control decision-making, the relationships between genotypes, IR bioassay results, and actual IR status in the field should be considered. While the V1016I and F1534C mutations are known markers of pyrethroid resistance in Ae. aegypti, the results of this study showed that the genotypes were not perfectly predictive of resistance phenotypes, even when both the V1016I and F1534C genotypes were considered. In the recorded frequencies of genotypes versus bioassay outcomes, resistant and susceptible phenotypes were observed for each genotype, although the resistant phenotype was still statistically associated with the mutant genotypes for both genes. This is likely due to other IR mechanisms, such as metabolic detoxification processes [19] that could influence IR status in these populations; however, these mechanisms were not considered in the current study. Additionally, factors such as temperature, larval nutrition, larval density and age have been shown to influence insecticide susceptibility in Aedes mosquitoes, leading to discrepancies between bioassay results and the actual outcomes of insecticide treatments in the field [43]. Further work on IR in Aedes populations could identify and possibly reconcile differences between results from the laboratory and the field. To comprehensively evaluate IR status, future studies should also investigate the effectiveness of insecticide application methods, intensity, timing, and coverage by households and the MoH, as well as the impact of larvicides, such as temephos, which is commonly used in this study area, as well as Bacillus thuringensis israelensis (Bti), which is a common control method in other parts of the world.
Geographic methods, particularly spatial statistics and modelling, are well suited for understanding the patterns and drivers of IR at meso- and local scales. Employing these approaches can lead to more targeted, efficient, and sustainable vector control efforts. Future research in this area should continue to explore the spatial and temporal variability in IR among Ae. aegypti populations. Recent work in Yucatan, Mexico, demonstrated that IR levels could vary significantly across neighborhoods within the same city [24]. Additionally, work on the temporal dynamics of resistance could be beneficial for vector control decision-making. For example, in a study on pyrethroid-resistant, field-derived Ae. aegypti, researchers demonstrated that susceptibility to pyrethroids could be restored within ten generations when the selective pressure of regular insecticide treatments was removed [25]. While this experiment was conducted in a controlled environment, similar work within the context of urban environments, such as the four cities included in this study, could help calibrate timing of insecticide class rotations, allowing for better long-term management of susceptibility.
In conclusion, the Ae. aegypti collected in these four cities in Ecuador showed varying levels of resistance to the insecticides tested, and these measures typically changed over the course of the three seasons during which sampling took place. Regular IR monitoring should be conducted as long as insecticide applications remain an integral component of vector control activities, particularly in areas where these operations are deployed to control arbovirus transmission. Beyond this monitoring process, appropriate alternative management strategies should be deployed when IR is detected. These strategies can include biological control and community mobilization to reduce Ae. aegypti breeding sites.
|
10.1371/journal.pgen.1000356 | Genetic Evidence for Single-Strand Lesions Initiating Nbs1-Dependent Homologous Recombination in Diversification of Ig V in Chicken B Lymphocytes | Homologous recombination (HR) is initiated by DNA double-strand breaks (DSB). However, it remains unclear whether single-strand lesions also initiate HR in genomic DNA. Chicken B lymphocytes diversify their Immunoglobulin (Ig) V genes through HR (Ig gene conversion) and non-templated hypermutation. Both types of Ig V diversification are initiated by AID-dependent abasic-site formation. Abasic sites stall replication, resulting in the formation of single-stranded gaps. These gaps can be filled by error-prone DNA polymerases, resulting in hypermutation. However, it is unclear whether these single-strand gaps can also initiate Ig gene conversion without being first converted to DSBs. The Mre11-Rad50-Nbs1 (MRN) complex, which produces 3′ single-strand overhangs, promotes the initiation of DSB-induced HR in yeast. We show that a DT40 line expressing only a truncated form of Nbs1 (Nbs1p70) exhibits defective HR-dependent DSB repair, and a significant reduction in the rate—though not the fidelity—of Ig gene conversion. Interestingly, this defective gene conversion was restored to wild type levels by overproduction of Escherichia coli SbcB, a 3′ to 5′ single-strand–specific exonuclease, without affecting DSB repair. Conversely, overexpression of chicken Exo1 increased the efficiency of DSB-induced gene-targeting more than 10-fold, with no effect on Ig gene conversion. These results suggest that Ig gene conversion may be initiated by single-strand gaps rather than by DSBs, and, like SbcB, the MRN complex in DT40 may convert AID-induced lesions into single-strand gaps suitable for triggering HR. In summary, Ig gene conversion and hypermutation may share a common substrate—single-stranded gaps. Genetic analysis of the two types of Ig V diversification in DT40 provides a unique opportunity to gain insight into the molecular mechanisms underlying the filling of gaps that arise as a consequence of replication blocks at abasic sites, by HR and error-prone polymerases.
| An important class of chemotherapeutic drugs used in the treatment of cancer induces DNA damage that interferes with DNA replication. The resulting block to replication results in the formation of single-strand gaps in DNA. These gaps can be filled by specialized DNA polymerases, a process associated with the introduction of mutations or by recombination with an undamaged segment of DNA with an identical or similar sequence. Our work shows that diversification of the antibody genes in the chicken B cell line DT40, which is initiated by localized replication-stalling DNA damage, proceeds by formation of a single-strand intermediate. These gaps are generated by the action of a specific nuclease complex, comprising the Mre11, Rad50, and Nbs1 proteins, which have previously been implicated in the initiation of homologous recombination from double-strand breaks. However, in this context, their dysfunction can be reversed by the expression of a bacterial single-strand–specific nuclease, SbcB. Antibody diversification in DT40 thus provides an excellent model for studying the process of replication-stalling DNA damage and will allow a more detailed understanding of the mechanisms underlying gap repair and cellular tolerance of chemotherapeutic agents.
| Homologous recombination (HR) contributes to genome maintenance by repairing double-strand breaks (DSBs) and single-strand lesions. It accomplishes this by associating the damaged DNA with intact homologous sequences (reviewed in [1]). Genetic studies of Escherichia coli indicate that DSBs are recognized by the RecBCD enzyme at the initial step of HR, while single-strand gaps are loaded with RecA with the help of the RecF, RecO and RecR (RecFOR) proteins [2] (reviewed in [3]). In yeast and vertebrate cells, however, it remains unclear whether single-strand lesions can also directly stimulate HR, or if their replication leads to DSBs, which then stimulate HR.
The process of DSB-induced HR is well characterized in the budding yeast [4]. First, DSBs are resected by a nuclease to generate a 3′ overhang. A major nuclease in this process is thought to be a complex containing three proteins: Mre11, Rad50 and Nbs1 (called the MRN complex) (reviewed in [5]). The role of the 3′–5′ exonuclease activity of purified Mre11 in DSB repair remains enigmatic, as DSB resection is of opposite polarity in vivo [6]. Recent studies indicate that the MRN complex requires another factor to function: CtIP, the ortholog of Sae2 and Ctp1 in S. cerevisiae and S. pombe, respectively [7]–[9]. Biochemical study demonstrated that Sae2, a cofactor of the MRN complex, can process a single strand nick, and expand it [10]. The single-strand DNA generated adjacent to the DSB is coated with polymerized Rad51, resulting in the formation of nucleoprotein filaments. The assembly of RAD51 at DNA damage sites is regulated by a number of RAD51 cofactors, including the tumor-suppressor gene BRCA1 (Breast Cancer Susceptibility Gene 1), BRCA2, and five RAD51 paralogs (RAD51B/C/D and XRCC2/3) (reviewed in [11],[12]). The Rad51-containing single-strand DNA filaments play a role in the search for homologous DNA sequences and subsequent strand invasion into homologous duplex DNA. The importance of the role of the MRN complex in genome maintenance is indicated by a marked increase in the number of spontaneously arising chromosomal breaks followed by cell death after depletion of Mre11 in DT40 cells [13], and is also indicated by the high incidence of tumorigenesis in certain hereditary diseases: ataxia-telangiectasia-like diseases (ATLD) and Nijmegen breakage syndrome (NBS), which result from hypomorphic mutations in the MRE11 and NBS1 genes, respectively [14]–[17].
A combination of HR and non-templated single-base changes contributes to Ig V sequence variation in chickens and in some mammalian species such as rabbits and cattle [18]. Similarly, the chicken DT40 B lymphocyte line undergoes templated HR-dependent diversification (hereafter called Ig gene conversion) as well as non-templated single-base substitutions (hereafter called Ig hypermutation) during in vitro passage [19]–[21]. HR introduces tracts of templated mutations to rearranged variable (V) regions [22]–[24]. An array of “pseudo-Vλ” regions, located upstream from the functional rearranged VJλ, provides donors for this non-reciprocal sequence transfer. Since donor and recipient segments have a ∼10% sequence divergence, sequential Ig gene conversion events are able to substantially diversify Ig V [24].
Both types of Ig V diversification are initiated by activation-induced deaminase (AID), which forms uracil from deoxycytidine (dC) [25]–[27]. Uracil is subsequently removed by uracil-DNA-glycosylase- (UNG) mediated hydrolysis, which generates abasic sites [28]–[30]. In UNG−/− DT40 cells, the rate of C to T transitions is more than ten times greater than in UNG+/+ cells, indicating that more than 90% of the AID-induced uracil is accurately eliminated, presumably by base excision repair [28]. Non-templated hypermutation is generated as a consequence of translesion DNA synthesis (TLS) past abasic sites [31]. It is currently unclear how Ig gene conversion is induced by abasic sites, although it is likely that the abasic sites are converted to either single-strand gaps or DSBs, which in turn stimulate HR with upstream pseudo-Vλ segments. Current evidence points towards single-strand gaps, rather than DSBs, as the main downstream intermediate of abasic sites in the induction of Ig gene conversion for the following reason. In cells deficient in BRCA1, BRCA2 or Rad51 paralogs, where Rad51 is not accumulated efficiently at DNA lesions, the impaired HR causes a shift of Ig V diversification from HR- to TLS-dependent hypermutation [20],[32],[33]. Cleavage of template strands containing abasic sites cannot occur prior to TLS past the abasic sites. Thus, a common substrate for both Ig gene conversion and TLS is likely to be a single-strand gap and/or a stalled replication fork [34].
We hypothesized that if Ig gene conversion was triggered by single-strand lesions but not by DSBs, it would not involve the MRN complex (which is currently proposed as being involved in double-strand-break resection to generate recombinogenic 3′ ends). To test this hypothesis, we generated nbs1 hypomorphic mutant DT40 cells, where Nbs1 null mutant cells were rescued by an NBS1p70 transgene. The resulting ΔNBS1/NBS1p70 cells shared a phenotype very similar to cell lines established from patients with Nigmegen-breakage syndrome [35], including significant reduction in the frequency of HR-dependent DSB repair. Unexpectedly, the defect of Nbs1 also suppressed Ig gene conversion by two orders of magnitude.
To further define the role of the MRN complex in Ig gene conversion, we next attempted to reverse the defective Ig gene conversion by ectopically overexpressing chicken Exo1 [36]–[40] or E.coli Exo1 (SbcB) [41]–[43]. Exo1 is an evolutionarily conserved double strand-specific 5′ to 3′ exonuclease, and involved in mismatch repair in the eukaryotic cells. Additionally, the eukaryotic Exo1 can promote HR by facilitating 3′ tail formation at DSBs [38],[39]. Although both eukaryotic Exo1 and SbcB expand single-strand gaps from single-strand breaks in mismatch repair, SbcB can digest single-strand DNA at an opposite direction, 3′ to 5′, and thereby suppress DSB induced HR by removing 3′ overhang at DSBs (reviewed in [3]). Remarkably, the ectopic expression of SbcB normalized Ig gene conversion, but overexpression of chicken Exo1 did not. Conversely, the ectopic expression of chicken Exo1, but not SbcB, increased the frequency of DSB-dependent gene-targeting [44],[45], presumably by promoting the resection of DSBs. These data argue against the possibility that SbcB promotes Ig gene conversion by processing DSBs. Hence, these data support the notion that single-strand gaps may be supported a common direct precursor of both Ig gene conversion and error-prone gap-filling. In addition, our study thus suggests that the MRN complex is involved in HR, probably in two different ways: by processing DSBs and by generating recombinogenic single-strand lesions.
The chicken NBS1 gene is located on chromosome 2, which is trisomic in DT40 cells. To completely inactivate the NBS1 gene, we generated deletion constructs containing different marker genes, a procedure designed to remove the entire reading frame of the NBS1 gene, including all 16 exons (∼30 kb) (Figure 1A). These targeting plasmids were sequentially transfected into wild-type (WT) DT40 cells, and the NBS1−/−/+ cells were isolated. To generate conditional NBS1-disrupted cells, we employed Cre-recombinase-mediated deletion of a chicken NBS1 transgene. NBS1−/−/+ cells were transfected with the transgene containing the WT NBS1 cDNA flanked by loxP sites on both sides (the loxP-NBS1p95 transgene) together with a Cre-ER expression vector [46]. The resulting NBS1−/−/+/loxP-NBS1p95 clones were transfected with targeting constructs to disrupt exons 1–16 or exons 13–16, which encodes the Mre11-binding domain of the third NBS1 allele (Figure 1B). We were only able to obtain targeted integration with the latter construct, because Nbs1 overproduction from the loxP-NBS1p95 transgene substantially reduced gene-targeting efficiency. The genotype of the NBS1−/−/Δ13–16/loxP-NBS1p95 (hereafter ΔNBS1/loxP-NBS1p95) clones was confirmed by Southern-blot analysis of HindIII-digested genomic DNA for the disappearance of a WT 5 kb band (Figure 1C). Western-blot analysis showed that ΔNBS1/loxP-NBS1p95 cells expressed levels of Nbs1p95 that were about 50 fold higher than the WT cells (Figure 1D). ΔNBS1/loxP-NBS1p95 cells tended to grow more slowly than did WT cells (Figure 1E), a phenotype that may be attributed to the overexpressed NBS1p95.
To investigate whether Nbs1p95 is required for cellular proliferation, ΔNBS1/loxP-NBS1p95 cells were treated with tamoxifen to activate the Cre recombinase, resulting in the deletion of the loxP-NBS1p95 transgene. ΔNBS1/loxP-NBS1p95 cells ceased proliferating four days after the addition of tamoxifen (Figure 1E), with substantial numbers of dead cells (data not shown). These observations indicate that NBS1 is required for cellular proliferation, as previously reported [47]. To investigate the cause of the cell death, we scored spontaneous chromosomal aberrations when the cells were dying. The tamoxifen-treated ΔNBS1/loxP-NBS1p95 cells indeed exhibited extensive spontaneous chromosomal breaks (Figure 1F), as did Mre11 deficient cells [13], indicating an essential role for Nbs1 in repairing lethal double-strand breaks.
We also made conditional Rad50-depleted cells and found that they too exhibited an increase in the level of chromosomal breaks before cell death (Figure S1). Thus, a loss of Mre11, Rad50 and Nbs1 has a very similar effect on the maintenance of chromosomal DNA in cycling cells, suggesting that the three molecules form a functional unit, as do the yeast ortholog proteins [5].
We wanted to test whether or not expression of Nbs1p70 could rescue the cells from cell death. To this end, we complemented ΔNBS1/loxP-NBS1p95 cells with an NBS1p70 transgene and generated ΔNBS1/loxP-NBS1p95/NBS1p70 clones. The Nbs1p70 protein contains an Mre11-binding site, but lacks both the FHA and BRCT domains (Figure 2A) [5]. Western-blot analysis verified the Nbs1p70 expression, which was about 30 times higher than the expression of endogenous Nbs1 (Figure 2B). To remove the loxP-NBS1p95 transgene, ΔNBS1/loxP-NBS1p95/NBS1p70 cells were exposed to tamoxifen for three days, and isolated clones were examined for the expression of the Nbs1 protein. All surviving colonies expressed Nbs1p70, but not WT Nbs1p95, showing that their genotype is ΔNBS1/NBS1p70 (Figure 2B). The resulting clones proliferated with slightly slower kinetics than did the ΔNBS1/loxP-NBS1p95 cells (Figure 2C). We therefore conclude that Nbs1p70 is sufficient to rescue NBS1-deficient cells. This conclusion implies that the viability of previously described Nbs1-deficient DT40 cells might be attributable to the leaky expression of an N-terminally truncated protein [48].
Two representative ΔNBS1/NBS1p70 clones were further studied for their HR capability by measuring their gene-targeting frequency and sensitivity to DNA-damaging agents. Table 1 shows the ratio of targeted-to-random integration events at two loci. No gene-targeting events were detectable in the ΔNBS1/NBS1p70 clones. We next measured cellular sensitivity to ionizing radiation and camptothecin, a DNA-topoisomerase-I inhibitor [49]. Ionizing-radiation-induced DSBs are repaired by the two major DSB repair pathways, HR and nonhomologous end-joining [50], whereas camptothecin-induced DSBs are repaired exclusively by HR [51]–[53]. Compared with WT cells, the ΔNBS1/NBS1p70 cells showed a significant increase in damage sensitivity, particularly to camptothecin (Figures 2D and E). This is consistent with previous reports showing that Nbs1 promotes HR-mediated DSB repair [47],[48].
The rate of Ig gene conversion was assessed by measuring the re-expression of surface immunoglobulin M (sIgM) in DT40 clones that carry a defined frameshift mutation in the light-chain Vλ gene [21]. Since the frameshift is eliminated by superimposed Ig gene conversion, leading to the production of Igλ, the rate of Ig gene conversion can be evaluated by measuring the kinetics of sIgM gain (Figure 3A). Thirty subclones from each genotype were analyzed for sIgM-gain after 3 weeks of clonal expansion [33],[54]. The median value of the fraction of sIgM+ cells was 1.84% for WT, 1.91% for NBS1−/−/+ and 0.75% for NBS1−/−/+/loxP-NBS1p95 cells (Figure 3B). The reduced Ig gene conversion rate in NBS1−/−/+/loxP-NBS1p95 cells may result from the toxic effect of the overproduced Nbs1p95 protein. Two ΔNBS1/NBS1p70 clones displayed a significant decrease in gene conversion, with only 0.1–0.2% of subclones gaining sIgM, a level close to the background of the flow-cytometric analysis. To accurately evaluate the Ig gene conversion rate, we exposed populations of cells to trichostatin A, a histone-deacetylase inhibitor that increases the Ig gene conversion rate ∼50 fold [55],[56]. Following culture for 3 weeks in trichostatin A, the sIgM gain was only elevated to 2.15% in the ΔNBS1/NBS1p70 cells, while the WT cells exhibited an increase from 1.84 to over 90% (Figure 3C). This suggests that the intact MRN complex might promote Ig gene conversion, as reported previously [57]. Alternatively, the accuracy of Ig gene conversion in the ΔNBS1/NBS1p70 cells might be reduced, leading to a decrease in the re-expression of sIgM.
To examine the accuracy of Ig gene conversion, we determined the VJλ-nucleotide sequences from unsorted cells treated with trichostatin A for 4 weeks (Figure 3D). In trichostatin-A-treated unsorted WT cells, at least 42 Ig gene conversion events were detected among the 40 analyzed Vλ segments (1.3×10−2 events per Vλ per division). In contrast, the number of Ig gene conversion tracts was only one in 40 analyzed Vλ (3×10−4 events per Vλ per division) in ΔNBS1/NBS1p70 cells. This 42-fold difference is comparable to the difference observed in the sIgM-gain assay (Figure 3C). Ig V sequence analysis showed that the accuracy of these events is unaffected, as neither aberrant recombination nor accumulation of point mutations was found in ΔNBS1/NBS1p70 cells. To characterize the nature of Ig gene conversion, we also analyzed the VJλ nucleotide sequences of sorted sIgM+ revertants from ΔNBS1/NBS1p70 trichostatin-A-untreated cell populations. The frame-shift mutation in Ig Vλ [21] was indeed eliminated by superimposed gene conversion in all 40 analyzed fragments derived from ΔNBS1/NBS1p70 cells (data not shown). Furthermore, we found no change in the pattern of gene conversion, such as length of gene-conversion tracts (84 nucleotides on average for both ΔNBS1/NBS1p70 and WT [56]) or usage of pseudo-V donor segments, and no aberrant recombination (data not shown). Thus, although the defective Nbs1 function reduces the rate of Ig gene conversion, it compromises neither its accuracy nor donor gene preference.
To analyze Ig V hypermutation in ΔNBS1/NBS1p70 cells, we increased the level of AID expression by introducing an AID transgene into DT40 cells through retroviral infection [31],[58]. We assessed Ig V diversification by determining the nucleotide sequence of Ig Vλ in unsorted cells at 14 days post-infection (Figure 3D). WT and ΔNBS1/NBS1p70 cells exhibited similar levels of non-templated hypermutation: about 5.0×10−4 per nucleotide per division (Figure 3F). Thus, a defect in Nbs1 does not affect Ig V hypermutation.
AID overexpression increased the rate of Ig gene conversion from 5.2×10−4 to 1.3×10−2 per Vλ per division in 40 analyzed Vλ in WT cells (Figures 3E and F). Surprisingly, the frequency of Ig gene conversion in ΔNBS1/NBS1p70 cells reached the level of the WT cells, i.e., 1.6×10−2 per Vλ per division in 40 analyzed Vλ sequences. Thus, the frequency of gene conversion was increased 25 fold in WT cells and 307 fold in ΔNBS1/NBS1p70 cells by the ectopic expression of AID. No aberrant recombination events were observed. We conclude that a defect in Ig gene conversion in ΔNBS1/NBS1p70 cells is completely normalized by the ectopic expression of AID. This observation suggests two scenarios, described as follows: DSBs might initiate Ig gene conversion in a manner similar to the way in which AID-dependent DSBs trigger Ig-class switch recombination (reviewed in [59]). Thus, higher levels of AID expression may result in multiple deamination events on both strands, with the ensuing incisions more likely to generate DSBs carrying the 3′ tails even in the absence of the intact MRN complex. Alternatively, Ig gene conversion might be initiated by single-strand gaps. In the latter model, the formation of multiple abasic sites and incisions in one strand results in the generation of recombinogenic single-strand gaps, after which Nbs1p70 is no longer required for the processing of single-strand lesions to stimulate Ig gene conversion.
There are two major DSB repair pathways: HR and nonhomologous end-joining (NHEJ). Two studies previously reported the negative effect of NHEJ on Ig gene conversion [60],[61], which suggests that DSBs are an intermediate in Ig gene conversion. However, the IgV sequence from unsorted populations show only a two-fold increase [60] or no increase [61] in the rate of Ig gene conversion in NHEJ deficient clones in comparison with WT cells. Furthermore, another study [20] and our own work did not reproduce their data (data not shown). In general, it is difficult to draw a conclusion from at best a two-fold difference due to possible clonal variations in DT40 cells. To determine the involvement of DSBs in Ig gene conversion more accurately, we performed two experiments: 1) Detection of deletions within Vλ in RAD54−/− and KU70−/−RAD54−/− clones [50] (Figure 4A), and 2) terminal deoxytransferase (TdT) expression (Figure 4B and C).
In the first experiment, the effect of Ku70 depletion on Ig V diversification was investigated in the RAD54−/− background, where HR is not completed despite the accumulation of Rad51 at sites with DNA damage [62],[63]. Since the loss of Rad54 is substantially suppressed by NHEJ in the repair of x-ray-induced DSBs [50], we assumed that if Ig gene conversion is initiated by DSBs, a majority of such breaks would eventually be repaired by NHEJ in RAD54−/− cells, as are x-ray-induced DSBs. Thus, the additional inactivation of Ku70 in RAD54−/− cells would result in the deletion of Vλ sequences, as illustrated by the extensive deletion of the V(D)J coding joint in NHEJ-defective B precursors [64]. To detect deletion of Ig Vλ, we determined the nucleotide sequences of Vλ in AID overexpressing WT, RAD54−/− and KU70−/−RAD54−/− cells (Figure 4A). The RAD54−/− and KU70−/−RAD54−/− cells exhibited only one (6.6×10−4 per Vλ per division in 36 analyzed Vλ sequences) and three (1.7×10−3 per Vλ per division in 43 analyzed Vλ sequences) single-nucleotide deletion events, respectively. There were no longer deletions. Thus, unlike the repair of x-ray-induced DSBs, this result does not support the idea that unrepaired AID-induced damage at the Vλ segment of RAD54−/− cells is subject to NHEJ-mediated DSB repair.
In the second experiment, we overexpressed TdT, which added nucleotides at DSBs in a template-independent manner during V(D)J-joining [65],[66]. TdT has been shown to access the Ig locus when expressed in a human cell line that undergoes constitutive Ig somatic mutation in vivo [67]. If DSBs are a frequent trigger for Ig gene conversion, TdT-mediated nucleotide additions should be readily demonstrated at Ig Vλ in DT40 cells expressing TdT. We therefore transfected a TdT expression plasmid into WT DT40 cells and performed an Ig Vλ sequence analysis. The TdT overproduction affected neither point mutation nor Ig gene conversion frequency (Figure 4B). In contrast to the effect seen in hypermutating Ramos cells [67], we could not detect any difference in insertion frequency between WT cells with or without TdT overproduction (Figure 4C). Furthermore, all the insertions were of a single base pair, with the exception of one sequence where a deletion of 19 base pairs was associated with the insertion of CCC, which could not be accounted for by a pseudogene donor (ACAACGTCCC..19 bp del…GACAACC). This is the only example within the analyzed 109 sequences that may reflect the activity of TdT. The absence of additional nucleotides at Ig Vλ indicates that DSBs are not intimately associated with Ig gene conversion.
In summary, these data support the hypothesis that the initiating lesions for Ig gene conversion are predominantly single-strand gaps rather than DSBs. Hence, AID overexpression that normalizes the impaired Ig gene conversion of ΔNBS1/NBS1p70 cells (Figure 3D) possibly does so as a consequence of the formation of multiple incisions in one strand, which promotes the generation of recombinogenic single-strand gaps even in the absence of the intact MRN complex. This hypothesis is also supported by a previous biochemical study, which demonstrates that AID processively deaminates C residues on a single-strand DNA [68].
If Ig gene conversion is triggered by single-strand lesions, then the MRN complex is likely to contribute to Ig gene conversion, possibly by converting small single-strand lesions to larger, more recombinogenic gaps. To test this hypothesis, we attempted to normalize the impaired Ig gene conversion of the Nbs1-deficient cells by overproducing nucleases whose activity is precisely characterized. These nucleases included Exo1 [36]–[40] and SbcB [41]–[43]. Using a retroviral vector, we introduced individual nuclease transgenes into DT40 cells and established overproducing clones. We cultured individual clones for 2 weeks and determined the nucleotide sequences of the Vλ segment. Remarkably, SbcB dramatically increased the rate of Ig gene conversion in ΔNBS1/NBS1p70 (Figure 5A–C). Unexpectedly, this increase was not observed in ΔNBS1/NBS1p70 cells overexpressing chicken Exo1, presumably because this exonuclease can work only in a physiological context such as during mismatch repair in the chicken cell line. The frequency of Ig gene conversion in SbcB overproducing Nbs1-deficient cells reached 4.2×10−3 per Vλ per division in analyzed 45 Vλ sequences, a level higher than the gene-conversion frequency of the WT cells (Figure 5B). Ectopic SbcB expression did not significantly change the position (compare Figures 3D and 5A) or pseudo-V usage (Figure 5D) of the Ig gene conversion. In contrast, the nature of the Ig gene conversion was distinctly different between trichostatin-A-treated WT cells and those overproducing AID (Figure 3D and 5A). Presumably, this is because, according to a previous biochemical study [68], overproduced AID can deaminate even “cold” spots at Ig V, thereby initiating HR from a wider range of nucleotide sequences than does the endogenous AID of DT40 cells. Thus, it is likely that SbcB promotes Ig gene conversion in the same physiological manner as does the MRN complex. SbcB has the 3′ to 5′ exonuclease activity specific for single-stranded DNA in vitro [41]–[43], and can thereby expand single-strand gaps to stimulate HR in vivo. Hence, we conclude that the MRN complex contributes to Ig gene conversion in a similar manner by increasing the size of single-strand gaps.
To test whether overproduced SbcB affects HR-dependent repair of DSBs in vivo, we measured the effect of SbcB overproduction on DSB repair. To this end, we measured I-Sce1-induced gene-targeting [44]. We inserted the S2neo fragment carrying the I-SceI recognition site [69] into the OVALBUMIN locus of DT40 cells, subsequently transfecting the 3′neo fragment [69] (gene-targeting vector in Figure 5E) together with an I-SceI expression plasmid. Since gene-targeting of 3′neo into S2neo leads to the restoration of the WT neomycin-resistance (neoR) gene, the efficiency of gene-targeting events can be analyzed by measuring the frequency of neoR colonies. As previously observed [45], the co-transfection of the I-SceI-expression plasmid increases the gene-targeting frequency of 3′neo by more than three orders of magnitude. To test whether SbcB affects DSB-induced gene-targeting, we measured gene-targeting frequency following transfection of both the 3′neo gene-targeting fragment and the I-SceI-expression plasmids, along with either a nuclease-expression-plasmid (SbcB or the chicken Exo1-expression plasmids) or a negative control vector into WT DT40 cells. The ectopic expression of SbcB had no impact on DSB-induced gene-targeting (Figure 5F). In contrast, overproduction of chicken Exo1 increased the frequency of gene-targeting events more than 10 fold. This observation argues against the involvement of overproduced SbcB in DSB repair.
We show in this study that DT40 cells deficient in the individual components of the MRN complex exhibit similar phenotypes, including extensive chromosomal breaks prior to cell death. This observation suggests that Nbs1 participates in HR as part of the MRN complex, as does the MRX complex in yeast. As expected, the lethality of Nbs1-deficient cells was rescued by the expression of the Nbs1p70 N-terminal-truncated protein. ΔNBS1/NBS1p70 cells showed a significant decrease in the rate of Ig gene conversion. In the following subsections we present evidence that suggests that Ig gene conversion may be initiated by AID-induced single-strand lesions and that the MRN complex contributes to Ig gene conversion presumably by processing these single-strand lesions to generate recombinogenic gaps.
Two mechanisms could underlie the AID-dependent initiation of Ig gene conversion. The first assumes that AID-dependent single-strand lesions are converted to DSBs (possibly by blocking replication in one of the two sister chromatids), which stimulate Ig gene conversion. The second states that AID-dependent single-strand lesions directly trigger Ig gene conversion. The first scenario is unlikely for five reasons. First, in brca1, brca2 and rad51-paralog DT40 mutants, which are defective in the accumulation of Rad51 at sites of DNA damage, inefficient repair of AID-induced lesions activates TLS associated with hypermutation at dC∶dG basepairs [20],[32],[33]. Thus, the AID-induced substrate for HR is also likely to be the substrate for TLS-dependent Ig V hypermutation. Since effective TLS requires that there is no cleavage of the abasic-site-containing strand, it seems therefore plausibe that unfilled gaps directly stimulate Ig gene conversion in HR-proficient cells. (Figure 6). Second, if AID directly causes DSBs in Ig Vλ, such breaks would likely be repaired primarily by NHEJ in HR-deficient cells. Although it has been shown that AID-mediated DSBs trigger Ig-class switch recombination, which is partially dependent on NHEJ-mediated DSB repair [70], we did not obtain evidence for the involvement of NHEJ in Ig gene conversion, even in RAD54−/− cells (Figure 4A), where a late step of HR is compromised [62],[63]. This observation conflicts with the critical role NHEJ plays in the repair of X-ray-induced DSBs, as evidenced by the significant increase in sensitivity to x-rays in KU70−/−RAD54−/− cells compared with RAD54−/− cells [50]. Third, overexpression of terminal deoxytransferase failed to add extra-nucleotide sequences at the Ig Vλ of DT40 cells (Figure 4B and C). This observation argues against the significant association of Ig gene converstion with DSBs, because N nucleotides are inserted at the DSBs, as observed in DSB-induced V(D)J recombination [65],[66]. Fourth, although chicken Exo1 overproduction significantly increased the frequency of DSB-induced HR (Figure 5F), as observed in yeast [38],[39], the overproduction of SbcB did not enhance DSB-induced gene-targeting. However, SbcB reversed the defective Ig gene conversion in the Nbs1-deficient DT40 cells. Moreover, it is believed that SbcB suppresses DSB-induced HR, because its 3′ to 5′ exonuclease activity may remove the 3′ protruded tails from DSBs (reviewed in [3]). Collectively, these data suggest that DSBs do not play a major role in triggering Ig gene conversion, and that it is more likely that single-strand gaps formed by the sequential action of AID, UNG and the MRN complex directly stimulate Ig gene conversion.
At one time, models for both DSB- and nick-initiated HR were proposed [71],[72] (reviewed in [1]). The finding of DSBs during meiosis, as well as the development of the restriction-enzyme-induced HR model, established the DSB as the main initiator of HR [73],[74]. However, accumulating evidence indicates that single-strand lesions are indeed responsible for the initiation of HR in both RecFOR-dependent HR in E. coli and in mutant V(D)J recombinase-induced HR in episomal plasmids [2],[75]. Adding to this evidence, our study indicates that Ig gene conversion is a form of HR that is directly stimulated by single-strand lesions on chromosomal DNA in higher eukaryotic cells. The question remains as to whether or not single-strand gap-induced HR effectively contributes to the release of the replication block in the absence of accompanying DSBs.
The notion that single-strand lesions directly stimulate Ig gene conversion indicates that, like SbcB, the MRN complex may promote HR by converting single-strand breaks to more recombinogenic substrates such as single-strand gaps. In fact, according to the nick-initiating HR model, the initial nick is expanded into a single-strand gap to trigger HR [72]. Moreover, the presence of such activity is suggested by the biochemical study of CtIP, a protein that physically interacts with the MRN complex [8],[10]. On the other hand, Larson et al. indicate that the MRN complex incises a strand near an abasic site [76]. However, if this activity plays a dominant role in the initiation of Ig gene conversion, one cannot explain why the subsequent defect in the accumulation of Rad51 at the incision in the rad51 paralog and brca mutant shift Ig V diversification from HR- to TLS-mediated hypermutation [20],[32],[33]. Nonetheless, it is possible that the incision activity accounts for a fraction of Ig gene conversion. A defect in this incision activity might be substituted by AID overexpression, as it could introduce multiple AP sites, which makes less effective AP endonuclease compensate for the defective incision activity of the mutant Mre11 complex in ΔNBS1/NBS1p70 cells.
Figure 6 presents two models for the participation of the MRN complex in Ig gene diversification. In both models, AID-mediated catalysis and subsequent hydrolysis of uracil lead to the formation of abasic sites. The first model assumes an endonuclease that can cleave the opposite strand of the abasic-site-containing strand (Figure 6A), while the second model hypothesizes single-strand gap formation as a result of stalled replication (Figure 6B). The MRN complex facilitates HR by increasing the length of gaps in both models. Quick and copious recruitment of Rad51 at DNA lesions triggers Ig gene conversion, whilst poor recruitment leads to translesion DNA synthesis past abasic sites by error-prone polymerases. In the second model, it is still unclear why, despite the 10% sequence divergence between pseudo-V donor and V(D)J recipient fragments, competition between equal sister-chromatid HR and Ig gene conversion (Figure 6B) does not fully inhibit homologous recombination in the latter [24]. Presumably, extensive processing of single-strand lesions by the MRN complex and SbcB allows for homologous recombination, whilst impaired processing inhibits both TLS and Ig gene conversion (Figure 6B). The overproduction of AID might form gaps between two adjacent abasic sites on one strand, thereby suppressing the defective processing of single-strand lesions in ΔNBS1/NBS1p70 cells (Figure 6B). Additionally, the MRN complex contributes to Ig gene conversions through its incision activity [76], and its defect in ΔNBS1/NBS1p70 cells is rescued by the formation of multiple AP sites in AID overexpressing cells.
All genomic fragments in the NBS1-targeting constructs were amplified from DT40 genomic DNA using LA-PCR (Takara Bio, Kyoto) with the primers indicated below. To make the NBS1Δ1–16 plasmid, the upstream and downstream arms were amplified with 5′-AGCGTCGACCCCGCGTATTTCAGCAGCCTG-3′ and 5′-AAAAGCTTTGGTTCCTCGGTGCTCCTCACC-3′ primers and 5′-ATCTGAAGCTTGCTCCACTGATATGTTTGC-3′ and 5′-AAGCGGCCGCTTTGTGATTCAAACACTGGA-3′ primers, respectively. The resulting amplified upstream fragment was cut at the NotI site (derived from genomic sequence) followed by Klenow treatment and subsequently a HindIII cut. The 2.5 kb blunt-end HindIII fragment was cloned into the XhoI (blunt ended with Klenow treatment) HindIII site of pBluescript II (Stratagene) (named the pBS/NBS1 5′ arm). Two oligonucleotides, containing either EcoRI-BamHI-BglII-SalI or BamHI-BglII-HindIII, were inserted into the EcoRI-SalI or BamHI-HindIII site of the pBS/NBS1 5′ arm plasmid. The 3.5 kb 3′ arm was inserted into the HindIII-NotI site of the pBluescript (pBS/NBS1 3′ arm). To make the NBS1Δ1–16 blasticidin (Bsr) gene-disruption construct, the BsrR marker cassette was cloned into the BamHI site of the pBS/NBS1 5′ arm plasmid (with EcoRI-BamHI-BglII-SalI sites), followed by the ligation of the resulting plasmid (cut with SalI and NotI) with the SalI-NotI fragment containing the 3′ arm from the pBS/NBS1 3′ arm plasmid (NBS1Δ1–16 Bsr). Similarly, a Puromycin- (PuroR) marker cassette was cloned into the BamHI site of the pBS/NBS1 5′ arm plasmid, followed by the insertion of the HindIII and NotI fragment of the 3′ arm from the pBS/NBS1 3′ arm between the HindIII and NotI sites (NBS1Δ1–16 PuroR). To make the NBS1Δ13–16 gene-disruption construct, the upstream arm was amplified with 5′-TTGGAGGTCGACAAGCAAAACTGATGACGG-3′ and 5′-AAAGGATCCTCTTGGACAGCTGACAACCAG-3′ primers. The 7.5 kb SalI- (in genomic sequence) BamHI fragment of the amplified fragment was cloned into the XhoI-BamHI site of the pBS/NBS1 5′ arm plasmid (named pBS/Δ13–16 5′ arm). A neomycin- (Neo) marker gene cassette was cloned into the BamHI site of the pBS/Δ13–16 5′ arm. The resulting plasmid was ligated with the SalI-NotI fragment of the 3′ arm used for the NBS1Δ1–16 Bsr construct (NBS1Δ13–16). A probe for Southern hybridization was amplified from DT40 genomic DNA using the primers 5′-AAGCTTGCATGCAAACCTTGTTTTATCTTC-3′ and 5′-TGACTGCACTCTGCTCATTCTGGTATCTTC-3′.
The following two expression vectors were generated: 1) pBluescript-loxP-chicken β-actin promoter-multiple cloning site-internal ribosomal entry site (IRES) enhanced green fluorescent protein (EGFP) gene-loxP (named the plox vector), and 2) pBluescript-chicken β-actin promoter-multiple cloning site (named the pβ-actin vector). Chicken Nbs1p95 cDNA was amplified from pBS-NBS1 by PCR with the 5′-AAGAATTCAGAAAGAACTAGAAGGTTAAG-3′ and 5′-TTTGGGCTCGAGTTACAGATCCTCTTCTGAGATGAGTTTTTGTTCTCTTCTCCTCTTCACATTAGG-3′ primers and cloned into the BglII-SalI site of plox (plox/NBS1p95). To make the NBS1p70 cDNA (Figure 2A), NBS1p95 cDNA served as template DNA for PCR amplification using primers 5′-AAGGATCCATGGATGAGCCTGCCATTGG-3′ and 5′-TTTGGGCTCGAGTTAAGCGTAATCTGGAACATCGTATGGGTATCTTCTCCTCTTCACATTAGG-3′, and the amplified fragment was inserted into the BamHI-NotI site of the pβ-actin plasmid (pβ-actin/NBS1p70).
Chicken Rad50 cDNA was amplified by a standard RT-PCR method with primers 5′-ATGGCCAAGATTGAGAAAATGAGCATCC-3′ and 5′-TTAATGAACGTATGAGCCAAGGGAGC-3′, and then cloned into pTRE2 (Clontech) (pTRE2/RAD50) (Accession #XM_414645). Two RAD50 disruption constructs, RAD50-Bsr and RAD50-HisD, were expected to delete exon11 to 13 encoding amino-acid sequences from 579 to 735. The 3.9 kb 5′ arm was amplified from DT40 genomic DNA using primers 5′-TGCCATCAAGAGGAATCCAACTGGCCGTTA -3′ and 5′-CTCAGTGCTTTTGCCATGAAGCCAGTCTTC-3′ and cloned into pBluescript KS(+). The resulting plasmid was inserted with the 1.4 kb SpeI-SacI genomic fragment including exon 14, which was excised from a phage clone derived from the chicken genomic DNA library, where it served as the 3′ arm in the RAD50 disruption construct. Lastly, marker cassettes, Bsr or HisD, were inserted into the BamHI site to generate the RAD50-Bsr or RAD50-HisD gene-disruption construct. The genomic 3.4 kb SacI-EcoRI fragment, which is located at downstream of the 3′ arm, was used as a probe for Southern-blot analysis.
Cells were cultured in RPMI1640, supplemented with 10−5 M β-mercaptoethanol, 10% fetal-calf serum and 1% chicken serum (Sigma, St Louis, MO) at 39.5°C. Methods for DNA transfection and genotoxic treatments are as described previously [77].
WT DT40 cells were sequentially transfected with NBS1Δ1–16 BsrR and subsequently with NBS1Δ1–16-PuroR-targeting constructs to obtain NBS1−/−/+ cells. They were then transfected with an expression vector containing Cre-estrogen receptor chimeric recombinase (pANMerCreMer [46]) together with the plox/NBS1p95 plasmid. The resulting NBS1−/−/+/loxP-NBS1p95 cells were transfected with the NBS1Δ13–16 gene-disruption construct to obtain ΔNBS1/loxP-NBS1p95. ΔNBS1/loxP-NBS1p95 cells were transfected with the pβ-actin/NBS1p70 vector to make ΔNBS1/loxP-NBS1p95/NBS1p70 cells. ΔNBS1/NBS1p70 cells were generated by exposing ΔNBS1/loxP-NBS1p95/NBS1p70 cells to 100 nM tamoxifen for 3 days followed by subcloning, as described previously [46].
WT DT40 cells were transfected with the RAD50-Bsr disruption construct to generate RAD50+/− cells. They were co-transfected with the pTRE2/RAD50 and pTet-off (Clontech) plasmids simultaneously to make RAD50+/−/tetRAD50 cells. These cells were transfected with the RAD50-HisD construct to generate RAD50−/−/tetRAD50 cells. Conditional inactivation of the RAD50 transgene was done using tetracycline as previously described [13].
Clonogenic survival was monitored by a colony-formation assay, as described previously [77]. To measure sensitivity to camptothecin (Topogene, Columbus, OH), appropriate numbers of cells were plated into six-well cluster plates containing the complete medium and 1.5% methylcellulose (Aldrich, Milwaukee, WI), supplemented with camptothecin. Colony numbers were counted at 7 and 14 days, and the survival percentage was determined in terms of the number of colonies of untreated cells. To measure ionizing-radiation sensitivity, serially diluted cells were plated in the medium containing methylcellulose, irradiated with a 137Cs γ-ray source and then incubated. Measurement of chromosome aberrations was carried out as previously described [77].
Methods described previously were used for the preparation of whole-cell extracts and western-blot analysis, with the following modifications. For western-blot analysis, the mouse monoclonal anti-human Nbs1 antibody (BD Transduction Laboratories catalog #611871) was used at a 1∶100 dilution, and HRP-conjugated donkey anti-mouse IgG antibody (Santa Cruz Biotechnology catalog #sc-2314) was used at a 1∶5000 dilution. Chicken Rad50 antiserum was raised in a rabbit against a whole protein of chicken Rad50. For the western-blot analysis, rabbit polyclonal anti-chicken Rad50 antibody was used at a 1∶100 dilution, and HRP-conjugated donkey anti-rabbit IgG antibody (Santa Cruz Biotechnology catalog #sc-2004) was used at a 1∶5000 dilution. For the western-blot analysis, rat monoclonal anti-mouse AID antibody (kindly provided by Dr. K. Kinoshita, Kyoto University) was used at a 1∶500 dilution, and HRP-conjugated donkey anti-rat IgG antibody (Jackson ImmunoResearch catalog #712-035-150) was used at a 1∶5000 dilution.
To analyze the frequency of targeted integration events at the OVALBUMIN [78] and HPRT [79] loci, their disruption constructs were transfected into cells. Following selection of clones resistant to appropriate antibiotics, Southern-blot analysis was performed.
We confirmed that ΔNBS1/NBS1p70 cells retained the same frame-shift mutation in the V sequence as do WT cells [21]. Generation frequency of surface IgM (sIgM) loss variants as well as sIgM-gain revertants were monitored by flow-cytometric analysis of cells that had been expanded for 3 weeks after subcloning and then stained with fluorescein isothiocyanate-conjugated (FITC) goat anti-chicken IgM (Bethyl, Montgomery, TX). At least 30 subclones were analyzed in each genotype. To enhance Ig gene conversion, trichostatin A (TSA, Wako Osaka, concentration: 1.25 ng/ml) was added to a mixture of sIgM-negative subclones from WT and the ΔNBS1/NBS1p70 #1 genotypes shown in Figure 3B. The fraction of sIgM+ revertants was monitored over time, as described previously [55]. In each analysis, the abundance of sIgM-positive cells was determined as the percentage of live cells whose FITC fluorescence fell at least eight fold more than the FITC fluorescence peak of sIgM negative cells. Ig gene conversion frequency of unsorted cells was calculated based on the number of gene-conversion events, of analyzed Vλ clones and of cell divisions.
For retrovirus infection, the pMSCV-IRES-GFP recombinant plasmid was constructed by ligating the 5.2 kb BamHI-NotI fragment from pMSCVhyg (Clontech) with the 1.2 kb BamHI-NotI fragment of pIRES2-EGFP (Clontech). Mouse AID [58] or chicken ExoI (Accession #AB084249) or SbcB cDNA was inserted between the BglII and EcoRI sites of pMSCV-IRES-GFP [58]. The preparation and infection of retroviruses were carried out as previously described [58]. Expression of the GFP was confirmed by flow cytometry. The efficiency of infection was more than 90%m as judged by GFP expression. Cells were sub-cloned into 96 well-plates a day after infection. After 2 weeks, clones displaying a bright GFP signal were determined by FACS analysis.
WT DT40 cells were transfected with a pSV2neo-based plasmid containing human TdT under control of the β-globin promoter and IgH enhancer [67] by electroporation as previously described. Clones were analyzed for TdT expression by indirect immunofluorescence microscopy using a mouse monoclonal anti-TdT (Dako) followed by anti-mouse Igκ conjugated to FITC. TdT-positive clones were expanded for 4 weeks, following which Ig-negative loss variants were sorted by FACS and the rearranged light-chain gene sequenced and analyzed as previously described [20].
DNA was extracted from three to five clones from genotypes at 14 days after AID, Exo1 or SbcB retrovirus infection, or at 28 days after TSA treatment. PCR-amplified fragments of Vλ segments were cloned into a plasmid and subjected to base-sequence analysis. Rearranged Vλ was amplified by PCR with Pyrobest DNA polymerase (Takara Bio) (30 cycles of 94°C for 30 s, 60°C for 1 min, and 72°C for 1 min) with 5′-CAGGAGCTCGCGGGGCCGTCACT-GATTGCCG-3′ and 5′-GCGCAAGCTTCCCCAGCCTGCCGCCAAGTCCAAG-3′ primers, as previously described [20]. PCR products were cloned into the TOPO pCR2.1 cloning vector (Invitrogen) and sequenced with the M13 forward (−20) or reverse primer using an ABI PRISM 3100 sequencer (Applied Biosystems). Sequence alignment using GENETYX-MAC (Software Development, Tokyo, Japan) allowed the identification of changes from the parental sequences in each clone. Differentiating between non-templated nucleotide substitutions and gene conversion was carried out as previously described [20]. The rate of hypermutation was calculated based on mutation frequency and number of cell divisions (42 cycles in WT and 37 cycles in ΔNBS1/NBS1p70 for 14 days).
107 cells were suspended in 0.1 ml Nucleofector Solution T (amaxa), and electroporated using a Nucleofector (amaxa) at program B-23. 2 µg of linear 3′ neo DNA and 2 µg of circular I-SceI expression vector (pcBASce), together with 2 µg of either control (pBluescript II KS+), SbcB or chicken Exo1 expression vector, were transfected. 3′ neo DNA was amplified by PCR from the SCneo neo substrate plasmid [69] using Phusion DNA polymerase (Finnzymes) (30 cycles at 94°C for 30 s, 60°C for 30 s, and 72°C for 2 min), with 5′-GGATCGGCCATTGAACAAGATGGATTGCAC-3′ and 5′-GGAAACAGCTATGACCATGATTACGCCAAG-3′ primers. The amplified fragment was used for electroporation, as previously described [44]. 24 hours after electroporation, the number of live cells was counted by FACS and transferred to 96 well-cluster trays with or without 2.0 mg of G418 per ml. Cells were grown for 7 days, and HR frequencies were calculated by the following equation: HR frequency (colonies/cell) = number of G418-resistant colonies/(plating efficiency of transfected cells in the absence of G418×number of live cells determined by FACS at 24 hour after electroporation) [44].
|
10.1371/journal.pcbi.1000476 | Amyloidogenic Regions and Interaction Surfaces Overlap in Globular Proteins Related to Conformational Diseases | Protein aggregation underlies a wide range of human disorders. The polypeptides involved in these pathologies might be intrinsically unstructured or display a defined 3D-structure. Little is known about how globular proteins aggregate into toxic assemblies under physiological conditions, where they display an initially folded conformation. Protein aggregation is, however, always initiated by the establishment of anomalous protein-protein interactions. Therefore, in the present work, we have explored the extent to which protein interaction surfaces and aggregation-prone regions overlap in globular proteins associated with conformational diseases. Computational analysis of the native complexes formed by these proteins shows that aggregation-prone regions do frequently overlap with protein interfaces. The spatial coincidence of interaction sites and aggregating regions suggests that the formation of functional complexes and the aggregation of their individual subunits might compete in the cell. Accordingly, single mutations affecting complex interface or stability usually result in the formation of toxic aggregates. It is suggested that the stabilization of existing interfaces in multimeric proteins or the formation of new complexes in monomeric polypeptides might become effective strategies to prevent disease-linked aggregation of globular proteins.
| The aggregation of proteins in tissues is associated with the pathogenesis of more than 40 human diseases. The polypeptides underlying disorders such as Alzheimer's and Parkinson's are devoid of any regular structure, whereas the polypeptides causing familial amyotrophic lateral sclerosis or nonneuropathic systemic amyloidosis correspond to globular proteins. Little is known about the mechanism by which globular proteins under physiological conditions aggregate from their initially folded and soluble conformations. Interestingly, several of these pathogenic proteins display quaternary structure or are bound to other proteins in their physiological context. In the present work, we show that protein-protein interaction surfaces and regions with high aggregation propensity significantly overlap in these polypeptides. This suggests that the formation of native complexes and self-aggregation reactions probably compete in the cell, explaining why point mutations affecting the interface or the stability of the protein complex lead in many cases to the formation of toxic aggregates. This study proposes general strategies to fight against diseases associated with the deposition of globular polypeptides.
| The formation of insoluble amyloid protein deposits in tissues is related to the development of more than 40 different human diseases, many of which are debilitating and often fatal. The polypeptides responsible for these disorders are not related in terms of sequence or conformation [1]–[6]. Some of these proteins and peptides are mostly unstructured. Examples include amylin, amyloid-β-protein and α-synuclein. In contrast, many other amyloidogenic proteins are globular in their native state, implying that they have a properly packed and cooperatively sustained structure under physiological conditions. This group includes ß-2-microglobulin, transthyretin, lysozyme, superoxide dismutase 1 and immunoglobulins. As a general trend, evolution has endorsed globular proteins with solubility in their biological environments [7]. However, it has been shown that, in vitro, under conditions where they become totally or partially unfolded, both these pathogenic proteins [8]–[11] and many globular polypeptides not related to disease [12]–[15] readily convert into aggregates and ultimately into highly structured amyloid fibrils. This self-assembly process is triggered by the destabilization and opening of the native structure, which exposes previously protected aggregation-prone regions that can nucleate the aggregation reaction and participate in forming the β-core of the mature fibril through specific intermolecular interactions [16]–[18]. Such amyloidogenic sequence stretches have been described in most of the polypeptides underlying neurodegenerative and systemic amyloidogenic disorders. The main intrinsic protein properties that promote the assembly of such sequences into fibrils have been recently defined [19], and several algorithms that predict amyloidogenic sequences with good accuracy are already available [3],[20],[21].
Although the study of protein aggregation from non-native states has provided a wealth of data on the physico-chemical determinants of amyloid formation, little is known about how globular proteins aggregate from their initially folded and soluble conformations under physiological conditions, where extensive unfolding is not expected to occur [22]. Deciphering this issue is important because the deposition of globular polypeptides is linked to devastating disorders, and there is an urgent need for therapeutic intervention.
Protein aggregation can be seen as an anomalous type of protein-protein interaction. In functional interactions, binding partners come together in a stable and precise orientation in seconds [23]. This efficiency relies on the structural features of the interacting surfaces. Perhaps the most significant characteristic of a functional protein-protein interface is the presence of small high-affinity regions within the interface, with a reduced number of residues accounting for most of the binding energy [24]–[26]. Several computational approaches have been shown to forecast such regions with high accuracy [27]–[34]. Statistical analysis of the structures of protein-protein interfaces has revealed that tryptophan, phenylalanine, and methionine and to a lesser extent leucine, valine, and tyrosine are preferentially conserved at interaction sites [35]. The same residues have been shown to be conserved in the aggregation-prone sequences of the human proteome [36]. This suggests an intriguing possibility: that amyloidogenic regions and interacting surfaces might overlap in globular proteins. Several of the folded proteins linked to amyloid diseases display quaternary structure or are bound to other proteins in their physiological context. If these interactions specifically cover amyloidogenic regions, they could play a role in protecting native-state proteins from aggregation. Alternatively, incorrect docking of interfaces might facilitate the assembly of overlapping amyloidogenic regions and therefore the formation of toxic protein aggregates of globular proteins. In the present work, we have used available computational approaches to predict aggregation-prone sequences and interacting residues in order to assess the extent to which these regions coincide in pathogenic and non-pathogenic proteins.
The prediction of regions responsible for aggregation based on the primary sequence of a protein has been tackled by several methods, from simple considerations of the properties of amino acids to complex molecular dynamics calculations [37]–[44]. Overall, most of these methods predict with reasonable precision the regions of proteins in the cross-ß core of amyloid fibrils. This accuracy allows the proposal that the aggregation propensity of a polypeptide chain is ultimately dictated by the sequence [45]. Here we have used four different algorithms in parallel to provide a consensus prediction of the amyloidogenic regions in globular proteins linked to deposition diseases (see Methods). We chose the algorithms implemented by Fernandez-Escamilla et al. (TANGO) [38], Conchillo-Sole et al. (AGGRESCAN) [40], Galzitskaya et al. [41], and Zhang et al. [43]. All of them use the primary sequence as input and assume that the detected regions need to be at least partially exposed to solvent in order to nucleate the aggregation reaction.
Identification of binding sites in polypeptides is a direct computational approach to deciphering biological and biochemical function. Although sequence-based approaches to identifying protein interfaces exist, their results are often unsatisfactory. Here, we have used three different structure-based methods whose algorithms are publicly available as web servers to produce a consensus prediction of the interaction interfaces in the globular proteins under consideration (see Methods). These structure-based methods were developed by Fernandez-Recio et al. (ODA) [32], Murakami and Jones (SHARP2) [31], and Negi et al. (InterProSurf) [33]. Although they are based on different principles and implement diverse computational strategies, all of them use the unbound three-dimensional structure of a globular protein as input.
Two levels of prediction were considered: i) residues predicted or shown to be both in aggregation-prone regions and at interfaces and ii) residues in aggregation-prone sequences that are close in space to the interaction surface (below 3 Å). The interaction predictions were compared with the experimentally determined contacts in the quaternary structure of the proteins or in complexes of the studied proteins with other polypeptides. The regions predicted to have high aggregation propensity were compared with fragments of the analyzed proteins shown experimentally to form amyloid aggregates or to be located in the core of the mature fibrils formed by these polypeptides. We have defined a parameter called Interface Proximity Index (IPI) to evaluate the degree to which an aggregation-prone region is closer to a given interface than to the rest of the protein surface (see Methods and Figure 1).
Amyloidosis related to β2-Microglobulin (β2-m) is a common and serious complication in patients on long-term hemodialysis [46]. Two aggregation-prone regions encompassing residues 22–31 and 60–70 were predicted for human β2-m (Figure 2). These regions neatly coincide with two secondary structure elements in β2-m: β-strand 2, formed by residues 21–31, and β-strand 6, formed by residues 61–71. Interestingly, most of the residues in these two regions appear to be solvent accessible (Table 1). In agreement with the prediction, the fragments 21–31 and 21–41 of β2-m self-assemble into fibrillar structures [47]. Also, a peptide corresponding to residues 59–79 and its shorter version 59–71 both form amyloid fibrils [48].
A main interaction cluster is predicted for human β2-m (Figure 2A). It involves Y26 and G29 in β-strand 2, residues H31-S33 in the loop connecting β-strands 2 and 3, residues D53-W60 in β-strand 5 and the adjacent loop, and finally, residues F62 and L63 in β-strand 6. Overall, 62% of the residues in regions with high aggregation propensity are less than 3 Å from predicted “hot spots” of interaction (Table 1), and 25% overlap with them. Specifically, residues at positions 26, 29, 31, 60, 62, and 63 are predicted to be important both for binding and for aggregation.
Class I major-histocompatibility-complex (MHC) molecules (HLA molecules in humans) are ternary complexes of β2-m, an MHC heavy chain, and a bound peptide [49]. The crystal structures of several of these complexes have been solved, providing a benchmark to evaluate the accuracy of the predicted interface. In HLA-A-class molecules, the interface of β2-m and the HLA heavy chain is well conserved [50] and typically comprises 16 β2-m residues: K6, Q8, 10Y, 11S, 12R, N24, Y26, H31, D53, S55, F56, W60, F62, Y63, D98 and M99. This includes 8 of the 15 interacting residues predicted for β2-m. Residues 24, 26, and 31 map to the first aggregation-prone region of β2-m, and residues 60, 62, and 63 map to the second one. Taking as an example the structure of one such HLA-A complex (PDB ID: 1DUZ) [51], 85% of the residues in β2-m aggregating regions are less than 3 Å from the interface in the complex (Table 1 and Figures 2B, 2D and 2E). The IPI values confirm that these regions are preferentially located close to the interface of the complex (Table 1 and Figure 1A).
Inside the cell, β2-m associates with the non-classical HLA class I molecule human hemochromatosis protein (HFE) [52]. Hereditary hemochromatosis is a genetic disorder characterized by defects in iron metabolism and associated with mutations in the HFE gene [53]. Some of these mutations prevent the binding of HFE to β2-m. There are 18 β2-m residues at the HFE/β2-m complex interface, according to its crystal structure (PDB ID: 1A6Z) [54] : I1, Q8, 10Y, 11S, 12R, N24, Y26, H31, D53, L54, S55, F56, W60, Y63, F62, L65, D98, and M99, including 9 of the 15 predicted interaction sites. Residues 24, 26, and 31 correspond to the first aggregation-prone region of β2-m and residues 60, 63, 62, and 65 to the second. Another significant feature of this complex is that 76% of the residues in regions with high aggregation potential are close to the interface with β2-m (Table 1 and Figure 2C). Therefore, the docking of the HLA heavy chain and HFE molecules on top of β2-m covers most of the residues in aggregation-prone regions because they are close to the interaction sites, as illustrated by their high IPIs (Table 1, Figure 1A and Figures 2F, 2G).
Aggregation of β2-m under physiological conditions is thought to be initiated by a cis-trans prolyl isomerization of the H31-P32 peptide bond [22]. The transition promotes repositioning of the hydrophobic side chains of F30, L54, F56, W60, F62, and Y63 as shown in the structures of the P32A and P32G mutants [55],[56]. Interestingly enough, all of these residues map in an aggregation-prone segment and/or at the interface. Although speculative, it is tempting to propose that conditions that promote the dissociation of β2-m complexes with the above proteins or related ones may uncover this region and facilitate its fluctuation towards amyloidogenic conformations. In fact, in vivo, β2-m is continuously shed from the HLA molecules in the cell surface into the serum and transported to the kidneys where it is eliminated. Renal failure increases the levels of circulating β2-m more than 50-fold and promotes its self-assembly and conversion into amyloid fibrils [57]. Consequently, dissociation of β2-m from the class I HLA complex effectively constitutes a critical initial step in its aggregation into amyloid fibrils.
Because the β2-m regions likely to be involved in aggregation are already located in preformed β-strands, local fluctuations may allow anomalous intermolecular interactions between these preformed elements, leading to the formation of an aggregated β-sheet structure without extensive unfolding. In this context, the formation of β2-m complexes both inside the cell and on the cell surface might play a protective role against β2-m aggregation, either by reducing conformational fluctuations or by preventing the exposure of dangerous amyloidogenic regions, or both.
Transthyretin (TTR) constitutes the fibrillar protein found in familial amyloidotic polyneuropathy (FAP), familial amyloidotic cardiomyopathy, and central nervous system amyloidosis. Around 100 different TTR mutations have been reported, many of which are amyloidogenic [58]. Native TTR is a homotetramer. Five aggregation-prone regions are predicted for the TTR monomer. They encompass residues 11–19, 26–34, 92–96, 105–112, and 115–121. In this case, the aggregation-prone sequences appear to coincide precisely with preformed β-sheet structures: A β-strand (11–19), B β-strand (26–36), F β-strand (91–97), G β-strand (105–112), and H β-strand (115–121). In concordance with the prediction, peptides 10–20 and 105–115, which map in the first and fourth aggregation-prone regions, have been shown to assemble into amyloid fibrils [59],[60].
A single interaction patch is predicted for the TTR monomer (Figure 3A). It involves 19 residues located in the A β-strand (L17, A19), in the loop between the A and B β-strands (V20–S23), in the α-helix (L82), in the loop between the helix and the F β-strand (S85-F87), in the F β-strand (E92), in the G and H β-strands, and in the loop between the G and H β-strands (L110, S112-T118). TTR is a dimer of dimers. In the dimers formed by the A and B or the C and D chains, the predicted clusters are contiguous, forming a large and continuous interaction patch. Of the residues in aggregation-prone regions in TTR, 41% are within 3 Å of predicted interaction sites (Table 1). With the exception of the I26-R34 fragment, all the regions with high aggregation propensity are located close to the predicted interface, and 30% of the residues in these segments overlap with predicted interaction sites. Residues 17, 19, 92, 110, 112, and the stretch 115–118 are predicted to be important both for aggregation and interaction events.
The crystal structure of the TTR tetramer (PDB ID: 1TTA) [61] reveals that the real interfaces between the four individual TTR chains involve residues L17, A19-S23, F87-E89, E92, V94-T96, Y105, L110, and S112-V122. In good agreement with the prediction, the interfaces include 16 of the 19 predicted interacting residues. Residues 17 and 19 map to the first aggregation-prone region, residues 92 and 94–96 to the third one, and residues 110 and 112–122 to the fourth and fifth stretches. Significantly, if we exclude the I26-R34 region (IPI<0), 90% of the residues in aggregating regions are close to the two interfaces of the TTR tetramer as confirmed by their overall high IPIs (Table 1, Figure 1B and Figures 3B, 3C). Accordingly, although these regions are mostly accessible to solvent in the monomer, they become protected in the native quaternary structure of TTR by the interaction of the TTR subunits (Figure 3D).
Dissociation of the TTR tetramer has been reported as a prerequisite for amyloidosis. The tetrameric structure dissociates into AB and CD dimers, but they are unstable in the absence of additional quaternary interactions, explaining why TTR exists in a primarily tetramer-monomer equilibrium [62]. The crystal structures of more than 10 FAP-related variants have been solved, showing that the mutants are essentially identical in tertiary and quaternary structure to the wild-type protein, precluding the presence of preformed conformational defects in the amyloidogenic mutants [63]. However, FAP-associated mutants are destabilized even when tetrameric. This destabilization favors tetramer dissociation to the amyloidogenic monomeric intermediate, exposing previously hidden, preformed, aggregation-prone β-strands. In this context, the overlap of interaction and aggregation surfaces in the AB and CD dimers appears to be an effective way to prevent TTR amyloidogenesis in physiological conditions. The success of this strategy is best exemplified by the behavior of the T119M TTR mutant. The presence of the T119M allele alleviates the effect of the aggressive V30M amyloidogenic mutation in patients carrying these two variants. It has been shown that heterotetramers that incorporate T119M subunits are more stable, dissociate at lower rates, and accordingly are less amyloidogenic [64].
Familial amyotrophic lateral sclerosis (fALS) is characterized by the presence of Copper-Zinc Superoxide Dismutase (SOD1) inclusions in spinal cords [65]. Native SOD1 is a homodimer. The SOD1 monomer displays four regions with high aggregation potential. They encompass residues 4–8 in ß-strand 1, 100–106 and 111–120 in ß-strands 6 and 7 and the loop connecting them, and residues 146–153 in ß-strand 8.
A total of 14 residues are predicted to be at the interface of the SOD1 monomer (Figure 4A). They correspond to E21, W32, G33, S105, S107, G108, H110, C111, I113-R115, G147, V148, and I151. Of the residues in aggregation-prone regions in SOD1, 61% are less than 3 Å from predicted interaction sites (Table 1), and 25% of them overlap the predicted interaction sites. In particular, residues 105, 111, 113, 114, 115, 147, 148, and 151 are predicted to be involved in both binding and aggregation.
According to the crystal structure of the SOD1 dimer (PDB ID: 2C9V) [66], the real interface between the two SOD1 subunits involves residues V5, V7, F50-T54, I113-R115, V148, and G150-Q153 (Figure 4B). Therefore, the interaction prediction is poor for the N-terminal part of SOD-1 but accurate for residues in the C-terminal region. Residues V5 and V7 are part of the first aggregation-prone region, S105 part of the third one, I113-R115 part of the fourth stretch, and V148 and G150-Q153 part of the last one. All the residues in the first and last aggregation-prone segments as well as residues C111-T116 are close to the dimer interface (Table 1). Accordingly, except for the 100–106 stretch (IPI<0), all the regions with high aggregation propensity in SOD display high IPIs (Table 1 and Figure 1D). Three out of the four cysteine residues in each SOD1 monomer (6, 111, and 146) are in those sequence stretches. C6 and C111 are present in the form of free cysteines whereas C146 forms a disulfide bond with C57. All of these regions are accessible to solvent in the monomeric form but become partially or totally protected upon dimer association (Figure 4C and 4D).
FALS has been shown to be associated with more than 100 different SOD1 mutations, which are scattered throughout the three-dimensional structure [67]. Among them, the A4V mutation has received special attention because it results in a rapidly progressing form of fALS [68]. Animal models suggest that the pathogenicity of the A4V SOD1 arises from an increased propensity to aggregate, forming amyloid fibrils or pores [69]. A4 is near the dimer interface and maps in the first aggregation-prone region. Hasnain and co-workers solved the crystal structures of dimeric forms of A4V and another FALS mutant, I113T [70]. I113 is also at the interface, in the third aggregation-prone region. Both variants display the same monomer fold and active-site geometry as WT, but their interfaces are destabilized. Ray and Lansbury have shown that a covalent link between the two A4V SOD1 subunits abolishes aggregation, suggesting that the monomer is an obligate intermediate along the aggregation pathway [71]. Other studies also support the idea that monomerization leads directly to aggregation and fibrilization [72]. However, other lines of evidence suggest that the cytotoxic properties of SOD1 are triggered by an incorrect connection of its cysteine residues. In support of this view, the toxicity of recombinant SOD1 in cultured cells is lost upon mutational removal of C6 and C111 [11], and nucleation of the aggregation reaction requires the presence of cysteine thiolates at both positions 57 and 146 [72]. In any case, it appears that the interface plays a protective role against aggregation in SOD1, by preventing the direct assembly of pre-formed and exposed aggregation-prone regions in the monomer, by stabilizing the monomer against conformational fluctuations that might expose amyloidogenic sequences, or by preventing the exposure and reshuffling of cysteine residues. Based on these observations, it has been proposed that the stabilization of the SOD1 dimer interface could become an effective approach to fight against fALS [71].
The light chains (LCs) of immunoglobulins have been implicated in the pathogenesis of amyloidosis in patients with monoclonal B-cell proliferative disorders (AL amyloidosis) [73]. When immunoglobulin molecules are secreted, two heavy chains (HCs) usually pair with two LCs to create a heterotetramer. Occasionally, free LCs are secreted, and these LCs can form homodimers. LC dimers can be innocuous, but they can also aggregate into pathogenic species. We have analyzed the aggregation propensity and interfaces of a non-pathogenic LC dimer (PDB ID: 2Q20) [74]. Five aggregation-prone regions are detected, encompassing residues 19–23, 31–38, 46–51, 71–78, and 84–89 located in the ß3, ß4 ß5, ß9, and ß10 strands, respectively (Table 1). The interface of the dimer involves 13 residues: D34, Y36, Q38, K42-P44, L46, E55, Y87, Q89, Y91, Y96, and F98. According to their IPIs, the second and fifth stretch are located preferentially at the interface of the complex, with 89% and 83% of their residues less than 3 Å from the interface, respectively (Table 1, Figure 1E and Figures 5A, 5B). It is important to note that both stretches map in preformed ß strands.
AL is distinct from other types of amyloidosis in that hypervariability yields a different set of mutations in each patient. Ramirez-Alvarado and co-workers have characterized an LC dimer isolated from an AL patient [74]. The pathogenic protein differs from its germline in seven residues. Only three changes are non-conservative, and all of them are located at the dimer interface: N34I, K42Q, and Y87H. The N34I and Y87H mutations occur precisely in the second and fifth aggregation prone regions in the protein. Ramirez-Alvarado and co-workers found that the mutant dimer has an interface that is rotated 90° from the canonical LC interface. The altered interface was accompanied by decreased thermodynamic stability of the dimer and accelerated fibril formation. This might result from the exposure and self-assembly of the above preformed aggregation-prone ß segments upon dimer destabilization or dissociation. Interestingly, the restorative mutation H87Y suffices to regain thermodynamic stability, delay amyloid formation, and restore the canonical dimer interface, illustrating a delicate balance between native and aberrant protein self-assembly.
Although AL is more frequent, in some systemic amyloidosis the amyloid deposits consist of an unusual form of IgG1 heavy chain (HC) [75]. The amyloid protein contains the complete heavy-chain variable (VH) domain contiguous to the third constant region (CH3) due to the total absence of the first (CH1) hinge and second (CH2) heavy-chain constant regions [75].
Using the structure of a complete human IgG1 antibody [76] as a model (PDB ID: 1HZH), we detected nine aggregation-prone regions in the heavy chain (Table 1). Four of the aggregation-prone regions are in the VH domain (29–38, 45–52, 87–93, and 100–106), three in the CH2 domain (275–281, 289–299, and 322–331), and two in the CH3 domain (390–397 and 435–442). Analysis of the structure of the oligomeric form of the antibody reveals that only the regions in the VH and CH3 domains of the heavy chain display high IPI values and therefore are adjacent to the interface in the native heterotetramer (Table 1, Figure 1C and Figures 5C, 5D). The truncated, pathogenic form of the IgG is found in monomeric form in urine, indicating that either it cannot associate or it dissociates from the light and heavy chains that block the exposure of the detected aggregating regions in a normal heterotetrameric IgG molecule. These sequence stretches are located in preformed ß strands and are ready for self-assembly reactions that might result in the observed amyloid deposits.
Human lysozyme forms amyloid fibrils in individuals suffering from nonneuropathic systemic amyloidosis. The disease is always associated with non-conservative point mutations in the lysozyme gene [77]. Four aggregation-prone regions were detected in human lysozyme, corresponding to residues 25–33, 57–66, 76–84, and 108–114. The first region maps in helix B, the second and third in the loop of the β-domain, and the last one around the short helix D (Table 1). In good agreement with the predictions, recent experimental data shows that the region comprising residues 26–123 is preferentially protected from proteolysis once it is incorporated into lysozyme amyloid fibrils [78].
Two different interaction clusters are predicted for human lysozyme (Figure 6A), one in the α-domain and the other in the ß-domain. The first involves residues in the loop of the β-domain: N60, R62-W64, N66, A73-N75, A76, and H78. The second cluster is located in helix C and around helix D and corresponds to residues A94, K97, R98, R107-W109, and W112. Residues K33 and W34 in helix B are also predicted to be involved in protein-protein interactions. Overall, 66% of the residues in regions with high aggregation propensity are less than 3 Å from predicted interaction sites, and 31% overlap with them. Residues 33, 60, 62–64, 66, 76, 78, 108, 109, and 112 might be implicated in both binding and aggregation reactions. Interestingly, residues I56, F57, W64, and D67, which are mutated in the four known single-residue familial variants associated with lysozyme amyloidosis, are comprised of or very close to protein segments with high aggregation propensity and/or interaction sites.
The mechanism of lysozyme aggregation under physiological conditions probably involves thermal fluctuations that transiently expose amyloidogenic regions [22]. These transitions are rare in the wild type protein, but they are more frequent in mutated forms related to amyloidosis. It has been suggested that residues 36–102 in the β-domain and helix C can unfold while the rest of the α-domain maintains a native-like conformation [9]. In particular, residues 78–80 have been proposed to have a high aggregation propensity and the lowest structural protection, and therefore the highest propensity to initiate aggregation [79]. This sequence includes predicted interacting residues in the loop of the ß-domain and also overlaps with the predicted 76–84 amyloidogenic region.
A single-domain fragment of a camelid antibody has been shown to inhibit the in vitro aggregation of the D67H amyloidogenic lysozyme variant [80]. The antibody epitope includes neither the site of mutation nor most of the protein region destabilized by the mutation; therefore it was suggested that the binding of the antibody prevents aggregation by restoring the structural cooperativity of the mutant protein through the transmission of long-range conformational effects [80]. The structure of the antibody-lysozyme complex (PDB ID: 1OP9) reveals that the epitope consists of 14 residues of the lysozyme molecule and encompasses residues located in the loop between the A and B helices in the α-domain (L15, G16, Y20), in the long loop within the ß-domain (A76, C77, H78, L79), and in the C-helix (A90, D91, A94, C95, K97, R98, R101) (Figure 6B). The epitope includes interaction residues in the first and second predicted clusters. Also, the residues in the loop of the ß-domain coincide with the 76–84 aggregation-prone region. Therefore, an alternative explanation for the protective action of the antibody could be that by docking on top of interaction clusters, it impedes the conformational fluctuation and exposure of the amyloidogenic region around residues 70–80 (Figure 6C).
A nice example illustrating how new binding interfaces can effectively inhibit amyloid formation has been recently reported for the Alzheimer's Aβ peptide. Two aggregation-prone regions comprising residues 16–21 and 29–40 are consistently predicted for Aβ (Figure 6D). The prediction is in excellent agreement with the experimental data in the literature indicating that these regions constitute the core of the Aβ fibrils [81]. Härd and co-workers have used the Z domain derived from staphylococcal protein A to evolve variants of this domain able to bind to Aβ with nanomolar affinity and abolish its aggregation (affibodies) [82]. The solution structure of one of these complexes illustrates how the affibody's protective effect is exerted by creating a new, continuous interface with Aβ that buries its two aggregation-prone regions within a large hydrophobic tunnel-like cavity (Figure 6E).
An important question to address is whether predicted interaction interfaces and aggregation-prone regions also coincide in monomeric and soluble proteins. Therefore, we have analyzed the predicted properties of four well-characterized soluble proteins: myoglobin, maltose binding protein, thioredoxin, and ubiquitin.
Human myoglobin is a compact protein not related to disease. Although after long exposure to high temperatures in vitro it unfolds and assembles into amyloid fibrils [15], it is a highly soluble protein in its native α-helical conformation. It displays four regions with high aggregation potential encompassing residues 8–15, 28–33, 67–76, and 110–117. This last segment partially overlaps with the peptide fragment 100–114 found to form amyloid structures in vitro [83]. A 12-residue interface is consistently predicted for myoglobin. It consists of residues L40, K42, F43, L89, S92, I99, P100, K102, Y103, I107, L137, and F138. Interestingly enough, only one residue (I111) in the predicted aggregating regions is close to the interface. In addition, its side chain is buried, resulting in a surface where predicted interaction and aggregation regions do not overlap (Figure 7A), a feature that might have evolved to resist aggregation.
Maltose binding protein (MBP) endows fused proteins with increased solubility indicating that it is by itself highly soluble [84]. However, because it is a relatively large protein (370 residues), 10 different aggregation prone regions are predicted, comprising a total of 82 residues. Similarly to the case of myoglobin, although 8 of these residues are close to the predicted interface, comprising residues F92, E153, F156, M321, E322, A324-I329 and W340, their side chains are not significantly exposed to solvent (Figure 7B).
Thioredoxin A (TRX) is another tag used to increase the solubility of recombinant proteins [85]. Three aggregation-prone regions comprising residues 22–27, 29–33, and 49–57 are detected in human TRX. The predicted interaction surface comprises residues T30-I38, D60, V71-T74, and A92. While the first and third aggregation stretches are at more than 3 Å of the predicted interface, the second one overlaps with it. Surprisingly, in contrast to myoglobin and MPB, this region is exposed to solvent (Figure 7C). This suggests that, as discussed in the previous section, it could be involved in protein assembly reactions. In fact, residues C32–C35 in this stretch constitute the consensus CXXC motif in the TRX active site. In agreement with this hypothesis, we found that in the solution structure of human TRX in a mixed disulfide intermediate complex with its target peptide from the transcription factor NF k-B, the second aggregation-prone region in TRX is part of the complex interface [86] (Figure 7D).
The question arises of why TRX does not self-assemble when it is free. It appears that evolution uses negative design to fight against protein deposition by placing amino acids that counteract aggregation at the flanks of protein sequences with high aggregation propensity [45]. These residues are called aggregation gatekeepers [87], and they reduce self-assembly using the repulsive effect of charge (Arg, Lys, Asp and Glu), the entropic penalty on aggregate formation (Arg and Lys), or incompatibility with ß-structure backbone conformation (Pro) [88]. Interestingly, P34 is adjacent in sequence to the TRX 29–33 aggregation prone region. P34 and the two basic, protruding K37 and K39 residues flank this region in the 3D-structure (Figure 7C), which overall would make self-assembly reactions far more difficult.
Ubiquitin is a small, soluble and highly conserved regulatory protein that is ubiquitously expressed in eukaryotes [89]. Three aggregation-prone regions are detected in ubiquitin, including residues 1–8, 42–47, and 67–74 in the ß1, ß3, and ß5 strands, respectively. In this case, the regions of the protein with the highest aggregation propensity overlap significantly with the predicted interaction interface (Figure 7E). This suggests that in principle, this surface is competent for protein assembly reactions. Importantly, it has been shown that ubiquitin binding motifs, such as CUE domains, bind precisely to a surface defined by the ß1, ß3, ß4, and ß5 strands of ubiquitin (Figure 7F) [90], illustrating again how aggregation-prone regions and interaction interfaces tend to overlap. In fact, biochemical and genetic analyses have defined the hydrophobic patch formed by the side chains of L8, I44, and V70 on the surface of ubiquitin as a key determinant for endocytosis and proteosomal degradation [91]. These three residues are located in each of the three aggregation-prone regions predicted for ubiquitin. Why, then, does ubiquitin not self-assemble when it is unbound in solution? An examination of the surface defined by the above ß-strands shows that ubiquitin uses negative design principles to avoid aggregation, placing a large number of positively charged residues on the edge of these strands and adjacent to them (Figure 7E). Upon binding to ubiquitin-binding domains, these basic residues are hidden at the complex interface.
It seems that the spatial coincidence of interfaces and sequences promoting self-assembly is not restricted to amyloidogenic proteins. To further confirm this extent, we analyzed the structure of 25 different eukaryotic proteins shown to form homodimers (Table 2 and Figure 8). As expected, the number of predicted aggregation-prone regions in a protein correlates with its size (R = 0.88). All the analyzed proteins present at least one aggregation segment in which half of the residues are closer than 3 Å to the interface, and 96% of them have at least one aggregation region in which >85% of the residues are adjacent to the interface (Table 2 and Figure 8). This supports the idea that the physico-chemical determinants of aggregation and native self-assembly might overlap significantly and is consistent with the observation that in homodimers, identical monomer subunits tend to associate by hydrophobic interactions [92]. After protein synthesis and folding, monomers probably associate rapidly into native homodimers due to the high local concentration of identical polypeptide chains, thus avoiding prolonged exposure of hydrophobic, aggregation-prone regions to solvent. Interestingly, in heterodimers, in which monomers spend a larger part of their lifetime in a non-associated state, the presence of gatekeeper amino acids (Lys, Arg, Glu, Asp, and Pro) at the complex interface is much greater than in homodimers [92], probably to prevent self-association between identical monomers.
During the revision of the present work, Vendruscolo and co-workers published a related study in which they used their algorithm Zyggregator to perform an extensive analysis of interfaces in protein-protein complexes [93]. Interestingly enough, they independently concluded that interface regions are more prone to aggregate than other surface regions. Also, in excellent agreement with our analysis on monomeric soluble proteins, they found that charged residues frequently disrupt hydrophobic patterns at interfaces and that regions of negative aggregation propensity tend to surround aggregation-prone regions, which suggests that monomeric and native oligomeric proteins have evolved similar strategies to prevent misassembly. In our study, the analyzed eukaryotic proteins were randomly selected from a dataset of non-redundant homodimers [92], without any previous knowledge of their 3D-structures. Interestingly enough, the aggregation-prone sequences near to the dimer interface are located in α-helices in ∼70% of the cases (Figure 8). This is in clear contrast with their location in globular amyloidogenic polypeptides, where they reside mainly in preformed ß-strands. Although the sample is not statistically significant, this observation might suggest that natural selection is acting against the presence of amyloidogenic ß-strands at homodimers interfaces. It is attractive to propose that, as shown here for amyloidogenic proteins, mutations at these protein interfaces and specifically at protective locations might lead to loss of function or toxic phenotypes in a significant number of, yet undescribed, human polypeptides.
In the present work, we have used computational tools to predict aggregation-prone regions and interaction sites in globular proteins related to depositional diseases and non-pathogenic polypeptides. From the comparison of the predictions with the structural and experimental data, it appears that protein-protein interaction surfaces and regions with high aggregation propensity overlap significantly in the quaternary structure of proteins.
The proximity and coincidence of protein-protein interfaces and aggregation-prone regions suggests that the formation of native complexes and the aggregation of their monomeric subunits probably compete in the cell. This implies that the molecular machinery that performs the vast array of cellular functions and the aggregates that might interfere with these functions promoting cell stress or even cell death are sustained by similar molecular contacts. It is likely that the specificity of native protein interfaces in protein complexes has evolved to minimize anomalous interactions and therefore detrimental protein aggregation reactions. In this sense, Vendruscolo and co-workers have recently identified disulfide bonds and salt bridges as specific interactions that can stabilize aggregation-prone interfaces in their native conformations in oligomeric proteins [93]. However, the balance between functional and aberrant self-assembly appears to be so delicate that point mutations that affect the interface or the stability of the complex, promoting a higher dissociation rate, usually lead to the formation of toxic aggregates, either through direct assembly of newly exposed aggregation-prone regions or by local unfolding of protein segments previously stabilized in the native structure of the complex.
Overall, the present analysis provides a rational to understand how globular proteins aggregate under physiological conditions, where they posses an initially folded and cooperatively sustained conformation and extensive denaturation is not expected to occur. The data strongly suggest that the stabilization of the interface in multimeric proteins, as in the case of TTR, SOD1, or LC immunoglobulins, and/or the blocking of conformational fluctuations and exposed amyloidogenic regions through the formation of new interfaces with other protein molecules, as in the case of lysozyme or Aß peptide, might be important strategies to delay the onset or slow the progress of conformational diseases caused by globular proteins.
The observed association between the failure to attain a native interface and the build up of harmful aggregates suggests that the range of genetic human diseases which ultimately might originate from the conversion of a soluble globular protein into toxic assemblies could be much larger than previously thought. Approaches combining the prediction of aggregation-prone regions from the linear protein sequence with the analysis of real or predicted protein interfaces in the 3D-structure might provide a means to identify physiologically and therapeutically relevant amyloidogenic sequences in the proteins linked to such disorders.
Aggregation-prone regions in the studied proteins were predicted using the primary sequence as input and a consensus of the output of four different available methods. The first algorithm we used is TANGO (http://tango.crg.es/). TANGO is based on the physico-chemical principles underlying ß-sheet formation, extended by the assumption that the core regions of an aggregate are fully buried [38]. The second algorithm employed was AGGRESCAN (http://bioinf.uab.es/aggrescan/). AGGRESCAN is based on the use of an aggregation-propensity scale for natural amino acids derived from in vivo experiments [40]. The third method, developed by Galzitskaya and co-workers, is based on the use of a packing density scale for natural amino acids and on the assumption that amyloidogenic regions are highly packed in the fibrillar structure [41]. The last approach was developed by Zhang and co-workers (ftp://mdl.ipc.pku.edu.cn/pub/software/pre-amyl/). It uses the microcrystal fibrillar structure of the prion hexapeptide NNQQNY [94] as a template and a residue-based statistical potential to identify amyloidogenic fragments of proteins [43]. All analysis was performed using the default parameters for each employed algorithm. In the present work, a sequence stretch in the analyzed proteins should comprise a minimum of five consecutive residues and be positively predicted by at least two of the above-mentioned methods to be considered an aggregation-prone region.
Interaction residues were predicted using the monomeric three-dimensional crystal structure of each of the studied proteins as input and a consensus of the output of three different algorithms. The first approach used to predict interaction surfaces was the Optimal Docking Area (ODA) method (http://www.molsoft.com/oda), which identifies continuous surface patches with optimal docking desolvation energy based on atomic solvation parameters adjusted for protein-protein docking [32]. Only the top ten ODA hot spots were considered. The second method we used was SHARP2 (http://www.bioinformatics.sussex.ac.uk/SHARP2). SHARP2 calculates multiple parameters for overlapping patches of residues on the surface of a protein. It considers the solvation potential, hydrophobicity, accessible surface area, residue interface propensity, planarity, and protrusion. Parameter scores for each patch are combined, and the patch with the highest combined score is predicted as a potential interaction site [31]. The patch size was selected by considering the interacting partner to be an identical protein, and only residues in the best-scoring patch were considered. The last algorithm used was InterProSurf (http://curie.utmb.edu/). This method is based on solvent-accessible surface area of residues in isolated proteins, a propensity scale for interface residues, and a clustering algorithm to identify surface regions with residues of high interface propensities [33]. Only the first five clusters were considered. All analysis was done using the default parameters for each algorithm. In the present work, a residue in the surface should be identified as at least by two of the above mentioned approaches to be considered an interaction site.
To evaluate whether the proximity of an aggregation-prone region to a given real interface is specific or the sequence stretch is as close to any other patch of the same size in the protein surface, we have defined the Interface Proximity Index: IPI
IPI = 1-(SP/IP)
IP = Interface Proximity = nR/nHS
nR = number of residues in the aggregation-prone region at less than 3 Å from the interface.
nHS = number of residues in the aggregation-prone region.
nS = number of residues in the aggregation-prone region at less than 3 Å from a randomly chosen protein surface that does not include the interface.
Each random surface was generated by an aleatory selection of a number of solvent exposed residues equal to the number of residues constituting the real interface. One hundred random surfaces were generated for each aggregation-prone region analyzed.
Solvent-accessible and buried residues in the monomeric complex subunits where identified using the PISA server at the European Bioinformatics Institute (http://www.ebi.ac.uk/msd-srv/prot_int/pistart.html).
An IPI≤0 indicates that the aggregation-prone region is equally or less close to the interface than to the rest of the surface. An IPI>0 indicates that the aggregation-prone region is closer to the interface than to the rest of the surface, e. g., an IPI = 0.5 indicates that the aggregation-prone region is half as far from the interface than from the rest of the surface. The maximum value for IPI is 1.
Figures were were generated with the Swiss-PDB viewer program (http://spdbv.vital-it.ch) and rendered with POV (Persistance of Vision).
|
10.1371/journal.pgen.1002951 | A Response Regulator Interfaces between the Frz Chemosensory System and the MglA/MglB GTPase/GAP Module to Regulate Polarity in Myxococcus xanthus | How cells establish and dynamically change polarity are general questions in cell biology. Cells of the rod-shaped bacterium Myxococcus xanthus move on surfaces with defined leading and lagging cell poles. Occasionally, cells undergo reversals, which correspond to an inversion of the leading-lagging pole polarity axis. Reversals are induced by the Frz chemosensory system and depend on relocalization of motility proteins between the poles. The Ras-like GTPase MglA localizes to and defines the leading cell pole in the GTP-bound form. MglB, the cognate MglA GTPase activating protein, localizes to and defines the lagging pole. During reversals, MglA-GTP and MglB switch poles and, therefore, dynamically localized motility proteins switch poles. We identified the RomR response regulator, which localizes in a bipolar asymmetric pattern with a large cluster at the lagging pole, as important for motility and reversals. We show that RomR interacts directly with MglA and MglB in vitro. Furthermore, RomR, MglA, and MglB affect the localization of each other in all pair-wise directions, suggesting that RomR stimulates motility by promoting correct localization of MglA and MglB in MglA/RomR and MglB/RomR complexes at opposite poles. Moreover, localization analyses suggest that the two RomR complexes mutually exclude each other from their respective poles. We further show that RomR interfaces with FrzZ, the output response regulator of the Frz chemosensory system, to regulate reversals. Thus, RomR serves at the functional interface to connect a classic bacterial signalling module (Frz) to a classic eukaryotic polarity module (MglA/MglB). This modular design is paralleled by the phylogenetic distribution of the proteins, suggesting an evolutionary scheme in which RomR was incorporated into the MglA/MglB module to regulate cell polarity followed by the addition of the Frz system to dynamically regulate cell polarity.
| Most cells are spatially organized with proteins localizing to specific regions. The ability of cells to polarize facilitates many processes including motility. Myxococcus xanthus cells move in the direction of their long axis and occasionally change direction of movement by undergoing reversals. Similarly to eukaryotic cells, the leading pole of M. xanthus cells is defined by a Ras-like GTPase and the lagging pole by its partner GAP MglB. We show that MglA and MglB localization depends on the RomR protein. RomR recruits MglA to a pole and MglB GAP activity at the lagging pole results in MglA/RomR localizing asymmetrically to the leading pole. Conversely, RomR together with MglB forms a complex that localizes to the lagging pole, and this asymmetry is set up by MglA/RomR at the leading pole. Thus, MglA/RomR and MglB/RomR localize to opposite poles because they exclude each other from the same pole. RomR also interfaces with the Frz chemosensory system that induces reversals. Thus, RomR links the MglA/MglB/RomR polarity module to the Frz signaling module that triggers the inversion of polarity. Phylogenomics suggests an evolutionary scheme in which the MglA/MglB module incorporated RomR early to impart cell polarity while the Frz module was appropriated later on to direct polarity reversals.
| The ability of cells to generate polarized distributions of signaling proteins facilitates many biological processes including cell growth, division, differentiation and motility [1]. The spatial confinement of the activity of signaling proteins lays the foundation for processes that require localized protein activity [2], [3]. For instance, directional migration of neutrophils during chemotaxis depends on the dynamic localization of the activated small GTPases Rac and Cdc42 to the front edge of cells where they stimulate the formation of cellular protrusions via actin polymerization while Rho activity is spatially confined to the rear end of cells to drive actomyosin contractility with retraction of cellular protrusions [4]. Similarly, chemotaxing cells of Dictyostelium discoideum exhibit actin polymerization based cellular protrusions at the front that are dependent of the localization of a small Ras-family GTPase [5]. In both systems, the subcellular localization of small GTPases is highly dynamic and changes in response to environmental conditions [4], [5]. Similar to eukaryotic cells, bacterial cells are highly polarized with proteins localizing to specific subcellular regions, often the cell poles [6]. Two major unresolved questions regarding cell polarity in general are how proteins achieve their correct subcellular localization and how this localization changes dynamically over time. In eukaryotic cells, members of the Ras-superfamily of small, monomeric GTPases have essential functions in regulating dynamic cell polarity [7]. Recent evidence suggests that the function of small Ras-like GTPases in dynamic cell polarity regulation is conserved from eukaryotes to prokaryotes [8].
Ras-like GTPases are binary nucleotide-dependent molecular switches that cycle between an inactive GDP- and an active GTP-bound form [9]. The GTP-bound form interacts with downstream effectors to induce a specific response. Generally, Ras-like GTPases bind nucleotides with high affinities and have low intrinsic GTPase activities [9]. Therefore, cycling between the two nucleotide-bound states depends on two types of regulators: Guanine-nucleotide exchange factors (GEFs), which function as positive regulators by facilitating GDP release and GTP binding, and GTPase activating proteins (GAPs), which function as negative regulators by stimulating the low intrinsic GTPase activity in that way converting the active GTP-bound form to the inactive GDP-bound form [9], [10].
If placed on a surface, cells of the rod-shaped bacterium Myxococcus xanthus move in the direction of their long axis with a defined leading and lagging cell pole [8], [11]. Occasionally, however, cells stop and then resume motility in the opposite direction with the old leading pole becoming the new lagging cell pole and vice versa [12]. These events are referred to as reversals and at the cellular level a reversal corresponds to an inversion of the leading and lagging cell poles [8], [11]. Recent evidence suggests that a signal transduction module consisting of the small, monomeric Ras-like GTPase MglA and its cognate GAP MglB is at the heart of the regulatory system that controls motility and the cell polarity axis in M. xanthus.
M. xanthus has two motility systems [11]. The S-motility system depends on type IV pili (T4P), which localize to the leading pole [13]. T4P are thin filaments that undergo cycles of extension, adhesion and retraction [14], [15]. During a retraction, a force is generated that is sufficiently large to pull a cell forward [16], [17]. The A-motility system depends on protein complexes often referred to as focal adhesion complexes (FACs) that are assembled at the leading pole and distributed along the cell body [18]–[20]. Each FAC is thought to consist of a multi-protein complex that spans the cell envelope [19]–[21]. In a moving cell, FACs remain stationary within respect to the surface on which the cell is moving [18]. The two motility systems function independently of each other; however, their activity is coordinated to generate force in the same direction [22].
During a reversal, the polarity of the two motility systems is inverted synchronously. Several T4P proteins localize in clusters at both cell poles and remain stationary during reversals [23]. In contrast, the PilB ATPase, which catalyzes extensions, primarily localizes to the leading pole, and the PilT ATPase, which energizes retractions, primarily localizes to the lagging cell pole. During reversals, PilB and PilT switch poles thereby laying the foundation for the assembly of T4P at the new leading pole [23]. In the case of the A-motility system, several proteins including AglQ, which is part of the A-motility motor [19], [21], AglZ, GltD/AgmU and GltF, which are part of the FACs, localize to the leading cell pole as well as to FACs between reversals [18], [21], [24]. During reversals, the polar protein clusters relocate to the new leading cell pole and, in parallel, the FACs are thought to change polarity [18], [19], [24]. Therefore, at the molecular level, a reversal involves a switch in the polarity of dynamically and polarly localized motility proteins.
MglA functions as a nucleotide-dependent molecular switch to stimulate motility and reversals at the cellular level [25]–[29]. MglA-GTP is the active and MglA-GDP the inactive form [26]–[28]. MglB is the cognate GAP of MglA [26]–[28]. Between reversals MglA-GTP localizes to the leading cell pole while MglA-GDP is distributed uniformly throughout cells [26], [28]. MglB localizes to the lagging cell pole [26], [28]. MglA-GTP generates the output of the MglA/MglB module and MglA-GTP is thought to stimulate motility at the leading cell pole by setting up the correct polarity of dynamically localized motility proteins and by stimulating T4P function and FACs assembly [26], [28]. MglB localizes to the lagging cell pole and excludes MglA-GTP from this pole by converting MglA-GTP to MglA-GDP and, thus, sets up the MglA-GTP asymmetry. In this way, MglA-GTP together with MglB define the leading/lagging polarity between reversals [26], [28].
The Frz chemosensory system induces cellular reversals but is not required for motility per se (Blackhart et al., 1985) The Frz system consists of seven protein [30] including the CheA histidine kinase FrzE and the FrzZ response regulator. Genetic and biochemical analyses have demonstrated that FrzZ is phosphorylated by FrzE and FrzZ serves as the output of the Frz system [31], [32]. The effect of Frz on reversals depends on MglA as well as on MglB [26], [28] and signaling by Frz induces the pole switch of MglA-GTP and MglB, thus, giving rise to an inversion of the leading/lagging polarity [26], [28].
We previously showed that the RomR response regulator, which consists of an N-terminal receiver domain and a C-terminal output domain, is essential for A-motility in M. xanthus [25]. Full-length RomR localizes in a bipolar, asymmetric pattern with a large cluster at the lagging pole and a small cluster at the leading cell pole. During reversals the polarity of the RomR clusters switches. The activity of response regulators is regulated by phosphorylation of a conserved Asp residue in the receiver domain [33]. A RomR variant in which this Asp residue in the receiver domain is substituted to Glu (RomRD53E), which is expected to partially mimic the phosphorylated state [34], causes a hyper-reversing phenotype while a substitution to the non-phosphorylatable Asn (RomRD53N) causes a hypo-reversing phenotype [25]. Because a cellular reversal involves the synchronous switch in polarity of both A- and S-motility proteins [25], these observations raised the question of the function of RomR in S-motility and in regulating the reversal frequency.
Here we re-examined the function of RomR in M. xanthus motility. We provide evidence that RomR is important for A- as well as for S-motility. Moreover, we show that RomR interacts directly with MglA and MglB. We show that RomR is a polar targeting determinant of MglA-GTP and that RomR together with MglB sets up the asymmetric polar localization of the MglA-GTP defining the leading cell pole. Similarly, we find that RomR sets up the asymmetric localization of MglB and that MglB and RomR are targeted to the opposite cell pole of MglA-GTP in an MglA dependent manner, thereby, defining the lagging cell pole. Thus, correct localization of MglA and MglB to opposite poles depends on RomR. For reversals, we show that RomR functions between the Frz chemosensory module and the MglA/MglB GTPase/GAP module. These observations in combination with phylogenomic analyses suggest that the MglA/MglB module together with RomR constitute the basic module for establishing cell polarity in gliding motility systems, and that the Frz system was incorporated at a later point to allow the dynamic inversion of the polarity axis during reversals. The paper by Zhang et al. [35] describes results similar to those reported here.
We previously demonstrated that RomR is required for A-motility based on the motility phenotype of a romR insertion mutant [25]. To determine the function of RomR in S-motility an in-frame deletion of romR (ΔromR) was generated in the fully motile strain DK1622, which serves as the wild type (WT) in this work. To assess A- and S-motility in the ΔromR mutant, motility was tested on soft (0.5%) agar, which is favorable to S-motility, and hard (1.5%) agar, which is favorable to A-motility [36]. S-motility is manifested by colony expansion with the formation of flares of cells at the edge of a colony and A-motility is manifested by colony expansion with the presence of single cells at the edge of a colony.
As shown in Figure 1A, the WT DK1622 formed the flares characteristic of S-motility on 0.5% agar, the ΔromR mutant was significantly reduced in flare formation and colony expansion, and the A+S− control strain DK1300 did not form these flares. On 1.5% agar, the WT displayed the single cell movements characteristic of A-motility at the edge of the colony whereas neither the ΔromR mutant nor the A−S+ control strain DK1217 did. Time-lapse microscopy of ΔromR cells at the colony edge on 1.5% agar and on 0.5% agar confirmed that the ΔromR cells did not display single cell movements on 1.5% agar and only displayed very limited movements on 0.5% agar (data not shown).
To confirm that the motility defect in the ΔromR mutant was caused by lack of RomR, we created a complementation construct in which a functional fusion between full-length RomR and GFP (RomR1–420-GFP) was produced from the constitutively active PpilA promoter at native levels (Figure S1) [25]. All motility defects were corrected by expression of RomR1–420-GFP (Figure 1B) [25]. From these analyses we conclude that RomR is important for S-motility in addition to A-motility.
Previous characterization of RomR described distinct regions: a response regulator receiver (REC) domain, and an output domain composed of a proline rich (Pro-rich) region and a glutamate (Glu-rich) region [25]. To more universally characterize RomR, we identified its homologs from a set of 1611 prokaryotic genomes. Similarity searches against this genome set using full-length RomR support that it is composed of two conserved regions (Materials and Methods). As expected, one conserved region corresponds to the REC domain. The output domain of RomR comprises two distinct regions: (i) a conserved α-helical C-terminal region (RomR-C) (Figure 1C and 1D) that corresponds to the previously described Glu-rich region and is not homologous to characterized domains; and, (ii) an unstructured region corresponding to the previously described Pro-rich region that links the two conserved regions (Figure 1C). Sequence analysis of all identified homologs showed that most maintain conservation of the RomR-C domain (Figure 1D; Figure 2) while the unstructured linker region was not conserved (Figure 1E). The linker regions show length and composition conservation within taxonomic groups suggesting that they may be associated with lineage-specific functions.
Previous studies [25] have shown that the REC domain alone cannot localize RomR to the poles but is important for reversals. In contrast, the output domain comprising the linker and RomR-C localize polarly and is important for stimulating motility. Informed by the RomR sequence conservation analyses, we carried out a detailed functional analysis of the individual parts of the RomR output domain fused to GFP. As mentioned, full-length RomR fused to GFP (RomR1–420-GFP) corrected the motility defects of the ΔromR mutant and displayed an asymmetric bipolar localization pattern (Figure 1B) consistent with previous observations [25]. The entire RomR output domain fused to GFP (RomR116–420-GFP), RomR-C alone (RomR369–420-GFP) and the linker alone (RomR116–368-GFP) also localized in an asymmetric bipolar pattern (Figure 1B). However, only the RomR116–420-GFP construct partially restored A- and S-motility in the ΔromR mutant (Figure 1B). Because the RomR-C construct RomR369–420-GFP accumulated at a lower level than native RomR (Figure S1), we examined a RomR-C construct that included a portion of the linker region (RomR332–420-GFP). RomR332–420-GFP accumulated at a level similar to native RomR (Figure S1) and showed asymmetric bipolar localization (Figure 1B). However, this construct was also unable to complement the motility defects of the ΔromR mutant (Figure 1B). From these analyses we conclude that RomR possesses two pole-targeting determinants, the linker region and RomR-C, which are individually sufficient for polar targeting. Moreover, both regions are required for motility.
In order to understand the potential interplay between RomR and other systems involved in motility, we compared its phyletic distribution to the distribution of mglA and mglB, in addition to genes that mark the presence of the Frz system (frzE), T4P (pilT) and gliding motility (gltF) in our genome set. The proteins of interest were identified using BLASTP searches, gene neighborhood analysis, and characteristic features (Materials and Methods). Informed by the analyses on which regions of RomR are conserved and functionally important, we used the REC and RomR-C portions of RomR to identify homologs. RomR was identified in 31 genomes whereas MglA (70 genomes) and MglB (60 genomes) are more widespread (Figure 2). Of the 60 genomes encoding both MglA and MglB, 26 also encode a RomR homolog (Figure 2). Thus, with the exception of five genomes, all genomes encoding a RomR homolog also encode MglA and MglB homologs. These five genomes support a close correlation between MglA, MglB and RomR: RomR in these five genomes have lost either REC or RomR-C, and none contain a complete, if any, MglA/MglB system (Figure 2). 10 of the 26 genomes encoding intact RomR proteins also encode a Frz system and all Frz encoding genomes encode homologs of MglA, MglB and RomR. The co-occurrence of Frz with RomR and RomR with MglA and MglB support a functional association between these proteins.
Genes for T4P and gliding motility were found in 476 and 12 genomes, respectively (Figure 2). Generally, MglA, MglB, RomR and Frz encoding genes co-occurred with genes for gliding motility suggesting a functional connection between these proteins. Similarly, all 26 genomes encoding intact genes for MglA, MglB and RomR also contained T4P encoding genes also supporting a functional connection between these genes.
To map the position of romR in the regulatory circuits controlling motility and reversals, we carried out genetic epistasis experiments, using motility and reversal frequencies as readouts for function. Motility assays confirmed that a ΔmglA mutant is non-motile [26], [28], unlike the ΔmglB or mglAQ82A mutants, which contain MglA locked in the active GTP-bound form, both of which display A- and S-motility and hyper-reverse [27] (Figure 3A and 3B). Next, we deleted romR in these three backgrounds to establish the relative order of the genes. The motility assays showed that mglA, mglAQ82A, and mglB are epistatic to romR as evidenced by the similar phenotypes shared between the single mutants and corresponding double mutants (Figure 3A). We analyzed the reversal frequencies of single cells in the ΔromR, mglAQ82A and ΔromR, ΔmglB double mutants and found that they displayed hyper-reversing phenotypes similar to mglAQ82A and ΔmglB single mutants (Figure 3B), respectively, which further supports the epistasis relationships observed in the motility assays. These data also demonstrate that the mglAQ82A and mglB mutations cause a bypass of the motility defects caused by the ΔromR mutation.
Previous work suggested that substitutions of D53 in RomR mimics the active phosphorylated state (RomRD53E) or the inactive non-phosphorylated state (RomRD53N) [25]. We confirmed that RomRD53N and RomRD53E both stimulate motility and that RomRD53N causes a hypo-reversing and RomRD53E a hyper-reversing phenotype (Figure 3B) [25]. Strains containing romRD53N or romRD53E in mglAQ82A or ΔmglB mutant backgrounds showed the hyper-reversing phenotypes similar to those of mglAQ82A or ΔmglB single mutants, respectively and no additive phenotype was observed (Figure 3B). Thus, the observed epistasis relationships are independent of the activation state of RomR.
The epistasis experiments combining the various mglA, mglB, and romR alleles suggest that romR acts in the same genetic pathway as mglA and mglB to stimulate motility and reversals. Moreover, the data are consistent with romR acting upstream of both mglA and mglB as a positive regulator and inhibitor, respectively. Because MglB is an inhibitor of MglA, an MglB inhibitor is formally similar to an MglA activator. Therefore, these experiments are consistent with three general models for how the effect of RomR on motility and reversals could be accomplished by (i) stimulating MglA; (ii) inhibiting MglB; or, (iii) a combination of the two.
Because frz acts upstream of mglA and mglB for reversals [26], [28], we tested whether romR lies between frz and mglA and mglB. The FrzZ protein is the direct output of the Frz system [31], [32]. To test the relationship between frz and romR, we combined a ΔfrzZ mutation, which causes a hypo-reversing phenotype [32], with different romR alleles.
Combining ΔromR with ΔfrzZ did not restore the motility defects caused by the ΔromR mutation (Figure 3A). A strain containing romRD53N, which is active for motility but not for reversals, and ΔfrzZ was motile and hypo-reversed similarly to the strains only containing ΔfrzZ or romRD53N (Figure 3B). A strain containing romRD53E, which is active for motility and causes hyper-reversals, and ΔfrzZ was motile and hyper-reversed with a frequency similar to that caused by romRD53E alone. In agreement with previous observations [26], [28], combining ΔfrzZ with mglAQ82A resulted in a strain that hyper-reversed with the same frequency as a strain only containing mglAQ82A. Thus, MglA is the most downstream part in the reversal circuit. These epistasis experiments suggest that romR and frzZ act in the same genetic pathway to stimulate reversals. Moreover, the data are consistent with frzZ acting upstream of romR and with frzZ acting as a positive regulator of romR for reversals.
The performed epistasis analyses support that MglA, MglB, RomR and FrzZ are part of a signaling network that regulates motility and reversals in M. xanthus. Previous studies of MglA, MglB, and RomR have demonstrated that all three proteins localize polarly. To understand how MglA, MglB and RomR interact to stimulate motility and reversals, we systematically determined the localization of MglA, MglB and RomR in the presence and absence of each other. We have been unable to construct a functional FrzZ fusion protein; therefore, FrzZ was excluded from these analyzes. First, MglA, MglB and RomR were localized using active fluorescent fusion proteins expressed at native levels in strains deleted for the relevant native copies [25], [26] (Figure S2). As previously observed, MglA predominantly localizes in a unipolar pattern, whereas MglB and RomR predominantly localize in a bipolar asymmetric pattern [26]–[28] (Figure 4A).
Next, we analyzed the localization of each protein in the absence of one other. We confirmed that MglA localization changes from unipolar to a predominantly bipolar symmetric pattern in the absence of MglB [26]–[28] (Figure 4A). In contrast, we found that MglA localized diffusely throughout the cytoplasm in the absence of RomR. When examining MglB localization, we found that MglB shifts from a predominantly bipolar asymmetric pattern to a bipolar symmetric pattern in the absence of RomR and a unipolar pattern in the absence of MglA (Figure 4A). RomR localization patterns showed a similar shift from predominantly bipolar asymmetric to unipolar in the absence of MglA, whereas it became more bipolar symmetric in the absence of MglB (Figure 4A). Therefore, all three proteins are mutually dependent for correct localization in all three pair-wise directions.
Lack of RomR causes diffuse localization of MglA. Because MglA-GDP localizes in a diffuse pattern [26] and MglA-GTP localizes polarly, we thought of four possibilities for how RomR could stimulate polar localization of MglA-GTP: (i) RomR acts as a GEF; (ii) RomR inhibits MglB GAP activity; (iii) RomR is an MglA polar targeting determinant; or, (iv) combinations of these activities. To explore these possibilities, we determined the localization of YFP-MglAQ82A, which is locked in the GTP-bound form and localizes in a bipolar pattern and with a central oscillating cluster in a ΔmglA mutant [27] (Figure 4B). In the absence of MglB, YFP-MglAQ82A localizes as in the ΔmglA mglB+ mutant [27]. In contrast, in the absence of RomR, YFP-MglAQ82A only localized to the central oscillating cluster (Figure 4B). Similarly, in the absence of RomR and MglB, YFP-MglAQ82A only localized to the central oscillating cluster (Figure 4B). Finally, we observed that in the absence of both RomR and MglB, YFP-MglA was primarily diffuse or formed a non-polar cluster and rarely formed polar clusters (Figure 4A). These localization patterns suggest that one function of RomR is as a direct polar targeting determinant of MglA; however, the data does not rule out the possibility that RomR may also regulate the nucleotide-bound state of MglA.
MglB-mCherry and RomR-GFP show a similar localization pattern in WT and in the ΔmglA mutant (Figure 4A). To determine whether MglB-mCherry and RomR-GFP colocalize, we constructed a strain expressing both fusion proteins. Consistent with the observations that RomR as well as MglB in moving cells localize with the large cluster at the lagging cell pole [26], [28], the two proteins colocalized in mglA+ cells with a bipolar, asymmetric localization (Figure 4C). MglB-mCherry and RomR-GFP also colocalized in the absence of MglA (Figure 4C). We previously showed that the unipolar RomR cluster in the ΔmglA mutant is at the pole containing T4P [25] and, thus, RomR and MglB both localize at the “wrong” pole in the absence of MglA. This observation in combination with the observation that in the absence of RomR, MglB becomes more symmetric and vice versa (Figure 4A) suggest that MglB and RomR depend on each other for bipolar, asymmetric localization and that MglA is important for establishing this pattern.
To test whether RomR interacts directly with MglA and/or MglB, we performed pull-down experiments. To this end we purified N-terminal His6-tagged MglB (His6-MglB) and C-terminal His6-tagged MglA (MglA-His6). When bound to a Ni2+-NTA-agarose matrix His6-MglB interacted with RomR in total cell extracts of WT M. xanthus as determined using α-RomR antibodies (Figure 5A). Similarly, when MglA-His6 was bound to the Ni2+-NTA-agarose matrix, it interacted with RomR in total cell extracts of WT M. xanthus (Figure 5A).
To discriminate between direct and indirect interactions between the three proteins, we purified N-terminally His6-tagged RomR (His6-RomR) and MalE-tagged RomR (MalE-RomR) and N-terminally GST-tagged MglA (GST-MglA). As shown in Figure 5B, GST-MglA bound to a glutathione-agarose column interacted with His6-RomR. In control experiments with purified GST, His6-RomR was not pulled-down. In a separate control experiment, a His6-PilP protein was not pulled-down by GST-MglA (data not shown). Thus, the interaction between GST-MglA and His6-RomR is specific and direct.
In a separate set of experiments, MalE-RomR bound to an amylose matrix interacted with His6-MglB (Figure 5C) but not with a His6-PilP control protein (data not shown). Moreover, purified MalE protein did not interact with His6-MglB. Thus, MalE-RomR interacts specifically and directly with MglB.
Here we report that M. xanthus motility is stimulated and regulated by two modules of signaling proteins: a polarity module consisting of the response regulator RomR, the small GTPase MglA, and the MglA GAP MglB, and a polarity inversion module consisting of the Frz chemosensory system with its output response regulator FrzZ. While the RomR/MglA/MglB polarity module is important for motility, the Frz polarity inversion module interfaces with the RomR/MglA/MglB module at the level of RomR to regulate motility by regulating the reversal frequency. Here we focused on understanding the network topology of the polarity module and how it interfaces with the polarity inversion module to ultimately regulate motility.
MglA-GTP functions to stimulate motility and reversals in the absence of MglB whereas the opposite is not the case. Therefore, MglA-GTP is the output of the MglA/MglB GTPase/GAP module (Figure 6). RomR, MglA-GTP and MglB are all polarly localized whereas MglA-GDP is not. We found that correct localization of the three proteins is mutually dependent in all three pair-wise interactions. Moreover, pull-down experiments using purified proteins and WT M. xanthus cell extracts or direct interactions studies with purified proteins together with previous results [26]–[28] show that the three proteins interact in all three pair-wise directions.
Based on the findings from the interaction and localization analyses, we suggest that RomR targets MglA-GTP to both poles and that MglB at the lagging cell pole is important for establishing the MglA-GTP/RomR asymmetry by means of its GAP activity. Thus, RomR is part of a MglA-GTP/RomR complex at the leading cell pole. Interestingly, MglA is neither polarly localized in the ΔromR mutant nor in the ΔromR, ΔmglB double mutant; however, the ΔromR mutant is strongly reduced in motility whereas the ΔromR, ΔmglB mutant is motile. We suggest that the crucial difference between the two strains is the presence and absence of the MglB GAP activity. In the ΔromR mutant, MglB is bipolar symmetrical and, consequently, the GAP activity is not confined spatially to a single pole and, therefore, MglA-GTP would be low. On the other hand, the ΔromR, ΔmglB mutant would not have GAP activity and, therefore, a sufficient level of MglA-GTP may accumulate to stimulate motility. In the ΔromR ΔmglB mutant, MglA is not polarly localized; nevertheless, this mutant is motile. Therefore, polar localization of MglA is not a strict requirement for motility.
The localization and interaction data suggest that MglB and RomR form a complex that is essential for establishing the bipolar asymmetric localization of the two proteins and that this asymmetry is established in an MglA-GTP/RomR-dependent manner. In total, these interactions generate a mutually-dependent circuit for asymmetric localization of the three proteins: (i) RomR targets MglA-GTP to the poles in the MglA-GTP/RomR complex, (ii) the MglB/RomR complex is essential for establishing the MglA-GTP/RomR asymmetry by means of the MglB GAP activity, and (iii) MglA-GTP/RomR is essential for establishing the MglB/RomR asymmetry.
Combining the localization and interaction data with the results of the epistasis experiments using motility and reversals as readouts, we suggest that between reversals RomR functions as a positive regulator of MglA by targeting MglA-GTP to the poles in the MglA-GTP/RomR complex and that RomR inhibits MglB (and in that way also activates MglA) by formation of the MglB/RomR complex that is targeted to the lagging cell pole in an MglA-GTP/RomR-dependent manner (Figure 6). The identification of the MglA/MglB/RomR polarity module for stimulation of motility provides a conceptual framework for detailed biochemical experiments to address whether RomR acts as a GEF on MglA and/or regulates MglB GAP activity.
The output of the Frz polarity inversion module is the FrzZ response regulator and the reversal-inducing activity of the Frz system depends on phosphorylation of FrzZ [31], [32]. Similarly, our data suggest that reversals are induced by RomR phosphorylation. Interestingly, the reversal frequency of the romRD53E mutant is two-fold lower than in the ΔmglB and mglAQ82A mutants possibly reflecting that RomRD53E is not a perfect mimic of phosphorylated RomR. Alternatively, the FrzZ signal is channeled to MglA and MglB in a pathway that is independent of RomR. Given that the romRD53N mutant has the same low reversal frequency as the ΔfrzZ mutant, we favor the former model. By combining our genetic data with previously published data [31], [32], we suggest that phosphorylated FrzZ acts as a positive regulator of RomR and that this effect likely depends on phosphorylation of RomR. In this model, RomR acts at the interface between the Frz polarity inversion module and the MglA/MglB/RomR polarity module (Figure 6).
This potential phosphorylation of RomR by an unknown mechanism induces a switch in the polarity of the MglA, MglB and RomR proteins. RomRD53N and RomRD53E both localize in a bipolar asymmetric pattern [25] suggesting that the effect of RomR phosphorylation is not directly on its polar localization or release. Clearly, detailed biochemical experiments will be needed to elucidate the interaction between FrzZ/RomR, MglA/RomR and MglB/RomR and how these interactions depend on the phosphorylation status of RomR. Our preliminary results suggest that the FrzE kinase does not phosphorylate RomR in vitro (Keilberg, D. unpubl). The widespread distribution of MglA, MglB and RomR in organisms lacking the Frz system suggests that the RomR phosphorylation state could be regulated by other mechanisms. Phosphorylated FrzZ could activate a yet to be identified histidine protein kinase, which would subsequently be involved in RomR phosphorylation, as has been described for the single receiver domain response regulator DivK in the activation of the histidine protein kinases DivJ and PleC in Caulobacter crescentus [37]. Alternatively, FrzZ and RomR may be part of a phosphorelay in which the phosphoryl-group would be transferred from FrzZ to RomR via a histidine-phosphotransfer protein as has been described for other phosphorelays [38]. Future experiments will be directed at distinguishing between these possibilities.
In bacteria many proteins localize to the cell poles [6]. Sophisticated mechanisms are employed to bacteria to facilitate polar binding of proteins: This polar localization can be mediated by trans-acting polar targeting factors as in the case of PopZ, which interacts directly with ParB and targets it to the cell poles in C. crescentus [39], [40]. Alternatively, proteins may localize to the cell poles based on recognition of membrane curvature as proposed for some peripheral membrane proteins in Bacillus subtilis [41], [42]. Understanding how MglA, MglB, and RomR recognize the cell poles will add to our understanding of the diversity of protein localization mechanisms and potential common traits they share.
The modular design of the regulatory circuits involved in motility and its control in M. xanthus are paralleled by the phylogenetic distribution of MglA, MglB, RomR and of the Frz system. With the exception of the M. xanthus proteins, the functions of these proteins are not known. Based on the analyses of the M. xanthus proteins, we suggest that MglA and MglB together with RomR may constitute a module for the spatial deployment of proteins, i.e. regulation of cell polarity (and giving rise to unidirectional cell movements without reversals in M. xanthus). Subsequently, the Frz chemosensory module was incorporated by some of these systems to establish a scheme for the dynamic temporal control of cell polarity (and giving rise to the irregular reversals observed in extant M. xanthus). As outlined in [43]–[46] the high degree of modularity of signaling systems makes these systems more evolvable in part because combining and integrating different modules allow for the comparatively simple evolution of signaling units with novel properties compared to building such units from scratch. The evolutionary scenario outlined here is in agreement with these concepts.
Plasmids were propagated in E. coli TOP10 (F−, mcrA, Δ(mrr-hsdRMS-mcrBC), φ80lacZΔM15, ΔlacX74, deoR, recA1, araD139, Δ(ara-leu)7679, galU, galK, rpsL, endA1, nupG) unless otherwise stated. E. coli cells were grown in LB or on plates containing LB supplemented with 1.5% agar at 37°C with added antibiotics if appropriate [46]. DK1622 was used as WT M. xanthus strain throughout and all M. xanthus strains used are derivatives of DK1622. M. xanthus strains used are listed in Table 1. Plasmids are listed in Table S1. Plasmid constructions are described in Text S1. Primers used are listed in Table S2. All DNA fragments generated by PCR were verified by sequencing. All M. xanthus strains constructed were confirmed by PCR. Plasmids were integrated by site specific recombination at the Mx8 attB site or by homologous recombination at the native site. The in-frame deletions of frzZ (ΔfrzZ) and romR (ΔromR) were generated as described [47] using pFD1 and pSL37, respectively. M. xanthus strains were grown at 32°C in 1% CTT broth [48] or on CTT agar plates supplemented with 1.5% agar. Kanamycin (50 µg/ml) or oxytetracycline (10 µg/ml) was added when appropriate.
Cells were grown to a cell density of 7×108 cells/ml, harvested and resuspended in 1% CTT to a calculated density of 7 ×109 cells/ml. 5 µl aliquots of cells were placed on 0.5% and 1.5% agar supplemented with 0.5% CTT and incubated at 32°C. After 24 h, colony edges were observed using a Leica MZ8 stereomicroscope or a Leica IMB/E inverted microscope and visualized using Leica DFC280 and DFC350FX CCD cameras, respectively. To quantify differences in motility, the increase in colony diameter after 24 h was determined. Briefly, the diameter of each colony was measured at two positions at 0 and 24 h. The increase in colony diameter was calculated by subtraction of the size at 0 h from the size at 24 h. Colony diameters were measured for three colonies per strain.
For microscopy, M. xanthus cells were placed on a thin 1% agar-pad buffered with A50 buffer (10 mM MOPS pH 7.2, 10 mM CaCl2, 10 mM MgCl2, 50 mM NaCl) on a glass slide and immediately covered with a coverslip, and then imaged. Quantification of fluorescence signals was done as follows. The integrated fluorescence intensity of polar clusters and of a similar cytoplasmic region was measured using the region measurement tool in Metamorph 7.7. The intensity of the cytoplasmic region was subtracted from the intensity of the polar cluster. These corrected intensities of the polar clusters were used to calculate the ratios between the polar signals in individual cells. If the ratio is ≤2.0, the localization is defined as bipolar symmetric, if the ratio is ≥2.1 and ≤10.0 the localization is defined as bipolar asymmetric, and if the ratio was ≥10.1 the localization is defined as unipolar. For each strain 200 cells were analyzed. For time-lapse microscopy, cells were recorded at 30-s intervals for 15 min. Images were recorded and processed with Leica FW4000 V1.2.1 or Image Pro 6.2 (MediaCybernetics) software. Processed images were visualized using Metamorph (Molecular Devices). Reversals were counted for >50 cells of each strain followed for 15 minutes and displayed in a Box plot.
Proteins were purified as described in Text S1. 0.5 mg of purified His6-MglB or MglA-His6 in buffer H (50 mM NaH2PO4 pH 8.0, 300 mM NaCl, 10 mM imidazole) was applied to a Ni2+-NTA-agarose column (Macherey-Nagel). M. xanthus cell lysate was prepared as follows: 200 ml of exponentially growing WT cells at a cell density of 7×108 cells/ml were harvested, resuspended in buffer H in the presence of proteases inhibitors (Roche) and lysed by sonication. Cell debris was removed by centrifugation at 4700×g for 20 min, 4°C and the cell-free supernatant applied to the Ni2+-NTA-agarose column with or without bound His6-MglB or MglA-His6. After two washing steps with each 10 column volumes of the buffer H, bound proteins were eluted with buffer H supplemented with 250 mM imidazole. Proteins eluted from the columns were analyzed by two methods: SDS-PAGE and gels stained with Coomassie Brilliant Blue R-250 and SDS-PAGE with immunoblot analysis using α-RomR antibodies [25].
To test for direct protein-protein interactions, 0.2 mg of purified prey protein (His6-RomR or His6-MglB or as a negative control His6-PilP) was mixed with 0.2 mg of purified bait protein (GST-MglA or MalE-RomR) and as a control with 0.2 mg of GST or MalE, respectively. Proteins were incubated with 0.5 ml sepharose beads (for MalE-tagged proteins: amylose beads; for GST-tagged proteins: glutathione beads) in buffer D (50 mM NaH2PO4 pH 8.0, 300 mM NaCl) for 5 h, 4°C. After washing the beads with 25 column volumes of buffer D, the elutions were performed with buffer D supplemented with 10 mM glutathione for GST-tagged proteins, and with 10 mM maltose for MalE-tagged proteins. Proteins eluted from the columns were analyzed by immunoblot analysis using α-GST antibodies (Biolabs), α-MalE antibodies (Biolabs), α-RomR antibodies [25] and α-MglB antibodies [26]. Immunoblots were carried out as described [46].
The following software packages were used with the described settings unless otherwise specified. The HMMER3 software package [49] was used in conjunction with the Pfam26 domain library [50] for domain architecture analysis with default gathering thresholds. In the event of domain overlaps, the highest scoring domain model was chosen for the final architecture. The JackHMMER method [51] was used for iterative similarity searches with a 0.0001 e-value inclusion threshold. For non-iterative similarity searches, we used BLASTP from the BLAST+ software package version 2.2.26 [52] and considered hits with e-values of 0.0001 or lower to be significant unless otherwise specified. Multiple sequence alignments were built using the l-ins-i algorithm of the MAFFT version 6.864b software package [53]. Phylogenetic trees were constructed using FastTree version 2.1.4 [54] with default settings or PhyML version 3.0 [55] with empirical frequencies and SPR topology searches. Secondary structure was predicted using the Jpred3 webserver [56].
All complete prokaryotic genomes 1609 were downloaded from the NCBI Refseq [57] database on April 4th, 2012. Due to our specific interest in Myxococcales, we also included the complete genomes of Stigmatella aurantiaca [58] and Corallococcus coralloides [59] from GenBank [60] as they were not yet available in Refseq at the time of genome collection.
The MglA and MglB sequences from M. xanthus (MXAN_1925 and MXAN_1926, respectively) were used in BLASTP queries against the genome set. All significant sequence hits were aligned using MAFFT and the core regions were extracted and used to build phylogenetic trees with FastTree. The tree representing 134 putative MglA homologs showed a distinct subfamily of 113 sequences that is associated with the characterisic intrinsic arginine finger [27] in comparison to a subfamily of 21 other putative small GTPases that lack it (Figure S3). We chose the subfamily of 113 sequences as our MglA set. In contrast, only 63 putative MglB homologs were collected by BLAST analysis, most of which are encoded near members of the MglA set. We used the core regions of the MglB homologs as BLASTP queries to identify more putative MglB partners of our MglA set. The collected sequences were aligned using MAFFT and the core regions were extracted and used to build a phylogenetic tree with FastTree (Figure S4). The tree of 86 putative MglB homologs revealed a subfamily of 71 sequences that were associated with our MglA sequence set based on genome context, and the members of this subfamily were chosen as our final MglB set. MglA and MglB sequences are listed in Table S3.
Initial BLASTP queries with the RomR sequence from M. xanthus (MXAN_4461) revealed it to be a multi-domain protein with two regions of conservation, an N-terminal receiver domain and a C-terminal domain that is not homologous to previously characterized domains. Given the ubiquity of receiver domains, we chose to use the C-terminal domain (369–420 of MXAN_4461) in a jackHMMER query against our genome set, which converged after three rounds. The results identified 28 significant hits, 27 of which have N-terminal receiver domains typical of response regulators. We extracted the receiver domain and C-terminal domains of the 27 response regulators sequences (Table S3) and used them as BLASTP queries against our database to identify potential divergent homologs. The queries with the C-terminal regions did not identify any new homologs, whereas the queries with the receiver domains identified 3599 homologs using our default gathering thresholds. Given this large data set, we chose to only gather hits of 1e-20 or lower from the BLASTP queries as this resulted in a set of only 133 sequences, which was more comparable to our previously defined MglA and MglB data sets. The 133 sequences were aligned using the e-ins-i algorithm of MAFFT. We used FastTree to build a phylogenetic tree from the receiver domain regions of the sequences because the remaining portions of the sequences could not be aligned. The resulting tree revealed a subfamily of 31 sequences most of which contain the previously defined C-terminal domain (Figure S5). Those lacking the domain were encoded in genomes from species closely related to their most similar sequence (e.g. two RomR sequences in members of Acidobacteria that lack the C-terminal domain group with a complete RomR sequence from another member of Acidobacteria), which supports their classification as RomR sequences. We chose these 31 sequences for our final RomR set. RomR sequences are listed in Table S3.
The Frz system was previously identified as a member of the ACF class of chemosensory systems [61]. We collected the core regions of all the CheA sequences from those analyses and built multiple sequence alignments for each class using MAFFT. Hidden markov models (HMMs) were built from each class specific alignment after being reduced such that no members of the alignment shared more than 80% identity. CheA sequences can be identified by the presence of HATPase_c and CheW domains from Pfam [62], and all sequences with HATPase_c and CheW domains were collected from our genome set. The sequences were compared to our CheA HMM library and assigned to classes based on the highest scoring model. All CheA sequences assigned to the ACF class were collected (164 sequences) and aligned using the e-ins-i algorithm of MAFFT. The core regions corresponding to the P3–P5 domains and the C-terminal receiver domain characteristic of this family were used to build a phylogenetic tree in PhyML. Sequences lacking any of these four domains or the N-terminal histidine phosphotransfer domain were predicted to be non-functional and removed from the analysis. We identified a FrzE specific subfamily in the tree based on Frz system features, genome context, and paralogy events (Figure S6). All FrzE sequences have a FrzZ encoded in nearby genes based on BLASTP queries using neighboring response regulator protein sequences. FrzE sequences are listed in Table S3.
Recent computational analysis of FAC proteins identified two distinct groups of genes: Group A genes that are only present in organisms that have gliding motility (members of Myxococcales and Bdellovibrionales), and Group B genes that have homologs in the Group A lineages in addition to Fibrobacter succinogenes and members of β/γ-proteobacteria for which gliding motility has not been observed [24]. We chose the Group A gene gltF as a marker for the presence of gliding motility because it is the most unique based on initial BLAST searches (many Group A genes are putative outer membranes proteins or proteins that contain TPR repeats, both of which result in non-specific BLAST hits). We used the MXAN_4868 GltF sequence as a query in a JackHMMER search, which identified 29 homologs that were present in all Myxococcales and Bdellovibrionales genomes consistent with previous observations [24]. All identified GltF sequences are listed in Table S3.
We used the retraction ATPase PilT as a marker for the presence of T4P. The PilT sequences from M. xanthus [63], Neisseria meningitidis [64], Pseudomonas aeruginosa [65], and Synechocystis sp. PCC6803 [66] share the same Pfam domain architecture, a single T2SE domain. We collected 3756 sequences from our genome set that matched this domain architecture, aligned them in MAFFT using default settings, and a phylogenetic tree was built from the alignment using FastTree (Figure S7A). This sequence set is expected to also include sequences for PilB and ATPases in type II secretion systems. To identify the branches corresponding to PilT, the PilT sequences from the four aforementioned organisms were used to identify a smaller set of 1219 PilT candidates. The 1219 sequences were realigned in MAFFT using default regions and the core region of the alignment corresponding to residues 5–327 of the M. xanthus PilT (MXAN_5787) was extracted and used to build a phylogenetic tree in FastTree (Figure S7B). Identification of characterized PilT proteins in this tree was used to identify a set of 547 PilT sequences (Table S3).
|
10.1371/journal.ppat.1000868 | Increased Asymmetric Dimethylarginine in Severe Falciparum Malaria: Association with Impaired Nitric Oxide Bioavailability and Fatal Outcome | Asymmetrical dimethylarginine (ADMA), an endogenous inhibitor of nitric oxide synthase (NOS), is a predictor of mortality in critical illness. Severe malaria (SM) is associated with decreased NO bioavailability, but the contribution of ADMA to the pathogenesis of impaired NO bioavailability and adverse outcomes in malaria is unknown. In adults with and without falciparum malaria, we tested the hypotheses that plasma ADMA would be: 1) increased in proportion to disease severity, 2) associated with impaired vascular and pulmonary NO bioavailability and 3) independently associated with increased mortality. We assessed plasma dimethylarginines, exhaled NO concentrations and endothelial function in 49 patients with SM, 78 with moderately severe malaria (MSM) and 19 healthy controls (HC). Repeat ADMA and endothelial function measurements were performed in patients with SM. Multivariable regression was used to assess the effect of ADMA on mortality and NO bioavailability. Plasma ADMA was increased in SM patients (0.85 µM; 95% CI 0.74–0.96) compared to those with MSM (0.54 µM; 95%CI 0.5–0.56) and HCs (0.64 µM; 95%CI 0.58–0.70; p<0.001). ADMA was an independent predictor of mortality in SM patients with each micromolar elevation increasing the odds of death 18 fold (95% CI 2.0–181; p = 0.01). ADMA was independently associated with decreased exhaled NO (rs = −0.31) and endothelial function (rs = −0.32) in all malaria patients, and with reduced exhaled NO (rs = −0.72) in those with SM. ADMA is increased in SM and associated with decreased vascular and pulmonary NO bioavailability. Inhibition of NOS by ADMA may contribute to increased mortality in severe malaria.
| Severe falciparum malaria is associated with impaired microvascular perfusion, lung injury and decreased bioavailability of nitric oxide (NO), but the causes of these processes are not fully understood. Asymmetrical dimethylarginine (ADMA), a competitive endogenous inhibitor of nitric oxide synthase (NOS), is an independent predictor of mortality in other critical illnesses, and can impair vascular function in chronic disease. ADMA can be produced by both the host and malaria parasites. The major novel findings of this study in malaria are that ADMA is an independent predictor of death in falciparum malaria, and is associated with decreased availability of nitric oxide in at least two organ systems affected by malaria parasites, the lining of blood vessels and the lungs. This study contributes to knowledge of regulation and availability of pulmonary and endothelial NO in critical illness and identifies pathogenic processes which may contribute to death in severe malaria. Therapies which increase the availability of NO or which reduce ADMA levels may have potential for adjunctive therapy of severe malaria.
| Plasmodium falciparum causes ∼1 million deaths annually [1], [2]. Despite rapid parasite clearance with the anti-parasitic drug artesunate, the mortality rate in severe malaria remains high [3], [4]. Endothelial activation, parasite sequestration, impaired microvascular perfusion and dysregulated inflammatory responses are all thought to contribute to severe and fatal malaria [5]–[9]. Increased understanding of these pathogenic mechanisms may identify targets for adjunctive therapies to further improve outcomes.
Severe malaria is associated with impaired nitric oxide (NO) bioavailability and blood mononuclear cell NO synthase (NOS) type 2 expression in both children [10], [11] and adults [6]. The concentrations of L-arginine, the substrate for NO production by all three NOS isoforms [12], are low in children and adults with severe malaria and likely contribute to the decreased NO production found in severe disease [6], [10], [13]. However, in adults with moderately severe malaria, L-arginine concentrations are at least as low as those seen with severe malaria, yet there is no impairment of vascular and pulmonary NO bioavailability as found in severe disease [6]. This suggests that factors other than substrate limitation contribute to impaired NO bioavailability in severe malaria.
Asymmetrical dimethylarginine (ADMA) is a non-specific endogenous NOS inhibitor which decreases vascular function in cardiovascular and renal disease [14], [15]. Protein-arginine-methyltransferases methylate arginine residues in proteins and ADMA is released when these proteins undergo degradation [14]. ADMA is primarily eliminated by the enzyme dimethylarginine-dimethylaminohydrolase-1 (DDAH-1) in the liver and kidney, with ∼20% being excreted in the urine [16]. In adult sepsis, elevated ADMA is independently associated with increased mortality, a likely consequence of non-specific inhibition of homeostatic NO production [17], [18]. Increased protein catabolism with hepatic and renal dysfunction in severe malaria has the potential to increase ADMA and impair NO production, but the importance of ADMA in the pathogenesis of malaria is currently unknown. Clarification of the role of ADMA in malaria is of particular importance given a recent genome-wide association study in children linking DDAH polymorphisms with risk of severe malaria [19], and the potential for the parasite as well as host to produce ADMA [20].
Acute lung injury is a common but little-studied complication of severe falciparum malaria associated with high mortality [21]. In sepsis and critical illness, acute lung injury and mortality are associated with decreased total and pulmonary NO [22], [23]. Pulmonary diffusion capacity and exhaled NO concentrations are both reduced in severe malaria [6], [24], however the causative factors have not been identified. The role of ADMA in impairing pulmonary NO bioavailability in severe malaria, or indeed any critical illness, is not known.
In a prospective longitudinal study of Indonesian adults with malaria, we evaluated the hypotheses that concentrations of methylated arginines are independently associated with a) disease severity, b) reduced exhaled NO and vascular NO bioavailability and c) increased mortality.
We measured asymmetrical dimethylarginine (ADMA), symmetrical dimethylarginine (SDMA) and L-arginine in 49 patients with severe malaria, 78 with moderately severe malaria and 19 healthy controls. In the SM patients, 20 patients had only one criterion for severe disease (coma in 14, hyperbilirubinemia in 4, respiratory distress in 2) with the remainder having >1 criteria. In total, 34 of the patients with SM were treated with intravenous artesunate and the remaining 15 received intravenous quinine [6]. All of the 78 MSM patients were treated with quinine with the exception of one who received artesunate. Exhaled NO concentrations (FeNO) could not be measured in those with coma, but were possible in 48% (11/23) of non-comatose SM patients, 88% (69/78) of MSM patients and 100% (19/19) of HCs. RH-PAT index was measured in all patients with malaria as well as HC. There were eight deaths among the patients with SM, and none in the MSM patients. Repeat RH-PAT and venous blood measurements were only performed in one and four of the eight fatal cases respectively. Baseline characteristics of study participants are summarized in Table 1.
ADMA and SMDA concentrations were increased in SM patients (0.85 µM; 95% CI 0.74–0.96 and 1.67 µM; 95% CI 1.24–2.09 respectively) compared to those with MSM (0.54 µM; 95%CI 0.5–0.56 and 0.58 µM; 95%CI 0.54–0.63) and HC (0.64 µM; 95%CI 0.58–0.70 and 0.53 µM; 95%CI 0.47–0.59); ANOVA p<0.001 for both ADMA and SDMA, Table 2, Figure 1A and 1B. L-arginine concentrations were significantly higher in HCs compared to patients with MSM and SM, with the L-arginine/ADMA ratio decreasing with increasing disease severity (p<0.001; Table 2).
Exhaled nitric oxide concentration (FeNO) was significantly decreased in SM patients compared to MSM and HC; (Table 2). FeNO was inversely correlated with plasma ADMA concentration (rs = −0.31, p = 0.003; Table 3) in all malaria patients, and in the subgroup of 11 patients with SM (rs = −0.72, p = 0.01; Table 3). FeNO also correlated inversely with HRP2 concentration (rs = −0.51, p<0.001) in malaria patients, but not in the SM group. FeNO remained inversely associated with ADMA, after adjusting for disease severity, creatinine and HRP2. FeNO was not associated with SDMA, the L-arginine/ADMA ratio, plasma hemoglobin or arginase.
As reported previously, the RH-PAT index was significantly lower in SM patients compared to those with MSM and controls (p<0.001) Table 2. In all malaria patients, there were moderate inverse associations between RH-PAT index and ADMA (rs = −0.32; p<0.001; Table 3) and SDMA (rs = −0.35; p<0.001) concentrations. After adjusting for factors previously shown to be associated with RH-PAT index including plasma hemoglobin [6], [7], [25], the inverse association with ADMA remained significant, with the final model including both ADMA and cell free hemoglobin (r = 0.40). In contrast, the L-arginine/ADMA ratio was not associated with the RH-PAT index. Longitudinally there was no association between the RH-PAT index and ADMA concentration or L-arginine/ADMA ratio in SM patients.
Patients with SM had significantly elevated plasma concentrations of Ang-2, ICAM-1 and E-selectin compared to those with MSM and HC; Table 2. Angiopoietin-2 and ICAM-1 were significantly correlated with both ADMA (rs = 0.48 and 0.42 respectively; p<0.001; Table 3), and SDMA (rs = 0.54 and 0.52; p<0.001), and this was also apparent in the subgroup of SM patients; Table 3. ADMA remained independently associated with Ang-2 and ICAM-1 after adjusting for confounding factors, including creatinine, plasma hemoglobin, parasite biomass and disease severity.
There was no significant correlation between ADMA or SDMA with E-selectin, and none between the L-arginine/ADMA ratio and Ang-2, ICAM-1 and E-selectin.
The plasma creatinine, total bilirubin, P. falciparum histidine rich protein 2 (HRP2) and venous lactate were increased in SM compared to MSM (Table 2). In all patients with malaria, there were correlations between ADMA and SDMA with creatinine (rs = 0.45; p<0.001; rs = 0.69; p<0.001; Table 3), total bilirubin (rs = 0.32; p<0.001; rs = 0.36; p<0.001; Table 3), HRP2 (rs = 0.46; p<0.001; rs = 0.62; p<0.001; Table 3) and lactate (rs = 0.3; p = 0.01; rs = 0.29, p = 0.02; Table 3). The associations remained significant in SM patients for each of these biomarkers of disease severity except for venous lactate (Table 3). There was no association between L-arginine/ADMA ratio and any biomarker. TNF was only measured in the SM patients, and was associated with SDMA (rs = 0.52; p = 0.001), but not with ADMA (p = 0.06).
In SM patients, ADMA and SDMA concentrations were significantly higher in the 8 patients who died (1.28 µM; 95%CI 0.88–1.74 and 3.76 µM; 95%CI 1.88–5.56, respectively) compared to the 41 survivors (0.77 µM; 95%CI 0.64–0.84 and 1.27 µM; 95%CI 0.99–1.54, respectively; p<0.001; Figure 1A and 1B).
Each micromolar increase in ADMA and SDMA concentrations was associated with an 18-fold (OR 18.8 95% CI 2.0–181; p = 0.01) and three-fold (OR 3.0; 95% CI 1.5–6.2; p = 0.002) increased risk of death, respectively. ADMA but not SDMA remained a significant risk factor for death after adjusting for other confounding factors, such as Ang-2, creatinine, parasite biomass, bilirubin, base deficit and lactate. A final model predicting a fatal outcome included ADMA, Ang-2, HRP2 and creatinine (Table 4). The L-arginine/ADMA ratio was not associated with risk of death.
The prognostic value of ADMA in predicting a fatal outcome was measured by the area under the receiver operating curve (ROC). ADMA (AUROC 0.85; 95% CI 0.71–0.99; Figure 2) was comparable to other reliable prognostic indicators, including Ang-2 (AUROC 0.84; 95% CI 0.71–0.96), HRP2 (AUROC 0.86; 95% CI 0.73–0.94), base deficit (AUROC 0.73; 95% CI 0.53–0.92), and TNF (AUROC 0.71; 95% CI 0.43–0.98), and a better predictor of fatal outcome than venous blood lactate (AUROC 0.63; 95%CI 0.41–0.83; p = 0.003).
In patients with severe malaria, there was no significant change in ADMA (Figure 3A) or SDMA concentrations during the course of hospitalization among the overall group, survivors or those with a fatal outcome. Among survivors, there was a daily increase in L-arginine/ADMA (β = 9.1, p<0.001; Figure 3B) but no increase in those who died.
ADMA is increased in severe falciparum malaria and is an independent predictor of mortality. Indeed ADMA was a better predictor of death than blood lactate, previously shown to be a reliable prognostic indicator of increased mortality in severe malaria [26]. Our study demonstrated that elevated plasma ADMA concentrations are independently associated with decreased exhaled NO concentrations, impaired vascular NO bioavailability, increased endothelial activation and parasite biomass. To our knowledge this is the first demonstration of a relationship between increased ADMA and impaired exhaled NO in any critical illness. Taken together, these findings suggest that ADMA, an endogenous inhibitor of all three nitric oxide synthase (NOS) isoforms, reduces NO bioavailability in at least two organ systems and may contribute to increased mortality in falciparum malaria.
In critically ill patients, elevated ADMA concentrations are likely to result from increased production and reduced elimination. The elevation in both ADMA and SDMA may result from increased host production of methylated arginines in severe malaria. The majority of circulating ADMA is taken up by the liver before being metabolized by dimethylarginine-dimethylaminohydrolase-1 (DDAH-1); approximately 20% is excreted unchanged in the urine [16]. Hepatic blood flow is known to be significantly impaired in severe malaria [27]. The correlation of bilirubin and creatinine with increased ADMA, suggests that similar to sepsis [28], decreased hepatic and renal elimination may also increase ADMA concentrations in severe malaria. The large parasite biomass in severe malaria may also be a potential source of ADMA, with Plasmodium falciparum possessing protein arginine methyltransferases capable of producing ADMA [20]. The significant independent correlation between parasite biomass and ADMA on admission suggests this may be occurring in vivo, although the persistently elevated levels in severe malaria after commencement of anti-malarial therapy suggest the importance of altered host production and clearance in the post-treatment period. Increased clearance due to increases in either hepatic blood flow or DDAH activity may explain the decreased ADMA concentrations in patients with moderately severe malaria. There are no clinical studies to date documenting increased DDAH activity in mild inflammatory diseases, but hepatic blood flow is known to be significantly increased in acute uncomplicated falciparum malaria compared to patients with severe disease and healthy individuals [27]. The converse may also be true, with lower plasma ADMA concentrations potentially increasing vascular NO bioavailability in moderately severe malaria [14], and possibly contributing to elevated hepatic blood flow.
The loss of DDAH-1 function in a murine model of sepsis increased ADMA, reduced NO signaling, and worsened vascular pathophysiology including endothelial function [29]. In human severe sepsis, a polymorphism in the DDAH-2 enzyme increased ADMA levels which were associated with increased severity of organ failure and early septic shock [30]. Recently, a genome-wide association study in children found that a polymorphism in the gene encoding DDAH-1 was associated with an increased likelihood of severe malaria [19]. These studies indicate that altered DDAH function may be a contributor to organ damage and increased mortality in severe malaria as well as in other critical illnesses.
SDMA does not inhibit NOS, but competes with plasma L-arginine for intracellular uptake by the cationic amino acid transporters (CAT). Unlike ADMA, it is not metabolized by DDAH and is almost exclusively eliminated by the kidneys [16]. In chronic disease SDMA has recently been shown to be an independent predictor for major cardiovascular events in certain chronic diseases [31]. We find that in malaria, SDMA concentrations are associated with mortality and decreased vascular bioavailability on univariate analysis, but not after adjusting for renal function. While the association between SDMA and disease severity is likely to reflect the degree of renal impairment and SDMA retention, it is possible that retained SDMA may also contribute to decreased NO bioavailability in severe malaria.
In critically ill adults with organ failure and severe sepsis, ADMA concentrations are associated with increased all-cause mortality and the severity of organ failure [17], [30]. Investigators have hypothesized that this may result from non-selective inhibition by ADMA of all three isoforms of NOS, particularly homeostatic NOS3 (endothelial NOS) [18]. This is similar to the postulated mechanism to explain the increased mortality with use of NG-monomethyl-arginine (NMMA), another non-specific NOS inhibitor, in a phase 3 clinical trial of sepsis [32]. In falciparum malaria, systemic NO production is impaired in severe disease and hypoargininemia is likely to be a contributing cause [6], [10], [11], [13]. Exhaled and vascular NO are both reduced in adults with severe malaria [6], but not in moderately severe malaria (MSM) despite similar degrees of hypoargininemia [6]. This may be explained by the higher ADMA in severe malaria and a greater competitive inhibition of NOS in SM compared to MSM, similar to clinical studies of healthy volunteers in which ADMA infusion reduced blood flow [15], [33]. In mouse studies, ADMA infusion alone reduces splenic blood perfusion, but when combined with hypoargininemia, causes a reduction in renal, hepatic and splenic blood flow with organ damage [34]. Regulation of microcirculatory flow is dependent on pre-capillary arteriolar vasodilatory responses which in turn are critically dependent on NO production [35], with both likely to be decreased by ADMA in SM. By decreasing functional capillary density, ADMA could further impair microcirculatory function already compromised by parasite sequestration in capillaries and post-capillary venules [36].
We have previously shown that hemolysis-related NO quenching by cell-free hemoglobin is associated with reduced vascular NO bioavailability in severe malaria [25]. In malaria, increased ADMA and cell-free hemoglobin were independently related to endothelial dysfunction, suggesting that inhibition of NOS and NO quenching both reduce vascular NO bioavailability.
NO has multiple regulatory functions that maintain endothelial quiescence in vitro, including inhibition of endothelial Weibel-Palade body (WPB) exocytosis and ICAM-1 expression [37], [38]. Plasma concentrations of angiopoietin-2 (Ang-2), an angiogenic factor stored in WPBs, predict increased mortality in malaria [7], and ICAM-1 is a major endothelial adhesion receptor mediating cytoadherence of parasitized red cells and microvascular sequestration [5]. We demonstrate that ADMA levels correlate with increased Ang-2, but the association between ADMA and increased mortality is independent of Ang-2, suggesting effects of NO inhibition in addition to increased WPB exocytosis.
Acute lung injury is a complication of severe malaria in adults associated with a high mortality rate [21], [24]. Gas transfer at the alveolar-capillary membrane and exhaled NO are both decreased in severe falciparum malaria [6], [24]. In an animal model of sepsis–associated pulmonary injury, non-selective NOS inhibition causes increased lung edema [39]. Clinical studies have shown decreased pulmonary NO concentrations in patients with acute respiratory distress syndrome, as well as an association between decreased NO production and a worse outcome in acute lung injury [22], [23]. The lung is a major source of ADMA and increased concentrations are associated with pulmonary arterial hypertension [40], [41]. In severe malaria, both ADMA and parasite biomass are strongly inversely associated with exhaled NO concentrations, suggesting that both factors impair pulmonary NO production in severe disease.
There are several limitations in our study. Measurement of exhaled NO was not possible in patients with coma and was possible in only half of severe malaria patients without coma. Our results may therefore not reflect the relationship between ADMA and exhaled NO in all syndromes of severe malaria. In patients who died, only 4 of 8 had at least one repeat blood sample, and the longitudinal data may not truly reflect the course of the methylated arginines in fatal cases. RH-PAT index is at least 50% dependent on endothelial NO release [42], but we cannot exclude an effect of ADMA on other vasodilators such as prostacyclin and endothelium-derived hyperpolarizing factor. Although we have measured plasma ADMA concentrations, the effects of ADMA are intracellular. Nevertheless, in vitro studies with endothelial cells have shown that increasing extracellular ADMA results in five-fold increases in intracellular concentrations. This suggest that intracellular concentrations of ADMA in severe malaria may be higher, and may be adequate for meaningful inhibition of for all three NOS isoforms (IC50s ∼2-5µM) [43]. The observational nature of the study does not allow us to conclude with certainty a direct role for ADMA in the pathophysiology of severe malaria. While the association of ADMA with mortality may reflect impaired renal and hepatic function, it remained significant after adjusting for these factors in a multivariable model. Furthermore, increased ADMA from impaired hepatic and/or renal clearance is not just a marker of organ dysfunction in critical illness, with retained ADMA having functional consequences on NOS activity.
In summary, the endogenous non-selective NOS inhibitor ADMA is elevated in SM and is an independent risk factor for mortality. ADMA is also associated with decreased FeNO and vascular NO bioavailability, as well as increased endothelial activation and parasite biomass. Therapies which increase NO bioavailability or which diminish ADMA levels represent rational approaches for interventional trials of adjunctive therapy in severe malaria.
The study was conducted at Mitra Masyarakat Hospital, Timika, Papua, Indonesia, a region with unstable transmission of multidrug resistant malaria [44], [45]. Written informed consent was obtained from all patients, if they were comatose or too ill, consent was obtained from relatives. The Ethics Committees of the National Institute of Health Research and Development, Indonesia, and Menzies School of Health Research, Australia approved the study.
Patients were ≥18 years old with moderately-severe (MSM) or severe (SM) Plasmodium falciparum malaria without P. vivax infection and with a hemoglobin level >60 g/L who had been prospectively enrolled in a study of endothelial dysfunction and exhaled NO [6]. Previous results from this study group have been published [6], [7], [13], [25]. Briefly, SM was defined as P. falciparum parasitemia and ≥1 modified WHO criterion of severity (excluding severe anemia). MSM was defined as fever within the preceding 48 hours, >1,000 asexual P. falciparum parasites/µL, no WHO warning signs or severe malaria criteria and a requirement for inpatient parenteral therapy because of inability to tolerate oral treatment. Healthy controls (HC) were non-related hospital visitors with no history of fever in last 48 hours, intercurrent illness or smoking in last 12 hours, or evidence of parasitemia [6]. Standardized history and physical examination were documented. Heparinized blood was collected daily, centrifuged within 30 minutes of collection and plasma stored at −70°C. Parasite counts were determined by microscopy. Hemoglobin, biochemistry, acid-base parameters and lactate were measured with a bedside analyser (i-STAT Corp). Patients were treated with anti-malarials and antibiotics using standard national protocols as previously described [6].
Solid phase extraction (SPE) of amino acids was followed by derivatisation with Accq-Fluor and separation on a Gemini-NX column at pH 9 [46]. The SPE method gives absolute recoveries of >80% for ADMA and symmetrical dimethylarginine (SDMA) and average relative recoveries of 102% for ADMA and 101% for SDMA. The HPLC method gives intra-assay RSDs of 2.1% and 2.3% and inter-assay RSDs of 2.7% and 3.1% for ADMA and SDMA respectively [46].
Plasma concentrations of cell-free hemoglobin and the endothelial activation markers, ICAM-1, E-selectin and angiopoietin-2 were measured by ELISA as previously reported in this population [6], [7], [25]. Total parasite biomass was quantified by measuring plasma histidine rich protein-2 (HRP2) by ELISA [6], plasma arginase by a radiometric method [6] and plasma TNF concentrations by flow cytometry, as previously reported [7].
Endothelial function was measured non-invasively using peripheral arterial tonometry (EndoPAT) by the change in digital pulse wave amplitude in response to reactive hyperemia, giving a RH-PAT Index as reported previously [6]. The RH-PAT index is at least 50% dependent on endothelial NO production [42]. Endothelial function was measured daily until death or discharge, or until the RH-PAT index was above an a priori cutoff (1.67) for two consecutive days [13]. Fractional concentrations of exhaled NO were measured by NO analyser (Aerocrine), as previously described, using American Thoracic Society guidelines and a flow rate of 250 ml/sec [6].
Statistical analysis was performed with STATA 9.2 software. Intergroup differences were compared by ANOVA or Kruskal-Wallis test, where appropriate. Pearson's or Spearman's correlation coefficients were determined depending on normality of distributions. Multiple stepwise linear regression was used to adjust for confounding variables. Longitudinal associations were assessed by mixed effects modeling using generalized estimating equations. Logistic regression was used to determine the association between death and ADMA concentrations. Variables hypothesized, as well as those shown in previous publications [6], [7], [25], to contribute to mortality, pulmonary NO and endothelial pathology were included in a multiple regression model if p<0.05 on univariate analysis and retained if they remained significant. Goodness-of-fit was assessed by the Hosmer-Lemeshow goodness of fit test and independent variables tested for interactions. The prognostic utility of continuous variables was measured using the area under the receiver operating curves (ROCs) and its 95% confidence intervals were calculated. A two-sided value of p<0.05 was considered significant.
|
10.1371/journal.pcbi.1006882 | ZnT2 is an electroneutral proton-coupled vesicular antiporter displaying an apparent stoichiometry of two protons per zinc ion | Zinc is a vital trace element crucial for the proper function of some 3,000 cellular proteins. Specifically, zinc is essential for key physiological processes including nucleic acid metabolism, regulation of gene expression, signal transduction, cell division, immune- and nervous system functions, wound healing, and apoptosis. Consequently, impairment of zinc homeostasis disrupts key cellular functions resulting in various human pathologies. Mammalian zinc transport proceeds via two transporter families ZnT and ZIP. However, the detailed mechanism of action of ZnT2, which is responsible for vesicular zinc accumulation and zinc secretion into breast milk during lactation, is currently unknown. Moreover, although the putative coupling of zinc transport to the proton gradient in acidic vesicles has been suggested, it has not been conclusively established. Herein we modeled the mechanism of action of ZnT2 and demonstrated both computationally and experimentally, using functional zinc transport assays, that ZnT2 is indeed a proton-coupled zinc antiporter. Bafilomycin A1, a specific inhibitor of vacuolar-type proton ATPase (V-ATPase) which alkalizes acidic vesicles, abolished ZnT2-dependent zinc transport into intracellular vesicles. Moreover, using LysoTracker Red and Lyso-pHluorin, we further showed that upon transient ZnT2 overexpression in intracellular vesicles and addition of exogenous zinc, the vesicular pH underwent alkalization, presumably due to a proton-zinc antiport; this phenomenon was reversed in the presence of TPEN, a specific zinc chelator. Finally, based on computational energy calculations, we propose that ZnT2 functions as an antiporter with a stoichiometry of 2H+/Zn2+ ion. Hence, ZnT2 is a proton motive force-driven, electroneutral vesicular zinc exchanger, concentrating zinc in acidic vesicles on the expense of proton extrusion to the cytoplasm.
| Herein we explored the mechanism of action of the human ZnT2 zinc transporter. ZnT2 is essential for zinc accumulation in breast milk and is therefore of paramount medical significance. Expanding on our previous study, we herein present energy calculations suggesting that ZnT2 functions as a proton/zinc antiporter. Our calculations consist of electrostatic and pKa calculations as well as zinc binding free-energy curves. Upon integration of our calculation results, we conclude that ZnT2 functions as an antiporter with a 2H+/Zn2+ stoichiometry, construct a Monte Carlo model to test this mode of ZnT2 transport activity, and validate our computational results experimentally using live human breast epithelial cells. These functional experiments reveal that ZnT2 cannot function in the absence of protons suggesting that it operates as a substrate-induced alternating-access transporter, displaying an apparent 2H+/Zn2+ stoichiometry.
| Divalent zinc ions are key and integral components of a multitude of proteins involved in a plethora of essential physiological processes including metabolism of nucleic acids, regulation of gene expression, signal transduction, cell division, immune- and nervous-system functions, wound healing, as well as apoptosis [1–4]. As such, it is crucial for cells to be efficiently shuttle zinc to- and from their surroundings, between organelles and distinct compartments. Mammalian zinc transporters belong to two families, ZnT and ZIP, canonically exporting and importing zinc, respectively. Specifically, ZnT2 plays a crucial role in concentrating zinc within secretory vesicles which were suggested to release zinc into breast milk during lactation [5]. In fact, exclusively breastfed infants nursed by mothers harboring loss-of-function mutations in ZnT2 suffer from transient neonatal zinc deficiency (TNZD), which leads to severe zinc deficiency in these infants (e.g. [6–12]). Furthermore, ZnT2 was shown to have a critical role in the development and normal function of the mouse mammary gland [13], as well as in involution [14,15]. Taken together, expanding our mechanistic understanding of ZnT2 function, will provide valuable information with important physiological and possible therapeutic implications.
ZnT2 was first identified by Palmiter et al., in 1996 [16], and was the second zinc transporter to be described in the literature. Unlike ZnT1 which is localized in the plasma membrane, ZnT2 is localized in intracellular vesicles where it plays a crucial role in zinc sequestration, hence conferring upon cells protection against zinc toxicity [16]. However, the driving force for ZnT2-dependent concentration of zinc inside secretory vesicles is currently unsettled. The first efforts to characterize the driving force of active zinc transport by ZnT2 made use of Bafilomycin A1 (BafA1), a specific inhibitor of the vacuolar-type proton ATPase (V-ATPase) which abolishes vesicular acidification [16]. Under this experimental setting, vesicular zinc accumulation was not disrupted by BafA1, leading to the conclusion that ZnT2 function is not dependent on a proton gradient [16]. However, it was later shown that YiiP, the bacterial homolog of the ZnT family, is a H+/Zn2+ exchanger displaying a 1:1 stoichiometry [17]. Furthermore, Ohana et al., proposed that the zinc transport function of ZnT5 proceeds via a H+/Zn2+ exchange [18] and a similar mechanism was suggested for ZnT1 [19]. Since the subcellular localization of ZnT2 maps to acidic intracellular organelles [5,15,16], a putative H+/Zn2+ exchange mechanism is plausible for ZnT2 as well [5,20]. Thus, the current study was undertaken to establish the driving force of ZnT2 and assess its stoichiometry.
In a recent study, we have delineated the putative zinc permeation pathway through the human ZnT2 [21]. Using computational and evolutionary considerations, we determined the residues contributing to the zinc binding site, and modeled the inward- and outward-facing (cytoplasmic- and vesicular lumen-facing, respectively) conformations of ZnT2. In the current study, we expanded our previous findings and computed pKa values for the titratable residues at the putative binding site of ZnT2 (see Fig 1) and calculated zinc binding free-energies. Our computations were integrated into a free-energy landscape and we constructed a mechanistic model for the function of ZnT2. We tested our model using Monte Carlo (MC) simulations and surmised that ZnT2 is a proton-coupled zinc transporter. We then proceeded to functionally validate our hypothesis by determining changes in vesicular-zinc accumulation after treatment with BafA1, a specific inhibitor of V-ATPase which abolishes vesicular acidification, using FluoZin3-AM, a zinc-sensitive fluorescent probe in live human MCF-7 cells. In addition, we used pH-dependent fluorescent probes, to evaluate the impact of ZnT2 overexpression and zinc accumulation on the intravesicular pH. Finally, we demonstrated experimentally that ZnT2 is indeed proton-gradient-dependent, and we predicted the H+/Zn2+ stoichiometry of transport to be 2:1. Taken together, the current study highlights the synergistic strength of combining modeling and computational techniques with experimental functional validation. Our findings reveal an important aspect of ZnT2 and expand our understanding of this important transporter at the molecular level.
The structural model for the simulation systems of ZnT2 in the inward-facing (IF) and outward-facing (OF) conformations were based on our recent study [21], in which PDB IDs 3H90 (x-ray at 2.9Å resolution [22]) and 5VRF (cryo-EM at 4.1Å resolution [23]) were used as templates for the OF and IF conformations, respectively (see S1 Text for a brief summary of the modeling process). For each system, the transporter homodimer was inserted into a 30 Å-thick particle grid emulating a hydrophobic membrane using the MOLARIS software package [24,25]. The gap between the protomers was treated as part of the membrane milieu, as this gap exists within the expected membrane space and it is suggested to be non-functional [23]. Then, the systems were hydrated with a 40 Å-radius water sphere, using explicit 3-particle water molecules, and were submitted to energy minimization using the steepest descent algorithm followed by a short local relaxation simulation of 100 ps using the MOLARIS software package. Relaxation was performed in the presence of a zinc ion at the putative binding site (based on the structures of 3H90 and 5VRF), to prevent charge repulsion between the binding site residues.
Binding energy curves were derived from our previous study [21]. Briefly, the zinc ion was positioned along the permeation pathway at 1 Å intervals along the z-axis. For each such position, 10 PDLD/S-LRA calculations (see below) were performed entailing an additional short relaxation step of 10–20 ps, allowing the transporter to relax around the new position of the zinc ion, while the zinc ion can freely move in the xy-plane. Several such trajectories were computed and the average value for all points was finally used in this study (see our recent study [21] for more details and for the complete data).
The binding energy calculation for each position point was performed using the scaled semi-macroscopic Protein Dipoles Langevin Dipoles approach (PDLD/S) of MOLARIS [24]. Water in this method is represented semi-macroscopically by Langevin dipoles. The energy is the average of the charged and uncharged states, following the linear response approximation (LRA), scaled using a dielectric constant ε = 8 for the protein. Convergence was achieved by running molecular dynamics (MD) simulations for the relaxation and averaging the results of the different conformations [26]. The calculations were performed for several protonation states. See S1 Text for more information on the PDLD/S-LRA method.
pKa calculations were performed using the PDLD/S-LRA method, as described above, by computing the difference in free energy between the protonated and unprotonated states of each residue. The other residues were kept charged or uncharged as indicated in Table 1. When a zinc ion was present, we used the position of the zinc ion in the X-ray template of the structural models of ZnT2 (PDBID: 3H90).
To assess the relative free energy of each protonation state, we computed the total electrostatic free energy of the cluster, following the formalism presented previously [21,27,28]. The total energy of each state is given by the sum of: (i) the solvation energy of the ionized residues (representing the energy cost of bringing the ionized residues and the zinc ion from the bulk to the interior of the protein, relative to the system with zero charges); (ii) the energy of ionization of the given residues (His or Asp) in water, based on bulk pKa values; (iii) the electrostatic interaction between the ionized residues (the Coulombic energy of bringing the ionized residues in the bulk from an infinite distance to their distance in the binding site, using a dielectric constant of 80); and (iv) the electrostatic interactions between the ionized residues and the zinc ion. When a zinc ion was present, we used the position of the zinc ion in the X-ray template of the structural models of ZnT2 (PDBID: 3H90). See our recent study for the complete dataset [21] and see S1 Text for more details.
The zinc ion was simulated using a 5-particle entity arranged as a tetrahedral complex, based on ligand-field theory [29]. See our previous study for the detailed parameterization [21] and see S1 Text for a longer description.
The MC simulation system consisted of a simplified transporter where ions discretely move between pre-designated sites. Specifically, the system was composed of: Two bulk pools of protons and zinc ions, barrier sites on both sides of the transporter, and binding sites for protons and zinc at H106 and H223 (each able to bind one proton, and both able to bind one zinc together). Using this discrete binding sites scheme, each site was assigned an energy based on the total free energy and the zinc binding energy (see Results). Particles were randomly chosen and movements were attempted in random directions, applying the Metropolis acceptance criterion [30]. Prior to beginning the production simulations, we submitted long MC simulations using a flat potential to reveal the energetic contribution of the entropy of the system (e.g. two binding sites for protons compared to only one binding site for zinc). The entropic contributions were very small (<1 kcal/mol) and were subtracted from the input potential; this process was successfully applied in previous studies [31,32]. The simulations were run for 1010 steps, except for the temperatures of 400K and 425K which were run for 2.2×1010 steps. It should be noted here that these high temperatures do not reflect a transporter seemingly functioning at such high temperatures, but is rather applied in order to improve the MC convergence (by effectively scaling the energy landscape); since the system is simplified, no unfolding can occur, and without energy scaling no events occurred during the simulations within a reasonable amount of time (i.e. weeks). The pools of ions were equilibrated (by transferring particles between them) to prevent mass-action-like effects, and the number of ions transported between the pools was counted as flux. The source code for this MC simulation system can be found in https://github.com/ralhadeff/computational-biology/tree/master/reduced-transporter-Monte-Carlo. The values described in the results section are the total number of protons or zinc ions (in absolute numbers) that have been transported from one side of the membrane to the other, as flux (note that the direction of flow of zinc ions and protons are opposite); the ratio is the flux of protons divided by the flux zinc ions (in absolute values); and the efficiency is the number of zinc ions transported divided by the number of conformational changes that occurred throughout a given simulation, multiplied by 2 (because each transport cycle requires 2 conformational changes, e.g. IF to OF and back to IF).
The DNA dye Hoechst 33342 and the zinc chelator N,N,N′,N′-tetrakis(2-pyridinylmethyl)-1,2-ethanediamine (TPEN) were purchased from Sigma-Aldrich Israel (Rehovot, Israel). The cell permeant viable fluorescent zinc probe FluoZin3-AM and LysoTracker Red DND-99 were obtained from Thermo Fisher Scientific (Waltham, MA). Zinc sulfate was obtained from Merck (Rosh-Ha’ayin, Israel). Bafilomycin A1 (BafA1) was obtained from Enzo Biochem, Inc. (Farmingdale, NY).
ZnT2-YFP and ZnT2-Ruby were generated as described previously [9,33]. Lyso-pHluorin was obtained from Addgene (Plasmid #70113).
Results are presented as means/medians ± S.D. Statistical comparisons were performed using Student’s t-test (Prism Graph Pad, Berkeley, CA), and a significant difference was demonstrated when p<0.05. Results from at least three independent experiments are shown.
Human MCF-7 breast-cancer cells were grown and transiently transfected as previously described [9]. FluoZin3-AM, a specific zinc indicator, labeled the zinc-containing vesicles that were detected solely in cells which overexpressed an active ZnT2 transporter. In contrast, cells transfected with an empty Ruby plasmid, showed very low levels of zinc accumulation that reflected the low number of FluoZin3-positive vesicles per cell. For the zinc transport function experiments, 18 hr after transfection, cells were incubated for 1 hr in growth medium containing increasing concentrations of BafA1 (0–1000 nM). Then, 1 μM FluoZin3-AM was added to the growth medium for 45 min, followed by double cell wash with PBS, trypsinization, and collection for analysis using a BD FACS Aria IIIu flow cytometer. Cells that were analyzed by a Zeiss LSM-710 confocal microscope were kept in BafA1-containing medium during the analysis.
At least 3 independent experiments were performed, and at least 10,000 cells were analyzed in each experiment.
MCF-7 cells tranfected with ZnT2-Ruby or empty Ruby plasmid were incubated in growth medium, growth medium containing 75 μM ZnSO4, or growth medium supplemented with 5μM TPEN for 1 hr. Then, 1 μM FluoZin3-AM was added for 1 hr, and the cells were analyzed for FluoZin3 fluorescence levels using flow cytometry. The mean FluoZin3 fluorescence levels were calculated only for live cells that expressed ZnT2-Ruby or empty Ruby.
MCF-7 cells tranfected with ZnT2-YFP or empty YFP plasmid were incubated in either growth medium, growth medium containing various concentrations of exogenously added ZnSO4, or growth medium supplemented with 5μM TPEN for 1 hr. Then, 100 nM LysoTracker Red was added for 1 hr, and the cells were analyzed for LysoTracker-Red fluorescence levels using flow cytometry. The mean LysoTracker Red fluorescence levels were calculated only for live cells that expressed ZnT2-YFP or empty YFP.
MCF-7 cells tranfected with ZnT2-Ruby were incubated either in growth medium, growth medium containing 75μM ZnSO4, or growth medium supplented with 5μM TPEN for 2 hr. Then, the cells were analyzed for Lyso-pHluorin fluorescence levels using flow cytometry. The mean Lyso-pHluorin fluorescence levels were calculated only for live cells that expressed ZnT2-Ruby.
A magnification of ×63 under immersion oil was used. Excitation wavelengths were: 405 nm for Hoechst 33342, 488 nm for FluoZin3 or Lyso-pHluorin, and 543 nm for Ruby-tagged ZnT2 proteins.
The putative binding site [34] of ZnT2 consists of 2 Asp residues and 2 His residues, namely H106, D110, H223, and D227 (see Fig 1). In our recent paper we calculated which protonation state for the binding-site residues is the most energetically stable in the presence of bound zinc [21]. Here we expanded these electrostatic calculations and present results for both cytoplasm-facing and vesicular lumen-facing conformations, based on the structural models used in our previous work, and using the PDLD/S-LRA method as well as Coulomb’s law, in the presence and absence of zinc (see Methods and S1 Text for more details). This scheme calculates the difference in the system’s free energy, as a function of the protonation state and is therefore a reliable indication of which state is most highly populated, i.e. the dominant state. At pH = 7, we find that both conformations are most stable when D110 and D227 are charged, while H106 and H223 are uncharged; i.e. all residues are deprotonated, resulting in a net charge of zero when including the divalent zinc ion (Fig 2). When zinc is absent, the energy difference between the states is small, with no dominant preference. However, one must consider the physiological context in which ZnT2 functions: taking up a zinc ion from the cytoplasm, where the pH is ~7.2, and releasing it into the acidic vesicular lumen, where the pH is ~5.5. These deviations from pH = 7 entail an energy shift (using the equation ΔG = 1.38×(pH-7) for kcal/mol at 300 K). The relative energies of the states at these pH values (7.2 and 5.5 for cytoplasmic-facing and vesicular lumen-facing states, respectively), present a decisively different situation (Fig 2). Zinc binding from the cytoplasm still presents the all-deprotonated state as the most stable state (pH 7.2 with zinc); however, after zinc is released to the vesicular lumen the most stable state becomes the one where H106 and H223 are protonated (pH 5.5 without zinc), for an all-charged state (where the net charge of the system is again zero). This result is somewhat expected, as His residues are empirically prone to become protonated in mildly acidic conditions, and provided us with the first indication that zinc is exchanged for protons. One issue that arises from this analysis is the lack of experimentally reliable information regarding the actual concentration of zinc that is available for the transporter. Addressing this issue is far from trivial due to experimental limitations [35,36]. As a consequence, the energies presented are not adjusted to the effective zinc concentration, and the absolute value of the energy can be misleading. We proceeded with the assumption that the effective zinc concentration is fairly high in the cytoplasm (where zinc-carrying metallothioneins are abundant [37,38]), which would result in a decrease in energy for the zinc bound states at the cytoplasm.
To further investigate the transport mechanism of ZnT2, we next calculated the pKa values of the four binding-site residues, in order to determine the feasibility of H106 and H223 binding protons. The calculations were performed using our PDLD/S-LRA method, which was successfully applied to many studies in the past (e.g. [31,39]). We calculated the pKa values of H106, D110, H223, and D227 at all protonation state combinations, in the presence or absence of zinc (Table 1). Both Asp residues are expected to remain unprotonated regardless of the conditions. Interestingly, both His residues show uncharacteristically low pKa values when zinc is present, due to the positive charge of the zinc ion, however, their pKa values increase dramatically in the absence of zinc, with values >7 for one proton and in the range of 5–6 for two protons (Table 1). This remarkable result suggests that two protons are exchanged for each zinc ion transported; the low pKa values in the presence of zinc suggest that zinc outcompetes both protons, whereas the pKa values that are comparable to the surrounding, including the acidic vesicular lumen, suggest protonation in the absence of zinc (see more details below). Based on the acidic pH of the vesicular lumen, we therefore propose that the zinc ion is being exchanged for two protons. We note that the pKa values themselves are not decisively pointing to multiple protons being exchanged, but the total free energy of the state, which takes into account more energy terms than the pKa calculation alone, does point at a 2:1 ion exchange ratio (see Fig 2).
We next evaluated the binding energy of zinc ions to establish a complete mechanism of transport which considers the kinetic parameters as well, i.e. the energy barriers. We performed rigorous binding free-energy calculations for zinc, using the PDLD/S-LRA method, at different protonation states of the binding site residues, and the results (including previously calculated data from [21]) are presented in Fig 3. It is immediately evident from our calculations that zinc binds with high affinity when the His residues are deprotonated, while binding is weaker with one proton bound and essentially counter-productive when the two His residues are protonated. Expectedly, the energy barriers are similarly affected by the number of protons bound, equivalent to the behavior of the intersection point in Marcus parabolas [40].
In summary, when taking into account the energy calculations provided by our charged-cluster total free energy, pKa value calculations, and zinc-binding free energies, we observe a more likely exchange mechanism of 2 protons for 1 zinc ion, with His 106 and His 223 playing the role of the proton-binding sites. To test this hypothesis, we proceeded to build a model that considers these energies and simulates the exchange cycle within the system. We then subjected the model to Monte Carlo (MC) simulations (Fig 4). The energies we used (Fig 3B) were derived from the state free energies (Fig 2), and the energy barriers were obtained from Fig 3A. The reason for choosing the total free energy for the energy of the different states in the MC model is because the total free energy takes into account more energy terms and is therefore (especially when dealing with uncertainties of modeling) more reliable and prone to convergence. We reiterate here that the total energy of zinc is not adjusted to the concentration of zinc (which is experimentally unknown) and therefore the model suffers in this respect from uncertainty. Nevertheless, evaluating the model and performing the experimental analysis below is still able to provide useful and testable predictions.
As evident from Fig 3 (compare A and B), we manually modified and increased the barriers on the closed side (cytoplasmic) of the lumen-facing model, since they were almost as low as the barriers on the open side and would account for a highly leaky transporter (see S1 Fig). This discrepancy is probably a consequence of the uncertainties involved in using a structural model rather than an experimentally solved 3D structure, as well as some deviations inherent to computational methods and convergence issues.
The MC model is described in more detail in the Methods section. Briefly, protons and zinc ions from bulk pools were allowed to bind and dissociate from the binding site (at H106 and H223) passing through the barriers, while the energy differences (shown in Fig 3B) were used for the MC conditioning, applying the Metropolis acceptance criterion [30]. The energy of the protons in the bulks was adjusted to the pH (by modifying the energy according to the equation mentioned above) and transfers were counted throughout the MC simulation. The energy difference and barrier for the conformational change were parametrically implemented due to the lack of experimental knowledge at this time. We chose a barrier of 18 kcal/mol (in the range of the expected barrier for the rates reported for YiiP [41]) to achieve reasonable data in the MC simulation (see the Discussion about temperature below) and an energy difference of 0 kcal/mol between the cytoplasmic-facing and lumen-facing conformations for simplicity. In S2 Fig we provide results for MC models using a conformational change barrier of 16 and 20 kcal/mol where one can see that if the conformational change barrier is very high, compared to the barrier for ions on the closed side (Fig 3B), leak will occur, otherwise, the transporter is not very sensitive to this parameter. We also tested the incorporation of a differential barrier for the conformational change, i.e. a small increase (2 kcal/mol) to the conformational change barrier if the total charge of the binding site cluster is non-zero. The rationale behind this modification is that the conformational change must proceed through an intermediate with no access to the bulk on either side of the membrane. This conformation would occlude the cluster from water. Thus, any state with a net charge would be energetically unfavorable due to lack of access to water stabilization. However, even without this condition, the conclusions remain, albeit at slightly lower efficiency and specificity (see S3 Fig).
The results of the MC simulation are presented in Fig 4, where we present the ion fluxes obtained using our model (Fig 4 left; where flux is the average number of ions that were transported from one side of the membrane to the other). We observe fluxes in opposite directions, indicative of an antiport function, with a flux ratio of 2.35±0.43 (Fig 4 right; average ratio of number of protons transported for each single zinc ion translocated), suggesting a mechanistic stoichiometry of 2H+/Zn2+. Our model provides a reasonable efficiency of 69±13% (counted as the number of zinc ions transported per two conformational-change events, for a full cycle). The MC simulations were performed at an elevated temperature (500 K; it should be emphasized that this does not represent a transporter functioning at such an elevated temperature but is rather a mathematical tool to scale the energies and increase the sampling in the MC simulation) because the barriers were otherwise too high to record enough events within a reasonable running time. Consequently, MC simulations were also performed at lower temperatures (albeit counting fewer overall events) and the results are presented in S4 Fig. Expectedly, at lower temperatures the barriers are less permissive (in the MC simulation) and therefore the stoichiometry is closer to 2 and the efficiency is higher (2.05 and 92%, respectively). We note that the simulations running at lower temperatures entail more steps in order to record sufficient events (see Methods). Additionally, for the 400K simulations we only included in the analysis simulations that successfully achieved at least one cycle (i.e. >2 conformational change events) accounting for 28 out of the 65 simulations submitted (at higher temperatures all simulations submitted completed several cycles each and were all included in the analysis). Finally, we point out that the free-energy landscape, on which the MC simulations are based, was calculated at room temperature.
We note that one aspect in our model that presents a difficulty is the low concentration of free zinc ions in the cytoplasm. The concentration of free zinc is exceedingly low, translating in low bulk energies for zinc ions in the model, and consequently higher energy barriers to bind zinc. Barriers for concentrations below μM range were too high to simulate, and we parametrically used higher concentrations of zinc as a proof-of-concept. We therefore suggest two possible explanations: (i) the energy barriers for zinc on the closed side of either conformation are underestimated by several kcal/mol and the conformational change barrier should be higher as well. This cannot be simulated in the MC simulation but extrapolation from our data predicts a viable transporter. (ii) the effective concentration (and therefore the energy of zinc ions) is much higher than the concentration of zinc ions in the bulk. This can be a result of the ample availability of zinc-carrying proteins in the cytoplasm, and direct interactions between ZnT2 and zinc-carrying proteins like metallothioneins, perhaps through site B mentioned in previous studies [21,22], allowing zinc to bind ZnT2 at an effective higher concentration.
We wish to emphasize that the model has its limitation, and a few parameters were manually inserted into the model (due to lack of experimentally determined values). Having said that, past modeling attempts using similar assumptions and parametrization have proven useful, and revealed genuine phenomena (e.g. [42]). Additionally, the modeling process supports the experimental validation, and none of the experimental protocols are based on any prior assumption arising from the model.
In summary, both methods reveal the same results, a proton/zinc antiporter functioning at an apparent 2:1 stoichiometry and at a reasonably high efficiency. Equipped with our model, we sought to functionally verify our predictions experimentally.
In order to provide experimental evidence that ZnT2 is a proton-driven zinc transporter, we examined whether upon alkalization of the vesicular pH, the ability of ZnT2 to accumulate zinc in these vesicles will decrease. We therefore used Bafilomycin A1 (BafA1), which abolishes vesicular acidification via potent inhibition of V-ATPase. We first determined at which BafA1 concentration one can obtain a significant vesicular alkalization. Towards this end, we used two distinct fluorescent pH sensors: (i) LysoTracker Red DND-99, which fluorescently labels acidic vesicles in live cells including lysosomes; and (ii) Lyso-pHluorin construct, which was transfected into human MCF-7 cells. Lyso-pHluorin is a variant of GFP that is targeted to lysosomes via a lysosomal membrane protein fused to a fluorescent tag localized in the lysosome lumen; hence, its fluorescence is markedly enhanced upon alkalization [43]. MCF-7 cells were transiently transfected with Lyso-pHluorin and 18 hr after transfection, the cells were collected and incubated in a growth medium containing increasing concentrations of BafA1. The cells were also incubated with LysoTracker Red and the cellular fluorescence levels of these two distinct dyes were determined using flow cytometry. Upon incubation with increasing BafA1 concentrations, the lysosomes as well as other V-ATPase-containing acidic vesicles, became increasingly less acidic as indicated by the BafA1 dose-dependent decrease in LysoTracker Red fluorescence (Fig 5A) and increase in Lyso-pHluorin fluorescence (Fig 5B). These results confirm that acidic vesicles including lysosomes, become alkaline upon treatment with BafA1.
ZnT2 was previously shown [9,21,44] to mediate the accumulation of zinc in intracellular vesicles in MCF-7 cells as indicated by the specific zinc probe FluoZin3 (Fig 6A). Furthermore, ZnT2 was previously shown to play a role in lysosomal zinc accumulation in both mouse mammary gland cells and in human HeLa cells [15,45,46]. We therefore determined the fraction of ZnT2 vesicles that co-localize with Lyso-pHluorin, indicating lysosomal localization of ZnT2 in MCF-7 cells using Imaris software for the confocal microscopy pictures. We found that 36±15% of ZnT2-containing vesicles co-localized with Lyso-pHluorin (S5 Fig). In order to assess whether vesicular zinc accumulation was proton-dependent, we treated MCF-7 cells with BafA1, which was shown above to markedly alkalize vesicular pH (Fig 5), and then determined FluoZin3 fluorescence levels. FluoZin3 fluorescence levels were significantly decreased (compared to untreated control cells) after treatment with BafA1 concentrations as low as 2.5–5 nM (Fig 6A and 6B). In addition, using an In-cell analyzer, we found that the average number of FluoZin3-containing vesicles did not increase upon the addition of 75 μM ZnSO4 to the growth medium; however, upon treatment with concentrations ≥ 2.5 nM BafA1, the average number of FluoZin3-containing vesicles significantly decreased when compared to cells that remained in growth medium (Fig 6C). Interestingly, BafA1 concentrations > 5 nM neither further decreased FluoZin3 fluorescence levels, nor the number of vesicles (Fig 6B and 6C), indicating that these low concentrations of BafA1 are sufficient to alkalize vesicular pH (Fig 5), thereby disrupting ZnT2-dependent vesicular zinc accumulation.
As mentioned above, overexpression of ZnT2 in MCF-7 cells results in high vesicular zinc accumulation. In order to assess the impact of increased extracellular zinc on intravesicular zinc accumulation, we determined FluoZin3 fluorescence levels using flow cytometry upon transient overexpression of Ruby-tagged ZnT2. We found that upon overexpression of ZnT2, cellular fluorescence levels of FluoZin3 were 7.5-fold higher than cells treated with TPEN, a specific zinc chelator; upon addition of 75 μM ZnSO4 into the growth medium, FluoZin3 fluorescence levels were 3.5-fold higher relative to cells that were grown in regular growth medium containing ~2 μM zinc (Fig 7A).
We next examined whether upon overexpression of ZnT2 in intracellular vesicles, with the addition of extracellular zinc, one can obtain vesicular alkalization, which can indicate a ZnT2-dependent proton export from the vesicles. Indeed, upon transient overexpression of ZnT2-YFP and increasing exogenous concentrations of ZnSO4, we observed a dose-dependent decrease in LysoTracker Red fluorescence, as determined by flow cytometry, which was significantly different from cells that were not exposed to additional exogenous zinc (Fig 7B). In contrast, cells transfected with an empty vector YFP, maintained the same LysoTracker Red fluorescence at the different zinc concentrations. Moreover, this decrease in LysoTracker Red fluorescence upon addition of 75 μM ZnSO4 was reversed when cells were treated with TPEN, a zinc chelator (Fig 7C). These results show that upon transient overexpression of ZnT2, the massive transport of zinc mediated by ZnT2 into intracellular vesicles, leads to a marked alkalization of the vesicular pH, and this effect is abolished upon zinc chelation by TPEN (Fig 7C). Furthermore, co-transfection of MCF-7 cells with Lyso-pHluorin and ZnT2-Ruby, showed a similar pattern; ZnT2 overexpressing cells exhibited high fluorescence levels of Lyso-pHluorin upon addition of 75 μM ZnSO4 and a markedly reduced fluorescence upon treatment with TPEN (Fig 7D).
In the present study we showed that upon incubation of cells transiently transfected with ZnT2 in zinc-containing medium, a marked vesicular zinc accumulation occurs, which in turn provokes an alkalization of these initially acidic vesicles. Consistently, disruption of the acidic pH of these vesicles by pharmacological inhibition of V-ATPase, abolished vesicular zinc accumulation. Furthermore, based on our computational energy calculations, we propose that ZnT2 functions as a vesicular proton-coupled zinc transporter with a stoichiometry of 2H+/Zn2+. Hence, these findings suggest that ZnT2 localized in acidic vesicles, mediates the active translocation of zinc from the cytosol into acidic vesicles, coupled to the movement of two protons in the opposite direction. This proposed mechanism of proton-coupled substrate transport into acidic vesicles is well-established for various transporters of the SLC superfamily, mediating the proton-dependent transport of an assortment of key physiological substrates. Specifically, vesicular transporters of the SLC18 family, comply with this mode of proton-coupled substrate antiport; for example, vesicular storage of monoamines including serotonin, dopamine, histamine, adrenaline and noradrenaline is mediated by the vesicular monoamine transporters (VMATs) 1 and 2 [47,48]. Hence, VMAT1 (SLC18A1) and VMAT2 (SLC18A2) mediate the packaging of these monoamines from the cytoplasm of neurons in presynaptic vesicles, a process involving the obligatory exchange of two protons (i.e. movement of protons from the acidic vesicles to the cytosol) per monoamine substrate. However, it should be noted that in contrast to the chemically inert zinc ion, it is essential that dopamine and noradrenaline are stored within acidic vesicles where they cannot be autoxidized. Specifically, unlike other biogenic amines important in the central nervous system, dopamine and noradrenaline are capable of undergoing a non-enzymatic autoxidative reaction giving rise to a superoxide anion that further decomposes to reactive oxygen species [49]. This autoxidative reaction was suggested to affect the incidence of Parkinson disease [49,50]. The vesicular acetylcholine transporter (VAchT/SLC18A3) is also a proton-coupled antiporter which mediates the concentration of acetylcholine within acidic vesicles in cholinergic presynaptic neurons, while consistently moving two protons to the cytoplasm per acetylcholine molecule transported [48]. Similarly, the vesicular inhibitory amino acid transporter (VIAAT) or vesicular GABA transporter (VGAT) aka SLC32A, is also a vesicular proton-coupled antiporter of glycine and gamma-amino butyric acid (GABA); VIAAT exchanges GABA or glycine for protons with a presumed 1:1 stoichiometry, although this suggested stoichiometry remains completely unsettled [48,51]. VIAAT is present on synaptic vesicles of inhibitory GABAergic and glycinergic neurons. In contrast, certain vesicular/endosomal transporters function in a proton-coupled cotransport mechanism; for example, as almost all iron in the circulation is bound to transferrin under physiological conditions, cellular uptake of iron predominantly proceeds via transferrin receptor-mediated endocytosis [52]. Following reduction by endosomal ferrireductases and release from transferrin, iron is exported from the acidic endosomal compartment to the cytosol in a proton-coupled transport mediated by the divalent metal transporter DMT1 (DMT1/Nramp2/SLC11A2) [53]. In enterocytes, DMT1, located at the apical side of the enterocyte epithelium, functions as a proton-iron cotransporter mediating the intestinal uptake of iron [53]. Whereas, in all other cells, DMT1 is found in intracellular membranes (i.e. endosomes), where it mediates the exit of endocytosed iron from endosomes into the cytoplasm in a proton-dependent manner [54,55]. Similarly, following the folate receptor-mediated endocytosis of reduced folates, the proton-coupled folate transporter (PCFT/SLC46A1) cotransports folates from acidic endolysosomes into the cytosol in a proton-dependent manner [56,57]. Hence, key physiological molecules such as neurotransmitters, vitamins and micronutrients may be transported into vesicles or exported out of acidic vesicles, in a proton-dependent manner in either a proton-coupled antiport or cotransport mechanism. Based on concentrative capacity calculations similar to the ones previously shown for VMAT1, 2 [51], a pH difference of ~1.5 units between the acidic vesicular lumen and the neutral cytoplasm would drive a ~100-fold concentrative ability for ZnT2; it should be noted that a divalent ion such as zinc requires the energy of two protons to achieve the same concentrative capacity as a monovalent ion would, when driven by only one proton. In this respect, our present study suggests that ZnT2 is an electroneutral antiporter exchanging two protons for zinc, a divalent metal cation, and thus any existing membrane voltage would not affect ZnT2’s concentrative capacity. Supporting our finding that ZnT2 is electroneutral, one should consider that extruding a single proton for the intravesicular import of a divalent cation would generate an unfavorable electrogenic transport, especially as the resting membrane potential of acidic vesicles is +20 to +40mV in the luminal side [51].
We find here that upon ectopic overexpression of ZnT2 and consequent vesicular zinc concentration, a marked vesicular alkalization occurred. This finding highlights the fact that during lactation in which ZnT2 expression is markedly induced, an intricate balance between the overexpression of vesicular ZnT2 and V-ATPase must be maintained in order to ensure that the acidic pH is maintained by V-ATPase as otherwise, alkalization will result in loss of the proton-motive force and consequent disruption of multiple vesicular secretory functions. In support of this hypothesis is a recent paper by Lee et al., [58] which showed that during lactation in a mouse model, the protein levels of V-ATPase are markedly elevated in mammary gland epithelial cells. This of course highlights the obligatory balance between overexpressed ZnT2 (as well as other transporters) which translocates protons to the cytoplasm during vesicular accumulation of zinc (and other physiological substrates) and proper V-ATPase levels. Therefore, it is not surprising that Lee et al., [58] discovered that ZnT2 directly interacts with V-ATPase; they further found that deletion of ZnT2 impaired vesicle acidification, biogenesis, trafficking, and secretory function. These novel findings underscore the intricate regulation of the balanced overexpression of vesicular transporters like ZnT2 and an accompanying proper elevation of V-ATPase levels to retain the crucial acidic pH of these secretory vesicles.
We herein sought to address the question of which ion provides the driving force for ZnT2 to actively pump zinc ions across the vesicular membrane. A priori, protons appear very likely to drive ZnT2-dependent zinc transport considering the high number of ionizable residues at the putative zinc-binding site, namely, 2 His and 2 Asp residues for ZnT2 [21]. Thus, in the current study we addressed the following key questions: do protons drive ZnT2-mediated zinc transport; if so, which specific residues bind the proton(s); and how many protons are being exchanged for each zinc ion translocated. Using free-energy calculations, we determined the pKa values of key residues, the relative free energy of the different protonation-state configurations of the zinc binding site residues and the corresponding binding energy of zinc. We reached the conclusions that H106 and H223 serve as proton-binding residues, and that ZnT2 functions as a H+/Zn2+ antiporter with an apparent stoichiometry of 2:1. In this context, one should consider the challenges involved in modeling some of these aspects. Specifically, reliable calculations of pKa values, especially for functional residues that are not readily available to the bulk, has long been a difficult task [59]. Additionally, the calculation of binding energies of ions with a high charge is very challenging due to the high solvation energy (-467 kcal/mol in the case of Zn2+) [60]; this high solvation energy needs to be compensated by interactions with the protein’s residues, however, these energy calculations could face convergence issues. To this end, our experience shows that using the semi-microscopic PDLD/S-LRA method (see Methods and SI) is very effective at overcoming these convergence issues.
We present a model for the mechanism of transport and verify it using MC simulations (Figs 3 and 4); our MC model is assessed for its robustness by modifying several parameters (S1–S4 Figs) and finally validated experimentally in live cells. Using fluorescent techniques, we showed that under alkalization of the vesicular pH, the zinc transport function of ZnT2 was disrupted, leading to the conclusion that ZnT2 functions as a proton-driven antiporter (see Fig 8 for a summarizing scheme; it should be noted that in this figure we manually adjusted the energies to account for the effective zinc concentration, explained above, for visual purposes).
ZnT2 functions as an obligatory dimer [33], but at this point we did not incorporate this information into our model. However, one can conceive several possibilities for how the dimers exhibit transport with the most reasonable ones being coupled dimers (i.e. both monomers undergo concerted conformational changes) or decoupled dimers (i.e. each monomer functions independently). The former is likely to increase the efficiency of the transport, provided that ZnT2 harbors a higher barrier for the conformational change if the cluster charge is non-zero (see above), because each monomer not only depends on its own charge but also on the total charge of the counterpart monomer, thus effectively increasing the penalty to undergo conformational changes that are not efficient. The underlying assumption here is that the ‘proper’ state of the system is the lowest in energy, which is true for our case, as shown in Fig 2.
Another point that we wish to bring forward is the potential existence of ligand-induced conformational selection driving ZnT2 transport cycle (e.g. similar in principle to vSGLT [61]). Whereas our MC simulations show that this is not strictly necessary (Fig 4), a model where the different ions drive the conformational change in the ‘correct’ direction would prove beneficial in terms of the rate of the reaction. In other words, if ZnT2 were to selectively prefer the cytoplasm-facing conformation, when it is bound to protons, and the lumen-facing conformation, when it binds a zinc ion, this would direct the conformation towards the fastest exchange rate, since as soon as the ion would bind on one direction, the equilibrium would shift, driving the exchange reaction forward, and so forth. Now, although we do not have direct observations for this (i.e. the difference in energy between the conformations has not been determined), it is encouraging to see that zinc binding is stronger in the lumen-facing conformation, which would suggest zinc shifts the equilibrium to some extent in that direction. Similarly, the pKa values for the His residues are on average lower in the cytoplasmic-facing conformation (Table 1), suggesting that protonation of the His residues should shift the equilibrium to the cytoplasm-facing conformation. Thus, the binding energies should in principle promote a behavior of ‘seesaw-like’ directional conformational changes.
In summary, using computational analysis we propose that ZnT2 functions as a vesicular proton-coupled zinc transporter with an apparent stoichiometry of 2H+/Zn2+, and provide experimental evidence for the proton-driven zinc transport of ZnT2. Our experimental and computational findings shed light on the molecular transport mechanism of ZnT2 and expand our knowledge regarding other zinc transporters.
|
10.1371/journal.pntd.0001676 | Clinical Relevance and Discriminatory Value of Elevated Liver Aminotransferase Levels for Dengue Severity | Elevation of aspartate aminotransferase (AST) and alanine aminotransferase (ALT) is prominent in acute dengue illness. The World Health Organization (WHO) 2009 dengue guidelines defined AST or ALT≥1000 units/liter (U/L) as a criterion for severe dengue. We aimed to assess the clinical relevance and discriminatory value of AST or ALT for dengue hemorrhagic fever (DHF) and severe dengue.
We retrospectively studied and classified polymerase chain reaction positive dengue patients from 2006 to 2008 treated at Tan Tock Seng Hospital, Singapore according to WHO 1997 and 2009 criteria for dengue severity. Of 690 dengue patients, 31% had DHF and 24% severe dengue. Elevated AST and ALT occurred in 86% and 46%, respectively. Seven had AST or ALT≥1000 U/L. None had acute liver failure but one patient died. Median AST and ALT values were significantly higher with increasing dengue severity by both WHO 1997 and 2009 criteria. However, they were poorly discriminatory between non-severe and severe dengue (e.g., AST area under the receiver operating characteristic [ROC] curve = 0.62; 95% confidence interval [CI]: 0.57–0.67) and between dengue fever (DF) and DHF (AST area under the ROC curve = 0.56; 95% CI: 0.52–0.61). There was significant overlap in AST and ALT values among patients with dengue with or without warning signs and severe dengue, and between those with DF and DHF.
Although aminotransferase levels increased in conjunction with dengue severity, AST or ALT values did not discriminate between DF and DHF or non-severe and severe dengue.
| Dengue is a global public health problem, as the incidence of the disease has reached hyperendemic proportions in recent decades. Infection with dengue can cause acute, febrile illness or severe disease, which can lead to plasma leakage, bleeding, and organ impairment. One of the most prominent clinical characteristics of dengue patients is increased aspartate and alanine aminotransferase liver enzyme levels. The significance of this is uncertain, as it is transient in the majority of cases, and most patients recover uneventfully without liver damage. In this study, we characterized this phenomenon in the context of dengue severity and found that, although liver enzyme levels increased concurrently with dengue severity, they could not sufficiently discriminate between dengue fever and dengue hemorrhagic fever or between non-severe and severe dengue. Therefore clinicians may need to use other parameters to distinguish dengue severity in patients during early illness.
| Dengue is a mosquito-borne arboviral infection endemic to most tropical and subtropical countries [1]. Elevation of the liver enzymes aspartate aminotransferase (AST) and alanine aminotransferase (ALT) is common in acute dengue illness, occurring in 65–97% [2], [3], [4], [5] of dengue patients, peaking during the convalescent period of illness (days 7–10) [2], [4], [6]. In dengue-endemic countries, dengue is an important cause of acute viral hepatitis [7].
Elevated AST and ALT levels have been associated with bleeding [2], [4], [6] and dengue hemorrhagic fever (DHF) [3], [8]. Liver failure has been recognized as a complication and unusual manifestation of dengue [9], [10] but occurred infrequently in 3 of 270 patients in Taiwan [6] and 5 of 644 patients in Vietnam [4]. In Malaysia, 8 of 20 pediatric DHF patients developed liver failure, 1 died, and the rest recovered completely [11]. In Singapore, AST or ALT levels were not independent predictors of DHF in 1973 adult dengue patients [12].
In 2009, the World Health Organization (WHO) revised its dengue guidelines and proposed severe organ impairment as one category of severe dengue in addition to severe plasma leakage and severe bleeding [1]. Severe liver involvement was defined as AST or ALT≥1000 units/liter (U/L). In Taiwan, AST>10 times the upper limit of normal (ULN) occurred in 11% of dengue patients [6], while in Brazil this occurred in 4% of their cohort [3]. In this study, we aimed to evaluate the clinical relevance of elevated AST and ALT levels and correlate liver aminotransferase levels with dengue severity according to WHO 1997 and 2009 classifications.
All laboratory-confirmed dengue patients identified from our hospital microbiology database and treated using a standardized dengue clinical care path at the Department of Infectious Diseases, Tan Tock Seng Hospital (TTSH), Singapore from 2006 to 2008 were retrospectively reviewed for demographic, serial clinical and laboratory, radiological, treatment, and outcome data. These cases were positive by real-time polymerase chain reaction (PCR) [13]. We included patients with only positive dengue serology in only subgroup analyses, as we did not have paired sera, and other etiologies for elevated AST and ALT could not be excluded without more extensive evaluation.
Cases were categorized using serial clinical and laboratory data from the entire clinical course as dengue fever (DF), DHF, or dengue shock syndrome (DSS) using WHO 1997 classifications [9]. Dengue fever classification requires fever and at least two of the following: headache, eye pain, myalgia, arthralgia, rash, bleeding, and leukopenia. Dengue hemorrhagic fever requires all of the following: fever, platelet count ≤100×109/liter, bleeding, and plasma leakage [9]. Dengue shock syndrome is a case of DHF with either tachycardia and pulse pressure <20 mmHg or systolic blood pressure <90 mmHg [9].
Cases were also categorized as dengue without warning signs (WS), dengue with WS, or severe dengue using WHO 2009 classifications [1]. Dengue (WHO 2009) requires fever and two of the following: nausea, vomiting, rash, aches and pains, leukopenia, or any warning sign [1]. Warning signs include abdominal pain or tenderness, persistent vomiting, clinical fluid accumulation, mucosal bleeding, lethargy or restlessness, hepatomegaly, or hematocrit rise (≥20%) with rapid drop in platelet count (<50,000/liter) [1], [14]. We modified the WHO 2009 warning sign of rise in hematocrit concurrent with rapid drop in platelet count by quantifying it as hematocrit ≥20% concurrent with platelet count <50,000/liter, as this was shown to correlate significantly with dengue death in our adult dengue death study [14]. Severe dengue includes severe plasma leakage, severe bleeding, and severe organ impairment [1].
We performed a subgroup analysis for median maximum AST and ALT values stratified by febrile (days 1–3 of illness), critical (days 4–6), and convalescent (days 7–10) phases as defined by WHO 2009 [1] and compared across dengue severity classification according to WHO 1997 [9] and 2009 [1].
We excluded severe dengue due to isolated elevation of AST or ALT≥1000 U/L from our definition of severe dengue outcome, as this would be a confounder in assessing the relevance of AST or ALT levels in defining dengue severity. Patients had AST/ALT taken at presentation and then throughout hospitalization at the physician's discretion. Maximum AST and ALT values recorded at a median of 4 days of illness (interquartile range [IQR]: 3–5 days) were used in this analysis. Those with pre-existing liver diseases were excluded. At TTSH, the ULN for AST is 41 U/L; for ALT, it is 63 U/L for males and 54 U/L for females.
We assessed the clinical relevance of elevated AST or ALT levels using four liver failure criteria—two for acute liver failure, and two that determine prognosis from chronic liver disease. The American Association for the Study of Liver Diseases (AASLD) recommends defining acute liver failure in a patient as: international normalized ratio (INR)≥1.5, any degree of altered mental status, and illness <26 weeks in duration without pre-existing cirrhosis [15]. The King's College criteria assess prognoses in those with acute liver failure; the criteria are: prothrombin time >100 seconds or 3 of the following: age >40 years, prothrombin time >50 seconds, serum bilirubin >18 mg/dL, time from jaundice to encephalopathy >7 days [16]. The model for end-stage liver disease (MELD) determines three-month mortality based on the following formula: 3.8×(log serum bilirubin [mg/dL])+11.2×(log INR)+9.6×(log serum creatinine [mg/dL])+6.4 [17]. The Child-Pugh criteria include assessment of degree of ascites, serum bilirubin and albumin, prothrombin time, and encephalopathy to determine one- and two-year survival [18].
The Mann-Whitney U and Kruskal-Wallis tests were used to determine statistical significance for continuous variables, and chi-square or Fisher's exact test for categorical variables. Statistical tests were conducted at the 5% level of significance. Receiver operating characteristic (ROC) curves showing the area under the curve (AUC) were generated to determine the discriminatory performance of aminotransferase values. All statistical analyses were performed using Stata 10 (Stata Corp., College Station, TX).
This was a retrospective study involving data collection from medical records. All patient data were anonymized during analysis. This study was approved by the Institutional Review Board, National Healthcare Group, Singapore [DSRB E/08/567].
From 2006 to 2008, 690 dengue PCR positive cases were reviewed. Males comprised 493 (71%) of the cases, and the median age of the cohort was 35 years (IQR: 27–43 years). A Charlson comorbidity index ≥3, which predicts increased one-year mortality [19], was noted in 5 (0.7%) patients. With WHO 1997 classification, 62% had DF, 31% DHF, and 7% DSS. With WHO 2009 classification, 14% had dengue, 62% had dengue with warning signs, and 24% had severe dengue. Hence, by WHO 1997 classification, 38% of patients with DHF/DSS needed close monitoring, while by WHO 2009 classification, 86% of patients with warning signs or severe dengue required close monitoring.
Median length of illness from onset to hospital presentation was 4 days (IQR: 3–5 days), while median length of hospital stay was 5 days (IQR: 4–6 days). Intravenous fluid was administered to 641 (93%) and platelet transfusion to 86 (12%). Intensive care unit (ICU) admission was required in 3 patients, and death occurred in 1 patient due to pneumonia.
Overall, 595 (86%) had AST above the ULN, and 316 (46%) had ALT above the ULN. Seven patients (1.0%) had severe dengue according to WHO 2009 criteria concurrent with AST or ALT≥1000 U/L while three additional patients had severe dengue due to AST or ALT≥1000 U/L only. Of the former seven patients, 86% had severe plasma leakage, 29% had severe bleeding, and none had severe organ impairment other than isolated AST or ALT≥1000 U/L. Among the 3 patients admitted to the ICU, AST or ALT values were above the ULN but below 1000 U/L.
No patients in our cohort developed acute liver failure under AASLD or King's College criteria. With Child-Pugh scoring, 2 (0.3%) belonged to Child-Pugh class C. With MELD scoring, predicted three-month mortality of 6% were identified in 68 (10%) patients in our cohort and 19.6% in 2 (0.3%) patients. The same two patients who were Child-Pugh class C also had a predicted 19.6% three-month mortality using MELD scoring; they both had DSS and severe dengue.
Median AST and ALT values for dengue without warning signs, dengue with warning signs, and severe dengue (Table 1) were 83.5 U/L, 92 U/L, and 124 U/L, respectively (p<0.001); median ALT values were 49 U/L, 53 U/L, and 73.5 U/L (p = 0.002). Table 2 shows median AST and ALT values for patients with DF, DHF, and DSS. Median AST values for these categories were 93 U/L, 103 U/L, and 137.5 U/L, respectively (p = 0.01), and median ALT values were 52 U/L, 60 U/L, and 74 U/L (p = 0.05).
In a separate analysis of our serology-positive cohort (n = 1487), median AST values for dengue without WS, dengue with WS, and severe dengue were 84 U/L, 114 U/L, and 147 U/L (p<0.001). Median ALT values were 56 U/L, 73 U/L, and 97.5 U/L (p = 0.01). For patients with DF, DHF, and DSS, median AST values were 105 U/L, 130 U/L, and 129 U/L (p<0.001), and median ALT values were 68 U/L, 78 U/L, and 85.5 U/L (p = 0.008).
In other hemorrhagic fevers, higher AST∶ALT ratios correlated with disease fatality [20]. In our PCR-positive cohort, median AST∶ALT ratios for DF, DHF, and DSS were 1.68, 1.68, and 1.88 (p = 0.29) and for dengue without WS, dengue with WS, and severe dengue, they were 1.60, 1.68, and 1.78 (p = 0.10), respectively.
The majority of our patients' maximum AST and ALT values were recorded during febrile (n = 258) and critical (n = 377) phases of acute dengue illness. By WHO 2009 dengue severity classification, the median AST and ALT values were significantly higher for severe dengue compared to dengue with and without warning signs during both the febrile and critical phases but not the convalescent phase (Table 3). By WHO 1997 classification, the median AST and ALT values were significantly higher for DHF versus DF and DSS in the febrile phase only but not critical and convalescent phases (Table 4).
In order to determine the reliability of AST and ALT values in defining dengue severity, ROC curves for AST and ALT against severe dengue excluding isolated transaminitis were determined (Figure 1). The AUC for AST was 0.62 (95% confidence interval [CI]: 0.57–0.67) and for ALT, 0.60 (95% CI: 0.54–0.64). This demonstrates that AST or ALT levels are insufficient to differentiate among the WHO 2009 dengue classifications. They were also poorly discriminatory between DF and DHF, as the areas under the curve (AUC) for AST and ALT were 0.56 (95% CI: 0.52–0.61) and 0.55 (95% CI: 0.51–0.59), respectively (Figure 2). In our serology-positive cohort, the AUC values for AST and ALT were 0.56 and 0.54 for differentiating between DF and DHF. The AUC values for severe and non-severe dengue were 0.64 and 0.60 for AST and ALT, respectively.
The box plots in Figure 3 for the distributions of AST values show considerable overlap among the liver enzyme values for those with dengue with and without warning signs, and severe dengue. Because there were extreme outliers in our cohort, only those with AST below 1000 U/L were included in these plots. Figure 4 shows overlapping AST values among those with DF and DHF. Similarly, considerable overlap was observed in ALT values for patients with dengue with and without warning signs, and severe dengue, as well as for DF versus DHF (data not shown).
Our analysis showed that liver aminotransferase levels were associated with but did not adequately differentiate between dengue severity. Although median AST and ALT values were significantly higher in those with DHF/DSS versus DF, and severe dengue versus non-severe dengue, very few (1.0%) had AST or ALT≥1000 U/L. Notably, none developed liver failure, and death occurred in only 1 patient (0.1%). The majority of patients recovered uneventfully.
The lack of acute liver failure in our study was not unusual, as the incidence of acute liver failure in dengue patients was 1.1% in studies by Trung and Kuo [4], [6]. The largest study to date reported no acute fulminant hepatitis [3]. In contrast to these adult studies, it is noteworthy that in dengue-endemic countries, dengue may be an important cause of acute liver failure in children [21], [22].
While some studies have shown that AST and ALT values differ between DF and DHF [3], [4], [8], few studies support AST or ALT as an independent predictor of DHF [23]. Two studies in Singapore found liver aminotransferase levels to be significantly elevated among DF and DHF patients [12] and survivors and non-survivors of dengue [24] on univariate analysis, but this association was lost after adjusting for confounders on multivariate analysis.
Trung et al. showed significant differences comparing other febrile illness, dengue without plasma leakage, and dengue with plasma leakage with and without shock during critical and convalescent phases for AST but during critical phase for ALT only [4]. We made the novel finding that liver aminotransferase levels may significantly vary according to dengue severity during the febrile phase. For DHF by WHO 1997 classification, both AST and ALT were significantly higher during the febrile phase compared to DF or DSS, and for severe dengue by WHO 2009, AST and ALT were significantly higher during the febrile and critical phases.
The impact of co-infection with hepatitis viruses or concomitant hepatotoxic drugs was not assessed in our retrospective study, although we did exclude those with known liver comorbidities. Kuo et al. found that hepatitis B or C did not increase the extent of liver aminotransferase elevation in a retrospective adult dengue study in Taiwan [6]. In contrast, Trung et al. found that hepatitis B co-infection modestly increased ALT levels without significant clinical impact in a prospective adult dengue study in Vietnam [4]. Tang et al. showed that dengue and hepatitis B co-infected patients showed an aberrant cytokine secretion profile compared with those with dengue alone. [25]. In Singapore, seroprevalence for hepatitis B was 2.8% [26] and hepatitis C 0.37% [27].
The etiology of elevated aminotransferase levels during acute dengue illness is unclear since AST is expressed in the heart, skeletal muscle, red blood cells, kidneys, brain, and liver, while ALT is secreted primarily by the liver [28], [29]. Because dengue infection can cause acute damage to these non-hepatic tissue types that express AST, raised aminotransferase levels may not be entirely due to severe liver involvement. It is therefore possible that the patients with high AST levels were also more likely to be classified as severe dengue under the 2009 criteria due to the common pathways to non-hepatic tissue damage, even though there is no association with poorer outcome.
Our retrospective study has some limitations. Aspartate and alanine aminotransferase values were tracked according to clinical judgment rather than at regular intervals during illness. We did not have dengue serotype data for each patient, but in 2006, DENV-1 was predominant in Singapore with a switch to DENV-2 in 2007–2008 [30]. Serology-positive cases were not included in primary analyses because our clinical laboratory used a rapid diagnostic test with potential for false positive results [31], we did not have paired sera to confirm dengue diagnosis [9], and not every patient with elevated AST or ALT was comprehensively evaluated for other etiologies of viral and non-viral hepatitis. Although serology-positive cases presented later during illness, we saw no difference in outcome. Five serology-positive patients (0.34%) required ICU admission versus 0.43% of PCR-positive cases, while four patients (0.27%) died in the serology-positive cohort, versus 1 patient (0.14%) among PCR-positive cases. However, relative data accuracy in our retrospective study was made possible by using a standardized dengue clinical care path. Another limitation of this study is the relatively few cases with substantially elevated liver aminotransferase levels. At the same time, since our cohort comprised primarily adults, additional studies in pediatric populations will be useful to confirm our findings.
In patients with DHF/DSS or severe dengue, early diagnosis and proper management may improve outcome in most patients without comorbidities. However, in resource-limited countries, patients with severe disease may present late to the hospital with shock, with or without organ impairment at the time of admission. Our study highlights that early diagnosis and proper management of dengue patients may lead to excellent prognosis without organ injury.
In conclusion, elevated aminotransferase levels were associated with DHF/DSS and severe dengue in our cohort of adult patients with confirmed dengue. However, no threshold values discriminated between DF and DHF or between severe dengue and non-severe dengue.
|
10.1371/journal.pntd.0001150 | Population Structure of Staphylococcus aureus from Remote African Babongo Pygmies | Pandemic community-acquired methicillin-resistant Staphylococcus aureus isolates (CA-MRSA) predominantly encode the Panton-Valentine leukocidin (PVL), which can be associated with severe infections. Reports from non-indigenous Sub-Saharan African populations revealed a high prevalence of PVL-positive isolates. The objective of our study was to investigate the S. aureus carriage among a remote indigenous African population and to determine the molecular characteristics of the isolates, particularly those that were PVL-positive.
Nasal S. aureus carriage and risk factors of colonization were systematically assessed in remote Gabonese Babongo Pygmies. Susceptibility to antibiotics, possession of toxin-encoding genes (i.e., PVL, enterotoxins, and exfoliative toxins), S. aureus protein A (spa) types and multi-locus sequence types (MLST) were determined for each isolate. The carriage rate was 33%. No MRSA was detected, 61.8% of the isolates were susceptible to penicillin. Genes encoding PVL (55.9%), enterotoxin B (20.6%), exfoliative toxin D (11.7%) and the epidermal cell differentiation inhibitor B (11.7%) were highly prevalent. Thirteen spa types were detected and were associated with 10 STs predominated by ST15, ST30, ST72, ST80, and ST88.
The high prevalence of PVL-positive isolates among Babongo Pygmies demands our attention as PVL can be associated with necrotinzing infection and may increase the risk of severe infections in remote Pygmy populations. Many S. aureus isolates from Babongo Pygmies and pandemic CA-MRSA-clones have a common genetic background. Surveillance is needed to control the development of resistance to antibiotic drugs and to assess the impact of the high prevalence of PVL in indigenous populations.
| Staphylococcus aureus is a bacterium that colonizes humans worldwide. The anterior nares are its main ecological niche. Carriers of S. aureus are at a higher risk of developing invasive infections. Few reports indicated a different clonal structure and profile of virulence factors in S. aureus isolates from Sub-Saharan Africa. As there are no data about isolates from remote indigenous African populations, we conducted a cross-sectional survey of S. aureus nasal carriage in Gabonese Babongo Pygmies. The isolates were characterized regarding their susceptibility to antibiotic agents, possession of virulence factors and clonal lineage. While similar carriage rates were found in populations of industrialized countries, isolates that encode the genes for the Panton-Valentine leukocidin (PVL) were clearly more prevalent than in European countries. Of interest, many methicillin-susceptible S. aureus isolates from Babongo Pygmies showed the same genetic background as pandemic methicillin-resistant S. aureus (MRSA) clones. We advocate a surveillance of S. aureus in neglected African populations to control the development of resistance to antibiotic drugs with particular respect to MRSA and to assess the impact of the high prevalence of PVL-positive isolates.
| Methicillin-resistant Staphylococcus aureus (MRSA) has emerged as a community-acquired pathogen in many countries throughout the world (community-acquired MRSA, CA-MRSA). CA-MRSA mostly causes skin or soft-tissue infections as well as deep-seated infections such as necrotizing pneumonia. Predominantly, CA-MRSA encodes the Panton-Valentine Leukocidin (PVL), a S. aureus exotoxin that induces lysis of monocytes and neutrophil granulocytes [1]. In African countries, the occurrence of CA-MRSA has been reported previously from Egypt [2], Mali [3], Algeria [4] and Nigeria [5].
Interestingly, population analysis of global methicillin-susceptible S. aureus (MSSA) isolates associated with PVL have recently indicated that PVL-positive MSSA and MRSA are phylogenetically related based on molecular epidemiological profiles and are dynamically interrelating [6]. Moreover, it was shown, that PVL-positive MSSA are a likely reservoir for the development of PVL-positive MRSA [6] via integration of Staphylococcus cassette chromosome mec (SCCmec) elements including the mecA gene conferring methicillin resistance. Indeed, it is striking that reports from African countries have recently described a high prevalence of PVL-positive MSSA isolates in Nigeria [7] and Mali [3] and have supported the hypothesis that at least one common European MRSA clone associated with PVL (sequence type ST152 according to multilocus sequence typing (MLST)) could originate from African MSSA clones [3]. Interestingly, a study on S. aureus colonization in Wayampi Amerindians in French Guiana revealed a predominance of ST1223 which is highly divergent from other global STs [8]. Ruimy et al. hypothesize, that this association of highly divergent clones in an isolated remote population may reflect the co-evolution of humans and S. aureus as well as human migration [8], [9].
Consequently, we raise the question, whether the “out-of-Africa” hypothesis, as shown for Helicobacter pylori [10], might also be true for PVL-positive S. aureus clones now emerging across the globe. To address this question, we aimed to collect systematically S. aureus isolates independent from the healthcare setting, which is associated with the dissemination of isolates adapted to the specific selection pressure of the hospital environment. Therefore, we performed a cross-sectional S. aureus carrier study among the indigenous Pygmy population in Gabon. One to five percent of the Gabonese population is comprised by Pygmy hunter-gatherers. Almost 50% belong to the Babongo tribe, most of them are living in Waka National Parc, Central Gabon.
Ethical clearance was obtained from our institutional review board (IRB, “Comité d'Éthique Régional Indépendant de Lambaréné”, Lambaréné, Gabon, protocol number: CERIL 15–09). As the majority of Babongos are illiterate and mainly speak the tribal language, we involved a local interpreter to provide detailed information about the study and to obtain a documented oral informed consent. We prepared a short written summary in French that described the information presented to the Pygmies. This document was signed or finger-printed by the participant, the researcher and a witness who spoke French and Babongo. The IRB approved the use of documented oral informed consent.
A cross-sectional survey of S. aureus nasal carriage in Babongo Pygmies was conducted as part of the German-African network on staphylococci and staphylococcal diseases (DFG PAK 296) and took place in the Ikobé region, Central Gabon in November 2009.
All Babongo or mixed Babongo-Bantu inhabitants of the Ikobé region were included if they provided a documented informed consent. Exclusion criteria were (i) infections of nostrils and (ii) a purulent rhinitis. Demographic data (self reported age, height, weight, sex and ethnic group) were recorded for each subject. Travel habits since birth and daily activities were recorded to assess risk factors for S. aureus carriage. Global positioning data of each village were taken by GPS-device (Garmin76 csx).
Nasal swabs were stored in cool boxes and inoculated on SAID agar plates (bioMérieux, Marcy l'Etoile, France) and Columbia blood agar plates in the laboratory facilities of the Medical Research Unit, Lambaréné within four days after sampling. Presumptive S. aureus isolates were identified by colony characteristics, catalase and latex agglutination test (Pastorex Staph-Plus, Bio-Rad Laboratories, Marnes-la-Coquette, France). Species identification and antibiotic susceptibility testing were performed by Vitek 2 automated systems (bioMérieux, Marcy l'Etoile, France). Molecular confirmation of S. aureus and determination of methicillin-resistance were performed as described [11]. To confirm susceptibility to penicillin, a blaZ PCR targeting the S. aureus penicillinase was performed additionally [12].
Panton-Valentine leukocidin (PVL) encoding genes (lukS-PV, lukF-PV) were detected [13]. Staphylococcal pyrogenic toxin superantigens (PTSAgs) were analyzed by detecting toxic shock syndrome toxin (TSST-1) encoding genes (tst) and the enterotoxins (sea, seb, sec, sed, see, seg, seh, sei and sej) [14], [15]. Exfoliative toxins (eta, etb and etd) and genes encoding members of the epidermal cell differentiation inhibitor (edin-A, edin-B and edin-C) were detected by gene amplification [14]–[17].
Capsular polysaccharide types 5 and 8 and accessory gene regulator subtypes (agr I–IV) were identified by multiplex PCR approaches [13], [18].
S. aureus isolates were typed based on sequencing of the hypervariable region of the S. aureus protein A gene (spa), spa types were assigned on the Ridom SpaServer (http://spaserver.ridom.de) curated by the SeqNet.org initiative [19]. Multilocus sequence typing (MLST) was carried out for each isolate [20]. Relatedness in allelic profiles was assessed using eBURST (version 3, http://eburst.mlst.net). To affiliate the S. aureus sequence types (STs) of the Pygmy population to known clonal complexes (CC), we compared our dataset with the whole MLST database of S. aureus using the stringent group definition of a minimum of 6/7 shared alleles.
Proportions of categorical variables were tested using Chi-square test and Fisher's exact test, where appropriate. Odds-ratio and the 95% confidence intervals were calculated to test for associations. The level of significance was α = 5%. All analyses were performed using the software “R” (http://cran.r-project.org, Version: 2.10.1) and package “epicalc”.
Nasal swabs were obtained from 99 Babongo Pygmies and 1 Babongo-Mitsogho living in the Ikobé region, Gabon. Study participants came from five camp-like villages (GPS coordinates in brackets) made up of about six to ten huts: “Village Tranquille” (S1°02.392′; E11°03.744′), “Tsibanga” (S1°02.577′; E11°05.661′), “Ossimba” (S1°02.433′; E11°04.655′), “Ndougou” (S1°02.316′; E11°04.417′), “Soga” (S1°03.020′; E11°10.963′) and “Egouba” (S1°01.638′; E11°08.123′). All villagers who had been in the villages during our visit met the inclusion criteria (n = 103). Three persons from Egouba refused to participate. The age distribution of participants showed a pagoda-shaped population pyramid with 45% of participants being ≤15 years old and 13% being ≥45 years old (Figure 1). Overall, 46% of participants were female. Of all participants, 65% have travelled at least to one of the nine capital cities of Gabon since birth. Twenty-six percent (n = 26) of the population had been hospitalized previously in dispensaries or primary-care hospitals (Table 1).
Culture of 100 nasal swabs identified 34 S. aureus isolates. The carriage rate was 33%. From one participant, two phenotypically different S. aureus isolates were isolated (white colonies with β-hemolysis and yellow-white colonies with faint β-hemolysis). Of all carriers, 42.4% were females (OR = 1.24, 95% CI = 0.49–3.15; p = 0.62). Carriage differed between the five villages ranging from 0% (Ndougou, n = 5 participants) to 45.7% (Village Tranquille, n = 46 participants, Table 1). There was an age-related carriage pattern with a peak colonization of 53.9% in participants between 10 and 20 years of age and a decreasing prevalence in subsequent age groups (Figure 1). No significant associations between S. aureus carriage and any recorded risk factor were detected (not shown).
Of the totality of 34 S. aureus isolates, 64.7% (n = 22) were susceptible to penicillin, blaZ PCR amplicons were only detected in penicillin resistant isolates (n = 12). In addition, 94.1% (n = 32) were susceptible to tetracycline and 88.2% (n = 30) to trimethoprim-sulfamethoxazole. All isolates were susceptible to oxacillin/methicillin, aminoglycosides, fluoroquinolones, macrolides, lincosamides (including inducible clindamycin resistance), nitrofurantoin, fosfomycin, rifampicin and vancomycin. Susceptibility to oxacillin/methicillin was further confirmed by the absence of mecA. Collecting and preparing medicinal herbs was not significantly related to a lower prevalence of antibiotic resistance (penicillin, tetracycline or trimethoprim-sulfamethoxazole) in colonizing S. aureus isolates (OR = 0.23, 95% CI 0.02–1.51, p = 0.078).
We detected 111 toxin encoding genes among 34 S. aureus isolates indicating a high prevalence of toxin co-possession. Table 2 shows the prevalence of the virulence factors tested and assigns these virulence factors to the total number of virulence genes, which were simultaneously detected. Overall, 73.5% (n = 25) of all isolates encoded one or more PTSAgs, with co-possession found in 15 isolates including 15 isolates (44%) that encoded the linked seg-sei loci. The tst gene was not detected but PVL-encoding genes were found in 55.9% (n = 19) of all isolates and were always co-detected with at least one other virulence determinant. The genes seh and etd were always co-detected with PVL-encoding genes (p = 0.238 and p = 0.113 respectively). Other virulence genes were only partially co-detected with PVL: sea (80.0%, p = 0.36), seb (14.3%, p = 0.03), seg/sei (60.0%, p = 0.67), edin-B (75.0%, p = 0.63). PVL was not co-detected with sec (p = 0.03) and eta (p = 0.44).
Among the accessory gene regulator subtypes, agr III was the most prevalent (61.76%, n = 21) followed by agr I and II (17.65%, n = 6 each) and agr IV (2.94%, n = 1, Table 3). Isolates encoding PVL-encoding genes were significantly less associated with agr I (OR = 0; 95% CI: 0–0.52, p = 0.002) but often co-occurred with agr III (OR = 3.1, 95% CI: 0.62–17.29, p = 0.107). sea was significantly associated with agr II (OR = 41.12, 95% CI: 2.74–2679.92, p = 0.002). There was no significant association of all other virulence genes with any agr type.
Detection of capsular polysaccharide (CP) encoding genes revealed a high prevalence of type 8 (CP8, 82.4%, n = 28) followed by type 5 (CP5, 14.7%, n = 5). One isolate was CP gene non-typable (2.9%). Interestingly, the PVL-encoding genes were co-detected with CP5 in 20% (OR = 0.16, 95% CI 0–1.91, p = 0.146), and with CP8 in 64.3% (OR = 8.44, 95% CI 0.79–447.45, p = 0.066). The heterogeneous distribution of CP was also reflected by a significant association of CP5 with agr I (OR = 41.12, 95% CI 2.74–2679.92, p = 0.002) and CP8 with agr III (OR = infinity, 95% CI 2.6–infinity, p = 0.001).
We identified 13 different spa types among 34 S. aureus isolates (Table 4). The most prevalent spa types were t1848 (23.5%, n = 8), t084, t148, t186 and t5941 (each 11.8%, n = 4). One participant carried two phenotypically different S. aureus isolates which had different spa types (t6025, t5941). Three spa types (t5941, t6020, t6025) have not been described before.
Ten different STs were found by MLST showing a Simpson's index of diversity (1-D) of 0.89 (Table 4). Among these, a hitherto unknown ST, designated ST1662 was detected. The most frequent ST was ST30 (23.5%, n = 8), exhibiting the following characteristics: t1848, agr III, CP8 and lukS-PV/lukF-PV-positive. PVL-encoding genes were found in isolates associated with spa types (ST) t1931 (ST1), t311 (ST5), t084 (ST15), t1848 (ST30), t5941 (ST80) and t159 (ST121, Table 4). STs did not cluster in distinct groups according to eBURST analysis. Interestingly, when comparing the STs of this study with the whole MLST database, all the STs of Babongo S. aureus isolates represented the predicted founders of their respective clonal complex, only the novel ST1662 was a singleton.
Detailed inter-village comparison revealed demographic differences between the six camp-like villages (Table 1). The number of participants was imbalanced ranging from two in “Egouba” to 46 in “Village Tranquille” which is the biggest Babongo camp and the residence of the Babongo leader in the Ikobé region. Except for “Egouba”, resistance to penicillin was equally distributed among isolates from different camps. Resistance to tetracyline and trimethoprim-sulfamethoxazole was only found in “Village tranquille” (Table 1). There was no predominance of a single spa type or ST in a certain village. However, the following spa types were only found in one village: t084, t159, t311 (“Village Tranquille”), t570 (“Tsibanga”), t6020 (“Ossimba”) and t127, t6025 (“Soga”, Table 1).
To our knowledge, this is the first investigation of S. aureus isolates from a semi-nomadic indigenous African population. It provides a characterization of susceptibility to antimicrobial drugs, virulence factors and the clonal structure of the isolates. The main findings of our study are the high prevalence of PVL-positive isolates and the same genetic background of Babongo S. aureus isolates as pandemic clones.
Our survey needs to be considered as a representative population-based study, because it covers more than 30% of the Babongo population [21], has a balanced distribution of sex and shows a pagoda-shaped population pyramid typical of a developing community (Figure 1). Due to the semi-nomadic lifestyle of the participants, we cannot give the exact total number of the population in each village, but the total Pygmy population in the study area is estimated to be 300 persons [21]. The S. aureus carriage rate of 33% among Babongo Pygmies is similar to those reported worldwide ranging from 25 to 35% [22], [23]. Carriage corrected for age groups showed the highest colonization in teenagers (54%). This is comparable to the nasopharyngeal carriage rate reported from Europe showing a peak prevalence of over 50% at the age of ten years [24]. The absence of S. aureus in Ndougou is probably due to the small sample size of this village (n = 5, Table 1) and the higher mean age of the participants as carriage declines in older age groups (Figure 1).
Resistance to beta-lactams was rare. Only 35.3% of the isolates were resistant to penicillin, no MRSA was detected. This high prevalence of isolates susceptible to penicillin might be an indirect marker of a limited use of antibiotic agents in this population. However, community-associated MRSA (CA-MRSA) can also emerge in remote populations as shown for Australian Aborigines and North-American Indians [25]–[27]. The toxin gene profile differed clearly from European carrier isolates. Whereas we detected similar rates of sea, sec, seh and eta, the prevalence of genes encoding other superantigens and exfoliative toxins were higher in the Babongo S. aureus isolates compared to isolates originating from Europe: seb (20.6 vs. 3.8%), etd (11.7 vs. 5.2%) and edin-B (11.7 vs. 6.2%) [28]. Surprisingly, genes encoding SED-SEJ and TSST-1, which are common among carrier and clinical isolates in Europe (approx. 7–15% and 15–25% respectively) [14], [28], [29], were not detected in this Pygmy population.
The distribution of the capsule types was biased towards CP8 vs. CP5 (82.4% vs. 14.7%) compared to asymptomatic carriers in Europe (approx. 60–75% vs. 10–35%) [23], [30], [31]. CP5 and CP8 have been shown to impact the virulence of S. aureus and the clinical course of infection [32], [33]. Capsular polysaccharide expression is part of the agr regulon, we showed significant association of CP5 with agr subtype I (p = 0.002) and CP8 with agr subtype II (p = 0.001). This distribution of CP among agr subtypes has also been shown in isolates derived from bovine mastitis [34].
Interestingly, more than 55% of the S. aureus isolated from Babongo Pygmies carried PVL-encoding genes. This prevalence is comparatively high as only 1–2% of clinical methicillin-susceptible S. aureus (MSSA) isolates from Europe are PVL positive [13], [35]. PVL is a bacteriophage-encoded pore-forming toxin, which causes necrosis of tissues and has cytocidal effects on human neutrophils [1]. The clinical role of PVL is not yet fully understood and its role as a virulence factor remains controversial. PVL can be associated with necrotizing pneumonia in humans [36]. A rabbit model of necrotizing pneumonia has clearly demonstrated that PVL both activates polymorphnuclear leucocytes (PMNs) and macrophages and induces necrosis of PMNs [37]. Infected rabbits had the same clinical features of necrotizing pneumonia as described in humans, i. e. lung necrosis, edema, hemoptysis and death [37]. The high prevalence of PVL-positive S. aureus strains could therefore be a risk for Babongo Pygmies to develop necrotizing infections. However, studies with different animal models have shown conflicting results concerning the role of PVL. Some animal studies suggested PVL as major virulence factor in a mouse pneumonia model [38]. Other studies indicate that phenol soluble modulins might enhance the cytolytic effect of PVL [39] or have shown that α-hemolysin (α-toxin) or a point mutation in the agr P2 promoter are responsible for an increased virulence of PVL-positive strains in mice [40], [41]. In addition, other experiments did not find any evidence for PVL as a virulence factor in a murine model [42], [43]. However, it is known that PVL acts differentially on neutrophils of various species as PVL has a strong cytotoxic effect on human neutrophils but not on murine neutrophils [1]. Thus, the impact of a high prevalence of PVL-positive strains in a healthy Babongo population is still unclear. Further prospective studies are needed to analyze if PVL has an impact on the incidence of S. aureus infections in a neglected population.
Noteworthy, the presence of PVL-encoding genes was not associated with one distinct clonal lineage, but was distributed among different STs and spa types. A high prevalence of PVL could be a common feature of Sub-Saharan S. aureus isolates as high prevalence of PVL has been also found in non-pygmy populations from South Africa (100%), Mali (100% in S. aureus ST152), and Nigeria (42.7%) [3], [7], [44]. In contrast to our investigation, all but one of these studies included clinical isolates and might be therefore biased. However, high prevalence of PVL encoding genes is frequently found in pandemic CA-MRSA-clones and certain MSSA lineages (ST1, ST5, ST30, ST80) appear to be a reservoir of CA-MRSA [6]. In our study, we found very common STs (ST1, ST30, and ST121) amongst the Babongo Pygmies, some of them are pandemic clones [6], [45]. This is surprising as the Babongo Pygmies split apart from other humans at least 50,000 years ago and still live in isolated areas [46]. However, it is unclear, whether the same genetic background of Babongo S. aureus isolates and pandemic clones reflects the global spread of distinct clones or if it is the result of separate evolutionary processes in different geographic regions where the same successful clones were independently selected.
As shown for Helicobacter pylori, bacterial polymorphisms may reflect human phylogeography and historical migrations [47]. The genetic diversity in H. pylori decreases with geographic distance from East Africa mirroring the migration of its human host [10]. Interestingly, the Simpson's index of diversity among S. aureus STs from Babongo Pygmies was higher (0.89) than in a comparably remote Amerindian community in French Guiana (0.82), but was still lower than in urban communities in France, Algeria, Moldavia and Cambodia (0.92, 0.93, 0.92, 0.91) [8]. This is possibly due to a higher exchange and transmission between people from different regions and communities in the urban setting. Comparing genetic diversity of S. aureus isolates from isolated population may contribute to the discussion whether S. aureus shows a similar co-evolution and phylogeographical distribution patterns as observed for H. pylori [8], [10]. To confirm this, more population-based carrier studies are needed from different geographic regions to address possible factors of S. aureus transmission between a given isolated population and its neighboring communities.
One limitation of our study is the small sample size which might be only increased if those inhabitants who went hunting in the deep rain forest would have been included. Healthy male subsistence hunter might therefore be underrepresented. In addition, we failed to collect confident data about the use of antimicrobial drugs to analyze its impact on the carriage of resistant isolates. This limitation was due to a poor documentation of antibiotic treatments in the personal health care files and due to difficulty in recalling reliably the intake of antimicrobial agents. Another limitation of our study is the missing data about the incidence of S. aureus-related infections to assess the impact of the high prevalence of PVL on developing invasive disease. To record more confident data about the use of antibiotic drugs for each participant and to survey the incidence of S. aureus infections, future prospective studies are warranted.
In conclusion, our study provides the first insight in S. aureus isolates from an African Pygmy population. While we found a high prevalence of PVL-positive isolates, its impact on the incidence of S. aureus infection in remote populations is not clear yet. Many S. aureus isolates had the same genetic background as pandemic CA-MRSA clones raising the question of a common ancestor. We recommend a close surveillance of S. aureus isolates in remote indigenous African population to control the emergence of resistant isolates and to investigate the role of PVL-positive S. aureus isolates in neglected communities.
|
10.1371/journal.pgen.1006858 | Lack of a peroxiredoxin suppresses the lethality of cells devoid of electron donors by channelling electrons to oxidized ribonucleotide reductase | The thioredoxin and glutaredoxin pathways are responsible of recycling several enzymes which undergo intramolecular disulfide bond formation as part of their catalytic cycles such as the peroxide scavengers peroxiredoxins or the enzyme ribonucleotide reductase (RNR). RNR, the rate-limiting enzyme of deoxyribonucleotide synthesis, is an essential enzyme relying on these electron flow cascades for recycling. RNR is tightly regulated in a cell cycle-dependent manner at different levels, but little is known about the participation of electron donors in such regulation. Here, we show that cytosolic thioredoxins Trx1 and Trx3 are the primary electron donors for RNR in fission yeast. Unexpectedly, trx1 transcript and Trx1 protein levels are up-regulated in a G1-to-S phase-dependent manner, indicating that the supply of electron donors is also cell cycle-regulated. Indeed, genetic depletion of thioredoxins triggers a DNA replication checkpoint ruled by Rad3 and Cds1, with the final goal of up-regulating transcription of S phase genes and constitutive RNR synthesis. Regarding the thioredoxin and glutaredoxin cascades, one combination of gene deletions is synthetic lethal in fission yeast: cells lacking both thioredoxin reductase and cytosolic dithiol glutaredoxin. We have isolated a suppressor of this lethal phenotype: a mutation at the Tpx1-coding gene, leading to a frame shift and a loss-of-function of Tpx1, the main client of electron donors. We propose that in a mutant strain compromised in reducing equivalents, the absence of an abundant and competitive substrate such as the peroxiredoxin Tpx1 has been selected as a lethality suppressor to favor RNR function at the expense of the non-essential peroxide scavenging function, to allow DNA synthesis and cell growth.
| The essential enzyme ribonucleotide reductase (RNR), the rate-limiting enzyme of deoxyribonucleotide synthesis, relies on the thioredoxin and glutaredoxin electron flow cascades for recycling. RNR is tightly regulated in a cell cycle-dependent manner at different levels. Here, we show that cytosolic thioredoxin Trx1 is the primary electron donor for RNR in fission yeast, and that trx1 transcript and protein levels are up-regulated at G1-to-S phase transition. Genetic depletion of thioredoxins triggers the DNA replication checkpoint up-regulating RNR synthesis. Furthermore, deletion of the genes coding for thioredoxin reductase and dithiol glutaredoxin is synthetic lethal, and we show that a loss-of-function mutation at the peroxiredoxin Tpx1-coding gene acts as a genetic suppressor. We propose that in a mutant strain compromised in reducing equivalents, the absence of an abundant and competitive substrate of redoxins, the peroxiredoxin Tpx1, has been selected as a lethality suppressor to favor channeling of electrons to the essential RNR.
| Cysteine residues are not very abundant in proteins, but they are over-represented in functional regions of proteins, such as surfaces and catalytic centers [1]. The thiol group of cysteines is subject of post-translational modifications altering its redox state; several of these oxidation states are reversible, such as sulfenic acid and disulfides. In particular, reversible thiol to disulfide switches happen as a consequence of cellular responses to oxidative stress, and several proteins with reactive cysteine residues undergo oxidations as part of their catalytic cycles (for a review, see [2]). Cells are provided with two major systems meant to control the thiol-disulfide status, the thioredoxin (Trx) and the glutaredoxin/glutathione (Grx/GSH) systems.
Trxs and Grxs catalyze thiol-disulfide exchange reactions, and share a motif known as the Trx fold [3]. Thermodynamically, both types of reductants use as the ultimate electron donor NADPH [4]. Electrons are therefore transferred from NADPH to final substrates through gradients in redox potentials. In the case of Trxs, Trx reductase is the intermediate between NADPH and Trx, while GSH reduces oxidized Grxs, GSH reductase being the link between NADPH and oxidized GSH.
Trx was first identified in 1964 as an electron donor for Escherichia coli ribonucleotide reductase (RNR), an enzyme required for DNA synthesis [5]. Grx was later discovered as an alternative electron donor for the same enzyme in E. coli mutants lacking Trx [6]. Many reports indicate that there is cross-talk between both branches of these electron transfer systems and certain redundancy, but it is also clear that there is substrate specificity.
From then onwards, it became clear that these oxido-reductases regulate a wide number of processes in eukaryotic and prokaryotic organisms, apart from DNA synthesis and repair, including antioxidant defense and redox regulation, sulfur metabolism or apoptosis; the substrates of Trxs and Grxs mediating these effects are peroxiredoxins (Prxs), GSH peroxidases, methionine sulfoxide reductases, phosphoadenylyl sulfate (PAPS) reductase or RNRs (for reviews on these and other functions of the electron donor cascades, see [2,7–12]). In most cell types, the only substrate of electron donors which is essential for survival (and not only for specific cellular processes such as cysteine biosynthesis or oxidative stress tolerance) is RNR.
RNR catalyzes the reduction of ribonucleosides into deoxyribonucleosides, and is therefore essential to provide the building blocks, deoxyribonucleotides (dNTPs), during DNA replication and repair. In eukaryotes, class Ia RNRs are composed of a large subunit, α, containing the catalytic site and two allosteric effector binding sites, that control which substrate is reduced (specificity site) as well as the rate of reduction (activity site) [13,14], and a small subunit, β, containing a stable diferric-tyrosyl radical cofactor (oxygen is required for the assembly of the diferric-tyrosyl radical cofactor in RNRs), which initiates nucleotide reduction through the transient oxidation of a cysteine to a thiyl radical in the catalytic site of the α subunit. During this process, two local cysteines in the large subunit provide reducing equivalents, and the disulfide bond generated between them, after isomerizing within the same α monomer towards a solvent-exposed position, is reduced by Trx or Grx to yield active RNR.
Balanced and sufficient pools of dNTPs have to be present during S phase of the cell cycle, and also to assist in DNA repair. In fact, several studies suggest that a correct supply of dNTPs during DNA replication is important for genome stability and for the prevention of cancer [15,16]. Inhibition of RNR activity by the radical scavenger hydroxyurea (HU) and other compounds has been used as a chemotherapeutic strategy of numerous cancer types [16,17]. RNRs are tightly regulated through many different mechanisms, which include allosteric and oligomeric regulation, transcription of the α and/or β-coding genes to modulate protein levels, inhibition of RNR catalytic activity and regulation of the subcellular localization of the RNR subunits (for reviews on RNR regulation, see [18,19]). In Schizosaccharomyces pombe, Cdc22 and Suc22 are the large and small subunits of RNR, respectively [20]. Most studies concerning regulation of fission yeast RNR activity have centered on the RNR inhibitor Spd1, which affects activity and subunit localization of α and β [21–23], and on the up-regulation of cdc22 transcription during the S phase and after DNA damage (for a review, see [19]). While the suc22 transcript does not fluctuate with the cell cycle, transcription of cdc22 is up-regulated by the MBF transcription factor, which triggers expression of genes required for the S phase [20,24,25]. Regarding regulation of cdc22 expression by checkpoint kinases under stress conditions, treatment with HU, which inhibits RNR, decreases the available pool of dNTPs and causes the formation of stalled replication forks and the activation of the DNA replication checkpoint driven by the Rad3 and Cds1 kinases; activated Cds1 phosphorylates and inactivates the Yox1 transcriptional repressor, promoting MBF-dependent cdc22 expression [26]. Regarding the regulation of RNR by cofactors and post-transcriptional modifications, changes in RNR subunit localization in response to iron bioavailability have been recently demonstrated in budding yeast [27].
Nevertheless, very little is known about the redox-dependent cell cycle regulation of RNR activity [28]. In fact, the identity of the S. pombe electron transfer components required for RNR recycling is unknown. Here, we identify Trx1 and Trx3 as the main electron donors of fission yeast RNR, we demonstrate that Trx1 expression is actually up-regulated at S phase at the transcript and protein levels, and that in the absence of Trx1 and Trx3 the DNA replication checkpoint is activated. With the expectation that a complete block of electrons flow should drive to cell lethality by blocking RNR at its oxidized form unless a continuous synthesis of RNR is triggered, we have managed to generate a synthetic lethal combination by deleting the Trx reductase and the Grx1-coding genes. A spontaneous suppressor of this synthetically lethal phenotype is a frame-shift mutation at the beginning of tpx1, the gene coding for the most abundant consumer of electrons in the cell, the Prx Tpx1. Our experiments suggest that in the triple knockout strain Δtrr1 Δgrx1 Δtpx1, the elimination of the main sink of electron favors the reduction of the essential substrate RNR.
S. pombe contains three genes coding for Trxs [29] and one for Trx reductase, trr1 (Fig 1A). Trx1 is the main cytoplasmic Trx [30]; Trx2 is localized to the mitochondria [31]; and Trx3/Txl1 has cytoplasm localization, although it is also associated with the proteasome [32–34]. Trx1 and, to a minor extent, Trx3 are the electron donors of the Prx Tpx1, essential for aerobic scavenging of peroxides and for signal transduction [35–37]. Thus, while cells lacking Trx1 are extremely sensitive to hydrogen peroxide (H2O2), Trx2 and Trx3 appear to be dispensable for the defense against oxidative stress [35]. Regarding the other branch of the disulfide reductases, fission yeast expresses only two dithiol Grxs, cytosolic Grx1 and mitochondrial Grx2, several monothiol Grxs such as Grx4 (involved in the iron starvation response) [38–40], endoplasmic reticulum-located Grx3 [41] and mitochondrial Grx5 (involved in mitochondrial iron-sulfur cluster assembly) [42], and one GSH reductase, Pgr1 (Fig 1A). Pgr1 has been reported to be essential at least during aerobic growth [43], but cells lacking the reductase can grow under semi-anaerobic conditions. These two cascades have to recycle enzymes which suffer disulfide formation as part of their catalytic functions, such as the essential Cdc22 and the non-essential Tpx1, Mxr1 or Met16 (Fig 1A).
To test the role of these cascades in the turnover of the large RNR subunit, Cdc22, we combined single or multiple deletion mutations with the expression of a tagged version of the protein, Cdc22-HA. This modification, performed at the chromosomal locus, did not affect cell fitness or cell tolerance to the RNR inhibitor HU (S1 Fig). As shown in Fig 1B, a DTT sensitive, slow migrating band was detected using anti-HA immune-blotting of extracts from asynchronous cultures of wild-type cells expressing Cdc22-HA, corresponding to 10.2 ± 1.6% of total Cdc22; the sensitivity to the dithiol DTT indicates that the slower migrating band corresponds to a disulfide-containing RNR. In extracts from cells lacking Trx reductase, a band between oxidized and reduced Cdc22 was detected, probably a transient intermediate between RNR and its electron donor. Importantly, cells lacking cytosolic Trx1, but not the mitochondrial Trx2, displayed 35 ± 2.7% of Cdc22 in its oxidized form. The lack of Trx3 did not significantly affect the amount of Cdc22 disulfide form, but it enhanced the ratio of oxidized-to-reduced form (50 ± 3.5%) in a Δtrx1 background. On the contrary, disruption of the Grx branch by deletion of the genes coding for Grx1-to-Grx5, or GSH reductase, Pgr1, did not exacerbate disulfide accumulation, unless added to the Δtrx1 Δtrx3 background (Fig 1C). The percentages of Cdc22 oxidation in these and other mutants of the Trx and Grx branches are indicated in Fig 1D.
To confirm that Trx deficiencies have an effect on Cdc22 activity, we measured dNTP levels of asynchronous wild-type, Δtrx1 and Δtrx1 Δtrx3 cultures (Fig 1E for dGTP, and S2A Fig for dATP), and detected small but significant decreases in the absence of Trxs. We also measured the percentage of cells at G1, S and G2 phases in asynchronous cultures from these cells. As shown in Fig 1F, cells lacking Trx1 and, to a larger extent, Trx1 and Trx3 displayed an enlarged population of cells at S phase, indicating that these strains completed DNA replication slower than wild-type cells.
We synchronized cultures of cells expressing Cdc22-HA and carrying mutations in several components of the reducing cascades by means of the cdc25-22 allele. Cdc25 is the G2-to-M activating phosphatase of the cyclin-dependent kinase of S. pombe, Cdc2. Upon shift to the non-permissive temperature, cdc25-22 cells are arrested at the G2/M transition and after dropping the temperature, cells are synchronically released from the arrest. As shown in Fig 2A, Cdc22-HA oxidation cycles in wild-type cells and in cells lacking Trx3, with a transient peak from 60 to 120 minutes after the release. This peak overlaps with that of the septation index (Fig 2B), which in fission yeast is concomitant with S phase. The peaks of Cdc22-HA oxidation and septation index are delayed and more sustained in cells lacking Trx1 (or Trx1 and Grx1). Strikingly, in cells devoid of cytosolic Trxs, Δtrx1 Δtrx3, Cdc22-HA oxidation does not cycle and the protein is maintained at its oxidized form at high levels, around 50–60%. In fact, these cells do not have a clear septation peak.
Next, and to confirm that the activity of Cdc22 was compromised in the mutants lacking Trxs, we measured dNTP levels in the previous synchronous cultures. As shown in Fig 2C (dGTP) and S2B Fig (dATP), while the levels of dNTPs increased in wild-type cells during cell cycle progression (60 and 100 min after release), cells lacking Trx1 or Trx1 and Trx3 displayed a reduction of dNTPs levels, pointing that these strains could have compromised DNA synthesis. In fact, when we analyzed the DNA content from the synchronous cultures, we indeed observed a delayed and extended S phase in Δtrx1 cells (Fig 2D). This is even more noticeable in cells in which all the cytosolic Trxs were absent: in Δtrx1 Δtrx3 S phase was not detected by FACS until 120 min after the release, which represents 60 minutes of delay when compared to wild-type cells.
Once confirmed that Cdc22 oxidation is exacerbated during catalysis, we tested whether the expression of its main electron donor was also up-regulated during S phase. As shown in Fig 3A, Trx1 protein levels were enhanced at S phase as determined in extracts from block and release experiments. To test whether this protein up-regulation was dependent on a transcriptional event, we used Cyclebase, a repository of published cell cycle experiments [44], to interrogate genome-wide studies on block and release experiments performed in fission yeast. As shown in Fig 3B, there is a small but consistent cell cycle regulation of trx1 mRNA. The combined peaktime of all published datasets occurs at the beginning of G1 (Fig 3C), a bit later than other S phase transcripts such as cdc22, cdc18, yox1 or nrm1 (S3 Fig). All these genes are up-regulated by the MBF transcription factor. To test whether the increase of trx1 mRNA is dependent on MBF, we analyzed its transcript levels upon HU treatment or in cells lacking the MBF repressor Yox1. As shown in Fig 3D, trx1 transcription does not seem to depend on the MBF complex, contrary to cdc22. Future work will help us elucidating who triggers the accumulation of trx1 mRNA at G1-to-S transition.
Inhibition of RNR activity by HU treatment triggers a DNA replication stress, probably through the decrease in dNTP concentrations and stalling of DNA polymerase at replication forks. In S. pombe, the DNA replication checkpoint is driven by the Rad3 and Cds1 kinases. One target of this cascade is the transcriptional repressor Yox1, which after phosphorylation by Cds1 is released from the MBF complex and its S phase promoters [26,45] (Fig 4A).
To test whether RNR inhibition by Trx deficiency can trigger replication stress, we first tested whether cells lacking Trx1 and Trx3 are sensitive to the presence of the RNR inhibitor HU. As shown in Fig 4B, Δtrx1 and Δtrx1 Δtrx3 cells are moderately and severely sensitive to HU, respectively, highlighting their defects in DNA synthesis. In wild-type cells, HU treatment exacerbates the accumulation of total and oxidized Cdc22-HA, as well as of the inhibitory phosphorylation of Yox1 (Fig 4C). Both Trx mutant strains, but specially Δtrx1 Δtrx3, display constitutive activation of the Rad3-Cds1 checkpoint cascade, as demonstrated by the presence of phosphorylated Yox1 even in the absence of HU stress in this strain background (Fig 4C), and by the enhanced levels of cdc22 mRNA under basal conditions (Fig 4D). We generated a Δtrx1 Δcds1 strain to demonstrate that the weak phosphorylation of Yox1 in Δtrx1 cells is dependent on Cds1 (Fig 4E). Δtrx1 Δtrx3 deletions are synthetic lethal with deletion of cds1, while a triple Δtrx1 Δtrx3 Δchk1 mutant is viable (S4 Fig); Chk1 is the effector kinase of the DNA damage checkpoint. We propose that the survival of cells lacking both Trxs depends at least partially on the Cds1-dependent transcriptional up-regulation of the cdc22 gene (Fig 4A).
So far, we have demonstrated that Trxs are the main electron donors of Cdc22, and that cells lacking Trx1 and Trx3 have important defects and activate the replication checkpoint. Taking into account that RNR is an essential protein, we attempted to induce lethality by combining a number of deletions of genes coding for components of the electron pathway cascades (see Fig 1A). Many of the mutants displayed severe growth defects, which could often be rescued by growing the cells in semi-anaerobiosis or in the presence of exogenous GSH (S1 Table, Fig 5A). Indeed, exogenous addition of GSH was sufficient to decrease the ratio of oxidized-to-reduced Cdc22 in mutants lacking Trxs (Fig 5B and 5C) and to alleviate some of their growth defects in liquid media (Fig 5D). This suggests that the Grx-GSH branch is a back-up mechanism of reduction of the essential RNR.
After exhaustive combination of gene deletions, only two crosses lead to lethality in fission yeast: double deletions of the trr1 (coding for Trx reductase) and gcs1 (codes for glutamate-cysteine ligase, the rate-limiting enzyme on the GSH biosynthetic pathway) genes, or the double knock-out trr1 and grx1 (coding for the only dithiol cytosolic Grx) (Fig 5E). We propose that in this Δtrr1 Δgrx1 strain background RNR would remain oxidized.
In spite of the results shown above, and using random spore selection, we unexpectedly obtained a single colony lacking Trr1 and Grx1 and therefore containing a suppressor mutation. To our surprise, we determined by sequence analysis that this mutation laid on the gene encoding the Prx Tpx1, introducing a one-base deletion at the 26th codon of the open reading frame and subsequently a frame shift and a stop codon at position 71 (Fig 6A). To confirm that the suppressor mutation was linked to loss-of-function of Tpx1, we performed tetrad analysis to select a triple Δtrr1 Δgrx1 Δtpx1 knock-out strain. As shown in Fig 6B by tetrad dissection, the double Δtrr1 Δgrx1 is synthetic lethal, while tiny colonies of the Δtrr1 Δgrx1 Δtpx1 strain were isolated under semi-anaerobic conditions and could be recovered on plates containing GSH. As shown in S5 Fig, the growth of this triple delete, Δtrr1 Δgrx1 Δtpx1, displays severe growth defects even under semi-anaerobic conditions, which can be partially overcome by GSH addition.
Tpx1 is probably the most demanding substrate of electron donors: there are more than 400,000 copies of the protein per cell [46], and Tpx1 is continuously catalyzing H2O2 detoxification during aerobic growth with the participation of Trx1, Trx3 and, probably, Grx1 [35,36]. The fact that strains such as Δtrx1 Δtrx3 Δgrx1 grow better under semi-anaerobic condition (Fig 5A) is an indication that Tpx1 may be competing with Cdc22 for reducing equivalents in cells devoid of the main cytosolic electron donors: the levels of peroxides during semi-anaerobic metabolism are lower than in the presence of oxygen, and therefore Tpx1 is not cycling and demanding electrons to the same extent.
To demonstrate that the absence of Tpx1 could positively impinge on the reduced-to-oxidized ratio of RNR, we measured the amount of oxidized and reduced Cdc22-HA in different strains expressing or not Tpx1. As shown in Fig 6C and 6D, deletion of tpx1 always reduced the percentage of oxidized Cdc22 in three different Trx-deficient strains. We conclude that tpx1 deletion allows the channeling of electrons into the disulfide-bonded RNR, and this is particularly relevant in redoxin mutants.
To test whether the competition between Tpx1 and Cdc22 for reducing equivalents could occur in a wild-type background, we forced temporal depletion of reduced Trx1 by Tpx1 during S phase. Exhaustion of reduced Trx1 by Tpx1 can only be accomplished when the Prx is actively scavenging peroxides, but an excess of H2O2 triggers Tpx1 over-oxidation and avoids Trx1 depletion [35]. Therefore, we applied mild oxidative stress in a continuous manner to S phase cultures, by synchronizing wild-type cells expressing Cdc22-HA using the cdc25-22 allele as shown in Fig 2, and adding or not 100 μM H2O2 at the onset of S phase, with subsequent additions of 25 μM every five minutes, to force Tpx1 oxidation and Trx1-dependent recycling (Fig 7). Trx1 oxidation was followed in extracts prepared in the presence of 4-acetamido-4′-maleimidylstilbene-2,2′-disulfonic acid (AMS) as described before [35]. AMS is a bulky thiol alkylating agent: while three moieties of AMS are incorporated in Trx1 when it is in the reduced form, only one AMS is incorporated when Trx1 is oxidized and two of its cysteine residues form a disulfide; slower migrating bands, corresponding to the transient mixed disulfides between Trx1 and its substrates, can also be detected by Western blot upon Trx1 oxidation. As shown in Fig 7A, the S phase-dependent oxidation of Cdc22 does not cause an apparent consumption of reducing equivalents, since the majority of Trx1 remains in the reduced form during the whole cycle. When a continuous addition of mild H2O2 is applied starting at 60 min (at the onset of S phase), a sustained oxidation of Trx1 is accomplished (Fig 7B), which is fully dependent on peroxide scavenging by Tpx1 [35]. Importantly, this Tpx1-dependent depletion of reduced Trx1 enhances the amount of oxidized Cdc22 (from 12% to 25%; Fig 7C), and the disulfide form is maintained for a longer period that in the absence of peroxides (Fig 7A, 7B and 7C). A small but significant cell cycle delay can be observed as a consequence of an elongated S phase, as demonstrated with the septation index (Fig 7D). In conclusion, if oxidative stress emerges during S phase, Tpx1 enzymatic activity jeopardizes the RNR-dependent synthesis of dNTPs through depletion of reduced Trx1.
Balanced pools of dNTPs have to be accumulated during S phase and after DNA damage and replication stress, and these DNA building blocks are synthesized on demand. In these two scenarios, replication and checkpoint activation, RNR activity is up-regulated through several different mechanisms. We have shown here that the catalytic disulfide formed at the large subunit of RNR, Cdc22, is reduced by the main cytosolic Trx, Trx1. Three important conclusions can be extracted from our work: first, the mRNA and protein levels of Trx1 are up-regulated during S phase, what demonstrates a new layer of regulation of RNR. Second, the other cytosolic Trx, Trx3, may support Trx1 in RNR recycling, so that cells lacking both electron donors suffer from severe replication stress which is partially overcome by the activation of the Rad3-Cds1 checkpoint. Third, the fitness phenotypes of mutants defective in electron donor capacity can be partially alleviated by depletion of another substrate of Trx1, the Prx Tpx1; elimination of an abundant competitor funnels electrons towards the essential RNR.
Regarding the cell cycle-dependent regulation of Trx1, all the experiments performed so far with synchronized S. pombe cultures highlight the smooth but consistent waves of trx1 transcripts, with a G1 peaktime (http://cyclebase2.jensenlab.org/) (Fig 3B). We have discarded the participation of the main transcriptional activator of G1-S phase genes, the MBF complex, in trx1 cycling (Fig 3D). Further work will be required to characterize this cell cycle-regulated event. Interestingly, it has recently been reported that colorectal cancer tissues display enhanced protein levels of both RNR and Trx1, and that inhibition of both proteins simultaneously produced a synergistic anti-proliferation effect in this model [47].
To the best of our knowledge, this is the first report demonstrating that eukaryotic cells carrying Trx deficiencies suffer from replication stress and constitutively trigger the DNA replication checkpoint. In E. coli, an interesting connection between electron donor supplies, activation of DNA replication by DnaA and transcription up-regulation of RNR was proposed by the group of Beckwith [48]. In Saccharomyces cerevisiae, it has been published that mutants lacking Trx1 and Trx2 display a longer S phase [49], that the total pool of dNTPs from asynchronous cultures is unaffected thanks to the novo synthesized RNR [50], but that dNTP levels of cells synchronized in S phase are significantly lower than those of wild-type cells [51]. In view of our results, this is probably due to a reduced pool of dNTPs to assist on DNA synthesis [28]. We show here that the checkpoint kinases Rad3 and Cds1 are constitutively active in Δtrx1 Δtrx3 cells, and that this activation is required for the survival of this strain, since the triple Δtrx1 Δtrx3 Δcds1 combination is lethal. We propose that in Δtrx1 Δtrx3 cells the main source of active/reduced Cdc22 is de novo synthesized protein, which arises from the constitutive up-regulation of cdc22 transcription in a Rad3-Cds1-Yox1 dependent manner.
The complete lack of Cdc22 recycling should drive cells to lethality. With this idea in mind, we have generated an extensive combination of deletion mutants in most components of the Trx and Grx branches, and one of these combinations resulted in synthetic lethality: Δtrr1 Δgrx1. The reason why other mutants, such as the quadruple Δtrx1 Δtrx2 Δtrx3 Δgrx1 and Δtrx1 Δtrx3 Δgrx1 Δgrx2 strains, could be isolated under semi-anaerobic conditions but not the aforementioned Δtrr1 Δgrx1 strain is still intriguing to us. It can be speculated that the lack of Trx reductase is more pervasive that the elimination of its substrates due to the accumulation of oxidized Trxs, which may invert their role towards thiol oxidases [52,53], or which may bind to Trx substrates with the same affinity as reduced Trx [54] and block their reduction by other electron donors. A similar result has been reported in other organisms such as E. coli, where lethality or severe sickness can only be accomplished by deletion of the Trx reductase coding gene in combination with a defect in the Grx branch [55]. In S. cerevisiae, deletion of the trx1, trx2, grx1 and grx2 genes has been reported to be lethal [56].
Prxs are probably the most demanding cellular substrates of electron donors: they are continuously catalyzing peroxide scavenging during aerobic metabolism, and they are among the most abundant proteins in most cell types. We reported before that Trx1 and, secondarily, Trx3 are the main electron donors of Tpx1, and cells lacking both cytoplasmic Trxs (Trx1 and Trx3) display constitutively oxidized Tpx1 [35]. Cells lacking Tpx1 cannot grow aerobically on plates; however, strain Δtrx1 Δtrx3 is still viable aerobically, suggesting a secondary role for the Grx/GSH system in Tpx1 reduction. Therefore, Tpx1 and Cdc22 compete for Trx1, Trx3 and, probably, another component(s) of the Grx/GSH cascade. This is not a problem in a wild-type background under most conditions: reduction of Tpx1 by Trx1 is hardly saturated, unless mild oxidative stress is applied, and only for a limited amount of time unless a continuous supply of peroxides is provided (Fig 7A and 7B); indeed, upon severe H2O2 stress, Tpx1 becomes hyper-oxidized to sulfinic acid and temporarily inactivated [35,37,57], which may be beneficial to avoid inhibition of RNR recycling. However, when electron donors become limiting by genetic interventions it is probably advantageous to promote reduction of an essential substrate, RNR, by eliminating a non essential one, a Prx. In fact, other processes improving the fitness of redoxin mutants as well as the oxidized-to-reduced ratio of RNR are semi-anaerobic growth (by minimizing the activity and electron consumption of Tpx1 in peroxide scavenging; Fig 5A) and GSH addition (by providing unlimited reducing power; Fig 5A and S5 Fig).
In our study, we present evidence for the existence of a novel electron donor for RNR, as the synthetic lethal phenotype of Δtrr1 Δgrx1 mutant can be rescued by eliminating Tpx1, the major competitor substrate for electrons. It was similarly proposed by Grant and colleagues that PAPS reductase could have an alternative hydrogen donor to Trx1 and Trx2 in budding yeast, since a Δtrx1 Δtrx2 strain grew on minimal media without sulphate under low-aeration growth conditions reducing the generation of reactive oxygen species [56], and probably minimizing the function of Prxs or GSH peroxidases. We have not identified yet the alternative electron donor of RNR in the Δtrr1 Δgrx1 Δtpx1 background. There are still two or three genes in fission yeast coding for monothiol glutaredoxins (Grx3, Grx4, Grx5), which have been proposed to participate in processes other than disulfide reduction. At least in mammalian RNR, a GSH-mixed disulfide mechanism for Grx-mediated reduction of RNR has been described [58]. Whether S. pombe monothiol Grxs are mediating the channeling of electrons to RNR in a GSH-dependent manner, or whether GSH itself can reduce the disulfide in Cdc22 will have to be elucidated.
Cells were grown in rich medium (YE) at 30°C as described previously [59]. When cells were crossed, we chose tetrad dissection or random spore analysis as indicated in the text. For tetrad analysis, asci were dissected by micromanipulation with a Singer Micromanipulator MSM 400 (Singer Instruments, UK). After growth of the dissected spores on YE agar plates under semi-anaerobic conditions, genetic markers were scored by replica-plating on YE-agar plates containing or not the indicated antibiotics, and placing the plates at 30°C under semi-anaerobic conditions in the presence or not of 2 mM GSH, as indicated. Anaerobic liquid cultures were grown in flasks filled to the top with medium at 30°C without shaking. Origins and genotypes of strains used in this study are outlined in Appendix S2 Table, and most of them were constructed by standard genetic methods. A strain with tagged Cdc22-HA, SB104, was constructed by replacing the cdc22-YFP::kanMX6 cassette of strain AWS16 (h+ cdc22-YFP::kanMX6 ade6-704 leu1-32 ura4-D18, kindly provided by A. Carr), with a cdc22-HA::natMX6 cassette and by cleaning the auxotrophies. The natMX6 cassette in SB104 was replaced by the hphMX6 cassette, resulting in strain SB110. Derived strains containing additional deletions were obtained by crossing SB104 or SB110 with the corresponding strains, and plating spores in appropriate media, with the exception of strain AD104 that was obtained by deletion of the pgr1 gene in SB110. Strains with tagged Yox1-13Myc were obtained by crossing appropriate strains with JA778 (h- yox1-13myc::kanMX6 ura4-D18) or JA779 (h+ yox1-13myc::kanMX6 ura4-D18).
Modified trichloroacetic acid (TCA) extracts were prepared blocking thiols with either iodoacetamide or AMS and separated in non-reducing denaturing electrophoresis as previously described [37]. Only when indicated, the reducing agent dithiothreitol (DTT) was added to the sample buffer prior to electrophoresis. Cdc22-HA and Yox1-Myc were immuno-detected with monoclonal house-made anti-HA or anti-Myc antibodies, respectively. Trx1 was immuno-detected with anti-Trx1 polyclonal antibody [60]. Anti-Sty1 polyclonal antibody [36] was used as loading control. Relative quantification of protein levels in Western blots was performed by scanning membranes with a Licor 3600 CDigit Blot Scanner (Licor Inc., USA) and using the Image Studio 4.0 software.
The dATP and dGTP levels were determined by the DNA polymerase-based enzymatic assay as described before [27]. In brief, the incorporation of dATP and dGTP into specific oligonucleotides, containing poly(AAAT) and poly(AAAC) sequences respectively, by the Klenow DNA polymerase was determined in the presence of excess [3H]-labeled dTTP.
We followed a previously published protocol for determining DNA content on isolated nuclei [61]. Briefly, 1x107 cells were fixed in 70% ethanol and nuclei were prepared. Isolated nuclei were treated with RNase A (37°C overnight) and DNA was stained in PBS solution containing 1 μM Sytox green.
Temperature-sensitive strains carrying the allele cdc25-22 were cultured in YE at the permissive temperature (25°C) in a shaker water bath until reaching OD600 of 0.3, shifted to the non-permissive temperature (36°C) for 4 hours and then allowed to resume the cell cycle by growing them at 25°C during 3 hours as described [62]. Full arrest at G2/M was checked by microscopy. 5 ml aliquots were taken from non-arrested cells and at different times after release to prepare TCA extracts. Cell cycle progression was monitored with fluorescence microscopy by measuring the septation index of calcofluor-stained cells and by flow cytometry.
Total RNA from S. pombe YE cultures was obtained, processed and transferred to membranes as described previously [63]. Membranes were hybridized with the [α-32P]dCTP-labelled cdc22, trx1 and act1 probes, containing the complete open reading frames.
For survival on solid plates, S. pombe strains were grown, diluted and spotted on YE plates containing or not HU at the indicated concentrations and plates were incubated at 30°C for 2–3 days as previously described [26]. To study the survival of strains on solid plates under aerobic or semi-anaerobic conditions, S. pombe strains were grown, diluted and spotted in YE, and plates were incubated at 30°C under aerobic or semi-anaerobic conditions. To grow cells in solid media in an semi-anaerobic environment, we incubated the plates at 30°C in a tightly sealed plastic bag containing a water-activated Anaerocult A sachet (Merck, Darmstadt, Germany) [36], or alternatively in a nitrogen-filled anaerobic chamber (Forma Anaerobic System, Thermo Electron Corp.). When indicated 2 mM GSH was added to YE agar plates.
Yeast cells were grown in YE from an initial OD600 of 0.2, with or without the addition of 2 mM GSH, using an assay based on automatic measurements of optical densities, as previously described [64].
|
10.1371/journal.pgen.1005889 | Determinants of Chromosome Architecture: Insulator Pairing in cis and in trans | The chromosomes of multicellular animals are organized into a series of topologically independent looped domains. This domain organization is critical for the proper utilization and propagation of the genetic information encoded by the chromosome. A special set of architectural elements, called boundaries or insulators, are responsible both for subdividing the chromatin into discrete domains and for determining the topological organization of these domains. Central to the architectural functions of insulators are homologous and heterologous insulator:insulator pairing interactions. The former (pairing between copies of the same insulator) dictates the process of homolog alignment and pairing in trans, while the latter (pairing between different insulators) defines the topology of looped domains in cis. To elucidate the principles governing these architectural functions, we use two insulators, Homie and Nhomie, that flank the Drosophila even skipped locus. We show that homologous insulator interactions in trans, between Homie on one homolog and Homie on the other, or between Nhomie on one homolog and Nhomie on the other, mediate transvection. Critically, these homologous insulator:insulator interactions are orientation-dependent. Consistent with a role in the alignment and pairing of homologs, self-pairing in trans is head-to-head. Head-to-head self-interactions in cis have been reported for other fly insulators, suggesting that this is a general principle of self-pairing. Homie and Nhomie not only pair with themselves, but with each other. Heterologous Homie-Nhomie interactions occur in cis, and we show that they serve to delimit a looped chromosomal domain that contains the even skipped transcription unit and its associated enhancers. The topology of this loop is defined by the heterologous pairing properties of Homie and Nhomie. Instead of being head-to-head, which would generate a circular loop, Homie-Nhomie pairing is head-to-tail. Head-to-tail pairing in cis generates a stem-loop, a configuration much like that observed in classical lampbrush chromosomes. These pairing principles provide a mechanistic underpinning for the observed topologies within and between chromosomes.
| The chromosomes of multicellular animals are organized into a series of topologically independent looped domains. This domain organization is critical for the proper utilization and propagation of the genetic information encoded by the chromosome. A special set of architectural elements, called boundaries or insulators, are responsible for both subdividing the chromatin fiber into discrete domains, and determining the topological organization of these domains. Central to the architectural functions of insulators are heterologous and homologous insulator:insulator pairing interactions. In Drosophila, the former defines the topology of individual looped domains in cis, while the latter dictates the process of homolog alignment and pairing in trans. Here we use two insulators from the even skipped locus to elucidate the principles governing these two architectural functions. These principles align with several longstanding observations, and resolve a number of conundrums regarding chromosome topology and function.
| The highly regular and reproducible physical organization of chromosomes in multicellular eukaryotes was recognized a century ago in cytological studies on the lampbrush chromosomes that are found in oocytes arrested at the diplotene phase of meiosis I [1–3]. At this stage, homologous chromosomes are paired. The two homologs display a similar and reproducible architecture. It consists of a series of loops emanating from the main axis, that are arranged in pairs, one from each homolog. In between the loops are regions of more compacted chromatin [2]. A similar physical organization is evident in insect polytene chromosomes [4]. As with lampbrush chromosomes, the paired homologs are aligned in precise register. However, instead of one copy of each homolog, there are hundreds. While loops are not readily visible, each polytene segment has a unique pattern of bands and interbands that depends upon the underlying DNA sequence and chromosome structure.
Subsequent studies have shown that the key features of chromosome architecture evident in lampbrush and polytene chromosomes are also found in diploid somatic cells [5–13]. One of these is the subdivision of the chromatin fiber into a series of loop domains. There are now many lines of evidence indicating that looping is a characteristic architectural feature. Biochemical evidence comes from chromosome conformation capture (3C) experiments, which show that distant sites come into contact with each other in a consistent pattern of topologically associating domains (TADs). While the first studies in mammals suggested that TADs have an average length of 1 Mb [14–16], subsequent experiments showed that the average is only about 180 kb [17]. In flies, TADs are smaller, between 10–100 kb [18,19]. Neighboring TADs are separated from each other by boundaries that constrain both physical and regulatory interactions. In mammals and also in flies, these boundaries typically correspond to sequences bound by insulator proteins like CTCF [17].
That TAD boundaries correspond to insulators is consistent with the known properties of these elements. Insulators subdivide the chromosome into functionally autonomous regulatory domains. When interposed between enhancers or silencers and target promoters, insulators block regulatory interactions. They also have an architectural function in that they can bring distant chromosomal sequences together, and in the proper configuration can promote rather than restrict regulatory interactions [20,21]. Moreover, insulators are known to mediate contacts between distant sequences (loop formation), and these physical contacts depend upon specific interactions between proteins bound to the insulators [22,23].
The notion that insulators are responsible for subdividing eukaryotic chromosomes into a series of looped domains raises questions about the rules governing loop formation in cis. One of these is the basis for partner choice. Is choice based simply on proximity, or is there an intrinsic partner preference? A second concerns the topology of the loop formed by interacting partners in cis. Do the partners interact to form a stem-loop-like structure, or does the interaction generate a circular loop (“circle-loop”)? The answer to this question will depend upon whether there is an orientation dependence to the interactions between two heterologous insulators. In flies, homologs are typically paired in somatic cells, not just in cells that are polyploid [24]. This means that the loop domains in each homolog must be aligned in precise register along their entire length. A plausible hypothesis is that both alignment and homolog pairing are mediated by insulator interactions in trans. If this is case, there are similar questions about the rules that govern trans interactions. Is there a partner preference in the interactions that mediate homolog pairing? Is there an orientation dependence, and if so, what is the topological outcome of the looped domains generated by insulator interactions in paired chromosomes in cis and in trans?
In the studies reported here, we have used insulators from the even skipped (eve) locus to address the questions posed above about the architecture of eukaryotic chromosomes. The eve domain spans 16 kb and is bordered upstream by the Nhomie (Neighbor of Homie, this study) insulator and downstream by Homie (Homing insulator at eve) [25,26]. eve encodes a homeodomain transcription factor that is required initially for segmentation, and subsequently in the development of the CNS, muscles, and anal plate [27,28]. It has a complex set of enhancers that activate expression at different stages and tissues [25,29–31], and a Polycomb response element (PRE) that silences the gene in cells where it isn’t needed [32]. In early embryos, the stripe enhancers upstream (3+7, 2, late stripes) and downstream (4+6, 1, and 5) of the eve gene activate transcription in a pair-rule pattern. Later in development, around the time that germband retraction commences, mesodermal (Me) and neuronal (CNS) enhancers turn on eve expression in a subset of cells in each of these tissues. These late enhancers continue to function once germband retraction is complete, while another enhancer (APR) induces transcription in the presumptive anal plate. Located just upstream of eve is CG12134, while the TER94 gene is downstream. Unlike eve, both of these genes are ubiquitously expressed throughout much of embryogenesis.
The Homie insulator has two striking properties [26]. First it directs homing of otherwise randomly inserting transgenes to a ~5 Mb region centered on the eve locus. Second, when the homed transgene carries a reporter, it is expressed in an eve-like pattern, the completeness of which diminishes with distance. Early stripe and later CNS expression are limited to 200 kb from eve, mesodermal expression has an intermediate distance dependence, while anal plate ring (APR) expression is seen at distances of several Mb. We showed previously that reporter expression at a site within the hebe gene 142 kb upstream of eve requires Homie [26]. Since other fly insulators mediate long-distance regulatory interactions by direct physical contact [22,33], we used high-resolution chromosome conformation capture (H3C) [34] to map contacts between transgenes at -142 kb and eve (see below).
The transgenes have an eve-promoter-lacZ (lacZ) reporter and Homie. One of them is inserted into the chromosome so that Homie is oriented in the same direction (→; Fig 1A, transgene #1) as the endogenous Homie in the eve locus, while the other transgene is inserted in the opposite orientation (←; Fig 1A, transgene #2). In the control transgene, Homie was replaced by DNA (Fig 1A, transgene #3). Fig 1A shows that the reporters in both Homie transgenes are regulated by the eve enhancers in a pattern which recapitulates that of endogenous eve. Thus, the orientation of the entire Homie:lacZ transgene in the chromosome doesn’t affect long-distance regulation. On the other hand, because of a hebe CNS enhancer located upstream of -142 kb, the expression pattern is not identical. In the transgene that is oriented so that Homie is closer to the eve locus than the reporter (Fig 1A, 2nd column: transgene #1), lacZ is regulated by both the hebe CNS enhancer (arrow in bottom panel) and the eve enhancers (all 4 panels). When the transgene is inserted in the opposite orientation so that the lacZ reporter is closer to the eve locus (Fig 1A, 3rd column: transgene #2), Homie blocks the hebe enhancer, and only the eve pattern is seen (all 4 panels). Finally, as expected, the reporter in the DNA control transgene (Fig 1A, right column: transgene #3) is not regulated by the eve enhancers (all 4 panels), but is regulated by the hebe enhancer (arrow in bottom panel). In this case, the reporter is separated from the hebe enhancer by DNA, not Homie. These results show that Homie induces a long-range interaction between a reporter transgene located many kilobases away and endogenous eve enhancers, and that this interaction is not sensitive to the orientation of the transgene in the chromosome. (However, this experiment does not test the orientation dependence of the reporter relative to the insulator, as this does not change between these two cases. This is tested below.) Furthermore, the long-range looping interactions between the transgene and the eve locus do not change the local enhancer blocking activity of the Homie insulator.
Since insulator bypass assays show that fly insulators pair with themselves [35–37], we expected that Homie:Homie pairing is responsible for long-distance regulation. However, as the transgene Homie might also interact with other eve elements, we used 11 primers spanning the locus (Fig 1B, arrows numbered 0–10) for H3C. Fig 1C shows the 3C results for the experimental and for the control DNA transgene, while in the inset we controlled for “non-specific” interactions using data from the DNA transgene as the reference. Whereas there is little interaction between the control transgene and the eve locus (Fig 1C green line), the experimental transgene shows significant interactions with two elements in the locus (magenta line). One is endogenous Homie. The other is located at the 5’ boundary of the eve Polycomb domain [38,39], and, from genome-wide chromatin immunoprecipitation studies [40], is bound in vivo by many insulator proteins. Based on these and findings below, we call this eve 5’ insulator Nhomie.
The experiments in Fig 1 demonstrate that reporter activation by the enhancers in the eve locus is independent of the orientation of the Homie-lacZ transgene in the chromosome. However, this doesn’t mean that reporter activation is independent of the relative orientation within the transgene of Homie and the reporter. To explore this possibility, we generated a transgene with two divergently transcribed reporters, lacZ and GFP (both are driven by the same eve basal promoter, see Materials and Methods). We then inserted Homie in both orientations between the two reporters. Fig 2A shows that in the endogenous eve locus, Homie is located downstream of the eve transcription unit in what we have designated as the “forward” 5’→3’ orientation (→). In transgene #4 (Fig 2A), using this same 5’→3’ convention for the relative orientation of the Homie insulator, the lacZ reporter would be located 5’ with respect to Homie. Thus, in this transgene the relationship between the reporter and Homie is just like the endogenous eve locus where the eve gene is located 5’ to Homie. The eve-GFP reporter is in turn located 3’ to the Homie insulator in the same relative position as the TER94 gene is with respect to the endogenous Homie. In transgene #5 (Fig 2A), the 5’→3’ orientation of Homie is flipped, so that GFP is now located 5’ relative to the Homie insulator, while lacZ is 3’. Each transgene was then inserted at -142 kb so that GFP is on the same side of Homie as the hebe enhancer, while lacZ is separated from the hebe enhancer by Homie (see diagrams in Fig 3A and 3B).
The two reporters in transgene #4 differ dramatically in their pattern(s) of expression (Fig 3A). In the case of the lacZ reporter, the eve enhancers activate expression in stripes in the early embryo, as well as in the CNS, mesoderm, and anal plate during mid-embryogenesis (green arrows). The lacZ reporter is not, however, activated by the hebe enhancer, as it is insulated by Homie. A quite different result is observed for the GFP reporter. First, unlike lacZ, it is not subject to regulation by the eve enhancers. Second, it is subject to regulation by the hebe enhancer (Fig 3A, black arrow). In transgene #5, the target for regulatory interactions with the eve locus is reversed (Fig 3B). Here, GFP is regulated by the eve enhancers (green arrows), while lacZ is not. And, since the orientation of the transgene in the chromosome remains the same, the hebe enhancer still activates GFP (Fig 3B, black arrow), while Homie blocks it from regulating lacZ.
These findings indicate that pairing interactions between the transgene Homie and the eve locus are orientation-specific. With respect to the endogenous Homie insulator (see below for Nhomie), the pairing interactions could be head-to-head or head-to-tail (Fig 3 diagrams). In the simplest topological model, head-to-head interactions predict that the lacZ reporter will be activated by eve enhancers when the 5’→3’ orientation of Homie in the transgene places this reporter 5’ of Homie, just as the eve enhancers in the eve locus are 5’ of the endogenous Homie (Fig 3A, transgene #4, “stem-loop” topology). The GFP reporter will be activated when the orientation of the transgene Homie is reversed (Fig 3B, transgene #5, “circle-loop” topology). The opposite pattern of activation is expected if Homie interactions are head-to-tail (Fig 3A and 3B, bottom diagrams). In each case, the topology of one of the variants is a stem-loop, while the topology of the other variant is a circle-loop. As can be seen from the expression patterns in Fig 3, it is head-to-head pairing between the transgene and endogenous Homie that fits the pattern of activation.
Why are the regulatory interactions in Fig 3 orientation-dependent, while those in Fig 1 are not? The difference lies in how we altered the orientation of the insulator in the two experiments. In the experiments in Fig 3, the 5’→3’ orientation of the Homie insulator in the transgene with respect to the two reporter genes was reversed. In contrast, in the experiments in Fig 1, the relative 5’→3’ orientation the Homie insulator with respect to the reporter was maintained (the reporter is 5’ with respect to Homie), while the orientation of the entire transgene was flipped. In the experiment in Fig 1 (illustrated in Fig 2B) head-to-head pairing between the Homie insulators in the transgene and the eve locus generates either a circle-loop (top, transgene #1) or a stem-loop (bottom, transgene #2). However, in both cases, the eve enhancers are brought into close proximity to the lacZ reporter. Note that as in Fig 3, the structure of the loops predicted for head-to-tail pairing of the Homie insulators in the Fig 1 experiments would place the lacZ reporter and the enhancers in the eve locus on opposite sides of the paired insulators, which would not be conducive for productive regulatory interactions (not illustrated).
Our 3C experiments identified a second element, Nhomie, in the eve locus that interacts physically with Homie at -142 kb (Fig 1C). We wondered whether Nhomie could also promote long distance regulatory interactions and function as an insulator. To test for these activities, we combined the Nhomie insulator with the lacZ reporter (Fig 2A). The Nhomie:lacZ transgene was inserted at -142 kb so that Nhomie is located between the lacZ reporter and the hebe enhancer (Fig 4A diagram, transgene #6). Since the relative orientation of Homie and the reporter was critical for productive regulatory interactions, we tested Nhomie in both orientations relative to lacZ. Using the same convention as was used for Homie, the 5’→3’ orientation of Nhomie in the endogenous locus places the eve enhancers and the eve gene 3’ of Nhomie. In transgene #6 (Fig 4A), the 5’→3’ orientation of Nhomie places lacZ is the same position relative to Nhomie as are the eve enhancers and eve gene in the endogenous locus: the reporter is located 3’ relative to Nhomie. In transgene #7 (Fig 4B), Nhomie is in the reverse orientation with respect to lacZ. In this case, the 5’→3’ orientation of Nhomie places the reporter 5’ with respect to the insulator.
Our experiments show that Nhomie shares many properties with Homie. Like Homie, it functions as an insulator and blocks the hebe enhancer from activating the reporter (Fig 4). It is also able to mediate long-distance regulation of the reporter by eve enhancers (Fig 4A). Moreover, as for Homie, these regulatory interactions depend upon the orientation of Nhomie relative to lacZ. However, the orientation of Nhomie with respect to the reporter that engenders robust activation is the opposite that of Homie. For Homie, the reporter is activated when it is located 5’ with respect to the orientation of the insulator, just like the eve gene is 5’ of the endogenous Homie. By contrast, for Nhomie, the reporter is activated when the orientation of the insulator places it 3’ relative to Nhomie; again, just as the eve gene is located 3’ relative to the endogenous Nhomie.
Our 3C experiments show that Homie at -142 kb physically interacts with Homie at the 3’ end of the eve locus (Fig 1C). It is clear from bypass experiments that self-interactions like that observed for Homie are not unusual, but instead are a characteristic property of fly insulators [35–37,41,42]. However, these transgene assays artificially juxtapose homologous partners in cis, as we have done here. In the endogenous setting, homologous partners are only present on the other homolog, and it is in this context that homologous interactions would be biologically relevant. Given that most fly insulators self-interact head-to-head, a plausible idea is that insulators are the elements responsible both for locally aligning homologs in precise register and for maintaining their stable association.
The classical evidence for homolog pairing in Drosophila is transvection [43–45]. Transvection is a regulatory interaction that occurs in trans rather than in cis, and requires local pairing of homologs. Typically two mutant alleles complement because the regulatory elements on one homolog activate the gene on the other homolog. Complementation is lost when pairing of the two alleles is disrupted by chromosomal rearrangements [46]. While a special combination of mutations is generally required to detect transvection, trans-regulatory interactions are clearly important for achieving appropriate levels of gene activity in wild-type flies [47].
The hypothesis that homologous insulator:insulator interactions are responsible for the pairing of homologs in register makes two predictions. First, placing homologous insulators in trans should promote transvection. Second, if the homologous interactions of the test insulator are orientation-dependent, transvection is expected to be greater when both copies are oriented in the same direction than when they are oriented in opposite directions. This is expected because self-pairing interactions are likely to be head-to-head rather than head-to-tail. There are two reasons behind this expectation. One is that the self-interactions detected in insulator bypass experiments are typically head-to-head, not head-to-tail [41,48,49]. The other is that head-to-tail self-interactions between endogenous insulators on each homolog would likely interfere with homolog alignment as well as transvection.
To test these predictions, we generated two transgenes, one containing the eve APR and mesoderm (Me) enhancers [25], and the second containing the lacZ reporter. The transgenes were inserted into a site far away from endogenous eve (on a different chromosome arm, at cytological location 23C4, where we do not see interactions with endogenous enhancers [26]), oriented so that both the enhancers and reporter are on the centromere-distal side of their respective transgene (Fig 5A). In the first experiment, the enhancer transgene had λ DNA, while the reporter had either DNA or Homie. Since there are insulator-like elements near the 23C4 attP site (one ~50 bp distal to the attP site, another ~8 kb proximal) [40], we expected to see some transvection [50,51] when either:lacZ or Homie:lacZ is trans to the:enhancer transgene. Fig 5A (top two panels) shows that the APR enhancer weakly activates lacZ (green arrows), while there is virtually no Me-driven expression (red arrows). As predicted, the presence of a forward-oriented Homie in the enhancer transgene substantially augments transvection (Fig 5A, 3rd panel). Not only is APR expression much stronger (green arrow), but Me-driven expression is also clearly observed (red arrow).
To confirm that stable pairing is head-to-head, we reversed Homie in the enhancer transgene (maintaining the overall transgene orientation in the chromosome). In this configuration, head-to-head pairing would introduce an S-shaped double loop. As illustrated in Fig 5A (“twisted pairing”), this would place the reporter on the opposite side of the paired insulators from the transgenic enhancers. This configuration would not be expected to increase enhancer-reporter interactions. Consistent with this prediction, reporter expression is about the same (Fig 5A, bottom panel) as in the negative controls carrying DNA (Fig 5A, top two panels). Alternatively, the need to form such a double loop might make this pairing interaction less stable than for the other orientation, when the head-to-head pairing reinforces the normal pairing of the homologs (“tightly paired” in the diagram). In fact, evidence below is more consistent with such “twisted pairing” interactions forming only transiently, or not at all (hence the red “X” in the diagram for “twisted pairing”). This is in line with the expectation, stated above, that head-to-head self-interactions between endogenous insulators mediate homolog alignment and pairing, while head-to-tail self-interactions are incompatible with smooth alignment and tight pairing.
To further explore the relationship between pairing direction and transvection, we generated dual reporters with divergently transcribed GFP and lacZ that have either DNA or Homie inserted between the reporters (Fig 5B). When the DNA:dual reporter is trans to the Homie-enhancer transgene, the APR enhancer weakly stimulates lacZ and GFP in the APR, while neither reporter is activated by the Me enhancer. The addition of Homie to the reporter (in the same orientation in the chromosome as that of Homie in the enhancer transgene) substantially enhances APR lacZ transcription, and turns on lacZ in the mesoderm. By contrast, there is only a slight increase in APR GFP expression, while mesoderm expression is detectable, but only weakly. The differences in transvection for the two reporters are consistent with the topology generated by head-to-head, not head-to-tail pairing (Fig 5B, “tightly paired”).
We also combined the dual reporter with an enhancer transgene in which the entire transgene containing Homie and the enhancers are flipped (Fig 5B, bottom panel). Head-to-head pairing of Homie would generate an S-shaped double loop (as diagrammed in Fig 5B, “twisted pairing”). In this case, there is little or no enhancement of transvection for either reporter, suggesting that the introduction of such a double loop between the paired homologs either is unstable or does not form (indicated by the red “X” in the diagram for “twisted pairing”).
We note that there are some subtle differences in the expression patterns for transgene combinations in which transvection is not significantly enhanced. This includes all the cases where our topology diagrams are labeled as “loosely paired” or “twisted pairing”. These differences may be due to a combination of several factors, such as differences in the size of the transgenes, weak or unstable interactions with insulators near the site of transgene insertion, or the shielding of transgenic reporters from position effects that weakly upregulate or downregulate reporter activity.
We next tested whether Nhomie self-interactions in trans also induce transvection. Nhomie was oriented in the single reporter transgene so that the lacZ reporter (diagrammed in Fig 6A) is 3’ with respect to Nhomie. It was then combined in trans with an enhancer transgene that had or Nhomie (in the same 5’→3’ orientation in the chromosome) so that the two enhancers are 3’ of Nhomie (Fig 6A). In the:Nhomie combination, the APR enhancer drives only weak expression, and activation by the Me enhancer is not seen. As would be predicted if head-to-head pairing aligns the enhancers and the reporter, lacZ expression is substantially elevated in the Nhomie:Nhomie combination. This conclusion is confirmed by the dual reporter assay. As shown in the lower half of Fig 6B, head-to-head pairing of Nhomie in the enhancer and dual reporter transgenes would juxtapose the Me and APR enhancers with the lacZ reporter, while the GFP reporter would be separated from the enhancers by the paired Nhomie insulators. In this configuration, the Me and APR should preferentially drive lacZ expression, not GFP expression, and this is what is observed.
While Homie-Homie or Nhomie-Nhomie self-interactions normally occur at the endogenous eve locus only in trans, this is not the case for Nhomie-Homie interactions. Heterologous interactions between neighboring insulators in cis are thought to be responsible for subdividing chromosomes into a series of topologically independent domains, and are expected to occur all along the chromosome. Like self-interactions, heterologous interactions are known to be specific [37,42,52,53], and consequently are likely also orientation-dependent. For heterologous insulators interacting in cis, we define their endogenous directionalities to be the same. That is, the arrows that represent them point in the same “forward” direction along the chromosome (as in Figs 1–4 for endogenous Nhomie and Homie). Using this convention, at the endogenous eve locus, head-to-tail interactions between Nhomie and Homie would generate a stem-loop, while head-to-head interactions would generate a looped circle or “circle-loop.” To test whether these two insulators can interact with each other independently of the eve locus, and (if so) determine their orientation dependence, we combined a Nhomie-lacZ reporter with two different Homie-enhancer transgenes. In the one in which the enhancers are 5’ of Homie (Fig 6C, top panel), head-to-tail pairing with Nhomie should align the enhancer and reporter, and favor transvection. When the enhancer transgene has Homie in the reverse orientation (Fig 6C, bottom panel), enhancer-reporter alignment would be favored by head-to-head pairing. Fig 6C shows that Nhomie and Homie can pair with each other in a foreign context (top panel), and that transvection is favored by head-to-tail pairing (top panel vs. bottom panel). These findings parallel those for self-pairing (Figs 5 and 6A and 6B), except that heterologous pairing is head-to-tail rather than head-to-head.
To confirm these results, we combined the dual lacZ, GFP reporter containing Homie with an enhancer transgene containing Nhomie. As illustrated in Fig 6D, head-to-tail pairing of Nhomie and Homie would juxtapose the enhancers with lacZ, while head-to-head pairing would juxtapose the enhancers with GFP. Consistent with head-to-tail pairing, lacZ transvection is stimulated, while GFP is not (compare 6D with the control in the upper half of 6B).
The insulator interactions in the transvection assay are local and likely facilitated by homolog pairing. To confirm that the eve insulators can interact specifically with themselves and with each other over large chromosomal distances, we took advantage of attP 25C1, located 2 Mb distal to 23C4. A Homie:lacZ transgene was inserted at 25C1. It was combined with an enhancer transgene at 23C4 containing either DNA or Homie (Fig 7A). No interaction between the transgenes is evident with the DNA control or when is replaced by the su(Hw) insulator. On the other hand, when both the reporter and the enhancer have a Homie insulator, the APR enhancer is able to activate lacZ expression (Fig 7A, upper left panel). This result is consistent with previous studies which showed that APR was the only enhancer in the endogenous eve locus that could act over distances >1 Mb with Homie-carrying transgenes [26]. As would be expected from the orientation dependence of insulator self-pairing, when Homie is inverted within the enhancer transgene (Fig 7A, upper right panel), expression is not seen, confirming that Homie-Homie pairing is head-to-head.
We also tested whether Nhomie can mediate distant regulatory interactions either with itself or with Homie. In the two transgenes used to test Nhomie self-interactions, the enhancers or reporter, respectively, are each located 3’ relative to the adjacent insulator. These Nhomie transgenes were inserted (separately) at 23C4 and 25C1, then crossed into the same animals. Fig 7B shows that Nhomie:Nhomie interactions can mediate long-distance activation of lacZ by the APR enhancer (Fig 7B, lower right). Nhomie also pairs with Homie, enabling the APR enhancer in the Nhomie transgene at 23C4 to activate a Homie-lacZ reporter at 25C1 (Fig 7B, upper left). As illustrated in Fig 7, these interactions are all consistent with the orientation dependence seen in the other assays, namely head-to-head self-pairing and head-to-tail heterologous pairing.
The experiments described above indicate that Nhomie and Homie must be able to physically pair with each other, and do so in a head-to-tail orientation. In the endogenous locus, head-to-tail pairing would generate a stem-loop containing the eve transcription unit and its associated enhancers and Polycomb silencers, linked together at the base by the Nhomie and Homie insulators. 3C experiments with Homie as the anchor confirm that Nhomie and Homie contact each other in the eve locus (Fig 1D).
The importance of insulators in organizing eukaryotic chromosomes has been recognized since their discovery in the 1980’s. However, the principles underlying their architectural and genetic functions have not been fully elucidated. With this goal in mind, we asked how these elements shape two critical architectural features of chromosomes. The first is homolog pairing. Homologs pair in flies from the blastoderm stage onward, and the consequent trans-interactions are important for proper gene regulation. The phenomenon of homolog pairing is not unique to Drosophila [24,54]. Homologs are paired in lampbrush chromosomes of invertebrate and vertebrate oocytes. The second is the looped domain organization [20,21,55]. Although there is now compelling evidence that insulators subdivide chromosomes into topologically independent looped domains (and that these domains play a central role in gene regulation), the topology of the loops is unknown. Moreover, while the loops must emanate from the main axis of the chromosome, the relationships between the loops, the insulators that delimit them, and the main chromosomal axis are not understood. As homolog pairing is more straightforward and the likely mechanism better documented, it is considered first.
Homolog pairing requires mechanisms for aligning homologs in precise register, and maintaining their stable association. While many schemes are imaginable, the simplest utilizes elements distributed along each homolog that have self-interaction specificity. Such a mechanism would be consistent with the persistence of local pairing and transvection in chromosomal rearrangements [44,56–60]. It would also fit with studies on the pairing process [56,61,62]. Self-association of pairing elements would locally align sequences in register, and ultimately link homologs together along their entire length. In this mechanism, self-association must be specific and also directional, namely head-to-head. This avoids the introduction of unresolvable loops and maximizes pairing for transvection.
In Drosophila, the homing of P-element transgenes, in which normally random insertion becomes targeted, suggested the ability of genomic elements to self-interact. Such a homing activity was found in the engrailed locus for a region that includes two PREs [63–65], and later studies showed that some insulators [26,66,67] and a promoter region [68] also possess homing activity. The self-interaction implied by homing suggests that these elements may facilitate homolog pairing. However, in contrast to PREs and promoters, insulators have consistently been found to engage in specific self-interactions (see below). Thus, among the known elements in the fly genome, insulators are the best candidates to align homologs in register and maintain pairing [20,21]. Moreover, genome-wide chromatin immunoprecipitation experiments (ChIPs) show that insulators are distributed at appropriate intervals along each chromosome [18,19].
A role in homolog pairing was first suggested by the discovery that the su(Hw) and Mcp insulators each can mediate regulatory interactions between transgenes inserted at distant sites [69,70]. The Fab-7 insulator can also mediate long-range regulatory effects [71]. Further evidence that self-association is characteristic of fly insulators came from insulator bypass experiments [35,36]. These experiments showed that bypass is observed when an insulator is paired with itself, while heterologous combinations are less effective or don’t give bypass [37,41,42,48,72,73]. Moreover, self-pairing is, with few exceptions, head-to-head.
That insulators mediate homolog pairing through specific self-interactions is further supported by our studies. Using a classical transvection assay, we found that Homie-Homie and Nhomie-Nhomie combinations stimulate trans-regulatory interactions between enhancers on one homolog and a reporter on the other (Figs 5, 6A and 6B). Moreover, the parameters that favor transvection dovetail with those expected for a pairing mechanism based on insulator self-interactions in trans. First, the two insulators must be in the same orientation. When they are in opposite orientations, transvection is not enhanced (or enhancement is much weaker, Fig 5). Second, the enhancers and reporter must be located on the same side (centromere proximal or distal) of the insulators (Figs 5, 6A and 6B). In addition to transvection, Homie and Nhomie also engage in highly specific and directional distant regulatory interactions (Fig 7).
While there is compelling evidence that insulator self-interactions are responsible for homolog pairing, many issues remained unresolved. Perhaps the most important is the nature of the code used for self-recognition and orientation. The best hint comes from bypass experiments using multimerized binding sites for Su(Hw), dCTCF, or Zw5. Homologous multimer combinations give bypass, while heterologous combinations do not. However, bypass is observed for composite multimers when they are inserted in opposite orientations (e.g., Su(Hw) dCTCF ↔ dCTCF Su(Hw)), but not the same orientation (e.g., Su(Hw) dCTCF →→ Su(Hw) dCTCF) [53]. These findings argue that the identity and order of proteins bound to the insulator determine its self-association properties.
The first direct evidence that insulators generate loops came from 3C experiments on the mouse β-globin and the fly 87A7 heat shock loci [23,74]. These studies suggested that physical interactions between adjacent insulators in cis could subdivide chromosomes into looped domains. Subsequent work has confirmed this conclusion [17]. However, while these experiments demonstrate that cis insulator interactions generate loops, they provided no information about the topology of these loops, or how they are arranged.
Cis interactions could, a priori, be either head-to-head like self-association in trans, or head-to-tail. The consequences are quite different. Head-to-head interactions generate a circle-loop, while head-to-tail interactions generate a stem-loop (Fig 8A and 8D, respectively). If heterologous insulators interact with only one specific partner, the circle-loop or the stem-loop will be linked to neighboring circles or stem-loops by loops without anchors. These unanchored loops would correspond to the main axis of the chromosome, and the circle-loops or stem-loops would then protrude from the main axis in a random orientation and at distances determined by the length and compaction of the unanchored loops.
On the other hand, if insulators in a chromosomal segment are able to interact with both of their neighbors, then the main axis of the chromosome in this region would be defined by the insulators. Quite different structures are predicted for head-to-head and head-to-tail interactions (Fig 8B and 8E). Head-to-head would give a series of variably sized circle-loops linked together at their base by an array of interacting insulators. The base would correspond to the main axis of the chromosome, and each circle-loop would extend from one side of the main axis to the other. If the direction of coiling were always the same, this would give a structure resembling a helix anchored to a rod (Fig 8B). If the direction of coiling were random, the structure would be more complicated and variable, since neighboring circle-loops could extend out from the main axis in either the same or the opposite direction (not illustrated). The loop-axis relationship would be more regular for head-to-tail insulator pairing in cis. Adjacent stem-loops would extend out from the main axis in opposite directions much like the lampbrush chromosomes formed when haploid sperm heads are injected into amphibian oocytes (Fig 8E) [75]. This stem-loop organization would also fit with the radial loop model proposed by Laemmli and others for the first level of folding of metaphase chromosomes [7,11].
Since our experiments show that Homie-Nhomie association is head-to-tail, the topology of the eve locus in vivo is a stem-loop, not a circle-loop. This finding raises a number of questions. Perhaps the most important is whether head-to-tail interactions are the rule rather than the exception. While the orientation dependence of homologous interactions has been extensively investigated, there have been no systematic studies on interactions between neighboring insulators. However, there are reasons to think that cis interactions are more likely head-to-tail than head-to-head. One is homolog pairing. As mentioned above, the circle-loops formed by head-to-head interactions can coil in either direction, either left-handed or right-handed. If coiling were random, then about half of the circle-loops on each homolog would be coiled in opposite directions. In this case, head-to-head pairing of homologous insulators in each homolog would generate a structure in which the circle-loops would point in opposite directions (Fig 8C, left circles). This topology would not be compatible with transvection. Coiling of the circle-loops in the same direction on both homologs would permit interdigitation of one circle-loop inside the other (Fig 8C, right circles); however, the chromatin fiber from the inside circle-loop would need to cross in on one side and out on the other. If the main axis of the chromosome in the paired region is defined by a series of interacting insulators in cis, then generating a topology permissive for transvection (not illustrated) would require coiling of successive homologous circle-loops on each homolog in the same direction, one inside the other (Fig 8C, right circles).
These topological issues aren’t encountered when heterologous insulator interactions in cis are head-to-tail. Head-to-head pairing of homologous insulators in trans would bring regulatory elements and genes in the two homologous stem-loops into close proximity. Alignment of the two homologs is straightforward whether or not the main axis of the chromosome is defined by a series of interacting insulators (Fig 8F illustrates one of these cases). Alternating loops extending upwards and downwards from the main axis of the chromosome would be directly aligned when homologous insulators pair head-to-head in trans.
While the requirements for aligning and pairing homologs would appear to favor stem-loops between heterologous insulators in cis in flies, homolog pairing does not occur in vertebrates except in specialized cell types [76]. This could mean that circle-loops formed by cis interactions between heterologous insulators are permissible in vertebrate chromosomes. However, even in organisms in which homolog pairing doesn’t occur in somatic cells, it seems possible that cis-pairing interactions more commonly generate stem-loops than circle-loops. First, following DNA replication and before mitosis (during the S and G2 phases of the cell cycle), sister chromatids are aligned. Maintaining this alignment may facilitate epigenetic mechanisms that template chromatin structures from one cellular generation to the next, such as the copying of histone modifications onto both daughter chromosomes. The simpler topology of stem-loops could facilitate this sister chromatid pairing, as well as their separation during mitosis. Second, recent studies on the relationship between loop domains and CTCF insulators showed that in more than 90% of the cases, the CTCF binding sites on opposite ends of a loop are in opposite orientation [17]. Thus, assuming that the orientation of pairing is such that the CTCF sites are aligned in parallel to form the loop, pairing between CTCF insulators at the ends of the loop would generate stem-loops rather than circle-loops. If insulators form the main axis of the chromosome, there is an additional explanation for such a bias. As shown in Fig 8B, head-to-head pairing in cis could generate a series of circular loops that extend out from the same side of the main axis. This configuration would be favorable for crosstalk between regulatory elements and genes in adjacent loops. By contrast, head-to-tail pairing, where adjacent stem-loops extend out in opposite directions (Fig 8E), would disfavor crosstalk, helping to explain how insulators block enhancer-promoter communication between adjacent loops.
See S1 Fig for the Homie and Nhomie regions used, control DNA, and tag sequences. Reporters contain the eve basal promoter, -275 to +106bp from eve +1 (TSS), either the lacZ or GFP coding region, and the eve 3'-UTR, +1300 to +1525 bp. Enhancers are: eve APR,+3.0 to +4.1 kb; eve Me, +5.7 to +6.6 kb, each cloned in plasmid attB∆2 [32] for transgenesis [77] using ΦC31 [78]. Target sites were: -142 kb from eve [26]; 23C4 (2L;3029226), generated by us; and 25C1 [77]. Sequence coordinates are Flybase version dm6 [79]. Two genomic fragments used in this study that span the insulator protein binding region we call Nhomie, based on genome-wide studies [40], were found to have indistinguishable function in our assays (S2 Fig). The corresponding sequences are given in S1 Fig.
RNA in situ hybridization and anti-β-galactosidase staining were as described [25]. In all cases, conclusions drawn were based on comparisons between control and experimental collections of embryos that were stained in parallel.
H3C analysis was performed as described [34], with the following modifications. Embryos (200 μl aged 0-6h at ~23°C) were cross-linked in either 2% or 3% formaldehyde for either 15 or 30 min (each gave similar results, and were included in the data presented), digested with 100U each of EcoRI (Roche) and MfeI (NEB) at 37°C overnight. About half of the material was ligated (Takara, 3500U) for >4 hr. at ~23°C, and incubated at 65°C overnight to reverse cross-links. Following RNase A (Roche, 40μg/sample) and proteinase K (Roche, 220μg/sample) digestions, purified DNA (20ng/reaction) was subjected to real-time PCR analysis using SYBR Green Master Mix (Roche).
All transgenes inserted at -142 kb used for 3C analysis had the same tag sequence, which was used as the anchor primer (Fig 1C), in combination with each of a series of accompanying primers from within the eve locus. To identify Homie-interacting regions within endogenous eve, an endogenous Homie fragment-specific primer was used as anchor (Fig 1D), along with the same series of accompanying primers. These sequences are given in S1 Fig.
PCR quantification was done as described [34], with the following set-up. The fragments in the eve locus created by EcoRI and MfeI digestion were cloned into anchor fragment-carrying plasmids, and served as standards for the expected ligation products. These plasmids were linearized and mixed with equimolar amounts of digested genomic DNA. Details of the various controls, such as the choice of primers and enzymes, were appropriate for each specific experiment [34]. Additional details are given in the legend to Fig 1.
|
10.1371/journal.pntd.0001733 | Seroprevalence and Risk Factors for Taenia solium Cysticercosis in Rural Pigs of Northern Peru | Taenia solium is a cestode parasite that causes cysticercosis in both humans and pigs. A serological survey was undertaken to assess the seroprevalence and risk factors associated with porcine cysticercosis in the rural district of Morropon, Peru. Pigs aged between 2 and 60 months were assessed by the Enzyme-linked Immunoelectrotransfer blot (EITB) assay to determine their serological status against porcine cysticercosis in a cross-sectional study. A total of 1,153 pigs were sampled. Porcine seroprevalence was 45.19% (42.31–48.06). The information about the animals and households was analyzed and risk factors associated with seroprevalence were determined by a multivariate logistic regression analysis. In the porcine population, the risk of being seropositive increased by 7% with every month of age (OR 1.07, 95% CI 1.05–1.09), and by 148% for pigs living in East Morropon (OR 2.48, 95% CI 1.82–3.37). Whereas, the presence of latrines in a household decreased the risk of being seropositive by 49% (OR 0.51; 95% CI 0.39–0.67). Sex and rearing system did not represent either risk or protective factors associated with the seroprevalence of porcine cysticercosis. The findings of this study could be used for further development of control programs that might focus on similar population groups within rural communities of developing countries where cysticercosis is endemic.
| Taenia solium causes taeniasis in humans and cysticercosis in humans and pigs. In humans the parasite may infect the central nervous system and cause neurocysticercosis. The World Health Organization (WHO) estimated that over 50,000 deaths per year are due to neurocysticercosis and the disease is also the main cause of acquired epilepsy. Pigs act as intermediate hosts for the parasite's transmission. Porcine cysticercosis causes economic losses to farmers of developing countries, because infected pork has reduced value or may be condemned. Previous studies have identified risk factors for T. solium infection in pigs in various parts of the world; however, findings are contradictory or not consistent. In this study, particular areas in which pigs lived and age (older pigs were at higher risk) increased the risk of being seropositive; whereas the use of latrines decreased their risk of being seropositive. The results of this study contribute to epidemiology of porcine cysticercosis in rural areas, which is relevant for establishing effective control programs in rural locations with similar characteristics.
| Neurocysticercosis is a disease that affects humans mainly in developing countries, causing serious morbidity and mortality [1]. T. solium infection in pigs causes production losses to farmers because infected meat has reduced value or may be condemned at slaughterhouses. In rural areas infected pig carcasses can be sold avoiding the legitimate commercial distribution [2]. Epilepsy caused by neurocysticercosis in humans incurs many economic and social costs. It affects workers within highly productive age groups reducing work productivity [3]. Stigmatization arises as a serious problem for the farmers/villagers with neurocysticercosis since they are relegated and do not have the benefit of being part of the normal community life [4], [5]. In Peru, epidemiological studies based on serological surveys using the Enzyme-linked Immunoelectrotransfer blot (EITB) have determined variable porcine cysticercosis seroprevalences in the three natural regions the country: coast, highlands and Amazon. The EITB test has been commonly used to determine the epidemiological characteristics of the taeniasis/cysticercosis complex [6]. Studies done in the Peruvian Amazon found seroprevalences of porcine cysticercosis that ranged from 28% to 49% [6]. In the Peruvian Highlands, a region with a high poverty rate, the disease is known to be hyper-endemic with seroprevalences up to 75% [1]. Studies in the Northern Coast of Peru found seroprevalences that ranged from 13% [4] to 30.8% [7].
Few studies on the risk factors for porcine or human cysticercosis in Peru have been done [7], [8]. These studies assessed the factors in the human and pig populations that are associated with the seroprevalence of porcine cysticercosis in rural villages of the Highlands and Coast of Peru. However, some social, economic, geographic and environmental characteristics are specific to particular locations and therefore risk factors may differ from communities located in different regions. A cross-sectional serological survey in pigs was undertaken to determine the seroprevalence of porcine cysticercosis and identify the risk factors for T. solium transmission. The survey was conducted in 14 villages located in the district of Morropon, Piura, Peru using the EITB as the diagnostic test. The particular region investigated was selected at the beginning based on limited, anecdotal knowledge which suggested a high rate of T. solium transmission in the region.
The study complied with the “National Health and Medical Research Council Australian Code of Practice for the Care and Use of Animals for Scientific Purposes” (7th edition, 2004) ethics standard. The study protocol was approved by the scientific boards at the Veterinary Faculty, the University of Melbourne, Australia and at the Veterinary Faculty, San Marcos National University, Peru. Study permissions were obtained from the Municipality of Morropon, from village leaders and from the pig owners. Due to a high level of illiteracy among villagers, the scientific board at the Veterinary Faculty, San Marcos National University approved the use of oral consent. Oral consent was obtained from household owners prior to them providing answers to the questionnaire and using their animals in the study; consent was recorded through the completion of the questionnaire.
Morropon is a province in the department of Piura, Peru. It is located in Northeast Peru, 82.3 km from Piura, the closest major city. The district of Morropon is the commercial center of the region and villages are located in the surroundings. The altitude is 131 meters above sea level. The climate is dry and hot from May to December, with heavy rain fall from January to April.
The sample size required for the study was obtained through the Sample Size formula for estimation of a proportion for infinite populations [9]. The referential prevalence used was 26% [10] and a 95% confidence level. The minimum number of animals calculated by the sample size formula was 296. This number of animals was required in order to represent a valid sample of the total population; however, due to the availability of resources and to not affect the villager's compliance by selecting some houses only, it was decided to undertake a census sampling, attempting to include every pig from all the villages in the study.
Sera samples for processing were sent to the diagnostic laboratory of the Instituto Nacional de Ciencias Neurológicas (Lima, Peru) to perform the EITB test. The EITB test used the same methodology described by Tsang et al. [11] and Gonzalez et al. [12]. The EITB assay for diagnosis of human or porcine cysticercosis identifies as being positive any sample having reactivity with any one of seven lentil-lectin, affinity-purified Taenia solium metacestode glycoprotein antigens (GP50, GP42-39, GP24, GP21, GP18, GP14 and GP13). The sensitivity and specificity of this assay in pigs were reported to be 100% [12].The sensitivity of this assay in humans is 98% and its specificity reaches 100% [13].
A serological survey of pigs was undertaken in the following villages: Alto Mambluque, Bocanegra, Coca, El Chorro, Faical, Franco, Franco Alto, Franco Bajo, La Bocana, Mambluque, Maray, San Francisco, Talanquera, and Zapotal located in the district of Morropon. These villages were selected based on various characteristics, such as socio-economic, accessibility, recent cases of human or porcine cysticercosis and acceptance and enthusiasm from villagers and village authorities. Pregnant sows and animals younger than 2 months were not included in the survey. All animals were ear-tagged to identify the animals and their blood samples. Animals were also vaccinated against Classical Swine Fever (CSF) as an incentive to the livestock owners to participate in the survey.
The household owners were interviewed and information about the household, living conditions and pig husbandry practices were recorded. The respondent was the owner of the household (husband or wife or any adults living in the household). Their oral consent was obtained prior to answering the questionnaire and taking blood samples from the pigs. The questionnaire was concurrently conducted while blood samples of pigs were taken. Specific information was recorded for each animal including: identification number, age and sex. Information about the households included: the presence of latrines, the rearing-system used and the village to which households belonged. The data was analyzed anonymously. Each household's data was kept confidential and not shared with any other household.
Data was entered on Microsoft Office Excel 2007 datasheets (Microsoft). Statistical calculations were performed using the computer program STATA 10.0 (StataCorp LP, USA). Descriptive analyses were based on frequencies and percentages for qualitative variables, and means with their confidence intervals for quantitative variables. Bivariate analyses were performed calculating odds ratios (OR) to assess the variables: sex, age, village, rearing system, and presence/absence of a latrine as potential risk factors.
For practical and statistical purposes, Morropon villages were assigned into three different areas: East Morropon, West Morropon and Yamango. The two criteria used to divide the area were geographical and road access. Yamango area included Faical, Coca, San Francisco, Mambluque and Alto Mambluque villages. East Morropon included Maray, El Chorro and Bocanegra villages. West Morropon included Zapotal, Franco, Franco Alto, Franco Bajo, Talanquera and La Bocana villages. Figure 1 shows a schematic map of the study area.
The aggregated effect of the sex, age (months), area, rearing system and presence/absence of a latrine over the binary dependent variable, EITB result, was modeled by a multivariate logistic regression analysis. The multivariate model included variables of epidemiological and biological interest. Adjusted relative risks for pig seropositivity to cysticercosis with the corresponding 95% confidence interval (95% CI) and p values were estimated by a multivariate logistic regression. A p value of less than 0.05 was considered to indicate statistical significance.
The number of animals sampled, the total seroprevalence and the seroprevalence per village are shown in Table 1. A total of 1153 animals were sampled belonging to 306 households. The population sample represented approximately 90% of the total pig population in the study villages. Seropositive animals to the EITB test were found in all 14 villages of study. From the 1153 sampled animals, 521 were positive, giving a seroprevalence of 45.19% (42.31–48.06). The seroprevalences per village ranged from 14.5% for Franco Alto to 81.2% for Bocanegra. When analyzed by area, East Morropon had the highest porcine cysticercosis seroprevalence, 56.6% (176/311).
The West Morropon area had the highest number of sampled pigs, 41.7% (481/1153), while East Morropon area had the smallest number of sampled animals, 27.0% (311/1153). The mean age of the population was 8.5±8.3 months, with ages ranging from 2–60 months. To have a gross overview of the age of the study population, age was arranged as an ordinal variable using ≤8, 9–16 and >16 months as reference ages (for the multivariate analysis age was modeled as a continuous variable). There was a higher proportion of “young” animals, with pig ≤8 months representing 70.8% (816/1153) of the total sample, pigs between 9–16 months represented 18.6% (214/1153) of the population, and pigs ≥16 months represented 10.7% (123/1153). There were slightly more female pigs, 55.5% (640/1153) in the population compared to males, 44.5% (513/1153). The majority of pigs were reared using the free-roaming system, 56.3% (649/1153) in contrast to pigs reared in confined conditions, 43.7% (504/1153). The seroprevalences in animals reared confined and free-roaming were 44.8% (226/504) and 45.5% (295/649) respectively. From the 306 households that participated in the serological survey, 43.1% (132/306) did not have latrines, and 42.5% (490/1153) of the sampled animals belonged to households without latrines. All interviewed families (306) knew about the occurrence of porcine cysticercosis in the area of study (known as “triquina” by the locals). However, less than 1% (2/306) associated the pig habit of coprophagy with the onset of the disease in the animals (data not shown). A summary of the various exposure and biological variables and the seroprevalence (EITB result) is presented in Table 2.
A multivariate logistic regression analysis was performed for all variables included in the study with the exception of the variable “village”. Instead of “village”, the variable “area” (the three areas into which Morropon district was divided) was included in the logistic regression model. The variable age (months) was analyzed as a continuous variable. The odds ratios with confidence intervals as well as the p values of a logistic regression to determine the risk factors for the EITB seropositivity are shown in Table 3. The analysis determined that the pig's age was a risk factor for reaction against the EITB test in seropositive animals (adjusted by sex, area, rearing method and presence/absence of latrine). The odds of being seropositive increased by 7% with every month of age in our study population (p<0.01). The multivariate analysis also determined that animals belonging to the villages in the East Morropon had a 148% increased risk of being seropositive compared to animals from West Morropon. The presence of latrines in households was a protective factor for the occurrence of being seropositive, finding a 49% decreased risk in houses with latrines versus houses without them (p<0.01). Finally, variables such as sex and rearing system did not represent either risk or protective factors associated with the seroprevalence of porcine cysticercosis.
This study found that the age of a pig and the area where the pig lived increased the risk of being seropositive to T. solium whereas the presence of latrines was found to decrease the risk of being seropositive. The seroprevalence of porcine cysticercosis was found to be 45.19% which is similar to the seroprevalence of porcine cysticercosis found in areas where the disease is considered hyper-endemic [1].
The finding that the presence of latrines was a protective factor to decrease the seroprevalence of porcine cysticercosis is not surprising as the use of latrines has been proposed to control cysticercosis worldwide by multiple authors [14], [15] and has been reported to be a protective factor for the occurrence of the disease [16], [17], [18], [19]. However, in other studies it was not associated with the occurrence of porcine cysticercosis [18] and in some cases it acted as a risk factor [20]. In our study, pigs were seen feeding on human feces near latrines. The access of pigs to human feces has been shown to be a risk factor for porcine cysticercosis [21]. Based on our observations, it appears that the presence or absence of latrines is as important as the knowledge required to use them properly.
In Morropon, an increase in the pig's age was a risk factor for being seropositive, which agrees with studies from Cameroon, Mozambique and Mexico [7], [22], [23]. However, Ngowi et al. [18] reported that the age was not a risk factor for the occurrence of porcine cysticercosis. It has been described that older animals have a higher chance of accessing human feces since younger animals have a disadvantage when foraging and scavenging [19]. Adult pigs also have a higher frequency of feces consumption compared to piglets [24]. In our study, the seroprevalences observed in young pigs may be also a reflection of the transfer of maternal antibodies [25]. Because of the association between seroprevalence and pig's age in rural communities, the effectiveness of control programs could be affected by the presence of older animals, as they might be reservoirs for the disease.
Peruvian rural families commonly use the pig as the equivalent of a savings account [26]. Pigs act as recyclers of waste thrown on the streets and walking paths, including feces. There is evidence that indicates that restraining pigs, which provides varying confinement (different locations and tethering length), may reduce the seroprevalence of porcine cysticercosis in rural areas [16], [27]. It has been reported elsewhere that using a free-range husbandry system increases the risk of acquiring cysticercosis [21], [23]. However, in this study the free-range rearing system was not a risk factor for porcine cysticercosis seroprevalence and could be explained by 1) the presence of human feces in the pig pens (due to open-air defecation); 2) the contamination of the environment by the presence of chickens in the households where pigs were raised (Jayashi, personal observations) since their feeding behavior promotes the spreading of Taenia spp. eggs [28]; and, 3) the inadequate construction of the pens that allows pigs to often “escape” their confinement and have access to human waste.
The villages in the area of East Morropon area are closer to the district center and have less harvesting areas. These properties had less land and animals and owners could be considered poorer than people living in villages at longer distances from the district center. Under these conditions the source of food is scarce and pigs are forced to scavenge for food and may more frequently ingest human feces, which could explain why pigs living in this area had increase odds for being seropositive.
This study had some limitations. Studies that have relevant information about the prevalence of porcine cysticercosis including our study are based on EITB serology rather than definitive post mortem examination [1], [4], [6], [20]. EITB has been widely used for epidemiological studies following Gonzalez's et al. [12] results, in which EITB was determined as being 100% specific and 100% sensitive in pigs. Based on these findings, EITB was considered as the gold standard test to diagnose porcine cysticercosis. However, some evidence shows that EITB serology does not correlate perfectly to necropsy results and in many cases animals are seropositive and necropsy is negative [10], [29]. It has been suggested that a positive serology result and a negative necropsy result is due to exposure of the animal to T. solium caused by an aborted infection which did not lead to mature cysticerci or detectable lesions [12]. A similar situation is presumed to exist with human serology [30]. Therefore, the risk of animal being seropositive in our study does not necessarily represent the risk of the animal being positive at necropsy.
This study describes the risks associated with the disease transmission in an area where porcine cysticercosis is highly prevalent. It expands the information available regarding the epidemiology of porcine cysticercosis in rural farming systems. The information obtained in this study was used for making the decision to proceed with a field vaccination trial in pigs to prevent cysticercosis in rural Peru. Nevertheless, this knowledge may also be used for further development of control programs that might focus in particular population groups within rural communities of developing countries where porcine cysticercosis and neurocysticercosis are endemic.
|
10.1371/journal.pgen.1003336 | MicroRNA-3148 Modulates Allelic Expression of Toll-Like Receptor 7 Variant Associated with Systemic Lupus Erythematosus | We previously reported that the G allele of rs3853839 at 3′untranslated region (UTR) of Toll-like receptor 7 (TLR7) was associated with elevated transcript expression and increased risk for systemic lupus erythematosus (SLE) in 9,274 Eastern Asians [P = 6.5×10−10, odds ratio (OR) (95%CI) = 1.27 (1.17–1.36)]. Here, we conducted trans-ancestral fine-mapping in 13,339 subjects including European Americans, African Americans, and Amerindian/Hispanics and confirmed rs3853839 as the only variant within the TLR7-TLR8 region exhibiting consistent and independent association with SLE (Pmeta = 7.5×10−11, OR = 1.24 [1.18–1.34]). The risk G allele was associated with significantly increased levels of TLR7 mRNA and protein in peripheral blood mononuclear cells (PBMCs) and elevated luciferase activity of reporter gene in transfected cells. TLR7 3′UTR sequence bearing the non-risk C allele of rs3853839 matches a predicted binding site of microRNA-3148 (miR-3148), suggesting that this microRNA may regulate TLR7 expression. Indeed, miR-3148 levels were inversely correlated with TLR7 transcript levels in PBMCs from SLE patients and controls (R2 = 0.255, P = 0.001). Overexpression of miR-3148 in HEK-293 cells led to significant dose-dependent decrease in luciferase activity for construct driven by TLR7 3′UTR segment bearing the C allele (P = 0.0003). Compared with the G-allele construct, the C-allele construct showed greater than two-fold reduction of luciferase activity in the presence of miR-3148. Reduced modulation by miR-3148 conferred slower degradation of the risk G-allele containing TLR7 transcripts, resulting in elevated levels of gene products. These data establish rs3853839 of TLR7 as a shared risk variant of SLE in 22,613 subjects of Asian, EA, AA, and Amerindian/Hispanic ancestries (Pmeta = 2.0×10−19, OR = 1.25 [1.20–1.32]), which confers allelic effect on transcript turnover via differential binding to the epigenetic factor miR-3148.
| Systemic lupus erythematosus (SLE) is a debilitating autoimmune disease contributed to by excessive innate immune activation involving toll-like receptors (TLRs, particularly TLR7/8/9) and type I interferon (IFN) signaling pathways. TLR7 responds against RNA–containing nuclear antigens and activates IFN-α pathway, playing a pivotal role in the development of SLE. While a genomic duplication of Tlr7 promotes lupus-like disease in the Y-linked autoimmune accelerator (Yaa) murine model, the lack of common copy number variations at TLR7 in humans led us to identify a functional single nucleotide polymorphism (SNP), rs3853839 at 3′ UTR of the TLR7 gene, associated with SLE susceptibility in Eastern Asians. In this study, we fine-mapped the TLR7-TLR8 region and confirmed rs3853839 exhibiting the strongest association with SLE in European Americans, African Americans, and Amerindian/Hispanics. Individuals carrying the risk G allele of rs3853839 exhibited increased TLR7 expression at the both mRNA and protein level and decreased transcript degradation. MicroRNA-3148 (miR-3148) downregulated the expression of non-risk allele (C) containing transcripts preferentially, suggesting a likely mechanism for increased TLR7 levels in risk-allele carriers. This trans-ancestral mapping provides evidence for the global association with SLE risk at rs3853839, which resides in a microRNA–gene regulatory site affecting TLR7 expression.
| Systemic lupus erythematosus (SLE [OMIM 152700]) is a complex and heterogeneous autoimmune disease with a strong genetic component that is modified by environmental exposures. Although the detailed etiopathogenesis of SLE remains unknown, excessive innate immune activation involving toll-like receptors (TLRs, particularly TLR7/8/9) and type I interferon (IFN) has been recognized as an important pathogenic mechanism in the disease [1]. Therapeutics targeting the TLR/IFN pathway are in development for the treatment of SLE, with ongoing clinical trials investigating monoclonal antibodies against IFN-α and inhibitors for TLR7/TLR9 (reviewed in [2]). Recent genome-wide association (GWA) and follow-up studies have revealed the association of a number of polymorphic variants in genes encoding components of the TLR/type I IFN pathway with susceptibility to SLE (reviewed in [3], [4]), providing insights at the molecular level to refine our understanding of this dysregulated pathway in the predisposition to SLE.
Our previous study identified a single nucleotide polymorphism (SNP), rs3853839, in the 3′ UTR of an X-linked gene TLR7 to be associated with SLE in 4,334 cases and 4,940 controls of Eastern Asian descent [5], providing the first convincing evidence for the genetic contribution of TLR7 to human SLE. Individuals carrying the risk G allele exhibited increased TLR7 transcripts and a more robust IFN signature than non-risk C allele carriers [5]. In this study, by fine mapping the TLR7-TLR8 region, we confirmed that the previously reported functional SNP rs3853839, located within a predicted binding site of miR-3148, was most likely responsible for observed association with SLE in three populations of non-Asian ancestry. We demonstrated a differential miR-3148 modulation explaining the effect of allelic variation at rs3853839 on TLR7 expression.
We conducted genotyping and imputation for genetic variants covering ∼80 kb of the TLR7-TLR8 region on Xp22.2. After applying quality control measures, 41 genotyped SNPs and 57–75 imputed SNPs/INDELs (insertion-deletion) (varying among different ancestries) were assessed for association with SLE in unrelated cases and healthy controls of European American (EA, 3,936 cases vs. 3,491 controls), African American (AA, 1,679 vs. 1,934) and Hispanic enriched for the Amerindian-European admixture (HS, 1,492 vs. 807) descent (Figure 1A).
The strongest association signal was consistently detected at rs3853839 in the three ancestries, including EA (minor allele frequency of 20.3% in cases vs. 17.2% in controls, P = 6.5×10−6, OR [95%CI] = 1.23 [1.13–1.35]), AA (19.8% vs. 16.7%, P = 1.1×10−3, OR = 1.24 [1.09–1.41]) and HS (44.8% vs. 37.3%, P = 7.5×10−4, OR = 1.26 [1.10–1.43]) (Figure 1B, Table 1). After Bonferroni correction for multiple comparisons, the association of rs3853839 with SLE remained significant in EA and HS, and reached a nominal significance in AA. Combining the EA, AA and HS datasets, the meta-analysis P value of rs3853839 (Pmeta = 7.5×10−11, OR = 1.24 [1.18–1.34]) exceeded the commonly used threshold of 5×10−8 for genome-wide significance (Figure 1C, Table 1). Thus, the association of rs3853839 with SLE previously identified in Eastern Asians was confirmed in three non-Asian ancestries.
Only six other SNPs within a relatively small interval of 5 kb spanning from TLR7 3′downstream to TLR8 intron 1 were consistently associated with SLE (P<0.05) in EA, AA and HS (Table S1), and remained significant trans-ancestral meta-analysis P values after Bonferroni correction (5.5×10−6≤Pmeta≤1.3×10−6, Table S1). Linkage disequilibrium (LD) analysis revealed low LD strength between rs3853839 and these SNPs across non-Asian ancestries (r2<0.26, 0.37, and 0.51 in EA, AA and HS, respectively), but these 6 SNPs are in strong LD with each other and could be defined as a block (Figure S1). Among them, non-synonymous SNP rs3764880 (Met1Val) located at TLR8 exon1 exhibited the strongest association (Pmeta = 1.3×10−6, OR = 1.15; Table S1). To distinguish whether the associations of these 6 SNPs with SLE were independent of rs3853839, we performed conditional haplotype-based association test. After conditioning on rs3853839, association signals detected at these 6 loci were completely eliminated in EA, AA and HS (Figure S1). In contrast, conditioning on rs3764880, a consistent association signal was detected at rs3853839 in EA and HS (Figure S1), indicating that the association signals detected at these 6 SNPs might be attributed to that of rs3853839.
Taken together, we confirmed rs3853839 as the only SNP in the TLR7-TLR8 region showing an independent association with SLE across all three non-Asian ancestries. A meta-analysis by combining all datasets of Asian and non-Asian ancestries showed compelling evidence of association with SLE at rs3853839 (Pmeta = 2.0×10−19, OR = 1.25 [1.20–1.32], Table 1). Given the location of TLR7 at X chromosome, we examined the allelic association of rs3853839 separately by gender. Of note, the sex-specific association of rs3853839 with SLE previously detected in Asian men [5] was not replicated in non-Asian ancestries (Table 1).
Given the convincing evidence for the trans-ancestral association of rs3853839 with SLE susceptibility, we then evaluated its effect on regulation of TLR7/8 expression. Messenger RNA (mRNA) levels of TLR7 and the two alternative TLR8 isoforms were measured by real-time PCR in PBMCs from healthy EA individuals (n = 62). TLR7 mRNA levels were significantly different among women (n = 41) carrying different genotypes of rs3853839 [P = 0.003, one-way analysis of variance (ANOVA)], in which the GG and GC carriers exhibited notably increased TLR7 mRNA levels compared with the CC carriers [P = 0.02 for GG (n = 5) vs. CC (n = 18) and 0.02 for GC (n = 18) vs. CC, respectively, Student's t test; Figure 2A) and the number of rs3853839 risk G allele was significantly correlated with increased TLR7 mRNA levels (R2 = 0.26, P = 8×10−4, linear regression test). Consistently, male G allele carriers (n = 5) also had significantly higher TLR7 mRNA expression than male C allele carriers (n = 16) (P = 0.01, Figure 2A). There was no significant association of rs3853839 genotypes with mRNA levels of two TLR8 isoforms in either women or men (Figure 2A). No sex differences in TLR7 or TLR8 mRNA levels were observed between individuals carrying the same genotype [GG women vs. G men: P = 0.41 (TLR7), 0.63 (TLR8a) and 0.50 (TLR8b); CC women vs. C men: P = 0.10 (TLR7), 0.91 (TLR8a) and 0.65 (TLR8b)]. These results were in accordance with our previous observations in Chinese [5], supporting the importance of rs3853839 in regulating TLR7 rather than TLR8 gene expression.
We assessed the intracellular expression of TLR7 and TLR8 proteins by flow cytometry in PBMCs from 7 pairs of healthy women (GG vs. CC) and men (G vs. C), respectively. Of the 7 pairs of individuals in each gender, 4 pairs were of EA descent and 3 pairs were Asians. Compared with C allele carriers, G allele carriers had significantly higher TLR7 protein levels in PBMCs (P = 0.038 and 0.009 in women and men, respectively; Figure 2B), especially in CD19+ B cells and CD14+ monocytes (Figure S2). No significant association between rs3853839 genotypes and TLR8 protein levels was observed in either total PBMCs or in specific cell subsets (Figure S3).
We next performed luciferase reporter assays to further confirm the functional effect of rs3853839 on TLR7 expression. PCR-amplified TLR7 3′UTR fragments with either the G or C allele of rs3853839 were cloned downstream of an SV40 promoter-driven Renilla luciferase gene in the psiCHECK-2 vector, which also contained a firefly luciferase gene to serve as an internal transfection normalization control (Figure 2C). Constructs were then transiently transfected into either HEK-293 or differentiated HL-60 (dHL-60, neutrophil-like cells) cells. After 24 hours, cell lysates transfected with the G-allele construct showed significantly higher luciferase activity than those transfected with the C-allele construct in both HEK-293 and dHL-60 cells (P = 0.026 and 0.009, respectively; Figure 2C). Taken together, consistent results from ex vivo and in vitro studies indicated that the SLE-risk G allele of rs3853839 conferred elevated TLR7 expression at the both mRNA and protein level.
To explore the mechanism of rs3853839 in regulating TLR7 mRNA turnover, we assessed allelic difference in TLR7 mRNA degradation by pyrosequencing. We first determined the rs3853839 G/C allele ratio in genomic DNA (gDNA) and cDNA from healthy EA women (n = 7) carrying the GC genotype. The mean G/C allele ratio in cDNA was significantly higher than the theoretical ratio of 1 as detected in gDNAs (P = 0.02, Figure S4), indicating a higher expression of the G- than the C-allele containing TLR7 transcripts in heterozygous PBMCs. The allelic specific expression analysis in EA was similar to our previous findings in Chinese [5], and confirmed the result of real-time PCR that the G allele of rs3853839 is associated with increased TLR7 mRNA expression. Then, PBMCs were cultured in the absence or presence of the transcriptional inhibitor actinomycin D (ActD), and the G/C allele ratio in cDNA (normalized to that measured in gDNA) was determined after 0, 2, 4, 6, and 24 hours, respectively. As shown in Figure 2D and 2E, the G/C ratio in cDNA appeared to change over time when PBMCs were incubated with ActD and exhibited a statistical difference at the 4 hour point (P = 0.04), implicating slower degradation of the G allele- than the C allele-containing TLR7 transcript in heterozygous PBMCs. The inhibitory effect of ActD on RNA synthesis was corroborated by a decrease in total TLR7 mRNA level at increasing time points after the addition of ActD in PBMC aliquots measured by real-time PCR (Figure S5).
MicroRNAs (miRNAs) that bind to target sequences located within the 3′UTR of mRNAs by base pairing have been shown to result in accelerated mRNA turnover or translation repression [6]. Single nucleotide change either within or around the sequence of miRNA target sites can potentially alter the base-pairing patterns and affect miRNA-mediated regulation [7], [8]. The updated TargetScan database (Release 6.2; http://www.targetscan.org) indicates that rs3853839 is located within a binding site of miR-3148, where the non-risk allele (C), but not the risk allele (G), is predicted to match miR-3148 at the second position (Figure 3A). We hypothesized that the C to G variation of rs3853839 could reduce the binding and regulation incurred by miR-3148, therefore, leading to dysregulated TLR7 expression. We first showed that transcript levels of miR-3148 and TLR7 were inversely correlated in PBMCs from 16 patients with SLE and 21 healthy controls (R2 = 0.255, P = 0.001; Figure 3B), suggesting the possible regulation of TLR7 expression by miR-3148. Next, to verify whether allelic variation of rs3853839 affects the interaction of miR-3148 with TLR7 3′UTR, psiCHECK-2 vectors containing TLR7 3′UTR segment with either the C or G allele of rs3853839 were cotransfected with various doses of miR-3148 or nontarget control mimic into HEK-293 cells. As shown in Figure 3C, we observed significant dose-dependent miR-3148-mediated decrease in luciferase activity for the C-allele construct (P = 0.0003 over all miR-3148-treated C-allele vector groups, ANOVA test), but not for the G-allele construct (P = 0.14). Cotransfection with miR-3148 at a concentration of 6, 12, and 48 nM, respectively, led to greater than two-fold reduction of luciferase activity in the C-allele than the G-allele construct [reduction in C-allele vs. G-allele construct: 13.2% vs. 4.8%, P = 0.023 (6 nM); 22.5% vs. 9.9%, P = 0.0012 (12 nM); 21.4% vs. 8.5%, P = 0.0031 (48 nM)]. These data supported the bioinformatic prediction that miR-3148 directly targets TLR7 3′UTR and the C to G variation of rs3853839 within the binding site alters the inhibitory effect of miR-3148 on modulating TLR7 expression.
Fine-mapping of the TLR7-TLR8 region with high-density genetic markers based on large scale genotyping and imputation confirmed SNP rs3853839 at TLR7 3′UTR as the most likely causal variant responsible for the association of TLR7-TLR8 region with SLE in populations of EA, AA and HS ancestry. In accordance with our previous observation in Asians [5], we detected elevated TLR7 expression at both mRNA and protein levels in PBMCs from EA homozygous risk G allele carriers, as well as a higher level of the risk than the non-risk allele-containing TLR7 transcripts in EA heterozygous PBMCs. The fact that two distinct ancestries share the same genotype-phenotype association implicates an important regulatory effect of rs3853839 on TLR7 expression. Toward this end, we have extended functional studies showing slower degradation of the risk allele-containing TLR7 transcripts in heterozygous PBMCs and regulation of TLR7 expression by miRNA-3148 that targets 3′UTR at the position of rs3853839. Finally, we showed that the presence of the risk G allele resulted in reduced suppression by miRNA-3148, suggesting a likely mechanism for increased TLR7 expression in risk-allele carriers.
The importance of TLR7 upregulation on mediating autoimmune responses has been addressed in murine models of SLE. The Y-linked autoimmune accelerator (Yaa) modifier, suggested mainly due to Tlr7 gene duplication, provides a prime example of TLR7 dysregulation leading to autoreactivity and inflammatory pathology [9]–[11]. Increasing Tlr7 gene dosage via generation of transgenic mice results in development of systemic autoimmunity, the severity of which directly correlates with the degree of Tlr7 overexpression [12]. Increased Tlr7 gene dosage promotes autoreactive lymphocytes activation, dendritic cells proliferation, and secretion of proinflammatory cytokines and IFN-α [12], which in turn upregulates TLR7 expression, leading to a feedback loop exacerbating autoimmunity [13]. In patients affected with SLE, up-regulated expression of TLR7 mRNA has been reported in PBMCs and B cells [14], [15]. Although a copy number variation (CNV) study in Mexican population showed increased TLR7 copies in childhood-onset SLE patients [16], no evidence for common CNVs at the TLR7-TLR8 region has been identified in individuals of diverse ancestries through our previous study by three independent methods including quantitative real-time PCR, PmeI pulsed-field gel electrophoresis and Southern blot [5], two recent studies using customized CGH platforms [17], [18] as well as other studies listed in the Database of Genomic Variants (http://projects.tcag.ca/variation; the latest version released in November 2010), suggesting that mutations similar to Yaa are not a frequent feature of human SLE. The current study identifying genetic variations conferred by a regulatory SNP in TLR7 expression and SLE susceptibility suggests that murine models provide profound clues to human genetics if we look beyond the specific mutations identified in the relevant pathways.
Unlike our findings in Asians that both sexes showed association [5], the impact of rs3853839 on risk for SLE was only observed for women in the non-Asian datasets (Table 1). Given the low prevalence of SLE in men, it is often challenging to collect a large enough number of affected men in a given population. Under the assumption that the associated G allele confers genetic risk with an odds ratio of 1.26 in EA, 1.22 in AA and 1.29 in HS subjects (ORs were determined in female datasets), and considering P<0.05 as the threshold of significance, the power estimate for female samples in each ancestry reaches more than 85%, whereas for male samples it is only 50% in EA, 19% in AA and 25% in HS dataset. Thus, there was clearly inadequate power to evaluate this association in AA and HS men. Despite a relatively robust sample size of EA men (344 SLE vs. 1,151 controls), a significantly higher G allele frequency was observed in male than female controls (20.0% vs. 16.5%, P = 0.005), contributing to the difficulty in assessing association with SLE in EA male subjects.
To our knowledge, the association of rs3853839 (or its tag SNP) with SLE has not been reported in four SLE GWA studies in European-derived populations [19]–[22] and three GWA studies in Asians [23]–[25]. According to the 1000 Genomes Project data, rs3853839 locates in a region with poor LD structure and cannot be tagged by any known SNP at the TLR7-TLR8 region with r2>0.65. The SNP rs850632, located at TLR7 3′downstream, shows the strongest LD with rs3853839 in Europeans (r2 = 0.38) and Asians (r2 = 0.65). However, neither rs3853839 nor rs850632 has been included in predesigned commercial genotyping arrays of those GWA studies, resulting in the absence of associations. Even if rs3853839 was genotyped, the published GWA studies might have inadequate statistical power to capture its association in the initial discovery analyses [5].
Evidence of other TLR7 polymorphisms associated with SLE has been reported, including two intronic SNPs (rs179019 and rs179010) found in Japanese population [26] and an exonic SNP (rs179008) in individuals from Southern Brazil [27]. The reported associations were modest due to limited sample size of these studies (less than 400 cases and 450 controls), and none of them have been confirmed by the current fine-mapping study using a large collection of EA, AA and HS cases-controls (Table S1). TLR8 polymorphisms have been described in infectious diseases [28], [29] with a genetic effect localized to a functional variant at exon 1 (rs3764880, Met1Val). The G allele of rs3764880, which abolishes a putative start codon within the alternative TLR8 transcript isoform a (Figure 1A), conferred a protective effect on susceptibility to pulmonary tuberculosis in Indonesian and Russian men [28], as well as on HIV disease progression in Germans [29]. Our data showed a significantly increased frequency of rs3764880-G allele in SLE than healthy controls in the three non-Asian datasets; however, its association with SLE was dependent on that of TLR7 SNP rs3853839. Other variants at the TLR7-TLR8 region showed either weaker association than rs3853839 in trans-ancestral meta-analysis or association uniquely in EA or HS. Taken together, these data support rs3853839 as the most likely polymorphism associated with SLE shared by multiple ancestries. Although imputation facilitated our ability to capture common variants (MAF>1%), further refinement in genetic effects of rare variants (MAF<1%) is needed by deep sequencing of this locus, especially the intergenic region between TLR7 and TLR8 that was not well imputed in this study.
Variations in 3′UTR regions may be important in gene regulation. To date, expression quantitative trait loci (eQTL) mapping has been widely used for characterization of SNPs that affect gene expression [30]. Although the TLR7 expression has been measured in previous whole-genome eQTL studies, currently only those using EBV-transformed lymphoblastoid cell lines of 1000 Genomes Project individuals provide publically available genotyping data of rs3853839. Based on the study by Stranger et al [31], we found that CG carriers of rs3853839 showed elevated TLR7 expression compared with CC carriers in YRI women (P = 0.012). In male individuals, the G allele of rs3853839 showed a trend of association with elevated TLR7 expression in CHB+JPT, CEU and YRI men, and the association was significant when combining all male data (P = 0.014). These findings are consistent with our results that rs3853839 alleles are associated with differential TLR7 expression.
An important finding of this study is that the SLE-associated variant rs3853839 confers a genetic effect on modulation of TLR7 expression by an epigenetic factor miR-3148. Accumulating evidence suggests that miRNAs are fine tuners of TLR signaling pathways [32]. Regulation by miRNA may occur at various levels of TLR pathways by targeting adaptor molecules, downstream regulators and cytokines (reviewed in [32], [33]). However, few studies point to TLR themselves (e.g. TLR2 and TLR4) being directly targeted by miRNAs [34], [35]. Using algorithms from TargetScan, only the newly identified human miR-3148 [36], which is not evolutionarily conserved among mammals, is predicted to bind TLR7 3′UTR sequences at the position of rs3853839. The inverse correlation of miR-3148 and TLR7 levels in PBMCs, along with functional validation by reporter gene assay, confirms an inhibitory effect of miR-3148 on regulating TLR7 expression and allelic variation of rs3853839 affecting miRNA-mRNA interactions. Further study will focus on investigating miR-3148 expression patterns in specific immune cell types, assessing biological impacts of changes in miR3148-mediated TLR7 expression on downstream immune responses, and evaluating roles of other miRNAs that target sequences in the vicinity of rs3853839. Of interest, an unconventional role for miRNAs has been identified as endogenous activators for RNA-sensing receptors (TLR7/8) in a cell- or tissue-type specific manner [37], [38]. Therefore, miRNA regulation in TLR7 signaling is more complicated than we expected and further functional studies showing the exact effects of miRNAs on TLR7 responses are warranted.
In summary, we have advanced our previous study by showing rs3853839 (at TLR7 3′UTR) as the most likely polymorphism responsible for the association of TLR7-TLR8 region with SLE in individuals of EA, AA and HS ancestry, and have characterized a differential miR-3148 modulation which explains the effect of allelic variation of rs3853839 on TLR7 expression. Our study highlights the importance of TLR7 as a shared genetic contributor to SLE in multiple ancestries, and provides evidence that microRNA acts as a negative regulator to control TLR7 expression, suggesting the possibility of miRNA-based therapies for amelioration of autoimmune diseases such as SLE where excessive TLR7 activation exists.
Written informed consent was obtained from all study participants and each participating institution had Institutional Review Board (IRB) approval to recruit samples. The overall study was approved by the IRB of the Oklahoma Medical Research Foundation (OMRF).
To test the association of TLR7-TLR8 with SLE, we used a large collection of case-control subjects from the collaborative Large Lupus Association Study 2 (LLAS2), including European American (4,248 cases vs. 3,818 controls), African American (1,724 cases vs. 2,024 controls), and Hispanic enriched for the Amerindian-European admixture (1,622 cases vs. 887 controls). African Americans included 286 Gullahs (155 cases vs. 131 controls), who are subjects with African ancestry. Cases were defined by meeting at least four of the 1997 American College of Rheumatology (ACR) revised criteria for the classification of SLE [39].
DNA samples were processed at the Lupus Genetics Studies Unit of OMRF. SNP genotyping was performed using an Illumina custom bead array on the iSCAN instrument for 47 SNPs covering the TLR7-TLR8 region on Xp22.2 and 347 admixture informative markers (AIMs). SNPs meeting the following criteria were included in the association analysis: well-defined cluster scatter plots, SNP call rate >90%, minor allele frequency >1%, total proportion missing <5%, P>0.05 for differential missing rate between cases and controls, and Hardy-Weinberg proportion (HWP) test with a P>0.01 in controls and P>0.0001 in cases.
Subjects with genotype missing rate >10% (due to low quality), shared identical by descent >0.4 or showing mismatch between the reported and estimated gender were removed. The global ancestry of each subject was estimated based on genotype of AIMs using principal components analysis [40] and ADMIXMAP [41], as described in another LLAS2 study [42], and then genetic outliers were removed.
Finally, a total of 13,339 unrelated subjects, including European Americans (EA: 3,936 cases vs. 3,491 controls), African Americans (AA: 1,679 vs. 1,934; composed of 92.5% of African Americans and 7.5% Gullahs) and Hispanics enriched for the Amerindian-European admixture (HS: 1,492 vs. 807), were analyzed for 41 genotyped SNPs of TLR7-TLR8.
Imputation was performed at 12.86–12.95 Mb on Xp22.2 using IMPUTE 2.1.2 [43], with SNP/INDEL genotypes of 381 Europeans, 246 Africans and 181 Americans from the 1000 Genomes Project (“version 3” of the Phase 1 integrated data, March 2012 release) as references in imputation for our EA, AA and HS subjects, respectively. Imputed genotypes had to meet information score of >0.9, as well as the quality control criteria as described above. After imputation, we obtained an additional 75 variants for EA, 57 for AA and 63 for HS (the number varied based on LD structure) for further analysis.
Total RNA was purified with TRIzol reagent (Invitrogen) from PBMCs and reverse-transcribed into cDNA with Superscript II Reverse Transcription kit (Invitrogen). The mRNA levels of TLR7 (NM016562.3) and TLR8 (isoform a: NM138636.4 and isoform b: AF246971.1) were measured by quantitative real-time PCR using TaqMan assays (TLR7 probe: Hs00152971_m1; TLR8 isoform a probe: Hs00607866_mH; TLR8 isoform b probe: Hs00152972_ml, Applied Biosystems). All samples were run in triplicate. Relative expression levels of TLR7 and TLR8 were normalized to the level of RPLP0, calculated by the 2−ΔΔCt method and Log10 transformed. The association of rs3853839 with mRNA levels of TLR7 or TLR8 was evaluated using ANOVA, Student's t and linear regression test.
To examine the correlation of miR-3148 and TLR7 mRNA levels, total RNA enriched in small RNAs were isolated from PBMCs using mirVanaTM miRNA isolation kit (Invitrogen), followed by reverse transcription with TaqMan MicroRNA Reverse Transcription kit (Applied Biosystems; for detecting miR-3148) and Superscript II Reverse Transcription kit (Invitrogen; for detecting TLR7), respectively. The miR-3148 level was quantified using Taqman MicroRNA Expression assay (Applied Biosystems), and the TLR7 level was measured using the same probe as described above. All samples were run in triplicate. Relative expression levels of miR-3148 and TLR7 were normalized to the level of snRNA U6 and RPLP0, respectively, calculated by the 2−ΔΔCt method and Log10 transformed. Association between transcript levels of TLR7 and miR-3148 was evaluated using linear regression test.
Four-color flow cytometry was performed to investigate intracellular expression of TLR7 and TLR8 in PBMCs from healthy EA and Asian individuals who were homozygous for rs3853839 (7 pairs of G-allele vs. C-allele carriers in each gender group). Freshly isolated PBMCs were incubated with 2% pooled human serum to block nonspecific binding to Fcγ receptors and then incubated with peridinin chlorophyll protein (PerCP)-conjugated anti-human CD3, allophycocyanin (APC)-conjugated anti-human CD19 and phycoerythrin (PE)-conjugated or fluorescein isothiocyanate (FITC)-conjugated anti-human CD14 (Miltenyi Biotec) to identify T cell, B cell and monocyte subpopulations, respectively. For intracellular staining, PBMCs were fixed in Fixation buffer (R&D Systems) for 10 minutes at room temperature, washed twice in Permeabilization/Wash buffer (R&D Systems) and stained with PE-conjugated mouse anti-human TLR7 mAb (R&D Systems) and FITC-conjugated mouse anti-human TLR8 mAb (Imgenex) for 1 hour at room temperature. Background fluorescence was assessed using appropriate isotype- and fluorochrome-matched control antibodies. Cells were collected and analyzed by FACSCalibur flow cytometer equipped with the manufacturer's software (CellQuest; BD Biosciences). Student's t test was used to compare protein levels of TLR7 or TLR8 in PBMCs from individuals of different genotypes.
The fragment of TLR7 3′-UTR bearing the G or C allele of rs3853839 was amplified by PCR from genomic DNA of subjects homozygous for the G or C allele using the following primers: 5′-TGTCTCGAGCCCTTCTTTGCAAAAC-3′ (forward) and 5′-AGAGCGGCCGCTAGTTGGCTCCAGCAAT-3′ (reverse). The PCR products were inserted into the downstream of the Renilla luciferase gene in the reporter vector psiCHECK-2 (Promega) by digestion using the restriction enzymes Not I and Xho I. The psiCHECK-2 vector also contained a firefly luciferase gene to serve as an internal transfection normalization control. All constructs were sequenced to assure proper orientation and authenticity in the vector.
HEK-293 (human embryonic kidney cell line) and HL-60 (human leukemic cell line) cells were obtained from the American Type Culture Collection (ATCC). HEK-293 cells were maintained in Dulbecco's modified Eagle's medium supplemented with 10% FBS, seeded on a 24-well plate at a concentration of 2×105 cells/well, and transiently transfected using Lipofectamine 2000 (Invitrogen) with 1 µg of either rs3853839 G or C reporter construct. HL-60 cells are predominantly a neutrophilic promyelocyte (precursor) and can be induced to differentiate to neutrophil-like cells when grown in RPMI 1640 medium with 15% FBS plus 2 mM L-glutamine, 25 mM HEPES and 1.25% DMSO [44]. Differentiated HL-60 cells seeded on 24-well plates (2×106 cells/well) were electroporated with 3 µg of report construct on a nucleofector device (Amaxa). The luciferase activity in total cell lysates was measured after 24 hours using a dual luciferase reporter assay system (Promega). Renilla luciferase activities were normalized to firefly luciferase activities. Each transfection was performed in quadruplicates and triplicates for HEK-293 and HL-60 cells respectively, and luciferase assays were repeated four times.
MicroRNA hsa-miR-3148 and nontarget control (NC) mimics were synthesized by Thermo Fisher Scientific. To test the effect of miR-3148, HEK-293 cells plated in 96-well plates were transiently cotransfected with 100 ng of each reporter construct (psiCHECK-2 empty vector, rs3853839-G or -C allele constructs) and increasing concentrations (1, 6, 12 and 48 nM) of miR-3148 or nontarget control mimic using Lipofectamine 2000 reagent (Invitrogen), and luminescence was measured 24 hours after transfection. Each transfection was performed in quadruplicates and repeated three times. Luciferase activity of reporter vectors was compared using Student's t test.
PBMCs isolated from EA healthy women with the GC genotype of rs3853839 (n = 7) were cultured in the absence or presence of 5 µg/mL ActD for 0, 2, 4, 6 and 24 hours. Using real-time PCR, we detected a decrease in total TLR7 mRNA levels over time with ActD incubation, which confirmed the transcriptional inhibition by ActD and allowed for detection of allelic differences in mRNA degradation. The G/C allelic ratio in the cDNA and gDNA after treatment of PBMCs with or without ActD were determined by pyrosequencing and calculated using software PSQMA 2.1 (Biotage) as previously described [5]. The G/C allele ratio obtained in TLR7 transcripts was normalized to that measured from gDNA of the same sample. A paired t test was used to compare the mean G/C allele ratio in TLR7 transcripts in PBMCs treated with ActD or vehicle control at each time point.
Associations of SNPs with SLE were assessed in each ancestral group under a logistic regression model adjusted for gender and the first three principal components estimated using AIMs. Conditional haplotype-based association tests were also performed by adjusting for gender and the first three principal components. The trans-ancestral meta-analysis was conducted on 40 genotyped and 14 imputed SNPs that were shared by the three ancestries with both a fixed and random-effects model. Homogeneity of odds ratios was evaluated using Cochrane's Q test. For each SNP, if the Cochran's Q test showed no evidence of genetic heterogeneity (P>0.05), a fixed-effects model was implemented; otherwise, a random-effects model was used. The Bonferroni corrected P-value threshold was adjusted to P<9.1×10−4 on the basis of the maximum number of tests across all populations (55 independent variants with r2<0.8). All analyses described above were performed using PLINK v1.07. Pairwised LD values shown in Figure 1 and Figure S1 were calculated using Haploview 4.2. Other data were analyzed using GraphPad Prism 4.0 software. A P value<0.05 was considered to be statistically significant.
|
10.1371/journal.pcbi.1000775 | On the Zwitterionic Nature of Gas-Phase Peptides and Protein Ions | Determining the total number of charged residues corresponding to a given value of net charge for peptides and proteins in gas phase is crucial for the interpretation of mass-spectrometry data, yet it is far from being understood. Here we show that a novel computational protocol based on force field and massive density functional calculations is able to reproduce the experimental facets of well investigated systems, such as angiotensin II, bradykinin, and tryptophan-cage. The protocol takes into account all of the possible protomers compatible with a given charge state. Our calculations predict that the low charge states are zwitterions, because the stabilization due to intramolecular hydrogen bonding and salt-bridges can compensate for the thermodynamic penalty deriving from deprotonation of acid residues. In contrast, high charge states may or may not be zwitterions because internal solvation might not compensate for the energy cost of charge separation.
| In the last two decades mass spectrometry has given an impressive contribution to biochemistry, protein science, proteomics and structural biology. This technique offers powerful insights into protein structure and dynamics along with useful information on the role of solvent in protein stability as it is able to preserve non-covalent interactions and globular structures during the proteins' flight inside the mass spectrometer. Unfortunately, the key issue of the charge state of ionizable groups, presumably different from that in solution, has not been elucidated yet. So far conflicting assumptions and conclusions have been drawn by several groups. In the present work a very accurate structural and energetic analysis of the protonation state of two peptides and a small protein in the gas phase was performed. Results suggest that internal solvation can stabilize charge separation with the formation of zwitterionic states.
| Predicting the structural properties of proteins in the gas phase is crucial to interpret mass spectrometry data, yet this is far from being understood [1]–[10]. So far, it has been established that (i) compact structures acquire smaller net charges than unfolded ones [11]–[13], (ii) secondary and tertiary structure elements play a crucial role for protein fragmentation [14]–[23], and (iii) hydrogen bonds (H-bonds) and salt-bridges [3], [24], [25] may stabilize the structures. However, how desolvation impacts on structural facets of proteins [2], [3], [8], [26]–[30], peptides [3], [31]–[35] and even single amino acids [36]–[45] is matter of a vivid debate.
A key point is the presence of charge separation. Whilst amino acids exist mostly in their zwitterionic form in the aqueous solution [31], [36], [40], [46], conflicting assumptions and conclusions have been drawn for the same molecules in vacuo [47]–[50]. For peptides and proteins, the key issue of the charge state of ionizable groups, presumably different from that in solution, is even less clear [2], [51]–[53]. One line of thought assumes neutral acidic functions for proteins analyzed in positive-ion mode (i.e., generating and detecting positively charged ions) and neutral basic sites in negative-ion mode. In other words, the number of ionized groups is generally assumed to be equal to the net charge of the protein ion [54]–[56]. Electrostatic energy calculations based on this supposition fail to reproduce experimental values of apparent gas-phase basicity (GPB) for folded protein ions [57]. The GPB of a basic species B is defined as the negative of the free-energy change, , for the gas-phase protonation reactionIf B is the conjugate base of an acid AH, then , where GA is the gas-phase acidity of AH. Analogously, the proton affinity is defined as the negative of the protonation enthalpy, PA = −.
In contrast, an increasing number of experimental [16], [24], [25], [58]–[61] and theoretical [62], [63] investigations carried out on peptides and small proteins indicate that zwitterionic states may survive in the absence of solvent if the structural features allow for adequate intramolecular solvation [64]–[67]. Recent ultraviolet photo-dissociation [16] and fluorescence [25], [61] experiments indicate the presence of stabilizing salt-bridge motifs in small biomolecules. Salt bridges exist also in protonated, gas-phase serine dimers [24] and have been predicted for arginine dimers [63], [68], [69]. These interactions add to other stabilizing contributions such as hydrogen bonds [3], [24], [25]. Molecular dynamics (MD) simulations on a minimalistic lattice model of a zwitterionic system [1] turned out to reproduce the experimental observation that compact structures acquire smaller net charges than unfolded ones [11]–[13]. On the basis of these simulations, it has been also proposed that steric and electrostatic shielding of charged residues in compact conformations are the major factors responsible for this structural facet. Energy calculations [2], [7], [52] and measurements [51], [70] on several well characterized proteins in their experimentally observed, most populated charge state suggest that the presence of zwitterions depends on the specific protein structure [2], [51]. Deprotonated aspartic and glutamic residues persist in the most abundant, positively charged protomer of insulin, the C-terminal domain of the ribosomal protein L7/L12 and ubiquitin, but not in tryptophan-cage and lysozyme [2].
Prompted by the current lack of understanding of the charge state of protein ions in vacuo, here we have carried out an exhaustive energy analysis on three systems largely studied in the gas phase both experimentally [16], [25], [59], [60], [71]–[77] and theoretically [52], [62], [63], [72]. These are the 8-residue peptide angiotensin II (AN) [74]–[76] and the 9-residue peptide bradykinin (BK) [16], [59], [62], [63], [71]–[73], as well as the tryptophan-cage (Trp-cage) [16], [25], [52], [60], [77] protein. The latter is a 20-residue mini-protein with a well defined secondary and tertiary structure in aqueous solution at ambient conditions. It consists of an -helix and a compact hydrophobic core formed by a Trp side chain from the -helix, surrounded by several hydrophobic residues (two prolines and one tyrosine) [78].
A complete exploration of the protomer space (i.e., all of the possible charge configurations compatible with a given charge state) of these biomolecules is performed coupling force field–based molecular dynamics and density functional theory (DFT) calculations. In contrast to previous computational studies [25], [31], [52], [62], [63], [79], [80], all of the charge states generated by ionized and/or neutral basic (R, K, H, Q, N-terminus) and acidic groups (E, D, C-terminus), featuring more than one protomer, are taken into account.
A computational protocol apt to this task has been developed, allowing for an exhaustive exploration of the conformational space of each protomer. Based on such protocol, we suggest that low-charge states are likely zwitterions. In those cases, H-bonds and salt-bridges stabilize largely zwitterionic states, considerably reducing the differences in the apparent GPB between basic residues and the conjugated base of acidic residues. At high net charge, instead, non-zwitterion states are most likely.
The sequences of BK, AN and Trp-cage are RPPGFSPFR, DRVYIHPF, NLYIQWLKDGGPSSGRPPPS, respectively. For each system, the following protonation sites were considered: , , N- and C-term for BK; , , , N- and C-term for AN ( in the neutral state can be protonated either in or , both tautomers were considered); , , , , and N- and C-term for Trp-cage. In the latter, protonation of was considered for the and charge states on the basis of experimental evidences [19].
BK and AN have no determined secondary structure and all of the calculations started with an all-trans backbone and side-chain conformation. Instead, the Trp-cage initial structure was obtained by a 20-ns MD simulation in aqueous solution at ambient conditions based on the NMR structure number 1 deposited in the protein data bank (PDB code: 1L2Y) [78] (see Text S1). The most probable protonation state in water [78] was chosen.
For the chosen set of protonation sites, all of the charge states which feature more than one protomer were taken into account. For these charge states, all of the possible protomers were considered, for a total of 100 protomers (see Tables 1, 2, and 3).
OPLS/AA force field-based [81], [82], constant-temperature MD calculations and geometry optimizations were carried out. The cutoff of electrostatics and van der Waals interactions was fixed at 0.7nm. In the MD simulations, the equations of motion were numerically integrated with a time step of 1.5 fs. All the hydrogen-bond lengths were kept fixed using the LINCS [83] algorithm. The temperature was controlled by the Nosé-Hoover thermostat [84]. The results of force field based MD simulations depend critically on the charge state used. Therefore, we performed a simulation for each protonation state. Specifically 8-ns MD simulations at high-temperature (700K for AN and BK, 350K for Trp-cage) were performed for each protonation state. The chosen temperatures were selected after several careful tests. In particular, for Trp-cage, a temperature of 350K turns out to allow for an exhaustive sampling of the side chains conformations without disrupting, in the relatively short simulation time, the secondary structure. The resulting trajectories were split into 5-ps, non overlapping time windows. For each window, the geometry of the lowest-energy MD conformation was optimized by a conjugated gradient scheme up to 0.1 kJ/molÅ residual force on any atom. This simulated annealing-like procedure yielded for each protomer a large set of conformations. The geometry of structures within 100 kJ/mol (60 kJ/mol for Trp-cage) from the lowest-energy force field conformer were refined at the ab initio level (see Section “Identifying relevant protomers of a given charge state”). With this criterion, 60 conformers (35 for Trp-cage), were randomly selected from equally spaced energy windows, one from each window, and re-optimized at DFT/BLYP level of theory.
The GROMACS [85] software package was used for all MD calculations.
Quantum-chemical geometry optimizations were performed within the framework of DFT. The Becke exchange [86] and Lee-Yang-Parr [87] correlation functionals (BLYP) were used in a hybrid Gaussian and plane wave approach [88]. The wave function was optimized by using an orbital transformation technique [89] and analytic Goedecker-Teter-Hutter [90], [91] pseudopotentials (PP). The TZV2P Gaussian basis set was used for valence electrons of all atoms, while the auxiliary electron density was expanded in plane waves up to a cutoff of 280 Ry.
The interaction between periodic images in the reciprocal space was removed according to the decoupling scheme presented in [92]. The calculations were carried out with the CP2K code [89], [93], [94], which has been shown to be very efficient for these systems.
The adopted DFT scheme was validated against more accurate (and more expensive) quantum chemistry methods. First, the relative energy of canonical and zwitterionic arginine conformers calculated with the present scheme agrees well with that obtained from all-electrons B3LYP, MP2, and CCSD calculations (see Table 2 in Text S1). Second, all of the 14 protomers of AN with total charge underwent all-electrons, single-point energy evaluations at DFT/B3LYP level with the 6–311++G(d,p) basis set using the Gaussian03 code [95] (Angiotensin II was chosen because it is the smallest of the three molecules studied and, in particular, the charge state was considered because it presents the largest set of protomers, and it is, therefore, a good benchmark case).These and the previous calculations provided the same energy ranking (see Table 3 in Text S1).
A final concern for using DFT for non-covalent systems is the underestimation of dispersion forces [96], [97]. This flaw of the current GGA functionals might influence the conformational energy, especially in the case of large molecular assemblies like those considered here. To quantify this error an estimate of the dispersion energy was performed for the DFT optimized structures using the OPLS/AA force field. The results of this calculation (see Tables 4, 5 and 6 in Text S1) indicate that the dispersion energy is not expected to change qualitatively the DFT energy ranking of protomers.
A standard procedure to identify the relevant protomers is currently lacking, even for peptides with more than a few amino-acids. On the one hand, the high complexity of the conformational space hampers an exhaustive search based on first-principle quantum chemistry (such as DFT) of the minimum-energy conformers. On the other hand, force field–based calculations [62], [63], [98], [99], or semiempirical quantum chemical methods [52], may not be accurate enough. For instance, Merck molecular force field [100] energies have been shown to correlate poorly with those calculated at the DFT/B3LYP level [62], [63]. In addition, the energies calculated by force fields do not take into account higher-order effects, which may play a role in our systems. DFT can, instead, take such effects into account.
However, if the empirically calculated conformer is much higher in energy than another (say with a greater than few hundreds of kJ/mol), it will be highly probable that the same ranking holds at the ab initio level (see Figure 1 in Text S1). Here, we seek such value by performing MD simulations based on the OPLS/AA, which offers the most complete set of base/conjugate acid pairs. The calculations on the three systems in vacuo provided several hundreds conformations, which then underwent DFT/BLYP [86], [87] geometry optimizations. Such quantum chemical scheme is extremely efficient for large molecules, as those investigated here [101], [102].
We found that less than 5% of the ab initio conformers located within 10kJ/mol from the energy minimum fall more than = 100kJ/mol (60kJ/mol for Trp-cage) above the OPLS/AA minimum (see Figure 2 in Text S1). Exploiting this fact, we used the ensuing protocol to identify the lowest-energy minimum for each charge state for each peptide: (i) generation of conformers for all possible protomers by OPLS/AA MD and simulated annealing-like calculations; (ii) elimination of conformers whose energy is larger than from the absolute minimum; (iii) DFT/BLYP geometry optimization of the conformers within ; (iv) ranking of the conformers based on their DFT energies.
Errors of this protocol are associated with (i) the accuracy of the DFT approach, (ii) limitations of sampling and ( iii) absence of entropy contributions. This points are discussed in the following.
We therefore conclude that the ranking obtained with our protocol provides a reliable identification of the most stable protomers.
We discuss here the salient structural data of the low-energy protomers identified with the protocol outlined above for each system and for every charge state that features, according to our choice of ionizable residues, more than one protomer. More details and additional observations can be found in Text S1. Structural data for each protomer of the considered charge states are reported in Tables 1, 2, and 3.
Our calculations suggest that most of the low-charge states are zwitterions, whilst high charge states might not. We now analyze the key factors for the stabilization of these two different states.
A computational protocol aimed at identifying the most stable species of angiotensin II, bradykinin, and tryptophan-cage has been developed and may be easily extended to other systems of similar size. The protocol provides results fully consistent with the experimental data. The results suggest that most of the low-charge states are zwitterions. Intramolecular interactions can stabilize zwitterionic states considerably, by reducing the differences in apparent GPB between basic residues and the conjugated base of acidic residues Based on a combined structural and energetic analysis, we suggest that salt-bridges provide a key energetic stabilization, in agreement with previous findings [3], [38], [51], [63], [116]. Indeed, the stabilization due to salt bridging might be such to reduce enormously the GPB of the biomolecules considered in the present study (up to 900 kJ/mol). H-bonding also has an important role in promoting charge separation. As a result, networks are formed where two (or more) salt bridges are clustered together, whenever it is possible.
Thus, we further corroborate the hypothesis that deprotonated carboxylate groups can be maintained in gas-phase peptide and protein ions produced by electrospray in positive-ion mode (and, vice-versa, protonated basic groups in negative-ion mode) [1], [2], [30], [38], [39], [62], [63], [111]. On the other hand, the formation of zwitterionic species in high charge states requires the protonation of residues with progressively lower GPB, which is accompanied by a large thermodynamic penalty that might not be compensated by internal solvation.
|
10.1371/journal.ppat.1007009 | Cryo-EM structure of infectious bronchitis coronavirus spike protein reveals structural and functional evolution of coronavirus spike proteins | As cell-invading molecular machinery, coronavirus spike proteins pose an evolutionary conundrum due to their high divergence. In this study, we determined the cryo-EM structure of avian infectious bronchitis coronavirus (IBV) spike protein from the γ-genus. The trimeric IBV spike ectodomain contains three receptor-binding S1 heads and a trimeric membrane-fusion S2 stalk. While IBV S2 is structurally similar to those from the other genera, IBV S1 possesses structural features that are unique to different other genera, thereby bridging these diverse spikes into an evolutionary spectrum. Specifically, among different genera, the two domains of S1, the N-terminal domain (S1-NTD) and C-terminal domain (S1-CTD), diverge from simpler tertiary structures and quaternary packing to more complex ones, leading to different functions of the spikes in receptor usage and membrane fusion. Based on the above structural and functional comparisons, we propose that the evolutionary spectrum of coronavirus spikes follows the order of α-, δ-, γ-, and β-genus. This study has provided insight into the evolutionary relationships among coronavirus spikes and deepened our understanding of their structural and functional diversity.
| Because of their structural and functional diversity, coronavirus spike proteins represent a model system for studying viral evolution. Their evolutionary relationships and history pose major puzzles in virology. A critical missing piece in understanding the evolution of coronavirus spikes is the unavailability of the tertiary structure of γ-coronavirus spikes. Here we determined the first cryo-EM structure of avian infectious bronchitis coronavirus (IBV) spike from the γ-genus. The IBV spike contains structural features that are unique to different other genera; hence it brings the spikes from different coronavirus genera into one evolutionary spectrum, with itself sitting in the middle of this spectrum. The order of the evolutionary spectrum is α-, δ-, γ-, and β-genus. Importantly, through comparisons between IBV spike and other coronavirus spikes, this study illustrates how coronavirus spikes have achieved structural and functional diversity through evolution to guide viral entry into host cells.
| As large enveloped RNA viruses, coronaviruses are capable of adapting to new hosts with relative ease through mutations and recombinations [1–3]. As a result, coronaviruses infect a wide range of mammalian and avian species, and have genetically evolved into four major genera: α, β, γ, and δ [4]. Coronaviruses from the four genera all contain envelope-anchored spike proteins that mediate viral entry into host cells [5, 6]. During viral entry, the spikes bind to host receptors through their S1 subunits and then fuse viral and host membranes through their S2 subunits. On the one hand, the spikes interact with host receptors and other host factors, hence needing to evolve for better adaptation to these host factors [7–10]. On the other hand, they are exposed to the host immune system, thereby needing to evolve to evade the host immune surveillance [11–14]. Consequently, the spikes are the most divergent among all coronavirus proteins [6]. The S1 subunits are particularly divergent, with little or low sequence similarities across different genera [15]. How coronavirus spikes have evolved to their current diverse structures imposes a major evolutionary conundrum.
Traces of protein evolution can often be found more reliably in their tertiary structures and related functions than in their primary structures, because proteins generally need to evolve within certain structural and functional constraints [16, 17]. To decipher the evolutionary puzzles surrounding coronavirus spikes, extensive structural studies have been carried out using both X-ray crystallography and cryo-electron microscopy (cryo-EM) [12–14, 18–28]. These studies have resulted in structure determinations of S1 from the α- and β-genera and spike ectodomains from the α-, β-, and δ-genera. Coronavirus spikes exist in two distinct conformations: the pre-fusion structures are present on mature virions and have a clove-like shape with three S1 heads sitting on top of a trimeric S2 stalk [12–14, 18–20]; the post-fusion structures are the membrane-fusion state and have a dumbbell-like shape with three S2 subunits forming a six-helix bundle structure [29–35]. Whereas the structures of S2 from different genera are similar to each other in both the pre- and post-fusion states, the S1 subunits from different genera diverge structurally and they also recognize a variety of host receptors [6, 7]. S1 contains two domains, N-terminal domain (S1-NTD) and C-terminal domain (S1-CTD), either or both of which can function as the receptor-binding domain. The S1 domains from different genera contain structural features that are unique to their genus. S1-CTDs are particularly diverse, with low or no structural similarity across different genera [15]. Overall, these previous studies have provided structural snapshots of coronavirus spikes from α-, β-, and δ-genera. However, because the structures of γ-coronavirus spikes were still missing, we lacked a clear picture of the evolutionary relationships among coronavirus spikes from different genera.
In the current study, we determined the cryo-EM structure of avian infectious bronchitis coronavirus (IBV) spike, the first such structure from the γ genus. The IBV spike possesses structural features that are unique to different other genera, suggesting that it falls in the middle of an evolutionary spectrum of coronavirus spikes. We also discuss how the structural evolution of coronavirus spikes has affected their functions as cell-invading molecular machinery. Overall, our study has filled in a critical gap in the structural, functional and evolutionary studies of coronavirus spikes, and deepened our understanding of viral evolutions in general.
IBV spike gene (virus strain M41; GenBank number ABI26423.1) was synthesized (Genscript) with codons optimized for insect cell expression. Its ectodomain (residues 20–1084) was cloned into pFastBac vector (Life Technologies Inc.) with a N-terminal honeybee melittin signal peptide and C-terminal GCN4 and His6 tags. It was expressed in Sf9 insect cells using the Bac-to-Bac system (Life Technologies Inc.) and purified as previously described [21]. Briefly, the protein was harvested from cell culture medium, and purified sequentially on Ni-NTA column and Superdex200 gel filtration column (GE Healthcare). IBV S1-CTD (residues 248–495) was expressed and purified in the same way as the IBV spike ectodomain, although it only contains a C-terminal His6 tag and does not contain the GCN4 tag.
For sample preparation, aliquots of IBV spike ectodomain (3 μl, 0.35 mg/ml, in buffer containing 2 mM Tris pH7.2 and 20 mM NaCl) were applied to glow-discharged CF-2/1-4C C-flat grids (Protochips). The grids were then plunge-frozen in liquid ethane using a FEI MarkIII Vitrobot system (FEI Company).
For data collection, images were recorded using a Gatan K2 Summit direct electron detector in the direct electron counting mode (Gatan), attached to a FEI Titan-Krios TEM, at Arizona State University. The automated software SerialEM [36] was used to collect ~2,000 total movies at 37,700x magnification and at a defocus range between 1 and 3 μm. Each movie had a total accumulated exposure of 53.66 e/Å2 fractionated in 50 frames of 200 ms exposure. Data collection statistics are summarized in S1 Table.
For data processing, whole frames in each movie were corrected for beam-induced motion and dose compensation using MotionCor2 [37] and ~1,400 best images were manually selected (we manually discarded micrographs with only carbon field of view or thick ice after motion correction as well as micrographs with defocus parameter higher than 4.5 μm after CTF estimation). The final image was bin-averaged to lead to a pixel size of 1.02 Å. The parameters of the microscope contrast transfer function were estimated for each micrograph using GCTF [38]. Particles were automatically picked and extracted using RELION [39] with a box size of 320 pixels. Initially, ~802,000 particles were subjected to 2D alignment and clustering using RELION, and the best classes were selected for an additional 2D alignment. ~5,000 best particles were applied for creating the initial 3D model using RELION. ~170,000 particles selected from 2D alignment were then subjected to 3D classification and the best class with ~100,000 particles were subjected to 3D refinement to generate the final density map. The final density map was sharpened with modulation transfer function of K2 operated at 300kV using RELION post-processing. Reported resolutions were based on the gold-standard Fourier shell correlation (FSC) = 0.143 criterion, and Fourier shell correction curves were corrected for the effects of soft masking by high-resolution noise substitution [40]. Data processing statistics are summarized in S1 Table.
The initial model of IBV spike ectodomain was obtained by fitting the seven parts (S1-NTD, S1-CTD, two parts of SD1, two parts of SD2, and S2) of the porcine delta coronavirus spike structure (PDB ID: 6B7N) individually into the cryo-EM density map of IBV spike using UCSF Chimera [41] and Coot [42]. Manual model rebuilding was carried out using Coot based on the well-defined continuous density of the main chain; the side chain assignments were guided by the densities of N-linked glycans and bulky amino acid residues. The structural model of the IBV spike in the pre-fusion state was refined using Phenix [43] with geometry restrains and three-fold noncrystallographic symmetry constraints. Refinement and model rebuilding in Coot were carried out iteratively until there were no further improvements in geometry parameters and model-map correlation coefficient. The quality of the final model was analyzed with MolProbity [44] and EMRinger [45]. The validation statistics of the structural models are summarized in S1 Table.
IBV pseudovirus entry assay was carried out as previously described [46]. Briefly, full-length IBV spike gene was inserted into pcDNA3.1 (+) plasmid. Retroviruses pseudotyped with IBV spike and expressing a luciferase reporter gene were prepared through co-transfecting HEK293T cells (source: American Type Culture Collection) with a plasmid carrying Env-defective, luciferase-expressing HIV-1 genome (pNL4-3.luc.RE) and the plasmid encoding IBV spike. The produced IBV pseudoviruses were harvested 72 hours post transfection, and then used to enter DF-1 cells (source: American Type Culture Collection) and HEK293T cells. After incubation for 5 hours at 37°C, the medium was changed and cells were incubated for an additional 60 hours. Cells were then washed with PBS and lysed. Aliquots of cell lysates were transferred to Optiplate-96 (PerkinElmer), followed by addition of luciferase substrate. Relative light units (RLUs) were measured using EnSpire plate reader (PerkinElmer). All the measurements were carried out in quadruplicates.
Recombinant IBV S1-CTD was assayed for its cell-binding capability using flow cytometry as previously described [13]. Briefly, HEK293T and DF-1 cells were incubated with recombinant IBV S1-CTD containing a C-terminal His6 tag (40 μg/ml) at room temperature for 30 minutes, followed by incubation with phycoerythrin (PE)-labeled anti-His6 antibody for 30 minutes. The cells were then analyzed for the binding of IBV S1-CTD using flow cytometry.
The total surface area and buried surface area of coronavirus S1-CTDs were calculated using the PISA server at the European Bioinformatics Institute (http://www.ebi.ac.uk/pdbe/prot_int/pistart.html) [47]. Specifically, for each trimeric spike protein, a PDB file containing all of the six S1 domains (including three copies of S1-CTDs and three copies of S1-NTDs) was submitted to the PISA server, and the total surface area and buried surface area for each S1-CTD were calculated. For the spike proteins used for the above analysis, all their S1-CTDs were in the “lysing down” state. The structures of MERS-CoV and HKU1 spikes were not included in the above analysis because the former contain at least one S1-CTD in the “standing up” state and the latter contains long stretches of missing residues in its S1 domains, both of which would interfere with the above analysis.
We constructed the IBV spike ectodomain (from IBV strain M41) in the pre-fusion state by replacing its transmembrane anchor and intracellular domain with a C-terminal GCN4 trimerization tag, followed by a His6 tag (Fig 1A). We expressed the protein in insect cells and purified the protein to homogeneity. We collected cryo-EM data on IBV spike ectodomain, calculated a density map at 3.93Å resolution (Fig 1B; S1 Fig), built an atomic model of the structure and refined it (Fig 1C and 1D). The final structural model contains all of the residues from 21 to 1022 (except residues 702–710) as well as glycans N-linked to 20 sites. Data collection and model statistics are shown in S1 Table.
The overall structure of IBV spike ectodomain resembles the pre-fusion structures of coronavirus spikes from the α-, β-, and δ-genera [12–14, 18–20]. It has a clove-like shape, with three S1 heads forming a crown-like structure and sitting on top of a trimeric S2 stalk. Each monomeric subunit of S1 contains two major domains, S1-NTD and S1-CTD, and two subdomains, SD1 and SD2 (Fig 2A and 2B). The S1-CTDs from three different subunits sit on the top and center of the spike trimer, whereas the three S1-NTDs are located on the lower and outer side to S1-CTDs (Fig 2C). SD1 and SD2 connect S1 to S2. The interface of trimeric S2 contains three central helices; each subunit of S2 contains one (Fig 2D and 2E). Each subunit of S2 also contains two heptad repeat regions, HR1 and HR2, and a fusion peptide (FP) (Fig 2D and 2E). In the post-fusion structure of trimeric S2, three copies of HR1 and three copies of HR2 would refold into a six-helix bundle structure, and FP would insert into the target membrane [29–35]. As in the structures of other coronavirus spikes, the HR2 region (residues 1022–1076) in the pre-fusion IBV spike is disordered (Figs 1A and 2D). The exact residue range of coronavirus FP remains unknown, although biochemical studies have identified a region in coronavirus S2 that associates with membranes and likely corresponds to FP (Fig 2D) [48, 49]. In the following sections of this paper, we will compare the structures and functions of IBV spike to those of the spikes from the other three genera, and discuss the evolution of coronavirus spikes.
IBV S1-NTD takes the same galectin fold as the S1-NTDs from the other three coronavirus genera, but it contains unique structural features (Fig 3A–3D). Its core structure is a twelve-stranded β-sandwich, which consists of two six-stranded antiparallel β-sheet layers stacked together through hydrophobic interactions (Figs 2B and 3C). The topology of the β-sandwich core is identical to that of human galectins (S2 Fig). Underneath the core structure is another β-sheet and an α-helix, which are also present in the S1-NTDs from the other three coronavirus genera. Above the core structure are some loops that form a partial ceiling-like structure. This structure is not present in human galectins or S1-NTDs from α- or δ-genus, but becomes a more extensive ceiling-like structure in β-coronavirus S1-NTDs (Fig 3A, 3B and 3D). Based on the structure and function of β-coronavirus S1-NTDs, we previously predicted that S1-NTDs from all of the genera have a galectin fold, and further correlated the galectin fold to their functions as viral lectins [15]. Recent structural studies, including the current one, have confirmed our previous structural predictions (S2 Fig). These studies also have unexpectedly revealed that the presence and extent of the ceiling-like structure on top of the core structure are unique structural features for S1-NTDs from different genera.
It has been known that IBV spike binds sugar [50]. A recent study further confirmed that the sugar-binding domain in IBV spike is its S1-NTD [51]. To date no structural information is available for the complexes of coronavirus S1-NTDs and their sugar ligand. Mutagenesis study showed that in the S1-NTD from β-genus bovine coronavirus (BCoV), the sugar-binding site is located in the pocket formed between the core structure and the ceiling [24]. In the structure of host galectins, despite no ceiling, the sugar-binding site is located in the same place [52]. Based on the structural similarity between the S1-NTDs from different coronavirus genera, the sugar-binding site in IBV S1-NTD might also be located in the pocket formed between the core structure and the partial ceiling (Figs 2B and 3C).
The structure determination of IBV S1-NTD provides insight into the structural and functional evolution of coronavirus S1-NTDs. We hypothesized that coronavirus S1-NTDs originated from host galectins based on the structural similarities between coronavirus S1-NTDs and host galectins [23, 24]. As host proteins, galectins are not recognized by the host immune system. In comparison, coronavirus S1-NTDs are under the host immune pressure to evolve. The gradual structural evolution of the ceiling on top of the core structure may have led to three functional outcomes. First, the ceiling could provide better protection to the sugar-binding site from host immune surveillance, which appears to be a common feature of viral lectins [53]. This hypothesis on protected sugar-binding sites in viral lectins is also consistent with the “canyon hypothesis” which states that receptor-binding sites on viral surfaces are hidden from the host immune surveillance [54]. Second, the ceiling is also involved in the quaternary packing of S1, which will be discussed later in this paper. Third, in the structure of S1-NTD from β-genus mouse hepatitis coronavirus (MHV), the outer surface of the ceiling has further evolved the capability to bind a protein receptor CEACAM1, while the presumed sugar-binding pocket has lost its capability to bind sugar [23]. Hence, the structural development of the ceiling is a possible indicator for the evolution of S1-NTDs in the direction of α- and δ-genera, then the γ-genus, and finally the β-genus. Furthermore, we performed quantitative structural comparisons of S1-NTDs from different genera by calculating the Z-score and r.m.s.d. between each pair of the proteins (Fig 3E). The result confirmed that S1-NTDs are relatively conserved among different genera, as reflected by the generally high Z-scores and low r.m.s.d. In terms of structural distances to α-coronavirus S1-NTDs, δ-coronavirus S1-NTDs are the closest, β-coronavirus S1-NTDs are the farthest, and γ-coronavirus S1-NTDs fall in the middle. Moreover, the structural similarity between α- and δ-coronavirus S1-NTDs is slightly higher than that between two β-coronavirus S1-NTDs, suggesting that S1-NTDs within β-genus have diverged slightly more than those between the α- and δ-genera. Taken together, S1-NTDs from the four genera form an evolutionary spectrum in the order of α-, δ-, γ-, and β-genus, with α-coronavirus S1-NTDs probably being the most ancestral (Fig 3A–3D).
The structure of IBV S1-CTD is significantly different from the structures of S1-CTDs from the other genera (Fig 4A–4D; S3 Fig). Its core structure is a β-sandwich containing two β-sheet layers: one is five-stranded and antiparallel, and the other is two-stranded and parallel (Figs 2B and 4C; S3 Fig). The interactions between the two β-sheet layers are present but minimal. In contrast, the core structures of α-coronavirus and δ-coronavirus S1-NTDs are both standard β-sandwich folds with extensive interactions between the two β-sheet layers: one is three-stranded and antiparallel, and the other is three-stranded and mixed (Fig 4A and 4B; S3 Fig). Even more drastically different are the β-coronavirus S1-CTDs, which contain only one five-stranded antiparallel β-sheet layer with the other layer turning into an α-helix and a coil (Fig 4D; S3 Fig). Despite these dramatic structural differences, the S1-CTDs from all genera share the same structural topology (i.e., connectivity of secondary structural elements) (S3 Fig). Moreover, the additional structural motifs on the edge of the core structure are also diverse among different genera (S3 Fig). In the IBV S1-CTD, two extended loops on the edge of the core structure function as putative receptor-binding motifs (RBMs) by potentially binding to an unknown receptor (see below) (Figs 2B and 4C). In both the α- and δ-coronavirus S1-CTDs, three short discontinuous loops are located in the same spatial region; they function as the RBMs in α-coronavirus S1-CTDs and putative RBMs in δ-coronavirus S1-CTDs (Fig 4A and 4B). In β-coronavirus S1-CTDs, a long continuous subdomain is located in this spatial region and functions as the lone RBM (Fig 4D). Structural variations of the RBMs in the S1-CTDs within each of the genera further lead to different receptor specificities [7]. In sum, IBV S1-CTD contains a weakened β-sandwich core structure and two extended RBM loops; the former structural feature falls between the β-sandwich cores of α- and δ-genera and the β-sheet core of β-genus, whereas the latter structural feature falls between the three short discontinuous RBM loops of α- and δ-genera and a single long continuous RBM subdomain of β-genus.
To investigate the function of IBV S1-CTD, we performed two assays to detect possible interactions between IBV S1-CTD and its potential receptor on the host cell surface. First, we carried out an IBV-spike-mediated pseudovirus entry assay in the presence of recombinant IBV S1–CTD (S4A Fig). To this end, retroviruses pseudotyped with IBV spike (i.e., IBV pseudoviruses) were used to enter host cells. In the absence of recombinant IBV S1-CTD, IBV pseudoviruses entered DF-1 cells (chicken fibroblast) efficiently, which was consistent with a previous report showing that DF-1 cells are permissive to live IBV (strain M41) infections [55]. As a negative control, their entry into HEK293T cells (human kidney) was inefficient. Recombinant IBV S1-CTD reduced the efficiency of IBV pseudovirus entry into DF-1 cells in a dose-dependent manner, likely because it competed with IBV pseudoviruses for an unknown receptor on the host cell surface. Second, we examined the binding of recombinant IBV S1-CTD to the host cell surface using a flow cytometry assay (S4B Fig). To this end, recombinant IBV S1-CTD was incubated with DF-1 cells, and subsequently cell-bound S1-CTD was detected using flow cytometry. Recombinant IBV S1-CTD bound to the surface of DF-1 cells efficiently, but not the surface of HEK293T cells. Taken together, IBV S1-CTD binds to a yet-to-be-identified receptor on the surface of chicken cells and hence functions as a receptor-binding domain (RBD). Thus, the S1-NTD and S1-CTD of IBV spike may both function as RBDs. Because coronavirus S1-CTDs from the α- and β-genera all use the additional structural features on the edge of their core structure as their RBMs, it is likely that the two extended loops in the same spatial region in IBV S1-CTD function as the RBMs.
Coronavirus S1-CTDs represent remarkable examples of divergent evolution of viral proteins. The core structures and the RBM regions of S1-CTDs are both divergent among different genera (Fig 4A–4D; S3 Fig). The core structures are β-sandwiches for α- and δ-coronavirus S1-CTDs, weakened β-sandwiches for γ-coronavirus S1-CTDs, and single β-sheet layer for β-coronavirus S1-CTDs. The RBMs are three short discontinuous loops for α- and δ-coronavirus S1-CTDs, two reinforced loops for γ-coronavirus S1-CTDs, and a single continuous subdomain for β-coronavirus S1-CTDs. Hence the S1-CTDs form an evolutionary spectrum, with α- and δ-coronavirus S1-CTDs on one end, β-coronavirus S1-CTDs on the other end, and γ-coronavirus S1-CTDs in between. We performed quantitative structural comparisons of S1-CTDs from all four genera (Fig 4E). The result confirmed that S1-CTDs are relatively poorly conserved among different genera, as reflected by the generally low Z-scores and high r.m.s.d. In terms of structural distances to α-coronavirus S1-CTDs, δ-coronavirus S1-CTDs are the closest, β-coronavirus S1-CTDs are the farthest, and γ-coronavirus S1-CTDs fall in the middle. The functional outcomes of the core structure evolution are not clear, but the evolution of the RBMs may have led to the following two functional outcomes. First, the diversity of the RBMs from three short loops to two extended loops and then to a long subdomain may allow coronaviruses to explore a wider variety of receptors. Second, the reinforced RBM regions in both β- and γ-coronavirus S1-NTDs facilitate quaternary packing of S1, which will be discussed later in this paper. Taken together, the S1-CTDs from different genera form an evolutionary spectrum in the order of α-, δ-, γ-, and β-genus, although the evolutionary direction could go either way (Fig 4A–4D).
Curiously, coronavirus S1 from different genera take two types of quaternary packing modes (Fig 5A–5D) [12–14, 18–20]. IBV S1 takes a cross-subunit quaternary packing mode where the S1-NTD and S1-CTD from different subunits pack together (Fig 5C). Specifically, in the trimeric IBV spike, one S1-CTD packs against two S1-CTDs from the other subunits as well as one S1-NTD from another subunit. The putative RBMs of IBV S1-CTD and the partial ceiling of IBV S1-NTD are both involved in the cross-subunit packing. As a result, the putative RBMs of IBV S1-CTD are partially concealed, disallowing their full access to the host receptor. Hence IBV S1-CTD in the current structure was captured in a “lying down” state, and would need to “stand up” on the spike trimer for efficient receptor binding. This potential conformational change of IBV S1 can minimize the exposure of the putative RBMs in its S1-CTD to the immune system, thereby functioning as a possible strategy for viral immune evasion. β-coronavirus S1 also takes the cross-subunit packing mode, with the RBM of its S1-CTD and the ceiling of its S1-NTD both involved in the cross-subunit packing (Fig 5D) [18–20]. In contrast, α- and δ-coronavirus S1 both take an intra-subunit packing mode where the S1-NTD and S1-CTD from the same subunit pack together (Fig 5A and 5B) [12–14]. The RBMs of α- and δ-coronavirus S1-CTDs are involved in the intra-subunit packing. Whether S1 packs in the intra-subunit or cross-subunit mode, the RBMs of S1-CTDs are concealed or partially concealed in their “lying down” state, and would need to switch to the “standing up” state for receptor binding. Overall, β- and γ-coronavirus S1 both take the cross-subunit quaternary packing mode, whereas α- and δ-coronavirus S1 both take the intra-subunit quaternary packing mode.
We examined whether the quaternary structures of coronavirus S1 can lead to functional differences of coronavirus spikes. First, in both β- and γ-coronavirus spikes, the RBMs of their S1-CTDs and the ceiling of their S1-NTDs have evolved to facilitate the cross-subunit packing. These additional structural features further evolved to gain other functions: the RBMs of S1-CTDs recognize diverse protein receptors, whereas the ceiling of the S1-NTDs either protects the sugar-binding site or recognizes a new protein receptor [7]. Second, to investigate the structural restrain on S1-CTDs that may hinder their potential conformational change, we calculated the total and buried surface areas of the S1-CTD on the spikes from different genera. The result did not reveal systematic difference between intra-subunit S1 packing and cross-subunit S1 packing in the buried surface area of S1-CTDs. However, it is worth noting that the S1-CTD from β-genus SARS-CoV has the smallest buried surface area (in both absolute value and percentage) (S2 Table). The relative small buried surface area of SARS-CoV S1-CTD indicates less structural restraint on the S1-CTD from other parts of the spike S1, possibly allowing the S1-CTD to switch to the “standing up” and receptor-accessible conformation more easily. The “standing up” conformation of SARS-CoV S1-CTD may also weaken the structural restraint of S1 on S2 (discussed in more detail later), potentially allowing membrane fusion to proceed more easily [56]. Indeed, frequent “standing up” of SARS-CoV S1-CTD has been observed [19]. Overall, compared to the intra-subunit quaternary packing of α- and δ-coronavirus S1, the cross-subunit quaternary packing of β- and γ-coronavirus S1 may have allowed their S1 to evolve additional functions in receptor recognition; moreover, the S1-CTD from β-genus SARS-CoV spike has a relatively small buried surface area, which may be responsible for its dynamic receptor-binding conformation.
The structure and function of IBV S2 are highly similar to those of S2 from the other coronavirus genera. In the pre-fusion structure of IBV S2, HR2 is disordered, whereas HR1 and FP each consist of several α-helices and connecting loops (the exact residue range of FP is not clear) (Fig 2D). In the post-fusion structure, HR1 would refold into a long α-helix, HR2 would refold into a mixture of α-helices and coils, three copies of HR1 and HR2 would pack into a six-helix bundle structure, and FP would also refold and insert into the target membrane (S5A Fig) [6, 29]. IBV S2 is locked in the pre-fusion state because of the structural restraint from S1. Specifically, because of the cross-subunit quaternary packing of trimeric IBV S1, HR1 and FP of IBV S2 are structurally restrained by two S1-CTDs from the other subunits and SD1 from another subunit, respectively (S5B Fig). The structural restraints from S1 on S2 can be weakened by the standing up of S1-CTDs (which allows receptor binding) and can be lifted completely upon proteolysis removal of S1. The packing between S1 and S2 in IBV spike is the same as those in β-coronavirus spikes [18–20]. However, in α- and δ-coronavirus spikes, the packing between S1 and S2 becomes different due to the intra-subunit quaternary packing of their trimeric S1: HR1 and FP are restrained by one S1-CTD and one SD1 from another subunit, respectively (S5C Fig) [12–14]. Other than the differences in S1/S2 packing, the structural and functional similarities of coronavirus S2 from different genera suggest evolutionary conservation of coronavirus S2.
The fast evolutionary rates of viruses, particularly RNA viruses, make it difficult to trace their evolutionary history [1–3]. Envelope-anchored coronavirus spike proteins guide viral entry into cells; they are the fastest evolving coronavirus proteins due to viral needs to engage diverse host receptors, maximize membrane-fusion efficiency, and evade host immune surveillance [7–14]. Coronavirus spikes from four different genera are divergent, and their evolutionary relationships pose a major puzzle in the virology field [6]. Because viral proteins need to function under certain structural and functional constraints, evolutionary information of viral proteins can be more reliably found in their tertiary structures and related functions than in their primary structures [16, 17]. Although extensive structural studies including both X-ray crystallography and cryo-EM have been done on coronavirus spikes, a critical piece that was still missing is the structure of γ-coronavirus spikes, preventing a clear understanding of the evolutionary relationships among coronavirus spikes [12–14, 18–28]. In this study, we determined the cryo-EM structure of IBV spike ectodomain, the first such structure from the γ-genus, which bridges the divergent structures of coronavirus spikes into an evolutionary spectrum and provides insight into the evolutionary relationships among coronavirus spikes.
Our study compares the structures and functions of coronavirus spikes from the four genera, and illustrates the structural and functional evolution of these proteins. First, coronavirus S1-NTDs from all genera share the same structural fold and possibly evolutionary origins with host galectins. From α- and δ- genera to γ genus and then to β genus, the S1-NTDs have evolved from simple galectin-fold core structure with an exposed sugar-binding site, to having a partial ceiling on top of the core structure, and to having an extensive ceiling to protect the sugar-binding site from host immune surveillance (the outer surface of the ceiling in one β-coronavirus can even bind to a novel protein receptor). The partial ceiling in γ-coronavirus S1-NTDs and the extensive ceiling in β-coronavirus S1-NTDs are also involved in the quaternary packing of S1. Second, coronavirus S1-CTDs from different genera are very diverse, but still form an evolutionary spectrum with α- and β-coronavirus S1-CTDs at two ends and δ- and γ-coronavirus S1-CTDs in the middle. The core structures of S1-CTDs have diverged from β-sandwich to weakened β-sandwich and then to β-sheet, whereas the RBMs have diverged from short loops to extended loops and then to a long subdomain. The functional significance of the core structure evolution is not clear, but the RBM evolution may allow the viruses to expand receptor recognition and also contributes to the quaternary packing of S1. Third, from α- and δ- genera to β- and γ-genera, the quaternary packing of S1 has diverged from simple intra-subunit packing to more complex cross-subunit packing. The cross-subunit quaternary packing of β- and γ-coronavirus S1 may have allowed their S1 to evolve additional functions in receptor recognition. Moreover, the relatively small buried surface area of the S1-CTD from β-genus SARS-CoV may allow the S1-CTD to be more dynamic for receptor binding. Finally, the S2 from all four genera are structurally and functionally conserved, although there are some differences in their S1/S2 packing. Quantitative structural comparisons also demonstrate that in terms of structural distances to α-coronavirus S1, δ-coronavirus S1 is the closest, β-coronavirus S1 is the farthest, and γ-coronavirus S1 is the intermediate. We also calculated the phylogenetic tree using the amino acid sequences of 29 coronavirus spikes from different genera, and the result showed that in terms of amino acid sequence distances to α-coronavirus spikes, δ-coronavirus spike is the closest, β-coronavirus spike is the farthest, and γ-coronavirus spike is the intermediate (S6 Fig). Taken together, coronavirus spikes from different genera form an evolutionary spectrum, with α-coronavirus spikes on one end, followed by δ-coronavirus spikes and γ-coronavirus spikes, and β-coronavirus spikes on the other end.
Because of their fast evolutionary rates, viruses are perfect model systems for studying evolution. Our study has demonstrated that despite structural divergence among coronavirus spikes, particularly in their S1 where low or little structural similarities can be detected, we can still trace the evolutionary relationships among these viral proteins through detailed analyses of their structures and functions. Our study also reveals that coronavirus spikes have evolved to remarkable diversity to expand their receptor recognition, facilitate membrane fusion, and evade host immune surveillance, while conserving basic membrane-fusion mechanisms. The evolutionary approaches that coronaviruses take and the evolutionary edges that they gain are good examples of viral evolution and deepen our understanding of evolution in general.
|
10.1371/journal.pcbi.1006126 | Mechanical unfolding reveals stable 3-helix intermediates in talin and α-catenin | Mechanical stability is a key feature in the regulation of structural scaffolding proteins and their functions. Despite the abundance of α-helical structures among the human proteome and their undisputed importance in health and disease, the fundamental principles of their behavior under mechanical load are poorly understood. Talin and α-catenin are two key molecules in focal adhesions and adherens junctions, respectively. In this study, we used a combination of atomistic steered molecular dynamics (SMD) simulations, polyprotein engineering, and single-molecule atomic force microscopy (smAFM) to investigate unfolding of these proteins. SMD simulations revealed that talin rod α-helix bundles as well as α-catenin α-helix domains unfold through stable 3-helix intermediates. While the 5-helix bundles were found to be mechanically stable, a second stable conformation corresponding to the 3-helix state was revealed. Mechanically weaker 4-helix bundles easily unfolded into a stable 3-helix conformation. The results of smAFM experiments were in agreement with the findings of the computational simulations. The disulfide clamp mutants, designed to protect the stable state, support the 3-helix intermediate model in both experimental and computational setups. As a result, multiple discrete unfolding intermediate states in the talin and α-catenin unfolding pathway were discovered. Better understanding of the mechanical unfolding mechanism of α-helix proteins is a key step towards comprehensive models describing the mechanoregulation of proteins.
| In order to migrate and survive, most cells need to be attached to their environment. Cells anchor to the extracellular matrix via transmembrane integrin, connecting it to contractile the cytoskeleton. Similarly, cell-cell contacts are formed via transmembrane cadherin, which also connects to the contractile cytoskeleton through scaffolding proteins. Examples of such proteins include talin and α-catenin, which connect integrin and cadherin respectively, to actin filaments of the cytoskeleton. Mechanical forces that are transmitted between the cell and its environment activate binding and regulate the functions of these scaffolding proteins at cell-extracellular matrix and cell-cell contacts. Functions of talin and α-catenin are tightly modulated by mechanical forces. The stretching of these proteins under mechanical load exposes buried binding sites for other partners, such as vinculin. We used steered molecular dynamics simulations and single-molecule atomic force microscopy to study how these proteins unfold under load. Our results suggest that α-helical talin and α-catenin unfold through stable 3-helix intermediates. These intermediates represent biologically active states, which may allow recruitment of other binding partners.
| Protein activity can be modulated by mechanical cues in addition to chemical stimuli and ligand binding. Mechanical forces can induce conformational changes in the protein structure, that leads to either a switch between the functional states of the protein, or allows for multiple functions[1,2].
α-helix folds are highly abundant among the structural proteins located at the cell-cell and cell-ECM contacts [3–5], among the essential muscle costamere complexes[6], as well as among the structures interconnecting the cytoskeleton[7,8] and cellular organelles[9]. Despite the numerous studies on mechanotransduction and the mechanosensitivity of proteins, the mechanisms associated with forced protein unfolding and mechanosignaling are not well understood especially where α-helical proteins are concerned. One reason for such a lack of understanding might be that α-helices are, in general, mechanically weaker compared to β-strand folds[10] and are therefore challenging to study experimentally. The low mechanical stability of α-helical proteins requires sophisticated experimental methods capable of force measurement in the range of piconewtons[11].
Among the α-helical structural proteins, several distinct folds have been identified. Talin-like proteins contain 5- and 4-helix bundles[12]. Proteins of the catenin family contain a 4-helix conformation with long α-helices interconnecting two bundles[13]. Both of these multi-helical protein folds respond to mechanical load by a dissociation of the bundles leading to mechanoregulatory function. The spectrin fold is formed of a 3-helix conformation with a long α-helix connecting the neighboring domains, forming a rigid rod of interacting 3-helix bundles. The spectrin-like conformation accounts for structural reinforcement in the cellular scaffolds[3,4,7]. Finally, single long α-helices and coils can also be found among structural α-helical proteins (PDB id 5KHT).
Focal adhesions and adherens junctions are fundamental mechanosensitive structures through which cells communicate with the extracellular matrix and with their adjacent cells, respectively. The processes of focal adhesion formation and maturation are regulated by mechanical signals[14,15]. Similar to the role of focal adhesions in the cell-ECM connection, adherens junctions are essential for cell-cell contacts. Adherens junctions are associated with the cadherin super-family of transmembrane proteins, which are connected through catenin-rich protein complexes to the actin cytoskeleton. The cadherins bind to the cytoplasmic protein β-catenin, which in turn binds to the filamentous F-actin binding adaptor protein, α-catenin[16,17].
Talin is a large focal adhesion protein that contains an N-terminal head domain, which is responsible for integrin binding. The larger talin rod domain consists of amphipathic α-helices arranged into 13 four or five-helix bundles (R1–R13) and a single helix dimerization domain (DD) at the C-terminal end (Fig 1A). Talin provides a link between the ECM, via the talin head-integrin interaction, and the cytoskeleton through the binding of actin filaments at actin binding sites located in the rod domain. Furthermore, the talin rod comprises up to 11 buried vinculin binding sites (VBSs), distributed along its structure. These binding sites are exposed by partial or complete unfolding of different bundles[18,19] as a result of mechanical stretching. In this way mechanical load regulates talin function by exposing the buried binding sites for certain binding partners such as vinculin[20], while simultaneously, epitopes for other binding partners become inactivated. An example of such epitope inactivation would be the talin binding partner RIAM which binds only to folded talin domains[21]. Thus, conformational changes of talin under mechanical load regulate the recruitment and activation of talin-interacting proteins. Interestingly, talin dimerization is also regulated by mechanical force [22].This property highlights talin as a key player in the transmission and sensing of mechanical signals between the extracellular matrix and cell cytoskeleton. These mechanical signals are central for a variety of cellular functions including spreading, migration, invasion and substrate sensing[23–25].
Similar to talin, α-catenin recruits vinculin, providing the mechanical connection between cell-cell adhesions and the cytoskeleton. α-catenin contains 5 α-helix domains: the dimerization domain (DD) functioning as β-catenin binding domain, three modulation domains I, II and III (MI, MII and MIII), and an F-actin binding domain (FABD) (Fig 1B). Vinculin binds the MI domain of α-catenin while the two adjacent domains (MII and MIII) inhibit vinculin binding to MI. It has been demonstrated that vinculin is recruited in a force-dependent manner to the cadherin/catenin complex upon force-dependent unfolding of α-catenin. This process resembles the binding of vinculin to the talin rod domain upon tension-dependent unfolding. This tension-dependent unfolding is thought to be central in cell-cell adhesion mechanosensing[9,26,27].
It has been previously hypothesized that the force-dependent unfolding of the vinculin binding domains of the talin rod and the α-catenin domains includes stable intermediate conformations[28–30]. However, this has never been previously studied in detail in computational simulation or in experimental setup suitable for analysis of single-molecule unfolding events. α-helical proteins are associated with key physiological and pathological processes in mechanobiology[31–33]. A number of diverse diseases from heart[34] and muscle[35], diseases of bone, vascular and nervous system, or skin[36] have been associated with mechanobiology. Hence, better understanding of their unfolding characteristics will open the possibility of answering a plethora of questions in biology and medicine. In this study, we investigate the molecular mechanisms of the unfolding of two α-helical proteins, talin and α-catenin, on an atomistic level by a combination of steered molecular dynamics (SMD) simulations and single-molecule atomic force microscopy (smAFM).
This study elaborates in closer detail on observations made during our previous work described in Haining et al. [28]. Previously, we have concentrated on the mechanical sensitivity of the talin rod subdomains during the initial domain breaking. We have reported that the talin subdomains are similar, however not identical in their ability to withstand mechanical load. All the tested talin subdomains unfolded in AFM setup over a range of mechanical forces between 10 and 40 pN. During the SMD simulations we repeatedly observed a possible unfolding intermediate in the unfolding trajectories which we intended to investigate further. Therefore, in the light of our previous findings, we have now selected two talin subdomains on either end of the mechanical sensitivity scale for the current study; weak R3 and strong R9. Talin R9 is an exceptional subdomain responsible for talin autoinhibition while it does not contain VBS. For that reason, we have also included subdomain R11 to investigate the unfolding intermediate in the terms of VBS activation.
In order to probe the existence of stable intermediates in the unfolding trajectory of α-helical mechanosensitive domains, we have selected two mechanoregulatory proteins, talin and α-catenin (Fig 1). We have selected the mechanically stronger talin 5-helix bundles R9 and R11, and the mechanically weaker 4-helix bundles of talin R3 and α-catenin modulation domains I to II (MI-MII).
We subjected two talin rod 5-helix bundles, R9 and R11, to end-to-end SMD stretching with constant velocity pulling simulation at 2 nm/ns. Previously, talin 5-helix bundles were more mechanically stable in our SMD simulations compared to 4-helix bundles. According to our previous study[28], R9 was one of the strongest rod bundles. Overall, unfolding of R9 and R11 showed two force peaks (Fig 2A), which correspond to breaking of the 5-helix bundle and 3-helix intermediate. This unfolding intermediate consists of core helices (H2–H4) after the dissociation of H1 and H5. For R9 the maximum unfolding forces detected in constant velocity pulling simulation at 2 nm/ns were 348 ± 24 pN and 349 ±39 pN (average force ± standard deviation). Unfolding peak forces for R11 were very similar to those for R9 (Fig 2A). Representative snapshots of the unfolding trajectories are shown in Fig 3, indicating changing of the bundle conformation during unfolding.
To confirm the unfolding intermediate, we designed point mutations in the R9 bundle forming disulfide bonds (clamps) in order to block the unfolding of the 3-helix core. These clamped R9 mutants were subjected to end-to-end pulling in simulation and experimental setup. We prepared three constructs including L1698C and A1748C cysteine mutations that protect 3-helix core from N-terminus (N clamp), A1720C and A1779C mutations that prevent unfolding of the 3-helix core from C-terminus (C clamp), and R9 construct with both N- and C-terminal disulfide clamps (N/C clamps) (S1 Fig). All R9 clamp mutants showed very similar unfolding of 5-helix state compared to the wild-type R9 (Fig 2B). The unfolding of the 3-helix state was effectively blocked in the R9 equipped with N/C clamps, while the constructs with only one disulfide clamp allowed unfolding of either H2 (R9 with C clamp) or H4 (R9 with N clamp) of the 3-helix core, but did not compromise the stability of the 3-helix intermediate (Fig 2B).
In additional experiments, we investigated talin R3 bundle and α-catenin modulation domains I to II (MI-MII), which are all 4-helix bundles. Talin R3 was easily unfolded to the 3-helix state by the separation of H4. The 3-helix intermediate (H1–H3) was more stable compared to the 4-helix bundle, the unfolding peak force for breaking the R3 3-helix state was 276 ± 20 pN (Fig 2E). Similarly, α-catenin was unfolded to 3-helix conformation by dissociation of H4 (four out of five simulations) in MII and H1 in MI (all five simulations). Further unfolding showed two force peaks for breaking the 3-helix state in MII (at 349 ± 33 pN) and MI (at 461 ± 68 pN) (Fig 2F). Collectively these data suggest that both 4- and 5-helix bundles unfold through stable 3-helix intermediate state. Furthermore, 5-helix bundles withstand mechanical load better than 4-helix domains, which are easily unfolded to the 3-helix state.
Described constant velocity SMD simulations of individual talin rod bundles and α-catenin were run five times each. Unfolding force profiles showed that our results are well reproducible (S2 Fig) and allowed us to calculate average peak force and a standard deviation (S1 Table).
To assess the effects of force penetration on the unfolding mechanisms in SMD simulations, we studied the unfolding mechanisms and the existence of the stable intermediates in linear protein chain consisting of two talin R9 monomers resembling the natural biological assembly of talin. We designed tandem construct possessing exactly the same mechanical stability, i.e. two talin R9 domains (R9WT−R9WT tandem). Furthermore, we analyzed the R9 tandems with disulfide clamps, protecting the 3-helix core from unfolding, in either first or second monomer, with respect to the fixed N-terminal and pulled C-terminal end. Thus, we prepared two tandem constructs with clamps, R9N/C clamps−R9WT and R9WT−R9N/C clamps, respectively. For R9WT−R9WT tandem, unfolding force showed four peaks, corresponding to the breaking of 5-helix states first, followed by dissociation of the 3-helix states in both α-helix bundles (Fig 2C). Because the pulling was applied to Cα atom of C-terminal residue, the second R9 in the tandem was closest to the point of pulling and unfolded to 3-helix state first. Both monomers of the R9WT−R9WT tandem unfolded to the 3-helix intermediate within approx. 30 ns of the SMD simulation with 2 nm/ns pulling velocity, and unfolded at approx. 50 ns. Both tandems containing one clamped monomer showed three unfolding force peaks (Fig 2D) lacking the peak corresponding to the breaking of the disulfide-clamped 3-helix structure. Indeed, the force penetration did affect the two different tandems with clamps resulting in different unfolding trajectories. For R9WT−R9N/C clamps tandem, both monomers had 3-helix conformations at ~ 30 ns, while for R9N/C clamps−R9WT, closest to the point of pulling monomer (R9WT) unfolded completely before R9N/C clamps molecule unfolded to 3-helix state (Fig 2D, Fig 3).
The investigation of the unfolding of a tandem provided us with a tool of studying the unfolding principles of multiple domains. Furthermore, the use of disulfide clamps in the tandem construct protecting the stable state provided us with a comparison force trace and additional proof of an intermediate unfolding conformation in both mechanosensitive proteins. Altogether, these findings indicate that the unfolding force required for the unfolding of the 3-helix intermediate state is similar to that needed for the unfolding of 5-helix state.
We utilized smAFM to characterize the unfolding patterns of R9 and R11 constructs and captured the 3-helix intermediate (Fig 4). Similarly to the SMD, we detected two unfolding events for each of the bundles, implying that unfolding occurs through a mechanically stable intermediate. Overall the bundle stability was higher in the case of R9 than R11, consistent with our previous results[28]. The distance between the two unfolding events is 20–25 nm, which is consistent with the collapse from 5 helices to 3 helices. Likewise, the distance from the second unfolding event to the HaloTag ruler of 25–35 nm is consistent with the subsequent collapse of the 3-helix intermediate. As such the smAFM data supports the picture derived from the SMD analysis: 5-helix bundles collapse via a stable 3-helix intermediate.
We also tested the R9 tandem construct to examine if the 3-helix intermediate could be detectable within a model of a polyprotein (Fig 5A–5D). We detected 4 unfolding events, consistent with a pattern of two bundles unfolding via a stable intermediate. The distances between the first and second event (~20 nm), second and third (~30 nm), third and fourth (~20 nm) and fourth and HaloTag (~30 nm) imply that, contrary to the SMD results, one α-helix bundle collapsed completely through a 3-helix intermediate before the second α-helix domain started unfolding. However, it is difficult to be certain, given the error margin of the peak locations and the likely stochastic nature of the unfolding process. When the 3-helix disulfide clamp was introduced into one of the tandem R9 domains, we saw a reduction in the number of unfolding events from 4 to 3 (Fig 5E–5H). This, along with the reduction of the overall unfolding length from 105 nm to 85 nm, demonstrates that the disulfide clamping was able to protect the 3-helix intermediate of the R9 from mechanical unfolding.
The forced unfolding of α-catenin modulation domains I to II (MI-MII) by AFM produced a pattern consistent with the SMD simulation (Fig 5I–5L). We detected two unfolding events with a maximum unfolding length of ~50 nm which implies that there is no unfolding event with respect to the reduction of both 4-helix bundles into 3-helix states. This is in agreement with herein presented SMD data and with our previous work reporting on the lower mechanical stability of the 4-helix bundles[28]. The two events observed with unfolding lengths of ~25 nm correspond to the mechanical unfolding of 3-helix intermediate states.
For the comparison of the 3-helix state mechanical stability we selected two mechanically diverse talin rod bundles, i.e. the mechanically weaker R3 and the mechanically stronger R9 for more detailed analysis. Although constant velocity SMD simulations are an excellent tool comparing the results with AFM analysis, they are less sensitive for the assessment of intermediate states as compared to constant force simulations. Therefore, we subjected the R3 and R9 bundles to constant force pulling simulations where, after screening of suitable force regime, constant force ranging from 160 pN to 200 pN for R3, and from 200 pN to 300 pN for R9 (Fig 6) was used. In constant force SMD, 4-helix R3 was weak even at 160 pN and rapidly unfolded to 3-helix state (within ~ 3 ns). After the separation of H4, H3 was slowly sliding relative to H1 and H2 (from ~ 3 ns to ~ 23 ns). The 3-helix intermediate did not unfold at 160 pN in 40 ns time window (Fig 6A), however, it unfolded completely (at ~ 72 ns) in extended 160 pN simulation (S3A Fig). R3 was extended also with constant force at 170 pN, 180 pN and 200 pN. Although it was completely unfolded after ~ 18 ns, ~ 13 ns and ~ 11 ns respectively, the stable 3-helix intermediate was observed in all trajectories (Fig 6). For strong R9 bundle, we first applied constant force of 200 pN and observed only partial uncoiling of terminal helices (H1 and H5) within 40 ns time window (Fig 6B), yet the 5-helix state remained intact. However, R9 unfolded to the 3-helix state (at ~ 86 ns) in extended 200 pN simulation (S3B Fig). The application of constant force of 220 pN or higher resulted in gradual unfolding of the bundle. During the unfolding, we recognized two stable states (5-helix and 3-helix states). In order to compare mechanical stability of 5- and 3-helix states, we used the disrupted 5-helix state of R9 and subjected it to stretching with constant force of 200 pN. Although 3-helix intermediate was relatively stable, it slowly unfolded over time under 200 pN while 5-helix bundle resisted the unfolding under the same force magnitude.
These results suggest that 3-helix state in R9 is a stable conformation. However, it is weaker as compared to the 5-helix state of R9. In more detail, the 3-helix state can be unfolded under lower force load once the 5-helix state of the R9 bundle is broken. On the other hand, the 3-helix intermediate state in R3 is the most stable conformation of the R3 bundle.
Numerous studies concerning α-helical mechanosensitive proteins have provided information on mechanically regulated switches between diverse binding partners and their associated functions. Perhaps one of the best known examples of this mechanoregulated protein-protein interaction is the tandem talin-vinculin, where mechanical stress applied to talin rod exposes binding sites for vinculin[37,38].
Our observation of SMD trajectories for talin multidomain constructs, described in our previous work[28], revealed a possible stable 3-helix intermediate during the forced domain unfolding. Because of the complexity of the SMD and smAFM data we obtained during the multidomain construct unfolding, we were not able to identify previously the 3-helix intermediate among the force traces directly. However, simultaneous domain unfolding was recognized for bundles of similar mechanostability in smAFM[28].
In this study, we investigated the detailed unfolding mechanisms of α-helical talin rod bundles and α-catenin MI-MII domains to probe the presence of any intermediate or partial unfolded states. Our results show that the two studied proteins unfold through a stable 3-helix intermediate. Constant velocity pulling used in SMD and smAFM revealed, that the unfolding force profiles for the 5-helix rod bundles R9 and R11 have two peaks, which correspond to the breaking of the 5- and 3-helix states (Fig 2A). The talin 4-helix bundle R3 and α-catenin bundles MI-MII also unfolded through stable 3-helix intermediates (Fig 2E and 2F). In addition, this 3-helix state was recognized as the most mechanically stable conformation for the 4-helix domains (Fig 6A).
Other studies have also provided indirect evidence of unfolding intermediates in alpha helical proteins. Investigations of the talin R3 subdomain have revealed a possible 3-helix intermediate capable of rapid or instant activation for vinculin binding. Specifically, the deletion of helix 4 of rod subdomain R3 (ΔR3H4) leads to super-active R3 localizing efficiently in cell-ECM contacts (S6 Fig). Rahikainen et al., 2017[23] showed that one or two destabilizing mutations in R3 H1 were sufficient to facilitate bundle unfolding, increasing the activation of vinculin binding and resulting in a strong cellular phenotype. The phenotype of the further destabilized state modified with four mutations was comparable to the super-active R3 potentially indicating a 3-helix intermediate. Further evidence of a 3-helix state in R3 domain is found in a recent study by Baxter et al. [39]. A 3-helix open state has been recognized after the dissociation of H1 from the R3 bundle under high pressure conditions. Similar effects were observed even for the talin R1 bundle, where the deletion of H5 resulted in the exposure of the VBS located in H4 and an active conformation of R1[40]. The authors also suggest that the deletion of H5, resulting in a 4-helix partial bundle, causes a destabilization of the R1 domain leading to partial unfolding. This observation is in line with our results; we showed that the 4-helix fold is a fragile conformation which does not require excessive mechanical force to unfold to a stable 3-helix state (Fig 2E). Finally, even previous computational studies suggested that only partial unfolding of talin subdomains described by minimal protein extension is sufficient for VBS activation. In more detail, R1 VBS was activated through torsional conformational change of the hydrophobic core orientation within R1 subdomain during an extension of less than 2 Å [41].
Inspection of the molecular characteristics of 5- (4- in R3) and 3-helix states did not reveal any dominant differences between these assemblies in terms of interactions or packing. The hydrophobic interactions appear to be the main factor in maintaining both these states, as shown in S4 Fig. In the previous study[28], we proposed two conserved residues that are important for maintaining R9 5-helix bundle stability, namely Leu1668 in H1 and Met1803 in H5. Further studies including experimental investigation of subdomains carrying mutations targeting the 3-helix core fragment would be needed to evaluate the contributions of individual residues for the mechanical stability of the 3-helix intermediate state.
Further unfolding of the 3-helix intermediate was observed in our experimental and simulation setup. Whether the complete unfolding of an α-helical domain takes place in vivo, or whether the 3-helix state is the final unfolding conformation remains unclear. Hints of both of these options can be found in the literature. As discussed earlier[40], S6 Fig, the deletion of terminal helix in R1 and R3 is sufficient for VBS activation and vinculin recruitment. Thus, we speculate that the 3-helix state is capable of vinculin binding. Vinculin binding to unfolded talin or α-catenin domains inhibits domain refolding under low mechanical load[20,27,42]. Simultaneously, vinculin binding to the 3-helix state bundle may protect it from complete unfolding[43]. Studies by Margadant et al.[43] show that the maximal length of talin is approx. 400 nm in living cells. This also supports the notion of partial unfolding even in the absence of vinculin. Interestingly, recently published work by Ringer et al.[11] revealed a force gradient across the talin rod domain. In the presence of vinculin, greater force was measured at the N-terminal end than at the C terminal end resulting in the bundle unfolding and activation for vinculin binding. As vinculin binds to activated talin and to actin, the force acting on the talin rod is divided and reduced towards the C-terminal end. We speculate that the reduced force might be insufficient to unfold the stable 5-helix subdomains located at the C-terminal end of the talin rod. However, here we showed that the intermediate resulting from 4-helix unfolding was mechanically weaker compared to the intermediate of a 5-helix bundle. Thus, the complete unfolding of 4-helix bundles at the N-terminal end of the talin rod may be possible. Based on the work by Yao et al., we may assume that complete yet reversible unfolding of R3 domain takes place under low force load. It was shown than under 4.8 pN of constant force load exerted on the full length talin, R3 occupies two distinct conformations with elongation of approx. 19 nm. However, which states these in fact are remains unclear [44]. We may hypothesize that only the end-to-end attachment to the pulling device and initial elongation under low force load is sufficient to collapse R3 into an activated 3-helix state. Such immediate conformational change would not necessarily result into an observable difference in the total end-to-end elongation compared to the R3 conformation in solution (S5 Fig) [41]. Similar elongation of approx. 19 nm between the 3-helix intermediate and the completely unfolded state was shown in our SMD results (Fig 6A). Here we also see that the 3-helix intermediate remained stable even though the end-to-end distance increased by 5 nm. This distance increase was caused by the uncoiling of H4.
Based on our observations of the talin unfolding mechanisms, we propose a model of multidomain α-helical protein unfolding under mechanical load, shown in Fig 7. In the absence of mechanical force, α-helical bundles remain in a folded conformation, capable of binding their ligands such as RIAM (R2, R3, R8 and R11) and DLC1 (R8) in the case of talin[45]. At low mechanical load, soft α-helix bundles, namely 4-helix domains, unfold to stable 3-helix intermediates. Since the activation of vinculin binding sites (VBSs) requires the unfolding of talin bundles[14], the formation of 3-helix states suggests that VBSs located at the terminal helices become available for vinculin. At the same time, the partial unfolding of the bundles leads to a disturbance of the binding sites for other binding partners located on the bundle surface, which abrogates their interaction[21]. With increasing force, additional bundles collapse to a 3-helix conformation switching from the mechanoregulatory role to a structural reinforcement role, similar to that of spectrin. In other words, it is possible that talin reacts to a range of small mechanical forces by the dissociation of certain bundles leading to a change in binding to other proteins. Such mechanoregulation would take place until the bundles reach a stable 3-helix spectrin-like conformation. At this point, the talin protein would assume a structural reinforcement role. Finally, at a high force, completely unfolded bundles would lose the ability to support ligand binding as well as structural function[20]. Such a model enables rich mechanosignaling through talin. A recent study by the Barsukov group has proposed the talin protein as being a hub for several different binding partners[45] where the proposed 3-helix intermediate state could be an essential component of binding regulation. Moreover, it has been recently shown, that the mechanical load across talin is not homogeneous, providing further variation in the regulation of talin functions[11]. Since the talin rod experiences a force gradient, once vinculin is bound, the local stress may become modulated and insufficient to unfold the 3-helix state. The 3-helix state may thus represent abundant talin rod subdomain conformation in living cells[11,43]. Experimental work is essential to confirm and refine these models.
While the activation of talin binding sites by mechanical force has been long studied, the detailed mechanism of the forced unfolding has not been previously discussed. The existence of a stable 3-helix intermediate may offer yet another level in the mechanoregulation process and in the cell’s response to mechanical stimuli. The existence of the unfolding intermediate also adds additional complexity to the assessment of the impact of mutations in the case of diseases. α-catenin truncating mutations have been detected in patients with hereditary gastric cancer[46] possibly increasing the disease susceptibility. Furthermore, α-catenin mutations have been directly associated with macular dystrophy[47]. The understanding of molecular mechanisms would shed light on the disease development and guide new treatment solutions. We may also speculate that the existence of a 3-helix intermediate, whether undergoing full unfolding or not, may provide an additional structural support. It is also possible that the 3-helix state functions as a molecular bumper reducing the impact of functional mutations present in the mechanosensitive protein. In other words, with additional level of mechanoregulation, the mutation effect on the cells behavior may be defused with only moderate effect on the cells fitness. Such a theory may be of importance in the case of talin which has been presented as a vital protein in cell and tissue biology. Yet, despite its important roles, only one mutation has been recognized as disease causing, in the talin-2 isoform located outside of the mechanosensitive region[48].
We show that α-helical proteins unfold via stable 3-helix intermediate states, representing biologically active states. smAFM and disulphide clamp mutations were used to confirm the models obtained with SMD. Our results suggest that talin is a central scaffolding hub in focal adhesions with multiple discrete unfolding states, acting as a sophisticated mechanosensor and an important regulatory switch. We further propose that the mechanical stability of α-helical domains as well as the mechanical stability of their unfolding intermediates should be considered when studying mechanoregulation models of α-helical proteins.
The following structures from RCSB Protein Data Bank were used as the protein models for the individual talin rod subdomains: R3 (id 2L7A residues 796 to 909), R9 (id 2KBB) and R11 (id 3dYJ residues 1975 to 2140). Talin R9 tandems were constructed using PyMOL, by creating a peptide bond between the last residue of the first R9 monomer and the first residue of the second R9 monomer. α-catenin including MI and MII domains (id 4IGG residues 275 to 506) was used in our simulations. The point mutations introducing cysteine residues into the talin rod subdomains in order to form the disulphide bonds (clamps) preventing the unfolding of 3-helix state, were designed and mutated using PyMOL.
MD and SMD simulations were performed using Gromacs ver 2016.1[49,50] at the Sisu supercomputer, CSC, Finland. The CHARMM27 force field[51] and explicit TIP3P water model[52] in 0.15 M KCl solution were used and the total charge of the system was adjusted by K+ and Cl- ions. The energy minimization of the system was performed in 10 000 steps using the steepest descent algorithm. The system was equilibrated in three phases using harmonic position restraints on all heavy atoms of the protein. The first phase of the equilibration was performed with NVT ensemble for 100 ps using the Berendsen weak coupling algorithm[53] to control the temperature of the system at 100 K. Integration time step of 2 fs was used in all the simulations. Following the NVT, the system was linearly heated from 100 to 310 K over 1 ns using an NPT ensemble at 1 atm of pressure. During this process, the Berendsen algorithm was used to control both temperature and pressure. For the final phase of equilibration and for all subsequent simulations, an NPT ensemble was maintained at 310 K using the V-rescale algorithm[54], and 1 atm as implemented in Gromacs 2016.1. The temperature coupling was applied separately for the protein and the solution parts. Each system was equilibrated up to 30 ns, with subsequent monitoring of the root mean square deviations (RMSD) of Cα atoms, considering the first approx. 5 ns as relaxation step. Hence, snapshots at 5 ns were used as starting structures for SMD simulations. Pulling vector was set between Cα of the first and the last residue of the appropriate domain. The movement of Cα of N-terminal residue was restrained with harmonic potential, while Cα of C-terminal residue was subjected to the constant velocity or constant force pulling. The pressure control was turned off for the pulling dimension (z-axes) in all SMD simulations as described in our previous work[28]. The constant velocity pulling SMD simulations were performed at 2 nm / ns with the spring constant set to 1000 kJ/mol nm2. In the constant force pulling SMD simulations, different force regimes were applied, 160 pN, 170 pN, 180 pN and 200 pN for R3 and 200 pN, 220 pN, 230 pN, 250 pN and 300 pN for R9. The system size in SMD was 227 thousand atoms for R3, about 800 thousand atoms for R9, R11 and α-catenin MI-MII, and about 1.2 million atoms for R9 tandems. Detailed composition of the systems used SMD simulations shown in S2 Table.
The constructs and experimental procedure for the smAFM were similar to those described before[28]. The talin fragment polyprotein constructs, including flanking I27, were synthesized and cloned in to pFN18a. The polyproteins were expressed in E. coli BL21-CodonPlus (DE3)-RILP competent cells, using the T7 promoter present in the plasmid. Protein expression was induced with IPTG when the culture reached an OD600 nm of 0.6. Cells were lysed by applying 0.2 mg/ml lysozyme for 30 minutes at 25°C, followed by sonication with an Sonifier cell disruptor model SLPe (Branson Ultrasonics Corporation, USA) and clarification of the lysate using centrifugation. The clarified lysate was subjected to Ni-NTA affinity chromatography beads in a batch process. The proteins eluted with imidazole were analyzed for purity with SDS-PAGE and used at a final concentration between 1–10 ug/mL.
Glass coverslips were functionalized with the chloroalkane ligand to HaloTag as previously described[28]. The glass coverslips were first cleaned using Helmanex III (1% in water), acetone, and ethanol washes. The surfaces were then prepped with O2 plasma cleaning for 15 min. Surfaces were then silanized using (3-aminopropyl)trimethoxysilane, diluted to 1% in ethanol. Surfaces were then washed with ethanol and then dried with N2. These amine-functionalized surfaces were then incubated with 10 mM succinimidyl-[(N-maleimidopropionamido)tetracosaethylene glycol] ester (SMPEG24 –Thermo) diluted in 100 mM borax buffer (pH 8.5) for 1 h. The final step involved incubating the surfaces overnight with 10 mMHaloTag Thiol O4 ligand in the same buffer. The surfaces were quenched with 50 mM 2-mercaptoethanol in water.
We used a commercial AFS-1 from Luigs & Neumann, GmbH, based on a device developed at the Fernandez Lab, Columbia University[55]. The cantilevers used were gold-coated OBL-10 levers from Bruker. The spring constants varied between 4 and 10 pN/nm as measured by equipartition theorem with the appropriate adjustments for cantilever geometry[56,57].Around 20 μL of protein solution was incubated on functionalized coverslips for 30 min prior to the experiments to allow for HaloTag binding. The cantilever was pressed into the surface with a force of ∼300 pN to bind the cantilever to the polyprotein. Force extension experiments were conducted at 400 nm/s retraction rate. Data analysis was carried out using Igor Pro (Wavemetrics). Unfolding peaks were identified by adjustable smoothing with a moving box average and then by searching for local maxima. The force of the peaks along with their unadjusted distance from the HaloTag benchmark was measured.
|
10.1371/journal.pntd.0005952 | Effectiveness of 32 versus 20 weeks of prednisolone in leprosy patients with recent nerve function impairment: A randomized controlled trial | While prednisolone is commonly used to treat recent nerve function impairment (NFI) in leprosy patients, the optimal treatment duration has not yet been established. In this “Treatment of Early Neuropathy in Leprosy” (TENLEP) trial, we evaluated whether a 32-week prednisolone course is more effective than a 20-week course in restoring and improving nerve function.
In this multi-centre, triple-blind, randomized controlled trial, leprosy patients who had recently developed clinical NFI (<6 months) were allocated to a prednisolone treatment regimen of either 20 weeks or 32 weeks. Prednisolone was started at either 45 or 60 mg/day, depending on the patient’s body weight, and was then tapered. Throughout follow up, NFI was assessed by voluntary muscle testing and monofilament testing. The primary outcome was the proportion of patients with improved or restored nerve function at week 78. As secondary outcomes, we analysed improvements between baseline and week 78 on the Reaction Severity Scale, the SALSA Scale and the Participation Scale. Serious Adverse Events and the need for additional prednisolone treatment were monitored and reported.
We included 868 patients in the study, 429 in the 20-week arm and 439 in the 32-week arm. At 78 weeks, the proportion of patients with improved or restored nerve function did not differ significantly between the groups: 78.1% in the 20-week arm and 77.5% in the 32-week arm (p = 0.821). Nor were there any differences in secondary outcomes, except for a significant higher proportion of Serious Adverse Events in the longer treatment arm.
In our study, a 20-week course of prednisolone was as effective as a 32-week course in improving and restoring recent clinical NFI in leprosy patients. Twenty weeks is therefore the preferred initial treatment duration for leprosy neuropathy, after which likely only a minority of patients require further individualized treatment.
| Nerve damage is a common and severe consequence of leprosy, and it can lead to impairment of sensory (sensibility) or motor (muscle) function of hand, feet and eyes. To prevent that the nerve damage becomes permanent, it is essential that new damage is treated in an early stage. Prednisolone, an immune suppressor, is the drug of choice. However, the ideal treatment duration is unknown. In our study, we compared a prednisolone treatment of 20 weeks with a longer treatment of 32 weeks. After the follow up period of 1.5 years, the percentage of patients with an improved or completely restored nerve function was the same in the two treatment groups, namely 78%. We can conclude that the longer treatment duration was not more effective, and that 20 weeks of prednisolone is sufficient to treat early nerve damage in most patients. We also found that in 15% of the patients neither one of the treatments was effective. For this specific group, more research needs to be done to find if alternative medication or perhaps even longer prednisolone treatment can be helpful.
| Leprosy is an infectious disease caused by Mycobacterium leprae. Since the introduction of antibiotic multidrug treatment (MDT) in the 1980’s, the number of leprosy diagnoses has decreased dramatically, and the disease was even declared eliminated as a public health problem at a global level in the year 2000 –i.e. less than 1 case per 10 000 inhabitants. Nevertheless, in 2015 a total of 210 000 new leprosy patients were diagnosed worldwide [1].
A main complication of leprosy is neuropathy, which often causes sensory and motor nerve function impairment (NFI). Untreated NFI can result in deformities of the hands and feet and may also affect the eyes. NFI can develop before MDT has started, but it can also arise during MDT and even several years after leprosy treatment has been completed [2,3]. The risk of developing new NFI within two years of starting MDT can be as high as 65% [4].
To prevent disabilities and deformities in leprosy patients, it is very important to detect and treat neuropathy in an early stage. Neuropathy is commonly treated with prednisolone [5], an immune suppressor that reduces the body’s immune responses towards M. leprae and relieves the pressure on the nerves by reducing inflammation and oedema [6,7].
Although ideally prednisolone therapy is adjusted to individual needs and response, this is not always feasible in field clinics, which often lack the treatment expertise of referral centres [8]. In these situations, the WHO recommends a standardized prednisolone treatment for 12 weeks [9].
Even though observational studies suggest that prednisolone can improve nerve function in 60–70% of nerves [2, 10–12], this effect has never been established in randomized controlled trials (RCT) [13]. There are indications, however, that a longer treatment duration may be more effective than the WHO standardized treatment. In an RCT in India, type 1 reaction (T1R) patients on a 20-week prednisolone course required less additional prednisolone than patients on a 12-week course [14]. Further research is needed to establish the optimal prednisolone regimen specifically for leprosy patients with NFI.
For this reason, we designed a study entitled “Treatment of Early Neuropathy in Leprosy” (TENLEP), comprising two RCTs aimed at determining how prednisolone treatment best can prevent permanent NFI. In this paper we describe the results of the Clinical trial, in which we evaluated whether a 32-week prednisolone course is more effective than a 20-week course in restoring and improving recent clinical neuropathy (<6 months) [15].
The TENLEP study was a multicentre, triple blind parallel-group clinical trial, conducted in six referral centres for leprosy in India, Nepal, Bangladesh and Indonesia. A more detailed description of the TENLEP study can be found in the study protocol paper [15].
Before trial initiation, approvals were obtained in each country from the appropriate ethical review committees: in India from the Indian Council of Medical Research and the Ethics Committee of the Foundation for Medical Research, Mumbai (FMR/IEC/LEP.01b/2011); in Nepal from the Nepal Health Research Council (Reg.no 14/2011); in Indonesia from the Komite Etik Penelitian Kesehatan RSUD Dr. Soetomo Surabaya (100/Panke.KKE/V/2011); in Bangladesh from the Bangladesh Medical Research Council- National Research Ethics Committee (BMRC/NREC/2010-2013/533). Written informed consent was taken from all patients, or their parents in case the patient was under the age of 18 years. The trial was registered in the Netherlands Trial Register (NTR2300) and in the Clinical Trials Registry India (CTRI/2011/09/002022 and 23).
Leprosy patients between 15 and 60 years of age with any recent peripheral NFI (less than 6 months onset) were eligible for the trial. Leprosy diagnosis and classification assignment was confirmed with physical exam by an experienced clinician, physiotherapist exam for neuropathy and slit skin smear microscopy. Ridley-Jopling classification was clinically assigned except in Anandaban hospital, where skin biopsy histopathology was additionally employed. NFI was established with voluntary muscle testing (VMT) and/or monofilament testing (MFT). Patients were excluded if they were pregnant, already receiving prednisolone treatment, suffered from other conditions that may affect the peripheral nervous system, or presented with a single skin lesion on the trunk as the only sign of leprosy. The sample size was calculated to be able to detect ‘restored or improved’ nerve function in 70% of the intervention group, compared to an assumed proportion of 60% in the control group. This one-tailed hypothesis, using 80% power, 5% significance and allowing for 20% loss to follow-up, lead to a sample size of 720 patients.
Prednisolone dose started at 45 mg/day for patients with low weight (≤50 kg) and at 60 mg/day for patients with high weight (>50 kg). The prednisolone dose was then slowly tapered during the treatment period, maintaining a plateau of 20–35 mg/day for 20 weeks in the 32-week arm–depending on weight group. Total dosage and dose over time are previously described [15]. To check the chemical composition of the prednisolone and placebo tablets, a random selection of packages from both treatment groups was evaluated at the start of the trial by the manufacturer (Rubicon), and by an independent Indian laboratory (Medibios Laboratories). After 20 months, the composition was checked again by the Royal Dutch Society of Pharmacists (KNMP). Treatment adherence was checked every month either verbally or by checking the medication package of the previous month.
Patients were randomly allocated to either 20 or 32 weeks of oral prednisolone, using a separate computer-generated blocked randomization sequence for each centre and weight group, blinded within pre-set patient numbered packaging. Patients were kept blinded, as the tablets appeared the same and treatment duration was kept equal by using placebo tablets. In addition, research staff and the statistician were kept blinded until all data analyses were performed. The key to treatment allocation was only broken if a patient had a serious adverse event (SAE) or required individualized treatment for a reaction or worsening NFI.
In each centre, monofilament testing and voluntary muscle testing were carried out by two trained assessors, except in the Indonesian centre where ten assessors have performed the tests. Assessments at baseline generally showed good inter-tester reliability [16]. For every patient, the sensory function of six nerves and the motor function of seven nerves were assessed on both left and right body side (total of 26 assessments). Sensory function was tested on three test-sites for each nerve, using a standard set of Semmes-Weinstein monofilaments, with the 200mg filament representing the normal threshold for the hand, and the 2g filament for the foot. For motor function assessment, the 0–5 Medical Research Council (MRC) scale was used [17]. The exact test methods and sites are previously described [13]. When the total monofilament score for a nerve was 3 or more, the sensory nerve function was considered impaired. A motor nerve scoring less than 5 on the Medical Research Council scale was also regarded as impaired. Follow-up assessments for VMT and MFT were carried out monthly during the treatment period (up to week 32), and at week 52 and 78. At baseline and the end of the study a Screening of Activity Limitation and Safety Awareness (SALSA) scale [18] and a Participation (P) scale [19] were completed for each patient. In addition, reaction severity was monitored with a Reaction Severity Scale (RSS) [20] at baseline, week 32, 52 and 78. When a patient did not show up for their follow-up appointment, a telephone call or in some cases a home visit was made in an effort to get the patient visit the clinic for assessments. When indicated in advance, a patient was allowed to miss one assessment, and the medication was provided for eight weeks.
As previously described, if nerve function deteriorated or new reaction symptoms developed, the trial treatment was stopped and an individually modified treatment was provided [15]. A decision tree was used to help clinicians decide whether or not patients should continue in the trial or receive individualized treatment in case of new NFI or reaction symptoms. Once individualized treatment was warranted, it was up to the clinician to decide on further patient management. The definition used to determine deteriorating MFT was: an increase of 6 points or more on the score per nerve since the last assessment, or an increase of 3 points or more on the score per nerve that was confirmed on the consecutive visit. VMT deterioration was defined as: a reduction in VMT score by two or more points or a reduction of 1 point on two consecutive visits. Removal from the trial treatment due to reactions was based on new or deteriorating reaction symptoms.
Patients who developed a serious adverse event (SAE) were removed from trial treatment and provided with individualized care. For both deteriorating and SAE cases, data collection continued up to the full 78 weeks.
The primary study outcome was the proportion of patients with restored or improved nerve function (of all nerves) as measured by MFT and/or VMT at 78 weeks. Secondary outcomes considered six variables: individual nerves, impairment counts, RSS, SALSA-scale and P-scale scores between baseline and week 78, SAE’s between the intervention and control groups, and the proportion of patients needing additional prednisolone with considering differences in timing, dose and duration. Definitions for restoration, improvement and deterioration in these respective categories for secondary and primary outcomes were as previously described and are depicted in Table 1 [15].
Data were entered at each centre in an Access Database and then combined to be analysed in Stata. Data were analysed according to the modified intention-to-treat principle: data of all randomized patients who matched the inclusion criteria were analysed for week 78, whether they had finished treatment or not, including all patients lost to follow-up. To handle missing data of patients lost to follow-up, the last observation carried forward method was used. For patients who received additional prednisolone, the assessment recorded at the time when additional prednisolone was first prescribed was carried forward. Only nerves with new impairments (<6 months) were included in the analyses. The primary and secondary outcomes were analysed using a Chi-Square test. A difference between treatment groups was considered significant when the p-value was < 0.05.
A total of 875 leprosy patients were enrolled in the trial, of whom 432 were randomized to the 20-week arm and 443 to the 32-week arm. Patients were recruited between February 2012 and October 2013, and the last follow-up data were collected in July 2015. The flow diagram in Fig 1 illustrates the number of patients followed up and the reasons for drop out. Seven of the randomized patients were excluded in the final analyses: one patient who had missing baseline MFT assessments and six patients who did not meet the inclusion criteria of having recent NFI. The number included in the analyses reported here was therefore 868.
The intake and distribution of several patient characteristics per research centre are presented in Table 2. Table 3 shows the baseline demographic and clinical characteristics of the total patient group. Differences in demographic and clinical characteristics did not reach statistical significance between the groups. At baseline, sensory function was more often impaired than motor function: for both groups the median number of nerves with impaired sensory function was 3, ranging from 0–12, while the median number of nerves with impaired motor function was 1 (0–13). The proportions of sensory and motor impairment per nerve are shown in Fig 2.
Of the 868 patients enrolled in the study, the trial treatment period—the first 32 weeks—was completed by 281 (65.5%) in the 20-week arm and 293 (66.7%) in the 32-week arm. Complete follow-up data until week 78 were collected for 230 (82%) of those patients in the 20-week arm and 219 (75%) patients in the 32-week arm (see Fig 1). At week 78, follow-up data were collected for 229 additional patients who did not complete trial treatment due to new NFI, new or recurrent reactions, SAEs or loss to follow-up. For the intention-to-treat analyses, the primary and secondary outcomes were analysed using the data of all patients who met the inclusion criteria (n = 868). A separate per protocol analysis was carried out including only patients who had completed treatment (n = 574).
The proportion of patients with restored or improved nerve function at week 78 was almost similar in both groups: 78.1% in the 20-week arm and 77.5% in the 32-week arm (p = 0.821). At week 32 and week 52, this outcome was not significantly different between the two groups either. Fig 3 shows the proportion of patients in each category of improvement. In the 20-week arm, more patients showed completely restored nerve function than in the 32-week arm, 23.5% against 18.7%. The per-protocol analysis, leaving out treatment non-compliers, resulted in an overall slightly higher proportion of patients with restored and improved nerves. However, again no significant difference was found between the groups: 81.9% in the 20-week arm and 81.7% in the 32-week arm (p = 0.960).
Table 4 presents the outcomes per nerve, for six sensory and seven motor nerves. The proportion of restored and improved motor nerves was overall higher than for sensory nerves. The sensory function of the radial, sural and posterior tibial nerves had the highest proportion of improvement and restoration. The motor function of the ulnar and posterior tibial nerves improved most between baseline and week 78. The median count of impairments at baseline and at week 78 is shown in Table 5. 75.7% of the patients in the 20-week arm showed an improvement in count of impairment, compared to 70.8% in the 32-week arm (p = 0.149). Table 5 also presents the results of the Reaction Severity scale, the SALSA scale and the P-scale for every assessment point. The proportion of patients with improved in RSS score at week 78 was higher in the 32-week arm (70%) than in the 20-week arm (65%), but this difference was not statistically significant (p = 0.09). At 78 weeks, the day-to-day situation for patients had barely improved: the scores of both SALSA and P-scale reduced slightly between baseline and week 78. The change over time did not differ significantly between the two arms (p = 0.638 for SALSA and p = 0.543 for P-scale).
Additional prednisolone, to treat new or deteriorating NFI and reactions, was required in 68 (16%) patients in the 20-week arm and 65 (15%) patients in the 32-week arm. This difference was not statistically significant. Interestingly, the majority (38/68) of these patients in the 20-week arm needed additional prednisolone between week 21 and week 32, while in the 32-week arm additional prednisolone was given mostly between week 33 and 52 (56/65). In the 32-week arm, the additional prednisolone given before week 32 was largely (78%) due to reaction—with or without accompanying worsening of NFI, while the additional prednisolone given after week 32 was more often for deteriorating NFI without reaction (55% of the cases needing additional prednisone did not have other reaction symptoms). Fig 4 shows the time until the first event requiring an additional prednisolone. The dose and duration of additional prednisolone treatment did not significantly differ between the two groups.
When zooming in on the group of patients that needed additional prednisolone (independent of treatment arm), it became apparent that their baseline characteristics differed from the group that did not need additional prednisolone. Table 6 gives an overview of these characteristics. Next to the presented variables, patients needing additional prednisolone during or after the trial period also showed significantly more often other reaction signs, i.e. inflamed and raised skin lesions, peripheral oedema, nerve pain and tenderness.
There were seven deaths, three in the 20-week arm and four in the 32-week arm; but none were related to trial treatment. SAE occurred in both groups, but were more often reported in the intervention group: 12 cases (2.7%) in the 32-week arm and four cases (0.9%) in the 20-week arm [21]. This difference was statistically significant (p = 0.04). The main reported SAE’s were hypertension (6), diabetes (6), and peptic ulcer (2): all commonly recognized side effects associated with prednisolone treatment. Minor side effects were reported in 66% in the 32-week arm and 68% in the 20-week arm. Detailed results on adverse events will be reported separately.
We used logistic regression to assess the (univariate) association between clinical and demographic characteristics recorded at baseline and the primary outcome of improved or restored nerve function. The five characteristics that demonstrated a significant correlation with the primary outcome were weight group, body mass index (BMI), gender, WHO disability grade and Eye Hand Foot (EHF) score. A patient with a higher BMI had higher odds on improved or restored nerve function at 78 weeks (OR = 1.09 (1.04–1.14)). The high weight group (>50 kg) had a 1.43 (1.09–1.86) higher probability of a positive primary outcome than the low weight group. Furthermore, women were 1.46 (1.06–2.10) times more likely to have a restored or improved nerve function than men. The subgroup analysis also demonstrated that patients having a grade 1 or 2 impairment at baseline were less likely to have restored or improved nerve function at week 78 (OR Grade 1: 0.73 (0.54–0.98) and OR Grade 2: 0.19 (0.12–0.28)). A higher EHF score also reduced the odds of a good outcome (OR: 0.73 (0.67–0.80). The following baseline characteristics had no influence on improved nerve function: WHO classification, Ridley-Jopling classification, BI (skin smear) and the presence or absence of severe Type 1 or Type 2 reaction.
Using a Cox Proportional Hazard model, we assessed whether there was a trend across the entire duration of follow up regarding primary outcome -i.e. not limited to 32, 52 and 78 weeks. The conclusion from this analysis was similar to what has been described above, there was no effect of treatment arm on primary outcome.
In both treatment arms in our study, a large majority (78%) of patients showed improvement and restoration of nerve function, indicating that 20-week prednisolone treatment is sufficient for most patients. However, around 15% of the patients required individualized further treatment because of reactions or deteriorating NFI. Notably, the proportion of patients needing additional prednisolone was similar in both groups (15% and 16%), and the duration and dose of the additional prednisolone treatment did not differ either. However, we did see a difference in timing of the need for additional prednisolone. New NFI and reactions were primarily reported in the first few weeks after prednisolone treatment had ended, thus after 20 weeks and 32 weeks respectively for the control and intervention arm. A similar rebound effect was observed in the Trials in Prevention of Disability (TRIPOD) study and in the study of Rao et al. [14,22]. The longer prednisolone treatment in our study seems to merely postpone the immune response in some patients.
The roughly 15% of patients that required additional prednisolone differed at baseline from the group that did not need extra prednisolone. This former group had significantly more MB patients, mainly BL and LL, higher average skin smear, higher EHF score, had T1R and T2R more often at baseline and showed more often increased signs of reaction, i.e. inflamed and raised skin lesions, peripheral oedema, nerve pain and tenderness. Further studies of an even longer prednisolone course than assessed in this study might be beneficial for these higher risk patients, as it seems that the immune response against dead M. leprae bacilli is too persistent to be permanently suppressed by a 32-week prednisolone course [23,24]. At present, however, our results reinforce that 20 weeks tapered prednisolone with an extended higher dose plateau is sufficient treatment for the majority of recent leprosy neuropathy cases, while extended individualized treatment will be needed for some patients, which may be associated with commonly recognized risk factors for complications.
Women overall had a better response to prednisolone then men. This was unexpected as previous studies show that women generally have longer delay in presenting with leprosy and may then have more severe NFI at time of diagnosis with less chance for recovery [25–27]. Though, in our study we only included patients with a delay of less than 6 months and therefore a pre-selection of women with recent NFI was made.
The group of patients needing additional prednisolone had significantly more often certain clinical baseline characteristics that are frequently associated with reduced outcomes, such as high BI, MB classification and reactions. When analysing the correlation between baseline characteristics and the primary outcome, i.e. improved/recovered vs. unchanged/deteriorated NFI, we had expected that patients with deteriorated NFI would also have significantly more often a high BI, MB classification and reactions at baseline. This was not the case, however. One explanation may be that the group of ‘unchanged’ patients diluted the group of patients with a deterioration, and made the association between these specific baseline characteristics and primary outcome less strong.
The mechanisms behind nerve damage in leprosy are generally not well understood and challenging to study. Nerve biopsies can give some insight in the pathophysiology of neuropathy, but are limited to a single point in time and do not provide information about the onset or further development of impairment [23]. Several mechanisms are described in literature though. In vitro, demyelination has been shown to occur when M. leprae invades the Schwann cells [28–30]. This process can take place in an early stage, when the immune system has not been activated yet. Secondary immune responses likely play a role via inflammatory cytokines [31–34] and T-cells [35–38], leading to demyelination or damaging of Schwann cells. Last, mechanical effects such as oedema and ischemia [23,39] can lead to axonal loss. Nerve regeneration can occur after the inflammation is controlled; however, the duration of the inflammation and nerve scarring may influence outcome [7, 39]. As a strong immunomodulator, prednisolone can impact secondary responses to reduce the inflammation, reducing oedema and scar formation [7]; thereby, providing time and space for nerve regeneration to occur if applied within <6 months of symptoms [2,8,40, 41]. When the delay in prednisolone treatment is too long (>6 months), the damage to the nerve is considered to be irreversible. From previous studies [5,42], it seems that NFI of shorter onset duration prior to treatment has better chances on improvement.
Observational studies in leprosy neuropathy have shown that prednisolone improves nerve function [2,10–12]. Although prednisolone is generally considered an important drug for improving NFI, it is not known by what mechanisms or to what extent prednisolone is responsible for the reported improvements, such as in placebo controlled randomized controlled trials like the TRIPOD studies [22,43]. Several studies have also indicated that a high proportion of patients (33–75%) experience spontaneous nerve function improvement even if left untreated, depending on the severity of NFI and type of nerve [22,44–46]. Even in 51% of placebo patients with old NFI (> 6 months of symptoms) improvement can be demonstrated without prednisolone treatment [45]. A placebo-controlled RCT could be an important next step to investigate the actual effect of prednisolone on the treatment of recent NFI. Alternatives for prednisolone should continue to be sought, as so far no treatments have been proven effective. For example, azathioprine and cyclosporine, immunosuppressants used to treat reactions, show only limited effect in studies inclusive of but not specifically for NFI [46–48]. Nerve decompression, a surgical method to relieve severe inflammatory pressure, is sometimes used to improve ongoing neuropathy in leprosy; however, controlled randomized trial evidence on the effectiveness of this method is lacking as well [49].
The strongest point of our study is the multi-centre, multi-country design which enabled representative sampling across leprosy demographics in South East Asia, where the bulk of global leprosy cases are diagnosed. This RCT was unique in maintaining a middle plateau for a longer time with a relatively high dose of prednisolone and demonstrated that a 20-week treatment was sufficient for the majority of recent neuropathy cases. This RCT was also able to demonstrate that baseline BMI of the patient was directly related to outcomes. This highlights the practical and clinical relevance for contextual realities, such as extreme poverty and malnutrition present within some leprosy-affected populations [50–52]. Such issues can represent individual complications for health and healing during a neuropathy episode.
At the conclusion of the study, three limitations became apparent. At the 78-week time point, only 72% of the patients in the 20-week arm and 78% of the patients in the 32-week arm provided data, exceeding the Evidence Based Medicine cut off of 20%, leading to a loss of validity [53]. In addition, treatment data was absent at other time points due to SAE, the need for additional prednisolone or temporary loss-to-follow up. To limit potential bias related to non-random loss of patients, data was analysed according to the intention to treat principle, using the last observation carried forward approach. By using this approach, the results of this study are likely conservative with the actual improving effect of prednisolone on NFI possibly even larger. A second limitation is that the duration of NFI prior to intake was self-reported, which may have allowed recruitment of patients with previous symptoms beyond the 6 month cut-off for inclusion and who were, therefore, less likely to demonstrate nerve improvement with treatment [5,41]. The third limitation is that Ridley-Jopling classifications were based on clinical aspects only, except for one centre that performed histopathology. The results of the logistic regression to evaluate the effect of RJ classification on the final result should therefore be interpreted with care.
We conclude that a 20-week course is sufficient to improve nerve function in 78% of patients with recent NFI. Future studies should focus on improved regimens or alternative treatments, as approximately 15% of patients with recent NFI require additional prednisolone. Moreover, alternative therapies could potentially reduce risks for developing steroid-dependency and other long term treatment side effects. Secondly, it is important to better unravel the pathophysiology of nerve function impairment and to study the actual effect of prednisolone on immune function as it relates to NFI. Understanding the mechanisms behind NFI could lead to alternative, more effective solutions for the treatment of NFI and the prevention of irreversible impairments and subsequent disabilities.
|
10.1371/journal.pcbi.1005685 | Multi-scale approaches for high-speed imaging and analysis of large neural populations | Progress in modern neuroscience critically depends on our ability to observe the activity of large neuronal populations with cellular spatial and high temporal resolution. However, two bottlenecks constrain efforts towards fast imaging of large populations. First, the resulting large video data is challenging to analyze. Second, there is an explicit tradeoff between imaging speed, signal-to-noise, and field of view: with current recording technology we cannot image very large neuronal populations with simultaneously high spatial and temporal resolution. Here we describe multi-scale approaches for alleviating both of these bottlenecks. First, we show that spatial and temporal decimation techniques based on simple local averaging provide order-of-magnitude speedups in spatiotemporally demixing calcium video data into estimates of single-cell neural activity. Second, once the shapes of individual neurons have been identified at fine scale (e.g., after an initial phase of conventional imaging with standard temporal and spatial resolution), we find that the spatial/temporal resolution tradeoff shifts dramatically: after demixing we can accurately recover denoised fluorescence traces and deconvolved neural activity of each individual neuron from coarse scale data that has been spatially decimated by an order of magnitude. This offers a cheap method for compressing this large video data, and also implies that it is possible to either speed up imaging significantly, or to “zoom out” by a corresponding factor to image order-of-magnitude larger neuronal populations with minimal loss in accuracy or temporal resolution.
| The voxel rate of imaging systems ultimately sets the limit on the speed of data acquisition. These limits often mean that only a small fraction of the activity of large neuronal populations can be observed at high spatio-temporal resolution. For imaging of very large populations with single cell resolution, temporal resolution is typically sacrificed. Here we propose a multi-scale approach to achieve single cell precision using fast imaging at reduced spatial resolution. In the first phase the spatial location and shape of each neuron is obtained at standard spatial resolution; in the second phase imaging is performed at much lower spatial resolution. We show that we can apply a demixing algorithm to accurately recover each neuron’s activity from the low-resolution data by exploiting the high-resolution cellular maps estimated in the first imaging phase. Thus by decreasing the spatial resolution in the second phase, we can compress the video data significantly, and potentially acquire images over an order-of-magnitude larger area, or image at significantly higher temporal resolution, with minimal loss in accuracy of the recovered neuronal activity. We evaluate this approach on real data from light-sheet and 2-photon calcium imaging.
| A major goal of neuroscience is to understand interactions within large populations of neurons, including their network dynamics and emergent behavior. This ideally requires the observation of neural activity over large volumes. Recently, light-sheet microscopy and genetically encoded indicators have enabled unprecedented whole-brain imaging of tens of thousands of neurons at cellular resolution [1]. However, light-sheet microscopy generally suffers from slow volumetric speeds (e.g. [2], but see also [3, 4]) and is usually applied to small and transparent brains. In scattering brains, current technologies with single-neuron resolution are usually based on slow, serially-scanned two-photon (2P) imaging methods that can only sample from O(102 − 103) neurons simultaneously with adequate temporal resolution [5]. Recent advances have enabled faster light-sheet imaging in cortex [6] and fast volumetric 2P imaging [7], but we must still contend with critical trade-offs between temporal and spatial resolution—and the need for even faster imaging of even larger neural populations.
Another critical challenge is the sheer amount of data generated by these large-scale imaging methods. A crucial step for further neural analysis involves a transition from voxel-space to neuron-source space: i.e., we must detect the neurons and extract and demix each neuron’s temporal activity from the video. Simple methods such as averaging voxels over distinct regions of interest (ROIs) are fast, but more statistically-principled methods based on constrained non-negative matrix factorization (CNMF) better conserve information, yield higher signal-to-noise ratio, recover more neurons, and enable the demixing of spatially overlapping neurons [8]. The methods described in [8] were not optimized for very large datasets, but NMF is a key machine learning primitive that has enjoyed more than a decade of intensive algorithmic optimization [9–12] that we can exploit here to scale the CNMF approach. We find that a very simple idea leads to order-of-magnitude speedups: by decimating the data (i.e., decreasing the resolution of the data by simple local averaging [13]), we can obtain much faster algorithms with minimal loss of accuracy.
Decimation ideas do not just lead to faster computational image processing, but also offer prescriptions for faster image acquisition over larger fields of view (FOV), and for observing larger neural populations. Specifically, we propose the following two-phase combined image acquisition/analysis approach. In the first phase, we use conventional imaging methods to obtain estimates of the visible neuronal locations and shapes. After this cell-identification phase is complete we switch to low-spatial-resolution imaging, which in the case of camera-based imaging simply corresponds to “zooming out” on the image, i.e., expanding the spatial size of each voxel. This has the benefit of projecting a larger FOV onto the same number of voxels; alternatively, if the number of voxels recorded per second is a limiting factor, then recording fewer (larger) voxels per frame implies that we can image at higher frame-rates. We are thus effectively trading off spatial resolution for temporal resolution; if we cut the spatial resolution too much we may no longer be able to clearly identify or resolve single cells by eye in the obtained images. However, we show that, given the high-spatial-resolution information obtained in the first imaging phase, the demixing stage of CNMF can recover the temporal signals of interest even from images that have undergone radical spatial decimation (an order of magnitude or more). In other words, CNMF significantly shifts the tradeoff between spatial and temporal resolution, enabling us to image larger neuronal populations at higher temporal resolution.
The rest of this paper is organized as follows. We first describe how temporal and spatial decimation (along with several other improvements) can be used within the CNMF algorithm to gain order-of-magnitude speed-ups in calcium imaging video processing. Next we investigate how decimation can enable faster imaging of larger populations for light-sheet and 2P imaging. We show the importance of the initial cell identification phase, quantitatively illustrate how CNMF changes the tradeoff between spatial and temporal resolution, and discuss how spatially decimated imaging followed by demixing can be interpreted as a simple compression and decoding scheme. We show that good estimates of the neural shapes can be obtained on a small batch of standard-resolution data, corresponding to a short cell-identification imaging phase. Finally we demonstrate that interleaved imaging that translates the pixels by subpixel shifts on each frame further improves the fidelity of the recovered neural time series.
Constrained non-negative matrix factorization (see Methods) relies on the observation that the spatiotemporal fluorescence activity (represented as a space-by-time matrix) can be expressed in terms of a product of two matrices: a spatial matrix A that encodes the location and shape of each neuron and a temporal matrix C that characterizes the calcium concentration within each neuron over time. Placing constraints on the spatial footprint of each neuron (e.g., enforcing sparsity and locality of each neural shape) and on the temporal activity (modeling the observed calcium in terms of a filtered version of sparse, non-negative neural activity) significantly improves the estimation of these components compared to vanilla NMF [8]. Below (cf. Fig 1), we describe a number of algorithmic improvements on the basic approach described in [8]: an iterative block-coordinate descent algorithm in which we optimize for components of A with C held fixed, then for C with A held fixed.
We begin by considering imaging data obtained at low temporal resolution, specifically a whole-brain light-sheet imaging recording acquired at a rate of 2 Hz using nuclear localized GCaMP6f in zebrafish. We restricted our analysis to a representative patch shown in Fig 2A, extracted from a medial z-layer of the telencephalon (pallium). (Similar analyses were also performed on patches from midbrain and hindbrain, with similar conclusions.) The neural centers were detected automatically using the greedy method from [8]. To ensure that the spatial components in A are localized, we constrained them to lie within spatial sub-patches (dashed squares in Fig 2A; see also Methods).
The first algorithmic improvement follows from the realization that some of the constraints applied in CNMF are unnecessary, at least during early iterations of the algorithm, when only crude estimates for A and C are available. Specifically, [8] imposed temporal constraints on C in each iteration: namely, C was modeled as a filtered version of a nonnegative neural activity signal S—i.e., CG = S, for an invertible matrix G—and therefore CG is constrained to be non-negative. We found that enforcing a simpler non-negativity constraint on C instead of CG (and then switching to impose the constraint on CG only once the estimates of A and C were closer to convergence) led to a simpler algorithm enabling faster early iterations with no loss in accuracy.
Next we found that significant additional speed-ups in this simplified problem could be obtained by simply changing the order in which the variables in this simplified block-coordinate descent scheme are updated [12]. Instead of updating the temporal activity and spatial shape of one neuron at a time (Fig 1A, black line) as in [8], which is known as hierarchical alternating least squares (HALS, [9]) or rank-one residue iteration (RRI, [14]), it turned out to be beneficial to update the activities of all neurons while keeping their shapes fixed, and then updating all shapes while keeping their activities fixed (Fig 1A, vermilion line). The ensuing method is a constrained version of the fast hierarchical alternating least squares (fast HALS, [15]) for NMF; one major advantage of this update ordering is that in each iteration we operate on smaller matrices obtained as dot-products of the data matrix Y with A or C, and there is no need to compute the large residual matrix Y − AC (which is of the same size as the original video) [10, 11]. (In the comparisons below we computed the residual to quantify performance, but excluded the substantial time spent on its computation from the reported wall time values.)
Next we reasoned that to obtain a good preliminary estimate of the spatial shape matrix A, it is likely unnecessary to use the original data at full temporal resolution [13]. Thus we experimented with the following approach: downsample temporally by a factor of k, then run constrained fast HALS (as described above) for 30 iterations, and then finally return to the original (non-downsampled) data and run a few more iterations of fast HALS until convergence. We experimented with three different downsampling methods: 1) selection of the k-th frame (this could be considered a kind of stochastic update rule, since we are forming updates based only on a subset of the data); 2) forming a median over the data in each block of k frames (applying the median over each pixel independently); and 3) forming a mean over each block of k frames. The mean approach (3) led to significantly more accurate and stable results than did the subsampling approach (1), consistent with the results of [12], and was about an order of magnitude faster than the median approach (2) with similar accuracy, so we restrict our attention to the mean approach (3) for the remainder of this work. (A further advantage of approach (3) relative to (1) is that (1) can miss fast activity transients.) Fig 1B shows the results obtained for a varying number of decimation factors k; we conclude that temporal decimation provides another significant speedup over the results shown in Fig 1A. The starting point of each line is the MSE for keeping the neural shapes obtained on decimated data fixed and solving for the time series on the full data. Refining the shapes on the full data further decreases the MSE, however by less then 1%, hence good shapes are obtained even using merely decimated data.
Besides downsampling methods, we also considered dimensionality reduction via structured random projections [16] or singular value decomposition (SVD, [17], see Methods). We compressed the data by the same factor k = 30, but found that both of these dimensionality-reduction methods were less efficient than simple decimation (Fig 1C). We further evaluated whether we can gain improvements by further compressing the decimated data via SVD or random projections, such that the reduced dimension is just slightly larger than the number of neurons. However, we did not obtain any improvements beyond plain decimation. We found this result to hold also for smaller patches that contained fewer neurons.
Further speed gains were obtained when applying spatial decimation (computing a mean within l × l pixel blocks) in addition to temporal decimation over the 30 preliminary fast HALS iterations (Fig 1C); see Algorithm 1 for full details. Strikingly, spatial decimation led not only to faster but also to better solutions of the biconvex factorization problem (where solution quality is measured by the residual sum of square errors, ||Y − AC||2, RSS), apparently because the spatially-decimated solutions are near better local optima in the squared-error objective function than are the non-decimated solutions.
In summary, by simplifying the early iterations of the CNMF algorithm (by removing the temporal deconvolution constraints to use fast HALS iterations on temporally and spatially subsampled data), we obtained remarkable speed-ups without compromising the accuracy of the obtained solution, at least in terms of the sum-of-squares objective function. But how do these modifications affect the extracted neural shapes and activity traces? We ran the algorithm without decimation until convergence and with decimation for 1 and 10 s respectively. Fig 2 shows the results for three neurons with overlapping patches. Both shapes and activity traces agree well even if the decimated algorithm is run for merely 1 s (Fig 2B and 2C) and are nearly identical if run longer (Fig 2D); hence, decimation does not impair the final obtained accuracy.
Our focus has been on speeding up CNMF, one computational bottleneck of the entire processing pipeline. For completeness, we report the times spent on each step of the pipeline in Table 1 and compare to the previous CNMF version of [8]. After loading, the data was decimated temporally by a factor of 30 to speed up the detection of the neural centers using the greedy initialization method from [8]. We further decimated spatially and ran fast HALS for 30 iterations before finally returning to the whole data and performing five final fast HALS iterations. Each trace was normalized by the fluorescence at the resting state (known as ΔF/F) to account for baseline drift using a running percentile filter. Finally, the fluorescence traces were denoised via sparse non-negative deconvolution, using the recently developed fast method of [18], which eliminated another computational bottleneck present in the original CNMF implementation (last row of Table 1).
We have shown that decimation leads to much faster computational processing of calcium video data. More importantly, these results inspired us next to propose a method for faster image acquisition or for imaging larger neural populations. The basic idea is quite simple: if we can estimate the quantities of interest (A and C) well given decimated data, then why collect data at the full resolution at all? Since spatial decimation by a factor of l conceptually reduces the number of pixels recorded over a given FOV by a factor of l2 (though of course this situation is slightly more complex in the case of scanning two-photon imaging; we will come back to this issue below), we should be able to use our newly-expanded pixel budget to image more cells, or image the same population of cells faster.
As we will see below, this basic idea can be improved upon significantly: if we have a good estimate for the spatial neural shape matrix A at the original spatial resolution, then we can decimate more drastically (thus increasing this l2 factor) with minimal loss in accuracy of the estimated activity C. This, finally, leads to the major proposal of this paper: first perform imaging with standard spatial resolution via conventional imaging protocols. Next perform the ROI detection and CNMF described above to obtain a good estimate of A. Then begin acquiring spatially l-decimated images and use the C-estimation step of CNMF to extract and demix the imaged activity. As we will see below, this two-phase imaging approach can potentially enable the accurate extraction of demixed neural activity even given quite large decimation factors l, with a correspondingly large increase in the resulting “imaging budget.”
The basic message of this paper is that standard approaches for imaging calcium responses in large neuronal population—which have historically been optimized so that humans can clearly see cells blink in the resulting video—lead to highly redundant data, and we can exploit this redundancy in several ways. In the first part of the paper, we saw that we can decimate standard calcium imaging video data drastically, to obtain order-of-magnitude speedups in processing time with no loss (and in some cases even some gain) in accuracy of the recovered signals. In the second part of the paper, we saw that, once the cell shapes and locations are identified, we can drastically reduce the spatial resolution of the recording (losing the ability to cleanly identify cells by eye in the resulting heavily-pixelated movies) but still faithfully recover the neural activity of interest. This in turn leads naturally to a proposed two-phase imaging approach (first, identify cell shapes and locations at standard resolution; then image at much lower spatial resolution) that can be seen as an effort to reduce the redundancy of the resulting video data.
We anticipate a number of applications of the results presented here. Regarding the first part of the paper: faster computational processing times are always welcome, of course, but more fundamentally, the faster algorithms developed here open the door towards guided experimental design, in which experimenters can obtain images, process the data quickly, and immediately use this to guide the next experiment. With more effort this closed-loop approach can potentially be implemented in real-time, whether for improving optical brain-machine interfaces [26], or enabling closed-loop optogenetic control of neuronal population dynamics [27, 28].
Highly redundant data streams are by definition highly compressible. The results shown in S1 and S2 Videos illustrate clearly that spatially-decimated image acquisition (the second phase of our two-phase imaging approach) can be seen as a computationally trivial low-loss compression scheme. Again, regarding applications of this compression viewpoint: reductions in memory usage are always welcome—but more fundamentally, this type of compression could for example help enable wireless applications in which bandwidth and power-budget limitations are currently a significant bottleneck [21, 29–31].
Regarding applications of the proposed two-phase imaging approach: we can potentially use this approach to image either more cells, or image cells faster, or some combination of both. In most of the paper we have emphasized the first case, in which we ‘zoom out’ to image larger populations at standard temporal resolution. However, a number of applications require higher temporal resolution. One exciting example is the larval zebrafish, where it is already possible to image the whole brain, but light-sheet whole-brain volumetric imaging rates are low [1] and current efforts are focused on faster acquisition [3, 4, 32]. Higher temporal resolution is also needed for circuit connectivity inference [33, 34] or the real-time closed-loop applications discussed above, where we need to detect changes in activity as quickly as possible. Finally, genetically encoded voltage indicators [35] may soon enable imaging of neuronal populations with single-cell, millisecond-scale resolution; these indicators are still undergoing intense development [36–39] but when more mature the resulting signals will be much faster than currently-employed calcium indicators and significantly higher temporal resolution will be required to capture these signals.
A number of previous papers can be interpreted in terms of reducing the redundancy of the output image data. Our work can be seen as one example of the general theme of increasing the ratio N/D, with N denoting the number of imaged neurons and D the number of observations per timestep, with demixing algorithms used post hoc to separate the overlapping contributions of each cell to each observed pixel. In a compressed sensing framework, [40] proposed to image randomized projections of the spatial calcium concentration at each timestep, instead of measuring the concentration at individual locations. In [41], [42], and [43], information is integrated primarily across depth, either by creating multiple foci, axially extended point spread functions (PSFs), or both, respectively. In contrast to these methods, [7] instead scanned an enlarged near-isotropic PSF, generated with temporal focusing, to quickly interrogate cells in a single plane at low spatial resolution. This approach is closest in spirit to the one-phase spatially decimated imaging approach analyzed in Figs 6–8, and could potentially be combined with our two-phase approach to achieve further speed/accuracy gains.
We expect that different strategies for increasing N/D will have different advantages in different situations. One advantage of the approach developed here is its apparent simplicity—at least at a conceptual level, we just need to ‘zoom out’ without the need for radically new imaging hardware. Throughout this work we have remained deliberately agnostic regarding the physical implementation of the spatial decimation; all of the decimation results presented here were based on software decimation after acquisition of standard-resolution images. Thus to close we turn now to a discussion of potential experimental caveats.
One critical assumption in our simulations is that the total recorded photon flux per frame is the same for each decimation level l. This is a reasonable assumption for light-sheet imaging (assuming we are not limited by laser power or by the peak or average light power on the sample): in this case, increasing the effective pixel size could be achieved easily, either isotropically with a telescope, or anisotropically, with a cylindrical lens or anamorphic prism pair. However, faster whole-brain light-sheet imaging requires faster shifts of the light sheet and imaged focal plane. This challenge can be solved by extended depth-of-field (EDoF) pupil encoding [3, 4, 32], remote focusing [44], or with an electrically tunable lens [45]. Higher light-sheet imaging rates can also be obtained with swept confocally-aligned planar excitation (SCAPE) microscopy [6]. In short, we believe our proposed two-phase imaging approach fits well with a variety of proven light sheet methods; for similar reasons, the two-phase approach would also fit well with light-field imaging methods [46–48].
In traditional two-photon imaging the situation is more complicated. The image is created by serially sweeping a small, diffraction limited point across the sample. Along the “fast” axis, the beam moves continuously, and the integrated signal across a line is constant, regardless of detection pixelation—the signal is simply partitioned into more or fewer bins. Along the “slow” axis, however, the galvonometers are moved in discrete steps, and low pixel numbers generally mean that portions of the image are not scanned, increasing frame speed, but concomitantly these ‘missed’ areas generate no signal. This consequently reduces the total number of photons collected. Thus to achieve the same photon flux over the larger (lower spatially sampled) pixels, while maintaining the same SNR, we require an enlarged PSF, which maps a larger sampled volume to each pixel. This approach was recently demonstrated to be effective in [7]; alternative strategies for enlarging the PSF could involve fixed diffractive optical elements [49] or spatial light modulator (SLM) systems [50]. Programmable phase-only SLMs offer the additional benefit of being able to dynamically change the size and shape of the excitation PSF, even between frames, which may help disambiguate closely spaced sources, and effectively control the recorded source sparsity.
In any instantiation, maximal imaging speed will be limited by the time required to collect enough photons for adequate SNR, which in turn is limited by photophysics and the light tolerance of the sample. In future work we plan to pursue both light-sheet and 2P implementations of the proposed two-phase imaging approach, to quantify the gains in speed and FOV size that can be realized in practice.
We also expect techniques for denoising, demixing, and deconvolution of calcium imaging video to continue to improve in the near future, as more accurate nonlinear, non-Gaussian models for calcium signals and noise are developed; as new demixing methods become available, we can easily swap these methods in in place of the CNMF approach used here. We expect that the basic points about temporal and spatial decimation discussed in this paper will remain valid even as newer and better demixing algorithms become available.
Light-sheet imaging of zebrafish was conducted according to protocols approved by the Institutional Animal Care and Use Committee of the Howard Hughes Medical Institute, Janelia Research Campus. Two-photon imaging of mouse was carried out in accordance with animal protocols approved by the Columbia University Institutional Animal Care and Use Committee.
The calcium fluorescence of the whole brain of a larval zebrafish was recorded using light-sheet imaging. It was a transgenic (GCaMP6f) zebrafish embedded in agarose but with the agarose around the tail removed. The fish was in a fictive swimming virtual environment as described in [51]. The closed loop setting, characterized by visual feedback being aligned with the recorded motor activity, was periodically interrupted by open loop phases. Whole-brain activity was recorded for 1,500 seconds with a rate of 2 volumes per second.
In vivo two-photon imaging was performed in a transgenic (GCaMP6s) mouse through a cranial window in visual cortex. The mouse was anesthetized (isoflurane) and head-fixed on a Bruker Ultima in vivo microscope with resonant scanners, and spontaneous activity was recorded. The field of view extended over 350 μm × 350 μm and was recorded for 100 seconds with a resolution of 512×512 pixels at 20 frames per second.
In the case of 2P imaging, the field of view contained N = 187 ROIs. It was observed for a total number of T = 2,000 timesteps and had a total number of D = 512×512 pixels. We restricted our analysis of the zebrafish data to a representative patch of size D = 96×96 pixels containing N = 46 ROIs, extracted from a medial z-layer of the whole-brain light-sheet imaging recording of T = 3,000 frames. The observations at any point in time can be vectorized in a single column vector of length D; thus all the observations can be described by a D × T matrix Y. Following [8], we model Y as
Y = A C + b f ⊤ + E (1)
where A ∈ R + D × N is a spatial matrix that encodes the location and shape of each neuron, C ∈ R + N × T is a temporal matrix that characterizes the calcium concentration of each neuron over time, b ∈ R + D , f ∈ R T are nonnegative vectors encoding the background spatial structure and global intensity, respectively, and E is additive Gaussian noise with mean zero and diagonal covariance. Our model assumes a linear relationship between fluorescence and calcium concentration as well as Gaussian noise. As emphasized in [8], more elaborate models for C can be incorporated in the alternating updates, but we did not pursue this generalization here.
For the zebrafish data we ensured that the spatial components are localized, by constraining them to lie within spatial patches (which are not large compared to the size of the cell body) around the neuron centers, thus imposing sparsity on A by construction. Because of the low temporal resolution of these recordings, the inferred neural activity vectors are not expected to be particularly sparse, and therefore we do not impose sparsity in the temporal domain. This leads to the optimization problem
minimize A , b , C , f ∥ Y - A C - b f ⊤ ∥ subject to : A , b , C ≥ 0 , A ( d , n ) = 0 ∀ d ∉ P n (2)
where Pn denotes the n-th fixed spatial patch. This problem is biconvex, i.e. solving for C and f with fixed A and b is a convex problem; likewise solving for A and b with fixed C and f is convex. As discussed in the Results section, we solve this problem by block-coordinate descent, first applied to much smaller decimated data and then using this solution as a warm start for the optimization on the full data. In the resulting Algorithm 1 we appended for concision b and f as an additional column or row to A and C respectively.
The matrix products A⊤Y and CY⊤ in Algorithm 1 are computationally expensive for the full data. These matrix products can also be performed on GPU instead CPU; whereas for the comparatively small 96×96 patches we did not obtain any speed-ups using a GPU, we verified on patches of size 256×256 that some modest overall speedups (a factor of 1.5–2) can be obtained by porting this step to a GPU.
For the decimated data the matrix products are cheap enough to iterate just once over all neurons and instead alternate more often between updating shapes and activities (instead of performing many iterations within HALSactivity or HALSshape in Algorithm 1). In early iterations our estimates of A and C are changing significantly and it is better to perform just one block-coordinate descent step for each neuron to update A (and similarly for C); for later iterations, and on the full data where it is more expensive to compute A⊤Y and CY⊤, we increase the inner iterations in HALSactivity or HALSshape.
Algorithm 1 is a constrained version of fast HALS. To further improve on fast HALS, [52] suggested to replace cyclic variable updates with a greedy selection scheme focusing on nonzero elements. This was unnecessary here because most nonzero elements are prespecified by the patches Pn; i.e., we are already focusing on the nonzero elements.
Because for the 2P data the observed imaging rate is much higher than the decay rate of the calcium indicator, we constrain the temporal traces C to obey the calcium indicator dynamics, to enable further denoising and deconvolution of the data. As in [8], we approximate the calcium concentration dynamics using an autoregressive process of order 2 (AR(2)),
C ( n , t ) = γ 1 C ( n , t - 1 ) + γ 2 C ( n , t - 2 ) + S ( n , t ) , (3)
where S(n, t) is the number of spikes that neuron n fired at timestep t. This equation can be conveniently expressed in matrix form as S = CG for a suitable sparse matrix G. We estimate the noise level of each pixel σd by averaging the power spectral density (PSD) over a range of high frequencies, and estimate the coefficients of the AR(2) process for each cell following [22]. Then we solve for A, b, C, f using the following iterative matrix updates:
minimize A , b ∥ A ∥ 1 , subject to : A , b ≥ 0 , ∥ Y ( d , : ) - A ( d , : ) C - b ( d ) f ⊤ ∥ ≤ σ d T ∀ d ∈ { 1 , 2 , . . . , D } (4)
minimize C , f ∥ C G ∥ 1 , subject to : C G ≥ 0 , ∥ Y ( d , : ) - A ( d , : ) C - b ( d ) f ⊤ ∥ ≤ σ d T ∀ d ∈ { 1 , 2 , . . . , D } . (5)
These updates are initialized with the results from constrained fast HALS (Alg 1). They impose sparsity on the spatial as well as temporal components, using the estimates of the noise variance as hard constraints to derive a parameter-free convex program. Following the approach in [53] the spike signal S is relaxed from nonnegative integers to arbitrary nonnegative values. The basis pursuit denoising problems in Eqs (4 and 5) can be solved with one of the methods described in [8]. However, a faster update of the temporal matrix C is achieved by using OASIS [18].
Every spatial component in A was normalized to have unit ℓ2-norm, with the corresponding temporal component scaled accordingly. Following [8] we then sort the components according to the product of the maximum value that the temporal component attains and the ℓ4-norm of the corresponding spatial footprints, to penalize overly broad and/or noisy spatial shapes.
In order to calculate the ΔF/F values we divided the fluorescence trace C(n, :) of each neuron by its baseline fluorescence that was obtained by projecting the rank-1 background bf⊤ onto the shape A(:, n) of the neuron, ( Δ F F ) t ≔ A ( : , n ) ⊤A ( : , n ) C ( n , t ) A ( : , n ) ⊤ b f ( t ). While we normalized A and C such that A(:, n)⊤A(:, n) = 1, the ΔF/F values do not depend on this specific normalization because the scaling factor cancels out.
To compress the data using truncated SVD in Fig 1C, we followed [17] and computed the eigenvectors V ∈ R M × T belonging to the M largest eigenvalues of the time by time covariance matrix Y⊤Y, which were then used to obtain the compressed data YV⊤. This method was faster than the randomized method due to [54]. The spatial background b for the compressed data was again initialized as 20% percentile of the original data and the temporal background as f ˜ = V 1 T. Because the compressed data and temporal traces can be negative we did not enforce the non-negativity constraint of C ˜ and f ˜, which was crucial as enforcing it indeed yielded worse results.
Random compression was performed as described in [16]. We applied the structured random compression algorithm [54] to Y and Y⊤ to obtain L ∈ R D × M and R ∈ R M × T. Specifically, we drew a Gaussian random matrix Ω ∈ R T × M and performed the QR decomposition of YΩ to obtain an orthonormal basis L. Analogously we obtained R for Y⊤. The iterated alternating fast HALS updates were C ← HALSactivity(L⊤Y, L⊤A, C) and A ← HALSshape(YR⊤, A, CR⊤), with L⊤Y and YR⊤ computed once initially.
Applying the code of [55] to the raw data we identified the neural shape matrix A1 and spatial background b1. We use the convention that the presence of a lower index l signifies an estimate and its value the decimation factor, i.e. index l = 1 denotes an estimate inferred without decimation. Further, we also obtained the denoised and deconvolved traces C1, S1 as well as f1. To emulate imaging with lower spatial resolution, spatial decimation was performed by converting A1 back into a 512 × 512 × N tensor (Y into a 512 × 512 × T video tensor) and calculating the average of non-overlapping patches of size l × l or l × 1 pixels for each of the N neural shapes (T timesteps). Converting the tensors back to matrices yielded the decimated neural shapes Al (data Yl). We proceeded analogously for the spatial background to obtain bl. The corresponding temporal traces were estimated by solving Eq (5) (with Yl replacing Y, Cl replacing C, etc.), initializing Cl and fl with the result of plain NMF that does not impose temporal constraints, i.e. solving minimizeC l , f l ∥ Y l - A l C l - b l f l ⊤ ∥ subject to Cl ≥ 0.
In order to obtain the results for 1-phase imaging without previous shape identification we solved Eq (4) for the decimated data Yl, initializing Al, bl by decimating A1, b1 and setting the temporal components to C1, f1. With increasing decimation factors an increasing number of shapes got purged and absorbed in the background, reflecting the fact that it would have been difficult to detect all ROIs on low resolution data in the first place. Using the obtained remaining shapes we again solved Eq (5) as above. The correlation values for purged neurons were set to zeros for the mean values reported in Fig 6.
To obtain some form of ground truth (Figs 6D–6F, 7C and 8C) we generated a simulated dataset Ys by taking the inferred quantities as actual ground truth: As ≔ A1, bs ≔ b1, Cs ≔ C1, fs ≔ f1. We calculated the residual Y − AsCs − bsfs⊤ and reshuffled it randomly but signal dependent for each pixel in time. We partitioned the residual for each pixel into 200 strata according to signal size and reshuffled it within each strata, thus retaining any potential link between noise variance and signal mean. The simulated dataset Ys was obtained by adding the reshuffled residual to AsCs + bsfs⊤ and the same analysis as for the original data was performed.
We performed additional control simulations that also took the inferred quantities as actual ground truth, Y * : = A 1 C 1 + b 1 f 1 ⊤, but did not rely on a reshuffling procedure. Instead, we either added Gaussian noise, Y d , t N ∼ N ( Y d , t * , σ 2 ), or Poisson noise ν + Y d , t P ∼ P ( ν + Y d , t * ), where ν is the photon count that is not due to calcium fluorescence (but rather dark counts and background light). Whereas the Gaussian noise had fixed variance σ2, the Poisson model results in heteroscedastic noise because its variance grows with the mean. The variance of the Gaussian noise was chosen to be equal to the average variance of the Poisson noise. The results shown in S2 Fig agree with those obtained by reshuffling the residual and are similar for Gaussian and Poisson noise, at least on average, though there is some spread if individual traces are considered (S2C Fig). Whereas the model (Eq 1) assumes Gaussian noise, it nevertheless performs well under Poisson noise, consistent with the results of [53].
Another control simulation merely took A1, b1 and f1 as ground truth. However, instead of taking the denoised fluorescence traces C1, which by construction followed the autoregressive model, the fluorescence traces C D were obtained from two datasets that combined electrophysiological recording and calcium imaging with GCaMP6f (11 cells) or GCaMP6s (9 cells) [23]. The calcium response kernel k ^ for each recorded cell was determined by solving the linear regression problem k ^ = arg min k ∥ y - s * k ∥ 2 where y is the noisy fluorescence data of a neuron, s its spike counts per bin, and * denotes convolution. Because only few cells were recorded, but for a longer duration than the 100 s of our two-photon dataset, we assigned to each of our ROIs the ground truth fluorescence trace by randomly selecting a recorded cell, taking a 100 s interval of its spike train that included at least three spikes, and convolving it with the cell’s kernel k. The trace was scaled such that its maximum value was equal to the maximal value of the inferred trace C1. Poisson noise was added, ν + Y D ∼ P ( ν + A 1 C D + b 1 f 1 ⊤ ), as described above. Results are shown in S3 Fig.
Projecting the noise of each pixel onto the neural shapes yields the noise of each neural time series. In practice the latter is estimated based on the noisy trace obtained by projecting the fluorescence data onto the shapes. For interleaved imaging (Fig 8) the shape of each neuron differs between alternating frames due to the varying pixelization. Therefore, instead of using one noise level for all timesteps, we estimated two noise levels σodd and σeven based on the PSD for all odd and even frames respectively. The residuals in the noise constraint of the non-negative deconvolution were weighted accordingly by the inverse of the noise level.
where y is the noisy fluorescence data of a neuron (cell index suppressed) obtained by subtracting the contribution of all other neurons as well as the background from the spatio-temporal raw pixel fluorescence data and projecting the odd and even frames of the result onto the considered neuron’s shapes aodd and aeven respectively. yodd and yeven denote the vectors obtained by taking only every second component of y starting with the first/second respectively. The denoised fluorescence c is denoted analogously. For simplicity we estimated the coefficients of the AR(2) process based on all frames without separating by noise level.
All analyses were performed on a MacBook Pro with Intel Core i5-5257U 2.7 GHz CPU and 16 GB RAM. We wrote custom Python scripts that called the Python implementation [55] of CNMF [8]. Our scientific Python installation included Intel Math Kernel Library (MKL) Optimizations for improved performance of vectorized math routines.
|
10.1371/journal.pntd.0004934 | Identifying Leprosy and Those at Risk of Developing Leprosy by Detection of Antibodies against LID-1 and LID-NDO | Leprosy is caused by Mycobacterium leprae infection and remains a major public health problem in many areas of the world. Challenges to its timely diagnosis result in delay in treatment, which is usually associated with severe disability. Although phenolic glycolipid (PGL)-I has been reported as auxiliary diagnostic tool, currently there is no serological assay routinely used in leprosy diagnosis. The aim of this study was to evaluate the effectiveness of two related reagents, LID-1 and LID-NDO, for the detection of M. leprae infection. Sera from 98 leprosy patients, 365 household contacts (HHC) and 98 endemic controls from Rio Grande do Norte, Brazil, were evaluated. A subgroup of the HHC living in a hyperendemic area was followed for 7–10 years. Antigen-specific antibody responses were highest in multibacillary (MB) at the lepromatous pole (LL/BL) and lowest in paucibacillary (PB) at the tuberculoid pole (TT/BT). A positive correlation for both anti-LID-1 and anti-LID-NDO antibodies was found with bacterial burden (LID-1, r = 0.84, p<0.001; LID-NDO, r = 0.82, p<0.001), with higher sensitivity than bacilloscopy. According to Receiver Operating Curve, LID-1 and LID-NDO performed similarly. The sensitivity for MB cases was 89% for LID-1 and 95% for LID-NDO; the specificity was 96% for LID-1 and 88% for LID-NDO. Of the 332 HHC that were followed, 12 (3.6%) were diagnosed with leprosy in a median time of 31 (3–79) months after recruitment. A linear generalized model using LID-1 or LID-NDO as a predictor estimated that 8.3% and 10.4% of the HHC would become a leprosy case, respectively. Together, our findings support a role for the LID-1 and LID-NDO antigens in diagnosing MB leprosy and identifying people at greater risk of developing clinical disease. These assays have the potential to improve the diagnostic capacity at local health centers and aid development of strategies for the eventual control and elimination of leprosy from endemic areas.
| Despite the substantial decrease in its prevalence, leprosy continues to be a worldwide challenge. Early diagnosis and treatment are important to interrupt transmission. Currently, there is no gold standard for the diagnosis of leprosy. Bacilloscopy and histopathology studies are complementary exams that provide high specificity but low sensitivity. It is therefore important to seek alternative tools to achieve rapid and accurate diagnosis. The state of Rio Grande do Norte, in Brazil, has municipalities’ that are considered hyperendemic for leprosy, such as Mossoró, the one included in this study. This city presented an average of new case detection rate (NCDR) of 45.4/100.000 inhabitants per year from 2001 to 2013, much higher than Brazil’s NCDR, which is currently 15.3. Here, we show that the utility of the recombinant antigens LID-1 and LID-NDO to diagnose MB patients and detect asymptomatic M. leprae infection. In addition, we showed that antibody levels were related to the clinical form of leprosy as well as to the bacillary load. Interestingly, we observed that serum levels of LID-1/LID-NDO antibodies can be used to predict leprosy development among HHC. The assays have the potential to eventually be implemented as point of care at local health centers.
| Leprosy is caused by Mycobacterium leprae infection and, despite the availability of free, effective multidrug therapy (MDT), it remains a major public health problem. Leprosy is the leading worldwide cause of non-traumatic peripheral neuropathy. Challenges to timely diagnosis result in delay in treatment, which leads to severe disability [1]. Of importance, one third of the leprosy cases develop immunopathologic reactions, which tend to be an additional cause of disability [2–4]. India and Brazil are the two countries with the largest number of cases [5].
Infection with M. leprae can evolve into a wide range of outcomes, from asymptomatic infection to disseminated disease. Presentation of clinical leprosy is also on a spectrum, varying between tuberculoid and lepromatous poles [6]. The tuberculoid pole (TT) is characterized by few, well-defined, hypopigmented, hypoesthetic lesions. Histopathologic analysis of lesions of TT patients usually has few or no bacilli [7]. In contrast, patients at the lepromatous pole (LL) present with numerous skin lesions, infiltration of skin and sometimes internal organs. Histology reveals foamy macrophages containing large numbers of bacilli within a disorganized lymphocytic infiltrate. Between these two poles, there are intermediate clinical forms as borderline-tuberculoid (BT), borderline-borderline (BB) and borderline-lepromatous (BL) [6]. People with TT leprosy present a strong Th1 cell mediated immune response and are usually seronegative for anti-M. leprae antibodies, while people with LL leprosy skew toward a Th2 pole and a strong antibody-mediated responses that do not control bacilli replication and are usually seropositive [8;9].
The diagnosis of leprosy is based on clinical examination, bacilloscopy and histopathology. Although, bacilloscopy and histopathology provide a high specificity, they have low sensitivity [10]. These tests also present technical and practical limitations because of their invasive nature, required materials, and need for specific technical expertise. The World Health Organization (WHO) developed a simpler classification to be applied in areas that lack the ability to carry on histopathological studies. Under WHO guidelines, patients are classified as paucibacillary (PB) when presenting with up to five lesions, and as multibacillary (MB) if they present with more than five lesions [11–13]. It is important to seek alternative and practical tools that can help to achieve the earliest possible diagnosis and therapy to interrupt both disease development and M. leprae transmission.
Serological tests can be used following finger-prick blood collection to evaluate antibody responses to M. leprae specific antigens. Phenolic glycolipid (PGL)-I has been used as the antigen [14–16]. Patients who have multibacillary leprosy produce large amounts of IgM directed towards PGL-I. The magnitude of anti-PGL-I IgM correlates well with the bacillary load [17]. Serologic responses are also detected against many proteins. The recombinant protein antigens ML0405 and ML2331 have exhibited high sensitivity for the detection of leprosy remove throughout the clinical spectrum when tested against large panels of sera from many different geographic regions (Philippines, Brazil and Japan) [18–20]. A fusion protein, LID-1 (leprosy IDRI diagnostic-1) was developed by fusing the ml0405 and ml2331 genes to produce a single chimeric protein with a better sensitivity than the original proteins alone [21;22]. Recently, LID-1 and PGL-I epitopes were conjugated to form LID-NDO. Prospective studies using LID-NDO showed high sensitivity and specificity [23].
The sensitivity of serological tests for leprosy varies depending on the geographic origin of the sera; therefore, response profile to a particular antigen needs to be evaluated in diverse populations [24]. As such, the primary objective of this study was to evaluate the ability of LID-1 and LID-NDO to discriminate the different clinical forms of leprosy in the state of Rio Grande do Norte, Brazil. We also evaluated whether positive responses against these antigens could provide a predictive value for leprosy development in household contacts of clinically identified leprosy cases.
Initially, to evaluate the ability of LID-1 and LID-NDO to detect the different clinical forms of leprosy and asymptomatic infection, we used samples collected at diagnosis and household contacts from different areas in the State of Rio Grande do Norte (Fig 1).
A total of 98 cases of leprosy were recruited from two leprosy referral centers in the state of Rio Grande do Norte, Brazil, between 2005–2014. Patients were recruited at Hospital Rafael Fernandes, Mossoró, and Hospital Giselda Trigueiro, Natal. According to the State of Rio Grande do Norte Health Secretariat database, Mossoró and Natal are the two municipalities with the largest number of leprosy cases in the state of Rio Grande do Norte [25;26]. Cases of leprosy were diagnosed as PB or MB by standard criteria and blood samples were collected prior to the initiation of MDT [11]. A subgroup of those patients (n = 50) which there were available biopsy results were characterized in accordance to Ridley and Jopling classifications [6].
Healthy household contacts (HHC) were recruited, as described in Moura et al (2013) [27]. Serum samples of these HHC were studied to assess the utility of serological assays to predict disease development. A total of 365 household contacts were clinically screened for leprosy, of whom 183 were contacts of PB and 167 were MB contacts. Serum samples from people residing in the endemic area, but with no history of contact with a leprosy case was used as endemic control (EC, n = 98). Epidemiological data for the groups are presented in Table 1. The mean age of leprosy cases was 45.6 years, 52% were female and 67.3% of the cases were MB. A subgroup of the HHC (n = 332) living in the hyperendemic area of Mossoró (recruited from 2006 to 2008) were followed for 7–10 years (Fig 1). In this cohort, through the analysis of Rio Grande do Norte leprosy database, it was identified those HHC that became a leprosy case in this interval of time.
Antibodies to LID-1 and LID-NDO were detected by enzyme linked immunoassay (ELISA), following previously established protocol [28]. The optical density (OD) for each sample in antigen-specific ELISA was obtained after subtracting the OD reading obtained in the placebo plate. Two control samples were included in each plate to normalize the overall data and account for inter-plate variations.
The protocol was assessed and approved by the Universidade Federal do Rio Grande do Norte Ethical Committee (CEP-UFRN) and by the Brazilian National Ethical Committee (CONEP/CNS/ Ministério da Saúde, Brasília). All participants or their legal guardians signed informed consent forms prior to sample collection.
We used analysis of variance (ANOVA) with the post-hoc Tukey’s test to evaluate mean differences between groups, assuming that OD values were normally distributed. The cut-offs used for sensitivity and specificity calculations were based on optimized OD thresholds generated by Receiver Operating Characteristic (ROC) analysis, assuming three different scenarios for the use of the antigens, which included: 1. The diagnose of multibacillary leprosy in a general endemic population; 2. The diagnose of multibacillary in a high risk population and 3. To identify asymptomatic M. leprae infected subjects. The diagnostic performances of LID-1 and LID-NDO tests were assessed by comparison of the Area Under the Curve (AUC) with DeLong’s test. We used the longitudinal setting (Fig 1) to estimate the predicted probability of a household contact (HHC) to develop leprosy, through simple logistic regression model using OD values as predictor variable. Additionally, we ran multiple logistic regression adjusting for age and sex, but no confounding effects were detected. All statistical analyses were performed in R, assuming a significance level of 0.05.
We evaluated the specific antibody profile to M. leprae antigens for leprosy cases, household contacts and endemic controls (Fig 1). The magnitude of the antibody responses varied according to the operational classification of patients (as either PB or MB), (Fig 2). For both antigens, MB patients presented higher levels of antibodies compared to all other groups (Fig 2A and 2B). PB patients presented low levels of antibodies, which was not different from those presented by EC or HHC of PB patients. As expected, HHC of MB patients presented a mean OD higher than that observed in HHC of PB patients, although this difference was only significant when LID-NDO was used (Fig 2B). MB patients showed a great variability in the levels of specific antibodies, which was readily apparent when samples were stratified by Ridley and Jopling classification (Fig 3A). By applying a polynomial model, we detected a linear increase of antibody levels across the spectrum from TT to LL, for both LID-1 and LID-NDO (Fig 3A and 3B). The comparison of LID-1 and LID-NDO mean OD between the different clinical forms of leprosy is presented in S1 and S2 Tables, respectively. There was an increment of 0.299 on the LID-1 mean from TT to LL poles (p < 0.001) (Fig 3A). For LID-NDO, this increment was of 0.319 (p<0.001) (Fig 3B).
To compare the ability of the two recombinant antigens to diagnose MB leprosy in ELISA assay and to establish a threshold for positive responses we compared data in both a 2x2 table which had the mean OD of the endemic control plus three times the standard deviation was used (S3 Table) and the ROC curve (Table 2 and Fig 4). The use of a 2x2 table showed that despite a high specificity obtained for both antigens, a relatively low sensitivity was found, especially when considering PB cases (either PB alone or combined with MB) (S3 Table). Therefore, we used ROC curves to determine the optimal threshold for identification of positive samples to compare the performance of the two recombinant antigens. Our analysis considered three different applicability for the ELISA to: 1. To diagnose MB cases in a general endemic population; 2. To diagnose MB cases in a high risk population (e.g. household contacts); and 3. To identify asymptomatic-infected individuals (Table 2). The sensitivity of ELISA to diagnose MB cases in the endemic population was 89% for LID-1 and 95% for LID-NDO; the specificity was 96% for LID-1 and 88% for LID-NDO.
The ROC curves used to compare the ability of LID-1 and LID-NDO ELISA to detect MB patients in either the general population or a high-risk population are presented in Fig 4. The AUC was 0.913 for LID-1 and 0.943 for LID-NDO, with no significant difference between the two antigens (p = 0.1796) to identify MB cases among the general population (Fig 4A). To identify MB cases among HHC, the antigens performance was also similar (p = 0.2861) (Fig 4B).
The scatter plot in Fig 5 represents the mean OD obtained for each sample with LID-1 and LID-NDO and the cut-offs obtained considering the scenarios 1 and 2. Despite the strong correlation between the levels of anti-LID-1 and anti-LID-NDO (r = 0.84, p<0.001), we observed that some PB patients were negative for LID-1 and positive for LID-NDO. However, all PB patients that were positive for LID-1 were also positive for LID-NDO. Considering scenario 1, the net sensitivity (95%) and net specificity (88%) were similar to those observed with the use of LID-1 or LID-NDO alone. However, considering the scenario 2 the net sensitivity (85%) and specificity (91%) were higher than those observed with the use of only one of the tests.
We observed strong positive correlations for both anti-LID-1 (r = 0.84, p<0.001) and anti-LID-NDO (r = 0.82, p<0.001) ELISA data with the bacterial index of leprosy cases (Fig 6A and 6B). However, the ELISA presented with a higher sensitivity than bacilloscopy exam. Among a total of 50 leprosy patients examined, bacilloscopy was negative in 28 cases (i.e. Bacterial Index = 0). In contrast, LID-NDO ELISA identified 5/20 and 5/8 of these PB and MB cases, respectively. These numbers yield a sensitivity to diagnose leprosy per se equal to 44% (22/50) for bacilloscopy but increased to 64% (32/50) for LID-NDO ELISA.
It is well documented that HHC of a leprosy case are at higher risk of being exposed to M. leprae and, consequently, of developing leprosy than the general population. We therefore, considered the distribution of specific antibodies levels, establishing a cut-off to define those HHC exposed to M. leprae and a cut-off to determine those HHC that presented positive responses equivalent to MB leprosy case. Irrespective of the antigen used, it is apparent that most of the HHC that was recruited in our study were previously exposed to M. leprae infection (Fig 7A and 7B). We also observed that some of the HHC presented high antibody levels, with OD similar to those measured in sera from MB patients (Fig 7A and 7B).
To determine if antigen-specific antibody responses could predict disease development, we followed 332 HHC from Mossoró for a minimum of 7 years. Among those HHC, 12 (3.6%) developed leprosy, as reported to the Minister of Health. The diagnosis of leprosy among these HHC occurred at a median of 31 (3–79) months after their recruitment into the study. Among the HHC that developed disease, at the time of recruitment 25% (3/12) presented detectable antibody responses against LID-1 and 33.3% (4/12) had responses against LID-NDO. These positivity percentages were approximately twice the percentages observed for all HHC recruited in the study (LID-1 = 11.9% and LID-NDO = 17.8%). It was observed that 50% (6/12) of the HHC that developed leprosy presented as MB. Interestingly, when we compared the mean OD of HHC that developed disease (LID-1 OD = 0.480 and LID-NDO OD = 0.879) against the mean OD of HHC that did not develop disease (LID-1 OD = 0.267 and LID-NDO OD = 0.492) a significant difference was observed between these groups (p = 0.01 for LID-1 and p = 0.003 for LID-NDO).
In order to evaluate the ability of the serological test to predict disease, we modeled the log odds of the individual being a MB as function of LID-1 and LID-NDO OD values. The predicted probabilities were then calculated from the fitted logit model, as
logodds(Yi=MB)=β0 + β1ODi
pi.hat=eβ0 + β1ODi1+eβ0 + β1ODi
(1)
For this analysis, we included household contacts (n = 332) and leprosy cases (n = 50, encompassing 8 PB and 42 MB) from the hyperendemic region of Mossoró. Fig 8A and 8B show the predicted probabilities (pi.hat) as a function of LID-1 and LID-NDO values, respectively. By computing the mean probability among all contacts, we projected a new case detection rate (i.e. incidence) of 8.3% based on LID-1 ELISA (Fig 8A) and 10.4% with LID-NDO ELISA (Fig 8B). The cut-off values for scenarios 2 (to diagnose MB cases in a high-risk population) and 3 (to identify asymptomatic-exposed individuals) were used to group individuals by risk (LR: low risk; MR: moderate risk; HR: high risk). The incidence rate for each group according to the antigen tested is in Table 3. The relative risk (RR) using LR as reference of a HHC classified as HR to develop disease was at least 7.7 times greater than those classified as LR. Interestingly, no HHC that developed leprosy in the longitudinal analysis was on the LR group (Fig 8A and 8B).
Despite the significant decrease in the global prevalence of leprosy, since the introduction of MDT, there are still many regions where the detection of new leprosy cases remains high [29]. Overall, active case finding studies indicate that the true prevalence of leprosy is still probably grossly underestimated. Diagnostic limitations hinder large-scale control programs aimed at the eventual elimination of this disease. In this sense, studies have shown that serologic tests may contribute to early diagnosis, even before the appearance of lesions [18;30;31]. In this study, we evaluated the presence of antibodies against the recombinant fusion protein LID-1 and the glycoprotein conjugate LID-NDO, in groups of individuals with leprosy or with prolonged exposure to M. leprae. Our data indicate that ELISA detecting antibodies against LID-1 and LID-NDO had high sensitivity and specificity to aid the diagnosis of leprosy and to identify people with asymptomatic M. leprae infection prior to lesion development.
We observed the highest antibody responses in MB with the clinical forms LL/BL and the magnitude of response declined throughout the clinical spectrum, with the lowest values observed in TT patients. Accordingly, the titers of specific antibodies correlated positively with bacillary indexes of patients [23;32]. These data confirm that these serum antibody responses can closely reflect infection levels and indicate that this could be a simpler, less invasive technique than skin slit smears to estimate M. leprae bacterial burden. Importantly, the serological tests were more sensitivity than bacilloscopy. Thus, these tests can certainly contribute to the accurate leprosy diagnosis, especially in areas where histopathological exams are not available.
Due to a shared environment, and likely an increased exposure, it is recognized that HHC represent a higher risk group for the development of leprosy. This is especially true when contact is with a heavily infected MB case [33–35]. As indicated by circulating antibody levels, our study suggests that a large proportion of the HHC evaluated were likely to have been previously exposed to M. leprae infection. This was expected since the participants were recruited in a hyperendemic area of Rio Grande do Norte [26]. We found that HHC of MB cases had, on average, higher levels of anti-LID-NDO than HHC of PB cases. Interestingly, HHC of MB patients also presented with higher level of antibodies than PB patients. This finding is in accordance with the results obtained for LID-1 and LID-NDO in another population in Brazil [23]. A possible explanation for this is that PB leprosy case mount an effective cellular immune response to control bacterial replication, and, consequently, mitigating antibody responses [8;9]. On the other hand, as our longitudinal study indicates, household contacts of MB patients may develop an antibody response, while either resolving infection or the possibility of subsequently developing clinical disease. In Leishmania infantum infection which can also have a spectrum of outcomes as M. leprae, an inverse correlation between antibody and cellular immune responses was observed [36].
Study of the seropositivity rates against PGL-I in Minas Gerais, Brazil, detected responses among HHC at a rate of 10.4% [34]. In our analysis, considering the cut-off used to define HHC positive for MB diagnosis, we found that seropositivity rates among the HHC of 11.9% and 17.8% for LID-1 and LID-NDO, respectively. This is a relatively high percentage, likely related to the high endemicity within the area from where this population was recruited [27]. In fact, when we considered the cut-off based on EC, the majority of HHC recruited into our study presented with antibody levels indicative of asymptomatic M. leprae infection, emphasizing the potential for leprosy to emerge and become a greater problem in the region.
Most people exposed to M. leprae develop a protective immune response and do not develop leprosy [37]. However, especially in hyperendemic areas, it is critical to develop strategies to control the transmission of M. leprae. Although not yet formally shown for LID-NDO, a previous study showed the application of LID-1 for the detection of cases up to a year before the recognition of lesions [18]. In our longitudinal study, among the 332 HHC recruited between 2006 and 2008 in the hyperendemic area of Mossoró, 3.6% were reported as developing leprosy at a later date. Interestingly, some of these HHC presented with positive serum antibody responses against LID-1 and/or LID-NDO years before the onset of clinical signs of leprosy. This suggests that LID-1 or LID-NDO could be used as a tool for early identification of M. leprae-infected individuals to increase vigilance to expedite the diagnosis of leprosy.
We compared the results obtained in our cohort with a model to predict leprosy development for HHC based on antibody levels. The probability of a HHC with positive serology to develop leprosy was 8.3% for LID-1 or 10.4% for LID-NDO. The difference between the value estimated by the model and the one observed in the cohort (3.6%) may arise due to passive case detection in the area, since studies show that the actual incidence of the disease can be much higher during active case finding [27;38]. Taken together, at a minimum, this indicates that these antigens could have helped in monitoring these individuals to provide the early diagnosis of new cases.
According to the levels of specific antibodies presented by HHC, we were able to define three groups based on risk of developing leprosy. Interestingly, when we investigated our cohort for HHC who developed leprosy we observed that they arose from the moderate and high risk groups, but not from the low risk. This reinforces the importance of conducting serological surveys in populations considered at risk (i.e. HHC) in endemic regions.
In summary, our data indicate that both LID-1 and LID-NDO ELISA represent important tools for the diagnosis of leprosy. Our data also indicate that these ELISA can be used to estimate the bacterial load of patients. In a leprosy endemic country like Brazil it is essential that new auxiliary techniques become available for disease control in various states and municipalities. Serological tests that do not require significant labor and can detect asymptomatic M. leprae infection may contribute to the control and eradication of leprosy.
|
10.1371/journal.pntd.0000822 | Design, Development and Evaluation of rK28-Based Point-of-Care Tests for Improving Rapid Diagnosis of Visceral Leishmaniasis | Visceral leishmaniasis (VL) is diagnosed by microscopic confirmation of the parasite in bone marrow, spleen or lymph node aspirates. These procedures are unsuitable for rapid diagnosis of VL in field settings. The development of rK39-based rapid diagnostic tests (RDT) revolutionized diagnosis of VL by offering high sensitivity and specificity in detecting disease in the Indian subcontinent; however, these tests have been less reliable in the African subcontinent (sensitivity range of 75–85%, specificity of 70–92%). We have addressed limitations of the rK39 with a new synthetic polyprotein, rK28, followed by development and evaluation of two new rK28-based RDT prototype platforms.
Evaluation of 62 VL-confirmed sera from Sudan provided sensitivities of 96.8% and 93.6% (95% CI = K28: 88.83–99.61%; K39: 84.30–98.21%) and specificities of 96.2% and 92.4% (95% CI = K28: 90.53–98.95%; K39: 85.54–96.65%) for rK28 and rK39, respectively. Of greater interest was the observation that individual VL sera with low rK39 reactivity often had much higher rK28 reactivity. This characteristic of the fusion protein was exploited in the development of rK28 rapid tests, which may prove to be crucial in detecting VL among patients with low rK39 antibody levels. Evaluation of two prototype lateral flow-based rK28 rapid tests on 53 VL patients in Sudan and 73 VL patients in Bangladesh provided promisingly high sensitivities (95.9% [95% CI = 88.46–99.1 in Sudan and 98.1% [95% CI = 89.93–99.95%] in Bangladesh) compared to the rK39 RDT (sensitivities of 86.3% [95% CI = 76.25–93.23%] in Sudan and 88.7% [95% CI = 76.97–95.73%] in Bangladesh).
Our study compares the diagnostic accuracy of rK39 and rK28 in detecting active VL cases and our findings indicate that rK28 polyprotein has great potential as a serodiagnostic tool. A new rK28-based RDT will prove to be a valuable asset in simplifying VL disease confirmation at the point-of-care.
| Visceral Leishmaniasis caused by Leishmania donovani is endemic in several parts of South Asia, East Africa, South and Central America. It is a vector-borne disease transmitted by bites of infected sand flies and often fatal in the absence of chemotherapy. Timely diagnosis is an essential first step in providing proper patient care and in controlling transmission. VL diagnosis in East Africa and Latin America are currently based on microscopic confirmation of parasites in tissue aspirates. The Kalazar Detect rapid test is widely used as a confirmatory test in India with very high accuracy, but sensitivity issues have severely limited its usefulness in the African sub-continent. Direct Agglutination Test is another confirmatory test used widely in East Africa and offers high sensitivity but is not field-friendly. We report on the design of a novel synthetic fusion protein capable of sequestering antibodies against three different Leishmania donovani antigens and the development of point-of-care tests for improving VL diagnosis. We believe the ease of use of these rapid tests and their high accuracy in detecting VL cases could make them useful as a first-line test, thereby eliminating the need for painful biopsies and ensuring better patient care.
| Leishmania parasites are transmitted to mammals by the bite of female phlebotomine sand flies and occasionally by the sharing of needles, by blood transfusion, or by congenital transmission. The life-cycle of Leishmania has two distinct forms: the flagellated promastigotes found in the gut of the arthropod vector and non motile amastigotes, which develop intracellularly in the mammalian host. Promastigotes injected into the skin during sand fly bite are internalized by dendritic cells and macrophages in the dermis where they lose their flagella as they transform into amastigotes. They multiply and survive within the phagolysosomes through a complex host-parasite interaction [1]. The prepatent period can vary from weeks to months and during this period disease symptoms may gradually appear and worsen with disease manifestations ranging from self-healing skin lesions, to diffuse cutaneous and mucosal manifestations and, in some cases, to severe visceral involvement of the spleen, liver and lymph nodes depending on the species of Leishmania.
Visceral leishmaniasis, known as kala-azar (Hindi for Black fever) in the Indian subcontinent and Africa, is the most severe form of the disease affecting approximately 500,000 adults and children worldwide. Following infection, the parasites disseminate through the lymphatic and vascular systems and infect other monocytes and macrophages in the reticulo-endothelial system, resulting in infiltration of the bone marrow, hepato-splenomegaly and sometimes enlarged lymph nodes (lymphadenopathy). Mortality of VL is high in the absence of treatment, which is generally lengthy, expensive and toxic. Clinical diagnosis relies on non-characteristic symptoms (long standing fever, cachexia, anemia and hepato-splenomegaly) which are only reliable in advanced cases and in epidemic situations. Parasitological diagnosis remains the reference standard in VL diagnosis, typically undertaken by microscopic examination of Giemsa-stained bone marrow, spleen or lymph node aspirates to detect amastigotes. This is not only invasive, but also suffers from low sensitivity, requiring both highly trained laboratory personnel to biopsy patients and high-powered microscopes that are typically available only in regional clinics. Thus, tissue aspirations for routine VL screenings are not feasible as a large-scale approach in remote field areas lacking electricity and even the most basic laboratory set-up.
The development of two serological tests for diagnosis of VL, the direct agglutination test (DAT) and the rK39 strip tests, have started to circumvent the need for tissue aspirates in Sudan and the Indian subcontinent, respectively. DAT is the first-line VL serodiagnostic test employed in Sudan [2] and, despite being highly sensitive and specific [3], [4], it is not optimum as a point-of-care test as it requires a laboratory capable of holding controlled temperature with overnight incubation. The rK39 strip test became possible after the identification of the k39 kinesin gene by an immunoscreen of a L. infantum expression library with sera obtained from visceral leishmaniasis patients [5]. VL patients mount a strong antibody response to the 39-amino acid, tandem repeat units in the gene, and the recombinant form of this gene, rK39, has been successfully used to develop an enzyme-linked immunosorbent assay (ELISA) [6], [7] as well as a point-of-care RDT [8], [9]. The rK39 RDT is a field-friendly, easy to use format that has been extensively tested in many countries. In a WHO supported multicenter trial, the FDA-approved rK39 RDT (Kalazar Detect- Inbios, Seattle) demonstrated excellent sensitivity (>95%) and specificity (>90%) in the Indian subcontinent (India and Nepal), but only moderate sensitivity (75 to 85%) and specificity (70–92%) in East Africa (Sudan, Kenya and Ethiopia) [10]. Reasons for the suboptimal performance of the rK39 RDT in Africa are not entirely clear and have been attributed to lower antibody levels to rK39 in infected individuals.
We previously identified k9 and k26, two Leishmania genes coding for hydrophilic proteins, and demonstrated that VL patients mount strong and specific antibody responses against K26, which can complement rK39 in a more accurate diagnosis of human VL [11]. Specific and independent antibody reactivity to each of the three antigens rK9, rK26 and rK39 have been studied and utilized in serodiagnosis of canine VL [12]. A multi-epitope, recombinant chimeric protein for serodiagnosis of canine and human VL was evaluated by fusing L. infantum k9 gene with single repeat units of k39 and k26 genes. ELISA with this fusion protein provided 96% sensitivity for canine VL compared to only 82% for human VL, with 99% specificity for both human and canine control groups [13].
We sought to improve serodiagnosis of VL by developing enhanced RDT prototypes that can detect >90% of the African VL cases, thus increasing diagnostic accuracy and overcoming limitations of the current rK39 RDT. The development of an affordable, simple, sensitive and specific point-of-care diagnostic test would clearly have a major impact on detection, control and treatment of visceral leishmaniasis patients. A synthetic gene, k28, was generated by fusing multiple tandem repeat sequences of the L. donovani haspb1 and k39 kinesin genes to the complete open reading frame of haspb2, thereby increasing antigen epitope density while providing complementing epitopes in the resulting recombinant protein. The recombinant fusion protein rK28 was evaluated on ELISA with a panel of active human Sudanese VL cases and also used to develop two prototype point-of-care tests, the results of which are presented. To our knowledge, this is the first study describing the development of a second generation rapid test for the serodiagnosis of human VL. Our results confirm that rK28 is an excellent serodiagnostic tool that has the potential to replace rK39 protein pending larger field trials.
The K28 gene was synthesized at Blue Heron Biotechnology, Bothell, WA using the Genemaker technology with a six-histidine tag downstream of the N-terminal methionine. The synthetic gene includes nucleotides 142–267 encompassing three 14-amino acid (aa) repeats of (L. donovani haspb1 gene, GenBank accession# AJ011810.1), nucleotides 2110–2343 encompassing two 39-aa repeats of (L. donovani k39 kinesin protein gene, GenBank accession# DQ831678.1) and nucleotides 1–400 (the complete ORF) of L. donovani haspb2 gene, GenBank accession# AJ011809.1). The 795 base pair (bp) product was subcloned directionally in the Nde I/Xho I sites of pET29 (Novagen, USA), and the transformants selected in XL-10 Gold cells (Stratagene, Santa Clara, CA.). Following sequence verification of the insert, the recombinant plasmid was subsequently transformed into E. coli HMS-174(DE3) for expression of recombinant protein. The k28 gene sequence has been submitted to GenBank under accession number HM594686.
Recombinant K28 protein (rK28) was produced by growing the transformed host cell HMS-174(DE3) in 2XYS with kanamycin using a fed-batch fermentation system. The production media was inoculated with a 10% inoculum culture in a log-growth phase. The culture was grown in a 10 L bioreactor (New Brunswick Scientific, Edison, NJ) to an optical density of 8–10 (A600), induced with IPTG (1mM final concentration) for 2 hours, and the cells harvested by centrifugation. Cell pellets were lysed in 50 mM Tris, pH 8.0 using an 110S microfluidizer (Microfluidics, Newton, MA) and cellular debris removed by centrifugation. The supernatant containing the expressed protein was combined with Ni-NTA agarose (QIAGEN, Valencia, CA) in a batch-bind mode and incubated overnight at 4°C. The resin was then packed in a column and washed with 10 column volumes of 20 mM Tris-Cl, pH 8.0, 250 mM NaCl, 0.5% CHAPS. Bound protein was eluted with 20 mM Tris-Cl pH 8.0 containing 400 mM imidazole. The eluted fractions were diluted 1∶3 with 20 mM Tris-Cl pH 8.0 and loaded onto a Q sepharose fast flow column (GE Healthcare Biosciences, Piscataway, NJ). The peaks eluted at 200 to 300 mM NaCl were combined; ammonium sulfate was added to make the final concentration 2 M, following which hydrophobic interaction chromatography was performed using Octyl Sepharose 4 Fast Flow (GE Healthcare Biosciences). Peaks eluted at low salt concentrations were combined, dialyzed into 20 mM Tris-Cl, pH 8.0, and sterile filtered through 0.22 µm filter. The final protein concentration was determined using BCA assay (Pierce Chemical, Rockford, IL). The lipopolysaccharide content of each protein preparation was measured by a Limulus amoebocyte lysate test (BioWhittaker, Walkersville, MD) and shown to be below 10 endotoxin units (EU)/mg of protein.
Disease-positive sera from 62 VL patients were obtained from the Gedaref state of eastern Sudan. The inclusion criteria for VL sera samples used in this study were, (i) all patients were parasitology positive, confirmed by microscopy of lymph node or bone marrow aspirates, (ii) all patients had clinical symptoms and were diagnosed as active VL cases. Microscopic confirmation of parasites was performed by trained technologists. The median age of the patients was 10 years. 32% of the recruited patients were females and 68% were males. Panels of negative sera were from 25 healthy endemic controls (EC) from the VL endemic region of Gedaref and 20 healthy non-endemic controls (NEC) from Khartoum, Sudan. Sera from patients with other infections included malaria- (n = 10), tuberculosis- (n = 10) and Salmonella- confirmed patients (n = 10) from Khartoum, Sudan (courtesy of Dr. Sayda El-Safi, Faculty of Medical Laboratory Sciences, Khartoum University, Sudan). Sera samples used in this study were collected as a part of routine diagnosis and treatment of patients in Gedaref and Khartoum, Sudan. Patients received standard treatment for the different disease indications as outlined by the Federal Ministry of Health-Sudan. All samples were subject to appropriate ethical clearance from the Faculty of Medicine, University of Khartoum and from the National Ethical Review Board at the Federal Ministry of Health-Sudan. The entire Sudanese negative-sera panel had no past history of visceral leishmaniasis. IRB approval was not sought for this study as banked sera from IRB approved protocols were used for both the ELISA and RDT testing. No personal identifiers were used nor any clinical investigation carried out as part of this study. Information about the patient's clinical diagnosis was available to us at the time this study was undertaken. Thirty normal human sera (NHS) from U.S residents with no history of international travel were used for testing non-specific reactivity of the recombinant proteins (Equitech-Bio, Kerrville, TX).
rK39, rK28, rK26, rK9 were titrated (200–25 ηg/well) with different dilutions of positive and negative sera (1∶100, 1∶200, 1∶400) using a checker board titration on flat-bottomed MediSorp™ and PolySorp™ Nunc MicroWell™ plates to determine the optimized ELISA conditions. A human immunoglobulin (IgG) standard curve was constructed using chrompure human IgG (Jackson ImmunoResearch, West Grove, PA) and used as a reference standard on every plate [14]. The first two columns on every plate were coated in duplicate with 4-fold dilutions of the standard curve (100, 25, 6.25, 1.563, 0.391, 0.098, 0.024, 0 µg/well) in 0.1 M bicarbonate buffer pH 9.6, 0.01% BSA, 0.1% sodium azide. The rest of the plate was coated with 25 ηg /well of the antigen in 0.1M bicarbonate buffer pH 9.6 at room temperature for 2 hours. The non-specific reactivity on the plate was blocked with 1% BSA in phosphate-buffered saline pH 7.2, 0.1% Tween 20 for a period of 2 hours at room temperature. The plates were washed in wash buffer (PBS, 0.1% Tween 20) four times and 100 µl of 1∶400 dilution of the sera in serum diluent (0.1% BSA in phosphate buffered saline pH 7.2, 0.1% Tween 20) added to the antigen wells and 100 µl of serum diluent added to the standard curve wells in duplicates and incubated at room temperature for an hour on a micro plate shaker at 500 rpm. The plates were washed in wash buffer and the bound antibodies were assayed using 100 µl per well of 1∶10,000 diluted Rec. Protein-G HRP (Zymed, San Francisco, CA) at room temperature for 1 hour. The enzyme reaction was developed with 100 µl per well of SureBlue TMB 1-component microwell peroxidase substrate (KPL, Gaithersburg, MD) for 5 minutes. The reaction was stopped using 50 µl/well of 1 M sulfuric acid, plates were read at 450 nm on a ThermoMax microplate reader and data analyzed using SoftMax Pro (Molecular Devices, Sunnyvale, CA). The final conditions that produced the best observed test results for the Sudanese panel of sera were 25 ηg/well of antigen, 1∶400 dilution of serum, and a 1∶10,000 dilution of recombinant protein G-HRP as the enzyme conjugate for detection on Medisorp plate surface. Extensive optimization experiments were performed with different parameters in order to determine the final conditions that could best differentiate between strong positive and borderline positive VL sera from truly negative sera. Reproducibility of the ELISA results were confirmed by having at least two independent operators (blinded to the identity of test samples) perform the same assay.
The human IgG standard curve was used as a reference standard to control for inter-plate variation, as well as to determine that the test was run properly. The standard curve was plotted on a 4-parameter curve fit, and an r2 value of 0.995 was required to validate the data from the plate. A receiver-operator characteristic (ROC) curve was used to evaluate all possible combinations of sensitivity and specificity and to determine an optimal cut-off that clearly discriminates between disease-positive and -negative sera [15], [16]. The ELISA test results from 105 non-VL sera {endemic controls (EC), non-endemic controls (NEC), US normal human sera (NHS) and other infection sera} were used as the negative data set, and the results from 62 VL-confirmed sera were used as positive data set. GraphPad Prism 4.0 software (GraphPad Prism Inc., San Diego, CA) was used to perform the statistical analyses. ROC curves were plotted using the software, and a table of sensitivity and specificity with all possible cut-offs were generated with 95% confidence intervals. The sensitivity of the test was determined as the fraction of the VL confirmed sera that were test positive, and specificity was calculated as the fraction of the EC, NEC, NHS and other infection patient sera that were identified to be truly test negative. The positive and negative predictive values of the tests were calculated. Area under the curve (AUC) was used as measure of diagnostic accuracy of the test providing a means to truly discriminate between disease-positive and disease-negative sera. Correlations in antibody responses were studied between individual component antigens and rK28. A nonparametric Spearman correlation was used to calculate the correlation coefficient.
Purified rK28 was provided to two manufacturers, EASE-Medtrend (Shanghai, China) and Chembio Diagnostic Systems (Medford, NY) to develop rapid tests. The EASE-Medtrend single lateral flow test utilizes a proprietary dynamic flow principle. The test antigen is immobilized on a nitrocellulose membrane within the test zone. The liquid conjugate is applied to the device through the reagent port, priming the device to facilitate the migration of serum applied in the sample port. The specific antibodies present in the serum are captured by the immobilized antigens and subsequently visualized in the form of a magenta-colored test line by the conjugate. In the control zone, a conjugate-binding reagent is immobilized on the membrane. A magenta line in the control zone appears in every valid test. The EASE-Medtrend rK28 based prototype will be referred to as K28-LF. A subset of Sudanese VL sera with low ELISA reactivity to rK39 were selected for studying the additive effect of antigens on the lateral flow format using the K28-LF RDT.
The Chembio immunoassay format is called the Dual Path Platform (DPP). It differs from conventional lateral-flow systems in that the test sample and the marker-detecting conjugate are delivered to the test line area independent of each other. The DPP assay has two laminated strips, connected to each other as a “T” shape inside a disposable plastic cassette. The first strip receives a sample and running buffer through the sample port. The sample migrates along the strip towards the second strip containing the test and control bands. Development of the assay is achieved by adding buffer to the development port. This step releases the conjugate (colloidal gold) and facilitates its migration to the test area. Antibodies, if present in the test sample, will bind to the capture reagent immobilized on the second strip, and the conjugate will react with this complex, making the test band detectable by visual evaluation. Irrespective of the presence of antibodies in the test sample, the control band should develop to assure correct DPP assay performance. The Chembio rK28-based prototype will be referred to as K28-DPP.
Testing of the two K28 RDT prototypes with larger sera panels were carried out in parallel in Sudan and Bangladesh. The rK28-DPP RDTs were tested in Sudan using 73 parasitology (LN aspirates) confirmed VL samples, 24 healthy endemic controls, 18 tuberculosis- and 20 malaria- confirmed sera. Evaluation in Bangladesh was carried out using the rK28-LF RDT and included 53 parasitologically (spleen aspirates) confirmed VL samples, 20 healthy endemic controls and 20 healthy non endemic controls. The rK39-based Kalazar Detect (InBios International Inc. Seattle, WA) RDT was used as a comparator in all studies.
Both rapid test formats use recombinant protein A-colloidal gold conjugate for detection and were performed according to manufacturer's specifications. The test results were determined in 10 minutes for the Inbios and Ease Medtrend strips and in 15–20 minutes for the Chembio DPP tests. Every sample was tested twice and the results were scored by the operator as well as independently by an individual blinded to the identity of serum samples. While microscopic confirmation of parasites is performed by well trained technologists, minimal training was required for individuals scoring the RDTs.
Sera samples used in this study were collected as a part of routine diagnosis and treatment of VL patients in Gedaref Hospital, Sudan and at the Rajshahi Medical College Hospital, Bangladesh. Study protocols for the collection were approved by the Institutional Review Boards of Khartoum University and Rajshahi Medical College. In Bangladesh and Sudan, written consent was obtained from all adult patients and parents or guardians of children. For samples from illiterate adult participants and children, verbal consent was read to and discussed with adults/guardians in presence of a literate relative, the consent form was signed by the literate relative and a fingerprint was obtained from the parent/guardian. IRB approval was not sought for this study as banked sera from IRB approved protocols were used for both the ELISA and RDT testing (retrospective study). No personal identifiers were used nor any clinical investigation carried out as part of this study. Information about the patient's clinical diagnosis was available to us at the time this study was undertaken.
The synthetic gene k28 was designed by fusing nucleotide sequences for three 14-aa tandem repeats of the L. donovani haspb1 gene [17], two 39-aa tandem repeats of the L. donovani kinesin gene [18] and the entire 133 aa of the L. donovani haspb2 gene (Figure 1A) [17]. The 795 bp nucleotide sequence cloned in pET-29a was used to express an N-terminal 6XHis-tagged recombinant protein in E. coli. The fusion protein was purified by affinity chromatography over a Ni-NTA agarose matrix. The 264-aa sequence (Figure 1B) encoded an acidic protein (pI 4.73) with a predicted molecular weight of 28.33 kDa. The protein migrated aberrantly around 40 kDa on SDS-PAGE (Figure 1C). Hypothetically, the mobility of rK39 and rK28 should be slower than the mobility of rK26. The predicted molecular masses of both rK39 (35.3 kDa) and rK28 (28.33 kDa) are indeed higher than that of rK26 (26 kDa) and yet they seem to migrate faster compared to rK26 (26 kDa). The slower mobility of rK26 is due to its higher proline content compared to either rK28 or rK39.This characteristic has been observed for other proteins with high acidity and high proline/lysine content including HASPB1, K26 and K9 [11], [17], [19].
In order to evaluate the antigen-specific antibody responses against rK28, rK39, rK26, and rK9, antibody ELISA's were optimized to obtain the best signal-to-noise ratio and develop a reproducible and robust assay that was capable of capturing antibodies over a biologically relevant assay range. Sudanese parasitology confirmed VL-positive and -negative sera (NHS, EC, NEC and other infection sera) were tested by ELISA on rK28, rK39, rK26, and rK9 to evaluate immunoreactivities (expressed as A450nm in Fig. 2) against individual proteins. The overall OD responses of individual VL-confirmed sera were quite similar for rK39 and rK28 antigens (Fig. 2A and B, respectively), and both were much higher than the responses to rK26 and rK9. The ELISA cut-off values of the 4 recombinant proteins rK28, rK39, rK26, and rK9 were 0.4151, 0.3043, 0.3149, and 0.1589, respectively, and were determined by ROC (Receiver-Operator Curve) analyses of the absorbance values at 450 nm (Table 1).
The sensitivity, specificity and area under the curve (AUC) were calculated for all 4 recombinant proteins (Table 1). rK28 was test-positive on 60 of the 62 VL-positive serum samples, yielding a sensitivity of 96.8%, while rK39 had a sensitivity of 93.5% (58/62). rK26 and rK9 both missed 6 out of 62 VL positive sera and had sensitivities of 90.3%. Specificity was calculated using a panel of 105 sera that included healthy endemic, healthy non-endemic, and other confirmed infectious disease sera, together with sera from healthy non-travelers from the United States. Based on the ELISA cut-off (Table 1), rK26 had the highest specificity of 97.1%, closely followed by rK28 with a specificity of 96.2%. rK39 had a specificity of 92.4% while the least specific was rK9 with a specificity of 82.9%. All healthy U.S donors (30 samples) were test-negative on rK39, rK28 and rK26 ELISA. rK9 was least specific, as 4/30 U.S donors reacted non-specifically on the ELISA (data not shown). The area under the curve (AUC) is a widely accepted metric for evaluating diagnostic accuracy [20]. The greater the AUC, the better the accuracy of the diagnostic test, and an AUC of 1 represents perfect accuracy [21]. The ROC curves obtained for the ELISA using absorbance values for rK28, rK39, rK26 and rK9 are shown in Figure 3. rK28 had the highest AUC (AUCrK28) with a value of 0.98. This was followed in order of accuracy by AUCrK26 = 0.97, AUCrK39 = 0.96, and finally AUCrK9 = 0.94.
rK28 is a fusion polyprotein comprising regions of L. donovani haspb1 (L. infantum k26 homologue), L. donovani kinesin (L. infantum k39 homologue) and L. donovani haspb2 (L. infantum k9 homologue). For many of the Sudanese VL sera tested, the relative absorbance observed were higher on the rK28 ELISA compared to rK39, rK26 or rK9. We also observed a subset of individual sera with very low reactivity to rK39, but much higher reactivity to rK28. The absorbance values (A450) of individual sera to rK28 were plotted against the absorbance values of the three individual proteins rK39, rK26 and rK9 and are shown as scatter plots (Figure 4). A nonparametric spearman correlation analysis was done to calculate the correlation coefficient r and two-tailed P values. The antibody levels measured by rK39 and rK28 (Spearman r = 0.8383, P<0.0001); rK26 and rK28 (Spearman r = 0.7141, P<0.0001); rK9 and rK28 (Spearman r = 0.4112, p = 0.0009) displayed a positive and significant correlation. Overall, there seemed to be a cumulative increase in the absolute magnitude of the antibody responses when using rK28 protein to capture serum antibodies. In order to investigate the cumulative effects of rK28, VL sera with low K39 reactivity were selected, and the responses against individual proteins titrated against 10-fold serial dilutions of the sera (1∶100–1∶1000000) in an ELISA (Figure 5). We observed that a majority of the VL sera tested in this manner had a much higher response to rK28 and in some cases a stronger response to rK26 (Figure 5C).This may be due to the fact that rK28 can capture circulating antibodies against all three component proteins (Haspb1, LdK39, and Haspb2) leading to a more robust signal. Therefore, it is likely that tests using rK28 protein could potentially diagnose individuals that are missed by rK39.
In order to evaluate rK28 on a point-of-care test, single lateral flow RDT prototypes (K28-LF) were developed by EASE MedTrend (Shanghai, PRC) based on their proprietary dynamic flow principle. Preliminary screening of 13 VL Sudanese sera (selected on the basis of low rK39 ELISA reactivity) demonstrated a much higher sensitivity for the K28-LF (92.3%) compared to a much lower sensitivity with Kalazar Detect (69.2%) (Table 2). Neither RDTs gave any false positive results with sera from U.S healthy subjects. To obtain a more realistic performance of the RDTs in a VL endemic region, we next evaluated both the rK28-LF and Kalazar Detect on a larger panel of VL confirmed sera in Bangladesh. 53 parasitology confirmed VL sera and 40 healthy endemic control sera with no history of VL were evaluated. Once again the K28-LF RDT provided more favorable results (98.1% sensitivity, 92.5% specificity) compared to Kalazar Detect (88.7% sensitivity, 100% specificity) (Table 3).
A second prototype test of rK28 (K28-DPP) using a distinct technology was developed by Chembio Diagnostic systems. Sera from 73 parasitology confirmed VL patients who were DAT or smear-positive and 62 negative sera (24 endemic controls, 20 malaria- and 18 tuberculosis-confirmed patients) with no history of VL were evaluated in Sudan. All sera were also tested with Kalazar Detect as a comparator (Table 4). The K28-DPP RDT proved to be superior and provided a sensitivity of 95.9% and specificity of 100% while the Kalazar Detect yielded a sensitivity of 86.3% and specificity of 96.4%. The DAT, which was performed on every serum sample used in this study, provided a sensitivity of 94.5% and specificity of 100%.
Aspirates from bone marrow, lymph nodes or the spleen are typically done to confirm the diagnosis of visceral leishmaniasis. Although the specificity is high, the sensitivity of microscopy varies and is greatly influenced by the experience of the individual making the smear, the quality of the smear, and the reagents used. Microscopy is available only in tertiary care or referral centers/hospitals in endemic countries and is a time-consuming procedure. These factors make it difficult to accurately diagnose VL patients in primary care settings. The identification of rK39 as a marker of active VL disease [5] followed by its use in a rapid test format [9] has revolutionized VL diagnosis in the Indian subcontinent. While the rK39-based rapid test has greater than 95–98% sensitivity in the Indian subcontinent and is now widely used as a means of confirming diagnosis of VL patients, its sensitivity is lower in the VL endemic regions of Africa, limiting its usefulness as a widely used point-of-care serodiagnostic test. Detailed studies of human leukocyte antigen (HLA) polymorphisms between VL subjects from the Indian and African subcontinents would be valuable to explain differences in epitope specificities between these populations.
This study was initiated with the goal of developing a highly sensitive, specific, cost-effective, and rapid point-of-care serodiagnostic test for VL diagnosis that would improve upon the rK39 RDT. Our strategy included design of a synthetic gene, k28, harboring sequences fused from three L. donovani tandem repeat containing genes (haspb1, LdK39 and haspb2). Previous work done by our group revealed that increasing the number of tandem repeat units exponentially increases the ability to capture antibodies in the serum [22]. We utilized this information and incorporated multiple tandem repeat regions of haspb1 and LdK39 in order to increase the antigen epitope density within the resulting fusion protein. The L. infantum homologues of these genes have previously been shown to have good serodiagnostic value for both human and canine VL [11], [12], [13]. The tandem repeat regions found in many protozoan proteins usually contain immunodominant B-cell epitopes capable of generating high levels of antibody response in infected individuals [23], [24], [25], [26].
Serological responses of Sudanese VL patients were tested by ELISA against rK28 and compared with individual L. infantum homologue proteins rK39, rK26, and rK9. In order to model a realistic scenario in a VL endemic country, our specificity data included healthy endemic and non-endemic control sera as well as sera from individuals with other diseases. Healthy human sera from the USA were also included as part of the specificity panel to evaluate if the recombinant proteins had any false positive reactivities. From the testing done on ELISA, rK28 was more sensitive and specific than rK39. The ROC curves also predicted a higher diagnostic accuracy for rK28 compared to rK39.
We next sought to determine whether VL sera with low/borderline ELISA reactivity against rK39 could be detected with greater accuracy using rK28. Our results showed that many of the rK39 low-reactive sera had higher reactivity to rK28 and in some instances to rK26. The benefit of using rK28 for VL diagnosis arises from its ability to capture circulating antibodies to 3 Leishmania antigens compared to rK39, which binds antibodies specific to a single antigen. The cumulative antibody binding observed with rK28 raises the intensity of the signal and makes the border-line positive low rK39 sera distinctly positive.
The higher sensitivity of rK28 in effectively identifying low rK39 sera prompted us to exploit this characteristic for developing rK28-based point-of-care rapid tests. rK28 was provided to two independent manufacturers for prototype RDT development to ensure that test format-specific constraints would not limit product development. Also, having multiple manufacturers creates a healthy competition promoting lower costs and better quality tests for clinicians.
Testing of 13 low rK39 reactive Sudanese VL sera with the rK28-LF prototype confirmed significantly higher sensitivity (92%) afforded by a rK28-based test in comparison to the Kalazar Detect test (69%). To evaluate accuracy of the rK28 RDT prototypes in detecting VL patients, independent studies were conducted with larger sera samples in Sudan and Bangladesh, two countries where VL is endemic. As the prototype tests were manufactured on a small scale for conducting pilot studies, the two RDT prototypes could not be tested in both countries. The rK28-DPP afforded 96% sensitivity in detecting DAT or smear-positive active VL patients in Sudan, while the rK28-LF RDT provided 98% sensitivity in detecting microscopy-confirmed active VL cases in Bangladesh. Overall, both rK28-based prototype tests proved more sensitive in detecting VL cases compared to the rK39-based tests. rK28-DPP tests also proved to be highly specific (100%) in Sudan while the rK28-LF was somewhat less specific (92.5%) in Bangladesh. Large-scale field studies in both countries for selection of the final test format are planned. Further testing in the field and close follow-up of healthy individuals in VL endemic areas who have tested positive on the rK28 tests, but lack clinical symptoms, will further illustrate characteristics of the fusion protein and help us determine whether rK28 is capable of acting as an early marker of infection. The samples used as a part of this study were VL patients confirmed by parasitology, therefore, the role of K28 rapid test in detecting parasitology negative and DAT negative VL patients is yet to be studied. This will be crucial in determining the true accuracy of the rapid tests.
Due to the lack of accurate and non-invasive field-applicable tests, VL patients (a majority of whom are children) undergo extreme pain and discomfort as a result of diagnosis by tissue biopsy. The K28 RDT's could become a crucial tool for VL diagnosis, providing an easy alternative to biopsies. Early diagnosis and treatment of VL are crucial for both the affected individual and for the community. Untreated VL patients act as a reservoir of disease, especially in Africa and the Indian subcontinent where the disease is anthroponotic. Early and accurate case detection and treatment are essential components in VL control and elimination programs. Identification of affected individuals using an affordable serodiagnostic test prior to using expensive confirmatory tests for parasite detection and subsequent initiation of treatment would greatly impact timely case management and disease control. Use of a low-risk, field-based diagnostic test to detect active disease with greater accuracy, as well as monitor sub-clinical infection rates would significantly impact population-based control of disease and potentially reduce time to cure for individual patients. Multicenter large scale field evaluation of these prototype formats, including the rK39-based RDT as a comparator, are being planned to enable selection of a rK28-based RDT.
In conclusion, we have designed a new synthetic fusion protein for improved serodiagnosis of VL. The rK28 protein affords higher sensitivity in detecting active VL cases compared to rK39 both on ELISA and RDT format. The development of an rK28-based point-of-care test has yielded promising results and will become a valuable tool in rapid diagnosis of VL in conjunction with complementary tools such as parasite circulating antigen detection tests and nucleic acid detection tests and permit addressing the under-reporting of this neglected disease.
|
10.1371/journal.pgen.1003156 | Post-Transcriptional Control of the Escherichia coli PhoQ-PhoP Two-Component System by Multiple sRNAs Involves a Novel Pairing Region of GcvB | PhoQ/PhoP is a central two-component system involved in magnesium homeostasis, pathogenicity, cell envelope composition, and acid resistance in several bacterial species. The small RNA GcvB is identified here as a novel direct regulator of the synthesis of PhoQ/PhoP in Escherichia coli, and this control relies on a novel pairing region of GcvB. After MicA, this is the second Hfq-dependent small RNA that represses expression of the phoPQ operon. Both MicA and GcvB bind phoPQ mRNA in vivo and in vitro around the translation initiation region of phoP. Binding of either small RNA is sufficient to inhibit ribosome binding and induce mRNA degradation. Surprisingly, however, MicA and GcvB have different effects on the levels of the PhoP protein and therefore on the expression of the PhoP regulon. These results highlight the complex connections between small RNAs and transcriptional regulation networks in bacteria.
| Regulation of bacterial gene expression participates in the ability of these microorganisms to quickly adapt to their environment. This regulation can occur at every level of gene expression. For instance, two-component systems are involved in transcriptional control, while small RNAs usually act at the post-transcriptional level. In this study, the pleiotropic small RNA GcvB is identified as the second small RNA regulator of the central PhoQ/PhoP two-component system, which highlights the connections between the different types of regulation.
| Gene regulation in response to environmental conditions is a key feature of bacterial cells, that allows their development in multiple and diverse niches. While this was originally thought to rely only on transcriptional control by proteins, it is now well established that mechanisms underlying the control of gene expression are much more diverse. For instance, numerous examples of post-transcriptional control have been reported that can be mediated by proteins, riboswitches or trans-acting small RNAs (sRNAs) [1].
Even though the first example of a chromosomally-encoded bacterial sRNA regulating the expression of a target-gene encoded at a different locus was described almost 30 years ago [2], it is only in the last decade that our understanding of the number and the role of sRNAs in bacterial physiology has greatly improved [3]. Among those, the Hfq-dependent sRNAs have been extensively studied. This class of sRNAs imperfectly pairs to target-mRNA(s), which in most cases occludes the ribosome binding site (RBS) of the target-gene and therefore down-regulates its expression through translational inhibition. This is often accompanied by degradation of the target-mRNA, either as a consequence of the translation inhibition and/or independently of this process [4]. It is also known that sRNAs can activate gene expression, again by increasing translation and/or stability of their target(s) through base-pairing interactions [5]–[8].
Hfq both prevents the sRNAs from being degraded, and facilitates and stabilizes sRNA-mRNA duplexes. As a result, a productive interaction between an Hfq-dependent sRNA and its target relies only on short and imperfect duplexes. Most, if not all, Hfq-binding sRNAs have multiple targets and, in parallel, a single target can be regulated by multiple sRNAs. This, in addition to the great number of sRNAs in bacterial species (>80 in E. coli for instance), contributes to the importance of these molecules in bacterial physiology. One of the best examples is probably the Hfq-dependent GcvB sRNA that has been shown to target more than 20 different mRNAs, most of them probably directly. Its transcription is activated by the product of the adjacent gene, GcvA, a regulator that also controls the gcvTHP operon as well as its own transcription [9]. Whereas GcvA negatively autoregulates its own synthesis, it can either activate or repress expression of gcvTHP, the glycine cleavage operon whose products catalyze the oxidation of glycine into carbon dioxide, ammonia and a one carbon-unit that will be transferred to tetrahydrofolate. Whether GcvA activates or represses gcvTHP operon expression depends on the presence of the GcvR protein and/or glycine. The GcvA/GcvR complex acts as a repressor; but in the presence of glycine, it is disrupted, allowing GcvA to activate the synthesis of the glycine cleavage system. In contrast, purines seem to promote repression. Similarly, gcvB transcription requires GcvA and is repressed by GcvR unless glycine is present. In addition, the Lrp global regulator has a positive effect on gcvTHP expression [10], but represses gcvB expression [9], [11].
As a result of this control by GcvA, GcvB is mostly present in fast-growing cells in rich medium [12]. It negatively controls expression of multiple targets involved in aminoacid transport and metabolism [9], [12]–[14]. As is often the case for sRNAs with several targets, a unique region of the sRNA, referred to as R1 for GcvB, pairs with almost all targets. This region is very well conserved, single-stranded and GU-rich [12]. So far, only 3 targets have been found to be regulated by GcvB independently of its R1 region: lrp, gdhA and cycA [13].
Also highlighting the importance of sRNAs is the fact that several of them target transcriptional regulators. This is true for instance for the master regulator of stationary phase RpoS, whose expression is post-transcriptionally controlled by at least 4 distinct sRNAs [8], [15]–[17]. Several two-component systems (TCS), such as EnvZ/OmpR or DpiA/DpiB, have also been shown to be repressed by sRNAs [18], [19]. Similarly, we have shown in a previous work that MicA, an RpoE-dependent sRNA known to repress the synthesis of multiple proteins [20], many of which are located in the outer membrane [21]–[23], was a direct regulator of PhoQ/PhoP synthesis [24]. This TCS is a central regulatory system in which the PhoQ sensor protein controls the phosphorylation status of the cognate response regulator PhoP, so that it is activated (i.e. phosphorylated) upon low magnesium conditions or in presence of antimicrobial peptides. Under such conditions, PhoP directly regulates dozens of genes involved in major cellular functions such as magnesium homeostasis, bacterial virulence, cell envelope composition and acid resistance [25]. Our previous findings linked therefore the expression of phoPQ operon to cell envelope stress through the regulatory sRNA MicA. In addition, they strongly suggested the existence of at least another Hfq-dependent sRNA controlling expression of phoPQ at the post-transcriptional level. In this study, we identify GcvB as such an sRNA, and address the mechanism as well as the physiological consequences of this control on the expression of the PhoP regulon.
Even though MicA is an Hfq-dependent sRNA, expression of phoP was found to be strongly activated at the post-transcriptional level by the deletion of hfq in both wt and micA deleted cells [24]. We thus hypothesized that one or several Hfq-dependent regulator(s) could affect phoP expression independently of MicA. Therefore, we transformed a strain carrying a PBAD-phoP-lacZ reporter fusion with a plasmid library overexpressing most of the known E. coli Hfq-dependent sRNAs from an IPTG-inducible modified Plac promoter [8]. Transcription of the phoP-lacZ fusion is driven by the PBAD promoter so that expression of this fusion should not be sensitive to control of phoP at the transcription initiation level. The transcription start site is expected to be identical to that of the proximal phoP promoter, P1, which is normally positively regulated by PhoP in E. coli [26]. The fusion encompasses only 66 nts of phoP mRNA, that correspond to a 36 nts 5′ leader followed by the first 30 nts of the ORF. The ß-galactosidase activity of the different transformants was assayed and the results are shown in Figure 1A. Of the 25 sRNAs tested, 4 modulated the expression of the fusion by more than 2-fold, SgrS and RydC positively and MicA and GcvB negatively. Since SgrS is involved in sugar metabolism [27], we suspected that its overproduction could affect expression from the arabinose-induced PBAD promoter. To test this possibility, we measured the SgrS-mediated repression of the same phoP-lacZ fusion when constitutively expressed from the Ptet promoter instead of PBAD. Since the pSgrS plasmid had no effect on this Ptet-phoP-lacZ fusion (Figure 1B), it is likely that it activated the PBAD- driven fusion at the promoter level and this was not investigated further. The RydC sRNA has been shown to activate (repress) the expression of fusions that are negatively (positively) regulated by Hfq, most likely by titrating Hfq [28]. One possibility is therefore that it acts on PBAD-phoP-lacZ in the same way, but further experiments are required for a definitive proof. The same may be true for sRNAs such as ChiX, that also activates the expression of the fusion almost 2-fold.
In the experiment shown in Figure 1A, pMicA repressed the expression of phoP-lacZ by 3.1-fold, which is in agreement with our previous results. Furthermore, this experiment also identified GcvB as a multicopy repressor of phoP-lacZ, since pGcvB was responsible for a 4.5-fold decrease in the ß-galactosidase activity of the fusion. This last result was confirmed using the Ptet-phoP-lacZ fusion (repression of 3.5-fold, Figure 1B), indicating that, as shown previously for MicA [24], GcvB most likely acts at the post-transcriptional level. Importantly, this repressor effect of GcvB was also visible when GcvB was expressed from the chromosome; a deletion of the gcvB gene was sufficient to increase expression of phoP-lacZ by 1.9-fold (Figure 1C).
One possible explanation for these results is that GcvB regulates the phoP-lacZ fusion by controlling the synthesis and/or activity of a post-transcriptional regulator of phoP. Since MicA is so far the only post-transcriptional regulator of phoP known to affect our phoP-lacZ fusion, we analyzed the effect of GcvB on phoP expression in the absence of MicA. In this context, overproduction of MicA and GcvB from a plasmid caused a 3.4- and 4.3-fold decrease respectively in the activity of the phoP-lacZ fusion (Figure 2A), which is similar to what was observed in micA+ cells. Consistent with this observation, deletion of gcvB resulted in a 1.7- or 2.3-fold activation of phoP-lacZ in wt or micA− cells respectively (Figure 2B). Therefore, GcvB acts independently of MicA to regulate phoP expression. In this experiment, deletion of micA has no significant effect on the expression of phoP, because transcription of MicA is dependent on the RpoE sigma factor, which is not activated under the experimental conditions of Figure 2B.
We had previously constructed a mutant form of phoP-lacZ (phoPmut-lacZ, where the 4 nts directly downstream of the AUG start codon are changed from CGCG to GCGC, Figure 3A) such that this fusion is no longer controlled by MicA [24]. Interestingly, this mutant fusion is still controlled by GcvB, since its expression is up-regulated by 2.2-fold in a ΔgcvB strain (Figure 2B, two last bars). This result suggests that the precise regions of the phoP mRNA that are required for MicA or GcvB action are different. Thus, GcvB acts on phoP, independently and apparently at a different site from that of MicA.
GcvB is a pleiotropic regulator, whose expression is highest in exponentially growing cells in rich medium as a result of its control by the GcvA transcriptional regulator. GcvB directly regulates more than 20 genes, the large majority of which are targeted via a very well conserved single stranded G/U rich region of GcvB, referred to as R1 (Figure 3A). Even though this might be partially due to an experimental bias given that the R1 region was used as a “bait” in a bioinformatic search for targets, this region is clearly required for the control of almost all targets identified so far [13]. Since sRNAs often regulate multiple targets via a single conserved region [18], [29], [30], we reasoned that the R1 region was also likely to be involved in the control of phoP, regardless of whether phoP was a direct or indirect target. We therefore measured the expression of the PBAD-phoP-lacZ fusion in the presence of a plasmid overexpressing either GcvB wt or a GcvB mutant in the middle of the R1 region (GcvBmutR1, see Figure 3B). In this experiment, the chromosomal copy of gcvB is deleted and the steady-state level of GcvBmutR1 was slightly lower than that of GcvB wt. Somewhat surprisingly, these two forms of GcvB repressed the expression of phoP-lacZ to a similar extent (Figure 3C, left panel); in contrast, a previously identified target of GcvB R1 region, livJ, was, as expected, less regulated by GcvBmutR1 than by GcvB wt (Figure 3D). While this does not completely rule out a possible role for the R1 region in the control of phoP (for instance, if pairing involves nts of R1 that are not affected by the mutR1 change or if alternative pairing(s) can take place with this mutant), this suggests that the role of R1 in phoP control is not as crucial as for the other targets of GcvB.
Because GcvB action on phoP was independent of MicA (see above), we next envisioned the possibility that it could directly pair with phoP to control its expression. If this interaction exists, we expected it to involve a region of GcvB outside of the R1. The TargetRNA program [31] was used to predict a potential pairing between GcvB and the phoP mRNA fragment encompassing nts −36 to +30 relative to the AUG (i.e. the region of phoP that is present on the phoP-lacZ fusion). According to this prediction (Figure 3B), the region between nts 148 and 174 of GcvB can imperfectly pair with phoPQ mRNA in the translation initiation region (TIR), which is the most frequent binding site for negatively acting sRNAs. Interestingly, this corresponds to a region of GcvB that was shown to be mostly single-stranded in solution [12] and is now referred to as region R3. The relevance of this putative direct interaction was tested in vivo by compensatory changes. While the PBAD-phoP-lacZ fusion was repressed by more than 4-fold upon overproduction of GcvB wt, this was not the case with the GcvBmutR3 variant, where nts 154 to 158 were changed from CUGUC to GACAG. Rather, expression of the fusion was increased by more than 2-fold (Figure 3C). The inability of GcvBmutR3 to repress phoP-lacZ expression is not due to an intrinsic instability, since it accumulates to a level similar to that of wt GcvB (Figure 3C). A possible explanation for the fact that GcvBmutR3 activates phoP-lacZ is that it could titrate Hfq when overexpressed, leading to changes in expression of Hfq-regulated genes, such as phoP (see [28], [32] for examples of competition for Hfq). In contrast, GcvBmutR3, but not wt GcvB, caused an 8-fold decrease in the activity of the compensatory mutant fusion (Figure 3C, right panel), clearly showing that GcvB and phoP mRNA directly interact in vivo. It is also worth noting that MicA efficiently repressed both the wt and the mutant fusion, which confirms that GcvB and MicA pair at different loci of phoPQ mRNA. In addition, when mutations in the R1 and R3 regions of GcvB were combined, the resulting GcvBmutR1R3 repressed the expression of phoPmutR3-lacZ, but not that of phoP-lacZ (Figure S1A). This again indicates that, at least when GcvB is overexpressed, its R1 region is not involved in the control of phoP.
To provide experimental support to the proposed phoP-MicA and phoP-GcvB base-pairing interactions (Figure 3B), a structural probing analysis of phoP mRNA alone or in the presence of either sRNA was performed in vitro using chemical probes (Figure 4, A and B). DMS (dimethyl sulfate), CMCT (1-cyclohexyl-3-(2-morpholinoethyl)carbodiimide metho-p-toluene sulfonate) and kethoxal (1-1-dihydroxy-3-ethoxy-2-butanone) respectively modify unpaired adenosine (and to a much lesser extent cytidine), uridine and guanosine residues. According to our probing data, the secondary structure of the 5′ region of phoP mRNA appears as a long irregular stem-loop which is hold by seven double-stranded elements named H1 to H7, separated by bulges or loops (Figure 4C). Upon addition of MicA, most nts from positions −13 to +11 of phoP mRNA, which include the nts forming the 5′ strands of H4, H5 and H6, display either a decreased reactivity towards the probes or correspond to RT-stops or -pauses which occur in a region rich in GC pairs (Figure 4, A and D). This model is also consistent with the fact that many nts located between positions +27 to +45 of phoP mRNA, which include all the nts forming the 3′ strands of H4, H5 and H6 in the absence of MicA, become more reactive in the presence of MicA (Figure 4, A and D). In conclusion, the interaction between phoP mRNA and MicA relies on (i) the disruption of at least three of these elements, namely H4, H5 and H6, and (ii) the formation of an extended base-pairing interaction between nts −15 to +11 of phoP mRNA and nts 4 to 31 of MicA, whereby both the Shine-Dalgarno (SD) sequence and the phoP translation start codon are base-paired (Figure 4D). Upon addition of MicA, further reactivity enhancements in phoP mRNA nts are observed outside of the proposed phoP mRNA-MicA duplex (see nts +53 and +54 in H1, −21 and +46 to +49 in H3, −16 which joins H3 to the duplex, +14 to +21 in H7 and its apical loop, Figure 4D). It is likely that these changes are due to either local breathing or even disruption of H1, H3 and H7, which are destabilized by the binding of MicA. Also, decreased reactivities are observed (see nts +12 which joins the duplex to H7 and +23 in H7, Figure 4D) for which we have no explanation.
In contrast, the duplex formed by phoP mRNA and GcvB seems shorter as it requires only the disruption of H3 and H4 to form; it is centered around and blocks the SD sequence, which is in complete agreement with the in vivo data. Indeed, nts displaying decreased reactivities towards the probes or corresponding to a region where GcvB-induced RT-stops or -pauses occur are clustered between nts −17 and −11 of phoP mRNA (Figure 4B and 4E). However, the duplex is likely to be subject to breathing as a certain number of nts located on both side of the cluster become more reactive in the presence of GcvB (see nts −21 to −19, −10 and −9, Figure 4E). Additional reactivity enhancements have been mapped outside of the proposed duplex (see nts −5, +2, +3, +5 and +31 to +36 in H5 and H6 and in the bubble located in between, Figure 4E), which are probably due to breathing of H5 which results from its destabilization by the binding of GcvB. Other regions of phoP mRNA located outside of the predicted duplex are subject to increase or decrease in reactivity in the presence of GcvB (see nts located in H6 and H7 and in the bubble in between, Figure 4E). They can be due to some rearrangement of the overall structure of phoP mRNA upon GcvB-binding and/or to a supplementary interaction between phoP mRNA and GcvB. For instance, nts +9 to +17 of phoP mRNA, several of which appear protected upon GcvB-binding, could theoretically pair with nts 89 to 97 of GcvB. While our in vivo data show that this putative supplementary interaction is not sufficient for control, it could nevertheless play a role in stabilizing the phoP mRNA-GcvB duplex or in increasing the kinetics of association. At this stage, its existence and importance remains to be experimentally addressed. Finally, the reactivities of nts +37 to +50, which form the 3′ strands of H3 and H4 in the absence of GcvB, could not be assessed with confidence because of the presence of several RT-stops or –pauses which are also present when phoP mRNA alone is reverse-transcribed in the absence of the probe (data not shown).
Target-mRNAs of negatively acting Hfq-dependent sRNAs are frequently degraded upon sRNA production. Therefore, to confirm the results obtained above by gene fusion, the levels of phoPQ mRNA were analyzed by Northern-Blot upon overexpression of MicA, GcvB or their mutant derivatives, using a chromosomal PBAD-phoPQ construct (Figure 5A). In this experiment, transcription of the phoPQ operon is again expected to start 36 nts upstream of the phoP start codon and has been put under the control of the PBAD promoter for two reasons: (i) to focus only on promoter-independent regulation and (ii) because of the low abundance of the phoPQ mRNA when expressed from its own promoters under the experimental conditions used here. With this construct, several specific bands are visible. The upper band migrates below a 3 kb RNA marker and most likely corresponds to the whole phoPQ mRNA, while the bands of lower molecular weight could result from either alternative transcription or processing events (Figure 5A). MicA, GcvB and GcvBmutR1 induce a decrease in phoPQ mRNA levels, but not MicAmut and GcvBmutR3, that have lost the ability to control phoP expression. This is in complete agreement with the results obtained with the phoP-lacZ fusions. Pairing of MicA and GcvB to the phoPQ mRNA is therefore likely to induce a degradation of this mRNA.
The effect of MicA or GcvB on the steady-state levels of the PhoP protein was then investigated by Western-Blot analysis, in a strain where phoPQ is expressed from its own promoter. As expected, MicA overexpression resulted in a strong decrease in the amount of PhoP (Figure 5B), while overexpression of MicAmut had no noticeable effect. When GcvB was overexpressed, PhoP levels were decreased, albeit to a much lesser extent than upon MicA overproduction. This is rather surprising since pMicA and pGcvB had a similar effect on the expression of phoP-lacZ (Figure 1, Figure 2, and Figure 3) and on the levels of phoPQ mRNA (Figure 5A). Interestingly, pGcvBmutR1, whose effect was also similar to that of pMicA and pGcvB in the previous experiments, is more efficient than pGcvB in down-regulating the levels of PhoP protein. Finally, GcvBmutR3 overproduction had no effect on the levels of PhoP, which is consistent with its inability to repress phoP expression (Figure 5B).
Therefore, control of phoP by GcvBmutR1 or MicA results, as expected, in a clear decrease of the PhoP protein levels. Surprisingly however, this decrease is only modest with wt GcvB, most likely because its R1 region has pleiotropic effects in the cell under the conditions used here, as discussed below.
We then tested whether MicA and GcvB can control the expression of the PhoP regulon by repressing PhoQ-PhoP synthesis. For this purpose, the expression of 4 genes whose transcription is directly activated by PhoP [25] was analyzed under conditions where either MicA or GcvB was overproduced (Figure 6). These 4 genes are ompT, mgtA, yneM and mgrR, that encode an outer membrane protease, a magnesium transporter, a protein of unknown function located in the outer membrane and a sRNA regulator of LPS modification respectively. As expected, MicA induced a >2.5-fold decrease in expression of these 4 targets as analyzed by either translational fusions (for ompT, mgtA) or by transcriptional fusions (for mgrR and yneM) to (Figure 6A). This decrease is most likely due to phoP regulation, since MicAmut, that does not regulate phoP, does not affect expression of these target genes. Similarly, the overproduction of GcvBmutR1, that also represses phoP, led to a ∼2-fold decrease in the expression of the 4 fusions, while overexpression of GcvBmutR3 does not, in agreement with its inability to control phoP. Finally, when the same experiment was carried out in the GcvB overproducing strain, no decrease was observed in the expression of the 4 members of the PhoP regulon that were tested. Instead, activity of mgrR- and ompT-lacZ was unchanged, while activity of mgtA- and yneM-lacZ was increased by 1.6 and 2.1-fold respectively (Figure 6A). These results obtained by gene fusion were confirmed when the levels of ompT or MgrR RNAs were analyzed by Northern-Blot (Figure 6B). Indeed, MicA and GcvBmutR1 induced a decrease in the level of both RNAs, whereas MicAmut, GcvB wt and GcvBmutR3 did not.
Therefore, MicA and GcvBmutR1 repress the PhoP regulon, by controlling expression of phoP. However, wt GcvB does not, which is consistent with the only modest decrease observed in PhoP levels upon its overproduction.
In most cases, negatively acting sRNAs base-pair with their target-mRNAs in the TIR and occlude the RBS, thereby preventing ribosome binding and translation initiation. This is frequently accompanied by a degradation of the target-mRNA, possibly as a consequence of translational block, or in a process that is directly induced by the sRNA pairing to the target-mRNA. Since MicA and GcvB both pair to phoPQ mRNA in the TIR of the first cistron and decrease phoPQ mRNA levels, toeprinting experiments were performed in order to determine whether they also inhibited ribosome binding. In these experiments, addition of 30S ribosomal subunit and initiator tRNA to a ∼200-nt phoP mRNA fragment transcribed in vitro induced an arrest of reverse-transcription, that was visible on a sequencing gel as a band at position +16, a classical toeprint position (Figure 7, A and B, lanes 3 and 2 respectively). When increasing concentrations of MicA were incubated with phoP mRNA prior to the addition of 30S and fMet-tRNA, the intensity of this band progressively decreased (Figure 7A, lanes 4–6). This was not observed when equal amounts of MicAmut were added instead (lanes 8–10), suggesting that it is the pairing of MicA to phoP TIR that inhibits ribosome binding. Similar results were observed with GcvB, whose addition inhibited the appearance of the +16 toeprint, even more efficiently than MicA (Figure 7B, lanes 7–8). Again, this is most likely due to the pairing between GcvB and phoP since GcvBmutR1, that should still pair with phoP, also inhibited the toeprint, while GcvBmutR3 and GcvBmutR1R3, that should not bind to phoP, inhibited the toeprint much less efficiently (Figure 7B, lanes 11–12, and Figure S1B). It is interesting that GcvB is much more efficient than MicA in inhibiting toeprint, and that GcvBmutR3 still inhibits toeprint to some extent. This might be related to the existence of a bipartite interaction between phoP and GcvB (see above).
Furthermore, as already observed in probing experiments (Figure 4), addition of MicA, but not MicAmut, to phoP mRNA in the absence of 30S subunits also induced stops or pauses of reverse-transcription, as indicated by the bands at positions +6 to +8 (Figure 7A, lanes 2, 4, 5 and 6). This corresponds to the 3′ end of the duplex between MicA and phoP (Figures 3A, 4D, [24]), indicating that this duplex is stable enough to induce pauses in reverse-transcription. In contrast, while GcvB pairs with phoP in vitro, given the results of the probing experiments and the toeprint inhibition, its binding does not induce pauses or stops of the reverse transcription that are sufficiently strong to be observed in this experiment. This is in contrast to what was observed in Figure 4 and this discrepancy is most likely due to the use of different experimental conditions in the probing and toeprint experiments. In fact, under conditions where the signal is highly amplified, reverse transcriptase stops are visible in the toeprint experiments (data not shown). Hfq protein was not included in these in vitro assays, because of the risk of non-specific interactions with RNA. The fact that, even in the absence of this chaperone, MicA and GcvB could both pair to phoP mRNA in vitro (i.e. in the absence of RNases) suggests that the requirement for Hfq in vivo is, at least in part, explained by its ability to protect MicA and GcvB from degradation.
In summary, both MicA and GcvB inhibit ribosome binding by pairing to phoPQ TIR. This translational block could be the step leading to the degradation of the target-mRNA in presence of the regulatory sRNAs observed in vivo.
In this study, we identify phoPQ mRNA as a new target of the E. coli GcvB sRNA. After MicA, this is thus the second sRNA regulator of this operon. Similar to MicA, GcvB directly controls phoPQ expression by pairing to the TIR of phoP, although at sequences slightly different from those of MicA. These pairings cause a steric inhibition of ribosome binding as seen by toeprint experiments and, possibly as a consequence of this translational control, induce degradation of the phoPQ mRNA. Furthermore, this work identifies a novel pairing region of GcvB, namely R3, essential for phoP control. Interestingly, this region was predicted in a computational approach as a potential target-binding region of GcvB (together with the R1) on the basis of its conservation and accessibility [33], and was proposed to participate in the control of cycA expression [34]. Whether GcvB controls yet additional genes through its R3 region remains to be investigated.
This region R3 is with R1 and R2 one of the most conserved in GcvB among enterobacteria (Figure S2A). However, phoP was not identified as a GcvB target in Salmonella in a recent study combining microarray analysis following GcvB pulse-expression and bioinformatic prediction based on complementarity to the R1 region [13]. While this could be due to the low abundance of phoPQ mRNA and to the fact that this control does not rely on R1, this could also indicate that the control of phoP by GcvB that exists in E. coli is not conserved in S. typhimurium. Consistent with this, predictions of pairing between GcvB R3 region and the TIR of phoP mRNA in Salmonella identified only 4 consecutive complementary nts at the most, which is probably too short to ensure specific binding. Furthermore, a preliminary analysis of potential interactions between GcvB R3 and the phoP TIR in different families of enterobacteriaceae suggests that GcvB could control phoP in species such as Klebsiella pneumoniae, Photorhabdus luminescens, Proteus mirabilis, Serratia proteamaculans, Shigella flexneri, Xenorhabdus bovienii and Rahnella (Figure S2B).
At first glance, it is quite surprising that, even though MicA, GcvB and GcvBmutR1 similarly repress phoP expression when followed by gene fusion to lacZ or mRNA levels, their effects on the PhoP protein are quite different. Indeed, while MicA and GcvBmutR1 strongly decreased the cellular level of PhoP, as expected, the effect of wt GcvB was much more moderate (Figure 5B). One noticeable difference in those experiments is that when the PhoP protein levels were assessed, phoPQ was expressed from its own promoter. In contrast, in both the experiments with gene fusion or mRNA levels, its transcription is driven by an heterologous promoter (PBAD), with phoPQ and phoP-lacZ mRNAs expected to originate at the same transcription start site than from P1. Therefore, one could hypothesize that wt GcvB would activate PhoP synthesis from transcripts originating from promoters P2 or P3 (upstream of P1), or by acting on phoPQ transcription, in addition to repress its expression post-transcriptionally. These possibilities were experimentally ruled out because (i) wt GcvB similarly repressed expression of a phoP-lacZ fusion whose 5′end is identical to the 5′end initiating from P1 or from P2 promoter (Figure S3A) and (ii) wt GcvB only poorly affects ompT expression (and even increases yneM expression) in a strain where all the phoPQ promoters have been replaced by PBAD (Figure S3B).
One could also envision that, when expressed from its own promoter under non-inducing conditions (as in Figure 5B), phoP expression is poorly affected by wt GcvB, whose R1 region can pair with other competing targets. This competition could not take place with GcvBmutR1, where the R1 region is mutated, in agreement with its ability to control phoP in all experiments. In contrast, when phoP expression increases, for instance because its transcription is driven by an induced PBAD promoter, it would now become available for repression by wt GcvB, because it would outcompete other GcvB targets, hence the stronger effect of wt GcvB on phoP in Figures 3C and 5A for instance. It will now be interesting to study how the induction of phoP from its own promoter and the expression of GcvB R1 targets will impact the control of phoP by GcvB; in other words, whether this model is physiologically relevant.
Yet another possibility to explain the difference between phoP-lacZ expression, phoPQ mRNA and the PhoP protein levels in presence of pGcvB is that, in addition to its negative effect on phoP expression at the translation initiation step, wt GcvB could stabilize the PhoP protein. Such a regulatory event would affect only PhoP levels, but not the activity of phoP-lacZ or phoP mRNA levels. This putative stabilization would be dependent on the R1 region of GcvB and could be mediated by one or several of its targets. Furthermore, one could wonder whether this stabilization is related to the phosphorylation status of PhoP protein. Because these R1 targets that could directly or indirectly control PhoP are likely to be multiple, it might be difficult to identify them by a genetic approach. Again, it is tempting to speculate that the expression of these targets, as well as the availability of the R1 region to regulate them (dependent on the expression of all GcvB targets and their relative affinity for the sRNA), will play an important role in the control of the PhoP regulon by GcvB. While under our experimental conditions, the amount of PhoP protein is not strongly affected by wt GcvB, there might be conditions where it could be. This would provide a mechanism to establish a hierarchy among the different GcvB-targets and to monitor the regulatory outcomes of GcvB-mediated controls.
In a previous work, we showed that MicA, which is induced by envelope stress, repressed the expression of phoPQ. This finding related the activity of this operon to the cell envelope status [24]. Our present findings show that GcvB relates amino acid (or peptide) uptake and metabolism to the expression of phoPQ. The induction of GcvB occurs in the presence of two different amino acids with two different mechanisms. First through the GcvA/GcvR repressing complex that is inactivated in the presence of glycine, and second through the global regulator Lrp, whose repression is alleviated in the presence of leucine [9], [11] (Figure 8). Although the raison d'être of such a connection between amino acid uptake/metabolism and PhoQ/PhoP regulon activity is not obvious, such a relationship has already been observed with several targets of this regulon. For instance, expression of mgtA and mgtCBR genes is derepressed under conditions of proline limitation due to the presence of a proline-rich open reading frame in their leader mRNAs [35], [36]. Another example is the proline transporter encoded by the proP gene that is also regulated by PhoQ/PhoP [37].
The complexity of the relationship between GcvB and the phoPQ regulon is highlighted by two experimental data. The first is the surprising way wt GcvB fails to strongly decrease PhoP levels as discussed above. The second is related to the recent results of a deep-sequencing study indicating that PhoP positively regulates GcvB levels in the cell [38](Figure 8). However, GcvB levels were unmodified when the magnesium concentration in the growth medium was varied. Therefore, GcvB could also modulate the degree of PhoQ/PhoP activation depending on the inducing signals.
Interestingly, there are already many examples of connections between sRNAs and TCS, as several sRNAs were previously shown to control TCS and conversely [39]. In addition, the negative feedback loop that exists between phoP and GcvB is reminiscent of other feedback loops involving sRNAs [40]. As in most cases however, the properties and possible advantages of this feedback loop in bacterial physiology remain to be experimentally addressed.
Our search for Hfq-dependent sRNAs regulators of the PhoQ/PhoP TCS was initially motivated by the fact that phoP expression was up-regulated in an hfq mutant strain independently of MicA. Interestingly however, even though GcvB is partially responsible for this Hfq-effect, expression of phoP is still higher in an hfq mutant than in hfq+ cells in the absence of both MicA and GcvB (data not shown). This suggests that there might be even more sRNAs controlling PhoQ/PhoP, which would allow the integration of yet additional signals to fine-tune expression of this central TCS.
Strains and plasmids used in this study are listed in Table 1, and sequences of the oligonucleotides in Table S1. Strains were grown aerobically in LB medium at 37°C. When needed, antibiotics were used at the following concentrations: ampicillin 150 µg/ml, tetracyclin 10 µg/ml, kanamycin 25 µg/ml or chloramphenicol 10 µg/ml. PCR amplification was performed using the Phusion DNA polymerase (New England Biolabs). IPTG (isopropyl-ß-D-thiogalactopyranoside) was used at a final concentration of 100 µM.
Replacement of gcvB gene by a kanamycin or tetracyclin resistance cassette was engineered by recombineering of a cassette amplified by PCR (with ΔgcvB::kanfor and rev, or ΔgcvB::tetfor and rev oligonucleotides) and flanked by homology regions upstream and downstream of gcvB into a strain carrying a mini-lambda allowing recombineering upon induction, such as NM300 or NM1200 for instance. These mutant alleles, as well as ΔmicA::tet [24] or phoP::kan [41] were then moved by P1 transduction when necessary.
Strains carrying gene fusions to lacZ were either obtained from different sources or constructed in this study by recombineering into strain PM1205 [19] or strain MG1508. In strain MG1508, a cat-sacB cassette following the PLtetO-1 promoter [42] is placed upstream of the lacZ gene in an MG1655 derivative that carries a mini-lambda.
Strain MG1173 carries a translational ompT-lacZ fusion, whose construction was described previously, at the lambda attachment site [18]. Strains KM112 and KM194 were also described elsewhere [43]. They contain respectively the promoter region up to nt+10 of MgrR, or the promoter region of yneM up to nt+91 (nt+1 is the transcriptional start site) upstream of lacZ chromosomal gene, starting 17 nts upstream of its ATG start codon. The translational Ptet-phoP-lacZ fusion was constructed by replacing the cat-sacB cassette of the Ptet-cat-sacB-lacZ construct in strain MG1508 with a PCR fragment encompassing nts −36 to +30 (relative to the ATG start codon) of phoP between homology regions to Ptet and lacZ respectively. This PCR fragment was generated with primers 5′Ptet-phoP and 3′phoP-lacZ. Recombinants were selected on LB-agar plates without NaCl supplemented with 6% sucrose, and further verified as in [24]. Similarly, construction of the P1-mgtA-lacZ fusion was done by recombineering of a PCR fragment carrying nts −330 to +30 of mgtA in MG1508, except that the homology regions were upstream of Ptet and within lacZ (see primers 5′P1mgtA and 3′mgtA-lacZ).
For strains MG1510 (PBAD-phoPmutR3-lacZ) and AC0067 (PBAD-livJ-lacZ), PCR fragments were generated with primers phoPmutR3for and 3′phoP-lacZ, or 5′LivJ-lac and 3′LivJ-lac respectively, then recombined into strain PM1205.
Construction of strains where the phoPQ operon, wt or interrupted by a kanamycin resistance cassette, is expressed from a PBAD promoter was as follows. First, the chloramphenicol resistance cassette followed by the PBAD promoter was amplified by PCR from plasmid pTM26 [44] with primers 5′Cm-PBAD-phoPQ and 3′Cm-PBAD-phoPQ. This product was then recombined in strain NM300 (or a derivative carrying the phoP::kan allele moved by P1 transduction using AB043 [41] as the donor strain). After selection on LB-chloramphenicol plates, the PBAD promoter and beginning of phoP gene were checked by sequencing, and resistance to kanamycin was verified for the PBAD-phoP::kan construct. These alleles were then moved by P1 transduction using DJ624 as the recipient strain and selecting for chloramphenicol resistant clones to create strains MG1516 and MG1517. MG1517 was checked for resistance to kanamycin.
Overnight cultures were diluted 500 fold in fresh medium (see below for exact medium composition) and grown to mid-exponential phase (OD at 600 nm∼0.4). The ß-galactosidase activity was then measured as in [45] and expressed in Miller units. Alternatively, the activity was measured in a 96-wells plate for some experiments. In this case, 100 µl of cells were mixed with 50 µl of permeabilization buffer containing 200 µg/ml of polymixin B [46]. After addition of 50 µl of ONPG, the absorbance at 420 nm was followed over-time and the ß-galactosidase activity was calculated as the slope of the resulting curve. It is expressed in arbitrary units.
Cells were grown in the following media: for Figure 1A, LB-Ampicillin-IPTG-Arabinose 0.002% or 0.02%; for Figure 1B, LB-Ampicillin-IPTG; for Figure 1C, LB; for Figure 2A and 3C, LB-Ampicillin-IPTG-Arabinose 0.02%; for Figure 2B, LB-Arabinose 0.002%; for Figure 3B, LB-Ampicillin-IPTG-Arabinose 0.02%, but IPTG was not included in the overnight cultures; for Figure 6A, LB-Tetracyclin-IPTG; for Figure S1, LB-Tetracyclin-IPTG-Arabinose 0.02%.
Values of ß-galactosidase activity given in the paper are the average of at least two independent experiments and are listed in Table S2.
RNA was extracted following the hot phenol method as previously described [47] and using 650 µl of cells. For experiment of Figure 3, RNA was extracted at the same time than samples were taken to measure ß-galactosidase activity. For experiments of Figures 5 and 6, cells were grown overnight in LB-Tetracyclin-IPTG-Arabinose (0.2% for Figure 5 and 0.02% for Figure 6), then diluted in fresh medium and grown to mid-exponential phase (A600∼0.4 for Figure 6B) or to stationary phase (A600∼2 for Figure 5) before RNA was extracted. For Northern analysis, a constant amount of RNA was separated on an 8% acrylamide TBE-urea gel (for MicA, GcvB, MgrR and SsrA RNAs) or on a 1% denaturing agarose gel (for ompT and phoP mRNAs) and transferred to an Hybond-N+ membrane. Detection was then performed using biotinylated probes and the Ambion brightstar detection kit following manufacturer's instructions.
Overnight cultures in LB-Tetracyclin-IPTG of strain MG1173 (wt) or MG1446 (phoP−), transformed with pBRplac and derivatives, were diluted 500-fold in the same medium and grown to mid-exponential phase. Cells were then pelleted and resuspended in SDS-sample buffer with DTT (New England Biolabs) at a final concentration of 15 OD600/ml. These samples were then boiled for 5 minutes and 15 µl were loaded on a 15% SDS-PAGE gel. Proteins were then transferred to an Hybond-C super membrane (Amersham) and the PhoP protein was detected using a 1∶1000 dilution of an anti-PhoP antiserum (from Mark Goulian) and the Immun-Star WesternC Chemiluminescent Kit (Biorad). EF-Tu was immunodetected from the same membrane. A representative blot from three independent experiments is shown.
PCR templates for the in vitro transcription of MicA or GcvB (and their derivatives) were prepared from the pBRplacMicA(mut) or pBRplacGcvB(mutR1, mutR3 or mutR1R3) plasmids, using the oligonucleotides 5′T7MicA(mut) and 3′T7MicA, or 5′T7GcvB and 3′T7GcvB respectively. Note that one or two G residues are added at the 5′end of MicA and GcvB respectively. For phoP, a PCR fragment corresponding to nts −36 to +169 of phoP mRNA preceded by two G residues was amplified from genomic DNA using oligonucleotides 5′T7phoP and 3′T7phoP. After purification, these PCR products were used in in vitro transcriptions reactions with the T7 RNA polymerase of Stratagene (for phoP) or the T7 Megascript kit from Ambion (for MicA and GcvB) following manufacturer's instructions. After phenol extraction and precipitation with ammonium acetate, RNA were purified using G-50 Microspin columns (GE Healthcare).
Toeprinting assays were adapted from Hartz et al. [48] as follows. 0.5 pmol. of phoP transcript were incubated with 2 pmol. of phoP-Cy5-probe#1, an oligonucleotide complementary to nts 52 to 71 of phoP ORF and labeled with a Cy5 group at its 5′ end, in a buffer containing 10 mM Tris-acetate pH 7.4, 60 mM ammonium chloride and 6 mM ß-mercaptoethanol. When required, sRNAs were added to this mix at the desired concentrations. These mixtures were denatured by heating at 80°C for 3 minutes, followed by a rapid cooling in an ethanol/solid CO2 mix. They were then thawed on ice and magnesium was added at a final concentration of 10 mM. For experiment of Figure 7A, an additional incubation step at 37°C for 10 minutes was performed at this stage. 190 µM dNTPs and 2.5 µM of initiator tRNAfMet were added together with 0.5 µM 30S subunits, and the mixtures were incubated at 37°C for 10 minutes. 1 unit of AMV RT (Finnzymes) was then added and cDNA synthesis was performed at 37°C for 20 minutes. Reactions were stopped by addition of formamide and EDTA, analyzed on a 6% sequencing gel together with sequencing reactions and RT stops were visualized using a typhoon fluorescent scanner set up for Cy5 detection.
1.25 pmol. of phoP transcript, mixed in water with 6.25 pmol. of sRNA when required, were denatured as above. After a slow thaw-out on ice, samples were incubated at 37°C for 10 minutes in order to allow sRNA and phoP message to pair. Samples were then diluted in a buffer containing 50 mM sodium cacodylate pH 7.5, 10 mM magnesium acetate and 50 mM ammonium chloride prior to DMS treatment, or in the same buffer except that sodium cacodylate is replaced by sodium borate pH 8.0 prior to CMCT or Kethoxal treatment. After a 10 minutes incubation at 25°C, 1 µg of L. lactis 23S rRNA was added, followed by 0.1 volume of DMS or kethoxal (stock solutions are 1/30 in ethanol or at 4 mg/ml in 20% ethanol respectively) and samples were incubated at 25°C for 5 minutes. For CMCT treatment, 0.1 volume of 100 mg/ml CMCT in the previous buffer containing sodium cacodylate was added and samples were incubated at 25°C for 10 minutes. Modified RNA were then precipitated with ammonium acetate (and 100 mM sodium borate pH 8.0 for samples treated with Kethoxal) and resuspended in water (or in 12.5 mM sodium borate pH 8.0 for samples treated with Kethoxal).
For reverse transcription, phoP-Cy5-probe#2, complementary to nts 87 to 106 of phoP ORF, was added to those samples at the final concentration of 1 µM. This was followed by the addition of 2 units of AMV RT (Finnzymes), together with 1 mM dNTPs and 4 mM DTT. cDNA synthesis and analysis was then performed as described above.
|
10.1371/journal.pgen.0040019 | The Genetic Structure of Pacific Islanders | Human genetic diversity in the Pacific has not been adequately sampled, particularly in Melanesia. As a result, population relationships there have been open to debate. A genome scan of autosomal markers (687 microsatellites and 203 insertions/deletions) on 952 individuals from 41 Pacific populations now provides the basis for understanding the remarkable nature of Melanesian variation, and for a more accurate comparison of these Pacific populations with previously studied groups from other regions. It also shows how textured human population variation can be in particular circumstances. Genetic diversity within individual Pacific populations is shown to be very low, while differentiation among Melanesian groups is high. Melanesian differentiation varies not only between islands, but also by island size and topographical complexity. The greatest distinctions are among the isolated groups in large island interiors, which are also the most internally homogeneous. The pattern loosely tracks language distinctions. Papuan-speaking groups are the most differentiated, and Austronesian or Oceanic-speaking groups, which tend to live along the coastlines, are more intermixed. A small “Austronesian” genetic signature (always <20%) was detected in less than half the Melanesian groups that speak Austronesian languages, and is entirely lacking in Papuan-speaking groups. Although the Polynesians are also distinctive, they tend to cluster with Micronesians, Taiwan Aborigines, and East Asians, and not Melanesians. These findings contribute to a resolution to the debates over Polynesian origins and their past interactions with Melanesians. With regard to genetics, the earlier studies had heavily relied on the evidence from single locus mitochondrial DNA or Y chromosome variation. Neither of these provided an unequivocal signal of phylogenetic relations or population intermixture proportions in the Pacific. Our analysis indicates the ancestors of Polynesians moved through Melanesia relatively rapidly and only intermixed to a very modest degree with the indigenous populations there.
| The origins and current genetic relationships of Pacific Islanders have been the subjects of interest and controversy for many decades. By analyzing the variation of a large number (687) of genetic markers in almost 1,000 individuals from 41 Pacific populations, and comparing these with East Asians and others, we contribute to the clarification and resolution of many of these issues. To judge by the populations in our survey, we find that Polynesians and Micronesians have almost no genetic relation to Melanesians, but instead are strongly related to East Asians, and particularly Taiwan Aborigines. A minority of Island Melanesian populations have indications of a small shared genetic ancestry with Polynesians and Micronesians (the ones that have this tie all speak related Austronesian languages). Inland groups who speak Papuan languages are particularly divergent and internally homogeneous. The genetic divergence among Island Melanesian populations, which is neatly organized by island, island size/topography, as well as their coastal or inland locations, is remarkable for such a small region, and enlarges our understanding of the texture of contemporary human variation.
| The populations in New Guinea and the islands immediately to the east (the Bismarck and Solomons archipelagos) are well-known for their great diversity in cultures, languages, and genetics, which by a number of measures is unsurpassed for a region of this size [1]. This area is referred to as Near Oceania, as opposed to the islands farther out in the Pacific, known as Remote Oceania [2] (see Figure 1). For simplicity, we refer only to the peoples of Near Oceania as “Melanesians,” although this term ordinarily encompasses additional groups to the east as far as Fiji, who are not covered in this study. Major parts of Near Oceania were settled from Southeast Asia early in modern human prehistory, between ∼50,000 and ∼30,000 years before present (YBP) [3–5]. Populations were relatively isolated at this edge of the human species range for the following 25,000 years. The early settlers in Near Oceania were very small groups of hunter-gatherers. For example, New Ireland, which is more than 300 km long, is estimated to have had a pre-Neolithic carrying capacity of ∼1,200 people or fewer [6]. There is evidence of sporadic, modest contact between New Guinea and the Bismarcks from 22,000 YBP, and with Bougainville/Buka in the Solomons only from ∼3,300 years ago [3,7].
By ∼3,300 YBP [3], at least one powerful new impulse of influence had come from Austronesian speaking migrants from Island Southeast Asia, likely associated with the development of effective sailing [8], that led to the appearance of the Lapita Cultural Complex in the Bismarck Archipelago. After only a few hundred years, “Lapita People” from this area had colonized the islands in Remote Oceania as far east as Tonga and Samoa, where Polynesian culture then developed [9].
The distribution and relations of Pacific language families reflect ancient settlement. Austronesian is a widespread and clearly defined linguistic family with more than 1,000 member languages, which has its greatest diversity, and likely origin, in Taiwan ∼4,000–5,000 years ago [10]. Some basic phylogenetic relations within Austronesian are sketched in Figure S1. All Austronesian languages spoken outside Taiwan belong to the Malayo-Polynesian branch, and almost all the Malayo-Polynesian languages of Oceania belong to the Oceanic branch. It is Proto Oceanic, the immediate ancestor of the Oceanic languages, that is associated with an early phase of the Lapita Cultural Complex. Proto Oceanic split into a number of branches as its descendants spread across Remote Oceania, including Proto Nuclear Micronesian and Proto Polynesian (a branch of Central Oceanic).
Almost all the other indigenous languages of Oceania are referred to as non-Austronesian, or Papuan. Most Papuan languages are found in New Guinea, with the remainder in nearby islands. This is a residual category of ∼800 languages. Most of these can be assigned to more than 20 different language families, but these families cannot be shown to be related on present evidence. There remain a number of “Papuan” isolates that cannot be grouped at all [11]. Trans New Guinea is the largest Papuan language family. It consists of ∼400 languages and dates to 6,000 to 10,000 YBP [12]. Other Papuan families including the ones in the Bismarck and Solomon archipelagos probably also go back at least to this period [13–15]. While it is reasonable to assume these different Papuan families had common origins further back in time, any evidence of such ties that is recoverable with standard methods of historical linguistics has been erased over the millennia. The concentration and number of these apparently unrelated language families and isolates is unsurpassed in any other region of the world [15].
Analyses of genetic variation at some informative loci, particularly the mitochondrial DNA (mtDNA) (reviewed in [16,17–19]), non-recombining Y-chromosome markers (NRY) (reviewed in [19,20]), and a small set of autosomal microsatellites [21] have provided divergent impressions of the population genetic structure of both Near and Remote Oceania. Because they have ¼ the effective sample size of autosomal markers, the mtDNA and NRY haplotypes have been particularly subject to the effects of random genetic drift, and each autosomal marker, no matter how informative, still represents a minute fraction of the total genetic variation among populations. Even so, these data have shown that the genetic variation in Near Oceanic populations is considerably greater than in Remote Oceanic ones, and that there are a cluster of haplogroups that developed in particular islands of Near Oceania between approximately 50,000 and 30,000 years ago.
However, a number of unresolved issues remain concerning the proper interpretation of these and other data that a comprehensive genomic sampling of neutral biparental markers across Pacific populations should clarify. A list of these includes: 1) to whom are these diverse Melanesian populations most closely related outside this region (East or South Asians, or perhaps even Africans, whom they physically resemble)? 2) how does the genetic diversity and differentiation of Near Oceanic populations compare with those in other regions? 3) is there a clear organization of the variation among groups in Near Oceania (i.e., either by language, by island, or distance from major dispersal centers)? 4) is there a genetic signature of Aboriginal Taiwanese/Southeast Asian or Polynesian influence in Melanesian populations, especially in the Bismarcks, where the Lapita Cultural Complex developed? and 5) are Polynesians more closely related to Asian/Aboriginal Taiwanese populations or to Melanesians?
Here we report the analysis of 687 microsatellite and 203 insertion/deletion (indel) polymorphisms in 952 individuals from 41 Pacific populations, primarily in the Bismarck Archipelago and Bougainville Island, and also including select sample sets from New Guinea, Aboriginal Taiwan, Micronesia, and Polynesia. The results show the reduced internal variation of Near Oceanic Melanesian populations and the remarkable divergence among them, and how this divergence is influenced by island size and topography, and is also correlated with language affiliation. We also detected a very small but clear genetic signature of “Asian/Polynesian” intermixture in certain Austronesian (Oceanic)-speaking populations in the region (by “genetic signature,” we mean an ancestral proportion in some groups inferred by the STRUCTURE analysis that predominates in another ancestral grouping). For global context, these data were compared with data from the Centre d'Etude du Polymorphisme Humain human genome diversity panel (HGDP-CEPH), composed of cell lines [22–24], especially its subset from East Asia. Figure 1A shows how undersampled the Pacific populations had been in the HGDP-CEPH dataset (as well as its emphasis on particular regions of Asia), and Figure 1B shows the distribution of our Pacific population samples, with its intensive coverage in Near Oceania.
Our sampling strategy concentrated on Papuan-speaking populations and their immediate Oceanic-speaking neighbors from the islands immediately to the east of New Guinea, in what is called Northern Island Melanesia, consisting of the Bismarck and Solomon Archipelagos (see Figure 1B). The three largest islands of the region were most intensively sampled—New Britain, New Ireland, and Bougainville—along with two nearby smaller islands (New Hanover and Mussau). Additional Pacific samples came from New Guinea (one set from the lowland Sepik region and one set from the Eastern Highlands), Micronesia (primarily from Belau), Polynesia (Samoans and one New Zealand Mãori group), and aboriginal Taiwan (Amis and the Taroko, a mountain Atayal group). The details of the sample locations and language family affiliations are given in Table S1 and in the Methods section.
Figure 2 shows the estimated values of θ (θ̂) calculated from expected heterozygosity (He) arranged from highest to lowest values, combining our Pacific populations and the HGDP-CEPH global set (the values of θ̂, He, and the average number of alleles per locus are given in Table S1). From Ohta and Kimura [25], under a stepwise model, the expected relationship between θ and heterozygosity (H) is
which rearranges to
For autosomal loci, θ is defined as θ = 4Neμ, where Ne is defined as the effective population size and μ is the per generation mutation rate. Assuming the mutation rate is constant across populations and that the stepwise mutation model is appropriate, θ̂ provides an estimate that is linearly correlated with effective population size. In contrast, H asymptotically approaches a value of 1 as the effective population size increases. Therefore, the use of θ̂ is more appropriate to represent differences in effective population sizes among populations (e.g., a θ ratio of 2 between two populations indicates twice the effective population size between the populations, while an H ratio of 2 does not).
The pattern of variation in Figure 2 is consistent with a series of successive founder effects that modern humans underwent in their expansions out of Africa (also shown by [26]). African populations have the highest values, followed in order by Europe/Central Asians, East Asians, Melanesians, and Native Americans. All the Pacific populations ranked together in a narrow band towards the low end of θ̂ values (between 4.8 and 2.9). Within the Melanesian set, inland populations generally had lower values of θ̂ than shore-dwelling groups, as shown. The three non-Pacific groups in the range between 4.8 and 2.9 were the Maya, Columbia, and Lahu. The Maya are known to have some European ancestry, which would explain their relatively high θ̂ for a Native American group; and the Lahu are an Asian population that was subject to particularly strong random genetic drift [24]. Columbia and other conglomerate groups made up of individuals from different populations (e.g., Bantu South, Sepik, Highlands, Micronesia, and Samoa) consistently had higher values of θ̂ than related groups. This combining of groups has caused inflated levels of diversity and effective population size estimates (i.e., there is more variation in a combined sample set than is typically contained in one from a clearly defined population).
Ramachandran et al. [26] investigated the correlation between geographic distance and genetic differentiation as measured by pairwise FST in the global HGDP-CEPH dataset, and found a linear relationship existed, with major deviations from the fitted line they believed consistent with admixture or extreme isolation. We analyzed this correlation by major region, adding our expanded Pacific dataset. The results, shown in Figure 3, show the extremely heterogeneous nature of the linear correlations and distributions from region to region. The sampled Melanesian populations were distributed across a comparatively small geographic area, but their range of pairwise FST values was extremely large. Only the Native American groups had an equivalent range of FST values, but these were unreliable since there were only five American populations distributed across very large distances.
To quantify the degree of variation within and among populations, an analysis of molecular variance (AMOVA) for the Pacific materials plus the HGDP-CEPH dataset was performed, with the results shown in Table 1. The global AMOVA results first presented in [24] for the HGDP-CEPH dataset were based on 377 microsatellites, included some first degree relatives, and included only two “Oceanic” populations (from the Nasioi of Bougainville and highland New Guinea). In the current analysis based on 687 microsatellites, the Americas had the highest among-population variation component, followed in order by Melanesia, Africa, Asia, and Europe. This pattern follows directly from their ranking in population heterozygosities or θ̂ [27].
As shown in Table 2, the microsatellite variation in Melanesia (New Guinea, New Britain, New Ireland, and Bougainville) was apportioned first by language group and then by island. While population variation among the different islands was considerable (refer to the 95% confidence interval), within-island variation among populations was more than three times greater. This was primarily due to the extensive variation within New Britain (with a 5% internal variance component), followed by Bougainville (3.7%), and New Ireland (2%, see Table S2). The variation among the three New Guinea samples in our series was lower, most likely because of their less rigorous population definitions (see the Methods section for sampling details).
Apportioning the molecular variance by language group (between Oceanic speaking and Papuan speaking populations) only accounted for 0.2 % of the total, which, as indicated by the very small 95% confidence interval, was still significant. Since the two language categories are scattered across the islands, geography and intermixture will confound possible language effects. While the microsatellite variation among the Oceanic-speaking populations was significant, it was much greater among the Papuan-speaking populations (many of which are located in the mountainous interiors of the larger islands).
To investigate individual and population similarities, we applied a Bayesian model-based clustering algorithm implemented in the STRUCTURE program [28] to our Pacific dataset combined with the HGDP-CEPH panel (also genotyped by the Marshfield Clinic). This program identifies groups of individuals who have similar allele frequency profiles. The great advantage of this clustering approach is that it avoids a priori population classifications, and instead estimates the shared population ancestry of individuals based solely on their genotypes under an assumption of Hardy-Weinberg equilibrium and linkage equilibrium in ancestral populations. It infers individual proportions of ancestry from K clusters, where K is specified in advance and corresponds to the number of posited ancestral populations; K can be varied across independent runs. Individuals can be assigned admixture estimates from multiple ancestral populations, with the admixture estimates summing to 1 across these population clusters.
Figure 4 presents the STRUCTURE analysis of our Pacific dataset plus the HGDP-CEPH Panel for 687 microsatellites and 203 indels on the 22 autosomes, on a total of 1,893 individuals from 91 populations. Each increase in K split a cluster that had been defined in an earlier run, and individuals from the same populations had very similar membership coefficients in the inferred clusters. Details of the STRUCTURE results are provided in the Table S3. Inclusion of our large Pacific dataset altered the sequence of splitting, but did not change, the five major global clusters that had previously identified with a smaller set of microsatellites: Sub-Saharan Africa, Western Eurasia, East Asia, “Oceania,” and the Americas [24]. The Taiwan Aborigines clustered with East Asia, while Polynesians and Micronesians had a mixed position between East Asians and Melanesians (“Oceania”). The Mãori had the suggestion of a minor proportion of European admixture, which had been indicated by the donors themselves.
There was a small but consistent “Asian/Polynesian” admixture estimate in specific Melanesian groups. Because clustering after K = 6 mostly involved Near Oceanic populations, we stopped the combined global analysis there, and analyzed the Pacific subset separately thereafter.
An unrooted neighbor-joining tree for the same HGDP-CEPH and Pacific samples, excluding the indels, was calculated from a matrix of pairwise FST “coancestry” distances (similar to Reynolds' D [29], see Table S4), and is shown in Figure 5. For comparison, the cluster colors for the K = 6 STRUCTURE run were superimposed on the tree. The results were compatible with the clusters identified with STRUCTURE. Branch lengths varied inversely with values of θ̂ or expected heterozygosity, so that populations with the longest branch lengths had the lowest values of θ̂. The longest branches belonged to the Native American and separate Melanesian groups. As with the STRUCTURE results, this unrooted FST based tree had Melanesians, East Asians, and Native Americans at the opposite end of the human tree from Africans and Europeans. Trees based on other population pairwise genetic distance matrices (Nei's chord distance [30], (δμ)2 [31], the proportion of shared alleles [32], and Cavalli-Sforza and Edwards' chord distance [33]) also indicated relatively large distances between Africans and Melanesians, and also consistently placed the Taiwan Aborigines between the East Asians and Polynesians/Micronesians (Figure S2).
We performed STRUCTURE analyses on a combined East Asia–Pacific dataset to explore in detail the relationships among Melanesians, Polynesians, Taiwan Aborigines, and East Asians, and to clarify the role of intermixture there. The samples included in this analysis were our Pacific set of 40 groups, and from the HGDP-CEPH panel, the “Papuans,” (identified here as “Highlands”), the East Asians, and French (the French were included to identify European admixture). The STRUCTURE results are shown in Figure 6, and the details on their reproducibility in Table S5. At K = 2 and K = 3, the Asia-Pacific clusterings mirrored the first five runs of the global comparison. Bougainville formed a cluster contrasting with central New Britain at K = 3; the New Guinea groups separate at K = 4; and a central New Britain cluster splits at K = 5. Then, at K = 6, a Polynesian cluster appeared, centered on the Mãori, with high ancestral proportions for the Samoan and Micronesian samples as well as the Taiwanese Aborigines. The former “East Asian” ancestral proportion in Melanesian populations converted almost entirely to “Polynesian” in this run. At K = 7, 8, and 9, more Melanesian clusters formed in New Britain and New Ireland. All but one of the Melanesian cluster foci are Papuan-speaking groups, primarily located in the interiors of the large islands (see Figures 7 and 8). The Mamusi, who are Oceanic-speaking neighbors of the Ata, are the exception. There is reason to suspect the Mamusi were originally a Papuan-speaking group (perhaps Ata speakers) who adopted an Oceanic language [34]. At K = 10, the “Europeans” were finally identified as a separate cluster. As shown in Table S5, runs at K = 11 and above became unstable and not reproducible.
The approximate percentage of “European” admixture is best seen in Figure 7, which gives average ancestral proportions by population. In the Mãori, the “European” ancestry was ∼12%, and for Samoans it was ∼5%. The Samoan and Micronesian results also suggested minor ties with East Asians and also Melanesians, specifically the “New Ireland” cluster (a number of Lapita sites have been found in the vicinity of New Ireland [3]). The Micronesians had low levels of inferred ancestry shared with populations in New Guinea, which is not far from Belau, where most of the Micronesian samples are from. This relationship is echoed in mtDNA results as well [35]. The typical ancestral proportions by population for a majority rule run are given in Table S6. As seen in Table S5, 15 out of 20 STRUCTURE runs on our Pacific dataset at K = 10 produced essentially the same group ancestry proportions as shown in Figures 6 and 7, with individual similarity coefficients ranging from 0.90 to 0.96, so these results are quite reproducible.
As in the global comparison, an “East Asian/Polynesian” estimated ancestry proportion for a number of Melanesian populations only occurred at frequencies of >5% in certain Oceanic-speaking (Austronesian) groups, and it is hereafter referred to as the “Austronesian” genetic signature. In Figure 7, the purple arrows point to those Oceanic-speaking groups in our Melanesian sample set that have this clear “Austronesian” signature. The probabilities were highest in the Kove and Saposa (just below 20%), followed by the Mussau at 15%, with the Teop, Mangseng, Nakanai (Bileki), Melamela, and Tigak having lower “Austronesian” signatures. In these Oceanic-speaking populations, the “Austronesian” ancestral assignment proportions never ranked higher than third, indicating their comparatively intermixed, and predominantly Papuan, genetic nature.
As a check on these results, particularly to verify the relationships of the Polynesians and Micronesians within our dataset, we performed a separate “supervised” STRUCTURE analysis [28,36], where the individual Mãori, Samoan, and Micronesian genotypes were distributed across eight representative populations (Taiwan Aborigines, East Asians, Europeans, and the Near Oceanic New Guinea, Ata, Baining, Kuot, and Aita). The results, shown in Figure S3A, underline the primary affinity of the Mãori, Samoans, and Micronesians to Taiwan Aborigines and secondarily to East Asians, with lesser suggestions of links to Europeans and New Ireland/New Britain (there is no suggestion of any Bougainville or Baining tie). In a second “supervised” STRUCTURE analysis where a ninth population was specified but not associated with a particular group a priori, the Polynesians/Micronesians constituted the largest proportion of this cluster (Figure S3B). Of the three populations in question, the Mãori had the smallest signal of external relationship, consistent with their extensive genetic drift, and the Micronesian group has the largest signal (to Taiwan, East Asia, New Guinea, and New Ireland/New Britain).
Figure 8 shows the distribution within Northern Island Melanesian populations of the STRUCTURE clustering probabilities for K = 10 in pie-chart form (some populations from the same language groups with very similar probability profiles were merged). Neighboring groups tended to share similar profiles. New Britain, the largest and most rugged island, had the greatest internal differentiation, with five different assigned clusters at >50% probabilities in different populations. Bougainville groups had two common cluster assignments, while there was only one common cluster in New Ireland.
Figure 9 shows the unrooted neighbor-joining tree for the East Asia–Pacific populations from a pairwise FST coancestry distance matrix for 687 microsatellites (the pairwise FST values are in Table S7). Bootstrap values for the branches, generated with the PHYLIP program from population allele frequencies for 100 different trees, are indicated by branch thicknesses. As shown, most of the trunk elements had high bootstrap values, as did a number of branches within Northern Island Melanesian groups. By contrast, the mainland East Asian group relationships were considerably more ambiguous, their branches were shorter, and only the Taiwan Aborigines had a strong internal branch. The tree branching again closely reflected the clustering in STRUCTURE, indicated by the corresponding colors from K = 10. The populations with the longest branches were those with the largest ancestral proportions assigned to single STRUCTURE clusters, and had the lowest heterozygosities. These populations tend to be Papuan-speaking groups in island interiors. The STRUCTURE analysis specifies the role and nature of admixture in a way that a population-based tree cannot.
The AMOVA, STRUCTURE, and population tree analyses were all driven by large distinctions in allele frequencies, rather than by the presence of private alleles in one population or another, since these generally occur in very low frequencies. In the first publication on the global HGDP-CEPH set of 377 microsatellites, Rosenberg et al. quantified continental relationships independent of the STRUCTURE analysis by showing the number of alleles that were only present in one continent, shared by two, by three, etc. [24]. The pattern of specific allele sharing was taken to indicate greater African heterogeneity, and that allele sharing was least for the Americas and for the two “Oceanic” groups.
With our enlarged dataset and microsatellite coverage, we also compared patterns of private alleles and allele sharing between regions (Table 3). We recovered 271 Melanesian-specific alleles, which in raw numbers actually exceeded those for Africa. Correcting for sample sizes, the rate of Melanesian-specific alleles was at the high end of the range for the major regions except for Africa. The number of alleles missing from only one continent, also given in Table 3, shows the dramatic effect of genetic drift on the American populations. The number of shared alleles between pairs of regions is shown in Table 4, with the correction for sample sizes in Table 5. All non-African regions including Melanesia shared the most alleles with Africa, indicating they were primarily subsets of African diversity. Melanesia shared more alleles with East Asia than with any other non-African region, but they cannot simply be viewed as an extension or subset of East Asian diversity. When Papuan and Oceanic speaking groups in Melanesia were analyzed separately, the Papuan-speaking groups showed greater isolation, as they shared fewer alleles with all other regions than did Oceanic speaking groups (unpublished data).
Our study suggests that in the Pacific, and specifically in Near Oceania, there is only a modest association between language and genetic affiliation. Oceanic languages were introduced and dispersed around the islands within the last 3,300 years, but there was apparently only a small infusion of accompanying “Austronesian” ancestry that has survived. Approximately one-half of the Oceanic-speaking groups in Melanesia had an identifiable “Austronesian” genetic signature (see Figure 7 and Table S8). In each case where there was such an “Austronesian” signature, at least two other cluster assignments had probabilities higher than the “Austronesian” one (see, in Figure 6, the Saposa and Teop of Bougainville; the Mussau and Tigak in New Ireland Province; and the Kove, Mangseng, Melamela, and Nakanai Bileki of New Britain). On the other hand, the Oceanic-speaking groups without the “Austronesian” signature were often genetically indistinguishable from their immediate Papuan-speaking neighbors (in New Britain, the Mamusi have no Austronesian signature, but they and the Nakanai Loso cluster closely with their Papuan-speaking Ata neighbors; the Nalik, Notsi, and Madak of New Ireland are genetically indistinguishable from their Papuan-speaking Kuot neighbors; the Tolai and Lavongai profiles suggest significant intermixture, but only between different Papuan-speaking groups). The result suggests that Oceanic languages were adopted by many formerly Papuan-speaking groups, while at the same time there was little genetic influence or marital exchange. At least in Near Oceania, rates of language borrowing and language adoption have been faster and more pervasive than rates of genetic admixture.
However it is measured, genetic variation is reduced within Melanesian populations (Figure 2), while the genetic divergences among them are very large (refer to Figures 6, 8, and 9 and to Tables 1–5). The size of the differences among the populations would appear to equal or surpass those among populations across East Asia, Europe, or even Africa. However, the large Melanesian population distinctions are a direct consequence of their very low levels of internal variation or heterozygosity. These low levels will directly inflate both the proportion of among group variation in AMOVA and also pairwise FST genetic distances (for a full discussion of this point, see especially [27] and also [ 26,37]). As population heterozygosities decrease, pairwise FSTs should increase because of this intrinsic mathematical relationship. This is illustrated by our global and Near Oceania datasets (Figure 10A and 10B). Those pairwise FSTs involving the Bantu South population (which has a heterozygosity approaching 1.0) are plotted against the heterozygosities of each population, and the resulting correlations approach 1.0.
Our Structure and tree analyses of the combined microsatellite datasets indicate that Melanesians are quite far removed from Africans, in spite of their superficial similarities in hair form and skin pigmentation [38]. In the initial analysis of the HGDP-CEPH dataset, the placement of the two Melanesian (“Oceanic”) groups was different. There, they split from Eurasia before Asians and Native Americans [39]. This also differed from the result of a genome-wide SNP study [40] on a very small world-wide dataset. The extreme positioning of Melanesians in our tree was not due to our over sampling. Rather, our extensive coverage of Melanesian variation has enabled a clearer resolution of their relationships with populations outside the region.
The pattern of Near Oceanic diversity has been made clear. The AMOVA analysis of the microsatellites showed that the larger and more rugged the island, the greater the differentiation among populations. The most divergent populations were in large island interiors while these same populations were internally the most homogeneous (as measured by reduced values of θ̂ and expected heterozygosity—Table S1). Genetic variation from one large Near Oceanic island to the next was also significant. While our coverage of microsatellite variation elsewhere in the Pacific was admittedly spotty, our data as well as other smaller scale microsatellite analyses [21,41] suggest that, excluding the large islands of Near Oceania, there is a gradual decline in variation as one moves from Asia eastward, and variation among populations in the Pacific otherwise is not nearly as great as that in the large islands of Near Oceania. As noted, New Guinea does not appear to have as much microsatellite/indel diversity among groups as New Britain. Our sample coverage and definition was less rigorous there, and we expect equivalent coverage in New Guinea would equal or surpass the divergence of our New Britain series.
The biogeographic pattern of population divergences in Near Oceania is most likely attributable to the restricted marital migration distances that have been documented most clearly for inland Bougainville groups [42], as well as for some New Guinea highlands populations [43]. Few people in small inland communities traditionally married and established households more than 1–2 kilometers from their birthplaces, while marital migration distances tended to be longer among shoreline communities. Nettle has argued that in ecologically rich tropical regions such as Near Oceania, small populations easily became self-sufficient, which in turn encouraged isolation and discouraged exchange [44,45], causing the development of extreme diversity among populations in both language and genetics. We suggest this was the underlying cause of the short marital migration distances among inland groups in Near Oceania, which in turn was responsible for the low population heterozygosities and resulting large genetic distinctions among groups [42].
Because they arrived first and came to occupy large island interiors, the Papuan-speaking groups are considerably more diverse than Oceanic-speaking groups, which tend, in large islands, to be arranged along the shorelines. The prehistoric record suggests there was a gradual reduction after initial settlement in the size of foraging zones of formerly mobile groups, associated with the filling up of the landscape [3, p. 16]. In many ways, these patterns and dynamics parallel the biogeography of birds and ants in the same region, where dispersal abilities of different species have dictated their patterns of diversity, and dispersal tendencies have, in many cases, contracted in island interiors over time [46,47].
Some known population relationships suggested the considerable age of the clusters identified by our STRUCTURE analysis. The Tolai of East New Britain, with an assignment profile similar to New Ireland groups, are known to have migrated from southern New Ireland over 1,200 years ago [42]. A major volcanic eruption in western New Britain 3,000 years ago isolated that section of the island, and the Anêm, along with the recently arrived and intermixed Kove, form a separate cluster. Although the two Baining groups of east New Britain formed a cluster of their own, it has been suggested from the mtDNA, Y, and X chromosome analyses that they have been separated by thousands of years [48] (see their long branch lengths in Figure 9). Also, the clustering of the Polynesians, Taiwan Aboriginals, and East Asians reflects ties older than 3,300 years. In the Pacific, the change in genetic clustering apparently has evolved over thousands of years, and in many cases tens of thousands. This is likely a function of small effective population sizes and the high degree of isolation/drift over these immense time periods.
There were indications from the mtDNA, NRY, and certain autosomal microsatellites that in Remote Oceania, where islands are generally smaller in size, genetic variation among human groups is comparatively reduced, which is a contrast to Near Oceania [17,19–21,49]. At some point, prehistoric Oceanic mariners apparently became so accomplished that the inter-island water crossings in the central Pacific were often no more of an impediment to travel than the (already occupied) rugged terrain of the larger island interiors in the western Pacific. In many areas, the ocean was transformed from a formidable barrier into a highway [50,51].
However, exactly where the (relatively homogeneous) Polynesians came from has remained controversial, and the number of proposed explanatory models for their origin form a continuum [49,52]. At one extreme is the “Entangled Bank” [53], which is essentially a null hypothesis for detecting clear signals of specific Polynesian ancestry anywhere to the west. It suggests that, although there certainly must have been a series of introductions and influences from Asia into the Pacific over the millennia, no decipherable signal has survived that can be identified as specifically ancestral to Polynesians, because of the complexities of human interactions from the outset [54]. Proponents argue that tree-like representations of population (or linguistic) relationships cannot be expected to develop regularly and are likely to be entirely inappropriate representations of population relationships in many, if not all, instances, since they so often ignore interactions between neighboring groups.
Models at the other end of the continuum assume contemporary genetic (as well as cultural) similarities can carry a clear signal of past population relationships. Primary among these is “The Express Train to Polynesia” model [55]. It proposes a rapid movement of the ancestors of the Polynesians from the vicinity of Taiwan to the Central Pacific, without extensive contact with indigenous Near Oceanic populations along the way.
With regard to human genetics, the published mtDNA evidence has generally been interpreted as supporting the “Express Train.” This is because a younger mtDNA haplotype (B4a1a1) is assumed to have been closely linked to the development and expansion of Polynesian populations. At present, the state of the evidence for this association is as follows: a) the precursor haplotype to B4a1a1 has been identified in Taiwan aboriginal populations [56]; b) the final development of B4a1a1 with the key mutation at nucleotide site 14022 seems to have occurred in eastern Indonesia or Near Oceania [17]; c) its frequency varies widely over Near and Remote Oceania before becoming ubiquitous in Central and Eastern Polynesian populations; d) in Near Oceania, it is common along many Oceanic-speaking coastal groups, as well as a number of Papuan-speaking groups, especially in New Ireland and Bougainville [17]; and e) its expansion dates are relatively recent, although old enough to suggest to some observers that it cannot be easily tied to the Polynesian expansion [17,56].
The “Slow Boat to Polynesia” model which is supported by NRY variant distributions, also assumes current genetic patterns in Oceania directly reflect prehistoric migrations and interactions. These NRY haplogroup distributions have been taken to suggest a very minor “Asian” contribution to current Polynesian populations, suggesting instead that Polynesians derived primarily from Melanesian (Near Oceanic) populations [19,57,58]. “Melanesian” NRY haplogroups were found to be very common in some Polynesian populations, while “Asian” NRY haplogroups were scarce in Melanesian populations [20,58], and low in their frequencies in the Central Pacific. However, recent studies have shown that the “Asian” NRY haplogroups are not as rare in Polynesia as initially thought, and are quite variable in frequency ([19], Table S2).
Because of their comprehensive nature, we believe the results of our autosomal microsatellite survey present a resolution to this issue with regard to human genetic relationships. The fact that the STRUCTURE cluster containing Micronesians, Samoans, and Maoris has a detectable signature only in Oceanic-speaking Melanesians and Taiwan Aborigines supports the position that an expansion of peoples from the general vicinity of Taiwan is primarily responsible for the ancestry of Remote Oceania, and that these people left a small but still identifiable signature in (some Oceanic-speaking) populations of Near Oceania. Scenarios for different male and female dispersals have been proposed to reconcile the divergent mtDNA and NRY patterns in Oceania [35,59], but the autosomal microsatellite results should now serve as the primary reference.
Although the Polynesians in our analysis were similar to Taiwan Aborigines and East Asians, they might be even closer to other populations not covered in our study, from Indonesia, the Philippines, or Southeast Asia. While there is a substantial body of evidence that indicates Taiwan is the primary point of Austronesian dispersal [60,61], there are now also suggestions of the importance of (Island) Southeast Asia as well [62,63]. The ties of particular Near Oceanic populations to those regions also remain poorly understood, but should be resolved with additional sampling from these regions and similar analyses.
To revisit the questions posed at the beginning, we can provide answers as follows.
1) To whom are these Melanesian populations most closely related outside the Pacific? Outside the Pacific, East Asian populations are apparently the closest (but still very distant) relatives of Melanesians. Africans and Europeans are the most distant.
2) How does the genetic diversity of Near Oceanic populations compare with groups in other regions? The within-group diversity in Melanesian populations is consistently very low, which acts to exaggerate the considerable among-group distinctions there. This great diversity in such a small region makes comparisons of human population structure from continent to continent problematic.
3) Is there a clear organization of the variation among Melanesian groups? The diversity among groups is primarily organized by island size and topographic complexity, with the inland Papuan-speaking groups the most isolated and differentiated. Shore-dwelling Oceanic-speaking groups are more intermixed (dispersal along the shorelines was easier).
4) Is there an identifiable genetic signature of Taiwanese/Southeast Asian or Polynesian influence in Near Oceanic populations, especially in the Bismarcks, where the Lapita Cultural Complex developed? There is a weak “Austronesian” genetic signature in only a portion of Oceanic-speaking populations in Melanesia, and none at all in Papuan-speaking groups (contradicting the results of mtDNA, but in accord with the NRY results).
5) Are Polynesians more closely related to Asian/Taiwanese populations or to Melanesians? Polynesians are closely related to Asian/Taiwanese Aboriginal populations, while they are very weakly associated with any Melanesian groups (the closest association there appears to be with New Ireland populations). This is in accord with mtDNA interpretations, but differs from the usual interpretation of the NRY results. The sailing capabilities of the ancestors of the Polynesians transformed the nature of their Diaspora and kept them relatively homogeneous.
Our Asia–Pacific sample set came from a variety of sources. The objective was to include between 15 and 25 unrelated individuals (minimally excluding reported first-degree relatives) from locales where individuals and their parents had all lived. These criteria were achieved in most instances. All of the samples except the cell lines were Whole Genome Amplified (Qiagen RepliG). Details are given below.
1. Samples from Northern Island Melanesia were collected in three field seasons (1998, 2000, and 2003) in collaboration with the Institute for Medical Research of Papua New Guinea. Besides a 10 ml blood sample, a simple genealogy and residency questionnaire was taken, including in most instances parent and grandparent names, residences, and native languages. All individuals gave their informed consent for participation, and the study was approved by the Institutional Review Boards of Papua New Guinea, Temple, Michigan, Yale, and Binghamton Universities. Among over 1,500 samples collected, 995 were chosen for submission to the Marshfield Clinic for microsatellite and indel analysis. As many Papuan-speaking groups as possible were included, along with neighboring Oceanic-speaking groups, focusing on New Britain, New Ireland, New Hanover, Mussau, and Bougainville. We included multiple locales in larger language groups where feasible; and picked samples from individuals whose family's residence histories suggested close identification with the sampling locale. People of mixed parentage (especially with one grandparent from a different language group or island) could not always be excluded if the minimum required sample size was to be achieved. A number of individuals who were born on the New Guinea mainland but had settled in Northern Island Melanesia were taken to constitute one additional sample—the “Sepik” —so that this sample is a conglomerate. DNA was extracted as previously described [17].
2. DNA was obtained from the Kidd lab collection of cell lines for: a) the Eastern Highlands of Papua New Guinea, primarily from the Gimi, which were collected in collaboration with the Papua New Guinea Institute of Medical Research, and also from Goroka Town; b) Micronesians, primarily from Belau, who drew each other's blood samples during their training in the Pacific Basin Medical Officer Training Program; and c) Samoans, who were in a combined collection from the Pacific Basin Medical Officer Training Program and from American Samoa. All individuals gave their informed consent for participation.
3. New Zealand DNA samples were collected from indigenous Mãori individuals residing in the North Island. Individuals were unrelated by first degree, had two Mãori parents by self-report, and belonged to one segment of the wider Mãori population. Ethical clearance was granted by the NZ National Ethics Committee. DNA was extracted from blood using Qiagen kits.
4. Taiwan Aboriginal samples comprise the Northeastern Taroko tribe from Hsiulin, part of the Atayal language group, and the Amis tribe living on the east coast of Taiwan and speaking Amis. All individuals were unrelated and had both parents belonging to the same tribe. Each individual gave informed consent to participation in population genetics studies and the project was approved by the Ethics Committee of the Hospital and the Department of Health of Taiwan. Blood samples were collected in acid citrate dextrose tubes. Genomic DNA was extracted from 500 μl of buffy coat using the QIAmp DNA kit (QIAml blood kit, Qiagen) by Loo Jun-Hun at the Transfusion Medicine and Molecular Anthropology Laboratory, Mackay Memorial Hospital, Taipei.
Each individual was originally genotyped for 751 microsatellite and 481 insertion/deletion autosomal polymorphisms. The microsatellites were drawn from Marshfield Screening Sets #16 and #54, and the indel markers were drawn from Marshfield Screening Set #101.
890 markers typed in our Pacific series (203 indels and 687 microsatellites) had been typed in the HGDP-CEPH Human Genome Diversity Cell Line as described in [23], although for some microsatellites, a change in primer length or position occurred between the HGDP-CEPH genotyping (2004) and our own (2006), or a change in allele calling occurred. Where the primer changed, allele sizes from one of the two data sets were adjusted (Table S9). The changes were done by comparing the same set of individuals (called “Nasioi” in our dataset, and “Melanesians from Bougainville” in the HGDP-CEPH dataset) duplicated in both studies. Two loci for which the allele size shift was ambiguous—GATA11C08 and GGAA10C09—were excluded. Of the 687 microsatellites remaining for the combined analysis with the HGDP-CEPH panel, 166 had primer changes between the datasets. All analyses utilized the 687 microsatellites, and in addition the 203 indels were used in the STRUCTURE analyses. The set of 957 individuals used here from the HGDP-CEPH panel is the “H971” subset of the original panel [64], without first-degree relatives, and with the Melanesian (Nasioi) removed, since these individuals were also present in our samples (one individual, number 857, was inadvertently deleted early in this analysis). Small African populations with single or two individuals were grouped into Bantu South (Herero, Ovambo, Pedi, Sotho, Tswana, and Zulu)
The expected heterozygosity and average number of alleles per locus were computed on the microsatellites with the GDA software [65], using the sample-size corrected estimator, as in [66]. FST was estimated on the microsatellites as in Equation 5.3 from [67] , using GDA, with 95% confidence intervals based on 1,000 bootstraps across loci. Indels were excluded from all analyses except STRUCTURE.
Cluster analysis of genotypes utilized the Structure versions 2.1 and 2.2 software package [28,36]. Results using Structure 2.1 and 2.2 were essentially identical. STRUCTURE was run with a Markov Chain Monte Carlo (MCMC) burnin of 20,000 steps, followed by an MCMC chain of 10,000 steps for clustering inference. Ten runs were performed at each K in most cases, except as noted in Table S3 (for K = 7) and Table S5 (for K = 10). When multiple runs at the same values of K produced discrepant results, we relied on majority rule (i.e., modal topography in cluster assignment) to pick the optimal result. For the combined global analysis, we terminated the STRUCTURE runs at K = 6, as explained in the Results, and for the Pacific set we terminated the analysis when it became unstable at higher values of K (i.e., when multiple solutions appeared). Details are provided in the Tables S3 and S5.
Individual similarity coefficients for pairs of runs were calculated as in [24] and Methods.
The neighbor-joining trees for Figures 5 and 9 were based on the FST distance matrices obtained with GDA. The bootstrap values for the Asia–Pacific dataset (Figure 9) were obtained based on allele frequencies using PHYLIP [68]. The neighbor-joining trees in Figure S3 were calculated using MSA [69] and drawn with Phylip.
Great circle geographic distances were calculated with the Haversine method as described in [26].
The results of the STRUCTURE runs were graphed with the software DISTRUCT [70]. |
10.1371/journal.pcbi.1005794 | A Bayesian approach to modelling heterogeneous calcium responses in cell populations | Calcium responses have been observed as spikes of the whole-cell calcium concentration in numerous cell types and are essential for translating extracellular stimuli into cellular responses. While there are several suggestions for how this encoding is achieved, we still lack a comprehensive theory. To achieve this goal it is necessary to reliably predict the temporal evolution of calcium spike sequences for a given stimulus. Here, we propose a modelling framework that allows us to quantitatively describe the timing of calcium spikes. Using a Bayesian approach, we show that Gaussian processes model calcium spike rates with high fidelity and perform better than standard tools such as peri-stimulus time histograms and kernel smoothing. We employ our modelling concept to analyse calcium spike sequences from dynamically-stimulated HEK293T cells. Under these conditions, different cells often experience diverse stimulus time courses, which is a situation likely to occur in vivo. This single cell variability and the concomitant small number of calcium spikes per cell pose a significant modelling challenge, but we demonstrate that Gaussian processes can successfully describe calcium spike rates in these circumstances. Our results therefore pave the way towards a statistical description of heterogeneous calcium oscillations in a dynamic environment.
| Upon stimulation a large number of cell types respond with transient increases of the intracellular calcium concentration, which often take the form of repetitive spikes. It is therefore believed that calcium spikes play a central role in cellular signal transduction. A critical feature of these calcium spikes is that they occur randomly, which raises the question of how we can predict the timing of calcium spikes. We here show that by using Bayesian ideas and concepts from stochastic processes, we can quantitatively compute the calcium spike rate for a given stimulus. Our analysis also demonstrates that traditional methods for spike rate estimation perform less favourably compared to a Bayesian approach when small numbers of cells are investigated. To test our methodology under conditions that closely mimic those experienced in vivo we challenged cells with agonist concentrations that vary both in space and time. We find that cells that experience similar stimulus profiles are described by similar calcium spike rates. This suggests that calcium spike rates may constitute a quantitative description of whole-cell calcium spiking that reflects both the randomness and the spatiotemporal organisation of the calcium signalling machinery.
| Transient changes in the intracellular calcium (Ca2+) concentration have long been associated with the activation of plasma membrane receptors [1]. Since the seminal work by Woods et al. [2] linking the frequency of cytosolic Ca2+ oscillations in hepatocytes to the concentration of various hormones, both experimental and theoretical studies have provided compelling evidence for encoding extracellular stimuli into intracellular Ca2+ oscillations [3–14]. In whole-cell recordings, Ca2+ oscillations are usually observed as sequences of spikes of the intracellular Ca2+ concentration.
A prominent feature of Ca2+ spike sequences is that they are random. Ca2+ spikes only occur with some probability that generally changes over time. For example, there are distributions of inter-spike intervals (ISIs) for agonist induced Ca2+ oscillations in HEK293 cells and spontaneous Ca2+ oscillations in astrocytes, microglia and PLA cells instead of a single value [15]. When astrocytes are transiently stimulated with ATP three times, with a recovery period between stimuli, the observed Ca2+ spikes display any number of response patterns, from no spikes to three [16]. To elucidate the principles that govern the translation of extracellular cues into changes of the intracellular Ca2+ concentration therefore requires faithfully capturing the stochasticity in Ca2+ spike generation.
To date, the main modelling approach to investigate stochastic Ca2+ spike generation has been based on numerical solutions of differential equations, both ordinary and partial [11, 12, 15, 17–30]. In these studies, the randomness of Ca2+ spikes results from the stochastic behaviour of Ca2+ releasing channels, such as the inositol-1,4,5-trisphosphate (InsP3) receptor (InsP3R), and their interactions. The random dynamics of the InsP3R is then either described by coupling a Markov chain for the InsP3R to the differential equations, or by assuming a Langevin-type equation. All of these approaches require detailed models of the InsP3R (with often a considerable number of rate constants), and other assumptions such as the number of InsP3Rs per cluster and the spatial distribution of InsP3Rs, Ca2+ pumps and Ca2+ buffers in the case of partial differential equations. However, such mechanistic detail, which has been instrumental in advancing our understanding of Ca2+ spikes, often comes at considerable computational costs.
It might be therefore desirable to change the perspective from the mechanistic bottom-up approach to a top-down view, in which cellular Ca2+ spikes are described directly. The mathematical concept that has proven particularly useful for this endeavour is the theory of point processes [31]. Indeed, in [15] a time-dependent conditional rate for the generation of a Ca2+ spike was introduced, and its two parameters (a time scale and an amplitude) were determined from experiments on four different cell types. Subsequent work [32, 33] demonstrated in more detail that for constant stimulation, the time scale was cell type specific, while different cells of the same type could be distinguished by their amplitude.
These modelling approaches have not confronted the issue of the dynamic nature of cell stimulation that occurs under physiological conditions. Cells in vivo experience a complex and dynamically-changing environment, where signals frequently arrive in a time-varying manner—such as transient release of neurotransmitters or oscillations in the concentration of circulating hormones. Diffusion of messengers through tissue (such as away from a blood vessel) also introduce spatial variation in signal strength, meaning that cell populations encounter a complex spatiotemporal pattern of stimulation. An efficient mechanism for modelling Ca2+ responses in such a heterogeneous population has not yet been devised. Our approach offers a solution to this issue and is based on combining point processes with Bayesian inference.
Bayesian concepts have been used with great success across various disciplines (see e.g. [34] for an overview). One advantage of utilising Bayesian ideas is that model parameters can be effectively constrained by observed data and can be estimated in a controlled fashion. For example, each set of parameter values comes with its own probability that informs us about how likely this set represents the observed data. In the past, the combination of Bayesian inference and point processes has been successfully applied to action potential spike trains in neurons [35, 36], but to our knowledge, this is the first time that Ca2+ spike sequences have been analysed in this way. While we can draw on these previous results, the substantial differences between action potential spike trains and Ca2+ spike sequences (e.g. the number of spikes per train or the time scales of spikes) have required significant attention.
A particular characteristic of Ca2+ spikes is that their generation depends on the cellular Ca2+ spike history and hence is non-Markovian. This results from both the stochastic nature of Ca2+ spike formation as well as the dynamic variation in the cellular signalling micro-environment. To model such history dependence, we follow a concept introduced in [37], which effectively turns a non-Markovian Ca2+ spike sequence into a Markovian one. This is achieved with the help of a so-called intensity function x(t), for which we provide a definition and more details in the Materials and Methods section. Since the intensity function is directly inferred from individual Ca2+ spike sequences, it is specific to each cell. This results from two facts. Firstly, Ca2+ spikes are shaped by the cellular composition of the Ca2+ signalling toolkit, i.e. the expression levels and spatial arrangement of Ca2+ channels, pumps, transporters and buffers. Secondly, each cell experiences a different signalling micro-environment as illustrated above. Autocrine and paracrine signalling modify the original signal further. In addition, any time-dependence of x(t) that originates from a time varying signal is compounded by dynamic changes of the Ca2+ signalling apparatus, such as Ca2+ store refilling, adaptation or desensitisation.
In the present paper we demonstrate how to estimate x(t) in the presence of such single cell variability. This will allow us to make progress in two main directions. Firstly, one of the most highly discussed questions in computational biology is concerned with scaling dynamics from single cells to tissue. Taking into account our current understanding of intracellular Ca2+ signalling, this will require the simulation of spatially extended single cells driven by fluctuating Ca2+ releasing channels. The computational costs for such studies are extraordinary, and how to sample the associated high dimensional parameter space remains an open challenge. On the other hand, generating Ca2+ spike sequences from an intensity function is computationally cheap and hence puts researchers in an advantageous position to obtain high quality statistical insights into tissue level dynamics. Secondly, while our approach is statistical, the Ca2+ ISI distribution that we determine as part of estimating x(t) has direct mechanistic interpretations. This may further our understanding of how exactly Ca2+ spikes emerge at the single cell level from the orchestrated action of the Ca2+ signalling toolkit. For instance, one line of argument suggests that Ca2+ spikes result from the co-ordinated interplay of a certain number of Ca2+ puffs. Our statistical analysis may provide quantitative estimates for this.
A main motivation for our study is the need to model the heterogeneity of Ca2+ responses observed in cell populations in a computationally efficient manner. To illustrate the problem, the results in Fig 1 show HEK293T cells challenged with a solution containing 100 μM carbachol. Fig 1B illustrates the large variability of Ca2+ responses observed between different cells even when the stimulus is stationary. While some cells only detect the onset of the stimulus (third trace), other cells exhibit Ca2+ spikes during the entire stimulation period. However, the spike characteristics vary significantly. Some cells show irregular Ca2+ spikes (first trace), while other cells settle into an almost regular pattern (fourth trace). It is also worth noting that the frequency of Ca2+ spikes spans a considerable range (cf. first and second trace) and that all cells display a decrease in peak amplitude, which is indicative of adaptation. Fig 1C and S1 Video provide further evidence for the large heterogeneity in the timing of Ca2+ spikes for a constant stimulus.
The first step in developing a top-down model that reproduces this heterogeneity is to determine the probability distribution that most accurately describes recorded ISIs. Throughout this study, Ca2+ spikes are treated as all-or-nothing events and any information on the width or amplitude of a Ca2+ spike is excluded. The reason why we can introduce an ISI distribution and hence treat successive ISIs as independent—instead of trying to fit a time-dependent ISI distribution—comes from the introduction of the intensity function x(t). We tested three possible candidate ISI statistics (see Materials and methods for details): an inhomogeneous Poisson (IP), an inhomogeneous Gamma (IG) and an inhomogeneous inverse Gaussian (IIG) distribution. The IP distribution often serves as starting point for analysing spiking behaviour as it is the most basic statistical distribution. While the parsimony of the IP distribution has undoubtedly helped in establishing a large body of mathematical results, real world data often exhibit more complex statistics. The IG distribution provides a natural extension of the IP distribution, in that it contains the IP distribution as a special case: putting γ = 1 in (5) recovers (10). The shape parameter γ endows the IG distribution with more flexibility, which has proven fruitful in numerous applications. Conceptually, spikes in general and Ca2+ spikes in particular have been described as first passage events [38]. One of the most fundamental models, which contains positive drift and random motion only, gives rise to the IIG distribution.
To ascertain which ISI distribution describes the results in Fig 1 best, we transformed the recorded ISIs using the time rescaling theorem and analysed the results in a Kolmogorov-Smirnov plot (see Material and methods). The Kolmogorov-Smirnov plot allows the visual inspection of whether two probability distributions are the same by plotting their respective cumulative distribution functions against one another. If the two distributions are identical, the cumulative distributions functions coincide and plotting one against the other results in a straight line with slope 1. Fig 2A shows results for the Kolmogorov-Smirnov plot for the data taken from Fig 1. Here, u corresponds to the cumulative distribution of the transformed data, which depends on the chosen ISI distribution, while s stems from a theoretical prediction based on the time rescaling theorem (see Materials and methods). If we have identified the correct ISI distribution, data points should cluster around a straight line with a 45° slope. Note that in our analysis, each cell was treated individually and no data amalgamation took place.
Results for the IP and IIG clearly deviate from a line with slope 1, while the data for the IG exhibit much less deviation. This suggests that an IG, but not an IP or IIG, better describes the ISI statistics of Ca2+ spikes. The box plots in Fig 2C provide further quantitative evidence. They demonstrate that the slopes obtained from individual cells are concentrated close to 1 and exhibit a significantly smaller variability for the IG compared to the IP and IIG. To further corroborate these findings, we analysed data from cells exposed to a lower concentration of carbachol (10μM). Fig 2B and 2D illustrate that again the IG distribution captures the ISI statistics most closely. In addition to Kolmogorov-Smirnov plots, we also tested our data with a quantile-quantile plot (see S1 Fig and Materials and methods). As with the Kolmogorov-Smirnov plot the correct ISI distribution leads to data points that accumulate around the 45° line. Panels A and B in S1 Fig show that this is the case for the IG, but not for the IP and IIG, hence confirming our results from the Kolmogorov-Smirnov plot. The box plots in panels C and D of S1 Fig show that again the slopes obtained from single cells for the IG model are much closer to one and exhibit less variability than those for the IP and IIG. Taken together, these results indicate that the patterns of Ca2+ spikes observed across the population of cells can most accurately be reproduced with an IG distribution. This approach will therefore be used as the basis of our analysis.
It is worth noting that when testing the different ISI distributions, we also obtain an estimate for the intensity function x(t). As illustrated by Eq (11), the probability for a specific ISI depends on x(t). Therefore, all ISI distributions in this study have to be understood as being conditioned on x(t). In the next section, we show how to estimate x(t) from measured Ca2+ spike sequences.
In addition to the ISI distribution, we also need to know the intensity function x(t) to fully describe Ca2+ spike sequences. In the following, we will use three different approaches to estimate x(t): peri-stimulus time histograms (PSTHs), kernel smoothing (KS), and Gaussian processes (GPs) combined with Bayesian inference. An illustration of a GP is shown in Fig 3. In contrast to a deterministic curve, a GP generates infinitely many curves (3 possible candidates are shown), and the statistics of these curves is the organising principle. At each time point, the values of a GP are Gaussian distributed, and the mean and standard deviation can change over time. To guarantee a controlled comparison between PSTHs, KS and GPs, we will fix an intensity function, generate surrogate Ca2+ spike sequences from it and then estimate how well the above methods recapture the original intensity function (see Materials and methods for details).
Before proceeding it is worth noting that the intensity function that we need to estimate is identical to the Ca2+ spike rate that we find from using either PSTHs or KS [36]. In other words, if we can obtain a high quality estimate for the Ca2+ spike rate from either PSTHs or KS, we have a very good estimate for x(t). However, this usually requires a large number of Ca2+ spike sequences, which is an issue that we will address below. Since we can identify the Ca2+ spike rate with the intensity function, we will use both terms interchangeably. Note, however, that this Ca2+ spike rate is different from the conditional intensity function defined in Eq (15), which is often used in generating spike trains.
A major objective of this model is to reproduce Ca2+ signalling patterns during complex stimulation conditions. For example, physiological patterns of hormone or neurotransmitter release are time-varying, rather than the step-changes used in typical experiments (such as Fig 1). We therefore tested the performance of candidates for x(t) for reproducing dynamically-changing signals—specifically sinusoidal oscillations. As a first choice, we considered a regularly oscillating intensity function xdet(t) = 0.5 cos(t) + 0.5 cos(0.5t) + 1. Fig 4A shows Ca2+ spike sequences generated from xdet, and Fig 4C reveals that both PSTH and KS capture the original intensity function very well.
By using a specific functional form for x(t) as in xdet(t), we make strong assumptions about the intensity function. A more flexible and versatile approach is based on GPs. In Fig 4B we plot Ca2+ spike sequences generated from one xGP candidate, while Fig 4D depicts the estimation of xGP from a PSTH and KS. As with xdet we find very good agreement between the original and estimated Ca2+ spike rate.
This strategy supposes that all Ca2+ spike sequences can be combined into a single large dataset, but in a physiological context this will not always be true. To be a more useful tool, the model should be able to simulate the diversity of responses expected in a more complex environment, where the stimulus varies in both space and time. Under these circumstances, the number of cells that receive an equivalent stimulus would be more limited, and so it is important to assess how the number of spike sequences available for parameter estimation affects the accuracy of predicting the Ca2+ spike rate. Consequently, we randomly picked groups of 1, 2, 4 and 7 Ca2+ spike sequences and computed the Ca2+ spike rate based on a GP and KS.
In Fig 5 we show results for xGP. Since the Ca2+ spike rate estimation obtained from KS depends on the bandwidths σ (see Eq (17)), we employed different σ values. For a single Ca2+ spike sequence (Fig 5A) the estimated Ca2+ spike rates differ visibly from the theoretical one, and the smallest bandwidth leads to spurious oscillations. As we increase the number of Ca2+ spike sequences the estimated Ca2+ spike rates capture the true Ca2+ spike rate more faithfully. The light blue area in each panel delineates the 95% confidence interval, which we obtain as a by-product from the GP optimisation. Overall, all estimates lie within this confidence interval except the one for σ ^ in Fig 5A (see Materials and methods for the definition of σ ^). We obtain similar results for xdet(t) as illustrated with S2 Fig.
We quantified the accuracy of predicting the Ca2+ spike rate by computing the normalised L2 norm of the difference between the known and estimated Ca2+ spike rate (see Materials and methods). Fig 6 shows that for a given method, the L2 norm decreases as we increase the number of Ca2+ spike sequences, which corresponds to better predictions. When we fix the number of Ca2+ spike sequences, GPs yield a better estimate. The improvement is particularly evident when comparing the Ca2+ spike rate estimation based on σ ^ at small numbers of Ca2+ spike sequences. We further tested that our results did not depend on the particular choice of Ca2+ spike sequences nor on the details of the surrogate generator. For the latter, we compared three different approaches: inverse sampling, a Bernoulli process and time rescaling. We generated a number of Ca2+ spike sequences with each method and then estimated the Ca2+ spike rate using the same methods as in Fig 6, i.e. KS with different bandwidths and GPs. S3 Fig shows box plots of the normalised L2 norm between the estimated and the true Ca2+ spike rate xdet. We find that for all three methods, the normalised L2 norm is generally smallest for GP estimates. To test the statistical significance of this result, we computed the corresponding p-values as shown in S1 Table using the non-parametric Mann-Whitney test. Based on the common assumption that a finding is statistically significant if p < 0.05, estimates using GPs perform statistically better than KS since the largest p value was 0.0375. We repeated the analysis for xGP and report the normalised L2 norm in S4 Fig. While the GP performs clearly better than KS with bandwidths of σ ^ and 35, the distributional results for bandwidths of 52 and 70 look similar to those of the GP. This is also confirmed by the p-values shown in S1 Table, where some exceed the threshold of 0.05. This indicates that KS can approach the performance of GPs. However, since there are no a priori estimates for this optimal bandwidths for a given scenario, GPs provide the more robust estimation method.
The results so far provide strong evidence that an intensity function derived from a GP allows accurate prediction of Ca2+ spike patterns even when the estimate is based on small numbers of Ca2+ spike sequences. We next applied our Bayesian approach to a more complex experimental system, designed to reproduce some of the stimulus heterogeneity expected in vivo.
We used a microfluidics chamber to deliver sinusoidal changes in carbachol concentration to HEK293T cells (see Materials and methods). The concentration of carbachol varied in both space and time. Fig 7A shows a complex concentration surface throughout the chamber at a fixed time point and illustrates how the sharp interface between high and low agonist concentration on the left side of the chamber widens as the flow progresses through the chamber. In Fig 7B we plot agonist concentration time courses sampled at four positions along the transverse direction of the microfluidics chamber, which demonstrates the stimulation heterogeneity that cells experience depending on their position within the chamber. We also include the corresponding Ca2+ spike sequences, which again display significant variability. The goal of this experiment was to generate an environment in which a population of cells is exposed to agonist in a manner that varies with the cells’ distance from the stimulus source and with a dynamic mechanism of delivery (by analogy to a circulating hormone diffusing from a blood vessel to underlying cells, for example).
This scenario is the context in which a Bayesian framework is most useful, as it can predict spiking Ca2+ responses to a complex but physiologically meaningful stimulus profile. In Fig 7C–7E, we show clustered stimulus curves, the corresponding Ca2+ spike sequences and the estimated intensity functions, respectively. We grouped cells that experienced similar agonist concentration profiles to allow for a meaningful comparison of the resultant intensity functions. To determine how similar stimulus traces are, we computed the weights of the three leading principal components (see Materials and methods). Results are shown in Fig 8, where we employed k-means [39] to detect possible clusters. Data points that belong to the same group are plotted in the same colour, and these colours correspond to those used in Fig 7C–7E. Overall, 4 distinctive groups represented the data best, containing 10, 13, 13 and 3 cells, respectively. It is worth noting that while clustering was performed on the stimuli time courses the Ca2+ spike sequences show a consistent pattern in that they are generally more similar within a given group than between groups.
The black lines in Fig 7C and 7E denote the mean stimulus and mean intensity function, respectively. We observed that the intensity functions broadly mirror the global behaviour of the stimulus. The mean intensity function for all responses showed a single peak with a similar time course to the stimulus. The amplitude of stimulus and intensity function are also well matched. However, there are also observable differences. For example, individual cells showed intensity functions with a more complex time course than the mean (such as multiple peaks). Furthermore, while the mean stimulus was symmetrical, the mean intensity functions could exhibit notable asymmetries. For example, for weak stimuli, the rising phase of the intensity function may be markedly slower than the falling phase (Fig 7C and 7E; yellow and green traces). This most likely reflects the excitable character of intracellular Ca2+ signalling [30, 40]. For weaker stimulation, it takes longer to reach the threshold for generating a Ca2+ spike, hence the intensity function grows more slowly. The quicker decrease results from the stimulus dropping below the Ca2+ spike generating threshold quickly after reaching its maximum, hence prohibiting further Ca2+ spikes. The faster increase of the intensity function for stronger stimuli (grey traces) lends further support for this interpretation, as the Ca2+ spike generating threshold is reached more quickly.
To illustrate the value of the mean intensity functions shown in Fig 7E, we generated surrogate Ca2+ spike sequences from them using the IG distribution and plotted them in Fig 9C. For ease of comparison, we also show the measured Ca2+ spike sequences in Fig 9B. We first note that the simulated Ca2+ spike sequences resemble the measured ones. This is also confirmed by the histrograms in Fig 9D, which exhibit large overlaps between the experimental and theoretical Ca2+ spike sequences. To quantify how similar the two histograms are, we computed the histogram distance given by
H ( R , S ) = ∑ i min ( R i , S i ) max ( ∑ i R i , ∑ i S i ) , (1)
where Ri and Si denote the histogram count in the ith bin of the recorded and simulated data, respectively. The histogram distance is bounded between 0 and 1, and the closer it is to 1 the more similar the histograms are. The histograms coincide if H = 1. We found that H = 0.86, 0.94, 0.91 and 0.89 (from top to bottom), which confirms our visual inspection that the histograms vary little between recorded and simulated Ca2+ spike sequences. Moreover, Ca2+ spike sequences generated from one intensity function exhibit a certain degree of heterogeneity, which is consistent with our experimental findings. Taken together, these results show that without explicitly including any information about the stimulus into the estimation of the intensity functions, our approach yields intensity functions that reflect the characteristics of the stimulus and that are consistent with experimentally recorded Ca2+ spike sequences.
A key task for all cell types is to faithfully respond to external stimuli. For signalling pathways that rely on the dynamics of the intracellular Ca2+ concentration, sequences of Ca2+ spikes have long been recognised as the likeliest encoding mechanism of extracellular cues. Detailed numerical simulations have provided a mechanistic understanding of how cells generate Ca2+ spike sequences and have demonstrated the emergence of ISI fluctuations from subcellular processes such as Ca2+ puffs. Conceptually, such modelling falls into the class of bottom-up approaches. In the present study, we have adopted a top-down perspective in that we have developed a modelling framework that directly describes the stochastic timing of Ca2+ spikes at the cellular level. Importantly, our data-driven approach implicitly accounts for subcellular details through the introduction of the intensity function x(t).
Our modelling approach is based on the idea of representing Ca2+ spike sequences as realisations of a point-process. In contrast to earlier applications of this concept, consecutive Ca2+ spikes in the current study do not necessarily exhibit the same statistics, i.e. the ISI distribution may become time-dependent. This is a direct consequence of the time-varying stimulation. However, through appropriately transforming the times of Ca2+ spikes and by using the time-dependent intensity function x(t), it is possible to describe Ca2+ spike ISIs with one distribution for the entire Ca2+ spike sequence.
We therefore began our investigation by testing different ISI distributions: inhomogeneous Poisson (IP), inhomogeneous Gamma (IG) and inhomogeneous inverse Gaussian (IIG). The IP process is a common choice as it is the simplest stochastic process and the only one for which the conditional intensity function q(t|yk, x) coincides with the intensity function x(t) (see Eq (15)). This greatly facilitates the mathematical analysis, which can draw on a large body of already established results. However, the IP process often fails to describe experimental spike trains, see e.g. Barbieri et al. [37]. We therefore turned to the more general IG process, which includes the IP process as a special case. In addition, we employed the IIG distribution. Our results in Fig 2 show that the IG distribution captures experimental ISIs very well, whereas both the IP and IIG distribution poorly represent the Ca2+ spike data. While these results all pertain to whole cell Ca2+ spikes, they also shed further light on the details of the subcellular processes that generate these Ca2+ spikes. The IG distribution with shape parameter γ results from sampling an IP process every γth spike. In other words, the ISI distribution for an IG is the same as if one measured the ISIs for γ successive spikes generated from an IP and then added all γ ISIs up to obtain a single ISI. Interestingly, one proposed mechanism for the generation of Ca2+ spikes is wave nucleation, where a critical number of Ca2+ puffs has to occur [19, 41]. Under the assumption that Ca2+ puffs are described by an IP process, the above arguments entails that γ can be linked to the number of Ca2+ puffs to trigger a Ca2+ spike. From a more general perspective Ca2+ wave nucleation can be considered as a first passage time problem, since we are interested in the first time that a critical number of Ca2+ puffs occurs. Importantly, the concept of first passage times is also at the heart of Ca2+ puff generation [42–44]. Assuming a continuous representation for Ca2+ spike generation, one of the simplest first passage time problems describes Brownian motion with positive drift to reach a fixed level for the first time [45]. The associated probability distribution is the IIG distribution, which we chose as our third candidate. The failure of the IIG distribution to capture the behaviour of Ca2+ spikes may point towards more complex subcellular dynamics than random motion and positive drift.
We used two different tests to determine the most likely ISI distribution, a quantile-quantile plot and a Kolmogorov-Smirnov plot. If we correctly identified the ISI distribution that is consistent with experimental data, the measured ISIs can be transformed to obey an exponential ISI distribution. Both the quantile-quantile plot and the Kolmogorov-Smirnov plot interrogate how closely the transformed ISIs are described by an exponential distribution. The fact that the two tests focus on different aspects of the distribution [46, 47] and that both identified the IG as the most plausible ISI distribution provides strong support for our findings. While our analysis suggests that the IG distribution describes the measured data best, we cannot rule out that other distributions that we have not tested, e.g. a log-normal distribution or a generalised exponential distribution, might yield equally good or even better results. If future work reveals another probability distribution than the IG distribution that is consistent with the data, the discussion will turn towards the stochastic processes that generate these distributions and how they reflect the physiology of Ca2+ signalling. We will touch on this point later in the discussion.
Our modelling framework rests on two pillars: an ISI distribution and a Ca2+ spike rate. While the ISI distribution may shine light on potential mechanisms that generate Ca2+ spikes, the Ca2+ spike rate encodes the speed of Ca2+ spike formation. To estimate the Ca2+ spike rate, we used the fact that it coincides with the intensity function x(t). Given that there are numerous ways to generally estimate intensity functions from experimental data, we tested three approaches with particular emphasis on Ca2+ spikes: Bayesian inference with GPs, KS and PSTHs. As Fig 4 illustrates, both KS and PSTHs yield excellent results when we have a large number of Ca2+ spike sequences that are all generated from the same intensity function. One of the reasons for these good estimates is that by pooling all Ca2+ spike sequences, the statistics become effectively Poissonian [31, 48] and that in this case an optimal bandwidth for KS [49, 50] and an optimal bin size for the PSTH [49, 51] is known. However, no a priori estimates for a bandwidth or bin size exist when only a small number of Ca2+ spike sequences with a few spikes each is available. As Fig 6 shows the relative error in estimating the true Ca2+ spike rate strongly depends on the bandwidth when Ca2+ spikes are generated from an IG process and estimates are based on only a few Ca2+ spike sequences. This severely limits the use of KS and PSTHs in estimating Ca2+ spike rates from experimental recordings since firstly the ISI statistics are not Poissonian, and secondly combining Ca2+ spikes from a large number of cells might not be possible as we will discuss below. We therefore turned to Bayesian inference using GPs. Importantly, the Bayesian approach already works for a single Ca2+ spike sequence. By using a prior distribution p(θ) for the hyperparameters θ, i.e. the parameters that describe the shape of the ISI distribution and that control the behaviour of the GP, we explicitly represent the uncertainty associated with each hyperparameter. This alleviates the need for fixing parameter values prior to the estimation of the Ca2+ spike rate as is the case for PSTHs and KS. Fig 6 shows that GPs yield better results than KS. Moreover, we also obtain confidence intervals from the GP optimisation (Fig 5), which allows us to judge the quality of the Ca2+ spike rate estimation a posteriori.
Spike rates are often estimated from pooled data. This practice is well founded if cells generate spikes with the same mechanism. When cells are stimulated, this approach also assumes that each cell experiences the same stimulus time course. For ligand-dependent signalling pathways, the last condition is usually met experimentally by exposing cells to a constant stimulus. However, this might not be the situation in vivo. We therefore used a microfluidics chamber to challenge HEK293T cells with a time-varying concentration of carbachol. As Fig 7 illustrates, different cells experience different concentration profiles of carbachol. Importantly, when computing the average over all stimuli time courses, not a single cell experiences this specific stimulation. To identify cells that are stimulated in a similar manner and hence can be compared with each other, we computed the weights of the three leading principal components of each stimulus time course, and then used a k-means algorithm. We then determined the most likely Ca2+ spike rate for each cell and computed the mean Ca2+ spike rate for cells within a given group (Fig 7). Overall, there is substantial variability in the Ca2+ spike rates with respect to their mean. This results from the variation in the Ca2+ spike sequences, which is illustrated by a comparison between the first and the third group. This variability in the Ca2+ spike rates leads to intriguing questions. On the one hand, the fluctuations could arise from the intrinsic stochasticity of Ca2+ spike generation. It is known that cells challenged repeatedly with the same stimulus (and allowing for recovery between successive stimulation) respond randomly (see e.g. [16]). Hence, the observed Ca2+ spike sequences constitute a sample of the possible cellular responses given a particular stimulus. On the other hand, the variability could stem from the composition of the Ca2+ signalling apparatus in each cell. At the subcellular level, a Ca2+ spike often corresponds to a travelling Ca2+ wave that is shaped by Ca2+ release from intracellular storage compartments such as the endoplasmic reticulum and Ca2+ sequestration by Ca2+ pumps [30, 52–54]. The spatial arrangement of Ca2+ releasing channels, Ca2+ pumps and Ca2+ buffers strongly affects Ca2+ waves and therefore Ca2+ spikes [11]. Given that even genetically identical cells express different numbers of the components of the Ca2+ signalling toolkit and arrange them in different spatial patterns, the variability in the Ca2+ spike rates could reflect single cell variability at the molecular level. The two sources for the variability of Ca2+ spike rates are not mutually exclusive, and a mixture of both is most likely to occur in vivo.
To bring order to such disparate Ca2+ spike sequences, recent studies have shown that when cells are challenged with a constant stimulus, a linear relationship exists between the mean and the standard deviation of ISIs [12, 15]. The slope of this relationship was shown to be robust to interventions at the molecular level (blocking Ca2+ pumps, energising Ca2+ release channels) as well as being cell type and agonist specific. A theoretical analysis revealed that the slope could be determined by a recovery timescale from global cellular inhibition after a Ca2+ spike. Therefore, one expects that each cell type and each agonist can be characterised by this timescale. In the present study, each cell possesses its own intensity function x(t). The organising principle that will lead to a cell type and stimulation specific description of Ca2+ spiking is given by the parameter values that describe the statistics of x(t), i.e. the hyperparameters of the GP. Put differently, for a given cell type stimulated with a specific agonist and application protocol, we expect one set of parameter values for the GP. Heterogeneous cell responses then originate from different realisations of the GP. To illustrate this concept, assume for the time being that the intensity function is constant. In a population, there will be a spread of Ca2+ spiking behaviour, with some cells only generating a few Ca2+ spikes, while others exhibit high Ca2+ spiking activity. Consequently, the intensity function for slow spiking cells is low, while it is large for high frequency cells. For independent cells, as is the case in this study, it is reasonable to assume that the values of the intensity function are normally distributed. This is equivalent to saying that the intensity function of each cell is a realisation of a GP. The assumption of one set of hyperparameters for a given experiment and cell type will allow us to quantitatively compare different experiments and answer questions such as how different cell types respond to the same stimulus, or how different stimuli shape Ca2+ spike sequences in a given cell type. We expect that similar responses will be mirrored in small differences between hyperparameter sets.
The last point raises the question of how transferable results are from one cell type to another and from one stimulation scenario to another. In addition, the ultimate goal of the modelling framework presented here is to apply it to physiologically relevant tissues. For example, gap junctional coupling between cells, or paracrine signalling, could potentially influence population heterogeneity. This would require an extension of the modelling framework towards network dynamics, which may be a formidable challenge, as predicting the network behaviour from single node dynamics is nontrivial, let alone inferring single cell dynamics from within a connected tissue.
In this work we have developed a mathematical framework to quantitatively describe the heterogeneous timing of Ca2+ spikes in a cell population subject to time-varying stimulation. At the heart of this new approach is the use of Bayesian inference to determine the most likely intensity function and hence the most likely Ca2+ spike rate for a given stimulus. As part of this estimation process, we found that the statistics of Ca2+ ISIs are best captured by an IG distribution. Importantly, knowledge of the intensity function and the ISI statistics suffices to completely describe Ca2+ spiking. Since generating Ca2+ spike sequences from an ISI distribution and intensity function is computationally significantly cheaper than solving partial differential equations for cellular Ca2+ transport, this approach is ideally suited for numerically studying large numbers of cells.
The estimation of inhomogeneous single cell behaviour also puts us in an ideal position to ascertain whether or not there is signal processing at the cell population level. Indeed, numerous examples exist where the average population behaviour is not shared by any cell (see e.g. [55]). These incongruous dynamics also warrant investigations into population invariances, where cell populations respond consistently in the same manner, albeit with completely heterogeneous single cell behaviour [56, 57]. By reliably estimating single cell Ca2+ dynamics, the present study provides a stepping stone towards answering these questions for intracellular Ca2+ signalling.
We here follow the exposition in [37] for the definition of the intensity function. Assume that Ca2+ spikes occur at times y1 < y2 < … < yN. Let p(v) denote the probability density for a general renewal process on v ∈ (0, ∞), i.e. p(v)dv is the probability for an event in [v, v + dv], and subsequent events are independent. For ya > 0, let y correspond to a time variable on (ya, ∞) and X be a one-to-one mapping X(y) = v of (ya, ∞) to (0, ∞). Conservation of probability then entails that
p ( y ) = | d v d y | p ( v ) = | X ′ ( y ) | p ( X ( y ) ) . (2)
In other words, the probability density for a Ca2+ spike at yi can be computed from the renewal probability density p if we know the mapping X. A convenient form of X is
X ( y ) = ∫ y a y x ( u ) d u , (3)
which satisfies the conditions above and where x is called the intensity function, which is the object that we need to estimate. Eq (3) can be interpreted as rescaling the original time y such that Ca2+ ISIs become independent and identically distributed in the new time [58]. Given two subsequent Ca2+ spikes times yi−1 and yi in the original time, the ISI in the new time is
X ( y i - 1 , y i ) = ∫ y i - 1 y i x ( u ) d u . (4)
Since it is only through the introduction of the intensity function x(t) that Ca2+ ISIs become Markov, we introduce the notation p(yi, yi−1|x), which corresponds to the ISI probability density given x(t). Note that formally the conditional ISI probability density is defined as the joint conditional probability density for spikes at yi and yi−1 (and hence no spike in [yi−1, yi]) given an intensity function x(t). We will employ three different choices for the ISI probability density: an inhomogeneous Gamma distribution
p ( y i , y i − 1 | x ) = γ x ( y i ) Γ ( γ ) [ γ X ( y i − 1 , y i ) ] γ − 1 e − γ X ( y i − 1 , y i ) , (5)
where γ > 0 denotes the shape parameter and Γ is the Gamma function; an inhomogeneous inverse Gaussian distribution
p ( y i , y i − 1 | x ) = x ( y i ) 2 π X 3 ( y i − 1 , y i ) exp { − ( X ( y i − 1 , y i ) − α ) 2 2 α 2 X ( y i − 1 , y i ) } , (6)
where α > 0 is the location parameter; and an IP distribution
p ( y i , y i - 1 | x ) = x ( y i ) e - X ( y i - 1 , y i ) . (7)
The time-dependent intensity function x(t) is modelled as a Gaussian Process (GP) [34, 59]. A GP is uniquely defined by its mean μ(t) and covariance function Σ(t1, t2). While there are many possible choices for Σ [34, 60], we employ the widely used squared exponential (SE) kernel
Σ ( t 1 , t 2 ) = σ f 2 e - κ ( t 1 - t 2 ) 2 2 + δ ( t 1 - t 2 ) σ v 2 , (8)
where κ measures the smoothness of the GP and σf controls its variance. The last term allows us to model additional noise sources. We originally included σ v 2 as a hyperparameter in the optimisation. However, we consistently found small values for σ v 2 and hence decided to fix it at a presentative value of σ v 2 = 10 - 4. We collect the spike times in a sequence of N Ca2+ spikes in a vector y = {y1, …, yN}. For consistency, we set y0 = 0. Through the introduction of an intensity function x(t), the joint probability density for a spike sequence y given x(t) factorises and reads as [36]
p ( y | x ) = p 1 ( y 1 | x ) p T ( T , y N | x ) ∏ i = 2 N p ( y i , y i - 1 | x ) . (9)
Here, p1(y1|x) represents the conditional probability density of finding the first spike at time y1. We also take into account that the observation time T usually exceeds the last spike time through the term pT(T, yN|x), which denotes the conditional probability that no spike occurs after yN. The statistics for p1 and pT are often based on an inhomogeneous Poisson (IP) process, i.e.
p 1 ( y 1 | x ) = x ( y 1 ) e - X ( 0 , y 1 ) , p T ( T , y N | x ) = e - X ( y N , T ) , (10)
where X is given by Eq (3).
For practical purposes, we discretise time with a time step Δ such that T = nΔ [36]. When working with experimental spike trains, we set Δ equal to the inverse of the recording frame rate. A spiking time yi can then be expressed as yi = li Δ for an appropriate l i ∈ N. By setting xi = x(iΔ) and using Eq (9) with e.g Eq (5), we obtain the probability density for a spike sequence for the inhomogeneous Gamma distribution as
p ( y | x , θ ) = x l 1 e − X ^ 0 , 1 e − X ^ N , n ∏ i = 2 N γ x l i Γ ( γ ) [ γ X ^ i − 1 , i ] γ − 1 e − γ X ^ i − 1 , i , (11)
where X ^ i , j = Δ ∑ k = l i l j x k and l0 = 0, ln = n. By introducing θ on the left hand side, we make explicit the dependence of the probability density on the hyperparameters θ, which in this case are θ = {γ, κ, σf}.
The most probable intensity function x*(t) given a spike train y is determined by x* = argmaxx≥0 p(x|y). Under the assumption that the nodal value x* is close to its mean, we have
x * ≈ ∫ x θ * p ( θ | y ) d θ = 1 Z ∫ x θ * F ( y , x θ * , θ ) d θ , (12)
where
x θ * = argmax x ≥ 0 p ( x | y , θ ) = argmax x ≥ 0 p ( y | x , θ ) p ( x | θ ) . (13)
To evaluate the first integral in Eq (12) we note that
p ( θ | y ) = p ( θ ) p ( y ) ∫ p ( y | x , θ ) p ( x | θ ) d x = F ( y , x θ * , θ ) p ( y ) , (14)
with F ( y , x θ * , θ ) = p ( θ ) p ( y | x θ * , θ ) p ( x θ * | θ ) / | Λ * + Σ - 1 | and Λ * = - L x 2 log p ( y | x θ * , θ ), where we used Laplace’s approximation for the integral as shown in S1 Appendix. We further introduced the notation L x 2 f to denote the Hessian of f with respect to x(t) and Z = ∫ F ( y , x θ * , θ ) d θ = p ( y ) / ( 2 π ) n / 2.
Let q(t|yk, x), t > yk denote the conditional intensity function, i.e. q(t|yk, x)dt is the probability for a spike in [t, t + dt] given an intensity function x(t) and the last spike at yk. We can express q(t|yk, x) in terms of the ISI probability density as [31]
q ( t | y k , x ) = p ( t , y k | x ) 1 - ∫ y k t p ( s , y k | x ) d s . (15)
The time rescaling theorem then states that the rescaled ISIs [31, 47, 61, 62]
τ k = ∫ y k - 1 y k q ( s | y k - 1 , x ) d s , (16)
are independent and identically distributed exponential random variables with mean one if y is a realisation from a point process with conditional intensity function q(t|yk, x).
Suppose there are K rescaled ISIs. For a quantile-quantile plot [46], we order the τk in ascending order giving rise to the new ISIs τ ˜ n. We then plot the quantiles of the distribution of the τ ˜ n against the quantiles of an exponential distribution with unit rate, which are given by τ ^ n = - ln ( 1 - s n ) with sn = (n − 0.5)/K.
For the Kolmogorov-Smirnov, plot [62], we define the random variable u k = 1 - e - τ k and then plot the ordered set of the uk against the cumulative distribution function of the uniform distribution, i.e. F(x) = x for 0 ≤ x ≤ 1, sampled at sn.
The Ca2+ spike rate is estimated from m spike sequences via kernel smoothing (KS) through [49, 63]
r = 1 m ∑ j = 1 m ∑ i = 1 N j f ( t - y i j , σ ) , (17)
where y i j denotes the ith spike time in the jth Ca2+ spike sequence yj, and Nj is the total number of spikes in yj. The function f represents the kernel, and we chose a Gaussian of the form
f ( t , σ ) = 1 2 π σ 2 exp ( - t 2 2 σ 2 ) . (18)
The parameter σ is referred to as the bandwidth of the kernel. In case we work with a large number of independent Ca2+ spike sequences yj, we can use an optimal bandwith [49, 50]. To evaluate how well a given method (e.g. Bayesian inference or KS) approximates the true Ca2+ spike rate used to generate surrogate data, we evaluated the normalised L2 norm as
L 2 = [ ∫ 0 t ( r ^ ( t ) − r ˜ ( t ) ) 2 d t ] 1 / 2 [ ∫ 0 t r ˜ ( t ) d t ] − 1 , (19)
where r ˜ and r ^ denote the known and estimated Ca2+ spike rate, respectively.
We arrange the stimuli experienced by individual cells in a matrix X such that each row corresponds to a single stimulus time course. We then compute the singular value decomposition of X, i.e. X = UΣVt, where t denotes transposition. The columns of V correspond to the eigenvectors of XtX, and Σ is a diagonal matrix that holds the singular values of X. The weights of the principal components of the stimuli time courses are the rows of XV = UΣ.
The k-means algorithm requires the number k of clusters as input and then determines the members of each cluster by minimising the error function [64]
E = ∑ i = 1 k ∑ x ∈ C i ‖ x - μ i ‖ 2 . (20)
Here, x are the data points, C1, …, Ck are the k disjoint clusters and μi is the centroid of the ith cluster. We varied k and visually inspected the clustering. For consistency, we also clustered the data using other algorithms such as mean shift, spectral clustering and density-based spatial clustering of applications with noise. While there were minor differences between the suggested clusters, the overall clustering structure remained the same.
To see which is the most likely ISI statistics, we apply the following protocol to every single cell from the experiment shown in Fig 1:
To test the performance of Ca2+ spike rate estimation, we generated surrogate data from an IG for the two different intensity functions xdet(t) and xGP(t) using inverse sampling, a Bernoulli process based on the conditional intensity function in Eq (15) and time rescaling [47, 65, 66].
A key factor in estimating Ca2+ spike rates from PSTHs and KS is the choice of a bin width and bandwidth, respectively. For a large number of Ca2+ spike sequences, optimal estimates exist [49–51], and we use them for Fig 4. In case of only a few Ca2+ spike sequences with a small number of spikes per sequence, as in Fig 5, no estimates for a bin width or bandwidth exist. We therefore employed a bandwidth that was approximately equal to the optimal bandwidth determined in Fig 4 as well as bandwidths 1.5 and 2 times larger than this. In addition, we used the same formal expression as for the optimal value, which resulted in the bandwidth σ ^. Note that σ ^ differs from the optimal bandwidth in Fig 4, since it explicitly depends on the number of Ca2+ spike sequences.
A bespoke perfusion system connected to a 3-port microfluidics device [67] was used to expose cultured HEK293T cells to varying concentrations of the muscarinic receptor agonist, carbachol. The HEK293T cell line was a gift from Dr N. Holliday, University of Nottingham, that had been frozen after passage 28 of the original stock. After thawing, cells were used for up to a further ten passages. Cells were seeded at a density of 105 cells/ml in the central micro-channel of the microfluidic devices, in DMEM D6429 growth media (Invitrogen, Paisley, UK) containing 10% fetal calf serum. Cells were loaded inside the microchannels with 1 μM of the Ca2+ indicator Fluo5F-AM for 30 min, followed by washout with imaging buffer (135 mM NaCl, 3 mM KCl, 10 mM HEPES, 15 mM D-glucose, 2 mM MgSO4 and 2 mM CaCl2) for at least a further 30 min. To stimulate the cells, the flow rates of two inlet channels into the microchannel were varied, allowing the interface between the two solutions to be shifted laterally across the chamber. One inlet stream contained the agonist (100 μM carbachol) and Alexa Fluor 594 (AF594, 2 nM; to allow monitoring of agonist concentration in proportion to AF594 fluorescence). The second inlet contained buffer alone. The interface formed between the two solutions due to laminar flow was shifted across the width of the microchannel by controlled changes in the fractional flow rates for each stream, with total flow being constant. In combination with the shifting interface position, the concentration gradient formed by diffusional collapse of the interface as the co-flow progresses through the channel length results in a spatiotemporal gradient in agonist concentration throughout the channel. This method enables the exposure of cells to pre-defined, time-varying changes in agonist concentration, from simple step-changes to complex waveforms. During dynamic stimulation with agonist, AF594 and Fluo-5F AM indicators were excited sequentially (100 ms exposure, 1 Hz frame rate) using a pE2 LED system (excitation peaks 470 nm and 565 nm; CoolLED, Andover UK). Emission was detected at 535 ± 50 nm and 565 ± 20 nm with an ORCA-R2 camera (Hamamatsu, Welwyn Garden City, UK).
A time-series analyser plugin to ImageJ (Wayne Rasband, National Institutes of Health, Bethesda, MD, available at http://rsb.info.nih.gov/ij) was used to manually define circular regions of interest (ROI) centred on each cell. Mean Fluo-5F emission intensity of pixels falling within each ROI was quantified and expressed as the ratio of fluorescence at time t divided by mean intensity from a 25 s window prior to the first increase in stimulus concentration (F/F0). The baseline window is selected as the window with minimum standard deviation from sliding 25 s windows taken from 0 to 120 s (before increase in stimulus concentration). Fluorescence of AF594 was quantified as the mean fluorescence intensity of pixels falling within each ROI being quantified; therefore each cell has a Ca2+ response measure and an associated stimulation profile.
|
10.1371/journal.ppat.1002375 | A TNF-Regulated Recombinatorial Macrophage Immune Receptor Implicated in Granuloma Formation in Tuberculosis | Macrophages play a central role in host defense against mycobacterial infection and anti- TNF therapy is associated with granuloma disorganization and reactivation of tuberculosis in humans. Here, we provide evidence for the presence of a T cell receptor (TCR) αβ based recombinatorial immune receptor in subpopulations of human and mouse monocytes and macrophages. In vitro, we find that the macrophage-TCRαβ induces the release of CCL2 and modulates phagocytosis. TNF blockade suppresses macrophage-TCRαβ expression. Infection of macrophages from healthy individuals with mycobacteria triggers formation of clusters that express restricted TCR Vβ repertoires. In vivo, TCRαβ bearing macrophages abundantly accumulate at the inner host-pathogen contact zone of caseous granulomas from patients with lung tuberculosis. In chimeric mouse models, deletion of the variable macrophage-TCRαβ or TNF is associated with structurally compromised granulomas of pulmonary tuberculosis even in the presence of intact T cells. These results uncover a TNF-regulated recombinatorial immune receptor in monocytes/macrophages and demonstrate its implication in granuloma formation in tuberculosis.
| Infection with mycobacteria results in a host response which results in the formation of granulomas, highly organized structures characterized by the presence of macrophages, which are considered to rely solely on invariant immune receptors. On the other hand, the presence of variable immune receptors is required for granuloma formation but this process is not solely dependent on T cells. Furthermore, TNF is required for the maintenance of the mycobacterial granuloma structure in humans. We now find evidence for subpopulations of human and mouse macrophages that express variable αβ T cell receptors (TCRαβ). Engagement of the macrophage-TCRαβ triggers CCL2 release and phagocytosis of baits directed to this receptor is enhanced. TCRαβ bearing macrophages accumulate in human tuberculosis granulomas and anti-TNF treatment of macrophages results in downregulation of the TCRαβ, which is associated with caspase 3 cleavage and suppression of TCRξ. Anti-TNF treatment reduces mycobacteria induced cluster formation of TCRαβ positive macrophages, which is in line with reduced granuloma formation in rag1–/–(T cell rag1+/+) and TNF–/–(T cell TNF+/+) chimeric mice. Consequently, both chimeras show reduced CCL2 staining after mycobacterial infection. In summary, we have identified a recombinatorial immunoreceptor in monocytes/macrophages and demonstrate its implication in mycobacterial infection.
| Macrophages are key players in major chronic inflammatory diseases including tuberculosis, atherosclerosis and rheumatoid arthritis. Based on their myeloid origin and professional phagocytic activity they are traditionally regarded as a pillar of innate immunity [1]. Tuberculosis is an infectious disease that in 2008 afflicted more than nine million individuals worldwide and claimed the lives of an estimated 1.3 million patients [2]. The disease is caused by mycobacteria that are efficiently contained by macrophages in highly organized immune structures, the tuberculous granulomas. Ample evidence indicates that the generation and maintenance of tuberculous granulomas require TNF [3], [4]. Moreover, reactivation of the disease by therapeutic TNF blockade is associated with disruption of the granuloma architecture that ultimately leads to spreading of the mycobacteria into the surrounding tissue [5].
Within the tuberculous granuloma, cellular immunity to mycobacteria is thought to be solely under the direction of T cells which orchestrate the macrophage host response to the pathogen. However, selective T cell depletion and reconstitution experiments in murine models of tuberculosis point to the involvement of variable host defense mechanisms in the control of mycobacterial infection beyond T cells [6]-[8]. The recent demonstration by our laboratory and others that neutrophils and eosinophils express T cell receptors (TCR) which are generated by V(D)J recombination has provided evidence for the existence of variable immune receptors outside lymphocytes [9]-[11].
These findings and the possibility that variable immune defense mechanisms outside T cells are implicated in the development of the tuberculous granuloma raise the question as to whether macrophages possess a molecular machinery for variable host defense. Here, we report that subpopulations of monocytes and macrophages express a recombinatorial TCRαβ which is TNF regulated and demonstrate a role of this novel immune receptor in the macrophage host response to mycobacteria and the formation of the tuberculous granuloma.
To assess the possibility that monocytes, like granulocytes [9]-[11], express the TCRαβ, we isolated human CD14+ monocytes from healthy donors (n = 12). Expression of the TCRαβ in peripheral blood monocytes was assessed in MACS-CD14+ purified cells by immunocytochemistry using antibodies to TCRα/TCRβ and MHC-II. Utilizing this approach, we consistently detected a ∼5% cell fraction that displayed bright TCRαβ+ expression in freshly isolated CD14+ monocytes which showed co-expression of MHC-II (Figure 1A). Purity of isolated CD14+ monocytes was routinely >99.5% as determined by flow cytometry (Figure S1A). We next characterized the TCRαβ expressing monocyte subpopulation in PBMC by flow cytometry. Consistent with immunocytochemistry, CD14+ cells from three normal subjects displayed positive staining for TCRβ in a 3–4% subfraction which did not exhibit staining for the T cell marker CD3 (Figure 1B, Figure S1B). We then determined expression of the ΤCRαβ in monocyte-derived macrophages from three healthy donors. For this, monocytes were differentiated into naïve, IFNγ activated and IL-4 activated macrophages, respectively, on glass slides for a period of 6 days [12] and stained for TCRαβ. To quantify the TCRαβ+ populations in adherent macrophages, laser scanning cytometry (LSC) was performed on the slides that were immunostained with Alexa 555-labeled secondary antibodies. LSC analysis demonstrated that a 5% subpopulation of naïve macrophages exhibited high fluorescence indicative of TCRαβ expression (Figure 1C, Figure S1C). In contrast, the fraction of TCRαβ bearing macrophages was significantly higher in the IL-4 (9%) and IFNγ activated macrophages (11%). Thus, activation of macrophages induces an increase in the subpopulation of TCRαβ expressing macrophages.
Immunoblot analysis confirmed the expression of the TCR α- and β-chains in monocytes and macrophages (Figures 1D). In line with this, immunogold electron microscopy using an antibody to the TCR α-subunit revealed specific staining in monocyte-derived macrophages (Figure 1E), which was detectable on the cell surface. Next we looked for evidence of TCRαβ expressing macrophages in vivo. For this, we performed double-immunostaining for the macrophage marker CD163 and the TCRαβ in bronchoalveolar lavage (BAL) fluid obtained from three individuals with normal BAL cytology. In fact, we found that a 5–15% subpopulation of alveolar macrophages showed positive staining for the TCRαβ (Figure 1F, Figures S1D and E) demonstrating that tissue macrophages are capable of expressing the TCR under physiological conditions.
We then tested whether the TCR is also expressed in macrophages from the mouse. Consistent with our findings in humans, immunocytochemistry demonstrated the presence of the TCRαβ in a 5–10% subfraction of spleen macrophages from C57BL/6J mice (n = 3) and RT-PCR revealed mRNA expression of both the TCR α- and β-constant chains in splenic macrophages (Figure 1G).
Together, these results reveal that subpopulations of peripheral blood monocytes and in vitro activated monocyte-derived macrophages constitutively express the TCRαβ. Moreover, they demonstrate the presence of ΤCRαβ bearing macrophages in normal human tissue, as exemplified for the lung, and provide evidence for TCRαβ expression by murine macrophages.
We next investigated whether the TCRαβ expressed by monocytes/macrophages represents a variable receptor. For this, we tested whether CD14+ monocytes and IFNγ macrophages have a rearranged TCRβ locus. Using a PCR assay based on the protocols established by van Dongen et al. [13] in combination with sequencing of specific amplification products, evidence for TCRβ locus genomic Dβ1→Jβ (Figure 2A i) and Vβ1→ Jβ (Figure 2A ii) recombination was found in both the monocyte and IFNγ macrophage fractions. Length spectratyping of the antigen-binding complementary determining region 3 (CDR3) is a well-established method for the assessment of TCR repertoire diversity in defined variable chains [14]. Representative Vβ13a CDR3 spectratype analysis in IL-4 or IFNγ primed macrophages from healthy individuals (n = 3) revealed monoclonal and oligoclonal repertoires and varied in the same donor depending on IL-4 or IFNγ activation (Figure 2B). Sequencing of the expressed Vβ13a CDR3β clonotypes in one subject (GenBank Acc. No. JF923737-JF923744) indeed revealed marked differences between IL-4 and IFNγ polarized macrophages (Figure 2C). Of note, quantitation of the expressed length variants, respectively, in all three individuals consistently demonstrated increased repertoire TCR Vα and Vβ diversity in IFNγ activated macrophages relative to monocytes and IL-4 macrophages (Figure S2A).
Evidence for TCRβ locus rearrangement in mature CD14+ monocytes and our previous observation of a rearranged TCRαβ in neutrophils [9] strongly suggested that TCR recombination occurs already at an early stage of myeloid development. To test this possibility, burst-forming unit-erythroid (BFU-E) and granulocyte/macrophage progenitor colonies (CFU-GM) were generated from CD34+ hematopoietic progenitor cells of two normal donors in two independent experiments. TCR Vβ mRNA expresssion profiling was performed on 10 randomly selected colonies from each individual. CDR3β length spectratyping revealed expression of single or few rearranged Vβ clonotypes in 50% (donor A) and 70% (donor B), respectively, of the CFU-GM analyzed (Figure 2D, Figure S2B). No TCR gene expression was observed in any of the BFU-E colonies tested (data not shown). The majority of the CFU-GM displayed a monoclonal expression pattern consistent with the clonogenic nature of the myeloblasts and monoblasts in this assay. This indicates that TCRβ locus rearrangement and expression of individual-specific Vβ repertoires occurs already during the early phase of in vitro myeloid lineage differentiation.
We next sought evidence for TCR α- and β-chain variability at the protein level. For this, the TCRαβ from macrophages of healthy donors was immunoprecipitated and subjected to MALDI-TOF mass spectrometry. Using this proteome profiling approach, we identified two peptides that showed partial sequence identity with known variable TCR α-chain fragments (Jα4, Jα39) and a total of four distinct peptides displaying partial sequence identity with variable TCR β-chain fragments (Jβ1.1, Jβ1.2, Jβ1.4, Jβ2.1) (Figure 2E). Three of the peptides (Jα4, Jβ1.4, Jβ2.1) spanned V→ J and J→ C junctions indicating that they originated from rearranged TCR α and β loci. These TCRαβ proteome profiling results are consistent with the presence of multiple TCR α- and β-chain variants in human macrophages.
In mice, TCR Vαβ mRNA profiling and CDR3 spectratyping of purified peritoneal macrophages confirmed the presence of diverse Vα and Vβ repertoires in wildtype macrophages. In contrast, no evidence for TCR Vαβ repertoire expression was found in macrophages from rag1–/– mice which are incapable of rearranging their immune receptor loci and thus lack a variable TCRαβ (Figure 2F).
In summary, the combined results from TCRβ VDJ rearrangement analyses, Vαβ CDR3 mRNA expression profiling and mass spectrometric TCRαβ peptide profiling indicate that the TCRαβ identified in human and murine macrophages is expressed as a variable receptor.
Given the presence of the TCRαβ ligand binding subunits in monocytes/macrophages, we tested whether these cells also express constituents of the TCR signaling pathway. RT-PCR expression profiling revealed expression of all critical components of the TCR signal transduction machinery including CD3ζ, ZAP70, LAT, Fyn, and Lck, respectively, in monocytes and macrophages from three randomly selected healthy donors (Figure 3A). To explore whether the TCRαβ complex is operative in macrophages, we next tested whether specific TCR engagement has an impact on the secretion of effector molecules by IFNγ macrophages. Analysis of the secretory pattern of a defined panel of cytokines, chemokines and growth factors (n = 15, Table S1, Figure S3) in response to canonical TCR stimulation with anti-CD3 antibodies revealed enhanced secretion of the major monocyte chemoattractant CCL2 (MCP-1) within 24 h (Figure 3B). No detectable effect was observed for any of the other studied macrophage effector proteins or the secretory protein CCL5, which served as a marker for potential T cell contamination, indicating that engagement of CD3 dependent macrophage-TCRαβ signaling triggers selective secretion of CCL2 (Table S1, Figure S3). Consistent with this, anti-CD3 antibodies induced CCL2 gene expression in IFNγ macrophages (Figure 3C). Thus, specific engagement of the TCRαβ in macrophages induces gene expression and secretion of the monocyte chemoattractant CCL2 indicating that the TCRαβ signal transduction pathway in macrophages is functional.
Next, we investigated whether the TCRαβ interferes with the phagocytic activity of macrophages. For this, IFNγ activated macrophages from two healthy donors were challenged with standardized phagocytosis baits (polystyrene beads, Ø 4.5 µm) for 15 min, 1 h and 10 h, respectively. To induce physical interaction of the baits with the macrophage-TCRαβ, the beads were coated with anti-TCRα/anti-TCRβ antibodies (Figure 4A). Identical beads coupled to equal amounts of nonspecific IgG antibodies or albumin (irrelevant protein) were used as controls. In addition, macrophages were challenged with albumin-coated beads in the presence of anti-TCRαβ antibodies that were not physically coupled to baits. As a positive control, baits were targeted to the known mediator of phagocytosis complement receptor 3 (CR3) utilizing antibodies to its subunit CD11b. Using this approach, we observed in both donors that the number of phagocytosing macrophages was significantly increased after 1 h when baits were directed to the TCR (1.4–3.0 fold vs. controls) (Figure 4B). This increase was already detectable after 15 min, however, did not reach statistical significance for all controls. A similar augmentation of phagocytosis was seen when beads were targeted to the CR3 (1.2–4.5 fold vs. controls). In addition, we found that after 10 h baits directed to the TCR had elevated bead/cell ratios relative to controls (1.3–1.8 fold). These findings suggest that binding of baits to the TCRαβ facilitates phagocytosis. Consistent with this, we noted that phagocytosis was unaffected when anti-TCRαβ antibodies were not physically linked to beads (Figure 4B,C). As expected, we found evidence for close proximity of ingested baits to the TCRαβ (Figure 4C).
In light of the observation that the TCR facilitates phagocytosis, we next investigated whether and to which degreee complete ablation of the TCR has a negative effect on the macrophage phagocytic capacity. For this, peritoneal macrophages from recombination defective rag1–/– mice (n = 7), which lack the TCRαβ (Figure 2F) were incubated with FITC-labeled Mycobacterium bovis Bacille-Calmette-Guérin (BCG). We used mycobacteria as baits because they represent classical macrophage pathogens and, like CCL2, are critically implicated in macrophage-driven granuloma formation [15]. Consistent with the findings in the bead targeting experiments phagocytosis of M. bovis BCG was significantly reduced in rag1–/– macrophages compared to recombination-competent rag1+/+ control mice (n = 7) after 6 hours of infection (Figure 4D). Together, these bait targeting and ablation experiments strongly suggest roles of the TCRαβ in the regulation of macrophage phagocytic activity.
Phagocytosis of mycobacteria by macrophages with subsequent granuloma formation are key features of host defense in tuberculosis infection. Given this and the above results suggesting implication of the macrophage-TCRαβ in phagocytosis, we next tested whether mycobacterial challenge has an impact on TCR expression in macrophages in vitro. We infected macrophages with FITC-labeled M. bovis BCG and noted that within 4–6 days macrophages routinely formed clusters (Figure 5). No clusters were formed in the absence of mycobacteria (Figure 5B). Bacilli frequently formed aggregates which may reflect fragmentation of M. bovis BCG. We found evidence for co-localization of BCG and the TCRαβ by immunocytochemistry (Figure 5A i) and immunogold electron microscopy (Figure 5A ii), however, this phenomenon was rare. Quantitative analysis of BCG-infected IFNγ macrophages from two healthy donors revealed a 4-fold increase of the percentage of TCRαβ+ macrophages relative to uninfected controls (8 vs. 32.5) (Figure 5B, Figure S4A). Consistent with this, we found increased expression of both the TCR α and the β constant chain genes in the macrophages that were infected with BCG. Moreover, CDR3 spectratyping of all 25 TCR Vβ chains showed a noticeable but not significant increase in the number of expressed Vβ CDR3 length variants (12 vs 18.5) in BCG-infected macrophages (Figure 5C). Ex situ Vβ clonotype analysis of randomly selected BCG/macrophage clusters from both donors revealed consistent expression of highly restricted TCR Vβ chain repertoires (Figure 5D, Figure S4B). In particular, we noted a bias toward the use of the Vβ1 chain which was expressed by 10 out of the 11 (91%) BCG/macrophage clusters analyzed. Collectively, these results demonstrate that in vitro M. bovis BCG infection induces formation of TCRαβ bearing macrophage clusters that express highly restricted Vβ repertoires.
To examine whether the macrophage-TCR is implicated in host defense against mycobacteria in vivo, we next screened for the presence of TCRαβ expressing macrophages in tuberculous tissue. Lung sections of patients with pulmonary tuberculosis (n = 13, Table S2) were immunostained for TCRαβand the macrophage markers CD68 and CD163, respectively. Ten out of 13 patients showed abundant staining for TCRαβ in well-circumscribed caseous granulomas. Typically, the innermost segment of the epithelioid cell corona, which represents the front line of cellular defense against mycobacteria within tuberculous granulomas, exhibited intense TCRαβ+ staining (Figure 6A i-iii). Controls showed no staining for CD2 (Figure 6A iv,v). Quantitative analysis of TCRαβCD68 immunofluorescence double-staining revealed that on average 87% of the macrophages expressed the TCRαβ in this zone (Figures 6A vi, vii). Additional single immunofluorescence staining for CD68 and CD163 confirmed that the predominant cell type in the inner host-pathogen contact zone was macrophages (Figure 6B, Figure S5A). Consistent with this, T cells and NK cells were typically localized in the peripheral corona zone of caseous granulomas as assessed by CD2 immunostaining (Figure 6B). We found no evidence for TCRαβ expressing macrophages in a lymph node from an individual with reactivated M. tuberculosis infection triggered by anti-TNF therapy (adalimumab) (Figure S5B). In keeping with immunohistochemistry, ex situ clonotype analysis revealed expression of TCR Vβ mRNA repertoires in small clusters (20–30 cells) of immunostained CD68+ macrophages that were laser microdissected from the inner epithelioid cell zone (Figure 6C). Similarly as observed in BCG infected macrophage clusters in vitro (Figure 5D), we noted that the epithelioid zone macrophages predominantly express the TCR Vβ1 chain.
The cytokine TNF is essential for host defense against mycobacteria and anti-TNF therapy may lead to disorganization of human tuberculous granuloma resulting in reactivation of latent tuberculosis [3], [16]–[19]. Since activation of the TCRαβ in macrophages results in CCL2 release, a key factor in granuloma formation, we examined whether TNFinhibition has a direct impact on the expression of the macrophage-TCRαβ. Uninfected or BCG infected macrophages were incubated in the presence of the anti-TNF antibody infliximab (50 µg/ml) or an isotype control antibody (anti-CD20, rituximab). Immunostaining revealed that TNF blockade inhibited TCRαβ expression relative to controls within 2h (Figure 7A). Consistent with this, the inhibitory effect could be completely reversed by re-exposure of macrophages to TNFfor 24 h indicating that macrophage-TCRαβexpression requires the presence of TNF (Figure S6A). TNF blockade also had an inhibitory effect on TCR β-chain mRNA expression (Figure 7B). Immunoblot revealed that TNF blockade not only suppressed expression of the TCRαβ ligand binding subunit but also that of the ζ-subunit (CD3) of the TCR signaling complex (Figure 7C). The latter is essential for TCRαβ stabilization on the cell surface [20] and its degradation is mediated by caspase 3 [21]. Immunoblot analysis and immunostaining showed that anti-TNF treatment results in an increase in cleaved caspase 3 in uninfected and BCG infected macrophages indicating that TNF blockade induces cleaved caspase 3 (Figure 7D, Figure S6B). Because TCRζ is required for stabilizing TCRαβ on the cell surface these findings identify TNF as a regulator of macrophage-TCRαβ expression.
It is well-established that abrogation of TNF results in defective tuberculous granulomas [7], [8], [22]. Given this and our finding that BCG induces TCRαβ expression, we next tested whether TNF blockade affects BCG triggered macrophage cluster formation in vitro. We found that treatment with the anti-TNF antibody infliximab significantly reduced the number and size of human macrophage clusters that formed during infection with M. bovis BCG in vitro (Figure 8A). The presence of the TCRαβ in mouse macrophages offers the possibility to study the role of the TNF/ TCRαβ/ CCL2 regulatory axis we identified in human macrophages in tuberculosis in vivo. For this, we determined whether deletion of the macrophage-TCR impacts granuloma formation and macrophage CCL2 release in a murine model of M. tuberculosis infection. Wildtype (wt) mice (n = 5) and TCR deficient rag1–/– mice (n = 7) that were reconstituted with wt CD3+ T cells were infected via aerosol with ∼100 M. tuberculosis bacilli [7]. The latter chimeric rag1–/–(T cell wt) mice develop lung tuberculosis in the presence of intact T cells but absence of TCR bearing macrophages and T lymphocytes are routinely detectable in the tuberculomas [8]. Four weeks post adoptive T cell transfer and M. tuberculosis infection all wt mice displayed compact, well-delineated granulomatous lesions in their lungs that were predominantly composed of macrophages and lymphocytes (Figure 8B i,ii). Rag1–/–(T cell wt) chimeras lacking the TCR in their macrophages developed granulomatous foci containing macrophages and lymphocytes as seen in the wildtype mice. However, these lesions were generally diffuse (Figure 8B iii,iv) and on average 1.5 fold larger than those of control mice indicating that rag dependent mechanisms outside the T cell system are required for proper granuloma formation in murine tuberculosis (Figure 8C). Importantly, immunostaining revealed abundant CCL2 staining in the tuberculous lesions of wt mice (Figure 8B i,ii) but CCL2 was routinely near absent in the chimeras lacking the macrophage-TCR (Figure 8B iii,iv). Collectively, these results demonstrate in vivo that ablation of the variable macrophage immune receptor in murine lung tuberculosis is associated with suppression of CCL2 and defective granulomas formation.
We finally tested the impact of TNF abrogation on CCL2 expression and tuberculoma formation in TNF deficient mice that were reconstituted with wildtype CD3+ T cells (TNF–/–(T cell wt) mice). Infection with M. tuberculosis infection in these chimeras resulted in disorganized tuberculous granulomas in which CCL2 was consistently absent (Figure 8B v,vi). Similarly as in the rag1–/–(T cell wt) chimeras, the granulomatous foci in TNF–/–(T cell wt) mice displayed a consistent increase in size (1.7 fold vs. controls) (Figure 8C). This together with our observation that TNF blockade inhibits expression of the macrophage-TCR strongly suggests that an intact TNF/TCR macrophage pathway is required for CCL2 production in the tuberculous granuloma. Consistent with this, we found no CCL2 expression in the tuberculous lymphnode of a patient who received therapeutic anti-TNF treatment (adalimumab) (Figure S7).
In this study, we report the existence of an as yet unrecognized recombinatorial TCRαβ based immune receptor in monocytes/ macrophages (macrophage-TCRαβ) and provide evidence for its implication in a major infectious disease - tuberculosis. We find that monocytes and macrophages have rearranged Vβ gene loci and consistently express diverse TCR Vαβ repertoires and show that macrophage-TCRαβ engagement induces the release of the monocyte chemoattractant CCL2 and demonstrate that the expression of the novel variable immune receptor depends on TNF. Furthermore, blockade of TNF in macrophages results in caspase 3 cleavage, TCRζ degradation and TCRαβ downregulation.
Traditionally, variable immune receptors are thought to be restricted to cells of lymphoid origin [1]. However, recent work from our laboratories and others demonstrating TCRαβ and TCRγδ expression in neutrophils and eosinophils [9]–[11], antigen receptor gene rearrangement in thymic granulocytes of mice [23], and NK cells which display immunological memory in response to viral infection [24] challenge this longstanding concept. These findings, which identify lymphoid features in myeloid cells, and the reciprocal demonstration that lymphocyte progenitors retain myeloid potential [25] and B cells in primitive vertebrates possess phagocytotic capabilities [26] extend the current concept of the strict dichotomy of the vertebrate immune system into non-specific/ myeloid and antigen-specific/ lymphoid immunity [1] and suggest a closer kinship between both lineages than commonly appreciated.
Combined in vitro evidence demonstrates a role for the macrophage-TCRαβ in the modulation of phagocytosis and the macrophage response during the initial phase of mycobacterial infection. In vivo, we find that the host/pathogen interface of human pulmonary tuberculomas is characterized by the massive presence of TCRαβ bearing macrophage-derived epithelioid cells and deletion of the macrophage-TCR in murine tuberculosis is associated with disorganized granuloma structure. Murine studies have previously suggested the possibility that variable immune receptors oustide the realm of T cells are involved in the control of mycobacterial infection [6]–[8]. Our demonstration that a TNF-dependent variable macrophage immune receptor is part of the host defense against mycobacteria provides evidence for this concept. In particular, our results suggest that macrophage based antigen-specific host response mechanisms are operative in the development of the tuberculous granuloma and thus add an all new aspect to the current understanding of tuberculoma formation (Figure 9). Furthermore, since it is well established that TNF is indispensable for the proper formation of tuberculous granulomas [7], [8], [22], [27], the finding that TNF blockade leads to suppression of the macrophage-TCRαβ provides a novel potential mechanism that may underly the reactivation of tuberculosis during therapeutic anti-TNF treatment [5].
TNF dependent regulation of the macrophage-TCR implies the involvement of the TNF receptor 1 and/or TNF receptor 2 signaling pathways in the control of expression of the novel immune receptor. At this point, it is unclear at which level the TNF receptor signaling cascade interferes with the macrophage-TCR machinery. It will be challenging to identify signaling components that connect TNF to the TCR in macrophages. Also, it will be interesting to see if TCR activation has an impact on TNF expression.
Given the near-ubiquitous presence of macrophages and their involvement in acute and chronic inflammation, it is likely that the macrophage-TCRαβ is implicated in further pathologies beyond tuberculosis. Candidate diseases include chronic inflammatory diseases such as atherosclerosis and autoimmune disorders such as rheumatoid arthritis. Efforts are underway to further explore the biological function of the recombinatorial macrophage receptor in the immune response.
Human monocytes/macrophages were isolated as previously reported [28] and purified by CD14+ Magnetic Cell Sorting (MACS) (Miltenyi Biotec, Bergisch Gladbach, Germany) Monocyte cell purity was routinely >99.5% as assessed by flow cytometry (Figure S1A) using the following antibodies: anti-CD14-FITC (DAKO, Glastrup, Denmark), anti-CD3-PE (BD Biosciences, Franklin Lakes, USA), anti-CD2-PE (BD Biosciences), mouse IgG1-FITC Isotype control and mouse IgG1-PE Isotype control. CD14+ monocytes were cultivated for 6 days in X-VIVO 10 serum-free and endotoxin-free medium (Cambrex, Verviers, Belgium) at a concentration of 5×105 cells/ml in the presence of IFNγ (1000 U/ml) (PeproTech, Rocky Hill, USA) or IL-4 (10 ng/ml) (Tebu Bio, Frankfurt, Germany) to induce differentiation into Th1 and Th2 polarized macrophages. X-VIVO 10 is an optimal medium for cultivating human monocyte-derived macrophages [12], [28] which does not contain exogenous growth factors or artificial stimulators of cellular proliferation. Burst-forming unit-erythroid (BFU-E) and colony-forming units containing granulocytes and macrophages (CFU-GM) were generated from human CD34 progenitor cells as previously described [29].
Infected granuloma tissue was obtained from the Department of Pathology, Universitätsmedizin Göttingen, Georg-August-University. Bronchial alveolar lavage fluid samples were excess material derived from three patients in the course of their medical care. The use of these specimens and mononuclear cells from healthy probands was approved by the Ethics Committee of the Faculty of Medicine Mannheim, University of Heidelberg, Germany and Ethics Committee of the Medical University of Göttingen, Germany. (Permit Number: 2007-254N-MA and 27/6/11). All patient samples were stricly anonymized. In accordance with the Declaration of Helsinki [30] no written informed consent was provided by study participants and/or their legal guardians.
The 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. Mouse procedures performed in this study were conducted at the Centenary Institute, after protocol review and approval by the University of Sydney Animal Ethics Committee (K75/3-2004/3/3878) and at the animal facility of the Klinikum Göttingen Georg-August-Universität (Göttingen, Germany) according to the Deutsche Tierschutzgesetz (LAVES Niedersachsen A-008/09), after protocol review and approval by the University of Göttingen. All surgery was performed under sodium pentobarbital anesthesia, and all efforts were made to minimize suffering by the attending veterinarian.
C57BL/6 mice and rag1 null mutant mice (rag1–/–) were bred and maintained at the animal facility, Klinikum Göttingen Georg-August-Universität (Göttingen, Germany). Macrophages from the spleen were collected and purified by CD11b-MACS for further analyses. Thioglycollate-elicited mouse peritoneal macrophages were collected according to standard protocols. Mice used in the M. tuberculosis infection experiments were maintained under specific pathogen-free conditions in the Animal Facility of the Centenary Institute of Cancer Medicine and Cell Biology (Newtown, Australia). Genetically modified chimeric mice were generated by transfer of 1×106 purified wildtype CD3+ T cells into TCR deficient rag1–/– mice and TNF null mice, respectively (n = 7 in each group) following whole body irradiation with 5 Gy as previously described [7]. At the time of T cell reconstitution wildtype, rag1–/–(CD3 rag1+/+) and TNF–/–(CD3 TNF+/+) chimeric mice, respectively, were infected via aerosol with ∼ 100 M. tuberculosis bacilli (H37Rv) utilizing a Middlebrook airborne infection apparatus. After four weeks all mice were sacrificed and paraffin embedded lung sections were used for immunohistochemical analyses.
PBMC were isolated from freshly collected whole blood or buffy coats from healthy individuals by Ficoll density gradient centrifugation as previously described [9]. Cells were first fixed with 4% paraformaldehyde and then blocked with normal horse serum (Jackson Immuno Research Laboratories). TCR staining was performed using the TCRβF1 antibody (clone 8A3), which was also utilized in the immunocytochemistry and immunoblot experiments. This framework antibody recognizes TCRαβ dimers. Mouse IgG1 and IgG2b antibodies serve as isotype controls and CD235a (purified) as a nonsense control antibody (all BD Bioscience). Antibody binding was detected with a secondary goat-anti-mouse FITC-labeled antibody (BD Bioscience). After saturating with normal mouse serum (Jackson Immuno Research Laboratories) the cells were stained with the following directly labeled antibodies: CD45 APC-H7, CD3 PerCP-Cy5.5, CD8-PE and CD14-APC (BD Bioscience). Analyses were performed on a FACS Canto using the FACS Diva software (BD Bioscience).
Monocytes/macrophages were stained with mouse anti-human TCRαF1/TCRβF1 antibodies (Thermo Scientific, Waltham, USA) in combination with Alexa Fluor 555 goat anti-mouse IgG (Invitrogen, Carlsbad, USA) and analyzed on an iCYS laser scanning cytometer (Compucyte, Cambridge, MA, USA). For assessment of apoptosis, annexin V/ PI staining was used. Contouring of cells was achieved by nuclear staining with DAPI. Photomultiplier tube settings for voltage, offset and gain were optimized (Figure S1C). Data were acquired and analyzed using the iCYS cytometric acquisition and analysis software (CompuCyte). For statistical analysis, the entire area of the microscopic slide was scanned and for every event pictures of each channel (red, blue, and scatter) were recorded and merged within a gallery. The count settings selected for signal area of DAPI staining were within the range 15–150 µm2 in order to exclude artefacts such as debris.
Modified RIPA buffer was used to extract whole cell lysates. SDS-PAGE and transfer to nitrocellulose membranes was conducted utilising the NuPAGE protein electrophoresis systems (Invitrogen). The primary monoclonal mouse anti-human TCRαF1/TCRβF1 antibodies (Thermo Scientific) and the polyclonoal antibodies for cleaved caspase 3 (ASP175) or CD3 zeta (Abcam, Cambridge, UK) were used. Mouse monoclonal antibodies to β-actin (Abcam) were used as loading control.
Immunoprecipitated TCRαβ bands were separated by SDS-PAGE and visualized by silver staining. The predicted 58kD single band was excised and destained. An in-gel digestion with trypsin was performed according to a standard protocol [31]. After o/n incubation at 37°C, peptides were extracted with a C18 affinity chromatography (ZipTip, Millipore) and eluted with 0.5% formic acid in 1∶1 (v/v) water: acetonitrile. The eluate (1 µl) was mixed with saturated solution (9 µl) of α-cyano-4-hydroxycinnamic acid in 50% CAN and 0.1% TFA, spotted onto a steel target and the droplet was air dried prior to MS-analysis. Peptide mass fingerprinting was performed on a MALDI-TOF-MS (Autoflex II, Bruker Daltonics) operating in the reflector mode. The MS spectra for peaks in the range of 1-3.5 kDa were generated by summarizing 350 laser shots (50 laser shots at 7 different spot positions). Spectra were analyzed using the flexanalysis software (Bruker) [31], [32]. Analysis of the MS spectra was performed utilizing the BioTools software (Bruker) in combination with an integrated online link to the Mascot database (www.matrixscience.com).
DNA from 106 monocytes, IFNγ-macrophages, HepG2 (negative controls) and PBMC (positive controls) was isolated using the Wizard Genomic DNA purification Kit (Promega). Screening for Dβ → Jβ and Vβ → Jβ rearrangements at the TCRβ locus was performed by PCR utilizing a modified non-multiplex approach according to the protocols by van Dongen et al. [13] and confirmed by sequencing. The customized primers used can be requested by the authors.
RNA from all monocyte/macrophage populations was prepared with TRI Reagent (Sigma) and transcribed into cDNA using the Reverse Transcription System (Promega). RT-PCR expression profiling of components of the TCR machinery and size spectratyping of the antigen binding TCR Vα/Vβ CDR3 regions were performed as previously reported [9]. Vαβ spectratypes of the human TCR CDR3 regions were assessed on a CEQ™ 8000 Genetic Analysis System (Beckman Coulter) using the D4-labeled primers D4-GCAGACAGACTTGTCACTGG (TCRα) and D4-TTGGGTGTGGGAGATCTCTGC (TCRβ), respectively. To determine the detailed CDR3 clonotypes for Vβ13a, specific RT-PCR amplification products were cloned into a pCR-TOPO vector (TOPO TA Cloning Kit, Invitrogen) using standard protocols. The cDNA sequences of the Vβ13a CDR3 regions were analyzed from at least 10 randomly picked clones. qPCR for the TCRβ constant chain and CCL2 was conducted using the IQ SYBR Supermix (Biorad, Hercules, USA). B2MG and GAPDH were used as housekeeping genes. Purity of monocyte/macrophage RNA was confirmed by PCR amplification of the leukocyte lineage markers CD2, CD8, CD14, CD64, CD68, CD163, MMP25 and MPO, respectively (exemplified in Figure 2D). Authenticity of all relevant PCR products was confirmed by sequencing. PCR runs were repeated at least twice. The sequences of additional PCR primers used in this study can be requested from the authors.
Before immunostaining all human and mouse tissue sections were deparaffinized and rehydrated. For immunostaining, 5 µm tissue sections or cells cultivated on coverslips were blocked with 5% goat serum in PBS (1% BSA), incubated with a combination of primary antibodies at 4°C overnight, washed in PBS for 15 min, and incubated with a combination of appropriate secondary antibodies. The following antibodies were used: FITC-labeled mouse anti-human CD68 (KP1, 1∶50) (DakoCytomation), mouse anti-human antibodies to TCRαF1 (clone 3A8) and TCRβF1 (clone 8A3) (1∶100, Thermo Scientific), hamster anti-mouse TCRβ (clone H57-597) (1∶50, BD Biosciences), mouse anti-human MHC Class II (clone 910/D7, 1∶200) (Acris Antibodies), anti-mouse F4/80 (AbD Serotec), rabbit anti-human CD163 (Santa Cruz), anti-mycobacterium tuberculosis-FITC (Acris, Herford, Germany), rabbit anti-human CD2 (Thermo Scientific) and anti-cleaved caspase 3 (ASP175) (Cell Signaling Technology, Beverly, USA). Goat anti-mouse IgG, Alexa 488 (Invitrogen), donkey anti-rabbit IgG, Cy3 (Dianova) (both 1∶400), goat anti-hamster IgG, Alexa 488, donkey anti-rat IgG, Cy3 (Jackson ImmunoResearch) and goat anti-mouse IgG, Alexa 555 (MoBiTec, Göttingen, Germany) were used as secondary antibodies. Mouse IgG1 and IgG2a (BD Biosciences), rat IgG2 (Biozol) and hamster IgG2 isotype control antibodies (BD Biosciences) were used as negative controls. For fluorescence imaging DRAQ5 (1∶2500) (eBioscience) and Vectashield Mounting Medium with DAPI (Vector Laboratories), respectively, were used for nuclear staining. Positive staining was either visualized by a Leica DMIRE2 microscope and the FW400 software or a Leica TCS SP-2 laser-scanning spectral confocal microscope equipped with a 63×1.32 objective (Leica Microsystems). Excitation sources were an argon laser (488 nm), a crypton laser (568 nm) and a helium/neon laser (633 nm). Data were acquired and analyzed using the Leica confocal software. Two- and three-color images were acquired using a sequential scan mode [28]. For light microscopy, samples were incubated with the TCRα/TCRβ antibodies, rabbit antibodies to CCL2 (ab7202, Abcam, Cambridge, UK) or rabbit anti-human CCL2 (ab9669, Abcam), respectively, in combination with Envision+ system-HRP anti-mouse or anti-rabbit HRP (Dako). Quantitation of TCRαβ+ cells in immunostained cytospin preparations and tissue sections was conducted by at least two blinded assessors and subjected to statistical analysis. For quantitation of electronic fluorescence microscopy images the NIH image J software was used.
IFNγ stimulated macrophages were incubated in the presence or absence of BCG for 6 h and subsequently fixed in 4 % formaldehyde and 0.2 % glutaraldehyde (0.1 M phosphate buffer). After washing the cells were scraped from the dish in 0.1 M phosphate buffer containing 1% gelatin, spun down and resuspended in 10% gelatin (0.1 M phosphate buffer) at 37°C. The cooled gelatin pellets were cut in small blocks, infiltrated in 2.3 M sucrose in 0.1 M phosphate buffer and mounted onto aluminum pins for ultramicrotomy before shock freezing. Ultrathin cryosections were picked in a 1∶1 mixture of 2 % methylcellulose and 2.3 M sucrose. For immuno-labeling sections were incubated with monoclonal antibodies specific for TCRαF1 (Thermo Scientific) and rabbit anti-mouse IgG antiserum (Rockland, Gilbertsville, PA, USA) followed by protein A-gold (10 nm). Alternatively, the primary mouse antibody was detected utilizing goat anti-mouse antibodies conjugated to gold (Aurion, Wageningen, The Netherlands). Sections were analyzed using a LEO EM912 Omega transmission electron microscope (Zeiss, Oberkochen, Germany) and digital micrographs were obtained with an on-axis 2048×2048-CCD camera (Proscan, Scheuring, Germany).
CD68+/TCRαβ+ cells were identified in 2 µm lung sections from patients with pulmonary tuberculosis by immunofluorescence microscopy (Carl Zeiss Microimaging, Göttingen, Germany). Single double positive cells or small cell clusters were then microdissected using a P.A.L.M. Laser Microdissection System with laser pressure catapulting (LPC) (P.A.L.M. Microlaser Technologies, Bernried, Germany). Total RNA from clusters of 20–30 cells was isolated using the invisorb RNA kit I (Invitek, Berlin, Germany) and cDNA was synthesized (Superscript III First-Strand cDNA Synthesis Kit, Invitrogen, Carlsbad, USA).
Attenuated Mycobacterium bovis BCG (Bacillus Calmette-Guérin, BCG-medac) bacteria, strain RIVM, were re-constituted with the supplied 0.9% NaCl solvent (BCG-Medac, Hamburg, Germany) and used for infection of macrophages (MΦ:BCG = 1∶10). For quantitation of phagocytosed bacilli, BCG mycobacteria were immunostained with a FITC-labeled antibody to M. tuberculosis (Acris). BCG phagocytosis was quantitated from a total of 20 randomly selected fluorescence microscopy images using the NIH image J software. The phagocytotic index (PI) was calculated as (percentage of macrophages containing at least one bacterium) x (mean area of bacterial staining per cell).
The areas (number of pixels) of the M. bovis BCG infected macrophage clusters were determined from electronic images using the image J software. From each individual 30 clusters were analyzed.
Isolated CD14+ cells were allowed to differentiate into macrophages for 6 days on semisolid agarose plates (0.4% agarose in X-VIVO 10 medium) in the presence of IFNγ (1000 U/ml) and M. bovis BCG. Single foci were collected using a 1 ml syringe containing 100 µl PBS. cDNA was prepared using the RNeasy MircoKit and Sensiscript RT Kit (Qiagen, Hilden, Germany) and subjected to TCR Vβ CDR3 spectratyping.
For selective activation of the TCR, freshly obtained CD14+ monocytes from two healthy donors were allowed to differentiate into macrophages in the presence of IFNγ. After 6 days IFNγ-polarized macrophages were washed twice with X-VIVO-10 medium (Cambrex) and co-stimulated with an endotoxin-free mouse anti-human CD3 (1 µg/ml, Beckman Coulter) monoclonal antibody as previously reported [9]. The culture supernatants were collected at various timepoints and the release of a selected panel of cytokines, chemokines and growth factors (n = 15) (Table S1) was assessed utilizing a customized Luminex-based multiplex Procarta Cytokine Assay Kit (Multimetrix, Heidelberg, Germany). All measurements were conducted at least in duplicate. In addition, CCL2 (Thermo Scientific) and CCL5 (Qiagen) in the supernatants were quantified using specific ELISAs.
For phagocytosis of baits targeted to the TCRαβ, IFNγ-polarized macrophages were co-incubated with polystyrene bead baits (Ø 4.5 µm, Invitrogen) coated with anti-TCRα/anti-TCRβ, anti-CD11b (BD Biosciences), nonspecific IgG isotype control antibodies or albumin for 15 min, 1 h and 10 h, respectively, at 37°C (MΦ:beads = 1∶1, 5 µg protein/107 beads). In addition, phagocytosis of albumin-coated beads was assessed in the presence of uncoupled anti-TCRα/anti-TCRβ antibodies. Quantitation of phagocytosed beads was conducted by bright field microscopy of at least 12 randomly selected fields of vision performed by two independent observers.
Student’s t test was used to compare the significance of differences between groups. Results were expressed as means ± SD. p<0.05 was considered statistically significant.
|
10.1371/journal.pcbi.1004719 | Radial Frequency Analysis of Contour Shapes in the Visual Cortex | Cumulative psychophysical evidence suggests that the shape of closed contours is analysed by means of their radial frequency components (RFC). However, neurophysiological evidence for RFC-based representations is still missing. We investigated the representation of radial frequency in the human visual cortex with functional magnetic resonance imaging. We parametrically varied the radial frequency, amplitude and local curvature of contour shapes. The stimuli evoked clear responses across visual areas in the univariate analysis, but the response magnitude did not depend on radial frequency or local curvature. Searchlight-based, multivariate representational similarity analysis revealed RFC specific response patterns in areas V2d, V3d, V3AB, and IPS0. Interestingly, RFC-specific representations were not found in hV4 or LO, traditionally associated with visual shape analysis. The modulation amplitude of the shapes did not affect the responses in any visual area. Local curvature, SF-spectrum and contrast energy related representations were found across visual areas but without similar specificity for visual area that was found for RFC. The results suggest that the radial frequency of a closed contour is one of the cortical shape analysis dimensions, represented in the early and mid-level visual areas.
| Current views suggest that neural representations of the visual environment are built from combinations of basis functions. For low-level visual feature analysis these basis functions are relatively well understood. It is not yet known, however, how primary visual features are combined into higher-level representations of visual objects. Psychophysical evidence suggests that visual shapes are perceived on the basis of the radial frequency components of the shape contour. We investigated whether human visual cortex contains representations of radial frequency components. Our results show that the neural mechanisms that utilize radial frequency are located in the early and intermediate visual areas, and provide further support for the idea of radial frequency based representations in shape perception. This suggests that radial frequency representations might be one link between low-level visual feature analysis and high-level object shape representations.
| To psychophysically investigate contour shape processing beyond local Gabor-like analysis Wilkinson et al. [1] introduced radial frequency patterns (Fig 1A), closed contour shapes formed by sinusoidally modulating the radius of a base circle (Fig 1C). Any closed shape, such as the outline of human face, can be constructed with multiple radial frequency components (RFC) [2]. Wilkinson et al. [1] showed that human observers are extremely sensitive in detecting shape deformation from circularity, with visual acuity exceeding the spatial resolution of the retina. Psychophysical studies have provided converging evidence that visual system relies on global shape analysis of these patterns [1, 3–10]. Experiments using psychophysical methods of adaptation [11, 12], masking [13], and sub-threshold summation [6] suggested that shape analysis of these patterns are RFC specific. This indicates that closed contour shapes are analysed—similarly to local spatial frequency and orientation—via narrow-band radial frequency channels. The neurophysiological evidence for RFC-based shape representations is, however, still missing.
In previous functional magnetic resonance imaging (fMRI) studies, shape representations have been studied with circular gratings and Gabor arrays. Radial and concentric gratings [14] as well as Gabor flow-fields that contain global shape [15] evoke stronger responses in mid-level areas V3 and hV4 than in primary visual cortex (V1) or area V2. Lateral occipital complex (LOC) is also associated with processing of visual objects [16–19] and contours [20]. Human fMRI results are consistent with single-cell recording studies in macaque monkeys that show selectivity for complex shapes in areas V2 [21–24] and V4 [23, 25–29] although complex shape units have been reported also from area V1 [23, 30].
The aim of this work is to test the hypothesis emerging from psychophysical evidence that intermediate shape analysis contains representations of contour RFC. We measured blood oxygenation level dependent (BOLD) responses from the human visual cortex to parametric variation of radial frequency and modulation amplitude of closed contours (Fig 1A). The measured BOLD-responses were analysed with multivariate representational similarity analysis (RSA) [31]. In RSA, correlations between activity patterns evoked by different stimuli are calculated to construct representational dissimilarity matrices (RDMs). To characterize response profiles for different visual areas, the measured RDMs were compared to model RDMs based on stimulus radial frequency, modulation amplitude, local curvature, spatial frequency spectrum and contrast energy. We used a searchlight approach [32] that makes no assumption about location, but instead the whole cortex is scanned voxel-by-voxel to find the stimulus-specific information. We found RFC specific response patterns and our results suggest that mid-level visual areas V2d, V3d, V3AB, and IPS0 contain radial frequency based representations of contour shapes.
During the measurements, the participants performed a demanding RSVP task at the fixation [18] to control for attention and to ensure that the participant did not attend to any specific shape or part of the contour. The percentage of correct responses varied, both individually and between runs, from 35 to 80%, but was significantly above the chance level of 25% correct (t(10) = 6.124,p < .001). The average of correct responses across participants was 50%.
First we investigated the activity evoked by our shape stimuli by calculating average activity within the searchlight sphere, each with radius of 3 voxels. This analysis is comparable to standard univariate voxel-wise analysis with smoothing. As expected, the stimuli (all modulated shapes averaged) evoked clear clusters of activity in all mapped visual areas (Fig 2A). Modulated shapes evoked slightly larger activity than circular shapes in all visual areas (Fig 2B), but the difference was statistically significant only in areas V3v, VO1, V3d, and V3AB (Table 1A). The difference in activation is probably due to adaptation (circles were presented more often) or contrast energy (circles had shortest perimeter). The response magnitude varied significantly across areas (Table 1B) and largest responses for both circles and modulated shapes were found in areas V3d, hV4, VO1, V3AB, LO and TO (Fig 2B). Thus our stimuli well activated areas known to be important for shape processing. The amount of activation slightly increased as a function of local curvature, but the increase was not statistically significant (Table 1B and Fig 2C). The amount of activation also slightly decreased as a function radial frequency, but the decrease was not statistically significant (Table 1B and Fig 2D). There were no significant interactions and the activation across visual areas did not depend on the local curvature or the radial frequency of the stimulus (Table 1B). In sum, the stimuli evoked clear responses across visual areas, but the amount of signal change did not depend on the stimulus parameters.
Next we investigated if any of the areas represented specific contour shape information. We tested whether the response profiles in any visual area resembled response profiles predicted on the basis of different stimulus parameters: radial frequency, concave curvature, convex curvature, and modulation amplitude. Within each searchlight we calculated a representational dissimilarity matrix (RDM) by cross-correlating the response patterns for the 16 modulated shapes, and compared the measured RDMs to model RDMs based on shape parameters (Fig 3). Visually comparing the measured RDMs (S1 Fig) with the model RDMs (Fig 3) does not reveal any model superior to other models. Visual inspection suggests, however, that there are differences across visual areas and non-random structures in the measured RDMs (S1 Fig) that might be explained by the stimulus parameters.
Next, the similarity between the measured and model RDMs was quantified by calculating the correlation between the measured RDMs and model RDMs within the spherical searchlights, and averaged across participants. Fig 4 shows these correlation maps in visual cortex for each model RDM. In general, the response profile maps (Fig 4) resembled the activity maps (Fig 2A), that is, the strongest correlations between brain and model RDMs were found across visual cortices approximately at the same retinotopic locations as the highest activity. However, LO2, TO1 and TO2, while robustly activated in the univariate analysis, were not captured by any of the models in the multivariate analysis.
The highest correlations (>three standard deviations above the mean) between the measured RDM and the RFC-model were found in visual areas V1, V2v, V2d, V3d, V3AB, IPS0, and some voxels in area LO1 (Fig 4A) suggesting that the activation patterns in these areas carry information about the RFC of the stimulus contour. For the amplitude of the shape modulation, no clear peaks in correlation were found in any visual area (Fig 4B). For convex curvature, clusters of high correlations were found in areas V1, V2v, V3AB, and in VO1 in right hemisphere (Fig 4C), and for concave curvature, high correlations were found in few voxels in V1 (Fig 4D).
We also calculated correlation maps for spatial frequency and contrast energy. As the radial frequency of the contour and the amplitude of the modulation are increased, the SF-spectrum of the stimulus shifts slightly to higher frequencies. The amplitude modulates the SF-spectrum more than the radial frequency. As expected, the pattern of activity in areas V1-V3, V3AB, IPS0, and V01 strongly correlated with the SF-spectrum of the stimuli (Fig 4D). Similarly the measured RDMs correlated with contrast energy in areas V1-V3, V3AB, and IPS0 (Fig 4E). Thus, the low-level visual factors—contrast energy and spatial frequency—predicted the dissimilarity of the response patterns non-selectively across visual cortex.
To further quantify the differences between response profiles across visual areas, we conducted a ROI analysis based on the probability atlas of visual areas [33]. In the ROI analysis, areas LO1 and LO2 that are a part of LOC [18], as well as areas TO1 and TO2, were combined. Further, we used the univariate activity maps (Fig 2A) as functional localizer, that is, for each individual we included only voxels that were clearly activated by the stimuli (t-value > 4.0). Separate repeated measures ANOVAs were conducted for every model.
The average correlation of the measured pattern and RFC-model did not differ across the left and right hemispheres (Table 1C and Fig 5A). However, the average correlation with RFC-model varied significantly across visual areas (Table 1C and Fig 5A). The highest correlations (significantly above zero, p < .05, t-test) between the measured RDM and RFC model RDM were in areas V2d, V3d, V3AB and IPS0, in both the left and the right hemispheres (Table 1D and Fig 5A). That is, in these visual areas the measured response patterns carried information about the radial frequency of the shape stimuli. The correlations in areas V1 and V2v were not significantly above zero. In areas V3v, hV4, VO1, LO1/LO2 and TO1/TO2, the average correlation with RFC-model was close to zero (Fig 5A). No significant interaction between the hemisphere and visual area (Table 1C) was found for RFC model suggesting that radial frequency is similarly represented in same visual areas in both hemispheres. Thus, the statistical analyses confirm the selective spread of pattern correlations for radial frequency only for certain visual areas, as was shown in the Fig 4A.
The effect of hemisphere was not significant for any model indicating robustness of the result. The correlations varied significantly across visual areas also for SF and contrast energy, but not for amplitude, concave curvature or convex curvature (Table 1C). The effect of SF and contrast energy was expected since these are the primary factors that drive the low-level neural responses. However, the modulation amplitude as such and the direction of local curvature seem not to be represented specifically in any certain visual area. The interaction between the hemisphere and visual area was not significant for any model, showing that response patterns were similar in the same visual areas in both hemispheres.
Since no significant effects of hemisphere were found, the correlations for each model RDM were averaged across hemispheres, and tested with t-tests (significantly above zero, p < .05) to statistically confirm the spread of correlations across visual areas (shown in Fig 4). We used probabilistic atlas to localize the visual areas and this might produce some uncertainty to classifying areas close to each other. Therefore, nearby visual areas were averaged as follows: V1, V2d/V3d, V3AB/IPS0, V2v/V3v, hV4/VO1, and LO/TO. The average correlation of the measured RDM with RFC model varied across areas and was significantly above zero in V2d/V3d and V3AB/IPS0 (Fig 5B). The average correlation with Amplitude model RDMs was constantly low and not significantly higher than zero in any of these areas. The average correlation to other models remained high across most of the areas: correlation with SF and energy models was significantly above zero in all areas, correlation with concave model was significantly above zero in all areas except LO/TO, and correlation with convex models was significantly above zero in all areas except hV4/VO1 and LO/TO (Fig 5B).
The clearest area specificity was found for radial frequency. This selectivity was further quantified by comparing (paired sample t-test) the average correlation in dorsal and ventral areas. Statistically significant difference between the ventral and dorsal part of areas V2 and V3, as well as between areas V3AB/IPS0 and hV4/VO1 was found only for RFC, but not for other models (Table 1E). The average difference across all dorsal (V2d/V3d/V3AB/IPS0) and ventral (V2v/V3v, hV4/VO1) areas was largest for radial frequency (Fig 5C), and this difference was significantly higher than for SF (Table 1E).
To further test the independent role of different stimulus parameters, a multiple linear regression analysis within each searchlight was conducted. The regression model contained all six model RDMs: radial frequency, amplitude, convex and concave curvature, contrast energy, and SF spectrum. The average R2 was 0.11, and varied between 0.09 and 0.13 depending on the visual area. Thus, all the models explained 11% of the total variability within the searchlights. Next we conducted regression analysis with leave-one-out method and calculated the change of R2 values relative to the full model. Removing the SF model from the regressors decreased R2 values ca. 40%. For other models, the decrease was 7–16%. Thus most of the explained variability was due to the SF model. This was expected since highest correlations were found for the SF model, and the SF model contains the same information of the stimulus variability as the other models combined. Next we replicated the regression analyses without the SF model. The average R2 of the full model decreased to 0.06. In the leave-one-out analysis, largest decrease of R2 values were found for convex curvature (29%) and radial frequency (23%). However, only for radial frequency the relative decrease of R2 values varied significantly across areas (Table 1F) and was more prominent in dorsal (V2d/V3d and V3AB/IPS0) than in ventral (V2v/V3v and hV4/VO1) areas. Thus, a similar specificity for RFC across visual areas was found in the regression analysis as in the correlation analysis.
Multivariate representational similarity analysis revealed that the RFC of the contour is represented in human visual areas V2d, V3d, V3AB and IPS0. Surprisingly, the areas hV4 and LO1/LO2, known to be important in global shape processing [15, 19, 20], while responding to the stimuli, did not show pattern selectivity for radial frequency. Low-level visual properties—SF spectrum and contrast energy–did not explain our results, since these parameters did not show similar specificity across visual areas as the radial frequency. Our results provide evidence for radial frequency based representations which could be used in contour shape processing, and in particular, we suggest that RFC representations are a mid-level link between local contour analysis in V1 and more comprehensive global shape analysis in areas such as hV4 and LOC. Alternatively, the radial frequency information in the early and intermediate areas may be utilised more directly in the higher level object sensitive areas with no additional middle steps.
Most previous fMRI studies on contour shape perception have compared BOLD-responses to different shapes, i.e., circular vs. parallel gratings [14] or global circular shapes vs. only local curvatures [15]. In the former study hV4 and FFA showed selectivity for concentric shapes, and in the latter study visual areas V3, VP and hV4 showed strongest responses to circularity. Consistent with these studies, our modulated shapes evoked larger responses than circular shapes in univariate analysis of mean signal change, i.e. in the overall fMRI response. The response magnitude did not, however, depend on the local curvature or radial frequency. In order to investigate the role of different stimulus parameters on shape representations, we investigated the multivariate similarity structure [31, 34] of the activity patterns evoked by parametric variation of the contour shapes. The multi-voxel pattern analysis is more sensitive than direct comparison of average responses within the visual area, because the multidimensional pattern of BOLD-responses across voxels contains more information about the response than the averaged one-dimensional measure. Further, the searchlight based approach [32] makes no assumptions where the stimulus specific activation patterns should be found. Our results provide further evidence that radial frequency is used in the contour shape analysis in the visual cortex. Furthermore, our results suggest that RFC based representations are located in visual areas V2d, V3d, V3AB and IPS0. For areas hV4 and LO1/LO2 we did not find evidence for RFC based representations.
Integration of local visual features to contours likely involves visual areas at different processing levels [35, 36]. The representations of contour convexity and concavity, as well as the representations of global shape are likely located in visual areas V3AB, hV4 and LOC [14, 15, 18, 20, 37–40]. In agreement with these studies, largest univariate responses in our study included areas V3AB, hV4 and LO1/LO2. However, the multivariate patterns were different between these areas. Our results suggest that the global closed-shape representations in V3AB are based on the radial frequency of the contour, but we did not find similar RFC based activity patterns in hV4 or LO1/LO2.
The lack of pattern specificity in hV4 and LOC was not due to our stimuli as such, since they robustly activated also these areas. One possibility is that shape representations or neurons encoding shapes in hV4 are so close to each other that voxel-level activity patterns measured with fMRI cannot discriminate them or the MVPA methods are not sensitive enough. This would suggest different structure for V3AB and hV4 neurons/representations since we did find significant voxel-level pattern correlations in V3AB. In primates, cell density [41] and microvascular density [42] vary across cortical areas which might affect BOLD-responses, and thus this is a possible explanation for the difference between the areas we found. Future studies could aim to image the RFC representations in these areas with smaller voxel size or higher spatial resolution using high-field fMRI. Second possible explanation for the difference between the visual areas is that fMRI might be particularly sensitive for feedback [43, 44]. Multivariate pattern reflects data distributed in large part of a functional area, whereas the univariate pattern is sensitive to more local changes. In this scheme the multivariate analysis might better see the feedback effects which typically have much wider distribution than the classical receptive field [45] and this would emphasize the early areas as well as give different distribution in the mid-level areas. Third possibility is that—since the RFC based representations are mainly limited to closed shapes—areas that represent more complex visual objects, such as hV4 and LOC, might simply use some other type of shape encoding.
Instead of individual functional localization of visual areas, we used probability atlas of visual areas measured in a separate study [33]. The average location of early visual areas (e.g., V1-V3) is more accurate than subsequent visual areas (e.g., hV4, LO1/LO2). Hence there might be more locational variability in the activity patterns in hV4 and LO1/LO2 across individuals and this might explain the absence of RFC specific activity patterns. However, we did find significant correlation between measured patterns and model RDMs for SF Spectrum and Contrast Energy also in areas hV4/VO1. In our searchlight analysis the activity patterns are smoothed with the spherical volume of searchlight, and we calculated an average within the ROIs. This analysis controls for small deviations in exact locations of activity patterns. Furthermore, all mapped visual areas, including hV4 and LO1/LO2, were activated by our stimuli and these univariate activation maps were used as functional localizers in the ROI analysis of RDM correlations. This emphasizes voxels across visual areas that were indeed processing our stimuli. Still, some locational uncertainties in our results remain. However, the locational uncertainties are more likely between nearby areas, e.g., between areas V2 and V3, than areas further away, such as between areas V3AB and hV4.
Another limitation of our study is the relatively low correlations found between measured and model RDMs. While the correlations were quite low, the results were systematic across participants, and the effect sizes (of the ANOVAs testing the effect of visual area) were large (Table 1). Instead of the correlation values as such, the structure of correlations across studied events was the main interest in our study. These structures reveal information of the representational geometry that can be compared to predictions based on different models, and we found clear differences between the models based on RFC of the contour and models based on the other stimulus parameters. Further, most of the previous RSA studies have investigated representational similarities across object categories. In contrast, we studied representational similarities of relatively similar shapes within a category, which could also explain the low correlation values we obtained.
The asymmetry between the dorsal and ventral areas for RFC model could be related to the ecologically justified and well known difference between the upper and lower visual fields [46]. Anatomically, there is slightly more cortex representing the lower than the upper visual field in macaque monkey V1 [47], physiologically stronger responses in MEG in humans [48, 49], and better behavioral performance in humans [46, 50]. In line with these earlier findings, our results suggest that radial frequency representations are biased towards cortical areas with lower visual field representations. This finding can be contributed by the relative cortical size of the lower vs. the upper visual field representations, and differences in the represented eccentricities in dorsal and ventral areas. However, we did not find similar asymmetry for contrast energy and spatial frequency. The anisotropy between IPS0/V3AB and hV4/VO1 areas cannot be explained by retinotopy, because all these areas comprise half-field representations, i.e. both the upper and lower visual fields [51]. Consistent with our results, lower visual field advantage was recently demonstrated for perception of RFC shape stimuli whereas no similar asymmetry was found for orientation or curvature discrimination [52].
The radial frequency and amplitude of the modulation determine the shape of an RFC pattern. However, several other parameters vary as the RF and amplitude of the stimulus is varied. Increasing the amplitude and the radial frequency increases the contour length and contrast energy and shifts the SF spectrum to higher SFs. The RDM models based on these low-level visual parameters did correlate with the measured patterns across visual areas but without similar specificity as was found for radial frequency. Thus the results found with the RFC-model are not due to these low level factors but reflect the different activity patterns evoked by the parametric modulation of RFC. For the amplitude of modulation the average correlation was constantly near zero, as expected. The amplitude as such is not a critical parameter for visual shape analysis. Slightly higher correlations were found for convex than concave curvature, but the correlations did not vary much across visual areas for these parameters. Higher correlations for convex curvature might indicate more critical role of convex than concave forms and angles in shape analysis, as previously suggested [4, 13, 28, 40, 53, 54].
In our experiment, all the stimuli were presented in four different orientations. As the orientation of the shape was varied, the shapes activated slightly different retinotopic locations. However, the RSA analysis was conducted for SPMT-images in which the shapes in different orientations had been averaged. Thus the role of different retinotopic locations was controlled already in GLM analysis. We did a separate control analysis with a different GLM model (shapes averaged across different amplitudes instead of orientations) and calculated correlation maps for the orientation and the retinotopic locations. As expected, the correlation map for the orientation as such did not reveal any peaks in any visual area. The correlation map for retinotopy revealed clear activity peaks across visual areas and was highly similar to the correlation map for the SF spectrum.
Recently, radial frequency patterns and multi-voxel pattern analyses have been used to study perception of RFC motion trajectories [55, 56]. The motion trajectory of a dot could be decomposed from areas V2, V3 and MT [55, 56]. In contrast, the shape of the static RFC patterns could not be decoded in these areas, but only in posterior parietal areas and in LOC [56]. In contrast, we found RFC specific response profiles in areas V2-V3, and V3AB. There are several differences in our and Gorbet et al. [56] experimental setups and data analysis which likely explain the different results. Most likely our setup was more sensitive to differences between RFCs because we had much higher number of stimulus presentations, and we compared response patterns to all different RFCs simultaneously, instead of comparing only two RFCs at a time [56]. Nevertheless, both our and Gorbet et al. [55, 56] studies agree in that radial frequency is used in the visual shape analysis in areas V2 and V3, but not in area hV4.
The prevailing theoretical view suggests that neural representation of the visual environment is built from a sparse set of basis functions whose combinations constitute the population code for perceptual representations [57, 58]. It is possible that the RFC representations, corresponding to the relatively simple combinations of Gabors, provide a set of mid-level basis functions for shape analysis. Our results provide further support for the idea of radial frequency based representations in shape perception, and suggest that the neural mechanisms that utilize radial frequency are located in the intermediate visual areas V2d, V3d, V3AB and IPS0, but not in areas hV4 or LOC. This result places the radial frequency representations to relatively early position in visual processing, presumably beyond Gabor analysis, but before object identification.
The ethics committee of the Hospital District of Helsinki and Uusimaa had approved the experiments (Coordinating ethics committee, Dnro 299/13/03/00/2010). The experiments were conducted according to the declaration of Helsinki and participants gave written informed consent before the measurements.
Eleven participants (one female), with normal or corrected-to-normal vision, participated in the study. First and last author participated as subjects; the rest of the participants were naïve to the purpose of the study.
The stimuli were radial frequency patterns (Fig 1A), which were constructed by sinusoidally modulating (Fig 1C) the radius of a base circle [1]. The spatial profile of the contour was 4th derivative of Gaussian and the peak spatial frequency of the contour was 1.57 c/deg (σ = 0.28 deg). The shapes were composed of four different radial frequencies, amplitudes (Fig 1A) and orientations (Fig 1B). The radial frequencies were 3 (triangle), 4 (quadrilateral), 5 (pentagon) and 6 (hexagon) cycle/perimeter. The minimum and the maximum local curvature of the contour depend on the radial frequency and the amplitude of the shape (equation 4 in [1]). The amplitude of the shapes was varied so that the maximum local curvatures at the peak (point of maximum curvature or convex curvature) and the trough (point of minimum curvature or concave curvature) of the radial modulation (Fig 1D) were roughly equal across different radial frequencies (Table 2). The amplitude varied between 0.0 and 0.46 in proportion to the radius, the maximum concave curvature varied between -0.1 and -4.8 deg-1, and the maximum convex curvature varied between 0.6 and 2.3 deg-1 (Table 2). Each stimulus was presented in four orientations, which corresponded to polar phases 0, 90, 180 and 270 deg (Fig 1B). In total there were 65 different stimuli (circle + 4 radial frequencies x 4 amplitudes x 4 orientations). The rms-contrast (the standard deviation of the luminance divided by the mean luminance) of the stimuli was 0.17, and same for all the stimuli. The radius of the base circle was 2.86 deg and the stimulus maximum diameter varied from 5.7 (circle) to 8.6 deg (radial frequency four with amplitude of 0.35).
The fMRI data were acquired with a Siemens MAGNETOM Skyra 3 T scanner (Siemens Healthcare, Erlangen, Germany) equipped with a 30- or 32-channel receive only head coil. Each measurement session started with a fast structural MR image with a 3D T1-weighted sequence (in-plane resolution 1.8x1.8 mm, and 1.5 mm slice thickness). Then eight experimental runs were measured using a gradient-echo echo planar imaging sequence (TR = 1800 ms, TE = 30 ms, flip angle = 60 deg, 64x64 acquisition matrix, FOV = 20 cm, 23 slices, 3.0 mm slice thickness, resulting in 3.1 x 3.1 x 3.0 mm voxel size).
The fMRI data were analyzed with SPM8 Matlab toolbox [59] and Freesurfer [60] software packages. The preprocessing comprised first the correction for the acquisition order of the functional images and then for the head motion.
Circles and 64 different contour shapes were presented in event-related design. In each run all shapes were presented once, except circular stimulus, which was presented 16 times. In addition, 20 rest trials were included. In total, one run consisted of 100 events (20 rests + 16 circles + 64 modulated shapes). The duration of each stimulus was 300 ms and the duration of each trial 2.4 seconds. Hence, the length of the run was 240 s (100*2.4s, corresponding to 135 volumes). Each run started with eight volumes, which were discarded from the analysis to reach stable magnetization. Stimuli were presented with abrupt on/offset. In total, eight runs were measured resulting in 800 stimulus presentations. The order of stimuli in each run was optimally randomized [61] and the order of runs was randomized across participants. Each radial frequency was presented 128 times (4 (amplitudes) x 4 (orientations) x 8 runs).
The shapes were presented at 10 deg eccentricity (at 43-cm viewing distance), one (same) shape in each quadrant in order to have separable upper/lower and left/right visual field responses in early visual areas. The participants performed a demanding RSVP attention task at the fixation [18]. Five different letters (Z,L,N,T, and X; Arial-font) were rapidly presented at fixation (150 ms/letter). The letter series contained 1–4 ‘X’-letters and the participants’ task was to count the ‘X’s and report the number of ‘X’s during a 1 s break in every 5.4 seconds. The letter task was used to control for participants attention. The letter task, instead of shape discrimination task, was also used to avoid ceiling effects in behavioral performance since the contour shapes were supra-threshold, far exceeding the contrast sensitivity and discrimination thresholds of these patterns.
The general linear model (GLM) analysis included a design matrix where the data were modelled with 17 effects of interest (1 for circle, 16 for different modulated shapes) and 8 nuisance regressors (1 for RSVP-letters, 1 for responses to attention task, and 6 for head motion parameters). Data were high-pass filtered at 1/200 Hz. SPMT-images were calculated for the 17 stimulus-related regressors, and corresponding BOLD signal changes for each voxel by dividing the parameter estimates with mean response. In the experiment each shape was presented in four different orientations. The orientation was omitted from the analysis and each shape was modelled with one regressor. Separate control analysis confirmed that orientation as such did not have significant effect on the measured activity patterns.
Next, we applied several searchlight analyses [32]. The radius of the spherical searchlight was 3 voxels, resulting in an average search volume of 100 voxels (similar results were obtained with smaller (70 voxels) and larger (270 voxels) searchlight volumes). First searchlight analysis was conducted to find visual areas that were activated by the contour stimuli. For a univariate activity map, comprising ca. 30000 voxels, we first centered the searchlight volume at each voxel, and then calculated within the volume the T-value of average signal change across the different shapes, i.e., stimulus-related regressors. The circle shape was omitted from the analysis because it was presented more often than other shapes, and it evoked weaker responses than modulated shapes. The same analysis was repeated for each participant. The result essentially corresponds to classical activity map but with smoothing by the searchlight volume.
In the subsequent searchlight analyses, we calculated a representational dissimilarity matrix (RDM; [31]) within each searchlight. The RDM comprised 1—Pearson correlation between the response patterns for the 16 visual shapes (the circle shape was again omitted from the analysis). The logic of the searchlight analysis is straightforward: If the multi-voxel response pattern, contained within the spherical searchlight, carries information about a parameter, such as the stimulus shape, then there should be high correlation between voxel response patterns for similar shapes and low correlation for dissimilar shapes. To find out what information the visual areas were representing, the measured RDMs within the searchlights were compared to different model RDMs [31, 62]. The comparison was quantified as Spearman correlation between the measured and model RDMs. Since the RDMs are symmetric over the diagonal, only the lower triangular parts of the matrices were used in the comparison. The searchlight RSA analysis were conducted using the RSA toolbox [63].
The RDM models for Radial Frequency Components (RCF), Amplitude, Concave and Convex Curvature were constructed by classifying the stimuli according to the parameter of interest (Fig 3; Table 2). In the RDM for RFC, for example, the stimuli are classified according to the radial frequency of the stimulus (RFC 3–6), and nearby radial frequencies are expected to be represented more similar than more distant radial frequencies. Thus, if the RFC of the shape is represented in any visual area, then the activity patterns for the same RFCs are expected to be more similar (higher correlation) than the activity patterns for different RFCs. A similar logic applies to other features of the stimuli. Separate models for concave and convex local curvature were used since previous studies have suggested different role of concavities and convexities in shape perception [53] and in fMRI [40]. To control for low-level visual feature similarity between the stimuli, two additional RDM models were calculated for Contrast Energy and Spatial Frequency (SF) spectrum of the stimuli (Fig 3). For the Contrast Energy model, total contrast energy (sum of squared pixel contrast values) for each stimulus was calculated and classified to create the model RDMs (Table 2). For the SF model, SF spectrum of each stimulus was calculated and cross-correlated to create model RDM.
In total 7 searchlight maps (activity map + 6 RSA correlation maps) were calculated for each participant. The individual searchlight maps were projected on the Freesurfer average cortical surface, and averaged across participants. The visual areas were identified from the Freesurfer average surface on the basis of probability atlas of visual areas measured in the separate study [33]. The statistical significance of the average signal changes and correlations across participants were tested with repeated measures ANOVAs and t-tests. Greenhouse-Geisser correction was used in ANOVAs when Mauchly's test of sphericity was significant.
|
10.1371/journal.pntd.0005075 | Use of Household Cluster Investigations to Identify Factors Associated with Chikungunya Virus Infection and Frequency of Case Reporting in Puerto Rico | Chikungunya virus (CHIKV) is transmitted by Aedes species mosquitoes and is the cause of an acute febrile illness characterized by potentially debilitating arthralgia. After emerging in the Caribbean in late 2013, the first locally-acquired case reported to public health authorities in Puerto Rico occurred in May 2014. During June–August 2014, household-based cluster investigations were conducted to identify factors associated with infection, development of disease, and case reporting.
Residents of households within a 50-meter radius of the residence of laboratory-positive chikungunya cases that had been reported to Puerto Rico Department of Health (PRDH) were offered participation in the investigation. Participants provided a serum specimen and answered a questionnaire that collected information on demographic factors, household characteristics, recent illnesses, healthcare seeking behaviors, and clinical diagnoses. Current CHIKV infection was identified by rRT-PCR, and recent CHIKV infection was defined by detection of either anti-CHIKV IgM or IgG antibody. Among 250 participants, 74 (30%) had evidence of CHIKV infection, including 12 (5%) with current and 62 (25%) with recent CHIKV infection. All specimens from patients with CHIKV infection that were collected within four days, two weeks, and three weeks of illness onset were positive by RT-PCR, IgM ELISA, and IgG ELISA, respectively. Reporting an acute illness in the prior three months was strongly associated with CHIKV infection (adjusted odds ratio [aOR] = 21.6, 95% confidence interval [CI]: 9.24–50.3). Use of air conditioning (aOR = 0.50, 95% CI = 0.3–0.9) and citronella candles (aOR = 0.4, 95% CI = 0.1–0.9) were associated with protection from CHIKV infection. Multivariable analysis indicated that arthralgia (aOR = 51.8, 95% CI = 3.8–700.8) and skin rash (aOR = 14.2, 95% CI = 2.4–84.7) were strongly associated with CHIKV infection. Hierarchical cluster analysis of signs and symptoms reported by CHIKV-infected participants demonstrated that fever, arthralgia, myalgia, headache, and chills tended to occur simultaneously. Rate of symptomatic CHIKV infection (defined by arthralgia with fever or skin rash) was 62.5%. Excluding index case-patients, 22 (63%) participants with symptomatic CHIKV infection sought medical care, of which 5 (23%) were diagnosed with chikungunya and 2 (9%) were reported to PRDH.
This investigation revealed high rates of CHIKV infection among household members and neighbors of chikungunya patients, and that behavioral interventions such as use of air conditioning were associated with prevention of CHIKV infection. Nearly two-thirds of patients with symptomatic CHIKV infection sought medical care, of which less than one-quarter were reportedly diagnosed with chikungunya and one-in-ten were reported to public health authorities. These findings emphasize the need for point-of-care rapid diagnostic tests to optimize identification and reporting of chikungunya patients.
| Chikungunya is a mosquito-borne virus that causes an acute febrile illness that often occurs with severe joint pain. The virus first arrived in the Western Hemisphere in late 2013 and has since caused epidemics in much of the Caribbean and the Americas. During the first months of the 2014 epidemic in Puerto Rico, we conducted household-based cluster investigations to identify factors associated with chikungunya virus infection and progression to disease. We found that using air conditioning and citronella candles in and around the home were associated with decreased rates of infection. Symptoms significantly associated with chikungunya virus infection included fever, joint pain, skin rash, and arthritis. Less than one-quarter of participants infected with chikungunya virus that sought medical care were diagnosed with chikungunya and one-in-ten were reported to public health authorities. This investigation demonstrates the importance of household-level behavioral interventions to avoid chikungunya virus infection, as well as the need for rapid diagnostic tests to improve identification of chikungunya patients.
| Chikungunya virus (CHIKV) is a mosquito-transmitted alphavirus that can cause an acute febrile illness characterized by potentially debilitating arthralgia [1]. Aedes aegypti and Ae. albopictus mosquitoes are the most common vectors of CHIKV and also transmit the four viruses that cause dengue (DENV-1–4) [1]. CHIKV previously caused outbreaks in Southeast Asian and African countries where large portions of the population (e.g., 38–75%) were affected [2–5], which may be attributable to high viremia in the host, high viral load in mosquitos, immunologically naive populations, and the absence of sustainable and effective vector control methods [6]. Although infection with CHIKV results in long-term protection from reinfection [7], it has been associated with persistent arthritis and/or arthralgia that may last several months [8, 9]. In areas where both CHIKV and DENVs circulate, misdiagnosis of chikungunya may be common, as patients with either disease may present with fever, myalgia, and arthralgia [10].
The first documented locally-acquired chikungunya case in the Western Hemisphere was reported in December 2013 on the Caribbean island of St. Martin [11]. Soon after, CHIKV spread to at least 45 countries and territories throughout the Americas where over 2 million suspected cases have been reported to date [12]. In the United States territory of Puerto Rico, the first laboratory-confirmed chikungunya case occurred in a patient from the San Juan metropolitan area who had illness onset in May 2014 and no history of recent travel [13]. The peak of cases reported through passive surveillance occurred in August 2014 [14], and to date >30,000 suspected chikungunya cases have been reported [15]. However, detection of anti-CHIKV antibodies in nearly 25% of blood donated during 2014 suggests a higher incidence of infection than was reported to public health authorities [16].
Because they are transmitted by the same mosquito vectors, CHIKV is thought to have similar transmission patterns as DENV, which often results in clusters of infected individuals in and around the households where infected individuals reside [17–20]. This is largely due to the anthropophilic nature of Ae. aegypti, which tend to disperse relatively short distances (<100 meters) and congregate around households [18]. Consequently, human movement has been identified as the primary mode of DENV dissemination beyond 100 meters [21]. Human population density, particularly in relation to urban centers, has also been associated with clustering of chikungunya cases [22].
Following the introduction of CHIKV into Puerto Rico, we conducted household-based cluster investigations to describe the spectrum of disease and factors associated with CHIKV infection, identify host factors associated with symptomatic infection, describe care-seeking behavior in individuals with chikungunya, and identify patient characteristics associated with accurate clinical diagnosis and case reporting of chikungunya patients.
The investigation protocol underwent institutional review at CDC and was determined to be public health practice and not research. As such, institutional review board approval was not required.
Puerto Rico, an unincorporated territory of the United States located in the Caribbean Sea, has an area of 3,424 square miles and in 2014 had an estimated population of 3,548,397 (1,036 residents per square mile) [23]. A cross-sectional investigation was conducted in which neighbors of chikungunya patients were offered enrollment in household-based cluster investigations. A convenience sample of laboratory-positive chikungunya cases was identified from suspected chikungunya cases that were reported to Puerto Rico Department of Health (PRDH) and tested laboratory-positive for CHIKV infection (“index cases”). Index case-patients or their parent or guardian were contacted by telephone within 30 days of the index case-patients’ illness onset and a home visit was scheduled. All household investigations were conducted between June 20 and August 19, 2014 (S1 Fig).
During each household visit, the head-of-household of the index case-patient’s household (the “index household”) and all households within a 50-meter radius of the index household were eligible for enrollment in the investigation. If the head-of-household agreed to participate in the investigation, all available members of the household were offered participation. Households were not revisited if the head-of-household was not home or declined participation. A questionnaire (S1 Appendix) addressing household characteristics was administered to the head-of-household, and an individual questionnaire (S1 Appendix) addressing demographics, travel history, and recent illnesses was administered to all participants. Parents or guardians answered individual questionnaires by proxy for participants aged <8 years.
Serum specimens were collected from all household investigation participants and transported to CDC Dengue Branch in San Juan, Puerto Rico for diagnostic testing. To detect evidence of CHIKV infection, all specimens were tested by rRT-PCR [24], IgM antibody capture (MAC) ELISA [25], and IgG ELISA [26]. Specimens were also tested for evidence of DENV infection, the results of which have been previously reported [13]. In summary, 5% of participants were positive for recent DENV infection, and none were positive for current DENV infection. Inclusion of DENV diagnostic test results in epidemiologic analyses did not appreciably affect the statistical significance of any findings, as there was minimal overlap of participants with evidence of infection with both CHIKV and DENV (i.e., 1 of 74 [1.4%]). Hence, DENV diagnostic test results are not included in the analyses presented herein.
Names and dates of birth of all CHIKV-infected participants were queried in surveillance databases at CDC and PRDH to determine if they had been reported as a suspected chikungunya case-patient.
Participants were individuals that provided a serum specimen and answered an individual questionnaire. Current CHIKV infection was defined by detection of CHIKV nucleic acid by rRT-PCR. Because CHIKV was first detected to be circulating in Puerto Rico in May 2014 and all household investigations were completed by mid-August 2014, recent CHIKV infection was defined by detection of either anti-CHIKV IgM antibody by MAC ELISA or anti-CHIKV IgG antibody by IgG ELISA. Participants were defined as being laboratory-positive for CHIKV infection if they had evidence of either current or recent infection. Participants were defined as laboratory-negative for CHIKV infection if they had no evidence of either current or recent CHIKV infection. For participants with current CHIKV infection that did not report any symptom of illness (n = 2), development of illness after interview was ruled out by follow-up phone call within 30 days of the household visit. Findings from multivariable and hierarchical clustering analysis of signs and symptoms associated with current or recent CHIKV infection were used to define symptomatic CHIKV infection.
General estimating equations (GEE) were used to model associations between individual health and household characteristics and binary outcomes of CHIKV infection status, correct chikungunya diagnosis, or asymptomatic infection. All GEE models were fit with a logit link and assuming an exchangeable correlation matrix. This method estimates the population-averaged effect, accounting for correlations in data of members from the same household and investigation cluster that might otherwise bias estimates [27]. Multivariate GEE analysis was performed to obtain a final model for the association between laboratory-positivity and symptoms reported among participants with illness in the past three months. Backward elimination was used in best-fitting model selection, removing variables from the full model that lowered the Quasilikelihood Information Criteria (QIC) relative to the full model [28]. Hierarchical cluster analysis, which uses a distance measure to identify similar clusters of variables and an agglomeration method to link clusters, was performed to analyze patterns of symptoms among participants with recent illness. Manhattan distance, a measure of similarity that sums the absolute differences among observations, was used due to the binary nature of outcomes. Ward’s method, which groups variables by minimizing the internal sum of squares, was used as the agglomeration method [29]. GEE analyses were performed using SAS 9.3 (SAS Institute Inc., Cary, NC), and hierarchical cluster analysis was performed using R version 3.2.3. ArcGIS version 10.2 (ESRI, Redlands, CA) was used for mapping household clusters.
A total of 21 household-based cluster investigations were conducted in the health regions of San Juan, Bayamón, Ponce, Arecibo and Caguas (S1 Table). Of 499 households eligible for participation, heads-of-household from 200 (46.2%) occupied households were available to be offered enrollment, and 137 (68.5%) accepted (Fig 1). Median rate of enrollment by cluster and health region was 66.7% (range: 37.5–100%) and 66.7% (range: 62.5–78.9%), respectively. Of the 410 residents of all enrolled households, 250 (61.0%) participated in the investigation. Participants tended to be older than all residents living in participating households (median age = 45 vs. 25 years, respectively).
Of the 250 household cluster investigation participants, 74 (29.6%) had evidence of CHIKV infection. Although infection rates varied by cluster both between and within health regions, all clusters had at least one infected individual apart from the index case-patient (Fig 2). This included 12 participants with current CHIKV infection and 62 participants recent CHIKV infection. Among those with current CHIKV infection, 9 (75.0%) were positive only by rRT-PCR, 1 (8.3%) was positive by rRT-PCR and IgM ELISA, and 2 (16.7%) were positive by rRT-PCR as well as both IgM and IgG ELISA. Of those with recent CHIKV infection, 53 (85.4%) were positive by both IgM and IgG ELISA, 5 (8.1%) were positive by IgM ELISA only, and 4 (6.5%) were positive by IgG ELISA only.
Duration of detection of diagnostic markers of CHIKV infection was plotted for all participants who had evidence of CHIKV infection by any method and reported recent symptoms of illness and a date of illness onset (n = 54) (Fig 3). All specimens collected before day four post-illness-onset (PIO) were positive by rRT-PCR. Detection of CHIKV nucleic acid by rRT-PCR decreased over time by day of specimen collection PIO, and by day 13 PIO no rRT-PCR-positive specimens were identified. Percent positivity by anti-IgM and IgG ELISA both increased according to day of specimen collection PIO. All specimens collected after week two PIO were IgM-positive, while all specimens collected after week three PIO were IgG-positive.
Following bivariate analysis, age and gender were not significantly associated with CHIKV infection (Table 1). Participants that reported having chronic joint disease or arthritis had nearly two-fold increased odds of having evidence of CHIKV infection. Reporting having had an acute illness in the past three months or having a household member that had an acute illness in the past three months were both associated with 14-fold increased odds of being laboratory-positive for CHIKV infection. No significant associations were found between CHIKV infection and housing type, having screened windows and doors, and reporting leaving doors or windows open regularly. Participants that reported using household air conditioning or citronella candles had two- or three-fold decreased odds of being laboratory-positive for CHIKV infection, respectively.
Following multivariable analysis that controlled for age and gender, female gender was associated with protection from CHIKV infection. Neither reporting having a chronic medical condition nor use of daily medications was associated with protection from CHIKV infection. Reporting having an acute illness or having a household member with an acute illness in the past three months both remained strongly associated with increased odds of CHIKV infection. Use of mosquito repellent and citronella candles remained associated with protection from CHIKV infection.
Of 99 participants that reported having an acute illness within the previous three months, 61 (61.6%) were laboratory-positive for CHIKV infection (Table 2). Median duration of illness in laboratory-positive participants was six days (range: 2–21 days). Following bivariate analysis, signs and symptoms associated with CHIKV infection in ill participants were fever, skin rash, arthralgia, and arthritis. Cough, rhinorrhea, and sore throat were associated with being laboratory-negative for CHIKV infection. No laboratory-positive symptomatic participants reported cough, rhinorrhea, or sore throat in the absence of fever or arthralgia. Following multivariable analysis, arthralgia and skin rash remained significantly associated with laboratory-positive symptomatic participants, and only retro-orbital eye pain remained significantly associated with laboratory-negative symptomatic participants. Headache, fever, arthralgia, myalgia, and chills tended to occur simultaneously more often among laboratory-positive participants, whereas cough, rhinorrhea, and sore throat occurred together more often among laboratory-negative participants (Fig 4).
Because of the prevalence of respiratory illness concurrent with chikungunya virus transmission, combinations of symptoms that grouped together following hierarchical cluster analysis and were most frequently reported among laboratory-positive participants following multivariate analysis were utilized to refine the definition of “symptomatic CHIKV infection” in order to minimize incorrect classification of symptomatically-infected participants (S2 Table). The maximal association of symptom combinations among laboratory-positive participants with concomitant minimization of association with laboratory-negative participants was arthralgia with skin rash or fever. This combination of symptoms yielded a symptomatic CHIKV infection rate of 62.5%, and was present among 6.8% of participants without evidence of CHIKV infection. This combination of symptoms was utilized in subsequent analyses to define “symptomatic CHIKV infection”.
Twenty-one (37.5%) participants, including two that had CHIKV nucleic acid detected by RT-PCR, were defined as having asymptomatic infection. Age was not significantly associated with asymptomatic infection (Table 3), nor was being a child (1 of 5 [20%] children with asymptomatic infection vs. 20 of 51 [39%] adults; OR = 0.30, 95% CI = 0.03–3.67). Neither sex nor reported chronic medical conditions was significantly associated with asymptomatic infection. Participants who reported having a household member with an acute illness within the previous three months more often had symptomatic infection (100% vs. 81%).
After again excluding the index case-patients, 22 (62.9%) of 35 symptomatic, laboratory-positive participants sought medical care. Seeking medical care for acute illness was associated with 3-fold increased odds of being laboratory-positive (Table 2). Neither hospitalization nor duration of illness was significantly associated with being laboratory-positive for CHIKV infection. No demographic or clinical characteristics were significantly associated with seeking medical care.
Of 22 laboratory-positive, symptomatic participants that sought medical care, five (22.7%) reported having been diagnosed with chikungunya (S3 Table). Neither age nor sex were significantly associated with correct reported diagnosis of chikungunya. All laboratory-positive, symptomatic patients diagnosed with chikungunya reporting having arthralgia in the hands, wrist, knee, ankle, and feet. Two (9.1%) laboratory-positive, symptomatic participants that sought medical care were reported to public health authorities.
By conducting household-based cluster investigations during the early months of the 2014 chikungunya epidemic in Puerto Rico that included 250 participants residing within 50 meters of a known chikungunya case-patient, we found that 30% of participants had evidence of CHIKV infection. Reporting having had an acute illness in the past three months and having a household member with an acute illness were associated with increased odds of infection, while use of either air conditioning or citronella candles were associated with decreased odds of infection. Symptoms significantly associated with CHIKV infection included arthralgia and skin rash. Nearly two-thirds of symptomatically-infected individuals sought medical care; however, less than one-quarter of these individuals were diagnosed with chikungunya, and one-tenth were reported to public health authorities as a chikungunya case. These findings demonstrate the utility of household-based cluster investigation to describe the epidemiologic and clinical characteristics associated with an emerging infectious disease and reasons for underreporting of clinically-apparent disease cases.
Serosurveys following chikungunya epidemics in Malaysia, Kenya, La Reunion, and Mayotte Island reported infection rates ranging from 37–75% [2–5]. Overall, 30% of participants in this investigation had evidence of CHIKV infection, which varied by cluster from 6.3% to 100%. These estimates likely do not reflect the final infection rates in these communities, as investigations were conducted during the first weeks of the epidemic in Puerto Rico where further CHIKV transmission likely occurred. A critical facet regarding interpretation of these results is that the objective of this investigation was not to determine the number of individuals infected with CHIKV during the indicated time frame, but rather to identify and compare the behaviors and characteristics of infected and uninfected participants. As such, the estimates of seroprevalence from previous studies and our findings are not directly comparable, as previous studies retrospectively measured rate of infection whereas this investigation collected a cross-sectional “snapshot” of infection rates during the initial stages of the epidemic.
Nonetheless, demographic and behavioral characteristics were able to be associated with susceptibility to or protection from CHIKV infection. Having a household member with acute illness in the last three months was strongly associated with increased the odds of infection, which supports the notion that, like DENV, CHIKV infections tend to cluster within households and neighborhoods [30]. Air conditioning use was associated with decreased odds of CHIKV infection, as has been reported in previous studies of DENV infection [31]. This finding may not be attributable to cooler temperatures in air conditioned homes but rather to buildings with air conditioning tending to have closed windows and doors and drier environments that result in lower rates of survival of Ae. aegypti mosquitos [32]. Use of citronella candles was also associated with reduced odds of CHIKV infection; however, the proportion of all participants using citronella candles was relatively small (17%), and thus likely did not contribute substantially to protection from infection on a population level. Past studies have shown varying and inconsistent levels of reduction of mosquito abundance associated with citronella candles [33, 34] as the quantity, concentration, and positioning of candles may play a role in their effectiveness [35].
By using findings from multivariable and hierarchical cluster analyses to identify arthralgia with fever or rash as being associated with CHIKV infection, we were able to more confidently define the rate of symptomatic CHIKV infection in this population as being 62.5%. Conversely, one-third of CHIKV-infected participants appear to have experienced asymptomatic infection, which is consistent with findings from past outbreaks that reported asymptomatic infection rates of 3–39% [36–38]; however, recent reports have suggested substantially higher rates of asymptomatic infection (e.g., 81%) [39]., Hence, further investigation including careful and potentially region-specific definitions of symptomatic infection is needed to determine factors influencing the rate of and progression to symptomatic CHIKV infection among diverse populations.
Most CHIKV-infected participants identified in this investigation that reported an acute illness in the past three months complained of characteristic symptoms of chikungunya: fever, arthralgia, myalgia, and skin rash [6]. Laboratory-negative CHIKV participants with recent illness were more likely to report symptoms of cough, rhinorrhea, or sore throat, suggestive of an upper respiratory infection. Symptomatic laboratory-positive CHIKV participants had three-fold increased odds of having sought medical care compared to participants that were laboratory-negative with reported recent illness. These observations together suggest greater disease severity of chikungunya as compared to common respiratory illnesses. Future studies should quantitate the burden of the chikungunya epidemic on health care resources in Puerto Rico.
Nearly two-thirds of symptomatically-infected patients sought medical care, demonstrating a relatively high rate of care-seeking behavior that may reflect the increased severity of arthralgia and myalgia as compared to patients with other etiologies of acute febrile illness. However, only one-quarter of chikungunya patients that sought care reported having been diagnosed with chikungunya, suggesting gaps in clinical suspicion of chikungunya. Other common diagnoses included more common etiologies of acute febrile illness including dengue and non-specific diagnoses such as viral syndrome. Because just one-tenth of clinically apparent chikungunya cases were reported as such to public health authorities, it is unclear how accurately the number of chikungunya cases reported to PRDH in 2014 reflects the true incidence of disease due to CHIKV infection. As with other reportable conditions for which passive case reporting is sub-optimal [40], including dengue [41, 42], the identified gaps in case detection via passive surveillance should be taken into consideration when making estimates of the burden of symptomatic and clinically-apparent chikungunya.
Strengths of this investigation included the ability to detect asymptomatic, sub-clinical, and clinically-apparent CHIKV infections, as well as the use of three different laboratory tests to identify current or recent CHIKV infection. It is therefore unlikely that any participants with CHIKV infection were not identified. Similarly, recall bias was likely to have been minimal since questionnaires captured events that had occurred within the prior three months. Conversely, a convenience sample of reported chikungunya cases was utilized to initiate cluster investigations, most of which were conducted in the San Juan metropolitan area. Moreover, factors that can influence both mosquito density (e.g., rainfall, temperature, humidity) [43] as well as the efficiency of CHIKV transmission (e.g., population density) [44] vary throughout Puerto Rico. For both of these reasons, our findings may not be representative of the entire population of Puerto Rico. Last, four laboratory-positive participants were defined as such by detection of anti-CHIKV IgG antibody only. Because lifetime travel history was not captured, it is possible that these individuals had been infected with CHIKV outside of Puerto Rico. Nonetheless, exclusion of these four individuals would not have significantly altered the observed associations.
Factors identified with protection from CHIKV infection identified in this investigation were household rather than individual behaviors, suggesting the importance of prevention practices in and around the household. Such behaviors should be encouraged in areas where Aedes mosquitoes are found. The clinical findings of this investigation highlight the need for increased capacity to identify chikungunya patients in out-patient settings. Due to the difficulty in utilizing signs and symptoms alone to differentiate patients with chikungunya from other febrile illnesses, clinical diagnosis and decision-making as well as case reporting would be aided by improved availability of rapid diagnostic tests.
|
10.1371/journal.pntd.0002628 | Cofactor-Independent Phosphoglycerate Mutase from Nematodes Has Limited Druggability, as Revealed by Two High-Throughput Screens | Cofactor-independent phosphoglycerate mutase (iPGAM) is essential for the growth of C. elegans but is absent from humans, suggesting its potential as a drug target in parasitic nematodes such as Brugia malayi, a cause of lymphatic filariasis (LF). iPGAM's active site is small and hydrophilic, implying that it may not be druggable, but another binding site might permit allosteric inhibition. As a comprehensive assessment of iPGAM's druggability, high-throughput screening (HTS) was conducted at two different locations: ∼220,000 compounds were tested against the C. elegans iPGAM by Genzyme Corporation, and ∼160,000 compounds were screened against the B. malayi iPGAM at the National Center for Drug Screening in Shanghai. iPGAM's catalytic activity was coupled to downstream glycolytic enzymes, resulting in NADH consumption, as monitored by a decline in visible-light absorbance at 340 nm. This assay performed well in both screens (Z′-factor >0.50) and identified two novel inhibitors that may be useful as chemical probes. However, these compounds have very modest potency against the B. malayi iPGAM (IC50 >10 µM) and represent isolated singleton hits rather than members of a common scaffold. Thus, despite the other appealing properties of the nematode iPGAMs, their low druggability makes them challenging to pursue as drug targets. This study illustrates a “druggability paradox” of target-based drug discovery: proteins are generally unsuitable for resource-intensive HTS unless they are considered druggable, yet druggability is often difficult to predict in the absence of HTS data.
| Parasitic worms like Brugia malayi cause widespread lymphatic filariasis (LF) in southeast Asia and sub-Saharan Africa. The adult worms causing most of the symptoms of LF are difficult to treat with existing drugs. As a possible step toward new LF drugs, we searched for inhibitors of the B. malayi cofactor-independent phosphoglycerate mutase (iPGAM), an enzyme thought to be critical to survival and development of this parasite. Despite testing over 100,000 compounds at each of two screening centers, we found only two compounds that consistently inhibited the B. malayi enzyme more strongly than the cofactor-dependent enzyme found in humans. These compounds have limited potency and are not especially great starting points for drug development. The 3-dimensional structure of iPGAM suggests that the active site is difficult to access from the surrounding solvent, which may partly explain our very low yield of inhibitors. We conclude that iPGAM may not be an ideal drug target in B. malayi or related organisms because it is difficult to inhibit with druglike compounds.
| For a protein to advance as a potential drug target, it should not only be important in pathogen survival and/or virulence, but also must be “druggable,” i.e., susceptible to modulation with drug-like compounds. Many otherwise promising proteins simply do not have binding pockets that lend themselves to therapeutic intervention [1], [2]. Historically, metabolic enzymes have been considered relatively druggable because (by definition) they bind small molecules, which can sometimes be mimicked by drugs [3]. Still, since enzymes' reactants may bear little resemblance to drugs in their hydrophilicity or other properties [4], a protein's amenability to therapeutic modulation cannot be guaranteed even if it is an enzyme. Another aspect of druggability concerns the chemical tractability of hit scaffolds identified through the HTS process. While assemblers of small molecule libraries strive to include tractable molecules, depending upon the source(s) of those libraries, not all scaffolds may be amenable to chemical modification.
Lymphatic filariasis (LF) is an infectious disease caused by the parasite nematodes Wuchereria bancrofti, Brugia malayi, and Brugia timori. Southeast Asia and sub-Saharan Africa harbor most of the world's ∼120 million current infections. It has been estimated that 40 million people suffer significant morbidity and/or disfigurement due to filariasis [5]. Most disfigurement is caused by adult-stage worms (macrofilariae), which are more impervious than immature worms (microfilariae) to current drugs such as diethylcarbamazine, ivermectin, and albendazole [6]. Thus, there is a strong need for new drugs, especially those that affect previously unexploited target proteins.
In the ongoing search for new LF drug targets, cofactor-independent phosphoglycerate mutase (iPGAM) has attracted significant interest. iPGAM catalyzes the interconversion of 2-phosphoglycerate (2-PG) and 3-phosphoglycerate (3-PG) in glycolysis and gluconeogenesis. The candidacy of iPGAM as a LF drug target is supported by several lines of evidence. RNAi knockdown of the C. elegans iPGAM, whose amino acid sequence is 70% identical to that of the B. malayi iPGAM, results in embryonic lethality or developmental defects (depending on the timing of the dsRNA injection), suggesting its functional importance in nematodes [7]. Selective inhibition of the parasite enzyme without harming the host should be possible, since mammals possess only a cofactor-dependent phosphoglycerate mutase (dPGAM), which differs greatly from iPGAM in structure, mechanism of action, and kinetic profile [8]. In particular, iPGAM is distinct from dPGAM in being catalytically active even in the absence of the cofactor 2,3-bisphosphoglycerate [9]. Finally, bacterially expressed iPGAMs from B. malayi and C. elegans have been purified and characterized [7], [8] and thus are readily available for high-throughput screening (HTS).
iPGAM's druggability – another key criterion in drug target prioritization, as noted above – has not yet been subjected to thorough experimental scrutiny, as far as we know. No potent inhibitors have been publicly reported to date, and the lack of a nematode iPGAM crystal structure further limits assessment of druggability. At the level of amino acid sequences, the closest iPGAMs with published structures are those from Bacillus stearothermophilus and Bacillus anthracis [10], [11], the former being 41% identical and 60% similar to the B. malayi iPGAM. The Bacillus structures show a monomeric protein with two domains: a phosphatase domain that removes the phosphate group from the glycerate substrate and a transferase domain that returns the phosphate to the substrate [10]. The two domains may twist to form open and closed conformations, with the open conformation apparently corresponding to an absence of substrate [11]. Thus, iPGAM's druggability could hinge partly on the fraction of time it spends in the open state, in which access to its active site is increased. However, this active site may not be especially druggable. The reactants (2-PG and 3-PG) are highly polar, and the nine amino acids that interact with them in the Bacillus stearothermophilus iPGAM (S62, H123, R153, D154, R185, R191, R261, R264, and K336) are all hydrophilic [10], [12]. Highly polar molecules (e.g., those with >5 hydrogen bond donors or >10 hydrogen bond acceptors [13]) are generally not “drug-like” in that they are poorly permeable through lipid membranes in the absence of a specific cellular transporter. The nine polar residues are all conserved in the B. malayi iPGAMiPGAM. Moreover, the iPGAM active site appears too small to accommodate additional, more hydrophobic moieties (Christophe Verlinde, personal communication). Thus, the active site of the B. malayi iPGAM is unlikely to be druggable in the sense of being bound by a sufficiently hydrophobic molecule.
The above analysis does not preclude the possibility of allosteric inhibition, however. In principle, allosteric inhibitors have the advantage of not needing to out-compete enzymes' substrates [14]. In practice, they have shown promise in studies of several infectious disease drug targets, including HIV reverse transcriptase [15], HIV integrase [16], hepatitis C virus NS5B polymerase [17], and Bacillus anthracis edema factor [18]. In the specific case of iPGAM, one can imagine an allosteric effector that forces the enzyme toward a closed conformation in the absence of substrate binding, thus preventing catalysis.
In the hope of finding B. malayi iPGAM inhibitors suitable for drug development, we performed high-throughput screens (HTS) of large compound collections at two sites: Genzyme Corporation (Waltham, MA, USA) and the National Center for Drug Screening (NCDS; Shanghai, China). In doing so, we took advantage of an innovative partnership (previously discussed in this journal [19]) between Novo Nordisk, the World Health Organization, and the NCDS. This partnership has enabled NCDS to screen a compound library formerly owned by Novo Nordisk, as exemplified by the present study and previous work [20].
iPGAM is dependent on divalent cations [8] and, like all enzymes, is responsive to changes in substrate concentrations. We screened for iPGAM inhibitors under conditions of abundant Mg2+ and abundant substrate (3-PG), which approximate normal cellular conditions. In particular, partial inhibition of glycolysis may lead to a buildup of [3-PG], and we hoped to discover inhibitors (whether competitive or allosteric) that would be effective even in the face of elevated substrate levels. Likewise, we did not pre-incubate the enzyme with compounds before adding substrate because we wanted to identify compounds that could inhibit iPGAM under physiological conditions, i.e., with substrate present.
As in previous work [7], [8], conversion of 3-PG to 2-PG (i.e., activity in the “forward”/glycolytic direction) was monitored via coupling of iPGAM to the downstream glycolytic enzymes enolase, pyruvate kinase (PyK), and lactate dehydrogenase (LDH). The primary assay readout was absorbance at 340 nm, reflecting NADH consumption by LDH, which in turn reflects upstream activity by PGAM, enolase, and PyK.
The two screening centers' workflows were somewhat different (Fig. 1). At Genzyme, compounds giving >40% inhibition against the C. elegans iPGAM were cherry-picked and re-tested; reconfirmed hits were then analyzed for chemical tractability [21]; compounds deemed tractable were tested against the B. malayi iPGAM and human dPGAM; and compounds showing relatively potent and selective inhibition of B. malayi iPGAM (IC50 <30 µM and lower than the IC50 vs. human dPGAM) were advanced to independent confirmation in a luminescence-based assay and efficacy testing with C. elegans larvae. At NCDS, the HTS was followed by dose-response assays of compounds that gave >20% inhibition in the initial screen. The best NCDS hit was then tested for activity against the P. falciparum dPGAM (since the human dPGAM was not readily available at the time), confirmed independently in the luminescence-based assay, and tested for efficacy against C. elegans larvae.
Primary efficacy testing was done against C. elegans rather than B. malayi for reasons of convenience. Any compounds showing good potency versus C. elegans in vitro would have been advanced to a model of B. malayi infection in gerbils [22].
Full-length, histidine-tagged iPGAM from C. elegans and B. malayi and full-length, histidine-tagged dPGAM from Homo sapiens and Plasmodium falciparum were expressed and purified as described previously [7], [8], [23]. 3-PG, adenosine diphosphate (ADP), nicotinamide adenine dinucleotide (NADH), PyK (from rabbit muscle, product P7768), LDH (from rabbit muscle, product L2500) and a PyK/LDH mixture (from rabbit muscle; product P0294) were purchased from Sigma-Aldrich (St. Louis, MO, USA). Enolase (from yeast; product #15515) was procured from Affymetrix (Santa Clara, CA, USA) and USB (Cleveland, OH, USA). Bovine serum albumin (BSA) was obtained from Amresco (Solon, OH, USA) and Sigma. 384-well plates came from BD Biosciences (Franklin Lakes, NJ, USA) and PerkinElmer (Waltham, MA, USA). Kinase-Glo was from Promega (Madison, WI, USA).
HTS assays were performed in 384-well plates (final volume: 50 µL per well). For the HTS at Genzyme, final assay concentrations were 30 mM Tris-HCl (pH 7); 5 mM MgSO4; 20 mM KCl; 1.5 mM 3-PG (3–4 times the Km; see Results below); 500 µM NADH; 3 mM ADP; 38.6 ng/mL C. elegans iPGAM; 2.5 Units/mL each of enolase, PyK, and LDH; 0.2% BSA and 10 µM of each compound tested (final DMSO concentration: 1%). For the HTS at NCDS, final concentrations were 30 mM Tris-HCl (pH 7.9); 5 mM MgSO4; 20 mM KCl; 3.5 mM 3-PG; 450 µM NADH; 2.5 mM ADP; 300 ng/mL B. malayi iPGAM; 2.265 U/mL enolase; 3.15 U/mL PyK; 4.71 U/mL LDH; 0.4 mg/mL BSA; and 5 µM of each compound tested (final DMSO concentration: 2%). At both locations the assay readout was absorbance at 340 nm. At Genzyme, absorbance data were taken as end-point readings on an Envision microplate reader (PerkinElmer) after 25 minutes of incubation at room temperature (∼20°C). At NCDS, data were collected kinetically (every 33 seconds) on a SpectraMax M2 microplate reader (Molecular Devices, Sunnyvale, CA, USA) during 15 minutes of incubation at room temperature following a 2-minute delay. HTS controls, representing the equivalent of enzyme inhibition, were wells with reduced [iPGAM] at Genzyme and wells with 50 µM or 200 µM tannic acid (discovered in preliminary Genzyme studies to reduce iPGAM activity) at NCDS.
To ensure robust, replicable results, hit compounds from the Genzyme and NCDS sites were shipped to the University of Washington for independent confirmation of inhibition of the B. malayi iPGAM using a separate batch of enzyme and a distinct assay readout. Catalytic assays were performed as described above except that ATP production by PyK was measured as luminescence at 528 nm following addition of Kinase-Glo, a luciferase-based reagent. A decrease in the slope of luminescence versus time was interpreted as inhibition of iPGAM.
Proprietary compound libraries housed at Genzyme and NCDS were screened. All compounds were pre-solubilized in 100% DMSO prior to use. The ∼220,000 compounds screened at Genzyme were taken from a library of ∼250,000 compounds from preferred vendors and internally synthesized compounds. This library emphasizes (A) drug-like or lead-like compounds, as judged by the criteria of compliance with the Rule of 5 [13] or Rule of 3 [24], compliance with Veber rules [25], and having similarity to known drugs according to fingerprint analysis; (B) heterocycles (∼2000 unique ring assemblies); and (C) natural product analogs (>7000). Over 40 screens have been conducted using this library, and most yielded usable chemical hits in other assays. The ∼160,000 compounds screened at NCDS were from a library of ∼325,000 synthetic compounds donated to NCDS by Novo Nordisk. The structural diversity of this library covers heterocycles, lactams, sulfonates, sulfonamides, amines, secondary amides, and natural product-derived compounds. Both libraries are intended to be relatively free of nonspecific aggregation-promoting compounds [26], but aggregators appearing as hits in the primary screen are generally filtered out during the hit confirmation process.
L1 arrested larvae were resuspended in S Basal medium, with E. coli supplied as food and test compounds added from DMSO stocks to final concentrations of 0 to 75 µM. Worms were incubated in liquid culture in multi-well plates (100 µL per well) at 20°C for three days and then scored for growth defects/arrest. At least 16 wells (each containing 20 to 40 worms) were scored for each concentration of each compound. Only one generation of worms was followed.
The iPGAM from B. malayi has a GenBank ID of AY330617 and a UniProt ID of Q4VWF8. The iPGAM from C. elegans has a GenBank ID of AY594354 and a UniProt ID of G5EFZ1.
Our attempts to find specific inhibitors of the B. malayi iPGAM met with extremely limited success. The Genzyme screen tested ∼220,000 compounds at a concentration of 10 µM; it identified 110 confirmable hits against the C. elegans iPGAM, but only one of these compounds (Fig. 2) passed all of the follow-up steps shown in Fig. 1. The yield of the ∼160,000-compound screen at NCDS was equally low. 233 compounds (0.15%) initially appeared to show ≥20% inhibition of the B. malayi iPGAM at a concentration of 5 µM (Fig. 3), but of these, only one compound consistently gave an IC50 <50 µM (Fig. 4).
In theory, such an exceedingly low frequency of confirmable hits could reflect problems with (A) the performance of the assay, (B) the makeup and/or handling of the compound libraries, and/or (C) the enzymes being screened. We now address each of these possibilities.
Regarding assay performance (A), the standard measure of HTS quality is the Z′-factor, which reflects the means and variability of inhibited and uninhibited samples. Z′-factors range from 1 to below 0, with values above 0.5 indicating a robust assay [27]. Our mean Z′-factors were 0.65 for the Genzyme HTS and 0.51 for the NCDS HTS. Visual inspection of the HTS data from Genzyme and NCDS (Fig. 5) likewise showed large, clean separations between positive and negative controls. Therefore our assay seemed adequate for detecting iPGAM inhibitors.
Regarding compound libraries (B), there are two potential issues: the chemical “space” covered and compound stability. While library builders seek to ensure broad coverage of potentially drug-like chemical space, a lack of hits could reflect limits in this coverage. We also cannot be 100% certain that there were no problems with the storage, handling, or dispensing of the compounds tested. However, these same Genzyme/NCDS libraries have been used to identify potent hit compounds in screens against other target proteins (e.g., [20], [28]). Therefore it seems unlikely that the low hit rate was caused by systematic degradation or incorrect dispensing of compounds.
Finally, regarding the enzymes screened (C), enzymes were expressed in E. coli from plasmids originally generated by New England Biolabs. SDS-PAGE analysis gave the expected molecular masses of ∼57 kDa (B. malayi) and ∼59 kDa (C. elegans), as previously reported [7]. Newly determined Km's for 3-PG of 0.37 mM (B. malayi iPGAM) and 0.38 mM (C. elegans iPGAM) were consistent with previously reported values [8] of 0.35 mM and 0.51 mM, respectively. Likewise, specific activities of the newly purified enzymes were similar to those reported previously (data not shown). Therefore, there is no indication that the enzyme stocks used in the HTS were misfolded or denatured, which would have hampered the search for inhibitors of the normal well-folded enzymes. However, it is notable that the Genzyme HTS used the C. elegans iPGAM rather than the B. malayi iPGAM, and that only a subset of the hits against the C. elegans iPGAM were subsequently tested against the B. malayi iPGAM. Given the strong amino-acid similarity (70% identity, 82% similarity) of the two enzymes [7], we would expect inhibitor profiles for these enzymes to be similar as well. Still, inhibitors specific for the B. malayi iPGAM might be missed in screening with the C. elegans iPGAM. The NCDS HTS directly addressed this possibility; it showed that using the B. malayi iPGAM in the primary screen did not appreciably increase the rate of hits against this enzyme.
Given these considerations, we believe that the low hit rates do not reflect any major experimental limitations, but instead may reflect iPGAM's poor druggability. There currently is no crystal structure for either the B. malayi or C. elegans enzymes; however, sequence homology [7] suggests that both of these are very similar to that of Bacillus stearothermophilus, for which the crystal structure has been resolved [12]. That structure suggests a peptide “gate” over the active site (confirmed in subsequent studies [29]) which may limit accessibility to potential inhibitors. Similar problems have been encountered for other enzyme targets. The crystal structure of cytosolic phospholipase A2 revealed that it contains an α-helical “lid” that can fold to cover the active site [30]. This feature has made it very difficult to obtain broad structural classes of inhibitors ([31] and John Leonard, personal communication).
The compounds shown in Fig. 2 may be useful as chemical probes in future studies of iPGAM, and thus represent an important outcome of our study. These compounds' potential application in drug development depends on several criteria, such as (A) potency against the target enzyme and target parasite, (B) chemical properties related to druglikeness and medicinal chemistry potential, (C) specificity of inhibition, and (D) activity of related compounds.
Regarding potency (A), neither hit compound had an IC50 vs. the B. malayi iPGAM of <10 µM, nor was either efficacious against C. elegans larvae at 25 µM. It has been proposed that, for anti-helminth hits to merit development into possible lead compounds, they should inhibit helminth motility by 50 to 100% at concentrations below 2 to 10 µg/mL [32], which would be 9 to 45 µM for a compound whose molecular weight is 223 (the average for the two compounds shown in Fig. 2). In this concentration range, our hits showed little or no activity against live worms. While factors like poor drug solubility, limited uptake into the worm gut, and limited transport across the cuticle could all affect activity in this assay, these compounds are not especially appealing as starting points for drug development.
Regarding chemical properties (B), both hits are reasonably drug-like and amenable to chemical synthesis and modification; e.g., they do not have limitations such as a high hydrophilicity or numerous chiral centers. Regarding off-target effects (C), PubChem BioAssay [33] shows that compound Genz-2 (PubChem CID 3614032) does not inhibit the other 9 targets against which it has been tested, while no bioassay data are available for NCDS-1 (PubChem 606970). Finally, of the five NCDS-screened compounds based on a scaffold of 1,10-phenanthroline, only the hit compound itself (5-amino-1,10-phenanthroline) gave any noticeable inhibition of the B. malayi iPGAM (D). This suggests that most changes to this compound's structure would eliminate its activity against iPGAM, and therefore that improvement of the hit compound's potency through medicinal chemistry would be difficult.
Thus, large-scale efforts at two different screening centers collectively failed to identify any high-priority compounds for drug development studies. In the absence of published empirical evidence that nematode iPGAMs can be potently and specifically modulated by drug-like molecules, advancing them as drug targets appears highly challenging and risky. We would also advise caution in the pursuit of non-nematode iPGAMs like the Trypanosoma brucei iPGAM [34], [35] as drug targets, since their druggability remains uncertain.
Our finding of poor nematode iPGAM druggability contrasts with a recent modeling study of the Wuchereria bancrofti iPGAM [36], which presents drug-like molecules proposed to be likely inhibitors. The authors developed a 3-D model of the W. bancrofti iPGAM, using the B. stearothermophilus iPGAM [12] as a template. They found that 63 residues were conserved among iPGAMs and that 53 residues contribute to binding pockets as defined by Q-SiteFinder [37]; the 19 residues belonging to both sets were termed the “common amino acid residues” (WB-iPGM19cr). When a virtual library of 2,344 small molecules were presented to the W. bancrofti iPGAM model in docking simulations, 65 were predicted to interact with at least one of the WB-iPGM19cr residues, and 8 of these 65 (each linked to exactly one WB-iPGM19cr residue) are considered to have good ADME/T properties and are “strongly recommended for further clinical trials.” We applaud this interest in the W. bancrofti iPGAM but note that none of the predicted inhibitors were tested experimentally.
Having invested considerable resources in screening nematode iPGAMs, only to find that they do not appear druggable, we must ask whether this disappointing outcome could have been predicted. The crystal structure of the B. stearothermophilus iPGAM does indicate a “gate” region peptide over the active site that could prevent access by potential inhibitors (see above [12], [29]). More recent studies on the crystalized structure of Tryanosoma brucei iPGAM also indicate that when crystalized in the presence of substrate, the substrate is buried and is not solvent-accessible, again suggesting a gate-like fold over the active site [35] that may increase the difficulty of finding inhibitors.
We also considered the possibility that allosteric inhibitors might be discovered as part of the screening process. A priori prediction of allosteric effects remains challenging; as summarized in one review, “In most cases, the novel allosteric binding sites could not have been predicted from the unliganded structure” [38]. When studying a protein without a solved crystal structure, such as the B. malayi iPGAM, the challenge increases further.
In conclusion, we suggest that target-based drug development suffers from a frustrating paradox: proteins are generally unsuitable for resource-intensive HTS unless they are considered druggable, yet druggability is often difficult to estimate in the absence of HTS data. Although improved druggability predictions [39] may eventually offer a way out of this paradox, a moderate level of risk currently appears unavoidable in the screening of many novel protein targets.
|
10.1371/journal.pntd.0006955 | Rabies vaccine initiation and adherence among animal-bite patients in Haiti, 2015 | Approximately 59,000 people die from rabies worldwide annually. Haiti is one of the last remaining countries in the Western Hemisphere with endemic canine rabies. Canine-mediated rabies deaths are preventable with post-exposure prophylaxis (PEP): wound treatment, immunoglobulin, and vaccination. In countries where PEP is available, variability in healthcare seeking behaviors and lack of adherence to recommended treatment guidelines could also contribute to these deaths. Yet, few studies have addressed these issues.
We examined animal-bite reporting and assessed adherence to treatment guidelines at nine healthcare facilities in Haiti. We analyzed individual-level, de-identified patient data (demographic characteristics, geographic location, healthcare facility type, vaccine administration, and bite injury information) using descriptive analyses and logistic regression to examine factors associated with receiving PEP.
During the 6 month study period, we found 2.5 times more animal-bite case-patients than reported by the national surveillance system (690 versus 274). Of the 690 animal-bite patients identified, 498 (72%) sought care at six PEP providing facilities. Of the case-patients that sought care, 110 (22%) received at least one rabies vaccine. Of the 110 patients, 60 (55%) received all five doses. Delays were observed for three events: when patients presented to a facility after an animal-bite (3.0 days, range: 0–34 days), when patients received their fourth dose (16.1 days, range: 13–52 days), and when patients received their fifth dose (29 days, range: 26–52). When comparing vaccination status and patient characteristics, we found a significant association for bite location (p < .001), severity rank score (p < .001), geographic location (p < .001), and healthcare facility type (p = .002) with vaccination.
High levels of underreporting identified here are of concern since vaccine distribution may, in part, be based on the number of animal-bite cases reported. Given that the Haitian government provides PEP to the population for free and we found animal-bite victims are seeking care in a timely manner─ reducing rabies deaths is an achievable goal.
| Human rabies deaths are preventable with timely provision of rabies post-exposure prophylaxis: wound treatment, rabies immunoglobulin, and vaccinations. In countries where resources are available, variability in healthcare seeking behaviors and lack of adherence to recommended treatment guidelines may also contribute rabies deaths. In our study, we examined animal-bite reporting to the national surveillance system, and patient’s adherence to treatment guidelines at nine healthcare facilities in Haiti, a high rabies burden country. Our study found 2.5 times more patients than reported by the national surveillance system. This underreporting can unintentionally impact rabies awareness and the allocation of resources to animal-bite victims. Furthermore, our study found that among patients who received care, a majority of them are seeking care in a timely manner, suggesting that reducing rabies deaths is an achievable goal in Haiti.
| Approximately 59,000 people die from rabies annually worldwide [1]. Yet, these deaths are preventable with timely provision of rabies post-exposure prophylaxis (PEP): wound treatment, rabies immunoglobulin (RIG), and vaccinations [2]. Many of these deaths occur in canine rabies endemic countries where PEP may be limited or inaccessible to animal-bite victims [3]. In countries where PEP is available, variability in healthcare seeking behaviors and lack of adherence to recommended treatment guidelines could also contribute to these deaths. Research has shown that characteristics such as age, gender, and geographic location of residence are associated with seeking care for animal bites and receiving PEP [4–11]. Ensuring adequate and timely PEP administration for animal-bite patients in contact with a suspect rabid animal is one cornerstone of the World Health Organization’s (WHO) goal of eliminating rabies deaths. Therefore, understanding the management of rabies exposures is crucial to preventing future deaths.
Haiti is one of the last remaining countries in the Western Hemisphere with endemic canine rabies [1, 12–16]. From 2009–2012, an average of four canine and seven human rabies cases were reported by Haiti’s national surveillance system (NSS), however, studies have shown these reports underestimate the true burden of rabies [1, 13–15, 17]. One community-based surveillance program with active bite-case investigation, which was operated in three Haitian communes, detected more rabid animals than reported for the entire country by NSS [13]. Additionally, findings from an active case investigation estimated the rabies-associated mortality to be 0.67 cases/100,000 persons compared to 0.07cases/100,000 persons reported nationally [17]. Furthermore, a modeling study suggests 130 human deaths per year are attributable to rabies in Haiti [1]. Underreporting can unintentionally impact rabies awareness and the allocation of resources such as PEP to animal-bite victims [13, 17]. Few studies have examined treatment practices and adherence to vaccine schedules for animal-bite patients.
For this present study, we were interested in understanding animal-bite reporting, and adherence to treatment guidelines at healthcare facilities. To achieve this, we piloted active case-finding for all animal-bite patients presenting to selected health facilities in Haiti.
We included nine healthcare facilities in our evaluation. These were selected by the Ministère de la Santé Publique et de la Population (MSPP). Three of these healthcare facilities were commune-level health centers and six facilities were hospitals. The facilities were located in six communes: Carrefour (2 sites; commune-level health center and hospital), Croix-des-Bouquets (1 site; hospital), Hinche (2 sites; commune-level health center and hospital), Leogane (1 site; hospital), Port-au-Prince (1 site; hospital), and Saint-Marc (2 sites; commune-level health center and hospital). All communes had at least one facility that provided PEP. Three communes (Carrefour, Hinche, Saint-Marc) also had facilities that did not provide PEP, but offered post-bite medical care. All facilities were located in high population density areas.
Rabies cell-cultured vaccines are donated to MSPP, often with assistance from PAHO, and are distributed to approximately 16 health facilities in Haiti (at least one facility per department). These vaccine are provided free of charge to patients when they are available. Stock-outs are frequent, in which case bite-victims must either purchase vaccines from private pharmacies at a cost of approximately $20 per dose, or travel to a different public facility to seek vaccine. Immunoglobulin is not provided at any public facilities, and are only available at private facilities for a fee or from the national stockpile with written permission from the Ministry of Health.
Our team visited each facility to conduct a retrospective review of all patients presenting for an animal-bite between April 1, 2015–September 30, 2015. We entered each patient’s individual level, de-identified data from the facilities’ paper-based animal-bite registries into an electronic database for analysis. Information from patients included demographic characteristics (age, gender), vaccine administration (dates of vaccination and dose number), and information related to the bite injury [bite location, bite type (e.g. multiple, single), animal, animal’s behavior]. WHO animal contact category was not collected at the facility, therefore it was unavailable for analysis. The data was reviewed and verified by a second individual for accuracy. Blank data were considered missing.
We defined patients receiving at least one of five intramuscular doses of the rabies vaccine as vaccinated. Patients receiving all five doses of vaccine were considered fully vaccinated. We categorized the nine facilities as PEP facilities (provided vaccine during the study period) or non-PEP facilities (provided wound treatment but no vaccine during the study period). None of the facilities provided rabies immunoglobulin during the study period. To categorize patient bite severity, we calculated a severity rank score for each patient as the sum of three variables: bite location (0 = missing, 1 = lower body, 2 = upper body, and 3 = multisite or head), bite type (0 = missing, 1 = single, and 2 = multiple), and animal’s behavior (0 = missing or calm, and 2 = aggressive). The severity rank score was then classified into: low (0 or 1 points), medium (2 or 3 points) and high (4 or greater points).
We calculated the number of days patients took to seek medical care and initiate vaccine. Deviations were defined by divergence from the recommended WHO PEP regimen and the degree of divergence was quantified by the number of off-schedule days. For example, a patient that waited to seek care three days after exposure would be considered a three-day deviation or a patient receiving their vaccine a day earlier than the recommended day would be considered a one-day deviation. Inaccurate dates such as documented date of vaccination occurring before the documented date of exposure were excluded from analysis.
For our analysis, we were interested in factors that influenced rabies vaccination. To identify associations with vaccinated (at least one dose) and unvaccinated patients, we used a chi-square analysis and multivariable logistic regression. Our independent variable was patient vaccination status: yes or no. Dependent variables in the logistic regression models included: age, gender, animal type, severity rank score, facility location, and type of healthcare facility. To arrive at a final model, we used a backwards elimination. Due to small sample sizes among stratified characteristics we did not include interaction effects. Analyses were performed in R v3.3.2 statistical analysis software and the mapping was performed in ArcGIS v10.3.
To compare the findings from our site visits and case-finding in nine healthcare facilities, we compared the total number of animal-bite cases reported to Haiti’s NSS during the same study period (April 1, 2015–September 30, 2015) for the same communes.
We obtained Institutional Review Board (IRB) approval from the Ministère de la Santé Publique et de la Population (MSPP) in Haiti and the IRB at the U.S. Centers for Disease Control and Prevention. The data were anonymous and de-identified.
During the six-month period, our study identified 690 animal-bite patients at nine preselected healthcare facilities in six different communes (Fig 1, Table 1). The majority of animal-bite patients (60%) sought care in two of the six communes: Carrefour (n = 280, 40.6%) and Port-au-Prince (n = 134, 19.4%). Of the six communes where animal-bite patients presented for care, two communes, Saint-Marc and Leogane did not report any bite cases to the national surveillance system (NSS). In contrast with our findings, Haiti’s NSS reported only 274 animal-bite cases during the six-month period from these communes. Overall, 2.5 times more case-patients were identified from this study compared to the NSS (274 versus 690).
Of the 110 patients that were vaccinated, 60 (54.6%) patients received all five doses (Table 2). Vaccine completion increased with increasing age: 54% 0–17 years, 60.7% 18–34 years, 70.8% 35–60 years, and 87.5% 61 years and older.
Patients living in Carrefour and Croix-des-Bouquets had vaccination completion rates over 70%, while Port-au-Prince and Saint-Marc were less than 60%. None of the 16 patients in Hinche and Leogane completed their fifth dose.
Patients sought care in a timely manner and generally followed the WHO PEP schedule (Table 3). Delays were observed for three events: when patients presented to a facility after an animal bite (3.0 days, range: 0–34 days), when patients received their fourth dose (16.1 days, range: 13–52 days), and when patients received their fifth dose (29 days, range: 26–52). No statistically significant delays were observed when data were stratified for characteristics such as age, gender, severity rank score, and area.
When comparing deviations from the WHO recommendations, we noted that patients residing in Leogane were reported to have the largest deviation of an average of 6.3 days (range: 0–67). The mean deviations were not statistically different by age, gender, severity rank score, and location (Table 4).
Of the 690 animal-bite patients identified, 498 (72%) sought care at the six PEP facilities (Table 1). Of these, 110 (22%) animal-bite patients received at least one rabies vaccine. Rabies immunoglobulin was not documented for any of the patients.
When comparing vaccination status and variables of interest, we found a significant association (p < 0.05) for bite location (χ2 = 53.06; p < .001), severity rank score (χ2 = 20.38; p < .001), geographic location (χ2 = 53.06; p < .001), and healthcare facility type (χ2 = 9.76; p = .002). We found no association between vaccination status for age group (χ2 = 4.98; p = 0.289), gender (χ2 = 3.56; p = 0.169), bite type (χ2 = 7.14; p = 0.028), animal type (χ2 = 7.94; p = 0.019), and animal behavior (χ2 = 5.19; p = 0.075).
Results from our logistic regression analysis, comparing characteristics of vaccinated and unvaccinated patients at the six PEP facilities (n = 498) are shown in Table 5. The odds of receiving vaccine were 1.64 (95% CI: 1.01, 2.68) times greater for male patients than female patients. As the severity rank score increased (low versus high, medium versus low), so did the odds of receiving vaccine. The odds of receiving vaccine was 8.10 (95% CI: 3.26, 22.91) times greater for patients with a high severity compared to a low severity score (Table 5). The odds of receiving vaccine was 2.72 (95% CI: 1.04, 8.11) times greater for patients with a medium severity compared to a low severity score. Additionally, the odds of receiving vaccine were higher [OR: 5.14 (95% CI: 2.77, 9.85)] for patients residing in Port-au-Prince than Carrefour.
We examined animal-bite treatment practices, patient adherence to vaccine, and animal-bite surveillance reporting in Haiti. We found patient adherence to the five-dose vaccine schedule was relatively high, but it was unclear why some high-risk patients went unvaccinated. Our active animal-bite case-finding uncovered issues with reporting compared to Haiti’s NSS; we identified 2.5 times more patients (690 patients from active case-finding versus 274 cases reported to NSS) with several communes not reporting to the NSS. Previous studies have also documented similar underreporting in Haiti [13, 17]. An active community investigation found two probable human rabies cases and 16 animal-bite victims that were not originally captured in NSS [17]. Underreporting is likely due to a lack of dedicated resources or a breakdown in reporting to NSS and limited healthcare seeking behavior among animal-bite victims. [13, 17]. Discrepancies in animal-bite cases may have severe consequences for vaccine allocation in areas that do not report or underreport cases.
In keeping with our findings, previous research found a higher frequency of animal bites among children [18]. Despite, the greater frequency of children found in our study, they had a lower initiation and completion rate compared to other age groups. In some instances, gender has also been documented as an important characteristic of healthcare-seeking behavior for animal bites and receiving vaccine [4]. Our study found a statistical difference between gender and vaccine administration but not vaccine adherence. Our experience working in Haiti leads us to believe that males are more likely to respond to an aggressive (potentially rabid) dog in a community, and may support why males in our study population are more likely to receive the first dose of vaccine.
Our logistic regression showed that independent predictors of receiving vaccination were gender, facility location, and severity rank score. While our logistic regression found higher odds of being vaccinated with a high severity rank score compared to a lower score, it is concerning that some high-risk patients did not initiate vaccination. For example, only half (n = 10) of patients with a multisite or head injury received vaccine. Given the lack of documentation, it is difficult to explain the reasons why this cohort did not receive vaccination. The dates suggest that there were no vaccine supply issues, as patients who were vaccinated presented at the same day or within a few days as those that did not receive vaccine. We believe a more comprehensive study is needed to better understand vaccine delivery and the rationale for not vaccinating high-risk individuals.
Adherence to the rabies vaccine schedule is associated with better survival after a rabies exposure. We found a surprisingly high adherence to completion of vaccine administration; 54% of animal-bite patients completed all five doses in our study. In contrast, lower adherence was reported in Iran (16.3%-18.7%), Tanzania (28%), and Cote d’Ivoire (47.3%) [6, 8, 19, 20]. Two studies in Nigeria had similar or higher completion coverage rates among animal-bite patients identified through a rabies lab registry and a pediatric hospital (60.7%) [21, 22]. However, we found a longer delay in seeking treatment after an exposure in our study population compared to similar studies in India, and Iran, where they reported a 24-hour delay of treatment after exposure among their study populations [5, 6, 10, 11]. Some studies found delays associated with age groups and gender, while our study did not [5, 6, 11]. The delays reported by those studies cited school-age patients not wanting to miss school in China, and women having less access to medical care due to cultural practices in Iran.
This study builds on the current rabies work in Haiti. Since the completion of this study, the Pan American Health Organization had developed and implemented a rabies training program for medical providers at health facilities. If this study was repeated at the same facilities, our findings could serve as a baseline on whether patient care, vaccine completion and adherence, and NSS reporting had improved. Three other studies found similar issues with adherence and compliance among bite-victims outside of Port-au-Prince [17, 23, 24]. Fenelon et al, reported that only 31% bite victims in Petionville (suburb outside of Port-au-Prince) initiated the vaccine series. Etheart et al, and Tran et al, found lack of compliance among bite-victims counselled by public health officials. Coupled with our findings, these studies suggests the need for comprehensive studies to understand the underlying factors for low compliance. This may help to identify strategies for increase vaccine uptake.
This study had limitations. First, the evaluation was based on a convenience sample, therefore may be potentially biased and not representative of the general population. For example, only one or two of the facilities in each commune were captured for this evaluation, therefore patients may have sought care elsewhere or were referred to another medical facility for vaccine. We attempted to address this by checking the names and birthdates for duplicate case-patients at each commune. We found one patient had sought care for treatment and vaccine at two different sites within the same commune. It is possible that other patients could have received vaccine or went home. Secondly, there was a large proportion of missing data for some of the variables. Similar studies have documented the same challenges with bite registries [7, 9].
In conclusion, we found important characteristics associated with vaccine administration and animal-bite exposures. The high levels of underreporting that we identified are of concern since the distribution of vaccine may in part based on the burden of animal-bite cases. Given that the Haitian government provides PEP to the population for free, we found animal-bite victims are generally seeking care in a timely manner. With improved surveillance and a better understanding of the underlying issues of low compliance ─ reducing rabies deaths is an achievable goal.
|
10.1371/journal.pcbi.1004122 | Kinetically-Defined Component Actions in Gene Repression | Gene repression by transcription factors, and glucocorticoid receptors (GR) in particular, is a critical, but poorly understood, physiological response. Among the many unresolved questions is the difference between GR regulated induction and repression, and whether transcription cofactor action is the same in both. Because activity classifications based on changes in gene product level are mechanistically uninformative, we present a theory for gene repression in which the mechanisms of factor action are defined kinetically and are consistent for both gene repression and induction. The theory is generally applicable and amenable to predictions if the dose-response curve for gene repression is non-cooperative with a unit Hill coefficient, which is observed for GR-regulated repression of AP1LUC reporter induction by phorbol myristate acetate. The theory predicts the mechanism of GR and cofactors, and where they act with respect to each other, based on how each cofactor alters the plots of various kinetic parameters vs. cofactor. We show that the kinetically-defined mechanism of action of each of four factors (reporter gene, p160 coactivator TIF2, and two pharmaceuticals [NU6027 and phenanthroline]) is the same in GR-regulated repression and induction. What differs is the position of GR action. This insight should simplify clinical efforts to differentially modulate factor actions in gene induction vs. gene repression.
| While the initial steps in steroid-regulated gene induction and repression are known to be identical, the same cannot be said of cofactors that modulate steroid-regulated gene activity. We describe the conditions under which a theoretical model for gene repression reveals the kinetically-defined mechanism and relative position of cofactor action. This theory has been validated by experimental results with glucocorticoid receptors. The mode and position of action of four factors is qualitatively identical in gene repression to that previously found in gene induction. What changes is the position of GR action. Therefore, we predict that the same kinetically-defined mechanism usually will be utilized by cofactors in both induction and repression pathways. This insight and simplification should facilitate clinical efforts to maximize desired outcomes in gene induction or repression.
| The initial steps by which steroid receptors induce or repress target gene transcription are the same. After steroid binding to the intracellular receptor, the resulting complex is activated/transformed to a form with increased affinity for DNA [1] and is concentrated in the nucleus, where it is recruited to DNA sequences that are usually near the regulated genes. Cofactors and comodulators assist or impede the transcriptional activity of DNA-associated steroid receptors [2,3]. Beyond this, it is currently not possible to predict the transcriptional outcome for any specific combination of gene, receptor, and cofactor/comodulator. In most cells, a given steroid-bound receptor will induce one set of genes while repressing another set under otherwise identical conditions. For certain genes, the same receptor-steroid complex activates transcription in one cell line while repressing it in another cell line [4]. Similarly, selected cofactors increase the activity of one steroid receptor while reducing the activity of another receptor [5]. In some cases, different cofactors cause the same gene to be induced or repressed [6,7]. In other cases, the same cofactor may interact with the same steroid receptor to augment the induction of one gene but increase the repression of another gene [6–10]. Thus no relationship between induction vs. repression and presence of particular promoter/enhancer bound factors and cofactors has yet emerged [11]. The DNA sequence to which the receptor is recruited, either by direct DNA binding or by tethering to another DNA-bound molecule, can often indicate the resultant activity of induction or repression respectively [11–13]. Even this categorization, though, is not precise as repression can occur from GR binding directly to DNA [14,15] and the outcomes can depend upon whether the cofactor binds to DNA-bound glucocorticoid receptor (GR) or GR is tethered to DNA-bound cofactor [16].
An unresolved question is whether the underlying mechanism of each transcriptional component is the same or changes with the direction of gene expression output (i.e., increase in induction vs. decrease in repression). The answers are vital because efforts to modify the responses with selected factor combinations, be it in isolated cells or human patients, will depend critically on whether the mechanism of each component is constant or varies with the specific mixture of factors. Current attempts to address these issues have been inhibited by insufficiently precise methods of analysis. Thus, many cofactors are classified as either coactivators or corepressors based solely upon their ability to increase or decrease respectively, the level of steroid-mediated gene expression [2,17]. Unfortunately, while such descriptions are operationally useful, they are mechanistically uninformative. It is well known from enzyme kinetics that an enzymatic inhibitor can increase the total response while an enzymatic activator can lead to decreased output [18–20].
A more precise and quantitative understanding of the mechanisms of gene transcription is required to resolve these issues. Mathematical modeling provides one solution; and, a theory has been developed recently to understand the underlying mechanisms of factor action during steroid-regulated gene induction. The theory is based on the fact that the dose-response curve for gene induction is non-cooperative with a Hill coefficient of one [20]. This shape of the dose-response curve has also been variously described as a Michaelis-Menten function, hyperbolic dose-response, first-order Hill plot, and first-order Hill dose-response. The essential feature is that it is in the mathematical family of linear-fractional functions.
The theory enables one to determine the kinetically-defined mechanism of factor action and the position of factor action in the many steps of the overall signaling cascade. This position is specified relative to both another competing factor and a steady-state analogue of a rate-limiting step called the concentration limited step (CLS) [20–27]. The concentrations of bound factors after the CLS are negligible compared to their unbound concentrations (which could be involved in other reactions as long as they are readily available). A factor can act kinetically like an enzymatic activator, which we call accelerator, or an enzymatic inhibitor, which we call decelerator, depending on how the factor participates in the reaction [23] (see Fig. 1). These classifications describe how the factor alters a specific reaction step independently of the direction of change in the observed product. In all cases investigated so far, the reporter acts as an accelerator at the CLS and thus serves as an invariant positional landmark in the otherwise poorly defined landscape of reactions in steroid-regulated gene induction [23,25–27].
Here, we extend the mathematical theory and model to GR-regulated gene repression as influenced by varying concentrations of competing cofactors. The relevant mass-action equations have been used to relate the graphs of several reaction parameters derived from the dose-response to the kinetically-defined mechanism and position of action of each of the two competing cofactors. Using the extensively characterized system of GR repression of AP1 induction in U2OS.rGR cells with a transiently transfected synthetic reporter [8,15,28,29], we show that four factors (the reporter gene, TIF2, and two small molecules [30]) have the same kinetically-defined mechanism and position of action as in GR-regulated gene induction. The difference between induction and repression is the position of GR action. This suggests that many transcriptional cofactors/comodulators are mono-functional and similarly modulate basic steps in gene expression irrespective of the directional change in gene product levels.
Experimentally, the dose-response of gene activity A in steroid-regulated repression has been found to be non-cooperative with a Hill coefficient of one (see Fig. 2): i.e.
A=Amax+Amin/IC50[S]1+1/IC50[S]
(1)
where Amax is the maximal activity with no added steroid, Amin is the minimum value of activity with saturating steroid concentrations, and IC50 is the concentration of steroid for half-maximal suppression. Added cofactors can change the parameters Amax, Amin and IC50 while preserving the shape of the dose-response (1). Our goal was to develop a theory for gene repression that explains why the dose-response has the shape given in (1) and what that implies for the actions of the added cofactors that change the parameters. We use the fact that the linear-fractional shape of the dose-response curve puts severe constraints on the possible biochemical kinetic schemes involved in gene repression, as it did for gene induction [20]. We then derive formulas that can be compared to the data to make predictions for the actions of cofactors.
Our data for the dose-response curve is based on the contributions from many copies of the induced gene in multiple cells. Hence, it is possible that the dose-response for a single gene could be different from the averaged population response. However, single-molecule imaging experiments for a gene induced by a nuclear receptor show that transcripts are produced in on-off stochastic bursts, where the bursting is well modeled as a stochastic process (i.e. random telegraph model [31,32]) and the burst probability is similar between cells. The frequency (stochastic intensity) of the stochastic bursts of individual genes follows a non-cooperative dose-response with respect to the inducing agonist [31,32]. Thus, the mean activity of a single gene follows the same non-cooperative dose-response and our theory explains the behavior of this mean.
We first present a theory for which gene induction has a non-cooperative dose-response since the theory for gene repression hinges directly on it. Gene expression involves the binding of molecules, protein, and DNA into complexes that lead to transcription. We model this as a sequence of complex building reactions, Yi−1+Xi↔qiYi, where we call the Y variables products, the X variables accelerators, and the q’s are equilibrium or affinity constants. Decelerators of various types can also inhibit each of these reactions. Fig. 1 shows the general reaction scheme at each step. The accelerator X in Fig. 1 acts like an enzymatic activator and the decelerator D acts like an enzymatic inhibitor [23]. The kinetic scheme is stochastic and characterized by a probability distribution for the reactants. However, to simplify the calculations, we consider the mean field limit obeying the law of mass action and assume that the gene activity is proportional to the mean concentration of one or a set of products. The dose-response curve is the gene activity as a function of the concentration of the initial product [Y0] (i.e., agonist steroid).
The theory can be illustrated by an example of induction with three reactions (i = 1, 2, 3) in the absence of deceleration (which we introduce later). Suppose the dose-response is given by [Y3] as a function of [Y0]. The goal is to calculate this function and determine conditions for when it is non-cooperative with unit Hill coefficient. In steady state, the concentrations obey the equilibrium conditions
[Yi]=qi[Xi][Yi−1],i=1,2,3
(2)
and the mass conservation conditions
[X1]+[Y1]+[Y2]+[Y3]=X1T[X2]+[Y2]+[Y3]=X2T[X3]+[Y3]=X3T
where XiT is the total concentration of accelerator i. Together they form a system of 6 equations and 7 unknowns. Therefore, any one concentration can be solved in terms of any other. In general, the dose-response for this system will not have unit Hill coefficient [20]. However, a non-cooperative dose response can arise if [Y2],[Y3]<<[Y1] and [Y3]<<[X3] so that the mass conservation equations become
[X1]+[Y1]=X1T(3a)[X2]+[Y2]+[Y3]=X2T(3b)[X3]=X3T(3c)
This form for mass conservation can be achieved if the concentration of X2 is limited with respect to its binding affinity, i.e., q2[X2]<<1, while the other factors are not limited. This can be achieved biochemically if the products have short lifetimes, which has been observed experimentally [33–35]. We call Step (3b) the concentration-limited step (CLS) [20]. All factors following the CLS are in excess (i.e., bound concentrations are negligible), implying that reactions after the CLS are pseudo-first order. Hence, the CLS is a step where the accelerator concentration is limited with respect to its binding affinity but the accelerator(s) following it are not limited. In the Methods, we give a detailed account of the CLS for an arbitrary number of reactions.
Substituting (2) into (3a)–(3b) gives
[X1]+q1[X1][Y0]=X1T[X2]+q[X2][Y1]+q3X3Tq2[X2][Y1]=X2T[X3]=X3T
Each equation is bilinear in the accelerator and product concentrations. When the accelerator concentrations are substituted back into the equilibrium Equations (2), the results are linear-fractional functions between adjacent products:
[Y1]=q1X1T[Y0]1+q1[Y0],[Y2]=q2X2T[Y1]1+q2(1+q3X3T)[Y1],[Y3]=q3X3T[Y2]
Linear-fractional functions form a group under function composition ensuring that the function of any product in terms of any other product is always linear-fractional [20]. Successive substitution of these functions yields the dose-response:
A≡[Y3]=q1X1Tq2X2Tq3X3T[Y0]1+q1[Y0]+q2(1+q3X3T)q1X1T[Y0]
where
Amax=q1X1Tq2X2Tq3X3Tq1+q2(1+q3X3T)q1X1T,EC50=1q1+q2(1+q3X3T)q1X1T
The dose-response can be derived for an arbitrary number of reactions as long as the mass conservation equations are bilinear as in (3) [20,22]. Although, our reactions are reversible and obey detailed balance, the theory can also incorporate dissipative irreversible steps (see Methods) [36]. Since [Y3] is proportional to [Y2] and this is true for all concentrations following the CLS, a more general form for the activity is the sum of all of these concentrations. Biophysically, this implies that the final product can arise from each step after the CLS independently.
As shown in Fig. 1, decelerators can interact with accelerators. Consider the competitive decelerator D interacting with X1 via X1+D↔X′1. If D is in excess, the addition of D leads to one additional equilibrium condition, [X′1]=q′[D][X1], and a modification to the mass conservation law for X1 (3a) to [X1]+[Y1]+[X′]=X1T. Solving the new equilibrium and mass conservation conditions then gives
[Y1]=q1X1T[Y0]1+q1[Y0]+q′[D]
The equations for [Y2] and [Y3] are unchanged. Since [Y1] remains a linear-fractional function of [Y0] in the presence of inhibition by D, the dose-response will also remain a linear-fractional function.
Activity can also be repressed by a reaction following the third reaction, Y3+X4↔Y4, which can suppress [Y3] by diverting the product to a less productive pathway. The new reaction changes mass conservation for X2 (3b) to [X2]+[Y2]+[Y3]+[Y4]=X2T which gives
[Y2]=q2X2T[Y1]1+q2(1+q3X3T(1+q4X4T))[Y1]
This is again linear-fractional, yielding
[Y2]=q1X1Tq2X2T[Y0]1+q′[D]+q1[Y0]+q2(1+q3X3T(1+q4X4T))q1X1T[Y0]
which is inhibited by X4T. Since [Y3]=q3X3T[Y2] and [Y4]=q4X4T[Y3], the most general expression of the activity that preserves non-cooperativity is the sum: A=∑i=24ai−1[Yi]. The activity as a function of the agonist and any other factor is linear-fractional with the form
A=V[Y0]1+W[Y0]
(4)
where
V=q1X1Tq2X2T(a1+a2q3X3T+a3q3X3Tq4X4T)1+q′[D],W=q1+q2(1+q3X3T(1+q4X4T))q1X1T1+q′[D]
(5)
V = Amax/EC50 and W = 1/EC50 are linear-fractional functions of the equilibrium constants and the concentrations of all factors (total concentration of accelerators and free concentration of decelerators) in the system. From (4) and (5), we see that activity can be repressed by the action of either a decelerator or an accelerator, provided the latter acts after the CLS and the contribution to the activity from [Y4] is less than that from [Y2] and [Y3]. We can write formulas for V and W for an arbitrary number of cofactors. Their functional forms are distinguished by the types of the cofactors (i.e., accelerators or one of 6 types of decelerator) and where they act in relation to each other and the CLS. The formulas for V and W for all the possible combinations of 2 factors are calculated in Dougherty et al. [22] and shown in S1 Table. They are always linear fractional functions of each accelerator or decelerator. They can be used to make predictions of the mechanisms of added factors on the basis of the graphs of V and W vs. the cofactor just as the Lineweaver-Burk plot is used in enzyme kinetics.
Finally, it is commonly accepted that transcription factors sometimes form oligomers before they act [17]. Our theory can be generalized to include factors acting through oligomers, provided that they do so non-cooperatively. For example, suppose that an accelerator X first forms an oligomeric complex with another factor Z in the reaction X+Z↔rXZ prior to interacting with a product Yi-1 in a reaction Yi−1+XZ↔qYi before the CLS. If Z is super-abundant compared to X then the concentration of the product as a function of the previous product has the non-cooperative form
[Yi]=rqXT[Z]1+r[Z][Yi−1]1+rq[Z]1+r[Z][Yi−1]
Hence, for accelerators acting through a hetero-oligomer, the limited factor, X, acts as an accelerator in our current theory while the action of the superabundant factor, Z, saturates with sufficiently high [Z]. A decelerator could likewise act through a hetero-oligomer.
We apply the theory to steroid-regulated gene repression by observing that gene activity in induction can be repressed by other factors and still maintain a linear-fractional form. Hence, a linear-fractional dose-response (1) can arise in repression if the steroid-receptor complex (GR) acts as either a decelerator at any position or an accelerator after the CLS of a gene initiated by some other inducer. We compute the dose-response for steroid-regulated gene repression (1) by substituting the steroid-receptor complex (GR) (or some activated form of GR) into the formulas for V and W in the dose-response (4) for a gene activated by another inducer. Note that the experimentally measured dose-response of gene activity with respect to the inducer need not be linear-fractional for this theory to hold. What is required is that downstream steps where the cofactors and GR act have the linear-fractional property.
It is well known that steroid binding to GR follows Michaelis-Menten kinetics in terms of [S] [20]. Suppose that GR acts as D in the above example. We substitute
[D]=[GR]=G[S]K+[S]
into Equation (5) to obtain
V=q1X1Tq2X2T(a1+a2q3X3T+a3q3X3Tq4X4T)(K+[S])[Y0]K+[S]+q′G[S],W=(q1+q2(1+q3X3T(1+q4X4T))q1X1T)(K+[S])[Y0]K+[S]+q′G[S]
Substituting this into (4) and clearing the fractions results in an activity that is linear-fractional in [S], which we can write as
A=T([S])U([S])=T(0)+T′∗[S]U(0)+U′∗[S]
(4’)
where
T=q1X1Tq2X2T(a1+a2q3X3T+a3q3X3Tq4X4T)(K+[S])[Y0]U=K+[S]+q′G[ S ]+(q1+q2(1+q3X3T(1+q4X4T))q1X1T)(K+[S])[Y0]
and the prime signifies derivative with respect to [S]. From this, we can thus surmise that
Amax=T(0)U(0),Amin=T′U′,IC50=U(0)U′
These three dose-response parameters are determined by the four quantities T(0), U(0), T′, and U′ which implies that the combination parameter:
AmaxIC50Amin=T(0)T′
is also linear-fractional. The theory predicts that these four dose-response parameters are always linear fractional and that there are four compatibility conditions between them: a) the numerator of Amax is equal to the numerator of Amax×IC50/Amin, b) the numerator of Amin is equal to the denominator of Amax×IC50/Amin, c) the denominator of Amax is equal to the numerator of IC50, and d) the denominator of Amin is equal to the denominator of IC50. These properties are not expected for arbitrary linear-fractional functions and provide a validity check for the theory.
Suppose we are only interested in the influence of an accelerator after the CLS (i.e., X3T or X4T). We can then write T and U above in terms of the accelerator XT, [S], and effective constants that depend on the parameters of the hidden reactions:
T=(B1+B2XT)(K+[S])U=K+[S]+q'G[S]+(B3+B4qXT)(K+[S])
From which we immediately obtain
Amax=B1+B2XT1+B3+B4XT
(6a)
Amin=B1+B2XT1+q′G+B3+B4XT
(6b)
IC50=(1+B+3B4XT)K1+q'G+B3+B4XT
(6c)
AmaxIC50Amin=K
(6d)
A cofactor can act like an accelerator or a decelerator before, at, or after the CLS. For two cofactors, such as GR and one other cofactor, there are 5 possible configurations for their action when not acting together at the same step (e.g., both before the CLS, one before and one at the CLS, etc.). There are 10 total positional combinations since GR can act before or after the other cofactor. There are 3 more configurations where GR acts at the same position as the other cofactor if one is a decelerator while the other is an accelerator. Thus, GR can act as a decelerator in all of these 13 configurations or as an accelerator after the CLS in 5 configurations. This gives a total of 18 possible configurations of GR and one other cofactor. A calculation for T and U can be made for each of these combinations and the results are in S2 Table. What these calculations show is how the dose-response parameters change as a function of differing amounts of added cofactor.
The experimental dose-response can be fit to the predictions for each of the 18 cases to see which fits best. However, many of the cases can be eliminated immediately based on qualitative properties of the curves. The dose-response parameters will always be linear-fractional functions with the form
y=a+bxc+dx
Depending on the parameters, y can appear like a constant, a linear function, a Michaelis-Menten function, or a general linear-fractional function that increases or decreases with x. We also know that y increases with x if ad<bc, decreases if bc<ad, and is a constant if ad = bc. The x value for half-maximal y (half-maximal concentration) is c/d, and a/b for 1/y. Using these properties, we see that Amin and Amax in Equations (6a) and (6b) for the three reaction example can either increase or decrease depending on the parameter values but if Amax increases then so must Amin. IC50 in (6c) is an increasing function because the denominator has an extra positive constant term. The graph of Amax×IC50/Amin in (6d) is a horizontal line. Similar predictions for all the possible combinations of GR and one cofactor are summarized in Table 1. The graph properties in Table 1 represent some sufficient conditions for the predicted mechanisms and position of action and do not represent a comprehensive list of all possible predictions.
Our experimental paradigm for steroid-regulated repression is the well-documented GR inhibition of phorbol myristate acetate (PMA) induction of a reporter construct (AP1LUC) with the human collagenase-3 promoter [15,28,29]. We usually measured the gene activity for four concentrations of the GR agonist Dex including EtOH in the presence of different concentrations of AP1LUC reporter plus one of three added cofactors: the plasmid for TIF2 or two small molecules (NU6027 and phenanthroline) recently identified in a high throughput screen as accelerators of GR transactivation [30].
We present three lines of evidence to support the application and validity of the theory. 1) Fig. 2 shows the dose-response curve for GR repression of PMA induction of AP1LUC without (Fig. 2A) or with added TIF2 (Fig. 2B). The curves (including EtOH) are well fit by the linear-fractional function in Equation (1) (solid lines) as required by the theory. We found that excellent fits of the dose-response data to Equation (1) could be obtained with four points (3 concentrations of Dex plus EtOH) (R2 = 0.984 ± 0.026 [S.D., n = 160 randomly selected plots], median = 0.993), from which we estimated the dose-response parameters Amax, Amin, and IC50 in the ensuing experiments. 2) Figs. 3–5 show plots of Amax, Amin, IC50, and Amax×IC50/Amin determined from four steroid concentrations for varying amounts of AP1LUC reporter and each of the added cofactors. The parameters are all well fit by linear-fractional functions (solid curves) as predicted by the theory. 3) We tested if these parameter graphs satisfied the four predicted compatibility conditions using Bayesian model comparison as detailed in the Methods. We found that the Bayesian Information Criterion (BIC) is lowest for the predicted model compared to two null models (see Tables S3, S4, and S5), which further validates the theory. Given the confidence that the theory is applicable, we used it to make predictions for the mechanism and position of action for the added cofactors as well as for the reporter and GR.
Since the dose-response for repression is derived from the dose-response for induction, how a factor affects the Amax produced by the inducer in GR-regulated repression should be the same as how it alters Amax in GR-regulated gene induction even if the inducer is different in the two cases. That this is so can be seen from the formulas for Amax in induction (S1 Table) and repression (S2 Table). In our system, the Amax in gene repression is the response to PMA alone and occurs in the absence of added steroid. All the cofactors we considered (other than the reporter) were found to be accelerators after the CLS in steroid-mediated gene induction [22,30]. The above example and S2 Table show that a graph of Amax vs. cofactor can inform us of where an accelerator acts in gene repression and these predictions are summarized in Table 1.
We used four concentrations of both AP1LUC and TIF2 in our competition assay to analyze TIF2 action in GR-regulated gene repression in U2OS.rGR cells. The graphs of Amax, Amin, IC50, and Amax×IC50/Amin vs. TIF2 (Figs. 3A-D) all have linear-fractional shapes (solid lines) as predicted. The data for Amax vs. TIF2 (Fig. 3A) are well fit by Michaelis-Menten functions that have an x-axis intersection coordinate of -46.2 ± 26.8 ng (S.E.M., n = 4) of TIF2 plasmid. As is true in the competition assays for gene induction [22], the interpretation of the graphs for gene repression requires that the x-axis values reflect the total amount of factor, i.e., the sum of endogenous plus exogenous factor (also see Methods). From quantitative Western blots (not shown), it was determined (assuming 50% transfection efficiency of cells [22]) that the endogenous TIF2 is equivalent to 2.7 ± 1.5 ng (S.E.M., n = 3) of plasmid. Thus the point of zero endogenous TIF2 is at -2.7 ng TIF2 plasmid, which is much more positive than the intersection point of the curves at -46, despite the large error range. As seen in Equation (6a), S2 Table (since T(0)>0 for XT=0 for all entries where k > CLS), and summarized in Table 1 (entry 5), this is consistent with TIF2 acting as an accelerator after the CLS.
Figs. 4 and 5 show that the dose-response parameters are also well fit by linear-fractional functions for the small molecules NU6027 and phenanthroline [30]. Amax and Amin versus both cofactors are lines that intersect at values more negative than the concentration of endogenous chemicals, which is zero. This behavior is consistent with both compounds being accelerators after the CLS (Figs. 4A&B and 5A&B and entries 5 and 10 of Table 1). The conclusions for TIF2, NU6027 and phenanthroline being accelerators acting after the CLS in gene repression are consistent with what was observed in gene induction [30].
The traces in the graphs of Amax vs. AP1LUC (Fig. 3E) for varying concentrations of TIF2 are all linear, intersecting at the origin. The linear plot of Fig. 3E is preferred over a nonlinear plot (BIC = 46.39 vs. 55.56 respectively). As can be seen from the example above, in the formulas of S2 Table where k = CLS and summarized in Table 1 (entry 2), this is consistent with AP1LUC acting as an accelerator at the CLS. Graphs of IC50 vs. AP1LUC (Fig. 3F) consist of near horizontal lines (e.g., a constant slope) that decrease in position with added TIF2 (average slope = -0.0016 ± 0.0038, S.D., n = 4 traces of graph). These plots, summarized in Table 1 (entry 17), are also diagnostic of the reporter AP1LUC being an accelerator (A) at the CLS.
S1 A&B and S2 A&B Figs. show that Amax and Amin versus AP1LUC for varying concentrations of NU6027 and phenanthroline are again linear through the origin. The graphs of IC50 vs. AP1LUC have a constant zero slope with NU6027 (= -0.0009 ± 0.0064, S.D., n = 4 traces) and with phenanthroline (= 0.0067 ± 0.0070, S.D., n = 4 traces) (S1 C and S2 C Figs). These imply that AP1LUC is an accelerator at the CLS in the presence of both NU6027 and phenanthroline. Hence, we find that the reporter is always an accelerator at the CLS in both gene induction and repression.
Generally, in order to determine the action of a factor, one measures the response to changes of that factor. However, this was not possible with GR because we could only observe the robust repression needed for accurate graphical analyses with the high amounts of stably transfected GR in our experimental system. However, we can still deduce the mechanism and location of GR by comparing the responses to changes in the cofactors to the formulas in S2 Table to see which behaviors for GR are compatible with the observed results.
Fig. 3 indicates that Amax increases while Amin, IC50 and Amax×IC50/Amin all decrease vs. TIF2. As we show in the Methods, this is mathematically possible only if GR acts as an accelerator after TIF2 and TIF2 is an accelerator after the CLS. The results are also summarized in Table 1 (entries 7, 19 and 22). Furthermore, from our posterior parameter estimates of our Bayesian model comparison test (see S3 Table), we find that the concentration of TIF2 at half-maximal Amax (parameter 1) is greater than the same for Amin (parameter 3). This condition is also true for 1/Amax (parameter 2) and 1/Amin (parameter 4). These conditions further support the above deductions that TIF2 is an accelerator after the CLS and GR acts as an accelerator after both the CLS and TIF2 (Table 1, entries 12 and 14).
Figs. 4 and 5 show that Amax and Amin are each augmented by increasing concentrations of both NU6027 and phenanthroline. Hence, these cofactors cannot uniquely predict the action of GR. However, they can still be used to test for consistency. Figs. 4C&D show that Amax×IC50/Amin and IC50 are both decreasing versus NU6027. According to entries 22 and 19 respectively of Table 1, these graphs support the inference that GR acts as an accelerator after NU6027, which acts as an accelerator after the CLS. The concentrations of NU6027 at half-maximal Amax and 1/Amax are larger than those of Amin and 1/Amin respectively (S4 Table), which is also consistent with the conclusion that NU6027 acts as an accelerator after the CLS and before GR.
Unlike TIF2 and NU6027, Amax×IC50/Amin and IC50 versus phenanthroline do not exhibit any obvious trends (Figs. 5C&D). However, an examination of the formulas for Amax×IC50/Amin and IC50 in S2 Table shows that there are parameter regimes where Amax×IC50/Amin and IC50 vary so slowly that they would appear constant when the factor acts after the CLS. There was also no significant difference between the half-maximal concentrations of Amax and Amin and their reciprocals for phenanthroline (S5 Table, mean posteriors). Hence, these data neither confirm nor contradict the above conclusion that GR acts as an accelerator after phenanthroline.
Therefore, in this system, the reporter (AP1LUC) and the added cofactors display the same kinetically-defined mechanisms of action, and at the same positions relative to the CLS, in GR-regulated gene repression and gene induction (Fig. 6). The only difference is that the position, but not mechanism, of GR action changes in gene repression from that in gene induction.
We introduce a theory for GR-regulated gene repression that is based on first principles. The theory is general but is mathematically solvable only when the dose-response curves for gene repression are linear-fractional (Fig. 2). The theory accommodates any number of pathway steps, transcription factors, and cofactors that alter the Amax, Amin, and/or IC50 of GR-controlled gene repression. The formation of multicomponent complexes is permitted as long as their concentrations are low or their lifetimes are short, which is biologically reasonable and has been observed for numerous factors [33–35]. The theory could also be generalized to include the action of factors through preformed hetero-oligomeric complexes with other factors. The competition assay for gene repression, like that for GR transactivation [20,22,23,27], informs the kinetically-defined mechanism of action of each factor (i.e., accelerator vs. one of six decelerators) and the position of factor action relative both to the other competing factor and to the CLS, which again appears to be an invariant marker in the overall reaction sequence (see below). The theory makes specific predictions regarding the graphs of Amax, Amin, IC50, and Amax×IC50/ Amin vs. one factor with increasing concentrations of a second factor (Table 1). The utility of the theory has been tested by examining the effects in the competition assay of four factors: AP1LUC reporter, p160 cofactor TIF2, and two recently identified pharmaceutical modulators of GR transactivation [30].
The theory assumes mass action kinetics in a well-mixed medium. It cannot account for stochastic effects of a single transcription event. However, the averaged amount of mRNA in imaging experiments of a single agonist-induced gene does follow Michaelis-Menten kinetics [31] and may be analyzable by stochastic models [32]. Understanding the connection between our theory and stochastic models is an important future step.
In every comparison, the AP1LUC reporter is found to act as an accelerator at the CLS. The CLS corresponds to that step where the concentration of the accelerator is limited compared to its binding affinity but the free concentration of accelerators in reactions after the CLS are in excess compared to their bound concentrations. The CLS can also be thought of as the step in the reaction sequence where an initially limited accelerator is replaced by another limited accelerator, such as GR being replaced by reporter gene. Thus, the species transmitting the input signal undergoes a “baton pass”, as in a relay race, to a new species at the CLS. In induction, the reporter gene also always, and uniquely, acts as an accelerator at the CLS [23,25–27]. Thus, reporter action at the CLS is an invariant signpost in both gene induction (with GREtkLUC) and gene repression (with AP1LUC) about which all other modulating factors and cofactors can be arranged. This leads us to predict that this will also be the case in future examples of gene repression.
Experiments with varying concentrations of GR cannot be performed in the present system because only the U2OS.rGR cells with high amounts of endogenous GR gave robust repression. Cells with very low levels of GR, which would permit adding increasing concentrations of GR, did not yield the large fold-repression needed for the high precision measurements of the competition assay (data not shown). Nevertheless, careful examination of the equations for the theory of repression revealed that the position and mechanism of GR action can be deduced from various graphical characteristics with respect to another cofactor. In this manner it can be determined that GR acts as an accelerator after both the CLS and each of the three factors (TIF2, NU6027, and phenanthroline). Therefore, the kinetically-defined mechanism of action of GR, like that of the other factors including the reporter, is again the same as that seen during gene induction. What has changed for GR is its position of action. In gene induction, GR acts before the CLS and every other factor so far examined [23,25–27]. In contrast, GR functions after the CLS and each of the factors studied here in gene repression. This is not unexpected. Something about GR action must change if the increased response by GR in gene induction is to become a decreased output in gene repression. Furthermore, as the induction of the AP1LUC reporter occurs with added PMA in the absence of GR, it is plausible that GR inhibition might occur at a step downstream of the induction reaction. In this respect, we note that GR represses TNFα induction of IL-8 gene expression by acting after transcription initiation [37]. An accelerator acting after the CLS can increase or decrease the total gene activity depending on the parameters. We propose that repression of AP1LUC results from GR acting as an accelerator to favor a less productive step after PMA-mediated induction, thereby decreasing the level of LUC expression (Fig. 6). It is important to realize that GR can be found to display identical mechanisms of action in induction and repression while interacting with different factors.
The mathematical models for both gene repression and gene induction by GR require the sequential binding of GR monomers as opposed to the binding of preformed dimers. This is supported by the observation of non-cooperative dose-response curves with unit Hill coefficient in both GR-mediated repression (Fig. 2) and induction [20]. The binding of preformed GR dimers would yield a dose-response curve with greater than unity Hill coefficient. The sequential binding of monomers in gene induction to form DNA-bound dimers is further supported by the activity of several dimerization-defective GR mutants [20] and by biophysical studies of other dimerizing proteins that are found to associate with DNA by the sequential binding of monomers to yield DNA-bound dimers [38–43]. Similar experiments with dimerization-defective GR mutants cannot be performed in the current system due to the weak response in cells with very low levels of endogenous GR. However, it currently appears that most, if not all, instances of GR-mediated repression may also involve dimers of GR that are bound or tethered to DNA [44,45]. We strongly suspect that, as in gene induction, the association of GR with the DNA of repressed genes also proceeds via the step-wise binding of GR monomers, as required by our mathematical model.
How to define the actions of cofactors for gene repression has been complicated by the fact that GR causes changes in total activity that are opposite those seen in gene induction [4–11,15]. Therefore, any classification of cofactor action relying on final activity is doomed to ambiguities. The present competition assay defines cofactor activity in an unbiased manner because it considers only what is happening at the position of cofactor action, independent of the direction of changes in final response. Such stepwise considerations can now be achieved independent of the biochemical processes involved and are essential if one desires to selectively modify GR-regulated repression (and induction) of specific genes.
Using a common set of definitions of cofactor action [23], we find that TIF2, NU6027, and phenanthroline are all accelerators after the CLS in both gene induction [21–25,30] and gene repression. Despite apparently identical mechanisms of TIF2, NU6027, and phenanthroline in GR-regulated gene repression, there are some graphical differences. This is due to a combination of where they act relative to each other and the specific reaction parameters, which cannot be defined exactly with our current data. Thus identical mechanisms of action do not require precisely identical graphs. Nonetheless, this permits great mechanistic simplifications. It suggests that at least those cofactors that are accelerators act in a constant and modular fashion, independent of the changes in the final product and the inducing agonist. Such common kinetically-defined mechanisms of cofactor action mean that new cofactor actions need not be invoked to account for the different responses in gene induction and repression because the underlying mechanisms are, in fact, the same. It also suggests that manipulation of cofactor action in induction vs. repression can simultaneously affect both pathways.
There have been many descriptions of binding partners for TIF2 under conditions where TIF2 reverses a transcriptional response [46–50]. However, it should be realized that those species to which a factor initially binds, or chemically modifies, are unlikely to constitute the position at which factor action is exerted, which is what is revealed by our competition assay. For example, paused RNA polymerase II action occurs at steps downstream from its initial binding [51,52] while protein modifications, such as histone acetylation, elicit their effects after the modifying protein has bound. Similarly, the involvement of different cofactor domains in, for example, repression vs. induction, does not require different mechanisms of action. Different domains may alter the strength of a specific cofactor-target interaction or cause similar modifications of different downstream targets [37,53,54]. Alternatively, the relative importance of two positions of cofactor action may be influenced by domain composition [25,26].
In summary, we have described a theoretical model for GR-mediated gene repression that defines factors by their action at a particular step of the overall reaction scheme, as opposed to the final outcome. The theory has been validated by experimental results with five factors. The mechanism and position of action of four factors is qualitatively identical to that previously found in gene induction. The difference is that the fifth factor, GR, is predicted to act before the CLS and various cofactors in gene induction and after the CLS and cofactors in gene repression, presumably by diverting the reaction scheme to, and accelerating, a less productive pathway. The apparent constancy of factor mechanism of action in gene induction and repression means that altering cofactor activities is predicted to simultaneously affect both GR-regulated pathways. Thus, new and uncharacterized pathways will not have to be considered, which will simplify approaches for maximizing desired outcomes. Finally, as for the theory for gene induction [20,22], the current theory for gene repression is general for any gene induction and gene repression process displaying a linear-fractional dose-response and thus could be of use to analyze the mechanisms of other inducible transcription factors.
Unless otherwise indicated, cell growth was at 37°C and all other operations were performed at 0°C.
Dexamethasone (Dex), PMA (phorbol myristate acetate), NU6027, and phenanthroline were purchased from Sigma (St. Louis, MO). Restriction enzymes and T4 DNA ligase were from New England Biolabs (Beverly, MA) and the dual-luciferase reporter assay was from Promega (Madison, WI).
TIF2 (Hinrich Gronemeyer, IGBMC, Strasbourg, France), and AP1Luc (Inez Rogatsky, Weill Medical College, Cornell University) were generously donated.
U2OS.rGR cells were grown in DMEM media supplemented with 10% FBS and 0.1 mg/ml G418 and seeded at 30,000 cells per well in a volume of 300 μl per well in 24-well plates as previously described [29] with the following modifications. One day after seeding in 24-well plates, cells in FBS-free DMEM were transfected with reporter (AP1Luc), 10 ng of phRG-TK Renilla (Promega) as an internal control, and the indicated amounts of plasmids for various factors in OPTIMEM plus XTREME Gene HP (Roche; 0.8 μl/well). Four hours after transfection, cells were refed with DMEM/10% FBS. The next day, cells were treated with PMA (10–25 ng/ml) and various dilutions of dexamethasone (Dex) and chemical. Sixteen hours later, the cells were lysed in lysis buffer and assayed for reporter gene activity using dual luciferase assay reagents according to the manufacturer’s instructions (Promega, Madison, WI). Luciferase activity was measured by a GloMax® 96 Microplate Luminometer (Promega, Madison, WI). The data were normalized to Renilla TK luciferase activity and expressed as a percentage of the maximal response with Dex before being plotted ± S.D. unless otherwise noted.
The basic protocol for gene induction [22,23] was followed except as noted in the text for 4x4 (all 16 combinations of 4 concentrations of both factor 1 and factor 2) assays with four concentrations of Dex, all in triplicate, for a total of 196 wells. All plots of the data assume a linear increase in factor plotted on the x-axis. When Western blots reveal a nonlinear relationship between the optical density of scanned protein band and the amount of transfected plasmid at constant levels either of total cellular protein, or of β-actin, the linear equivalent of expressed plasmid must be determined as previously described [22] (see also below).
The nonlinear plot of OD vs. ng of transfected plasmid is first fit to the Michaelis-Menten formula
Amax=m1*plasmid/(m2+plasmid)
to obtain constants m1 and m2. The functional equivalent of the transfected plasmid that gives a linear OD vs. plasmid plot is then obtained from the formula of
Plasmid(linear)=m2*plasmid/(m2+plasmid)
The x-axis value of amount of plasmid in the various graphs is then this “corrected plasmid” value.
A linear-fractional dose-response arises when the mass conservation equations are bilinear. The general bilinear form for an arbitrary number of reactions is:
[X1]+[Y1]=X1T[X2]+[Y2]=X2T⋮ } Pre-CLS
[Xj]+[Yj]+[Yj+1]+⋯+[Yk]=XjT CLS
[Xj+1]=Xj+1T⋮[Xk]=XkT } Post-CLS
[Xk+1]+[Yk+1]=Xk+1T⋮ } Pre-CLS
[Xl]+[Yl]+[Yl+1]+⋯=XlT CLS
[Xl+1]=Xl+1T⋮ } Post-CLS
where there can be multiple CLS steps. The bilinearity of these equations is easily confirmed by recursively substituting the equilibrium conditions for the products. The reactions can be divided into pre-CLS, CLS, and post-CLS steps. In pre-CLS steps, only the accelerator and its immediate product appear in the mass conservation equation. This can be achieved if all downstream products have much lower concentration than the accelerator and its product, i.e., [Yj]<<[Yi], [Xi], j>i, which is satisfied if [Yi+1]<<[Yi], [Xi], i.e., products become successively smaller. Substituting in the equilibrium conditions gives qi+1Xi+1T[Yi]<<[Yi] or qi+1Xi+1T<<1. Hence, accelerators are limited with respect to their affinities at pre-CLS steps and the CLS is the last step of this sequence for which the accelerator is limited. At post-CLS steps, the free concentration of accelerator is equal to the total concentration, i.e., the bound concentration is negligible. This is satisfied if [Yj]<<[Xi], j≥i, which implies that the concentration of accelerator is in excess compared to the products and the reactions are pseudo-first order.
The crucial point for preserving linear-fractional dose-response is that the equilibrium conditions have the form of (2) and the mass conservation conditions have bilinear form. Any kinetic scheme, reversible or irreversible, that obeys a similar set of equations at steady state will have a linear-fractional dose-response. For example, consider the irreversible reaction P+X →P′ + X and P′→ P. In this “hit-and-run” scheme, the accelerator X interacts transiently with P and modifies it. The modified P then relaxes back to the original state as a first order kinetic process. This reaction could occur for example if P is in an “excited” state, X nudges it to a lower energy state P′, and energy is expended to pump P′ back to the excited state. The concentrations have kinetics
d[P′]dt=kf[P][X]−kr[P′],d[P]dt=−kf[P][X]+kr[P'],d[X]dt=0
which has a steady state solution [P′] = q[X][P], where q = kf/kr, and conservation equations [P]+[P′] = PT, [X] = XT. The equations combine to yield [P′] = q XT[P], thereby mimicking a post-CLS reaction.
Conversely, we could have the irreversible reaction P+X →X′ + P and X′→ X, where the product from the previous reaction now transiently interacts with an accelerator and modifies it to create a new product X′. In steady state, the new product obeys [X′] = q[X][P] with mass conservation condition [X]+[X′] = XT, or [X]+q[X][P] = XT. The equations combine to yield
[X′]=qXT[P]1+q[P]
If this reaction were followed by a reaction of the form X′+U→X″+U, X″→ X′ then the mass conservation law for X changes to [X]+[X′]+[X″] = XT. Combining with the additional steady state equation [X″] = q′UT[X′], we obtain
[X′]=qXT[P]1+q[P]+qq′UT[P]
which mimics a CLS step. Hence, in this irreversible hit-and-run scheme, accelerators that successively modify a product are akin to post-CLS steps and a reaction where the product switches roles and modifies a factor to make a new product is akin to a CLS. Reversible and irreversible reactions could be combined as long as the steady state conditions resemble the equilibrium conditions (2) and the mass conservation conditions have bilinear form.
Experiments find that Amax increases while Amin decreases with an increase in TIF2, which acts as an accelerator after the CLS. Experiments also find that IC50 and Amax×IC50/ Amin decrease with increases in TIF2. We show below that this behavior puts severe constraints on the possible mechanisms and positions where GR can act. Specifically, we need to determine for which cases it is possible for Amax > Amin, and Amax increases while Amin, IC50 and Amax×IC50/ Amin all decrease for increases of a post-CLS accelerator. We examine each possible case for TIF2 acting after the CLS individually.
Unless otherwise noted, all experiments were performed in triplicate multiple times. KaleidaGraph 4.1 (Synergy Software, Reading, PA) was used to determine a least-squares best fit (R2 was almost always 0.95) of the experimental data to the theoretical dose-response curve, which is given by the Equation (1). This was done by first estimating Amax directly from the data for zero Dex. The data points were subtracted from Amax and the resulting curve was then fit to a Michaelis-Menten function from which Amin and IC50 were determined. The values of n independent experiments were normalized, averaged, and then plotted and analyzed as described in the Supporting Information. The Bayesian Information Criterion was used to determine whether the best fit of Amax vs. AP1LUC data is obtained with linear or Michaelis-Menten plots.
The theory predicts that the four dose-response parameters are linear-fractional functions of the cofactor and satisfy four compatibility conditions. To test this, we fitted the theoretically predicted model, with two other “null” models to the data and used the Bayesian Information Criterion to see which model was best. The models tested were
1)Predicted Model:
Amax=a5[R]a1+[X]a2+[X],Amin=a6[R]a3+[X]a4+[X],IC50=a7a2+[X]a4+[X],Amax=a8a1+[X]a3+[X]
2)Permuted Model:
Amax=b5[R]b1+[X]b4+[X],Amin=b6[R]b2+[X]b3+[X],IC50=b7b3+[X]b2+[X],Amax=b8b4+[X]b1+[X]
3)Unconstrained linear-fractional model:
Amax=c9[R]c1+[X]c5+[X],Amin=c10[R]c2+[X]c6+[X],IC50=c11c3+[X]c7+[X],Amax=c12c4+[X]c8+[X]
where [X] is the concentration of the added cofactor, the a’s, b’s, c’s are parameters to be fitted to the data, and [R] is the concentration of the reporter, which we know is an accelerator at the CLS. The predicted and permuted models have the same model complexity with 8 free parameters, while the unconstrained model has 12 free parameters. The predicted model is a subset of the unrestricted model. The half maximum of the dose-response parameters are given by the parameter in the denominator. For example, the cofactor concentration for half maximum for the predicted model of Amax is given by a2, and that of Amin is given by a4. Conversely, the concentrations of half maximum of 1/Amax and 1/Amin are given by a1 and a3, respectively.
We used a Metropolis-Hastings Markov Chain Monte Carlo (MCMC) method [55] to compute the Bayesian posterior probabilities of the various parameters. The likelihood function we used was proportional to exp(−χ2/2) where
χ2=−∑p=1k∑i=1n(yi,pdata−yi,pmodel)2σi2
The sum is over all n data points i for each dose-response parameter p out of k total, and the error variance σi,p2 is the experimentally determined standard error variance of replicate experiments. The results for the maximum-likelihood values, and the mean and standard deviation for posteriors for TIF2, NU6027, and phenanthroline are in Tables S3 to S5. The reported results are for 2×107iterations for the models with 8 parameters and 3×107 for the unrestricted model after an even longer transient period to ensure convergence. The longer time for the unrestricted model was to compensate for the extra parameters. We used the Bayesian Information Criterion, BIC = x2 + k ln n, to test which model best fit the data accounting for model complexity.
|
10.1371/journal.ppat.1004492 | The Relationship between Host Lifespan and Pathogen Reservoir Potential: An Analysis in the System Arabidopsis thaliana-Cucumber mosaic virus | Identification of the determinants of pathogen reservoir potential is central to understand disease emergence. It has been proposed that host lifespan is one such determinant: short-lived hosts will invest less in costly defenses against pathogens, so that they will be more susceptible to infection, more competent as sources of infection and/or will sustain larger vector populations, thus being effective reservoirs for the infection of long-lived hosts. This hypothesis is sustained by analyses of different hosts of multihost pathogens, but not of different genotypes of the same host species. Here we examined this hypothesis by comparing two genotypes of the plant Arabidopsis thaliana that differ largely both in life-span and in tolerance to its natural pathogen Cucumber mosaic virus (CMV). Experiments with the aphid vector Myzus persicae showed that both genotypes were similarly competent as sources for virus transmission, but the short-lived genotype was more susceptible to infection and was able to sustain larger vector populations. To explore how differences in defense against CMV and its vector relate to reservoir potential, we developed a model that was run for a set of experimentally-determined parameters, and for a realistic range of host plant and vector population densities. Model simulations showed that the less efficient defenses of the short-lived genotype resulted in higher reservoir potential, which in heterogeneous host populations may be balanced by the longer infectious period of the long-lived genotype. This balance was modulated by the demography of both host and vector populations, and by the genetic composition of the host population. Thus, within-species genetic diversity for lifespan and defenses against pathogens will result in polymorphisms for pathogen reservoir potential, which will condition within-population infection dynamics. These results are relevant for a better understanding of host-pathogen co-evolution, and of the dynamics of pathogen emergence.
| Understanding pathogen emergence is a major goal of pathology, because of the high impact of emerging diseases. Pathogens emerge onto a new host from a reservoir, hence the relevance of identifying the determinants of host's reservoir potential. Host lifespan is considered as one such determinant: short-lived hosts will invest less in defenses, being more susceptible to infection, more competent as infection sources and/or will sustain larger vector populations, and thus, are effective reservoirs for long-lived host infection. Evidence for this hypothesis derives from analyses of different hosts of multihost pathogens, and here we examine whether it holds at the within-species level by comparing two genotypes of the plant Arabidopsis thaliana that differ in life-span and in tolerance to its natural pathogen Cucumber mosaic virus. Experiments showed that defenses to the virus and its aphid vector were less efficient in the short-lived genotype that, according to model simulations, was an effective reservoir under a large range of conditions. Reservoir potential, though, was modulated by the demography of host and vector and by the genetic composition of the host population. Thus, within-species genetic diversity for lifespan and pathogen defense will result in differences in reservoir potential, which will condition infection dynamics and host-pathogen co-evolution.
| Understanding the complex interplay of factors resulting in pathogen emergence has been a major goal of evolutionary ecology for the last twenty years, as emerging infectious diseases often have a high impact in human and animal health, agriculture and conservation [1]–[5]. Generally, emergent pathogens are multi-host pathogens that spill-over onto a new host population from one or more epidemiologically connected populations in which the pathogen can be permanently maintained, that is, from a reservoir host sensu Haydon et al. [2]. Hence, identifying the causes that determine the reservoir potential of a host, i.e., its ability to sustain pathogen populations for transmission to the target host, is central for understanding emergence and, more generally, infection dynamics. Emergent pathogens are often vector-transmitted [1], [3], [4], so that parameters associated with the triple interaction host-pathogen-vector should be taken into consideration for predicting host reservoir potential. For vector-transmitted pathogens, three epidemiological parameters have been underscored as modulating host reservoir potential: i) the probability that a vector acquires the pathogen when feeding on an infected host (host competence), ii) the probability that a host is infected by a feeding vector that carries the pathogen (host susceptibility), and iii) the ability of the host to sustain vector populations [6]–[8].These parameters vary among host species and genotypes [2], [6]–[9], and knowing which factors determine such variation will facilitate identifying the potential of a host as a reservoir for pathogen emergence.
Host lifespan has been identified as a trait related to host reservoir potential. The rationale for linking host lifespan and reservoir potential is that evolution of defenses against pathogens has a fitness cost for the host in terms of other advantageous life history traits, such as fecundity or survival [10]–[13]. Because the probability of becoming infected is higher in long-lived host individuals than in short-lived ones, disease prevalence will be higher in populations of long-lived hosts. Thus, long-lived hosts will be under higher selection pressures for developing costly defenses than short-lived hosts. Model analyses under different scenarios do indeed show that long-lived hosts generally will invest more in defense against pathogens [14]. Major forms of defense are resistance, which has been defined as mechanisms that reduce infection and/or pathogen multiplication in the infected host [15], [16], and tolerance, defined as mechanisms that reduce the negative impact of infection on host fitness, i.e., that reduce pathogen virulence [15], [17]. By the same rationale, longer-lived hosts would be expected to have developed defenses that would reduce their ability to sustain vector populations. However, the relationship between lifespan and defense is a complex one and, because the demographic turnover of short-lived hosts is high and that of long-lived ones is low, it may be influenced by demographic factors, notably by population density and composition [14], [18]. This is in agreement with experimental data showing a link between host density and defense evolution [19]–[21]. Despite the attention received from theoreticians, experimental analyses of the relationship between host lifespan and defense are scarce. For plants, there is evidence that short-lived species of grasses (Poaceae) have a high reservoir potential of a generalist plant virus for long-lived hosts [22], [23]. Also, it has been experimentally shown that defenses reducing susceptibility to infection by that virus was lower in short-lived (annual) that in long-lived (perennial) species of grasses, while competence was not explained by lifespan [8], [24].To our knowledge, the relationship between host lifespan and reservoir potential has not been analyzed at the within-species diversity level.
Here we address the hypothesis that there is a trade-off between defense to a vectored pathogen, and to its vector, and host lifespan, which would explain pathogen reservoir potential. Specifically we ask the following set of questions: i) is host lifespan related to competence, susceptibility and the ability to sustain vector populations? ii) is host lifespan related to reservoir potential?, and iii) does reservoir potential depend on the density and genetic composition of the host population?. For this we used an experimental approach for estimating realistic values of the parameters determining host reservoir potential, as a basis for model analyses of the factors that modulate reservoir potential. We focused our research on a plant-virus system: the wild plant Arabidopsis thaliana L. Heynh. (Brassicaceae), its natural pathogen Cucumber mosaic virus (Bromoviridae), and the virus insect vector Myzus persicae Sulzer (Aphididae), an interactive assembly of biological components found in nature. A. thaliana (from here on, Arabidopsis) has been for a long time the model organism of choice for plant molecular genetics, and is increasingly used in analyses of plant ecology, including the evolutionary ecology of plant-pathogen interactions [e.g. 25–29]. Arabidopsis is an annual species, presently distributed worldwide after experiencing an expansion from its native range in Eurasia and North Africa [30]. The genetic structure of Arabidopsis in the Iberian Peninsula has been studied in detail, demonstrating that it is a centre of genetic diversity for this species [31]–[33]. Demographical analyses carried on in the Iberian Peninsula indicated that Arabidopsis plants flower and complete their life cycle in spring, and that populations are built of two or one cohorts of plants that either germinate in the autumn and overwinter as rosettes, or germinate in the spring [29], [34]. Also, populations are genetically heterogeneous and are composed of short-lived genotypes, which do not require vernalization to complete their life cycle, and of long-lived ones, that usually require vernalization and belong to the autumn cohort [32], [35]. Cucumber mosaic virus (CMV) is a single-stranded, messenger sense RNA virus, with a three-partite genome encapsidated in isometric particles. CMV has an extremely broad host range, infecting over 1200 species in more than 100 plant families. CMV is efficiently transmitted by more than 75 species of aphids (Hemiptera: Aphididae) [36], in a non-persistent, stylet-borne manner, i.e., CMV does not infect the insect vector, instead, particles are retained in the maxillary stylets, allowing the aphid to transmit the virus for a short time (less than 6 h) after acquisition. The green peach aphid (M. persicae), a cosmopolitan aphid species, is an important component of the aphid populations feeding on plants in a variety of habitats in Spain (http://www.cabi.org/isc/datasheet/35642), including those in which wild Arabidopsis populations are present. M. persicae is one of the most efficient vectors for CMV, and is frequently used in transmission experimentation [37]. Analyses of virus infection in six wild Arabidopsis populations from different habitats of Central Spain have shown that CMV was the major viral pathogen, with prevalence reaching over 70%, according to the population and the year [29 and unpublished data]. Importantly, while maximum virus prevalence occurred at early developmental stages of the plants, plants remained infected and infectious for the rest of their life. CMV is also transmitted through the seed, with efficiency varying largely according to the plant species and genotype, and it has been reported in Arabidopsis to vary between 2 to 8% [38]. It has been shown that Arabidopsis genotypes differ in their ability to sustain populations of M. persicae [39]. Also, Arabidopsis genotypes differ in resistance and tolerance to CMV, and previous work in our group has shown that tolerance to CMV infection was correlated positively with lifespan in Arabidopsis [40], [41].
We used an experimental approach comparing two Arabidopsis genotypes that differ largely in lifespan, for the following traits: competence as a source for infection, susceptibility to infection, ability to sustain vector populations and rates of seed transmission. These data were then used as a basis to estimate realistic values of the parameters for the analysis of host reservoir potential using a model in which the dynamics of infected plants and of viruliferous vectors were jointly considered. Our results show that the short-lived and the long-lived genotypes differed in susceptibility and in their ability to sustain vector populations, which resulted in a higher reservoir potential for the short-lived host. However, reservoir efficiency of the short-lived genotype was modulated by the density of host plant and vector populations, and by the genetic composition of the host population.
The CMV strain LS-CMV used in this work was derived from biologically active cDNA clones [42] (Accession numbers AF416899, AF416900 and AF127976 for genomic RNAs 1, 2 and 3, respectively). In vitro transcripts were multiplied in Nicotiana clevelandii plants, virions were purified as in Lot et al. [43] and viral RNA was extracted by virion disruption with phenol and sodium dodecyl sulphate. Two genotypes of Arabidopsis, Landsberg erecta (Ler) and Llagostera (Ll-0), were chosen for the very different span of their life-cycle and tolerance to LS-CMV infection [40], [41]. The genotype Columbia glabrata 1 (Colgb1) was used as an additional virus source for aphid transmission of CMV to Ler and Ll-0. These three Arabidopsis genotypes were initially multiplied simultaneously under the same greenhouse conditions to minimize maternal effects. For experiments, Arabidopsis seeds were surface-sterilized, plated on one-half-strength Murashige and Skoog basal salt mix medium [44], 1% (w/v) sucrose, 0.8% (w/v) Bacto agar, and stratified in the dark at 4°C for 96 h. Plates were then transferred to a growth chamber at 22°C under long day conditions (16 h light/8 h dark) and 65–70% relative humidity. Five days-old seedlings were transplanted into 96 well trays with a mix 3∶1, peat∶vermiculite, and after 10 days were transferred into individual 10 cm diameter pots containing the same substrate, in order to minimize spatial and resource limitation. Plants were grown in a growth chamber at 22°C under normal light (220–250 µmol.S−1.m−2) and long day conditions with 65–70% relative humidity. Generally 2–3 days post-transfer, when plants had 4–5 leaves (stages 1.04–1.05 as in Boyes et al. [45]), three rosette leaves per plant were mechanically inoculated with a total of 15 µl of 100 µg/ml suspension of LS-CMV RNA in 0.1M Na2HPO4. Only 15 µl of 0.1M Na2HPO4 were applied to mock-inoculated controls. Ten days post-inoculation (dpi), three circles of 4 mm diameter were cut from three randomly chosen systemically infected leaves. In this sample, systemic infection was confirmed by ELISA with PathoScreen CMV Kit (Agdia, Elkhart, IN, USA), and virus accumulation was quantified by quantitative PCR (qRT-PCR). For this, total RNA extracts were obtained using TRIzol reagent according to manufacturer's protocol (Life Technologies, Carlsbad, CA, USA), and then utilized with Brilliant III Ultra-Fast SYBR Green QRT-PCR Master Mix following manufacturer's recommendations (Agilent Technologies, Santa Clara, CA, USA) in a final volume of 10 µl. Assays were performed in triplicate on a LightCycler 480 II real-time PCR system (Roche, Indianapolis, IN, USA). Primers CMV-CP LS Q Fwd (TAAGAAGCTTGTTTCGCGCATTC) and CMV-CP LS Q Rev (CGGAAAGATCGGATGATGAAGG) were designed to amplify 106 nt of the LS-CMV coat protein (CP) gene (GenBank: AF_127976) and primers UBI/PE4Fwd (AATGCTTGGAGTCCTGCTTG) and UBI/PE4 Rev (CTTAGAAGATTCCCTGAGTCGC) amplified 107 nt of the peroxin4 mRNA (GenBank: NM_122477) used as the loading internal control. Quantification was expressed as pg of viral RNA per ng of total RNA. No-template reactions were included in each trial. Thermal parameter for RT-PCR amplification were 50°C for 10 min, 95°C for 3 min and 40 cycles of 95°C for 5 s and 60°C for 10 s. Dissociation curves were generated to ascertain that only one single product was produced and detected in each case.
Time to bolting and to flowering was estimated as the number of days from the end of stratification until the appearance of the reproductive meristem and the first open flower, respectively. Rosette longevity was estimated as the number of days from the end of stratification until 50% of rosette leaf-senescence. Total plant lifespan was estimated as the time in days from the end of stratification until 50% of siliques had shattered. To quantify leaf mass per unit area (LMA, g×m−2), leaves were harvested at flowering and digitally captured for area measurement using the software ImageJ 1.47v (NIH, USA, http://rsbweb.nih.gov/ij/index.html). Then, leaf dry weight was determined after maintaining the leaves at 60°C until constant weight was reached. Rosette relative growth rate (cm/(day×cm)) was estimated by digital capture and further measure of rosette diameter increase every 2 days for a period of 12 days post inoculation. LMA and relative growth rate are parameters that relate to quick-return and slow-return physiological phenotypes, which may influence host competence, susceptibility, and the ability to support vectors [8], [24].
To assay germination potential, the same day as siliques were ripe, seeds were collected and plated onto germination media without surface-sterilization. Seeds were then either directly transferred to a light/dark regime as previously mentioned or they were submitted to stratification prior to the light/dark protocol. The percentage of germinated seeds was determined 4 days post-transfer to the light/dark cycle. To analyze the effect of virus infection on seed germination, a period of 10 months dormancy (storage in the dark at room temperature) was applied to the seeds before carrying out the regular aforementioned surface-sterilization/stratification protocol. Seed transmission of LS-CMV was estimated by qRT-PCR, testing fifteen biological replicates, consisting of a mix of six 5-day-old seedlings. Virus transmission rate for a single seed was then estimated using the expression reported by Gibbs and Gower [46], , where p is the probability of virus transmission by a single seed, y is the number of positive samples, n is the total number of samples assayed (n = 15), and k is the number of seedlings per sample (k = 6).
Experiments were performed using a clonal population of the CMV aphid vector M. persicae derived from a single virginiparous apterous female collected in a pepper crop at Alcalá de Henares (Madrid, Spain) in 1989. In order to obtain apterous or alate adult aphids adapted to Arabidopsis, two aphid colonies were maintained at ICA-CSIC (Madrid, Spain) (Lat. 40°43′97″N, Long. 3°68′69″W, Alt. 710 m) on an equal mixture of Ler and Ll-0 plants in rearing cages in environmental growth chambers under different conditions: i) For producing apterous aphids, the colony was maintained at low aphid density at 23/18°C (light/dark) temperature, 14 h/10 h (light/dark) photoperiod and 60–80% relative humidity, and ii) Alate aphids were reared at 20/16°C (light/dark) temperature, 12 h/12 h (light/dark) photoperiod and 60–80% relative humidity. Newly emerged alates of 0–48 h of age were used for experiments.
Host plant preference assays were performed with non-viruliferous alate aphids in several dual-choice assays within a 1 m3 arena. Aphids were offered as choices either mock inoculated Ler and Ll-0 plants, or LS-CMV-infected Ler and Ll-0 plants, or mock-inoculated and LS-CMV infected plants of either genotype. Twenty plants, ten of each choice class, were randomly placed in each arena during the trial. Two hundred winged aphids were placed in a flight platform similar to the one described by Fereres et al. [47] and released 0.5 m above the test plants. Settled alate individuals, as well as nymphs produced on each plant, were counted 24 h after their release.
To estimate the capacity of mock-inoculated and LS-CMV-infected Ler and Ll-0 plants to sustain vector populations, the intrinsic rate of natural increase of aphids was determined. Apterous adults of the M. persicae colony of equal age and weight were used. For synchronization, three apterous adult females were placed on each Arabidopsis test plant and confined using a plastic cup. The following day, adults were removed and three newly born nymphs were kept on each plant, which reached adulthood 7–8 days later. At that point, only one adult was kept per test plant, and every 24 h the newly born nymphs were counted and removed to avoid crowding effects that could influence the reproductive potential. The intrinsic rate of natural increase of the aphid population (rm) was calculated according to the equation proposed by Wyatt and White [48], , where Nd is the number of progeny produced by an adult during d days, d being the pre-reproductive period (number of days from birth to first reproduction). Between 25 and 34 plants per treatment, for a total of 116 plants, were used in this study.
Transmission assays were performed as described by Fereres et al. [49]. Briefly, groups of 30–40 apterous individuals of M. persicae were collected and after a pre-acquisition starvation period of one hour were released on LS-CMV infected source plants. All infected plants used as sources for transmission exhibited a saturated optical density (OD) in an ELISA test for CMV. Following a 10 min acquisition access period, groups of 5 aphids were transferred to each of the 3 week-old Arabidopsis target plants. A 24 h inoculation access period was allowed before spraying plants with imidacloprid (Confidor, Bayer). Plants were then transferred to an aphid-free growth chamber at 22°C under long day cycle. Three weeks post-inoculation, plants were tested by ELISA for LS-CMV infection. Samples were considered positive when their OD was 3 times higher than the negative control's OD after 24 h of incubation. A total of 37 source plants were sequentially tested, with transmission assayed to 252 Ler and 252 Ll-0 target plants. Transmission rate by a single aphid vector was estimated based on the Gibbs and Gower equation [46] (see above), from the number of positive samples (y), the total number of samples assayed (n), and the number of aphids used per transmission trial (k = 5).
All the above experiments involving M. persicae were performed twice, and results are presented as the mean values of both experiments.
Statistical analyses were performed using the statistical software package Statgraphics Centurion version 15.1.02 (StatPoint technologies, Inc., Warrenton, VA). The data sets were analyzed using analysis of variance (ANOVA) and transformation by was applied when necessary for data normalization.
The two genotypes of Arabidopsis used in this work, Ler and Ll-0, presented large differences in their life history. Lifespan (from germination until senescence, see Material & Methods) of Ler plants was of 57.25±0.70 days (d), while it was significantly longer, 87.88±1.06 d, for Ll-0 (Table 1) (F1,16 = 581.15, P≤10−5). Thus, lifespan defined a short-lived (Ler) and a long-lived (Ll-0) genotype. Differences in lifespan between Ler and Ll-0 were also associated with differences in each of any other temporal life-history trait measured (F1,16≥234.37, P≤10−5), such as time to bolting, flowering time or time to first silique shattered (Table 1). In addition to differing in temporal life-history traits, Ler and Ll-0 plants also differed broadly in morphology. As reported [40], [41] the allometric relationship between vegetative (rosette) and reproductive (inflorescence) parts was much larger for Ll-0, which had large rosette built of more than 150 leaves, as compared with the small rosettes of Ler plants, with less than 10 leaves (Figure S1). Also growth rate estimated by rosette diameter increase per day was higher in Ll-0 than in Ler, but the leaf mass per unit area did not differ between both genotypes (Table 1).
Ler and Ll-0 plants also differed in seed germination potential. While seed germination rate was similar for Ler and Ll-0 after a ten month dormancy period (Table 1), seeds of Ler had a germination rate of 95% as soon as siliques reached maturity. This was not the case for Ll-0, which seeds showed a germination rate of 0% at silique maturation. These data indicate a lack of stratification or dormancy requirements for Ler seeds, and the need of a dormancy period for Ll-0 seed germination.
To test if lifespan was related to LS-CMV multiplication in Ler and Ll-0 genotypes, virus accumulation in systemically infected leaves ten dpi was analyzed. Virus accumulation did not differ significantly in Ler and in Ll-0 plants (5.35±1.74 pg/ng and 2.57±0.28 pg/ng, respectively,F1,16 = 2.48, P = 0.1376).
Next, we tested host susceptibility to horizontal transmission, i.e., the probability of the host becoming infected by a viruliferous M. persicae vector. Transmission experiments were performed using Ler and Ll-0 plants as targets for transmission, and Ler, Ll-0 and Colgb1 plants as sources for virus acquisition by aphids. Data (Table 2) show that Ler and Ll-0 plants were similarly susceptible to LS-CMV aphid transmission when the aphids acquired the virus in plants of the same genotype (9.67±2.89 and 8.72±1.91% transmission for Ler and Ll-0, respectively, F1,74 = 0.59, P = 0.4431). Nonetheless, susceptibility to transmission in Ll-0 plants was dependent on the inoculum source, being significantly lower (F2,37 = 4.35, P = 0.0208) when aphids acquired the virus in Ler plants (3.83±0.85%) than when the virus was acquired in Ll-0 or in Colgb1 plants (8.72±1.91% and 10.48±2.13%, respectively). On the contrary, susceptibility of Ler plants was not affected by the genotype of the plants in which the virus was acquired (F2,37 = 0.28, P = 0.7548).
Last, susceptibility to vertical transmission through the seed was similar for Ler and Ll-0 plants, the rate of single seed transmission being of 1.98±0.70 and 2.47±0.82%, respectively (values from eight assayed mother plants, F1,16 = 0.21, P = 0.6537).
The effects of LS-CMV infection differed in the two Arabidopsis genotypes analyzed (Table 1). Infection resulted in an increase of the lifespan of Ll-0 plants to 99.25±2.19 d, in comparison with mock-inoculated plants (F1,16 = 22.07, P = 0.0003), while the lifespan of infected Ler plants did not differ from that of mock-inoculated plants (58.5±0.5 days, F1,16 = 2.11, P = 0.1685). The longer life-span of infected Ll-0 plants is the result of reduced growth rates. Thus, infection resulted in a reduction of the relative rate of rosette growth for Ll-0 plants (F1,10 = 10.53, P = 0.0118) but not for Ler plants (F1,10 = 0.11, P = 0.7476), and in a delay in the development of the inflorescence, as shown by the time to flowering and to silique maturation (Table 1). Neither the leaf mass per unit area (F1,6≤0.41, P≥0.5587) or the seed germination rate after a 10 month dormancy (F1,16≤0.16, P≥0.6930) were affected by infection in either genotype.
Two parameters related to host competence, i.e., to the host capacity as a source of virus for vector transmission, were also estimated for Ler and Ll-0 plants: their capacity to attract vectors and to sustain vector populations.
Results from dual free-choice experiments showed that a similar number of winged morphs of M. persicae had settled on mock-inoculated Ler and Ll-0 plants, and a similar number of nymphs were produced in both genotypes after 24 h, which is an indication of the time an aphid spends on the plant on which it has landed (Table 3, Cage 3, F1,40 = 0.92, P = 0.3424 and F1,40 = 1.47, P = 0.2333 respectively). When mock-inoculated and infected plants were compared, it was found that more aphids settled on mock-inoculated than on LS-CMV-infected plants of both genotypes (Table 3, Cage 1 and Cage 2). While this preferential settlement was only marginally significant for each plant genotype, Ler or Ll-0, when analyzed separately (F1,40 = 1.72, P = 0.1981 and F1,40 = 3.37, P = 0.0743 for Ler and Ll-0 respectively), it was significant when data from both genotypes were pooled together (F1,80 = 4.86, P = 0.0304), indicating that virus infection reduced aphid preference after 24 h. Similarly, more nymphs were produced in mock-inoculated plants than in infected ones (F1,40 = 1.92, P = 0.1738 and F1,40 = 9.20, P = 0.0043 for Ler and Ll-0 plants, respectively, the difference being significant when data from both genotypes were pooled, F1,80 = 4.40, P = 0.0391). Interestingly, more winged adult aphids were recovered from infected Ler than from infected Ll-0 plants (Table 3, Cage 4; F1,40 = 4.53, P = 0.0398), and those winged adults had produced more nymphs after 24 h (F1,40 = 6.98, P = 0.0119).
We also estimated the intrinsic rate of natural increase (rm), and other reproduction-related parameters, of the M. persicae population on mock-inoculated and LS-CMV-infected plants of Ler and Ll-0 genotypes. Results in Table 4 show that the aphid's pre-reproductive period was affected neither by plant genotype nor by infection status (F3,116 = 0.99, P = 0.3996). On the other hand, a trend was observed towards higher numbers of total nymphs produced, daily fecundity and rm from adults feeding on infected plants as compared with those from mock-inoculated plants. Although this trend was not statistically significant for all parameters (see Table 4), it was consistent for both genotypes (F1,53≥2.52, P≤0.1182 and F1,63≥2.37, P≤0.1290 for Ll-0 and Ler, respectively) and statistically significant when data from both genotypes were pooled together (F1,116≥5.41, P≤0.0217). Interestingly, when plants were considered according to genotype, and not to infection status, it was found that Ler plants were able to better support aphid population growth than Ll-0 plants, as rm values differed significantly (0.3697 vs. 0.3506, F1,116 = 3.89, P = 0.0509).
To analyze the differences between short- and long-lived genotypes of Arabidopsis in reservoir potential, and the effects of host plant density and population composition in this trait, we used a model in which the dynamics of the plant and vector populations were jointly considered, based on that proposed by Madden et al. [50]. In this model, the rate of plant infection depends on the interaction of susceptible non-infected plants with infectious vectors, hence on the density of infected plants and infectious vectors. Our model differs from that of Madden et al. [50] in that: i) We considered the plant population divided into two classes, susceptible non-infected (S), and infected (I). We did not consider a separated class of infected-non-infectious plants (i.e. we do not consider a latent period) since the rate of colonization of host tissues and organs by the virus, which determines transmissibility, is not genotype-dependent (our unpublished results). Neither did we consider a class of recovered plants, as CMV causes systemic chronic infections so that plants, once infected, remain so until the end of their life cycle. Also, infected plants were considered to be a source of infection for aphids until the end of their life cycle, because aphids fed on stalks and cauline leaves until senescence (our unpublished observations). ii) The aphid vector population was divided into two classes, non-viruliferous (X), and viruliferous aphids (Z), i.e., aphids that had acquired the virus by feeding on infected, I, plants and were able to transmit it by feeding on non-infected, S, plants. Since CMV is transmitted non-persistently, there is no latent period for transmission. iii) A major difference with Madden et al. [50] model is that virulence, ν, expressed as the effect of infection on plant mortality) was considered in our model. For a single host, the dynamics of the model is described by the following equations:(1)(2)(3)(4)These equations represent the variation with time (days) of the density of non-infected susceptible plants, S (Eq. 1), of infected plants, I, (Eq. 2), both expressed as plants×m−2, and of virus-free aphids, X, (Eq. 3) and viruliferous aphids, Z, (Eq. 4), expressed as aphids×plant−1. K indicates the maximum density of the plant population. Parameter m indicates the per capita plant mortality rate, and parameter v indicates the variation of plant mortality rate due to infection (i.e., virulence). The transmission rate is decomposed into two parameters, Φ, which indicates the number of susceptible plants visited per day by an aphid, and b, which represents the probability of virus transmission per vector visit. Similarly, parameter a indicates the probability that a vector acquires the virus at each visit. Because transmission is non-persistent, viruliferous vectors lose the virus with a rate τ, returning to the X class of virus-free vectors. Last, α indicates the per capita mortality of aphids, which is not affected by their viruliferous state.
This model was extended to two hosts and its dynamics for Host 1 are described by the set of equations:(5)(6)(7)(8)Subscripts 1 and 2 denote the host. Equations 5–8 describe the variation with time (days) of the density of susceptible plants, Si (Eq. 5), of infected plants, Ii, (Eq. 6), expressed as plants×m−2, and of virus-free aphids, Xi, (Eq. 7) and viruliferous aphids, Zi, (Eq. 8), expressed as aphids×plant−1, i = (1,2), and the model differs from the single host model (Eq. 1–4) in allowing for aphids to visit Host 2 from Host 1 (parameter Φ12), from which acquisition of virus and transmission to Host 1 from Host 2 occurred (parameter b21). A similar set of equations describes the dynamics for Host 2 (not shown); the only differences would be the substitution subscript numbers 2 for 1 and vice versa.
Definitions and values of model parameters are shown in Table 5, in which Host 1 and Host 2 represent a short-lived and a long-lived genotype, respectively. Accordingly, model parameters were estimated from the experimental data obtained for Ler and Ll-0, respectively, as in the previous sections. Because in Arabidopsis CMV infection is not lethal, virulence (νi) was estimated as the variation of the plant's lifespan by infection, DI, as compared with the lifespan of uninfected susceptible plants DS, following Day [51]. Mortality rates relate to lifespan by m = 1/DS, for non-infected plants, and (m+νi) = 1/DIi for infected plants. Data in Table 1 show that CMV infection does not affect the lifespan of Ler, while it delays the completion of Ll-0 life cycle, resulting in a negative virulence. Note that the lifespan of susceptible non-infected plants in Table 5 is shorter than the lifespan of mock-inoculated plants in Table 1, as we considered that plants would not attract aphids until they had reached the four-leaf stage. Similarly, the mortality rate of aphids, α, was estimated as the inverse of their lifespan, which was of an average of 40 d in the conditions in which experiments were performed. Transmission probability by single aphids during each visit of an S plant from an I plant (parameter b) was according to the data in Table 2. Following Madden et al. [50] we considered that the probability of acquisition of a non-persistently transmitted virus (parameter a) is the same as the probability of transmission, and that the aphid remains viruliferous for a maximum of 6 h (parameter τ). The frequency of aphid visits to S plants from I plants was not estimated experimentally, and was given arbitrary values varying between 0.01 and 1. These values may be realistic, because epidemiological studies of CMV in different regions of Spain for different years indicate transmission rates of 0.008–0.122 days−1 [52]. However, because mock inoculated and infected plants of genotypes Ler and Ll-0 showed different vector preference, the probability of aphid visits from infected Ll-0 to healthy Ler plants was considered to be 0.63 times the probability of visits from infected Ler to healthy Ler plants, infected Ll-0 to healthy Ll-0 or infected Ler to healthy Ll-0 plants, according to the data in Table 3.
For simulations, we considered a monomolecular growth of the population of susceptible plants, θ = rp (K−T), where T = S+I, is the total plant population density, K is the maximum density of the population, and rp is its growth rate. To make the plant population constant during a growing season, we set rp = 1 to get a constant value of T = K. Simulations were performed for 5<K<500, according to data on the variation of plant density in wild Arabidopsis populations in Central Spain over sites and years [29, and unpublished results]. Similarly, we made the aphid population constant by considering its growth, ψ as monomolecular according to ψ = ra (Q−P)/K, where P = X+Z is the total aphid population, Q is its maximum density, and ra = 1 for a constant aphid population density per plant. Q was made to vary between 0.1 and 5 aphids per plant, which are realistic values of density of aphid populations in Central Spain [53]. For all simulations the initial condition was that 2% of the plants were CMV infected (i.e., values of I1 and I2 of 0.02Ki), according to the experimentally determined rates of vertical transmission of LS-CMV in Arabidopsis (this work). All simulations were done in R.
For all model simulations the frequency of short-lived Host 1 and of the long-lived Host 2 in the plant population was made to vary between 0 and 100%. The threshold values of the density of the aphid population, Q, and of the probability of aphid visits to healthy plants from infected ones, Φ, necessary for the occurrence of a CMV epidemic, depended on the total plant population density (K). For K = 5 epidemics occurred at Φ≥1 and Q≥3, while for K = 50 and K = 100 epidemics occurred for Φ≥0.5 and Q≥1, and for K = 250 or higher, epidemics occurred for Φ≥0.1 and Q≥3, or for Φ≥0.5 and Q≥1 (Fig. 1 and Table S1). Hence, plant and aphid population densities are primary factors determining CMV infection. Note that although Φ was made to vary independently of Q, it is known that aphid mobility increases as aphid population density increases [50], so that in a real situation Φ would not be independent of Q.
The incidence of CMV (i.e., the fraction of infected plants, I) at equilibrium was compared first for single-genotype populations of either the short-lived Host 1 or the long-lived Host 2. It was found that incidence was higher for the long-lived host for a broad range of values of plant and aphid population density (K, and Q values) and of probability of aphid visits to non-infected plants from infected ones (Φ values). Incidence was higher for the short-lived host only for plant and aphid densities, and probability of aphid visits to plants, that would result in low transmission rates, e.g., for any Φ and Q value if K = 5, and for low Φ and Q values for higher K, e.g., Φ = 0.5 if K = 50, Φ = 0.5 if Q = 1 when K = 100, or Φ = 0.1 for Q = 5 or Q≥3 when K = 250 or K = 500, respectively (Fig. 1 and Table S1). Hence, the highest competence and capacity to sustain aphid populations of the short-lived genotype is countered by the longer infectious period of the long-lived one when host and/or aphids are abundant and transmission is more effective.
Then, CMV incidence at equilibrium was compared for populations that were built of mixtures of short- and long-lived hosts in different proportions. For heterogeneous host populations, model simulation analyses showed that for any values of plant and aphid population density, and of the probability of aphid visits to non-infected plants from infected ones, K,, Q, and Φ, the total incidence in the plant population, and the incidence for each plant genotype, was always higher than for single-host genotype populations (Fig. 1, and Table S1). This result indicates that when the population is heterogeneous, there is a balance between the different epidemiological parameters, such as transmission rate, vector preference or infectious period, that result in higher incidence. However, for all explored conditions, the total incidence increased with increasing frequency of the short-lived host in the plant population, and incidence was always higher for the long-lived than for the short-lived host (Fig. 1 and Table S1). These two results indicate that the short-lived host was always a better source of infection for the long-lived one than vice versa. Thus, for all analyzed conditions, the short-lived host was a more efficient reservoir than the long-lived one. The difference between CMV incidence in the long-lived and short-lived hosts, however, depended on the density of the aphid population, Q, and their mobility, Φ, so that for any plant density value the difference increased with increasing Q if Φ≥1, but had a minimum at Q = 3 for low Φ values (e.g., 0.5). Thus, the efficiency as a reservoir of the short-lived host depended non-linearly on the rates of transmission (Fig. 1 and Table S1). The difference between CMV incidence in the long- and short-lived hosts also depended on the density and genetic composition of the host population: for K = 5, the difference had relative maxima for different frequencies of the short-lived host according to the value of Q; for K = 50, Φ = 0.5 and Q = 1, the difference in CMV incidences in both hosts increased monotonically with the frequency of the short-lived host in the population. For any other values of K, Φ and Q, the difference between incidences in the long- and short-lived hosts decreased monotonically with the frequency of the short-lived host in the population (see Fig. 2 for examples). Thus, the efficiency of the short-lived host as a reservoir depended in a complex non-linear way on the genetic composition of the host population, modulated by the density of the host and vector populations and by the rate of aphid mobility, hence modulated by the factors determining transmission efficiency.
We considered next the values of CMV incidence in each host, and of their difference, after 150 iterations, which informs about the evolution of the epidemics before equilibrium. Also, this is a temporal frame more in line with the Arabidopsis life cycle of about 5 months. After 150 iterations the data (Fig. 2 and Table S2) show a more dramatic variation of CMV incidence according to values of plant and aphid density and aphid mobility, and according to the frequency of the short-lived host in the population, than at equilibrium. Notably, the difference between incidences in the long- and short-lived hosts showed relative maxima for a wider range of values of plant and aphid density and aphid mobility. The difference could even show negative values, either with relative minima or monotonically increasing with increasing proportion of the short-lived host in the population, depending on the density of the host and vector populations. It is interesting to underscore that when plant population density was high (high values of K), and both aphid population density and aphid mobility were low (low values of Q and Φ), and the proportion of the short-lived host in the population was low, the incidence in the short-lived host was higher than in the long-lived one. This result indicates that under this set of conditions the epidemic progressed faster in the short-lived host at earlier times, and faster in the long-lived host at later times. Thus, this result underlies the role of the short-lived host as a reservoir.
Vectored viruses comprise a large fraction of emerging pathogens of humans, animals and plants, and important research efforts are devoted to understand the network of ecological and evolutionary factors that determine virus emergence [2]–[5], [54]. Generally, emerging pathogens can be permanently maintained in reservoir hosts [2], [3], and identification of the determinants of host reservoir potential is central to understand emergence and infection dynamics. Within this context, the relationship between host lifespan and reservoir potential is an understudied field. It has been proposed that long-lived hosts will invest more in costly defenses against pathogens because they are more exposed to infection than short-lived hosts. Consequently, short-lived hosts will be more susceptible to infection, more competent as sources of infection and/or will sustain larger vector populations, hence being effective reservoirs for the infection of long-lived hosts [8], [14], [55]. Evidence for a link between host lifespan and reservoir potential derives mostly from the comparison of species within the host range of a pathogen [6], [7], [22], [23], [56]–[58], and the underlying mechanisms have rarely been analyzed [e.g. 8,24,58]. To our knowledge, the relationship between host lifespan, defense and reservoir potential, has not been analyzed at the intra-species level, in spite of abundant evidence of genetic diversity resulting in polymorphisms for lifespan and defense within species.
It has been shown that tolerance to CMV infection in A. thaliana correlates positively with lifespan, the longer-lived genotypes showing a higher tolerance to CMV infection [40], [41]. Higher tolerance was, at least in part, due to the ability of long-lived genotypes to modify their developmental schedule upon infection, so that more resources were allocated to reproduction than to growth [40], [41]. In this study we focus on two Arabidopsis genotypes that represent extremes of lifespan and tolerance, i.e., Ler and Ll-0 [40], [41], as representatives of the short- and long-lived genotypes that co-exist in wild Arabidopsis populations in different regions of the Iberian Peninsula [29], [32], [35]. Ler and Ll-0 plants differed sharply in the temporal parameters of their life-cycle, in their morphology and in other relevant life-history traits (Table 1 and [40], [41]). Among these, it is noteworthy that seeds of the short-lived genotype, Ler, did not require a period of dormancy before germination, which will allow them to have more than one generation per growing season under field conditions [29], [34]. This will not be the case for the long-lived Ll-0-like genotypes that may require a period of dormancy for germination and vernalization for flowering. CMV infection did not affect germination rate of any genotype tested, and CMV vertical transmission rate did not differ for the short- and long-lived genotypes. Thus, it is tempting to speculate that CMV will follow different strategies to increase its fitness in short- and long-lived genotypes. In the long-lived genotype, which has only one generation per year, infection increases the lifespan (i.e., the infectious period), which is not modified in the short-lived genotype, which may have more than one generation per year. Interestingly, in this system we cannot equate short and long lifespan with the physiological phenotypes of Quick- or Slow-Return, since rosette relative growth rate was higher in the long-lived Ll-0 genotype than in the short-lived Ler, but leaf mass per unit area did not differ between both genotypes. This unexpected result is at odds with evidence derived from the comparison of different plant species [8], [24], [59], [60], and suggests that the trade-off between lifespan and development and reproductive rates [61] may not apply at the level of within-species diversity, or at the temporal scale at which the lifespan of Ler and Ll-0 differ. The different performance of aphids in each genotype, regardless of similar LMA, could be explained by differences in nutrient composition in the phloem sap [62], which would not translate into detectable differences in LMA.
Using these two Arabidopsis genotypes we analyzed if lifespan, in addition to correlating positively with tolerance to CMV [41], also correlated with susceptibility, competence and the ability to sustain vector populations; three parameters that determine reservoir potential [8]. Ler and Ll-0 plants were similarly competent sources of infection, as well as similarly susceptible to infection when transmission assays were performed between plants of the same genotype. These results agree with the non-significant difference in virus accumulation between Ler and Ll-0 plants, transmission rates of CMV in different hosts having been shown to increase with virus titer until saturation at high virus titers [63], [64]. On the other hand, the susceptibility to infection of Ll-0, but not of Ler, depended on the genotype of the plant from which the aphid vector acquired the virus. Hence, the susceptibility of the long-lived genotype will be, on the average, lower in a heterogeneous plant population than that of the short-lived genotype. The capacity to sustain aphid vector populations was higher in the short-lived Ler genotype than in the long-lived Ll-0 one, when both mock-inoculated and CMV-infected plants were considered. In addition, more aphids landed and settled, and more nymphs were generated, in CMV-infected, but not in mock-inoculated, Ler than in Ll-0 plants, indicating that the capacity to sustain vector populations was differentially modified for each genotype by virus infection, and was higher in the short-lived genotype. It is noteworthy that the results of our aphid host preference experiments agree with previous reports involving different host plant species infected with CMV, in which it was shown that after an initial attraction to CMV-infected plants, aphids migrated and settled into mock-inoculated ones at some time between 30 and 60 min after landing [65], [66]. All these results taken together indicate that two of the three epidemiological parameters related to reservoir potential, susceptibility to infection and capacity to sustain vector populations, were higher in the short-lived than in the long-lived genotype, which would indicate a higher reservoir potential of the short-lived genotype. Thus, our results agree with predictions on the relationship between host lifespan and defense evolution [14], [55] and with evidence derived from the comparison of different hosts of multi-host parasites infecting both plants and animals [6]–[8], [56]–[68]. It is important to point-out that most previous evidence for a relationship between lifespan, defense and reservoir potential, was derived from the comparison of different species with large differences in lifespan, of the order of years, e.g., rodent species with life expectancies of 2 vs. 4–8 years, or annual vs. perennial plant species [8], [56]. Our results extend that evidence to a much narrower time-scale in lifespan differences, in the order of weeks, as expected for the variability in lifespan at the within species level for a short-lived plant. Although our experimental results derive from the study of only one short- and one long-lived genotype, they indicate that a link between lifespan and defense at the within-species level will result in polymorphisms for reservoir potential within a heterogeneous single-species host population. There is growing evidence for covariance between host traits determining reservoir potential and local extinction risk [55]. If longer lived, less susceptible and competent host genotypes are also at higher local extinction risk is a question to be explored, as it could determine both disease dynamics and the evolution of apparently unrelated traits in host populations.
In terms of reservoir efficiency, the lower defenses, i.e., higher susceptibility and capacity to sustain vector populations, of the short-lived Arabidopsis genotype could be balanced by the longer infectious period of the long-lived genotype, and the balance might be modulated by the demography and genetic composition of the host population. To explore how the efficiency as a reservoir is affected by host demography and by the genetic composition of the host population, we developed a simple epidemiological model which is essentially a simplification of that proposed by Madden et al. [50] with the addition of a virulence parameter to take into account the effect of infection on host plant mortality. The model was run for a set of realistic parameters derived from the above-discussed experiments with Ler and Ll-0, and for a wide range of host plant and aphid population densities according to field observations [29], [53]. The negative virulence in the long-lived Host 2 deserves some consideration. The negative virulence value result from the increase in lifespan of long-lived Arabidopsis genotypes upon infection, which associated with resource reallocation resulting in tolerance [41], and is linked to a decrease in growth rate (Table 1). Under field conditions in which plants will compete, the slower-growing infected long-lived plants could be at a competitive disadvantage, hence suffering from a positive virulence as was shown in experiments in which plant density was manipulated [21]. Model simulation analyses were done considering different virulence values, the negative virulence value shown in Table 5, virulence 0, or a positive virulence of the same amount, with no substantial difference in the results. Thus, even in the case that negative virulence were an artifact of green-house experimentation, it will not affect the outcome of model simulation analyses.
When the model was run for single-genotype populations of either short-lived or long-lived hosts, it was found that the short-lived host population sustained higher incidence of CMV only at low plant and/or aphid populations densities, i.e., under conditions of low transmission rates. Thus, the higher reservoir potential of short-lived genotypes due to lower defense was partly countered by the longer infectious period of long-lived genotypes when transmission was highly efficient (i.e. higher K, Q and Φ). In agreement with this result, we found that the equilibrium incidence of CMV was always higher for mixed-genotype host populations than for single-genotype populations of either short- or long-lived hosts, underscoring the role in determining the reservoir efficiency of both host defense and infectious period. The total CMV incidence in mixed-genotype populations increased with increasing frequency of the short-lived genotype, and incidence was always higher in the long-lived genotype subpopulation, indicating that for the realistic parameters and within the wide range of conditions considered, the short-lived genotype is always a better reservoir than the long-lived one. How good a reservoir was the short-lived genotype, however, depended in a complex way on the density of the host and aphid populations, as determinants of the probability of transmission. These non-linear effects were more pronounced when we analyzed the predictions of the models after 150 iterations, which may approximate the Arabidopsis growth period, rather than the equilibrium values. After 150 iterations, the difference between CMV incidence in long- and short-lived hosts showed maxima, or even became negative, for a given K value according to the Q and Φ values (Fig. 2 and Table S2). Data from 150 iterations also showed that the relative rate of epidemic growth in each host varied with time. The complex relationship found between efficiency as a reservoir of the short-lived genotype, the density of host and aphid populations, and the genetic composition of the host population, is in agreement both with model predictions on the variation of defense with host population traits [14], [18] and with our experimental results with the same host-pathogen system [21].
In summary, in this work we show that the hypothesis stating that there is a correlation between host lifespan and investment in defenses against pathogens, developed and tested for the different hosts of multi-host pathogens, also holds for two genotypes of a single host species, which may differ in lifespan at much smaller temporal scales. Analyses of more short- and long-lived genotypes would be required for generalization, but our results indicate that the less efficient defenses of short-lived genotypes result in their higher reservoir potential. However, the reservoir potential of short-lived genotypes may be balanced in heterogeneous host populations by the longer infectious period of long-lived genotypes. Model simulations under realistic parameter ranges showed that this balance is modulated according to complex, non-linear relations, by the demography of both the host and vector populations, and by the genetic composition of the host population, an important conclusion that often is not considered in analyses of the evolutionary ecology of pathogen emergence. Thus, within-species genetic diversity for lifespan and defenses against pathogens will result in polymorphisms for pathogen reservoir potential, which will condition within-population infection dynamics. These results should be taken into account in the future in joint analyses of the population genetics of traits determining host defense and lifespan, to get a better understanding of the evolution of defense against pathogens in host populations.
|
10.1371/journal.ppat.1008005 | Somatic hypermutation to counter a globally rare viral immunotype drove off-track antibodies in the CAP256-VRC26 HIV-1 V2-directed bNAb lineage | Previously we have described the V2-directed CAP256-VRC26 lineage that includes broadly neutralizing antibodies (bNAbs) that neutralize globally diverse strains of HIV. We also identified highly mutated “off-track” lineage members that share high sequence identity to broad members but lack breadth. Here, we defined the mutations that limit the breadth of these antibodies and the probability of their emergence. Mutants and chimeras between two pairs of closely related antibodies were generated: CAP256.04 and CAP256.25 (30% and 63% breadth, respectively) and CAP256.20 and CAP256.27 (2% and 59% breadth). Antibodies were tested against 14 heterologous HIV-1 viruses and select mutants to assess breadth and epitope specificity. A single R100rA mutation in the third heavy chain complementarity-determining region (CDRH3) introduced breadth into CAP256.04, but all three CAP256.25 heavy chain CDRs were required for potency. In contrast, in the CAP256.20/27 chimeras, replacing only the CDRH3 of CAP256.20 with that of CAP256.27 completely recapitulated breadth and potency, likely through the introduction of three charge-reducing mutations. In this individual, the mutations that limited the breadth of the off-track antibodies were predicted to occur with a higher probability than those in the naturally paired bNAbs, suggesting a low barrier to the evolution of the off-track phenotype. Mapping studies to determine the viral immunotypes (or epitope variants) that selected off-track antibodies indicated that unlike broader lineage members, CAP256.20 preferentially neutralized viruses containing 169Q. This suggests that this globally rare immunotype, which was common in donor CAP256, drove the off-track phenotype. These data show that affinity maturation to counter globally rare viral immunotypes can drive antibodies within a broad lineage along multiple pathways towards strain-specificity. Defining developmental pathways towards and away from breadth may facilitate the selection of immunogens that elicit bNAbs and minimize off-track antibodies.
| Broadly neutralizing antibodies (bNAbs) develop in some HIV infected individuals, partly due to their complex evolutionary pathways that are characterized by extensive somatic hypermutation (SHM). Furthermore, bNAbs within a lineage may form a minor subset, amidst many strain-specific “siblings”, indicating that minor sequence differences between lineage members can significantly affect neutralization. Here, we define mutations that limit breadth in two “off-track” members of the CAP256-VRC26 bNAb lineage, and show that these occur with relatively high probability. A dominant autologous virus with a globally rare V2 sequence appears to have selected for an off-track antibody, providing a mechanism for the development of this antibody during infection. These data highlight the complex interdependencies between high levels of SHM and breadth, as mutations that neutralize autologous viruses may limit heterologous breadth. Consequently, strategies to increase SHM by repeated vaccinations will require careful antigen selection to focus the humoral response to globally common epitopes, limiting off-track responses.
| Antiretroviral therapy has transformed HIV from a progressive fatal infection to a manageable chronic disease [1]. However, a preventative vaccine is urgently needed as drug resistance, access to medication and adverse side effects hamper the utility of antiretroviral drugs. Despite intensive research, strategies to induce protective immune responses by vaccination have achieved little success [2]. While neutralizing antibodies are elicited during HIV infection, the response is typically strain-specific. However, ~20% of individuals develop broadly neutralizing antibodies (bNAbs) which potently neutralize diverse global viruses in vitro and protect non-human primates from infection [3–8]. Therefore, understanding the mechanisms behind the elicitation and evolution of bNAbs is central to the development of an HIV vaccine.
During infection, the humoral response targets the HIV envelope (Env) glycoprotein which consists of three heavily glycosylated non-covalently linked gp41-gp120 protomers. While strain-specific antibodies recognize exposed and variable sites, bNAbs target relatively conserved and occluded sites, including the membrane proximal external region (MPER), gp120-gp41 interface, CD4-binding site (CD4bs), N332 supersite and a quaternary V1V2 epitope at the trimer apex [9]. Many V1V2 directed bNAbs have atypically long (>24 amino acids) third complementarity determining regions of the heavy chain (CDRH3) which provides access to glycan shielded and buried epitopes [10–14]. Additionally, bNAbs frequently exhibit unusually extensive somatic hypermutation (SHM), which suggests multiple rounds of affinity maturation. Understanding how these unusual features arise is critical to recapitulating this process by vaccination.
The degree to which SHM and affinity maturation occurs is influenced by HIV diversity and antigen load [15]. A number of longitudinal studies describing the “arms race” between viral quasispecies and bNAb lineages have provided insights into the evolution of breadth [16]. These studies have identified bNAb-initiating Envs that engage antibody precursors as well as emerging viral variants, which select bNAb intermediates [12,17–23]. We have extensively studied virus/antibody co-evolution in the superinfected CAP256 donor from whom 33 members of the CAP256-VRC26 V2-targetting bNAb lineage were isolated [11,12,22]. The evolution of this lineage was elucidated through analysis of longitudinal viral and antibody deep sequences from 59–206 weeks post infection [11,12,22]. Viruses from six weeks post infection are sensitive to neutralization by CAP256 lineage members but by 59 weeks post infection most viruses are resistant. Similar to other V2-directed bNAbs, the CAP256-VRC26 lineage members have very long (35–37 amino acid, aa) anionic and tyrosine-sulfated CDRH3s that recognize the electropositive V2 apex [24]. Modelling and mapping studies show that several residues within the CDRH3 are critical for epitope recognition, including the tyrosine sulfated YYD motif at the apex of the CDRH3 [24]. This motif is present in the unmutated common ancestor (UCA) of the CAP256-VRC26 lineage, 27 of the 33 lineage members as well as other V2-directed bNAbs such as PG9/PG16 [10,11]. In addition, the CDRH1 has been shown to stabilize the CDRH3, and residues in CDRH2 have been implicated in glycan recognition [25,26]. Within strands B and C of the viral Env protein, the CAP256-VRC26 lineage’s footprint includes residues N160, R166, D167, K168 and K169 and is similar to other V2 glycan bNAbs (PG9/PG16). The breadth, potency and extent of SHM of CAP256-VRC26 lineage members ranges from 2–63%, 0.003–5 μg/mL and 4.2–18%, respectively [11].
SHM, which is mediated by activation‐induced cytidine deaminase (AID), is essential for breadth as bNAb germline revertants generally lack substantial binding and neutralization [10,27–30]. Mutations that are selected during affinity maturation may increase antibody-antigen affinity directly at the paratope or indirectly by improving stability [31,32]. SHM is a semi-random process as AID hot-spots (WRCH/Y, W = A/T, R = A/G and H = A/C/T) and cold-spots (SYC, S = C/G and Y = C/T) are distributed throughout the antibody variable regions. Consequently, mutations associated with breadth are unlikely to occur if they are in cold-spots and/or require more than one nucleotide change [33,34]. Additionally, neutral mutations can accumulate via co-selection and constitutive background mutational noise [32]. Although increased SHM is generally associated with neutralization breadth, exceptions exist within the CAP256-VRC26 and other bNAb lineages. We have previously described these non-broad but highly mutated CAP256 antibodies as “off-track” [22]. Such off-track antibodies have been recovered from bNAb lineages in several individuals. 1AZCETI5 is a clonal member of the CD4bs CH103 bNAb lineage, and shares VDJ genes, CDRH3 length and extent of SHM with bNAb CH103, but the former only neutralizes autologous viruses [18]. The PGDM1400 bNAb, a member of the V2 glycan targeting PGDM lineage with 83% breadth has similar genetic properties to PGDM1406, but the latter exhibits limited breadth (6%) [35]. Similarly, within the PGT121 N332 targeting bNAb lineage are two heavy chains (HC) (22H and 6H) that are related to broad members but exhibit no neutralization [36]. As monoclonal antibody isolation methodologies are designed to recover bNAbs, the proportion of off-track antibodies in bNAb lineages is unknown. However, the finding that 40% of related antibodies recovered by flow cytometry from the donor of the CD4bs bNAb, VRC-PG04, were off-track, indicates that such antibodies may be common [37].
Collectively, studies into co-evolution and bNAb targets have identified candidate antigens to initiate antibody lineages and guide the development of breadth. Indeed, native-like Env trimers administered in combination or sequentially elicits tier 2 neutralizing antibodies in mice and rabbits, providing a strong rationale to test similar immunogens in people [38–42]. Importantly, the mutations crucial for breadth and potency as well as those that confer an off-track phenotype in bNAb lineages are largely unknown. Identification of these mutations will inform the design of vaccine antigens that elicit breadth but dampen the evolution of off-track responses. The high degree of sequence identity between CAP256-VRC26 bNAbs and their off-track relatives, the richly populated lineage and the availability of contemporaneous viral sequences is an ideal opportunity to compare off-track antibodies with bNAbs to determine why these antibodies lack breadth.
Here we define key residues that restrict neutralization breadth and potency in two off-track antibodies within the CAP256-VRC26 lineage. We found that in one off-track/bNAb pair, residues in each of the three CDRHs were necessary for potency and that a key breadth-conferring mutation in the CDRH3 was highly improbable. In the second pair, we show that the CDRH3 was completely responsible for modulating neutralization and that the mutations leading to the off-track phenotype were relatively probable. Lastly, through epitope mapping we identified viruses that guided the evolution of an off-track antibody toward a globally rare immunotype. This study highlights the stochastic nature of SHM and shows that breadth-restricting mutations typically arise with a greater probability than breadth-conferring mutations. This work provides insights into the evolution of breadth and mature strain-specificity, informing the design of immunogens that minimize the elicitation of off-track mutations while guiding evolution to globally conserved sites.
To determine the genetic determinants that controlled the off-track phenotype, we selected two bNAbs with high sequence identity to two off-track antibodies. Off-track antibodies were defined as having low neutralization breadth despite their capacity to neutralize early autologous viruses [12], extensive sequence maturation and identity to related bNAb lineage members. The CAP256.25 bNAb was paired with the off-track antibody CAP256.04 and bNAb CAP256.27 was paired with the off-track antibody, CAP256.20. The two pairs of antibodies cluster together on a maximum likelihood tree (Fig 1A), consistent with our previous observations that antibodies with considerably different breadth may differ by relatively few amino acids [22]. The HC and light chain (LC) of CAP256.04 (30% breadth, 0.27 μg/mL potency, 9% SHM) both share 92% sequence identity with the HC and LC of bNAb CAP256.25 (63% breadth, 0.003 μg/mL potency, 12% SHM) (Fig 1A). Within the second pair, the HC and LC of CAP256.20 (2% breadth, 1.87 μg/mL potency, 16% SHM) shares 93% and 96% sequence identity with the CAP256.27 HC and LC (59% breadth, 0.047 μg/mL potency, 16% SHM), respectively (Fig 1A). A total of eighteen amino acids differ between the HC and 11 amino acids between the LC of CAP256.04 (orange) and CAP256.25 (brown) which are displayed as yellow spheres in an alignment of the two antibodies (Fig 1B). Seventeen amino acids (yellow spheres) differentiate the HCs and four the LCs of CAP256.27 (purple) and CAP256.20 (light purple) (Fig 1B).
As neutralization by members of the CAP256 bNAb lineage has been largely attributed to the HC [12], we first confirmed that the LC was not responsible for the off-track phenotype of CAP256.04 and CAP256.20. We generated chimeras between the HC and LC of each off-track/bNAb pair (25HC+04LC, 04HC+25LC and 27HC+20LC, 20HC+27LC) and tested their neutralization against both CAP256 infecting viruses (the primary and superinfecting viruses) and 14 heterologous viruses. The titers did not differ significantly (Wilcoxon signed rank test) between the chimeric antibodies and the natural antibody from which the HC was donated (Fig 1C). These data indicate that the LC has no impact on the neutralization breadth within these pairs, consistent with structural analyses showing no light chain contacts with the viral epitope [26], and we therefore focused on the HC in subsequent studies.
To determine which regions of the HC restrict the breadth of CAP256.04, we constructed several chimeras and point mutants between CAP256.25 and CAP256.04, focusing initially on the CDRHs. The sequence variation between the CDRHs is shown in Fig 2A and includes charge and hydrophobicity differences between the two antibodies (R28S, D30N, G31R, H53Y, K57D, A100rR and N100ddF) and an insertion (G100w) in CAP256.25 that is absent from CAP256.04 (Fig 1B). We first transplanted all the CAP256.25 CDRHs (12 amino acid changes in total) into CAP256.04, creating CAP256.0425H123 (Fig 2B). Antibodies were tested against a panel of 16 viruses that were sensitive to CAP256.25 (geometric mean potency of 0.002 μg/mL against this panel), of which 8 were neutralized by CAP256.04 (with a 230-fold lower geometric mean potency of 0.17 μg/mL) (Fig 2B). The neutralization breadth and potency of the CAP256.0425H123 chimera (which neutralized all 16 viruses with a geometric mean potency of 0.008 μg/mL) matched that of CAP256.25 (0.002 μg/mL) (Fig 2B). In the reverse experiment, introducing the framework (FR) regions from CAP256.25 into CAP256.04 (CAP256.0425FR123) had no effect on breadth (Fig 2B). These data suggest that the mutations in the CDRs, rather than the FW regions restrict the breadth of CAP256.04.
To dissect the role of individual CDRHs, each was individually swapped between the two parental antibodies. CAP256.0425H1 neutralized two additional viruses (ZM214 and ZM249) that were resistant to CAP256.04 (Fig 2B). In addition, we observed virus-specific increases in potency for some sensitive viruses (with a 6-18-fold improvement for Du156, Q259, and Du422 and a 250-fold increase in potency for CAP45). In contrast, CAP256.0425H2 neutralized the same number of viruses as CAP256.04 (Fig 2B), though again slight virus-specific differences in potency were observed, with CAP45 becoming resistant to CAP256.0425H2 while PVO gained sensitivity (Fig 2B). The largest effect of a single CDRH swap was observed for the CDRH3 which differed by six amino acids. CAP256.0425H3 neutralized six additional viruses (ZM109, PVO, ZM214, ZM249, Q461 and Q842) (Fig 2B) although with similar potency to CAP256.04 (0.39 μg/mL and 0.17 μg/mL, respectively). This was confirmed in the reverse experiment, where the CAP256.04 CDRH3 was introduced into CAP256.25 (CAP256.2504H3) resulting in reduced breadth (Fig 2B). Overall, these data suggest that the breadth of the CAP256.04 off-track phenotype is modulated by the CDRH3, but potency is attenuated by a combination of all three CDRHs.
As the CAP256.25 CDRH3 conferred additional breadth into CAP256.04, we assessed which residue(s) were responsible for this effect (Fig 2A). Mutating CDRH3 residues that altered the charge (K102N) or aromaticity (F100ddA) of the CDRH3 failed to improve the neutralization of CAP256.04 (S1 Fig). In contrast, the R100rA mutation, which removed a positive charge from the CDRH3, improved the breadth of CAP256.04 from 50% to 94% of this panel, with seven viruses resistant to CAP256.04 becoming sensitive to CAP256.04 R100rA (Fig 2B). Interestingly, this single mutation recapitulated the breadth introduced into CAP256.04 by the entire CAP256.25 CDHR3. The reverse mutant, CAP256.25 A100rR similarly resulted in a substantial loss of breadth (S1 Fig). The combination of R100rA with the CAP256.25 CDRH1 did not confer additional breadth beyond that introduced by R100rA (Fig 2B) and, surprisingly, the addition of the CDRH2 to either CAP256.04 R100rA or CAP256.0425H1 R100rA reduced the breadth of these antibodies (Fig 2B).
To model the interactions between V1V2 and CAP256.04 or CAP256.25, we fitted the sequence of CAP256.04 to the crystal structure of CAP256.25 in complex with the CAP256-34 week trimer which shares key contact residues for the CAP256-VRC26 lineage, including N160, R166, D167, K168 and K169 with CAP256.15-wk SU (Fig 3A), and which likely triggered the lineage [22]. The model showed that the N130 glycan projects from the trimer and is situated proximal to the CDRH1. The CAP256.25 R28 residue was predicted to hydrogen bond with the N130 glycan while CAP256.04 S28 was not predicted to interact with this glycan. Surprisingly, the introduction of S28R to CAP256.04 R100rA and CAP256.0425H3 (S1 Fig) decreased the breadth of these antibodies by five and four viruses, respectively, although the potency remained unchanged. Together with R100rA, the combination of S28R and N30D increased the breadth compared to S28R alone, but two viruses (PVO.04 and AC10.29) remained resistant to neutralization (S1 Fig). The conformation of the CDRH3 is partly supported by the intermolecular interactions between this loop and the CDRH1 and CDRH2 [25,26]. Consequently, the interaction between R28 and the CDRH3 may shift the CDRH1 toward the N130 glycan, which may be countered by the N30 residue (Fig 3A, top panel). Furthermore, the model showed that 100r was positioned in close proximity to the viral residue K168, and that K168 was hydrogen bonded to the CDRH3 residue E100c (Fig 3A, bottom panel). Therefore, due to electrostatic repulsion, R100r likely displaces K168, disrupting the interaction between K168 and E100c. In contrast, CAP256.25 has a small, uncharged Ala at position 100r which would not obstruct the adjacent bond. This suggests that an R100rA mutation may reduce electrostatic repulsion and therefore increase antigen affinity, providing a mechanism for the enhanced breadth associated with this mutation.
To estimate the probability of the occurrence of mutations prior to selection that are associated with the off-track phenotype, we utilized the ARMADiLLO software [34]. We first assessed the probability of the CDRH1 T28 residue (found in the CAP256.UCA) mutating to either 28R (as in CAP256.25) or 28S (in CAP256.04, Fig 2A). We found that the CAP256.25 T28R mutation was improbable (0.41%), requiring two base changes (ACC to AGG), whereas T28S (in CAP256.04) was ten times more probable (5.1%), requiring only one change (ACC to TCC, Fig 3B). Similarly, two base changes were necessary for the improbable (1.1%) mutation S30D (in CAP256.25), whereas S30N (in CAP256.04) was a probable (14.7%) event as it occurred in an AID hot-spot. The combination of T28R and S30D (in CAP256.25), which was necessary for neutralization when transplanted into CAP256.04, was a highly improbable event (0.004%, Fig 3B). Furthermore, the CDRH3 Q100rA mutation, that confers breadth to CAP256.25, was predicted to occur very rarely with a 0.019% probability, requiring two base changes (CAA to GCA), one of which was in an AID cold-spot (Fig 3B). In contrast, we found that Q100rR that results in the off-track phenotype was predicted to occur with a much higher probability of 4.7%, requiring only a single nucleotide change that was not located within an AID cold-spot (Fig 3B). Furthermore, the improbability of the evolution of T28R, N30D and Q100rA was reflected in next-generation sequences of the CAP256 lineage [22,43], where T28R, S30D and Q100rA together accounted for <2% of the sequences. Altogether, this suggests that a single relatively probable mutation in the CDRH3 primarily limits the neutralization breadth of CAP256.04.
Next, we sought to determine which sites were responsible for the off-track phenotype of CAP256.20 compared to the related bNAb, CAP256.27 (Fig 4A). Of the 6 amino acids that distinguish these CDRH3s, several were charge changes, with an overall CDRH3 charge of -7 and -3 in CAP256.27 and CAP256.20, respectively. First, we transplanted the CDRH3 from CAP256.27 into CAP256.20 (CAP256.2027H3) and tested this chimera for neutralization breadth. In contrast to the CAP256.04/.25 pair, transfer of only the CAP256.27 CDRH3 into CAP256.20 introduced both bNAb-like breadth and potency into the off-track antibody, increasing breadth from 1/16 to 14/16 viruses and potency from 2.92 to 0.05 μg/mL (Fig 4B). In the reverse experiment, replacing the CDRH3 of CAP256.27 with that of CAP256.20 (CAP256.2720H3) almost completely abrogated neutralization (Fig 4B).
To determine which of the six residues within the CDRH3 of 256.20 limited breadth, we focused on three residues which altered the electronegativity of the loop, R100d, H100j and V100n (Fig 4A), and mutated these sites alone and in combination. The V100nE mutation, which slightly decreased the charge of the CDRH3 only marginally improved neutralization breadth by one additional virus (the autologous superinfecting virus, CAP256.15-wk SU) (Fig 4B). R100dW also decreased the charge and increased breadth by an additional three viruses (CAP256.15-wk SU, CAP210 and ZM197). Similarly, introduction of an H100jD mutation conferred neutralization of the same three additional viruses but with greater potency than R100dW (geometric mean potency of 0.04 compared to 0.33 μg/mL, respectively). The combination of R100dW and H100jD increased potency and breadth to a total of 12/16 viruses, while V100nE, R100dW and H100jD together introduced the same bNAb-like broad neutralization and potency as the entire CDRH3 transplant (Fig 4B). In the reverse experiment, mutating positions 100d, 100j and 100n in the CDRH3 of CAP256.27 to match the sequence of CAP256.20 knocked out neutralization against previously sensitive viruses (S2 Fig). Therefore, of the 17 residues that differentiate the HC of CAP256.27 and CAP256.20 only three amino acids, all within the CDRH3, are responsible for limiting the breadth of CAP256.20.
Comparison of the probabilities of the breadth-conferring and restricting mutations between CAP256.27 and the CAP256.UCA showed that the G100nE bNAb mutation was improbable (0.89%), since two base changes were required (Fig 5A). In contrast, the W100dR, D100jH and G100nV mutations that took CAP256.20 off-track are highly probable (10%, 14% and 3%, respectively), as only single base changes were necessary, with the H100j mutation in an AID hot-spot (Fig 5A). The probability of retaining all three mutations that confer breadth (W100d, D100j and G100n) in the absence of selection is 0.33%. Overall, the relatively higher probability of these mutations which restrict breadth highlights the potential ease with which members of bNAb lineages may mature along undesirable pathways.
To explore the basis of these results, we fitted the amino acid sequences of CAP256.20 and CAP256.27 HCs to the crystal structure of CAP256.25 HC in Swiss-PdbViewer (v4.1.0), and then aligned this to a CAP256-34 week trimer structure (Fig 5B, [26]). The YYD motif of CAP256.27 was predicted to hydrogen bond with two K166 residues on two protomers and R166 and K169 on the remaining protomer (Fig 5B, top left). The mutations in CAP256.20 likely disrupt these interactions by placing a positive charge (H100j) proximal to K169 and possibly preventing sulfation of the preceding Tyr [44], which would knock out contacts with K166 (Fig 5B, top right).
Furthermore, the hydrogen bond between E100n, present in CAP256.27, and the N160 glycan was disrupted by the CAP256.20 V100n substitution (Fig 5B, bottom left). In addition, the CAP256.20 R100d mutation, places a positive charge next to K169 resulting in electrostatic repulsion (Fig 5B, bottom right) and possibly preventing the backbone interactions between these residues. Therefore, key interactions between the highly electronegative CDRH3 of CAP256.27 (-7) and the electropositive V2 epitope are abrogated by the more electropositive CDRH3 of CAP256.20 (-3), resulting in the off-track phenotype.
We were interested in determining which viral immunotypes (or epitope amino acid variants) drove CAP256.20 away from breadth. The K169 immunotype, which forms part of the CAP256 epitope, is fairly conserved within subtype C viruses (65.9%) (www.hiv.lanl.gov) (Fig 6A) [12]. In contrast, the 169Q immunotype is present in only 6.5% of subtype C viruses (Fig 6A) but dominated the viral population in donor CAP256 across 21 of 28 time points from six to 206 weeks post infection (Fig 6B and S3A Fig). We hypothesized that this globally rare immunotype, which predominated in CAP256, contributed to the evolution of CAP256.20. Introduction of 169Q mutations into eleven heterologous viruses slightly improved the neutralization of four viruses by CAP256.20 (S3B Fig), suggesting a preference of CAP256.20 for 169Q, though other viral determinants clearly contribute to neutralization sensitivity/resistance.
As autologous viruses, rather than heterologous viruses, select mutations during SHM, we determined the sensitivity of autologous viruses containing either a 169Q or 169K immunotype to neutralization by CAP256.20 and CAP256.27. We identified two viruses that naturally contained a 169Q (CAP256.42-wk 5 and CAP256.48-wk 10), and were sensitive to CAP256.20 and CAP256.2720H3. Introduction of Q169K mutations into these viruses knocked out neutralization by these antibodies, and their matched CDRH3 chimeras (Fig 6C). In the reverse experiment, introduction of a 169Q into three autologous viruses (CAP256.34-wk 80, CAP256.59-wk 10b and CAP256.15-wk SU, containing 169K/I and resistant to CAP256.20 and CAP256.2720H3) resulted in increased sensitivity to these antibodies (Fig 6C). In all five viruses, the 169K immunotype was associated with increased neutralization potency by CAP256.27, CAP256.2027H3 and the additional broad CAP256.20 mutants and chimeras (S3C Fig), consistent with previous studies [11,12]. These data indicate a CDRH3-mediated preference of CAP256.20 for the 169Q immunotype, which limited the evolution of breadth in this antibody.
A major focus of HIV vaccine design is based on a deep understanding of the development of bNAbs during HIV infection. Studies of HIV/antibody co-evolution have provided a template for the maturation of breadth that is now the basis of B-cell lineage vaccine strategies [12,17–23]. However, much less is known about the maturation of antibodies within bNAb-containing lineages that fail to acquire breadth. We have previously shown that these include both “dead-end” antibodies (that fail to acquire breadth and exhibit low SHM), and “off-track” antibodies that acquire substantial SHM, but little breadth, and are the focus of this study. Here we used previously identified strain-specific “off-track” antibodies that are closely related to broad members of the CAP256-VRC26 bNAb lineage to probe the genetic and viral contributors to their maturation [22]. In two pairs, we identified key breadth-restricting mutations, and defined the probabilities of their occurrence. Furthermore, we show the preferential neutralization of a globally rare immunotype by CAP256.20 impeded the evolution of breadth, providing a mechanism for the development of off-track responses in infection. Together these data provide insights into the challenges associated with driving maturation of antibodies towards breadth, a central question in HIV vaccine design.
Studies of bNAb lineages have highlighted the substantial plasticity in their maturation, enabling the development of breadth through multiple pathways [11,12,21,43]. This suggests that immunization strategies promoting high levels of SHM along diverse pathways may enhance bNAb development. Our data suggests that similarly, the development of off-track antibodies can occur by multiple pathways, and through the acquisition of very few mutations. Indeed, within the CAP256.20/27 pair, only three mutations were sufficient to divert CAP256.20 away from both breadth and potency. Furthermore, the enhanced breadth of CAP256.04 R100rA, which contains a single CDRH3 mutation, in contrast to the six mutations in CAP256.0425H3, suggests substantial mutational “noise”. Additionally, we found an increase in breadth when the CAP256.25 CDRH2 or single CDRH1 mutations were introduced into CAP256.04. These data confirm that extensive SHM does not necessarily result in breadth and suggests that many mutations do not impact or negatively affect breadth.
The maturation of bNAbs towards breadth includes a requirement for functionally relevant but improbable mutations [34]. Consistent with previous studies, we find that the key breadth-conferring mutations in the CAP256-VRC26 lineage, such as Q100rA and G100nE, are highly improbable [34] and that mutations that restricted breadth were in many cases relatively probable. Interestingly, the two pairs of CAP256 antibodies provide distinct examples of how the probability of mutations could shape on- versus off-track maturation. In CAP256.20/27, the potential for high probability breadth-limiting mutations (e.g. W100dR, D100jH and G100nV) is evident in pulling the B-cell off-track. In contrast, in the CAP256.04/25 pair, an important improbable mutation associated with breadth, Q100rA, represents a potential bottleneck that CAP256.25 has overcome to achieve breadth, but CAP256.04 has not. The more probable R100r mutation is also present in broader lineage members, suggesting compensatory mechanisms and that multiple pathways to breadth exist. Together these data indicate that, in some instances, the pathway to the off-track phenotype offers less “resistance” compared to the requirement for highly improbable mutations associated with breadth, and will require careful selection of immunogens to avoid this phenotype [34].
Several studies, including our previous work in the CAP256-VRC26 lineage, have shown how exposure to diverse viral variants contributes to the development of breadth [11,12,22]. This study extends this work to define the mechanisms that select the off-track phenotype within a single lineage. The enrichment of the globally rare 169Q immunotype in CAP256 and the preferential neutralization of this immunotype by CAP256.20, implicates 169Q autologous viral variants in the evolution of this off-track antibody. Early CAP256-VRC26 lineage members (CAP256.01 and CAP256.24) and other off-track antibodies (CAP256.12 and CAP256.13) are unable to neutralize 169Q viruses. However, most lineage members, especially those with breadth, are able to neutralize this immunotype, though to a lesser extent than K169 viruses. This indicates that whereas most lineage members tolerate 169Q, CAP256.20 is uniquely reliant on the globally rare Q169 immunotype, explaining the strain-specificity of this antibody. Structurally, this may be a consequence of the reduced dependence of bNAbs on specific side chain residues within the epitope, compared to an increased dependence of off-track antibodies such as CAP256.20 on such side chains. These data highlight the fact that evolution towards breadth, a desirable outcome from a vaccinology perspective, is distinct from antibody maturation to counter circulating autologous viral variants.
This study provides evidence that affinity maturation to counter globally rare viral immunotypes can drive antibodies within a broad lineage away from breadth. As with the maturation of breadth, off-track antibodies can develop through multiple evolutionary pathways. Furthermore, limited breadth despite high levels of SHM can occur by the introduction of few, but relatively probable, mutations. The inherently stochastic nature of affinity maturation may make avoiding off-track antibodies challenging. Furthermore, additional research is needed to determine if the expansion of ‘off-track’ B-cell lineages prevents or limits the maturation of bNAbs. However, defining pathways towards and away from breadth will facilitate the selection of immunogens that elicit bNAbs and minimize off-track antibodies. Our data suggests that by selecting sequential immunogens that present globally conserved epitopes, the elicitation of off-track antibodies can be minimized. As immunization strategies improve to allow targeting of specific antibody residues, these data inform the design of immunogens to elicit V2-directed bNAbs.
CAP256 is a participant enrolled in the CAPRISA 002 Acute Infection study, established in 2004 in Kwa-Zulu Natal, South Africa. The CAPRISA 002 Acute Infection study was reviewed and approved by the research ethics committees of the University of KwaZulu-Natal (E013/04), the University of Cape Town (025/2004) and the University of the Witwatersrand (MM040202). CAP256, an adult, provided written informed consent. The specificity of her plasma has been described, monoclonal antibodies isolated and autologous Env evolution characterized [11,12,22,45,46].
Exchanging the CDRH3s between monoclonal antibodies was achieved with a two-step overlapping PCR. The CDRH3s were amplified with the AccuPrime Pfx DNA Polymerase and reaction mix (Thermo) with primers that were complementary to the recipient antibody. The product was gel extracted (1%, 1x TAE) and purified with the QIAquick Gel Extraction Kit (Qiagen) and used as the mutagenesis primer for the second PCR with the QuikChange Lightning Multi Site-Directed Mutagenesis Kit (Stratagene). Genes with CDRH1, CDRH2 and CDRH3 exchanges were synthesized (GenScript) and excised with Age1 and Sal1 (Thermo) according to the manufacture’s recommendation. Heavy chain fragments were separated in an agarose gel (1% 1x TAE) and ligated (Roche) into the CMVR expression vector as per the manufacturer’s protocol. Mutations in both antibody and viral Env genes were introduced by site-directed mutagenesis with the QuikChange Lightening Multi Kit (Stratagene) per the manufacturer’s instructions and plasmid sequences were confirmed by Sanger sequencing with the ABI PRISM Big Dye Terminator Cycle Sequencing Ready Reaction kit (Applied Biosystems, Foster City, CA) and resolved on the 3500 genetic analyzer.
The TZM-bl cell-line was obtained from the AIDS Research and Reference Reagent Program and the 293T cell-line was obtained from Dr. George Shaw (University of Pennsylvania, Philadelphia, PA). The cell-lines were cultured at 37°C (5% CO2) in Dulbecco’s Modified Eagle Medium (DMEM) supplemented with 10% heat-inactivated foetal bovine serum (FBS), 50 μg/mL gentamicin (Sigma) and 25 mM HEPES, at confluency monolayers were disrupted with 0.25% trypsin and 1 mM EDTA (Sigma). The 293F cell-line (Life Technologies) was maintained in serum and antibiotic free FreeStyle 293 Expression media at 37°C (10% CO2) in an orbital shaker (130 rpm).
Cloned Envs and pSG3ΔEnv backbone plasmids (NIH AIDS Research and Reference Reagent Program) were co-transfected into 293T cells with 1:3 PEI MAX transfection reagent (Polysciences). Following 48 hours, filtered pseudovirus supernatants were adjusted to 20% FBS and stored at -80°C.
Equal quantities of antibody heavy and light chain plasmids were co-transfected into 293F cells using PEI MAX. Monoclonal antibodies were purified from cell-free supernatants after six days using protein A affinity chromatography. Concentration and buffer exchange (1x PBS) was completed with Vivaspin concentrators (Sartorius).
Neutralization was measured as previously described by a reduction in luciferase gene expression after single-round infection of TZM-bl cells with Env-pseudotyped viruses [47,48]. Titers were calculated as the reciprocal antibody dilution (IC50) causing 50% reduction of relative light units (RLU).
The crystal structures of CAP256.04 (PDB: 4ORG) was aligned to the crystal structure of CAP256.25 and CAP256-34 week trimer in PyMOL 2.0.2. No structures were available for CAP256.20 and CAP256.27, therefore the amino acid sequences of these antibodies were fitted to the crystal structure of CAP256.25 in Swiss-PdbViewer 4.1.0 and aligned in PyMOL. Protein model representations were created with Swiss-PdbViewer v4.1.0 and PyMOL v2.0.2. Graphs were created in GraphPad Prism 6, sequence alignments were made in BioEdit v7.2.5 and phylogenetic trees were constructed in MEGA v6.06.
The Antigen Receptor Mutation Analyzer for Detection of Low Likelihood Occurrences (ARMADiLLO) program estimates the probability of amino acid substitutions prior to antigenic selection [34]. Briefly, 104 simulated mature sequences per site of interest are generated with the S5F mutability and substitution models through comparison of the nucleotides sequences of the CAP256.UCA and a given mature lineage member [33]. The probability of particular amino acid substitutions is then estimated by the frequency of the observed site in the set of simulated sequences. Here, a mutation of <2% probability is classified as “improbable” as this reflects a frequency ≤2 B-cells per germinal centre harbouring that particular mutation [34]. Probabilities of combinations of mutations are derived from the products of individual probabilities.
|
10.1371/journal.ppat.1003608 | Multicellular Bacteria Deploy the Type VI Secretion System to Preemptively Strike Neighboring Cells | The Type VI Secretion System (T6SS) functions in bacteria as a contractile nanomachine that punctures and delivers lethal effectors to a target cell. Virtually nothing is known about the lifestyle or physiology that dictates when bacteria normally produce their T6SS, which prevents a clear understanding of how bacteria benefit from its action in their natural habitat. Proteus mirabilis undergoes a characteristic developmental process to coordinate a multicellular swarming behavior and will discriminate itself from another Proteus isolate during swarming, resulting in a visible boundary termed a Dienes line. Using transposon mutagenesis, we discovered that this recognition phenomenon requires the lethal action of the T6SS. All mutants identified in the genetic screen had insertions within a single 33.5-kb region that encodes a T6SS and cognate Hcp-VrgG-linked effectors. The identified T6SS and primary effector operons were characterized by killing assays, by construction of additional mutants, by complementation, and by examining the activity of the type VI secretion system in real-time using live-cell microscopy on opposing swarms. We show that lethal T6SS-dependent activity occurs when a dominant strain infiltrates deeply beyond the boundary of the two swarms. Using this multicellular model, we found that social recognition in bacteria, underlying killing, and immunity to killing all require cell-cell contact, can be assigned to specific genes, and are dependent on the T6SS. The ability to survive a lethal T6SS attack equates to “recognition”. In contrast to the current model of T6SS being an offensive or defensive weapon our findings support a preemptive mechanism by which an entire population indiscriminately uses the T6SS for contact-dependent delivery of effectors during its cooperative mode of growth.
| Bacterial Type VI Secretion Systems (T6SS) function as contractile nanomachines to puncture target cells and deliver lethal effectors. Little is known about the lifestyle or physiology dictating when bacteria normally express their T6SS. Proteus mirabilis undergoes a characteristic developmental process to coordinate multicellular swarming behavior and discriminates itself from other Proteus isolates during swarming, by a visible boundary termed a Dienes line. We report this phenomenon, first noted in 1946, requires the lethal action of the T6SS, T6SS-dependent effectors, and immunity proteins. T6SS-mediated lethality is unique to morphologically distinct swarmer cells, and when it occurs requires direct contact. Using this multicellular system, we report that the pre-formed T6SS strikes neighboring cells upon cell-cell contact. Our findings support a preemptive mechanism by which an entire population indiscriminately uses the T6SS during a cooperative behavior and that social recognition in bacteria is immunity to lethal T6SS attack.
| The type VI secretion system (T6SS), a recently discovered Gram-negative secretion pathway [1], [2], delivers effectors upon direct contact with a target cell [3], [4] through a contractile puncturing device. Death of the target cell is the primary outcome that follows the delivery of the lethal effectors, which are translocated from the attacker cell cytoplasm into the periplasm of the target cell via a T6S apparatus in a contact-dependent process [5], [6]. It is well known that bacteria release bactericidal agents such as bacteriocins and antibiotics into the extracellular environment as a means to effectively eliminate bacterial competitors [7]. However, many bacteria survive and replicate in their natural habitat using a multicellular life-style that requires direct contact and cooperativity. As a consequence, we would expect bacteria that have developed multicellular behavior to benefit from a mechanism that depends on cell-cell contact to discriminate between one another and eliminate non-self bacteria, potential cheaters, or competitors from the population. In this regard, T6SS activity, like multicellular behavior, requires direct cell-cell contact. Hence, lethal action of the T6SS would provide a specific advantage for bacteria to discriminate, recognize, and kill competitors that would otherwise interfere or benefit from a cooperative behavior. Surprisingly, despite the strict requirement for cell-cell contact to achieve multicellular behavior and direct contact being required to trigger the contractile T6SS, any relationship between the two remains an open question [8].
Proteus mirabilis, a Gram-negative bacterium, undergoes a characteristic developmental process and associated multicellular behavior known as swarming. This unique morphological cycle, differentiating from short (2 µm in length), rod-like vegetative swimmer cells into hyperflagellated, elongated (40 µm in length) swarmer cells, allows Proteus spp. to swarm rapidly and uniformly in multicellular rafts across a surface, resulting in a bull's-eye pattern on an agar surface. In 1946, Louis Dienes observed that actively swarming strains of Proteus have the remarkable ability to form a visible, macroscopic boundary or “Dienes line” at the intersection of bacterial swarms [9], [10]. Formation of the Dienes line occurs only between swarms of different Proteus isolates, demonstrating the ability of one strain of Proteus to distinguish itself from another [9], [10]. More than 200 Dienes types have been reported, a diversity that rivals O-serotypes [11]. Although the Dienes phenomenon has been a practical tool in diagnostic settings to type clinical isolates [12], [13], only limited information has been obtained regarding the biology dictating Dienes line formation with the important exception that it has been shown to be dependent on direct cell-cell contact [14].
A recent study [15] identified a six-gene operon in P. mirabilis BB2000, termed identity of self or the ids genes, and proposed that these genes encode the molecular determinants of self-recognition involved in the Dienes phenomenon. Five of the six ids genes were suggested to mediate self-recognition by an unknown diffusible signal [16], [17]. While the mechanism that governs strain recognition and Dienes line formation in Proteus remains elusive, it was speculated that lack of self-recognition does not result from killing [15], [16]. While these mechanisms are not understood, the largely descriptive studies on social recognition in Proteus uniformly agree that the Dienes recognition phenomenon occurs only between cells that have undergone the morphological and developmental transformation from planktonic cell to highly differentiated swarm cells that are exhibiting multicellular cooperative behavior [18], [19]. Whether in broth or on a surface non-permissive for swarming, undifferentiated Proteus that are capable of, but are not actively exhibiting, cooperative behavior, do not discriminate themselves from other strains or demonstrate social recognition [9], [10], [14].
Using a genetic screen, we discovered that the Dienes recognition phenomenon, first noted more than 65 years ago, requires the action of the T6SS. Here, we demonstrate that the presence of a Dienes line correlates with bacterial killing showing that one P. mirabilis strain will selectively kill a different P. mirabilis strain. A functional T6SS as well as one specific Hcp-VgrG effector-encoding operon among five such operons are required for killing. Thus, our findings elucidate the mechanism that explains social recognition and Dienes line formation. We show that “recognition” is the ability to survive attack from an opposing cell's T6SS. We further investigated the T6SS using our multicellular model system applying the exceptional tools recently developed for studying the T6SS in planktonic cells [20], [21], [22] because the biochemical and genetic studies that have described how bacteria use the T6SS were generally performed under conditions that do not favor natural expression of the specialized secretion system [5], [6], [20], [21], [22], [23], [24]. In contrast to the main conclusion drawn from studying planktonic bacteria, that the T6SS is used for interbacterial competition by functioning solely as either an offensive or defensive weapon [20], our findings provide compelling evidence for a contrarian model to describe how the T6SS is deployed under physiologically relevant conditions during a phase in which differentiated Proteus cells interact with one another to swarm. We propose a preemptive antagonism model for T6S that adequately describes contact-dependent delivery of lethal effectors in the context of a natural developmental process that coordinates multicellular behavior.
It has been recently proposed that the ids genes of P. mirabilis are sufficient for self-recognition and Dienes line formation [16]. However, as a part of this study, we report that ids mutants constructed in the well characterized prototype strain HI4230 have no defect in self or non-self identity (Fig. 1A, white arrows). Non-identical strains, HI4320 and BB2000, form a Dienes line (Fig. 1A, black arrow), while identical strains of HI4320 merge (Fig. 1B, white arrow). We used an approach similar to the one used to identify ids in BB2000 [15], by first identifying a single gene in strain HI4320 involved in the Dienes phenomenon using a transposon screen of P. mirabilis HI4320. Only one of 1920 insertion mutants, identified as 9C1, formed a Dienes line with its wild-type parental strain HI4320 on swarm agar (Fig. 1B, black arrow).
Using our genetic approach, however, we identified additional genes required for the Dienes phenomenon by re-screening the insertion mutants for those that lost the ability to form a Dienes line with mutant 9C1 (Fig. 1C, white arrow). Twelve mutants merged with 9C1, suggesting a role in strain recognition (Fig. 1D–F). Remarkably, 8 of 12 mutants have transposon insertions in genes localized to the same operon (Fig. 1G, red triangles) as the original 9C1 mutant (Fig. 1G, yellow triangle). The four additional mutants that merged with 9C1 (Fig. 1F) have transposon insertions in four of the 17 contiguous genes (Fig. 1G, blue triangles) that encode highly conserved structural components of the T6SS (Table S1) [25]. These genes reside on an adjacent but divergent gene cluster beginning with PMI0749 (COG3516), which encodes a homolog of the Vibrio cholerae VipA T6SS sheath protein (Table S1).
PMI0750 and PMI0751 belong to the hcp (COG3157) and vgrG (COG3501) superfamilies, respectively, and encode proteins that could assemble to form the puncturing device of the T6SS [1]. Located directly downstream and within the same potential operon as PMI0750 and PMI0751 are genes PMI0752–PMI0758. We predict that these genes encode T6SS-dependent effectors because 9 transposon mutants within this putative operon were identified that affected either recognition or killing. Furthermore, we predicted that PMI0754, PMI0755, and PMI0758, which were not identified in the screen, also play a role in T6S-dependent recognition or killing. Indeed, strains with mutations in each of these three genes (Fig. 1G, green triangles), constructed in strain HI4320; 0754::kan, 0755::kan, and 0758::kan, merged with 9C1 (Fig. 1H). Thus, independent disruption of every gene within the identified operon by either transposon insertion or by targeted knock-out abolishes Dienes line formation with the susceptible 9C1 mutant suggesting that the entire hcp-vgrG effector operon, encoded by PMI0750–PMI0758, is required for complete function.
As will be discussed, the observation that the entire hcp-vgrG effector operon is required for Dienes line formation and that recognition ultimately arises from the reaction occurring when two bacterial cells use the T6SS to deliver effectors to one another suggests directionality; two strains attempting to kill one-another. To explore this possibility, we first sought to understand whether the Dienes recognition phenomenon results from killing. We hypothesized that bacteria in the area of the Dienes line would be permeable to the fluorescent DNA stain SYTO 9 due to the loss of membrane integrity and viability resulting from the lethal action of the opposing swarm's T6SS. Indeed, bacteria in the Dienes demarcation can be brightly stained by the vital dye (Fig. S1), which indicates a loss of viability and killing, occurs at the boundary between swarms.
To quantify killing, competition assays were conducted between P. mirabilis HI4320 and mutant 9C1. Wild-type HI4320 killed mutant 9C1 by at least 7-logs when plated together on agar to permit swarming and co-cultured for 18 h (Fig. 1I). Further, we find 106 fewer 9C1 than were co-inoculated with HI4320, which must be explained by killing. Insertion mutants of the hcp-vgrG effector operon and the T6SS merged with mutant 9C1 and were unable to kill 9C1 (Fig. 1D–F; Table 1). Dienes line formation was restored when mutants of the hcp-vgrG effector operon were complemented with a plasmid carrying the entire hcp-vgrG effector locus (Fig. 1J and 1K). As expected, mutant 9C1, when complemented with the hcp-vgrG effector operon (9C1+), reverted back to a wild-type Dienes phenotype and was able to kill 9C1 (Fig. 1I and 1J). Complementation analysis confirms that the hcp-vgrG operon encodes factors required for Dienes line formation and for killing or for immunity to killing (i.e., recognition). The remarkable conservation between the newly identified P. mirabilis T6SS and the best-characterized T6SSs from V. cholerae (Fig. 1L) and P. aeruginosa (Fig. 1M) strongly suggests that the T6SS could function to deliver lethal effectors to target cells during interbacterial competition in P. mirabilis.
Despite killing of one strain by another, the macroscopic Dienes line is fixed at the midpoint between swarms. To observe the events immediately following the merger of two swarms, where it is expected that direct contact and T6S-dependent interactions occur, we performed live-cell microscopy. Cultures of P. mirabilis HI4230 producing green fluorescent protein (GFP) and cultures of mutant 9C1 producing dsRed were spotted on opposite sides of a swarm agar plate and were observed in real-time as the bacterial strains swarmed toward one another (Movie S1). The strains merge and seamlessly interdigitate as cells from opposing swarms first come into direct cell-cell contact (Movie S1).
To obtain greater resolution and specificity of the lethal T6S-dependent interaction and visualize the aggressor during associated killing, we constructed a strain containing a fusion between the gene encoding HI4230 VipA (PMI0749) and the gene for super folder green fluorescent protein (sfGFP) in a fashion similar to methods used in V. cholerae studies [22]. Seminal work by the Mekalanos Laboratory has rapidly extended understanding of the biomechanics and timing for assembly, disassembly, and contraction of the T6SS nanomachine in both V. cholerae and P. aeruginosa [1], [2], [20], [22]. These pioneering studies have validated a number of tools that can be applied to any T6SS of interest, given the high degree of conservation of core T6SS genes across a wide-range of bacterial taxa. Specifically, it is now established that cycles of T6SS assembly and contraction, or T6SS activity, can be directly visualized using fluorescent protein fusions to orthologs of the V. cholerae T6SS sheath protein VipA [20]. Fusion of sfGFP to the T6SS sheath protein (VipA) encoded by PMI0749 causes a punctate pattern of fluorescence emitted from sfGFP (Fig. S2) in HI4320 during swarming and infiltration of 9C1 expressing dsRED (Fig. S2). This is in contrast to the diffuse fluorescence emitted when sfGFP is independently expressed (Fig. S2). Actively swarming HI4230 cells expressing VipA::sfGFP were removed from the agar plate and individual swarmer cells were observed to emit punctate green staining over the length of the cell (Fig. 2A). The green staining was not diffuse throughout the swarm cell nor observed at only one location (Fig. 2A), suggesting that VipA is localized in discrete regions along the length of the elongated cell and the T6SS is assembled at numerous sites (Fig. 2A). In contrast to V. cholerae (2–3 µM in length) with a single site of T6SS assembly [20], [22], the much longer wild-type swarmer cell of P. mirabilis (∼40 µM in length) has many more T6SS assembly sites; approximately 12 areas have intense green staining (Fig. 2A).
Examination of the T6S-dependent interaction between 9C1 expressing dsRed and wild-type HI4230 expressing VipA::sfGFP revealed a one-sided invasion by the dominant HI4320 (Fig. 2B–D). By 10 min post-merge, massive infiltration of 9C1 dsRed by HI4320 VipA::sfGFP (Fig. 2C) was observed. In contrast, 9C1 dsRed swarmer cells were rarely observed on the opposite side of the swarm merge points even at this early time post-merge (Fig. 2D). This could reflect the fact that HI4320 efficiently kills 9C1 when their multicellular swarms come into contact. It is important to note that the individual wild-type HI4320 VipA::sfGFP- expressing swarmer cells (green) located within the 9C1 mutant swarm (red) are not co-localized (yellow) but rather appear aligned (green) directly against mutant 9C1 swarmer cells (red) (Fig. 2C). An apparent increase in VipA::sfGFP signal appears in HI4320 at the vicinity with the greatest degree of contact between mutant 9C1 and advancing wild-type swarm (boxed area, Fig. 2D). This was unexpected because the VipA::sfGFP fusion is induced by arabinose, which is uniformly present (10 mM) throughout the plate. We hypothesize that the physical interaction between opposing swarm populations at the edge of the leading swarm front triggers rapid assembly and disassembly of the multiple VipA subunits that compose the T6SS sheath [22] because it has been previously demonstrated that when a cell expresses a fusion of a fluorescent protein to an orthologous VipA, the intensity of green staining (sfGFP) directly correlates to greater T6SS sheath assembly at the area of inter-strain contact that reflects delivery of the cell-puncturing device, encoded by hcp and vgrG, and effector proteins [20], [22].
To examine this infiltration in real-time, we imaged the area approximately 3 mm beyond the merge point on the 9C1 dsRed side where there are no observable swarms of multicellular HI4320 expressing VipA::sfGFP (Movie S2). Only individual wild-type HI4320 swarmer cells were observed deep within the 9C1 swarm population and appeared to maintain the ability to migrate rapidly within the opposing swarm (Fig. 2E and 2F) (Movie S2). To test the hypothesis that the lethal activity of the HI4320 T6SS against mutant 9C1 prevents the susceptible strain from infiltrating the wild-type swarm, we examined the intersection between HI4320 VipA::sfGFP and HI4320 expressing dsRED, and HI4320 VipA::sfGFP and 9C1 expressing dsRED, on the same agar plate (Fig. 3A). While numerous dsRED-expressing HI4320 are readily observed penetrating the HI4320 VipA::sfGFP swarm (yellow boxed area, Fig. 3B), it was not possible to detect dsRED-producing 9C1 within the wild-type swarm (yellow boxed area, Fig. 3C). Close examination of the infiltrating cells in real-time reveals single HI4320 swarm cells, producing the T6SS sheath VipA fused to sfGFP, in direct contact with a 9C1 target cell within the infiltrated swarm (Fig. 3D), demonstrating a role for T6S-dependent activity during multicellular behavior.
To assess operon structure, we performed RT-PCR to identify cDNA representing a transcript that spans the junctions between the ORFs within the putative hcp-vgrG effector operon (Fig. 4A). It was possible to detect and amplify cDNA representing each junction between all of the predicted ORFs from the identified operon when reverse transcriptase was included (+) in the cDNA synthesis reaction (Fig. 4A). The same products were observed when genomic DNA extracted from HI4320 was used as the template (g); however, it was not possible to amplify the same products when reverse transcriptase was excluded from the cDNA synthesis reaction (−) (Fig. 4A). Thus, the hcp-vgrG gene cluster representing genes PMI0750–PMI0758, is co-transcribed as a single operon; however, the possibility that additional promoters are present within the operon cannot be ruled out based on these results.
To gain clues as to how the identified effector operon confers both killing and immune functions, we employed bioinformatic analyses on the predicted proteins encoded by the operon. It is known from studies in other bacteria that Hcp and VgrG, encoded by the first genes within the operon, are structural components of the T6SS that form a hollow tube (Hcp subunits) that could be capped with a puncturing tip (VgrG) [24]. It is inside this hollow tube made from numerous copies of the Hcp monomer, where lethal effectors presumably reside. Despite the remarkable structural similarity between the Hcp-VgrG T6SS effector delivery complex and the bacteriophage T4 gp5–gp27 cell-puncturing device, the possible T6SS puncturing tip, VgrG, lacks a lysozyme domain [24]. The third gene of the primary hcp-vgrG effector operon, PMI0752, has predicted structural homology to T4 gp5 lysozyme-containing domain (Fig. 4B). Having predicted lysozyme activity encoded in a protein separate from VgrG, could represent a similar strategy used by bacteriophage that lack a canonical lysozyme domain within their orthologous gp5–gp27 puncturing device [24], [26]. Transposon insertion into this putative lysozyme-encoding gene results in the loss of 9C1 killing (Fig. 1D and 1I, 17A1), yet 9C1 is not affected by production of the PMI0752 gene product, even when it contained the HI4320 PhoA signal peptide (Fig. S3). This result suggests that the PMI0752-encoded lysozyme fulfills the T6SS cell wall-puncturing function, being required for delivery of another effector rather than a performing a lethal effector function itself.
While PMI0750–PMI0752 appear to encode three proteins that together might comprise the T6SS puncturing device with lysozyme activity to breach the target cell wall, the remaining proteins encoded by the operon have less obvious relationships to other well-characterized proteins. PMI0753 has a modeled structure with homology to proteins that contain an Ig-like β sandwich fold with thiol protease activity (Fig. 4B). It is likely that PMI0754 encodes a protein that is not delivered into the target cell (i.e., not an effector), because it is predicted to contain a signal peptide and would be secreted into the periplasm of the host bacterium via the general secretory pathway. PMI0754 is also predicted to possess 3-oxoacyl-(acyl carrier protein) synthase activity (Fig. 4B) and could serve to acylate an immunity or effector protein to tether the protein within the periplasm, as has been suggested to occur for immunity proteins in P. aeruginosa [6]. PMI0755 encodes a protein of unknown function that contains a predicted α/β TIM barrel with putative structural homology to amylase/proteinase/hydrolase proteins (Fig. 4B) and appears to contain a C-terminal Endo VII/HNH superfamily nuclease domain.
Mutant 9C1 has a transposon insertion in PMI0756 and as a result is susceptible to T6SS-dependent lethal activity delivered by parental wild-type cells. PMI0756 is predicted to have a 7-bladed propeller fold and has a predicted hydrolase domain (Fig. 4B). PMI0757 encodes a predicted α-α superhelical protein that has modeled structural homology to Armadillo (ARM) repeat proteins. It is notable that both 7-bladed propeller proteins and ARM repeat proteins often promote protein-protein interactions in eukaryotic cells [27], [28]. The last gene in the identified hcp-vgrG effector operon is predicted to be a thioredoxin isomerase (Fig. 4B). In bacteria, thioredoxin activity is generally important for proper inter- or intra-molecular disulphide formation among exported proteins. Since the T6SS delivers proteins from the cytoplasm into the periplasm of a target cell, bypassing the originating cell's own periplasm, the putative thioredoxin encoded by PMI0758 could be important for proper folding of effectors upon access to the target cell periplasm. In support of PMI0758 being required for lethal effector function, mutation of the potential thioredoxin (PMI0758::kan) abolishes killing of 9C1 (Fig. 1H).
Since the transposon insertion in PMI0756 (i.e., 9C1) in HI4320 results in 9C1 being susceptible to the lethal action of the wild-type T6SS, it follows that 9C1 has lost an immunity function. However, disruption of PMI0756 also results in the inability of 9C1 to exert lethal action on itself (Fig. 4C, 9C1 pBAD), despite 9C1 being susceptible to T6SS-dependent killing. Complementation of 9C1 with the entire hcp-vgrG operon on a plasmid restores both immunity against wild-type and the ability to kill and form a Dienes line with 9C1 containing an empty vector, demonstrating that there are no mutations outside of the hcp-vgrG operon that are responsible for the phenotype displayed by the 9C1 mutant (Fig. 1I and 1J). To determine the contribution of PMI0756 to immunity and killing, the PMI0756 gene from wild-type HI4320 was cloned into a plasmid containing an arabinose-inducible promoter and introduced into 9C1. In the absence of arabinose, a Dienes line formed between 9C1 pBAD0756 and wild-type HI4320, while no demarcation was observed between 9C1 pBAD0756 and 9C1 containing an empty vector (data not shown). Arabinose induction of PMI0756 in 9C1, however, restores 9C1 immunity to wild-type HI4320 killing as shown by the absence of a Dienes line (Fig. 4D, white arrows), but does not provide 9C1 with the killing function required to form a Dienes line with 9C1 containing empty vector because a Dienes line is not formed between their respective swarms (Fig. 4E).
That PMI0756 restored 9C1 immunity to wild-type HI4320 but not killing of susceptible 9C1, and insertions downstream of PMI0756 [namely 12B5 (PMI0757::tn) and PMI0758::kan] lose the ability to kill 9C1 but remain immune to wild-type HI4320, it is likely that the transposon insertion in PMI0756 in 9C1 affects production of one or both of the downstream gene products that are required for killing (Fig. 4C). Indeed, while PMI0756 is necessary and sufficient to restore 9C1 immunity against parental HI4320 (Fig. 4D–F, white arrows), all three genes encoded by PMI0756, PMI0757, and PMI0758 are required to restore both 9C1 immunity against HI4320 and 9C1 killing of 9C1 lacking PMI0756 (Fig. 4D–G, black arrows). In support of a specific role in killing, loss of PMI0757, PMI0757 and PMI0758, or PMI0758 alone, abrogates 9C1 killing by HI4320 (Fig. 4C) because mutant 12B5 can be complemented to form a demarcation with 9C1 only when PMI0757 and PMI0758 are both present (Fig. 4H and 4I). PMI0758::kan also retains immunity to wild-type HI4320 (Fig. 4J, 0758::kan pBAD) but requires induction of PMI0758 to restore killing and Dienes line formation with 9C1 (Fig. 4K, black arrows).
It is notable that insertions within the identified hcp-vgrG effector operon, in any gene upstream or downstream of PMI0756, abrogates the ability to kill and form a Dienes line with 9C1 but does not cause loss of immunity to wild-type (Fig. 1D–F) (Fig. S4), suggesting that: 1) each of the gene products from PMI0750 (hcp) through PMI0755, in addition to PMI0757 and PMI0758, provide T6SS-dependent killing function; 2) that PMI0756 is required for both immunity and killing; and 3) that PMI0756 is the sole gene product encoded by this hcp-vgrG locus that is necessary for immunity against the lethal T6SS-dependent activity of the putative effector operon.
While it was possible to isolate mutants unable to kill an immune-deficient mutant (9C1) and thus, identify T6S-dependent effectors, it is not possible to understand the biology that dictates use of T6S in a natural system since once immunity is restored, the deficient strain will not kill the parent strain, as they are otherwise isogenic. We presume the natural Dienes phenomenon, which we have now shown requires the lethal action of the T6SS, could result from both strains killing each other (i.e., preemptive antagonism) following the initial signal of direct cell-cell contact. To determine if the interaction observed between wild-type swarms depends on the activity of the T6SS, we disrupted the T6SS (Fig. S5) in another P. mirabilis isolate, strain BB2000 (BB2000 ΔT6) using allelic replacement. When strain HI4320 and strain BB2000 swarms meet, a Dienes line is visible (Fig. 5A) since by definition the Dienes line forms between non-identical multicellular swarms [9], [10]. Interestingly, when one of the opposing strains, either BB2000 or HI4320 is missing the action of the T6SS (BB2000ΔT6 or HI4320ΔT6), the Dienes line is indistinguishable from the one that forms between HI4320 and BB2000 wild type strains that are both competent for T6S-dependent killing (Fig. 5A, black arrows). However, when both HI4320 and BB2000 lack their respective T6SS (BB2000ΔT6 and HI4320ΔT6), the demarcation line does not form and the non-identical swarms merge in a manner identical to HI4320 or BB2000 merging with themselves (Fig. 5A and 5B, white arrows). Thus, the experiments confirm that killing is T6-dependent and swarm cells of both strains are capable of killing each other following direct contact. Even the 9C1 mutant, which is unable to attack using the primary effector operon, appears to use the T6SS with another effector since it makes a demarcation with BB2000 lacking a functioning T6SS (Fig. 5B, dashed arrow). That 9C1 is capable of forming a Dienes line with the T6SS-deficient BB2000 supports indiscriminate firing of the T6SS during swarming and that 9C1 is resistant to self-killing, both are necessary for a preemptive antagonism model for the T6SS.
Due to the discrete organization of the identified effector genes linked to genes that encode Hcp and VgrG structural components (Fig. 4B) into a single transcriptional unit (Fig. 4A) distinct from the conserved T6SS gene cluster (Fig. 1G), we sought to determine if additional orphan effector operons could be identified by scrutinizing the genome for hcp/vgrG gene pairs. The prototype P. mirabilis HI4320 genome contains the ids genes (PMI2990–PMI2996) [15], [17], which we now propose to be an hcp-vgrG orphan effector operon. Three orphan hcp-vgrG operons, in addition to idsA-F, were also found: PMI0207–PMI0212, PMI1117–PMI1121, and PMI1332–PMI1324 (Fig. 6A and B). Examination of the predicted amino acid sequences reveals that the Hcp proteins of the hcp-vgrG effector operons are highly homologous, with the exception of the truncated protein encoded by PMI1332 (Fig. S6). The VgrG proteins have high percent amino acid sequence identity at the N-termini with decreasing homology beginning approximately 100 residues from the C-termini (Fig. S7). All of the hcp-vgrG effector operons have remarkably similar organization, where the first two genes of the effector operon are the hcp-vgrG pair followed by up to 7 genes that encode non-structural T6SS proteins with putative effector and immunity functions (Fig. 6B). Because the ids operon does not influence Dienes line formation in strain HI4320 but does in BB2000 [15], we examined the HI4320 genomic sequences upstream of all hcp-vgrG effector operons for promoter differences that could explain disparate expression. Alignment analyses of the five predicted promoter regions indicate a high degree of similarity (Fig. 6C).
Because the multiple sequence alignments show a high degree of conservation between the identified hcp-vgrG effector operon and the four orphan effector operons, we reasoned that all five hcp promoters could be functional. To test this hypothesis and to observe hcp expression during swarming, transcriptional hcp promoter-luciferase fusions were constructed for each hcp-vgrG effector operon and examined hourly for luminescence during swarming (Fig. 6D). A constitutive sigma-70 em7 promoter (Pem7) driving expression of luxCDABE [29] was included as a control. As the P. mirabilis strains containing the individual hcp promoter-luciferase fusions swarm outward across the swarm agar plate, the outer most edge of the bacterial swarm zone quantitatively emits the brightest light, in contrast to the constitutive em7 construct that shows highest luminescence where cell density is the greatest (Fig. 6E). These findings indicate that actively swarming bacteria at the periphery are expressing the hcp promoter-luciferase fusion at higher levels than bacteria in the preceding swarm rafts (Fig. 6D and 6E). All five hcp genes and associated effector operons are expressed during multicellular swarming. We also found expression of all five reporter constructs occurs during liquid culture (Fig. S8). However, expression of the T6SS-related genes during batch culture is difficult to evaluate since the natural Dienes phenomenon and killing occurs exclusively in cells that have differentiated into swarm cells, neither of which occurs during culture in liquid medium. It is notable, that the constitutive control em7 construct displays a linear relationship between luminescence and cell density as expected both in broth (Fig. S8) and on agar plates (Fig. 6E).
Because all of the HI4320 hcp-vgrG effector operons and the T6SS are expressed during swarming, we sought to determine a link between possession of hcp-vgrG effector operons and Dienes type using a collection of P. mirabilis clinical isolates. We hypothesized that variability in number and type of hcp-vgrG effector operons or the nucleotide variability within the same hcp-vgrG island available in a strain could dictate its ability to kill and overcome an opposing strain. To determine which hcp-vgrG effector operons are present, we performed multiplex PCR on chromosomal DNA of 16 P. mirabilis isolates, including HI4320 and BB2000. We screened for one effector gene localized to each of the hcp-vgrG effector operons and PMI0742 of the T6SS. PMI0742 and PMI0756 were present in all of the P. mirabilis isolates examined (Fig. S9), suggesting that the T6SS-linked PMI0750–PMI0758 effector operon may be the central hcp-vgrG effector operon for T6S-dependent killing. Due to these findings, we have termed the PMI0750–PMI0758 operon the primary hcp-vgrG effector operon (pef).
To determine if the presence of a specific hcp-vgrG effector operon has an effect on Dienes line formation, we examined the 16 P. mirabilis isolates against each other on swarm agar. While most isolates only merged with themselves and formed a Dienes line with all other isolates (red squares; Fig. S9, panel A), we found that some non-identical isolates do not form Dienes lines (green squares; Fig. S9, panel A). For example, P. mirabilis isolates DI120 and DE121 merge together and both are able to merge with KU140 but only DI120 will also merge with RZ130 and ST111, which are also able to merge with each other (Fig. S9). This clearly shows that the Dienes line can determine if two isolates are different. However, lack of a Dienes line does not indicate that two isolates are identical. Because these isolates have the same complement of hcp-vgrG effector operons (Fig. S9, panel C), the presence or absence of a specific hcp-vgrG effector operon does not determine Dienes type and thus dictate what strains will or will not form a Dienes line. We hypothesize that the Dienes phenomenon may involve other non-T6S factors in addition to the number or variability of hcp-vgrG effector operons a strain possesses. For example, previous studies have shown that contact-dependent growth inhibition (CDI) functions in Escherichia coli intercellular competition [30], [31], [32].
First described in 1946 [9], the macroscopic demarcation known as the Dienes line forms at the boundary between swarms of genetically related but non-identical isolates of Proteus mirabilis [10]. This phenomenon has been exploited as a powerful epidemiological tool in clinical microbiology laboratories to track outbreaks in hospital units [12], [13]. Here we show that Dienes line formation and the specificity of this direct cell-cell contact-dependent reaction precisely correlates with killing and this killing is dependent upon the single T6SS and its primary effectors. Upon initiation of swarming differentiation, the T6S apparatus is assembled and appears to fire when opposing swarms meet. Each strain is immune to killing itself. The dominant strain infiltrates deeply beyond the boundary of the two swarms and continues to assemble and discharge the T6SS. Dienes line formation, underlying killing, and immunity to killing, can be assigned to specific contiguous genes. These observations define the mechanism by which P. mirabilis uses contact-dependent delivery of effectors for interbacterial competition during a developmental process that coordinates multicellular behavior. Hence, we sought to use this multicellular model to study T6SS under conditions in which no genetic manipulation was necessary to activate assembly of the T6SS and trigger its firing. We propose that the P. mirabilis T6SS can serve as such a universal model to study this ubiquitous secretion system that mediates interbacterial competition.
That killing of an isogenic mutant is sufficient to produce a line of demarcation supports the notion that the biologically relevant non-isogenic bacterial T6SS interaction can result from strain A killing strain B, strain B killing strain A, or strain A and strain B killing each other. Our findings indicate that when two non-identical multicellular participants make contact, the failure to recognize the opposing cell as self is a result of the lethal action mediated by the effectors delivered by the T6SS. Thus, a lack of killing following cell-cell contact of two swarming populations active for T6SS-dependent killing equates to “recognition” of self. However, a great limitation of using mutant strains that are otherwise isogenic with the wild-type parent strain is that while the wild-type can kill a mutant strain, the complemented mutant will never kill the parent. But, by using non-isogenic, wild-type isolates, we confirmed that T6SS-mediated killing among multicellular bacteria is not unidirectional. That is, wild type swarm cells attempt to kill each other using T6SS-mediated injection of effectors immediately upon cell-cell contact. For example, when the swarms of strains HI4320 and BB2000 meet, a Dienes line forms. When a T6SS deletion mutant of HI4320 is substituted for its wild type strain, a Dienes line still forms. When a T6 deletion mutant of BB2000 is substituted for its wild type, a Dienes line again still forms. Only when HI4320ΔT6 is pitted against BB2000ΔT6, does no line form, indicating that HI4320 can kill BB2000 in a T6SS-dependent manner and BB2000 likewise can kill HI4320 in a T6SS-dependent manner. Only when the T6S systems of both strains are inactivated does the killing stop, preventing Dienes line formation. Thus, in the wild, while one strain may be dominant over another, all swarming strains of P. mirabilis, upon cell-cell contact, are likely preemptively deploying the T6SS to kill competitors, or for that matter, kill any neighboring cell within its own swarm. It is also possible that the benefit gained from preemptive deployment is to preserve a clonal population rather than to benefit by eliminating your competitor.
Interestingly, growth inhibition between colonies on agar medium has also been observed between different isolates of P. aeruginosa, E. coli, and Salmonella spp. [33], [34], [35]. However, this phenomenon, known as colony incompatibility, is thought to be due to bacteriophages, bacteriocins, or through production and excretion of antibiotics [33], [34]. A phenomenon similar to colony inhibition has been noted in E. coli [30], [32], suggesting universality of multicellularity, cell-cell contact, and competition with non-identical strains of the same species [8]. However, until now, an appropriate multicellular system has not been described that would aid in answering the open questions of why bacteria use contact-dependent delivery of effectors to eliminate one's neighbors [8]. Would contact dependence allow discrimination or are T6SS-dependent interactions involved in developmental processes such as multicellularity or organized communities [8]?
Recent work [15], described the ids, or “identify of self” genes, responsible for self-recognition in P. mirabilis BB2000 and proposed that the function of the ids operon could be related to a diffusible bar code scanning mechanism of recognition that precedes the decision to attack [16]. In that work, T6S was ruled out because, while idsA and idsB were noted to have some homology to hcp and vgrG, the authors were unable to identify any genes near the ids locus that encode the structural components of the T6SS [15], [16]. In our study, we identified not only the T6SS genes themselves but also, immediately adjacent, specific genes in the primary hcp-vgrG effector operon as required to kill the immunity-deficient mutant 9C1. Because T6S-dependent killing occurs when two non-identical strains make contact, we hypothesized that the genes downstream of hcp and vgrG encode effectors that are injected by the T6SS into the recipient bacterium's periplasm that ultimately leads to death for the target cell. By interrogating the P. mirabilis genome, we found not only the pef operon linked to the T6SS, but four additional conserved orphan hcp-vgrG effector operons (including the ids operon), which are unlinked to the single T6SS in P. mirabilis. Our finding that ids is an orphan T6SS effector operon, found only in a minority of strains, explains why the genes encoding structural components of the T6SS apparatus were not identified nearby the ids genes [16]. In addition, a limitation of the bar code scanner or self-recognition model [16] proposed for the Dienes phenomenon is that the mere presence of a line between swarm populations is not sufficient to identify the dominant or susceptible strain.
While it is unclear that the specific number or combination of effector operons controls whether or not a multicellular population of P. mirabilis forms a Dienes line with another swarm, it is clear that the Dienes reaction requires the T6SS to be active in only one of the two populations. Therefore, we speculate that certain strains may possess immunity to specific Hcp-VgrG effector operons while lacking the ability to kill with the respective operon. While we have shown that HI4320 and BB2000 are capable of killing one another and, thus, HI4320 is not immune to BB2000, the existence of multiple idsE homologs has been proposed to provide HI4320 immunity against BB2000 [15]. Yet, it was not reported whether BB2000 or other strains that are not fully sequenced and annotated, like HI4320, also possess multiple idsE copies. The existence of multiple “orphan” hcp-vgrG effector operons within a bacterial genome is not uncommon [36]. One study [1] discovered three IcmF-associated (TIGR0334) homologous protein (IAHP) loci in the Pseudomonas aeruginosa PAO1 genome that encode hcp and vgrG homologs, one of which has been proposed to be involved in virulence [1]. The organization of orphan T6SS effector operons with cognate hcp and vgrG genes that are not encoded nearby the T6SS genes has been observed in a number of bacterial genomes for which sequences are available [25], [36]; our findings of four orphan T6SS effector operons in a single genome suggests that novel lethal effector and immunity genes may be readily identified in other bacteria, linked with their respective hcp-vgrG pair. Further, our findings demonstrate that, despite high (>90%) amino acid sequence identity between the multiple Hcp and VgrG proteins in HI4320, the orphan pairs are unable to substitute for the primary Hcp and VgrG to provide the ability to kill the immune-deficient 9C1 strain.
Immunity to self-killing (i.e., “recognition”) strongly supports the preemptive model for T6SS-dependent antagonism and also explains why isogenic wild-type strains do not normally kill one another when in cell-cell contact during multicellular cooperation of specialized swarm cells. Our findings clearly show that it is loss of immunity that allows an otherwise isogenic mutant strain to be killed, however, by definition, isogenic wild type strains do not kill themselves because they would only exist in the absence of the multicellular behavior that has evolved to use the T6SS to gain a specific advantage. Indeed, it is difficult to explain how loss of immunity in one isogenic pair is sufficient to observe naturally expressed T6SS lethality when two strains are mixed and co-inoculated unless the secretion system is preemptively armed to fire upon direct cell-cell contact.
Our data support the idea that one gene from the primary hcp-vgrG effector operon, pefE (PMI0756), encodes a dual function protein that is necessary to confer immunity to preemptive attack by cognate T6SS-delivered effectors. This protein is also required for the lethal activity of the respective primary operon-encoded effectors. Indeed, we have shown that PefE alone can confer immunity to killing by the primary Hcp-VgrG delivered effectors, while pefE and the entire pef operon are required for T6SS-mediated lethality. Future studies will reveal if each of the T6SS-dependent pef gene products delivered by the T6SS function independently as cognate pairs [37] or perhaps assemble in the target cell to form a multi-protein complex or functional effector “package”. We speculate the dual function of PefE is a failsafe mechanism that has evolved to prevent self-killing in the event that loss of immunity occurs. It is notable that insertions within the identified hcp-vgrG pef effector operon, in any gene upstream or downstream of PMI0756, abrogates the ability to kill and form a Dienes line with 9C1 but does not cause loss of immunity to wild-type, which suggests that there is a second promoter driving expression of the failsafe pefE.
We cannot rule out the possibility that pefE encodes a protein that provides an immune function by binding a single cognate effector. It is possible that the pef ‘effectors’ work sequentially, and PefE inactivates a single toxic protein required for antagonistic effects. It could also be possible that a complex representing multiple pef gene products assembles in the target cell, and PefE interferes with the formation of the complex. Alternatively, the mechanism for immunity could also be to block or prevent the preemptive T6SS strike. The dual-function of PefE (that is, immunity and killing) suggests it may also act as a chaperone or otherwise interact with another protein to provide killing function for the attacker cell.
Interestingly, we have been unable to identify toxin-immunity pairs within the HI4320 genome orthologous to those effectors currently described in P. aeruginosa [6]. It is possible that the currently described effector-immunity pairs from P. aeruginosa may be accessory effectors as the Pseudomonads have large genomes to support a generalist lifestyle, unlike Proteus with a more compact genome. However, our findings suggest that T6S-mediated interbacterial competition is fundamentally similar between these microorganisms. In support of this, we have identified a PefE orthologous protein encoded within the HSI-1 T6SS locus of P. aeruginosa. Since PefE appears to provide an immune function in Proteus, and the ortholog is encoded within a T6SS locus, we reason it may perform a similar function when P. aeruginosa naturally expresses its T6SS for antagonizing a natural competitor. Furthermore, the fundamental similarity between T6SS's could also be represented by lysozyme-like effectors; PMI0752 (pefA) encodes a putative bacteriophage-like gp5 lysozyme domain. Bacteriophage, including T4 that is structurally related to the T6SS, requires lysozyme activity to breach the cell wall of the target bacterial cell and to inject their DNA, yet it remains unresolved how the T6SS puncturing device would deliver lethal effectors without lysozyme activity. Since the T6SS appears to target bacterial cells, we hypothesize that the T6SS, like bacteriophage puncturing devices, also requires lysozyme activity to deliver a lethal payload. Consistent with the T6SS requiring lysozyme activity, both P. aeruginosa and V. cholerae have been shown to encode T6SS effectors with lysozyme activity [6], [23].
Our model has provided evidence to answer a number of open questions regarding the role for bacterial T6SS during competition, in general, and its potential relationship to multicellularity [8]. Studying T6S in a multicellular model has overcome a number of limitations that arise from studying the T6SS in bacteria under conditions where the bacteria being studied do not normally express the T6SS, do not require cell-cell contact, or the natural competitor is not known. For example, it has been shown that in V. cholerae, a planktonic bacterium, the global regulator RpoN positively regulates T6SS-dependent hcp and vgrG expression, but has no effect on expression of the core, cognate T6SS structural genes, including vipA [38]. Therefore, forced expression of T6SS activity may incompletely activate T6S-dependent processes and make interpretations of results difficult. Furthermore, if the bacterial physiology dictates repression of T6SS activity, then it would be difficult to distinguish true immunity from the absence of an effector target that might occur due to incomplete or forced expression of T6SS activity during times when the bacteria do not normally express that function. Forced expression of T6SS activity could also lead to recognition [15], [16] and decision-based models [20] for T6S function due to incomplete activation in one or more bacterial strains.
Several lines of evidence in the present study point toward a preemptive model for T6SS-mediated antagonism that is intimately linked to a cellular differentiation pathway in bacteria that culminates in the formation of specialized cells that function to coordinate a cooperative and multicellular behavior. The preemptive model for T6S in bacteria is also consistent with the generally accepted notion that T6SS-dependent activity is cell-cell contact-dependent and deployment of the pre-assembled contractile puncturing device into the target cell is extremely rapid [22]. Because T6S is dependent on cell-cell contact, it would seem beneficial for bacteria exhibiting contact-dependent multicellular behavior to employ the T6SS to discriminate, “recognize”, and kill competitors rather than indiscriminately secrete bactericidal agents when competing for resources in their natural habitats. Indeed, our findings indicate that function of theT6SS and the pef gene products encoded by the primary hcp-vgrG effector operon are coordinately linked to multicellular swarming in P. mirabilis. It is uncertain whether the interdependence of T6SS-dependent killing and multicellularity truly functions for bacterial competition or simply to maintain homogenous population during the development of multicellularity. We hypothesize that during multicellular swarming of P. mirabilis, the T6SS is continuously functioning in a preemptive manner in actively swarming populations to prevent cheaters from benefiting from the cooperative behavior [39], [40]. The T6SS may restrict non-identical populations from swarming within each other's multicellular population. Thus, kin selection may be advantageous to preserve uniform bacterial swarming populations to maximize cooperation during migration to new resources or away from predators in their natural habitat.
P. mirabilis HI4320 was cultured from the urine of a nursing home resident with catheter-associated bacteriuria [41]. A transposon library of 1920 P. mirabilis HI4320 insertion mutants [42] was screened for Dienes line formation on swarm agar [lysogeny broth (LB) medium containing NaCl (10 g/L) and 1.5% agar] by spotting the plate with wild-type HI4320 and mutants (for example see Fig. 1C). For the transposon screens, bacteria were plated by spotting 5 µl of an overnight LB broth culture onto swarm agar, incubated at 37°C, and observed within 18 h. To identify genes containing insertion mutations, arbitrary PCR on genomic DNA was used to amplify the 3′ end of the transposon and flanking chromosomal DNA as previously described [42]. PCR products were cloned into the pCR2.1-TOPO vector (Invitrogen), maintained in E. coli TOP10 (Invitrogen), and sequenced to identify the transposon insertion site. P. mirabilis HI4230 mutants 0754::kan, 0755::kan, and 0758::kan were constructed using the TargeTron kit (Sigma-Aldrich), and T6SS mutants were constructed by allelic replacement. Antibiotics were added as necessary at the following concentrations: kanamycin, 25 µg/ml; ampicillin, 100 µg/ml; and chloramphenicol, 20 µg/ml; for pBAD constructs, 10 mM L-arabinose was added to swarm plates to induce pBAD VipA::sfGFP.
Dienes line formation between P. mirabilis mutant 9C1 and the identified transposon insertion mutants was restored by complementing the insertion mutants with pGEN containing a DNA fragment carrying the T6SS-linked hcp-vgrG effector operon (PMI0750–PMI0758). This 9-kb fragment was PCR amplified with Phusion High-Fidelity Polymerase (Thermo-Fisher Scientific), digested with restriction enzymes SphI and NotI (New England Biolabs), and ligated to the linearized pGEN-MCS vector [43]. The resulting construct was transformed into the P. mirabilis insertion mutants. Individual genes and partial operons were amplified from HI4320 genomic DNA and cloned under control of the arabinose inducible promoter in pBAD-MycHisA (Invitrogen).
A suspension (5 µl) containing a 1∶1 ratio of strains was spotted onto swarming agar. Following incubation overnight at 37°C, the entire swarm was collected from the agar plate, serially diluted, and plated on LB medium containing NaCl (0.5 g/L) and 1.5% agar containing Amp (100 µg/ml) for the parental strain or kanamycin (25 µg/ml) for the 9C1 mutant to determine CFU/ml. The output ratio was compared to the input ratio to quantify killing of mutant 9C1. The cat gene was cloned into linearized pGEN-MCS vector and transformed into the transposon mutants to distinguish these (AmpRCamR) from mutant 9C1 containing pGEN-MCS (AmpR). 9C1 killing is reported as [(CFU of test strain/CFU of 9C1)output]/[(CFU of test strain/CFU of 9C1)input].
To capture live images of Dienes line formation, P. mirabilis HI4230 and the 9C1 mutant were spotted (5 µl) from overnight LB cultures onto opposite sides of an agar plate, allowed to swarm at 37°C, and imaged directly on swarming agar under phase contrast and fluorescence microscopy. For fluorescence studies, P. mirabilis HI4230 was transformed with a plasmid expressing sfGFP. PMI0749, which encodes vipA, was amplified from HI4230 chromosomal DNA; sfGFP was amplified by PCR and VipA was fused with sfGFP, separated by a DNA linker encoding 3×Ala 3×Gly, and cloned into plasmid pBAD-myc-HisA [22]. pGEN encoding Red fluorescent protein (dsRED) [44] was transformed into mutant 9C1. For vital staining, 5 µl of 5 µM SYTO 9 (Invitrogen) was applied directly to bacteria on swarm agar for 5 min prior to visualization. Microscopy experiments were performed in the Center for Live Cell Imaging (CLCI) at the University of Michigan Medical School using an Olympus IX70 inverted microscope with FITC and Texas Red filter sets (Olympus). Images were collected using a CoolSNAP HQ2 14-bit CCD camera (Photometrics). All devices were controlled through Metamorph Premier v6.3 software (Molecular Devices). Initial analysis of the imaging data and the preparation of image overlays, montages and movies, were performed using Metamorph v7.7 software. If necessary, deconvolution of the images was performed using Huygens v4.1 software (Scientific Volume Imaging BV).
To observe expression of the multiple hcp-vgrG effector operons during swarming, a 500-bp fragment immediately upstream of the translational start of hcp was PCR amplified from chromosomal HI4320 DNA with EasyA Polymerase (Stratagene), digested with restriction enzymes PmeI and SnaBI (New England Biolabs), and ligated into pGEN-lux [29] to create transcriptional hcp promoter-luciferase fusions. The constitutive reporter Pem7-lux and a promoterless construct (Pneg-lux) were used as controls [29]. LB cultures were incubated overnight and diluted 1∶100 into fresh LB medium containing ampicillin (100 µg/ml). OD600 of a 200 µl sample volume was read every h as well as the luminescent emission (100 µl sample volume) using a Synergy HT plate-reader operating KC4 software (Bio-Tek, Winooski, VT). Luminescence was plotted as a function of cell density as measured by OD600 over time. For swarming studies with the hcp promoter-luciferase fusions, 5 µl of an overnight bacterial LB culture was spotted in the center of a swarm agar plate. Following incubation at 37°C, luminescent emission was captured using the ChemiDoc XRS system (Bio-Rad). Further analysis was conducted to obtain density plots of the luminescence using the Discovery Series Quantity One software (Bio-Rad).
Nucleotide sequences of P. mirabilis HI4320 genes were obtained from the Kyoto Encyclopedia of Genes and Genomes (KEGG) [45] and saved as SeqBuilder files (DNASTAR). Sequence files were entered into MegAlign (DNASTAR) and subjected to multiple alignment using Clustal W [46]. Alignment data of the predicted promoters of the five hcp-vgrG effector operons and PMI0742 are presented as a phylogenetic tree to view predicted evolutionary relationships.
Sixteen P. mirabilis clinical isolates [47] were examined and recorded for Dienes line formation with one another by spotting a 5 µl volume from a bacterial overnight culture onto swarm agar and incubating at 37°C for 18 h. P. mirabilis HI4320 mutant 9C1, transposon mutant 14H2, which has a disruption in PMI0742 of the T6SS, and mutants of idsB and idsD, PMI2991 and PMI2993, respectively, were also included in these tests as controls.
Multiplex PCR (Qiagen) was performed by amplifying the chromosomal DNA of 16 P. mirabilis isolates for the presence of an effector gene representative of each of the hcp-vgrG effector operons. Primers were designed for P. mirabilis HI4230 genes PMI0210, PMI1120, PMI0756, PMI1329, and PMI2993 of each hcp-vgrG effector operon which resulted in variable sized PCR products; 250-bp, 350-bp, 600-bp, 850-bp, and 1-kb, respectively. PMI0742 of the T6SS was also included and amplified with forward (F) primer, 5′CTCAAGAGCCGGTGATCCATCCTGAAAAAC3′ and reverse (R) primer 5′GTAATTGTCTTGGTGCAGCCGAAAGTG3′ resulting in a 450-bp PCR product. PMI0210 was amplified with F primer 5′TTATTGCTTGGCGAGGCTCTCAGG3′ and R primer 5′GCCAAAGCTAAAGCTCCTCCTAAGCTATG3′, PMI0756; F primer 5′TGCTTAAAACCGAAAGAACAAGGGATGC3′ and R primer 5′CCATTCCAACACTGTAAACGGTAGTC3′, PMI2993; F primer 5′AATTAACGGAACAAATAGTACCAAATCTGC3′ and R primer 5′GCCAAGCCGCTGTGATAACCAAC3′, PMI1120; F primer 5′GCGTCAGCAGGTCTATGAATATAG3′ and R primer 5′CATAACGATAACGGGTGGTTTTTC3′, PMI1329; F primer 5′TATTGTTGTTTGGCGAGGAACGGC3′ and R primer 5′TGAGTGGTCTCCACCACCAGTTAC3′
|
10.1371/journal.pntd.0007146 | The role of TLR9 on Leishmania amazonensis infection and its influence on intranasal LaAg vaccine efficacy | Leishmania (L.) amazonensis is one of the etiological agents of cutaneous leishmaniasis (CL) in Brazil. Currently, there is no vaccine approved for human use against leishmaniasis, although several vaccine preparations are in experimental stages. One of them is Leishvacin, or LaAg, a first-generation vaccine composed of total L. amazonensis antigens that has consistently shown an increase of mouse resistance against CL when administered intranasally (i.n.). Since Toll-like receptor 9 (TLR9) is highly expressed in the nasal mucosa and LaAg is composed of TLR9-binding DNA CpG motifs, in this study we proposed to investigate the role of TLR9 in both L. amazonensis infection and in LaAg vaccine efficacy in C57BL/6 (WT) mice and TLR9-/- mice. First, we evaluated, the infection of macrophages by L. amazonensis in vitro, showing no significant difference between macrophages from WT and TLR9-/- mice in terms of both infection percentage and total number of intracellular amastigotes, as well as NO production. In addition, neutrophils from WT and TLR9-/- mice had similar capacity to produce neutrophil extracellular traps (NETs) in response to L. amazonensis. L. amazonensis did not activate dendritic cells from WT and TLR9-/- mice, analysed by MHCII and CD86 expression. However, in vivo, TLR9-/- mice were slightly more susceptible to L. amazonensis infection than WT mice, presenting a larger lesion and an increased parasite load at the peak of infection and in the chronic phase. The increased TLR9-/- mice susceptibility was accompanied by an increased IgG and IgG1 production; a decrease of IFN-γ in infected tissue, but not IL-4 and IL-10; and a decreased number of IFN-γ producing CD8+ T cells, but not CD4+ T cells in the lesion-draining lymph nodes. Also, TLR9-/- mice could not control parasite growth following i.n. LaAg vaccination unlike the WT mice. This protection failure was associated with a reduction of the hypersensitivity response induced by immunization. The TLR9-/- vaccinated mice failed to respond to antigen stimulation and to produce IFN-γ by lymph node cells. Together, these results suggest that TLR9 contributes to C57BL/6 mouse resistance against L. amazonensis, and that the TLR9-binding LaAg comprising CpG motifs may be important for intranasal vaccine efficacy against CL.
| Leishmaniasis is a major neglected tropical disease, being responsible for more than 20 million deaths per year. The high mortality rate highlights the difficulties and ineffectiveness of the current prophylactic approaches and treatments currently available. Therefore, the development of an effective vaccine would be highly advantageous to circumvent these problems. Despite the many vaccines preparations that have been studied in the last few years, none have shown satisfactory efficacy to be approved for human use. Immune receptors, including the TLR family, are known to be important for host defense during parasitic infections, such as leishmaniasis, and also for vaccine efficacy. In this work, we investigate the role of TLR9 during Leishmania amazonensis infection in vaccinated and non-vaccinated mice. We used a C57BL/6 TLR9-/- mouse model and a first-generation vaccine preparation (LaAg) composed of a total lysate of L. amazonensis. We demonstrate that TLR9 is important for controlling leishmaniasis infection caused by L. amazonensis and is involved in the efficacy of the LaAg vaccine. These findings will certainly help in the development of a better vaccine against leishmaniasis.
| Leishmaniasis is a group of chronic, non-contagious diseases caused by flagellate protozoa of the genus Leishmania [1]. Leishmania (L.) amazonensis is an etiological agent for a broad spectrum of leishmaniases in South American countries [2], including Brazil, where it is a causative agent of localized cutaneous leishmaniasis, diffuse cutaneous leishmaniasis and, rarely, visceral leishmaniasis [2]. Most cases of L. amazonensis infections in Brazil are concentrated in the north of the country (Amazon Forest Region). All medications used in the treatment of leishmaniasis are toxic and expensive. There are also difficulties in controlling the disease due to the great biological parasite diversity, the different clinical forms of the disease, including those severe forms resistant to chemotherapy, thus making prevention through vaccination the best strategy.
Currently, there is no vaccine approved for human use against leishmaniasis; however, several vaccine preparations are being studied. The Leishvacin® (or LaAg) vaccine is a first-generation vaccine composed of total proteins, lipids, carbohydrates, RNA and DNA of L. amazonensis and has been studied for years [3,4]. The efficacy of the mucosa as administration route of LaAg vaccine has already been tested in both oral and intranasal routes [5,6]. Studies show that intranasal administration of LaAg provides greater protection to BALB/c mice challenged with L. amazonensis and is more advantageous due to the easier application, and lower doses of antigen required as compared to the oral route. The protection achieved by intranasal immunization was accompanied by the development of a long-term immune memory and adaptive immunity [5,7].
Toll-like receptors (TLRs) are transmembrane proteins that recognize pathogen-associated molecular patterns (PAMPS) [8]. The TLRs play an important role during Leishmania infections. TLR9 recognizes unmethylated CpG DNA sequences, which are commonly found in bacteria and Leishmania [9], but not in mammalian cells where these sequences are normally methylated [10]. It has been shown that the activation of TLR9 promotes a host-protective response. For example, TLR9-dependent activation of dendritic cells (DCs) by DNA from L. major favors Th1 cell development and lesion resolution [11]. TLR9 signaling is essential for the innate natural killer (NK) cell response in murine cutaneous leishmaniasis caused by L.major [12]. Similarly, in visceral leishmaniasis caused by L. major, the activation of NK cell also requires TLR9 [13]. DC activation by L. braziliensis has also been shown to be dependent on TLR9 in vitro [14]. Furthermore, TLR9-/- mice inoculated with L. braziliensis exhibited transiently increased lesion sizes and parasite burdens in comparison with those of control mice [14]. DNA from L. mexicana activates murine bone marrow-derived macrophages leading to the production of proinflammatory cytokines, such as TNF-α and IL-12, as well as the overexpression of mRNA for TLR9 [15].
Despite all this knowledge, our understanding of TLRs and cytoplasmic pattern recognition receptors (PRRs) that recognize and respond to Leishmania is rather limited [16]. In this study, we investigated the role of TLR9 during L. amazonensis infection in C57BL/6 (WT) mice and C57BL/6 TLR9-/- mice, the adjuvant effect of the DNA containing CpG motifs present in LaAg vaccine and their efficacy when administered by the mucosal route, which presents high expression of TLR9 [17].
C57BL/6 WT mice were acquired from Universidade Federal Fluminense (UFF) and Universidade Federal do Rio de Janeiro (Fundação BioRio). Tlr9−/− mice in the C57BL/6 background were generated by and obtained from Dr. S. Akira (Osaka University, Japan). All animals were maintained in our own animal facility at UFRJ using sterilized bedding, filtered water and commercial feed ad libitum. Experimental groups consisted of C57BL/6 WT or TLR9-/- mice at 6–8 weeks of age. The Health Sciences Center Ethics Committee of Federal University of Rio de Janeiro (Comissão de Ética no Uso de Animais do Centro de Ciências da Saúde da Universidade Federal do Rio de Janeiro) approved the animal use under the protocol number IBCCF 157.
The parasites used in this study were L. amazonensis (MHOM/BR/75/Josefa) originally isolated from human cutaneous leishmaniasis [18], maintained at 26°C in M199 medium (Sigma) containing 10% heat-inactivated fetal bovine serum (SFB, Cultilab) and hemin (5μg/ml, Sigma). To ensure infectivity, amastigotes were isolated from lesions of pre-infected BALB/c mice and promastigotes were only used until the fifth passage of the culture.
Macrophages were isolated from the peritoneal cavity of mice following the injection and withdrawal of RPMI (Gibco, NM, USA) medium. The cells were counted with Trypan blue and transferred to a 24 well plate at a concentration of 5×105 cells/well, then incubated for 1 h at 37°C and 5% CO2 to allow the macrophages to adhere to the plate. Thereafter, the plate was washed with PBS three consecutive times to remove the non-adherent cells, and 400 μL RPMI with 10% FBS supplemented with glutamine, pyruvate, and non-essential amino acids was added. After 24 h, the wells were washed again with PBS to remove B1 lymphocytes and 300μL RPMI with 10% FBS was added.
Macrophages (5×105 cells/well) were infected with 2.5×106 of stationary-phase L. amazonensis (ratio of 5:1). After 4h, the wells were washed with PBS three times and left at 37°C and 5% CO2. After 48 h, the supernatants were recovered and the plate was washed again with PBS three times. Finally, the plate was fixed and stained with a fast panoptic kit (Laborclin, RJ, Brazil). The infection was analyzed by optical microscopy (CX31, Olympus, Japan). One hundred (100) macrophages were counted per well and it was evaluated whether the macrophages were infected or not (percentage of infection) and the number of intracellular amastigotes in the infected macrophages were counted.
After in vitro infection, the supernatants were collected and placed in a 96 well plate (100 μl/well) followed by Griess reagent (100 μl/well) (SIGMA). The plate was incubated for 10 min and the optical absorbance was read at 540 nm.
Neutrophils were obtained from mouse bone marrow as previously described [19]. Briefly, bone marrow cell suspension obtained by femur and tibia flushing were centrifuged in a Percoll gradient (58%, 65% and 72% v.v.; GE Healthcare, Little Chalfont, UK), and 74.5 ± 2.59% Ly6G+ neutrophils were obtained through this method.
Neutrophils (5×105 cells/well) were either left untreated (Nil) or incubated with increasing amounts of promastigotes (1:1, 1:5) for 4 h at 35°C and 5% CO2. Culture supernatants were collected and extracellular DNA was analyzed using Quant-IT dsDNA Picogreen kit (Invitrogen). Fluorescence was detected using a SpectraMax Paradigm microplate reader (Molecular Devices). Data was presented as fold increase over control.
Neutrophils (1×105) seeded on a poly-L-lysine-coated glass slide were incubated with CSFE (0.5 μM, Invitrogen) stained-promastigotes (1 parasite/1 neutrophil ratio) for 4 h at 35°C and 5% CO2. Cells were fixed with 4% formaldehyde, washed with PBS, blocked with mouse serum for 60 min, washed twice with PBS and blocked again with 4% PBS-BSA for 30 min at room temperature. The cells were treated with anti-histone H1 (1:400; EMD Millipore, Billerica, MA, USA), followed by anti-mouse AF 546-Mab 3864 (1:400; EMD Millipore) and DAPI (Sigma) for 60 min in each step. Confocal images were taken using a Leica DMI-8 microscope.
Naïve mice were euthanised and spleens from WT and TLR9-/- mice were removed and treated with ACK. 1×106 cells from the spleens were cultivated with 5×106 L. amazonensis promastigotes stained with CFSE, or spleen cells were incubated with medium alone, as a as control. After 24 h, cells were stained for CD11c (PerCP), MHCII (PE), and CD86 (APC) and analysed by flow cytometry. Cells were gated on CD11c+, and further analysed based on the expression of MHCII+ and CD86+, and finally infection of these cells was determined by CFSE- or CFSE+ expression.
Leishmania amazonensis promastigote antigens (LaAg) were prepared as previously described [20]. Briefly, stationary-growth phase promastigotes were washed three times in PBS and subjected to three cycles of freezing and thawing. LaAg was lyophilized, stored at -20°C and reconstituted with PBS immediately prior to use. For quality control, protein content was quantified using the Smith method with the bicinchoninic acid protein kit (Sigma-Aldrich), expecting a protein level around 60% and SDS-PAGE was performed routinely as described before [3, 4]. The quantification of the DNA content was performed by ultraviolet spectrophotometry after DNA purification using the Kit Extractme Genomic DNA (Blirt DNA Gdansk). The protein content was 647 μg and the DNA content was 0.8 μg in 1 mg of LaAg.
Mouse immunization was carried out by instillation of 10 μg of LaAg in 20 μl PBS (10 μl per nostril) using a micropipette adapted with a polystyrene microtip. A booster dose was given 7 days later [3]. The control mice received just PBS. Seven days post-boost, the animals were infected in the right hind paw with 5×105 L. amazonensis stationary phase promastigotes in 20 ul PBS. For hypersensitivity, the size of footpad was evaluated 18 h, 24 h, and 48 h after infection using a pachymeter and expressed as the difference between the thicknesses of infected and contralateral PBS-injected paws [20]. For evaluation of lesion growth, lesion sizes were measured once a week using a pachymeter and expressed as the difference between the thicknesses of infected and contralateral non-infected paws. The parasite load was determined at the end of the experiments, when the infected foot was skinned and individually homogenized in 1 ml PBS using a tissue grinder. Tissue debris were removed by gravity sedimentation for 5 min. The homogenates were then submitted to limiting dilution assay (LDA).
Quantification of antibody production was carried out using ELISA (Goat Anti-Mouse IgM-UNLB:Cat. No.1021-01, Goat Anti-Mouse IgG Fc-UNLB:Cat. No. 1033–01, Goat Anti-Mouse IgG1-UNLB:Cat. No. 1071–01, SouthernBiotec). First, total Leishmania amazonensis antigen (LaAg) diluted in PBS (5 μg/mL) was added to the plate for coating overnight. On the second day, the content was discarded and the plate was blocked with Block Buffer (PBS with 5% heat-inactivated fetal bovine serum (HIFCS, GIBCO Laboratories, Grand Island, NY, USA) and 0.05% Tween 20) for 1 h. The plate was then washed three times with Wash Buffer (PBS with 0.05% Tween 20) and the samples were diluted in Block Buffer and added to the plate. After 1 h, the plate was washed five times with Wash Buffer and the secondary antibody specific for each isotype of interest (one for each plate) was added. After 1 h, the plate was washed seven times with Wash Buffer and TMB was added. Finally, the reaction was blocked with HCl.
For in situ production [21], infected paws were isolated as mentioned above. The paw homogenates were centrifuged (10 min, 20,000 × g, 4°C) and the supernatants were collected. For antigen stimulation, lesion-draining popliteal lymph nodes were excised 72 h post infection and single-cell suspensions were prepared. The cells were plated at a concentration of 1×106 cells/ml and were stimulated with 50 μg/mL of Leishmania major antigens (LmAg) for 72 h at 37°C with 4% CO2 (For in vitro assay, we used LmAg instead of LaAg since the latter induces apoptosis [4]). IFN-γ, IL-10 and IL-4 cytokines were quantified in the supernatant by ELISA following the manufacturer’s instructions (R&D Systems).
Lymph node cells were cultured for 4 h at 37°C in the presence of PMA (20 ng/ml), Ionomycin (1 μg/ml) and brefeldin A (Sigma-Aldrich). The cells were surface stained with Anti-CD3-PerCP, anti-CD8-APC-Cy7 and anti-CD4-PE-Cy7 (Biolegend) and were fixed and permeabilized for 1 h using the Foxp3/Transcription Factor Fixation/Permeabilization Kit (e-Bioscience, Santa Clara, USA). Intracellular cytokine staining was performed with anti-IFN-γ-APC and IL10-Pe (Biolegend). At least 10,000 gated CD4+ lymphocyte events were acquired in a BD FACSCanto™ II (BD Biosciences New Jersey, USA) and the data was analysed with FlowJo X software.
Experiments were performed two or three independent times, and the result of one representative experiment is shown. The lesion sizes caused by the infection were statistically analyzed by Two-way ANOVA following Bonferroni’s post hoc test. The results provided in the remaining figures were tested by Student’s t-test. We used GraphPad Prism v.5 software, and the results were considered significant when P ≤ 0.05.
To assess the L. amazonensis infection profile of macrophages from C57BL/6 (WT) and TLR9-/- mice in vitro, peritoneal cells were infected and analyzed after 48 h as described in the methodology. Using light microscopy, the infection of WT macrophages (Fig 1A) and TLR9-/- macrophages (Fig 1B) was evaluated, and a similar infection profile was observed between both groups. The analysis of the total number of amastigotes (Fig 1C) and the ratio of amastigote/macrophage (Fig 1D) also presented similar results. Likewise, the production of nitric oxide (NO) showed no significant difference between WT and TLR9-/- macrophages (Fig 1E).
The formation of neutrophil extracellular traps (NETs) derived from WT and TLR9-/- bone marrow neutrophils in response to L. amazonensis (Fig 2) was assessed. NET release was analysed using the immunofluorescence microscopy after staining for anti-DNA/Histone H1. No difference was observed in the NET release by WT (Fig 2A) and TLR9-/- neutrophils (Fig 2B) stimulated with L. amazonensis. The release of dsDNA using the DNA picogreen assay of neutrophils from WT and TLR9-/- neutrophils was also measured (Fig 2C), and, again, no difference was observed. These results suggest that TLR9 does not participate in NET induction by L. amazonensis infection in vitro.
Based on the role of TLR9 in the activation of dendritic cells, the percentage of the DC population, through the expression of MHCII and CD86, from the spleen of WT and TLR9-/- mice was evaluated by flow cytometry following L. amazonensis infection. As was demonstrated before [22, 23, 24], Leishmania amazonensis does not have the ability to activate DC, as can be observed in the similar percentages of double positive MHCIIhi+CD86+ in WT, regardless of whether the cells were infected or not, as identified through CFSE expression from parasites, or a non-infected medium control was used (Fig 3A). In TLR9-/-, a similar profile was observed (Fig 3A). There was no difference in CFSE+ DCs between WT and TLR9-/- samples, indicating that there is no difference in phagocytosis of L. amazonensis (SF3).
To evaluate the role of TLR9 during in vivo infection, WT and TLR9-/- mice were infected with L. amazonensis. No difference was observed in the parasite load 7 days post-infection (dpi) (Fig 4) indicating the absence of an early response dependent on TLR9. Lesion size of infected mice was monitored weekly which showed a similar progression in both groups until 42 dpi (Fig 5A). After which time, the lesion size in TLR9-/- mice increased significantly more, with a peak around 60 dpi, followed by a partial resolution of the lesion with chronic parasite persistence in both groups until the last day (Fig 5A and 5B). Interestingly, although presenting similar lesion sizes at 120 dpi, the parasite load was significantly higher in TLR9-/- mice (Fig 5B). In another set of experiments, mice lesions were monitored until the peak of infection (67 dpi), confirming the higher lesion size in infected TLR9-/- mice (Fig 5C). The parasite load evaluated at 67 dpi in these groups was also significantly higher in TLR9-/- mice when compared with WT mice (Fig 5D).
Production of immunoglobulins by both WT and TLR9-/- mice at 67 dpi was assessed by ELISA. TLR9-/- mice produced higher amounts of IgG and IgG1 (Fig 6B and 6C) compared with WT mice, but there was no difference in IgM levels (Fig 6A). These results suggest an involvement of IgGs in the increase of infection susceptibility to L. amazonensis by TLR9-/- mice.
In order to further identify possible causes of the differences in lesion size and parasite load at the peak of infection between the groups, the production of cytokines in mice paws at 67 dpi was also assessed by ELISA. The results showed a decrease in IFN-γ production by TLR9-/- mice in comparison with WT mice (Fig 7A). However, the production of other cytokines, such as IL-4 and IL-10, showed no significant difference between the two groups (Fig 7B and 7C).
Due to the difference observed in IFN-y in the paw, flow cytometry was used to evaluate the percentage of IFN-γ produced by CD4+ T cells and CD8+ T cells present in the draining lymph node at 67 dpi. The results demonstrated a decrease in the percentage and number of IFN-γ producing CD8+ T cells in TLR9-/- mice when compared with WT mice (Fig 8A and 8B). The percentage and numbers of IL-10-producing CD8+ T cells was not significantly different between the two groups (Fig 8C and 8D). The same cytokines were also evaluated in CD4+ T cells and the results were similar between both groups (Fig 9A–9D).
It is known that TLR9 is highly expressed in nasal mucosa [25] and that the LaAg vaccine has DNA in its composition [4]. Our results suggest that TLR9 participates in the induction of adaptive immune response against L. amazonensis, which lead us to consider the hypothesis that the recognition of the LaAg DNA by TLR9 in the mucosa could act as an adjuvant, improving LaAg vaccine. Thus, to assess whether there would be a relationship between the effectiveness of the LaAg vaccine and the activation of TLR9, the profile of L. amazonensis infection in LaAg-vaccinated WT and TLR9-/- mice was investigated (Figs 10A and 11A).
In the peak of infection, vaccinated TLR9-/- mice partially controlled the lesion in comparison with control TLR9-/- mice injected with PBS (Figs 10A and 11A). Vaccinated TLR9-/- mice showed no difference in relation to PBS WT mice, however, they demonstrated larger lesions in comparison with the LaAg-vaccinated WT mice (Figs 10A and 11A). The parasite loads at 52 dpi suggest that the LaAg vaccine efficacy is partially dependent on TLR9 activation in the peak of infection (Fig 11B). This dependency was more evident at 166 dpi, as vaccinated TLR9-/- mice were unable to control parasite burden (Fig 10A and 10B).
To investigate the lack of protection that controls the parasite load in vaccinated TLR9-/- mice, the delayed hypersensitivity response after LaAg immunization and using the infection as challenge was evaluated. Vaccinated WT mice showed a strong delayed hypersensitivity response at 18 h, 21 h and 24 h, compared with control WT mice injected with PBS, control TLR9-/- mice and LaAg vaccinated TLR9-/- mice (Fig 12). Vaccinated TLR9-/- mice presented a small but significant increase in thickness compared to the control TLR9-/- mice only 21 h after challenge. Then, 3 days after infection, the draining lymph nodes were antigen-stimulated ex vivo. In WT mice, those that were vaccinated were able to produce IFN-γ (Fig 13A); however, the same was not observed in TLR9-/- mice. No difference was observed in the production of IL-4 (Fig 13B) and IL-10 (Fig 13C). Taken together, we observed that vaccinated WT mice were more competent to induce cellular response in comparison with vaccinated TLR9-/- mice.
TLRs serve as the first line of defense for innate immune cells and are important for protecting against Leishmania infections [8–16]. The importance of TLR signaling to protect against Leishmania infections is further established by studies that have as direct targets several TLRs using specific agonists and knockout mice [16]. It has already been described that TLR9 can be activated by non-methylated CpG DNA found in Leishmania [9], but not in mammalian cells, in which these sequences are normally methylated [10].
Neutrophils and macrophages are among the cells of the innate immune system that respond to Leishmania infection. These cells have several PRRs that help them to protect the host environment from invading pathogens. Thus, we decided to evaluate the in vitro infection by L. amazonensis in TLR9-/- cells in comparison with WT C57BL/6 mice. Our results showed no significant difference between the infection of macrophages from TLR9-/- and WT mice in addition to NO production (Fig 1A–1E). Likewise, the infection by L. amazonensis did not induce formation of NETs (Fig 2A–2C). Although other works analyzed the role of TLR9 in different Leishmania species [12,14, 26, 27], none of them examined the in vitro macrophage infection and neutrophil NET formation induced by L. amazonensis. We also evaluated the infection of dendritic cells by Leishmania amazonensis, however, it was not possible to observe the activation of infected dendritic cells in WT and TLR9-/- (Fig 3). Although L. amazonensis has the ability to impair DC activation and it is not related to TLR9. Indeed, the same phenotype was observed before on dendritic cells from WT infected with L amazonensis [22, 23, 24].
To evaluate the role of the TLR9 receptor in vivo, WT mice and TLR9-/- mice were infected with L. amazonensis. The parasite load 7 days post-infection showed no difference between the groups (Fig 4), indicating the lack of an early response dependent on TLR9. The groups also presented a similar profile of lesion progression during the infection (Fig 5A–5D). However, the TLR9-/- mice showed a slight increase in the lesion size at the peak of infection, which coincided with an increase in parasite load (Fig 5). These results are similar to those found in studies with Leishmania major [12], Leishmania braziliensis [14], Leishmania infantum [28] and Leishmania guyanensis [27]. These results suggest the participation of TLR9 in the immune response against leishmaniasis.
Antibodies have a pathogenic role in infection by L. mexicana [29, 30] since knockout mice for antibody receptors are able to resolve infection against those parasites. IgG1 antibodies are considered the pathogenic antibodies in L. mexicana infection [30]. Antibody production is also associated with pathogenesis in the L. amazonensis model, since BALB/JhD mice, a lineage that lacks B cells, is more resistant to infection [31] or using BALB/XID mice [32]. For this reason, we analyzed the levels of serum immunoglobulins of both WT and TLR9-/- groups. TLR9-/- mice produced higher amounts of IgG and IgG1 when compared with WT mice (Fig 6B and 6C), but the levels of IgM were similar between these groups (Fig 6A). In the first moment, we expected to observe the reduction of IgG and IgG1 in TLR9-/- may be due to direct effect on B cells since the importance of TLR9 to activation and production of antibody [33, 34] that was observed in L donovani infection when endossomal TLR-/- mice (Unc931bLetr/Letr mice) presented a reduction of IgG production [35]. However, we observed an increase of specific antibodies IgG in TLR9-/- that corroborate with a recently finding that demonstrate that TLR9 activation inhibits proliferation, differentiation and production of IgG production of Follicular B cells [36]. Besides, the lack of TLR9 was directly related with the increase of IgG1 in other models [37, 38] corroborating our data. We suggest a pathogenic role for the antibody production, since TLR9-/- mice presented higher levels of IgG1 along with larger lesions and higher parasite loads, which corresponds with previous data [29, 30, 31, 32].
It is already known that TLR9 activation induces a Th1 type immune response [39]. The ability of CpG (antigen that activates TLR9) to induce Th1 polarization made it an interesting new target for treating allergy and infectious diseases [40]. Therefore, we analyzed the Th1 cytokine that could be involved in the infection and in the increase of parasite load and lesion size in TLR9-/- mice. We observed a reduction of IFN-y in the infected tissue of TLR9-/- mice (Fig 6A). Flow cytometric analysis evidenced that the lack of TLR9 decreases the production of IFN-γ only in CD8+ T cells (Figs 7A, 8A and 8B) during L. amazonensis infection. It has previously been demonstrated that TLR9 does induce IFN-γ producing CD8+ T cells [41]. In our model, the number of IFN-γ producing CD4+ T cells was similar between the groups (Fig 9A and 9B), indicating that they did not fail to induce Th1 response (only IFN-γ producing CD8+ T cells). Similar findings showing that the activation of TLR9 increases IFN-γ producing CD8+ T cells, but not Th1 cells, have already been reported [42], corroborating our data. In addition, no difference was observed in the production of IL-4 and IL-10 (Figs 7B, 7C, 8C, 8D, 9C and 9D).
The participation of IFN-γ modulated by TLR9 is still unclear, since infections caused by different species of Leishmania result in different profiles. In L major, the reduction of IFN-γ produced by NK, but not by CD4+ T cells, was associated with increased number of parasites pathogenesis [12]. However, in another study using L. major, the production of IFN-γ was also reduced in CD4+ T cells [11]. In both studies, the production of IFN-γ by CD8+ T cells was not investigated. Independently, in these two studies, IFN-γ was associated with protection. Furthermore, it was shown that, in the absence of TLR9, a decrease of lymph nodes hypertrophy was observed [43] that is directly related to the absence of effector cells. In L. braziliensis, a different phenotype with increased hypertrophy in TLR9-/- was observed, which increased the number of IFN-γ producing CD4+ T cells and IFN-γ producing CD8+ T cells. In L guyanensis, the lack of TLR9 allowed an increase of the Th2 response [27]. In our study, we did not observe any modification in lymph node hypertrophy (SF4). The reduction of IFN-γ was observed in the lesion site and in IFN-γ producing CD8+ T cells. Recently, it was demonstrated that TLR9 expression is predominant in lesion of patients with L. amazonensis in comparison with L. braziliensis [44], this could be an explanation for the differences in the immunological profile between these species. Further studies are necessary to gain a complete understanding of the role of TLR9; however, in general, TLR9 is related to the production of IFN-γ by different cell types that is associated with the control of lesion size and parasite load.
The LaAg vaccine, which comprises the total lysate of L. amazonensis, contains fragments of Leishmania genomic DNA, proteins and other cell debris [4]. These DNA fragments, that contain CpG sequences, activate TLR9, thereby developing a protective immune response against infection. Our group has been studying the use of adjuvants to increase vaccine effectiveness in order to obtain a better combination promoting a greater effect of the vaccine against leishmaniasis [45]. Furthermore, the use of TLR9 agonists as mucosal vaccine adjuvants has been studied to understand the best way to activate this receptor and increase the immune response protection that it can trigger [38].
LaAg vaccine induced protection via the intranasal route [3], although it failed to induce protection by the intramuscular route [6,46]. As it was confirmed in this work, following vaccination and challenge with L. amazonensis the infection in the paw had a reduced lesion size compared to the unvaccinated group (Fig 10A), showing that the administration by intranasal route of the LaAg vaccine is systemically effective. The choice of the administration route is very important, since leishmaniasis is a disease with broad spectrum, present in all continents and has an increasing number of cases worldwide [7]. The use of the intranasal route is an easy-to-administer technique, which facilitates mass immunization campaigns, does not require sterile injection and, therefore, has no personnel training costs. This is an interesting prospect, since leishmaniasis is considered a neglected disease that affects several countries with limited financial resources and would improve vaccine coverage worldwide [39]. It is also evident that it increases treatment compliance for promoting a painless immunization.
Although the only approach to respiratory mucosal vaccination that successfully achieved commercial use is nasal immunization against influenza, the major challenge for nasal vaccines is still the lack of licensed nasal adjuvants to enhance mucosal and systemic immunogenicity [47, 48]. Recently, it was demonstrated that TLR9 is present in nasal mucosa [17, 41]. Based on this and on the presence of DNA content in LaAg vaccine, we decided to evaluate the involvement of TLR9 in intranasal efficacy. We demonstrated the partial involvement of TLR9 in LaAg vaccine in the lesion control and parasite load (Figs 10 and 11). Although some studies investigate the profile of the immune response during vaccination against leishmaniasis, this is mainly focused on IFN-γ and iNOS [49], and few studies investigate the participation of receptors, specially TLR9, in vaccination mechanisms against leishmaniasis. From what we know at present, this is the first report that associates the importance of TLR9 receptor in LaAg vaccine efficacy against leishmaniasis. We showed that intranasal immunization with LaAg is dependent on TLR9 to induce effector function demonstrated by an increase of delayed hypersensitivity (Fig 12) and production of IFN-γ (Fig 13) by ex vivo antigen stimulation only in vaccinated WT mice but not in TLR9-/- mice. Cellular response [50] and production of IFN-γ [51] have already been associated with a mechanism to control L. amazonensis infection. The same mechanism was described in the efficacy of LaAg intranasal vaccine, respectively, cellular response evaluated by hypersensitivity [5] and production of IFN-γ, [5,45] which supports our data shown here.
Understanding the mechanism of action and efficacy of first generation vaccines is important to improve second generation vaccines. It has been demonstrated that the yellow fever vaccine YF-17D activates multiple TLRs, including TLR9 in plasmacytoid and myeloid DCs [52]. The typhoid fever and BCG vaccines, which carry the same formula since their creation in 1911 and 1921, respectively, have DNA as adjuvant to activate TLR9. It is clear that the use of DNA as adjuvant is not a novel mechanism and, more recently, other studies demonstrated the positive effects of DNA as adjuvant in vaccines [53, 54]. It was proven that the use of BCG-DNA as adjuvant enhances the immune response in three viral swine diseases and also in Taenia solium cysticercosis, where the BCG-DNA or CpG-ODN increases the levels of IgG2, IFN-γ, the percentage of CD8+ T cells and specific proliferation of peripheral blood mononuclear cells [55, 56]. However, plasmid DNAs vaccines work through mechanisms other than the TLR9; it has been shown that CpG motifs found in plasmids don’t significantly increase the TLR9 response in DNA vaccines with both TLR9-/- and wild-type mice presenting antibody and cellular immune responses to DNA vaccine similar in terms of magnitude and type [57].
Adjuvants are substances capable of increasing and improving the quality of the immune response, and they are often present in the subunit vaccines. Some adjuvants currently used are TLR ligands, such as AS04 (a TLR4 activator) [58]. Furthermore, TLR9 agonists were described as highly effective adjuvants for viral vaccines, capable of inducing specific and functional antibody responses with low doses [54]. CpG adjuvants have been reported as being effective and safe for intranasal vaccines [55, 56]. Here, for the first time investigating the LaAg vaccine, TLR9-/- mice were vaccinated and their immune responses were evaluated. Our study indicates that TLR9 is associated with the mechanism of protection of intranasal LaAg vaccine and suggests that the use of a TLR9 agonist could be considered an adjuvant by intranasal route against leishmaniasis. The participation of TLR9 in LaAg intranasal vaccine efficacy through induction of systemic IFN-γ production is the main contribution of this manuscript.
In this work, we observed that TLR9-/- animals are more susceptible to infection, presenting a larger lesion size in the peak of infection, and also an increase of parasite load in the peak of infection and in the chronic phase. Moreover, the effectiveness of the LaAg vaccine is partially dependent on the activation of the innate immunity receptor TLR9, mainly in the chronic phase where the absence of the TLR9 receptor lead to no protection in vaccinated mice.
|
10.1371/journal.pcbi.0030198 | Evolutionary Models for Formation of Network Motifs and Modularity in the Saccharomyces Transcription Factor Network | Many natural and artificial networks contain overrepresented subgraphs, which have been termed network motifs. In this article, we investigate the processes that led to the formation of the two most common network motifs in eukaryote transcription factor networks: the bi-fan motif and the feed-forward loop. Around 100 million y ago, the common ancestor of the Saccharomyces clade underwent a whole-genome duplication event. The simultaneous duplication of the genes created by this event enabled the origin of many network motifs to be established. The data suggest that there are two primary mechanisms that are involved in motif formation. The first mechanism, enabled by the substantial plasticity in promoter regions, is rewiring of connections as a result of positive environmental selection. The second is duplication of transcription factors, which is also shown to be involved in the formation of intermediate-scale network modularity. These two evolutionary processes are complementary, with the pre-existence of network motifs enabling duplicated transcription factors to bind different targets despite structural constraints on their DNA-binding specificities. This process may facilitate the creation of novel expression states and the increases in regulatory complexity associated with higher eukaryotes.
| Networks are a simple and general way of representing natural phenomena that range in scale from the social interactions between people to the organization of circuits on a microchip. Many networks have been found to contain repeated patterns of connections between small groups of nodes. These patterns, termed network motifs, are thought to be involved in controlling the flow of information through the network. This article investigates the processes that led to the formation of the two most common types of motif in the network controlling gene expression in baker's yeast. Around 100 million y ago, yeast's ancestor underwent a whole-genome duplication, which resulted in the organism containing four copies of each gene rather than the usual two. The duplicated genes that remain in the yeast genome are used to infer the two mechanisms that give rise to network motifs. These are rewiring of interactions between genes, and the duplication of proteins that control gene expression (transcription factors). These two processes are complementary with the rewiring mechanism enabling duplicated transcription factors to regulate the expression of different genes. It appears likely that these two processes are involved in enabling the increases in complexity that are associated with multicellular life.
| One of the most fundamental questions in biology is how incremental evolutionary changes lead to the observed complexity in biological systems. The advent of genome sequencing and associated functional genomic technologies have provided the first evidence for the origins of complexity on an organism-wide scale. Modularity is an emergent property of biological networks that has been observed in metabolic [1], protein–protein interaction [2], and transcription factor networks (TFNs) [3]. Several explanations have been put forward for the evolution of modular biological systems, which include robustness to mutational [4] and environmental perturbations [5], insulation against cross-reactivity between alternative signalling cascades [6], and selection for survival in multiple environments [7].
Parallel studies of small, artificial TFNs have demonstrated that alterations in network topology and components can be used to create a wide range of dynamic properties such as bistability and oscillations. However, relatively few local topologies are widely observed in natural networks [3,8]. For example, although a circuit composed of two inhibitory transcription factors (TFs) arranged in a feedback loop has been shown to act as a stable memory element in the lambda phage virus and artificial systems [9], this topology is uncommon in both the Escherichia coli and Saccharomyces cerevisiae transcriptional networks so far uncovered [3,8]. An outstanding question is whether the absence of these and other local topologies is a result of mechanistic or functional constraints on network evolution.
In this article, transcription regulatory interactions in the yeast S. cerevisiae were defined using the large-scale chromatin immunoprecipitation (ChIP-on-chip) dataset of Harbison et al. [10] These interactions were used to define a network with nodes representing genes and directed edges binding of a protein encoded by a TF gene to the promoter of a target gene. We begin by investigating several growth models for the formation of bi-fan motifs, which involve a pair of TFs that bind the promoters of two target genes, as shown in Figure 1. The bi-fan motif is typically embedded in extended structures that we term the bi-fan array, involving a pair of TFs that both regulate a larger number of common target genes. Figure 1 illustrates how the number of bi-fan motifs within an array grows quadratically as target genes are added. In later sections, we demonstrate a specific structural relationship between bi-fan arrays and the feed-forward loop (FFL) motif, and a common origin for many of these network structures.
The topology of the bi-fan motif suggests several evolutionary mechanisms for its formation, including duplication of either TFs or target genes [11]. It is also possible that the motifs could have arisen from rewiring of regulatory interactions as a result of cis-sequence evolution in genic promoter regions or the trans-evolution of the protein sequences encoding TFs. The cis-sequence evolution refers to mutations in noncoding regions that alter the binding affinity of TFs for a particular promoter, thus affecting the expression of genes in close proximity on the chromosome [12,13]. Conversely, trans-evolution typically involves mutations in the sequences encoding TFs that alter, for example, their DNA-binding specificity. These trans-changes have the potential to alter the expression of large numbers of genes [12,13]. In this article, the relative contributions of these mechanisms are investigated by defining a common evolutionary origin for pairs of genes using the whole-genome duplication (WGD) event that occurred in S. cerevisiae after its divergence from Kluyveromyces waltii [14,15].
We investigated the organisation of bi-fan motifs in the yeast TFN using two algorithms that have been used previously for detecting motifs in directed networks [3,8]. These algorithms fix both the in-degree and out-degree of each node and then randomly replace the edges in the network. This approach can then be used to detect motifs that occur more frequently in the native network than a large ensemble of random networks (see Methods for further details). Although the original methods for detecting network motifs involved exhaustive enumeration of all small (typically 2- to 6-node) subgraphs in the network, previous work [3,16] suggests that bi-fan motifs are embedded in larger structures within the yeast and E. coli TFNs. In fact, it is possible to show (see Methods for details) that the overrepresentation of bi-fan motifs in any directed network is associated with the array structures shown in Figure 1.
Bi-fan arrays were identified in the yeast TFN by searching for pairs of TFs with a number of shared targets that exceeded the number found in the randomized networks with p < 10−4. A description of the p-value calculation is included in the Methods section. A total of 442 bi-fan arrays were identified at this strict significance threshold. These arrays account for a total of 1.25 × 105 (68% of the total) bi-fan motifs compared with an expected number of 7.3 × 103 under the null model. The overrepresentation of bi-fan motifs in the Saccharomyces TFN (shown in Table 1) can therefore be attributed to a relatively small number of bi-fan arrays that, on average, regulate a large number of target genes. The following two sections investigate the influence of gene duplication on formation of the bi-fan array structure.
Two approaches were used to identify genes that have arisen from duplication. The first method involves using genes that were created from the most recent WGD in the evolution of S. cerevisiae [14,15]. These data are likely to be of very high fidelity because of the requirement for genes to reside in regions of doubly conserved synteny with the K. waltii genome [15]. Another advantage of defining common origin using WGD data is that duplication of all genes occurred simultaneously, and duplicates initially possessed very similar promoter regions. This provides a means to estimate the relative cis- and trans-conservation rates upon gene duplication, as shown in Table 2.
Table 2 shows that the trans-conservation rate is relatively high, which is caused by nine of the 17 WGD duplicates forming statistically significant bi-fan arrays. These arrays contain a substantial proportion of the network's bi-fan motifs. Conversely, the cis-conservation rate for all promoters duplicated by WGD is low, with relatively few bi-fan motifs arising from conserved interactions. In the case of promoters of genes that are diverging rapidly, the conservation rate is only slightly above that expected for randomly selected promoters and indicates substantial plasticity in promoter binding.
It is also possible to rule out more recent single-gene duplications as a significant source for bi-fan motifs, as these have been estimated to occur very infrequently in S. cerevisiae, at a rate λ = 1–6 × 10−5 per gene per million y [17]. An upper bound for the number of single-gene duplications that have occurred since the divergence of S. cerevisiae from K. waltii can be calculated by assuming that the rate of duplication is at the upper limit and that the rate of loss is zero. The number of gene duplications is then given by the exponential growth model
where NG = 3,500 is the approximate number of single-copy genes in S. cerevisiae, and T = 100–150 million y is the time since WGD [17]. Equation 1 suggests that the number of single-gene duplications that have occurred since WGD, NG, is less than 35. Conservation at the levels shown in Table 2 would not result in a large number of bi-fan motifs originating from target gene duplication.
WGD is a feature in the evolution of most known eukaryote organisms, including chordates [18]. However, fewer than 10% of yeast proteins originated from the latest WGD in the Saccharomyces lineage. More ancient gene duplications account for the majority (90%) of proteins encoded in the yeast genome [19]. For this reason, we identified duplicates with a more ancient common origin using domain assignments from the Pfam HMM library [20] (see Methods for further details). The results shown in Table 2 have demonstrated that the promoter-binding patterns of duplicate target genes are likely to have diverged on time-scales longer than 100–150 million y, so the analysis is restricted to TFs with common origin identified with the structure of their DNA-binding domains. These results indicate that a total of 27 bi-fan arrays involve TFs with structurally similar DNA-binding domains, accounting for a total of 14.4% of the bi-fan motifs. 239 bi-fan arrays containing 49.2% of the motifs involve two nonhomologous TFs with the remainder involving at least one TF with an unknown structure. This suggests that more ancient TF duplications have also contributed to the formation of bi-fan motifs in the network (see Figure S1).
In summary, the redundancy of duplicated TFs results in the formation of bi-fan arrays, although the majority of these network structures do not arise directly from gene duplication. Conversely, the duplication of target genes does not appear to contribute greatly to formation of bi-fan arrays because the network is subject to greater cis-plasticity. This difference also arises from the different statistical properties of the (compact) in-degree distribution and the (power-law) out-degree distributions [21]. Taken together, these results suggest that the two major processes that contribute to the formation of bi-fan motifs are duplication of TFs and the accumulation of common target genes, as depicted in Figure 2A–2B.
The colocalization of nonhomologous TFs at genic promoters is likely to involve a combination of two physical mechanisms. The first mechanism involves the presence of binding sites for the two TFs that occur independently in the same set of genic promoters [22]. This process could also enable cooperative binding if a TF displaces nucleosomes that occlude the binding site of a second TF [23]. The plasticity in the promoters of duplicated genes, shown in Table 2, suggests that bi-fan arrays could have arisen from mutations in promoter regions and subsequent selection for TF binding at numerous dispersed loci. The second mechanism involves protein interactions between the TFs that enable cooperative binding to DNA. For example, mitogen-activated protein kinases without intrinsic DNA-binding affinity are localised to actively transcribed genes during the stress response in yeast via interactions with other proteins [24]. It has also been shown previously [8] that protein–protein interactions tend to occur between pairs of TFs that form bi-fan motifs, and we have confirmed that this property also applies to the bi-fan array structure (Figure S1). In the following section, we investigate how gain and loss of protein–protein interactions could cause duplicated TFs with similar DNA-binding specificities to bind different targets in vivo.
The existence of bi-fan arrays involving nonhomologous TFs suggests that TF duplication could also increase the frequency of these network features. For example, duplication of a TF that forms a regulatory complex would create two further bi-fan arrays, as depicted in Figure 2C. These network features appear as triplets of TFs that form bi-fan arrays with each other, and where two members of the triplet are related by WGD. The network includes 39 of these triplets, containing a total of 2.47 × 104 bi-fan motifs.
The statistical significance of the triplets of bi-fan arrays involving a pair of TFs originating from WGD can be computed by constructing a null model where the 442 bi-fan arrays are fixed and the 17 WGD relationships are added randomly to the network. This approach can then be used to compare the frequency of these network topologies to that in a large number of randomized networks. The expected number of triplets in the random model is 2.96 with p < 10−6, demonstrating that these network features are a statistically significant property of the network. Further details are provided in Figure S2. Since the WGD duplications occurred simultaneously [14], can be identified with high confidence [15], and were not succeeded by a large number of subsequent duplications [17], it is possible to assign half of the bi-fan motifs in these arrays to trans-regulatory interactions that were conserved after gene duplication. This accounts for a further 9.9% of the bi-fan motifs, and suggests that almost one-fifth of the motifs in the 442 bi-fan arrays can be attributed to a single WGD event.
A notable feature of the TFs duplicated by WGD is their very similar consensus DNA-binding specificities. Examples include the TFs MSN2p and MSN4p, which bind the stress response element AGGGG [25] and the leucine zippers YAP1p and YAP2p, which both bind the canonical sequence TTAGTCAGC. These are not isolated examples; almost all pairs of TFs that originate from WGD have similar DNA-binding motifs where these are known [10]. It is therefore not surprising that binding cross-reactivity causes duplicated TFs to occupy similar sets of promoters with the associated conservation of common bi-fan arrays. A more pertinent question is therefore which physical mechanisms enable these TFs to bind different targets in vivo.
The most likely mechanism for the divergence of promoter occupancy is that one of the duplicated TFs binds DNA cooperatively with another TF or cofactor via protein–protein interactions [26] or the modification of chromatin structure [23]. The second TF, which lacks such an interaction, cannot bind these promoters with high affinity. A specific example is provided by the forkhead TFs FKH1p and FKH2p, which bind overlapping sets of promoters and have identical DNA-binding preferences in vitro. It has been shown experimentally that differential promoter occupancy is achieved in vivo by FKH2p binding DNA cooperatively with the second TF, MCM1p [27]. This process is recapitulated by our analysis, which indicates that FKH2p forms a bi-fan array with MCM1p, but that this interaction is not shared by FKH1p. Our analysis also implicates the cell-cycle regulator SWI6p as being involved in creating the differential promoter occupancy between the two forkhead TFs.
The processes by which the TFs diverge in promoter binding propensities can be understood in terms of conventional models for the functional divergence of gene duplicates [28,29]. Immediately after duplication, the derived TFs are involved in an identical set of bi-fan arrays to the ancestral TF. The gain of an interaction that enables cooperative DNA-binding in one member of the pair is known as neofunctionalization, with subfunctionalization involving the loss of such interactions, depicted in Figure 2D. Of the two mechanisms for functional divergence, subfunctionalization is likely to be the dominant source of binding diversity, since the loss of a protein interaction may involve only a few degenerative mutations in one of the TFs, whereas gain requires formation of a novel interaction and subsequent accumulation of target genes [28–30]. This is supported by the rates of sequence evolution [15] in duplicated TFs. In the two pairs of whole-genome–duplicated TFs that have accelerated evolutionary rates compared with their K. waltii orthologue (the cell-cycle regulators FKH1p and FKH2p, and the stress response genes SKN7p and HMS2p), the faster-evolving proteins are involved in bi-fan arrays with fewer partner TFs than the more slowly evolving paralogue (see Table S3).
In summary, many bi-fan motifs in the Saccharomyces TFN originate from WGD. We have provided evidence that the functional divergence of duplicated TFs, which is likely to be involved in the generation of novel expression states, can be understood in terms of the patterns of gain and loss of bi-fan motifs within the overall structure of the network. The following section investigates the influence of WGD on the formation of FFL motifs.
Having suggested putative evolutionary models for the formation of bi-fan motifs in the S. cerevisiae TFN, we now turn our attention to the FFL. Although the FFL has a topology that appears distinct from the bi-fan motif, the presence of bi-fan arrays suggests another simple mechanism for formation of large numbers of FFL motifs. This process is depicted in Figure 3. In total, there are 43 stastically significant bi-fan arrays that form at least one regulator–regulator interaction, accounting for a total of 1,773 (61.2% of the total) FFL motifs in the TFN. Since these pairs of transcription regulators are expected to be involved in only 36 FFLs, these network features are sufficient to explain the deviation from the null model. The yeast WGD data indicate that four FFL arrays arise directly from WGD containing 334 (18.8%) FFL motifs. A further 11 FFL arrays, containing 299 (16.8%) FFL motifs, involved one of the bi-fan arrays conserved after TF duplication. In none of these cases were the FFL-forming interactions conserved between duplicated TFs.
We investigated whether FFLs were a statistically significant feature of the network given its bi-fan structure by randomizing edges between transcription regulators while holding interactions between transcription regulators and nonregulators constant (see Methods). This procedure fixes the vast majority of edges present in bi-fan arrays but involves rewiring of the regulatory interactions between TFs that could give rise to FFLs. Table 1 and Figure 4 show that the FFL topology remains statistically significant under this null model. Figures 4 and 5 show the frequencies of FFLs and bi-fan motifs as pairs of directed edges are swapped randomly, and demonstrate the sensitivity of the number of FFLs to rewiring of a small number of regulator–regulator interactions. Figure 5 confirms that the number of bi-fan motifs is affected only weakly by randomization of interactions between transcription regulators.
The majority of FFL motifs in the yeast TFN result from one or two direct regulatory interactions existing between TFs that form a statistically significant bi-fan array. Although experiments involving randomization of edges between TFs while other parts of the network are fixed suggest that the FFL motif remains overrepresented in natural networks, independently of the presence of bi-fan arrays, it is also possible that the FFL-forming edges could arise from some other nonselective process such as gene duplication. To investigate this question, we used a generalized linear model [31] to fit the probability of a directed regulatory interaction between TF, a, and a second TF, b, as a function of several local network properties (see Methods for full list). This statistical model was used to identify the network variables that are informative in predicting whether such an interaction occurs.
The final model indicates that the probability of forming a regulatory interaction increases with the out-degree of node a and the number of targets shared by the pair of TFs (i.e., the size of the bi-fan array), but that interactions are suppressed if the second TF b directly (auto-) regulates its own transcription. Figure 6 shows a measure of the error of optimized linear models involving subsets of these variables, and indicates that the out-degree has the greatest influence on the probability of forming a regulator–regulator interaction. This would be expected under a neutral model; however, the importance of the second term indicates that there is a propensity toward formation of FFLs from bi-fan arrays in the yeast TFN. This supports there being positive selection toward formation of the FFL motif and the signal-processing properties associated with this topology [32].
The previous sections have demonstrated that network motifs are typically organized in larger structures that are likely to have originated from two specific growth models. In this section, we investigate whether network motifs originating from duplication of TFs also contribute to more global properties of the network such as its overall modularity [33]. This feature of the TFN was investigated by using a divisive algorithm for partitioning the network into densely connected groups of nodes, which constitute modules, with sparser connections between groups [34]. The network was partitioned into 18 modules with an overall modularity score Q = 0.50, which suggests significant community structure [33].
The dendrogram in Figure 7 shows a representation of the division path of the algorithm and enriched functional annotations associated with all genes in the extant modules (see Text S1). The algorithm defines a hierarchy of modular structures, with the more “coarse-grained” solutions also representing relevant network structures [34]. In this case, the five coarsest granularity partitions represent the broad functional classes of small molecule transport, cell cycle/reproduction, protein synthesis, protein degradation, and metabolism. Figure 7 also shows enrichment of structural families within each module, and indicates that members from several structural families of DNA-binding protein are not distributed uniformly.
The most recent WGD in Saccharomyces can be used to investigate whether duplicated TFs diverge from the ancestral network module, and whether the duplication has contributed to the overall modularity of the network. This latter property is quantified by calculating the change in the modularity upon deletion of each node, which allows identification of modular (ΔQ > 0) and nonmodular TFs. Of the 15 pairs of TFs where both members bind a significant number of promoters under the conditions assayed by Harbison et al., 11 are members of the same module (p < 0.01 under permutation of module labels). In nine of the pairings, both TFs contribute positively to the modularity of the network, suggesting that gene duplication is involved in the formation of modular networks (the scores are tabulated in Table S3).
There are three further pairs of duplicated TFs in which the sign of ΔQ differs between the duplicates, and in which the membership of bi-fan arrays has diverged asymmetrically. If subfunctionalization, which in this context involves the loss of common bi-fan arrays, is the dominant source of functional divergence [30], these examples suggest that the TF that retains the majority of the ancestral functions remains a global (nonmodular) regulator, and that the mutations lead to specialization of its duplicate. Interactions between TFs that lead to creation of FFL arrays also tend to increase network modularity, since the majority (31 out of 43) involve intramodule connections (p < 0.01).
We have shown that the overrepresentation of bi-fan motifs in any directed network is associated with bi-fan array structures rather than individual network subgraphs. This property has been observed empirically in the original article describing network motifs in E. coli, which showed that bi-fan motifs are organised in dense overlapping regulons which consist of small numbers of TFs and operons that have particularly dense connectivity, and which also have few connections to the rest of the network [3]. Other work in E. coli has shown that clustering individual bi-fan motifs by overlap of any of their components leads to recovery of the network's largest fully connected component, and that a similar property can be observed for FFLs [16].
Many of the bi-fan arrays and the motifs within them can be attributed to the WGD event that occurred recently in the evolution of Saccharomyces, with the overwhelming majority of these structures arising from duplication of TFs. These represent a subset of the duplicative bi-fan arrays within the network, suggesting that many more of these network structures may also arise from divergent mechanisms of network evolution. It is possible that structural or sequence similarity could be used to detect more complex bi-fan architectures arising from ancient TF gene duplications. However, this is complicated by the rapid sequence divergence of TFs [15,17,35] and the potential for a particular network topology to be created by several alternative combinations of TF duplication and edge rewiring. It is clear, however, that the TFs arising from WGD have a larger number of shared targets and conserved network motif properties than more ancient duplicates. An outstanding question is whether this property is caused solely by the late occurrence of WGD in Saccharomyces or is also affected by the different effects of gene dosage in single-gene duplication and WGD events [36].
Although many bi-fan arrays originate from TF duplication, there is evidence that this topology also arises from environmental selection via the accumulation of DNA-binding motifs in promoter regions [22] or protein–protein interactions between TFs [8,24]. A mixture of these two effects is known to be a feature of mechanisms for combinatorial control of gene expression [26,37]. This article has also provided evidence that the cooperative binding of TFs to DNA is also likely to be involved in creating the functional divergence of duplicated TFs, as depicted in Figure 2C–2D. This mechanism may be particularly important for enabling increases in regulatory complexity to occur in unicellular organisms where redundant duplicate proteins cannot persist in the genome as a result of genetic drift [38], and consequently the fixation rate of single-gene duplications is very low [17].
The analysis of target genes indicates that the conservation of the TFs bound to duplicated promoters is related to the rate of sequence divergence of their associated genes, independently of molecular clock–based assumptions of the age of the duplication event [39,40]. This analysis also demonstrates that the cis-conservation is typically low and is restricted either to recent duplicates or the small number of genes that are stabilised by gene conversion [15,17]. Target gene duplication does not therefore make a substantial contribution to the formation of network motifs in the yeast TFN, contrary to other studies of Saccharomyces TFN evolution [11].
The rapid divergence in the promoters of duplicate genes is in agreement with other studies showing that gene expression evolves much more rapidly than an organism's gene content [12,13]. This result provides an explanation for a recent study of motif evolution [41], which found that the protein constituents of individual network motifs do not tend to co-occur across several very divergent yeast species. It was thus suggested that the motifs themselves are nonconserved and therefore not critical to the functionality of the network. However, the rapid cis-changes presented in Table 2 and the presence of positive selection toward motif formation suggest that the motif structures may be present in the comparison genomes, although their identity is likely to have changed on these relatively long time-scales. This is supported by the convergent evolution of similar network structures across diverse organisms, such as that observed between the human embryonic stem cell regulators SOX2, OCT4, and NANOG [42].
FFL motifs arise from a small number of regulatory interactions between TFs that form statistically significant bi-fan arrays. Our analysis indicates that there is likely to be positive environmental selection for the high/low-pass filtering properties of the FFL motif [3, 32] independently of the bi-fan array topology. As a result, FFL motifs could act as both a source and a consequence of duplicative bi-fan arrays in the course of network evolution. An outstanding question concerns the chronology of FFL formation, as it is not clear to what extent the existence of an FFL-like topology accelerates the accumulation of target genes or whether FFLs arise from existing bi-fan array structures, as depicted in Figure 3.
The static representation of the yeast TFN, representing a union of DNA-binding interactions across numerous environmental conditions, can be partitioned into modules that represent specific biological functions. Some structural families of DNA-binding proteins are not distributed uniformly across the network modules and are also involved in a larger number of bi-fan arrays with members of their own family. There are two potential causes for this observation. The WGD data indicates that TFs duplicated by WGD tend to occupy the same network module and share far more common targets than more ancient duplicates. It is therefore possible that proteins within a particular family underwent lineage-specific expansions more recently than other families. This appears to be the case for the YAP TFs, of which between two and three TF pairs originate from WGD [15,43]. The other possibility is that constraints on the diversity of binding sites available to a particular family of TFs [44,45] lead to a slower divergence of promoter binding, as exemplified by the GATA-binding family of Zinc-finger TFs.
In summary, the TFN contains many features that reflect the evolutionary history of the organism (i.e., divergent evolution), suggesting that its structure does not necessarily reflect an optimal “design” [46], and that evolutionary constraints contribute to both the modularity and network motifs that are present in the network. However, there is also strong evidence for the involvement of natural selection in the formation of network motifs beyond the neutral duplication–divergence model. The motif concept also provides a framework for understanding the mechanisms that have enabled increases in regulatory complexity to occur in a simple eukaryote, and which are also likely to apply to higher organisms.
The TFN was generated using the original gene-mapped ChIP-on-chip data from Harbison et al. [10]. The raw binding profiles were thresholded at a p-value of 10−3. TFs were classed as bound to an intergenic region if the binding profile was below the threshold in any of the assays carried out under alternative growth conditions. This included around 11,000 unique interactions between regulators and promoter regions.
Randomization of the networks was carried out using modified versions of the two algorithms used in [3,8]. Both these methods ensure that the networks' degree distributions remain unchanged by fixing both kin and kout for each node [47] while randomly rewiring edges. One of the algorithms involves repeatedly swapping nonisomorphic pairs of directed edges until the network is sufficiently randomized. The second algorithm involves specifying a set of in and out stubs for each node. Directed edges are then added from each out stub to a randomly selected in stub while again preserving the networks' in- and out-degree distributions. The two algorithms for generating null networks were found to produce identical results, provided that a sufficient number of iterations were carried out in the edge-swapping algorithm.
The number of bi-fan motifs within the TFN, fbi-fan, can be rewritten in an alternative form, which suggests that this particular motif is, in general, associated with array structures such as that shown in Figure 1
where the summations are over the NT TFs, or nodes with nonzero out-degrees, and where k(xi, xj) is the number of targets shared by TFs xi and xj. Equation 2 implies that for bi-fan motifs to be overrepresented in the network, there must be pairs of TFs (xi, xj) that have a greater number of shared targets than under an equivalent null model of the network.
The standard approaches to generating null network models [3,8,47] involve randomization of directed edges while preserving the in- and out-degree of each node. This null model provides an additional constraint on Equation 2
where kiin is the in-degree of node i and N is the total number of nodes in the network. Intuitively, Equation 3 represents the frequency of “mono-fans” in the network (i.e., two TFs binding to the same target). The left-hand side of Equation 3 represents the frequency of “mono-fans” in terms of the number of shared targets for each pair of TFs, which may vary in different randomizations of the network. The right-hand side represents this quantity in terms of the (fixed) in-degree sequence.
The constraint in Equation 3 indicates that a high degree of overlap for a subset of the TFs, required for overrepresentation of bi-fan motifs, implies a lower number of shared targets for other pairs of TFs. This suggests that bi-fan motifs are characteristic of networks with a modular or community structure [3,33].
Bi-fan arrays were identified by searching for pairs of TFs with a number of shared targets that exceeded the number found in 9,995 of the randomizations of the network. Figure 8 indicates the number of bi-fan arrays identified at the highest significance thresholds. Since there are a total of 176 TFs with kout ≠ 0 in the ChIP-on-chip dataset [10], there are a total of 1.54 × 104 comparisons. A total of 595 arrays were recovered at this threshold, with an expected number of 15.4 for a random network.
The number of targets shared by pairs of TFs in the randomized networks is well approximated by a Poisson distribution, which was used to estimate p-values for the bi-fan arrays identified to be significant from the bootstrap estimates (see Text S1). A total of 442 of the bi-fan arrays were significant at the threshold, which is the stringent threshold used in further analyses. A total of 297 bi-fan arrays were found at the p < 0.05 threshold after a Bonferroni correction for the multiple hypotheses tested.
The Pfam domain assignments were verified using the Saccharomyces Genome Database (http://www.yeastgenome.org), which also provided annotations for three additional TFs (INO4p, XBP1p, and CUP1p) that were missed by Pfam. The basic leucine zipper predictions were manually subdivided into the YAP and AP-1 families using definitions from the literature [48]. The two largest families of TFs in yeast, the classic Zinc-finger and the Zn-Cys binuclear cluster domain, are short, ancient domains that typically form one of many contact points between the TF and DNA [49,50]. Consequently, the shared presence of these domain types is not necessarily indicative of recent divergence or similar DNA-binding specificity. These families were therefore subdivided using sequence clustering. The BLASTclust program was used with sequence identity set to 25% and the alignment length parameter set to 0.25. This procedure may result in more distant duplicates being missed but increases the statistical significance of any homologous bi-fan arrays identified from analysis of the yeast TFN (groupings can be found in Text S1).
Several generalized linear models [31] were used to fit the probability of a regulatory interaction between a pair of TFs, f(πi), as a function of local network properties.
where xi = [x1, x2, …, xj] is the vector of network properties, β and α are the parameters of the model, and f(·) is the link function. Several link functions, including linear, logistic, and log–log, were compared using the deviance and the Hosmer-Lemshow criterion [31]. The log–log model provided the best fit under both measures and was used to model the full set of network variables.
The initial set of variables were the out-degree of node a, kaout, the out-degree of node b, the number of targets shared by the pair of TFs, kabarray, the expected number of shared targets, and binary variables representing a feedback or autoregulatory interaction at node a, autoregulation at node b (kbauto), transcription regulation of node a by node b, homology, and genome duplication. Backward stepwise elimination was then used to remove uninformative variables (see Text S1 and Figures S3 and S4 for further details), and resulted in the following model,
indicating that the probability of forming a regulatory interaction between TFs increases with the out-degree of node a and the number of targets shared by TFs a and b. Conversely, interactions are suppressed if the second TF b directly regulates its own transcription.
The modularity of the network is defined using the criterion Q, which is defined for undirected networks, but can be applied to the Saccharomyces TFN by considering each edge as undirected [33],
where the sum is over the number of identified modules, Nm, L is the number of edges in the network, ls is the number of intramodule edges, and ds is the sum of the degrees of the nodes in module s. Intuitively, a cluster contributes a large ΔQ to the network's overall modularity if the number of intramodular connections is much larger than the number expected in an equivalent network with edges placed at random (a null model that corresponds exactly to the randomization procedures used in this article [47]).
The standard approach to module identification is to seek a partition of the network such that the modularity, ΔQ, is maximised. In this study, a spectral module detection algorithm [34] is used, which involves solving a series of eigenvector problems on a characteristic modularity matrix. The algorithm divides the network recursively into disjoint binary partitions until no further increase in the modularity is recovered. The division of the network can then be used to calculate the sensitivity of Q to the deletion of nodes from the network, ΔQ. |
10.1371/journal.pntd.0006522 | Immunogenicity and efficacy following sequential parenterally-administered doses of Salmonella Enteritidis COPS:FliC glycoconjugates in infant and adult mice | In sub-Saharan Africa, invasive nontyphoidal Salmonella (iNTS) infections with serovars S. Enteritidis, S. Typhimurium and I 4,[5],12:i:- are widespread in children < 5 years old. Development of an efficacious vaccine would provide an important public health tool to prevent iNTS disease in this population. Glycoconjugates of S. Enteritidis core and O-polysaccharide (COPS) coupled to the homologous serovar phase 1 flagellin protein (FliC) were previously shown to be immunogenic and protected adult mice against death following challenge with a virulent Malian S. Enteritidis blood isolate. This study extends these observations to immunization of mice in early life and also assesses protection with partial and full regimens. Anti-COPS and anti-FliC serum IgG titers were assessed in infant and adult mice after immunization with 1, 2 or 3 doses of S. Enteritidis COPS:FliC alone or co-formulated with aluminum hydroxide or monophosphoryl lipid A (MPL) adjuvants. S. Enteritidis COPS:FliC was immunogenic in both age groups, although the immune responses were quantitatively lower in infants. Kinetics of antibody production were similar for the native and adjuvanted formulations after three doses; conjugates formulated with MPL elicited significantly increased anti-COPS IgG titers in adult but not infant mice. Nevertheless, robust protection against S. Enteritidis challenge was seen for all three formulations when three doses were given either during infancy or as adults. We further found that significant protection could be achieved with two COPS:FliC doses, despite elicitation of modest serum anti-COPS IgG antibody titers. These findings guide potential immunization strategies that may be translated to develop a human pediatric iNTS vaccine for sub-Saharan Africa.
| Non-typhoidal Salmonella enterica (NTS) serovars Enteritidis and Typhimurium (including monophasic variant I 4,[5],12:i:-) are significant causes of invasive bacterial disease amongst infants and toddlers in sub-Saharan Africa, and currently, there are no approved NTS vaccines. We have demonstrated previously that immunization with S. Enteritidis core and O-polysaccharide (COPS) conjugated to the flagellin protein (FliC) from the homologous serovar protected adult mice from fatal infection with a Malian S. Enteritidis blood isolate. The target population for iNTS vaccines in sub-Saharan Africa, however, are young infants. In the current study, we evaluated S. Enteritidis COPS:FliC vaccination during murine infancy or adulthood. We found that COPS:FliC was immunogenic in both adult and infant mice and that co-formulation with adjuvant impacted the magnitude and quality of the immune response. Despite these differences, all vaccine formulations protected against experimental challenge in both age groups. Furthermore, robust efficacy was attainable after only two COPS:FliC doses, coinciding with the appearance of COPS-specific antibodies. The results from this study suggest that S. Enteritidis COPS:FliC is a promising pediatric vaccine candidate for use in sub-Saharan Africa and may help inform potential immunization strategies for iNTS COPS:FliC conjugate vaccines.
| In sub-Saharan Africa, hospital-based blood culture surveillance of febrile pediatric admissions has revealed that invasive non-typhoidal Salmonella enterica (iNTS) infections caused by serovars Typhimurium, I 4,[5],12:i:-, and Enteritidis are widespread among children less than 5 years of age [1, 2]. Recent studies estimate an annual incidence of 200–400 cases per 100,000 child years in some areas, accompanied by case fatality rates of 12 to 30% [3, 4]. Genomic and phenotypic analyses have revealed several unusual traits for sub-Saharan iNTS isolates including the predominance of multi-locus sequence types not typically found in North America and Europe, gene loss in a manner analogous to typhoid and paratyphoid fever serovars, and diminished inflammatory activity in cell culture and animal models [2, 5–8]. Elucidation of the reservoir of infection and the predominant modes of iNTS transmission in sub-Saharan Africa has proved elusive, hampering efforts to implement environmental control interventions. Development of an effective iNTS vaccine for sub-Saharan Africa, therefore, remains an important public health priority, and is considered epidemiologically and immunologically feasible given the predominance of only three serovars and the established efficacy of typhoid fever vaccines as a precedent [9, 10]. Although predominantly an intracellular pathogen, Salmonella are susceptible to antibodies during extracellular periods prior to invading host cells or following release from infected cells.
The putative role of humoral immunity in protection against iNTS disease is supported by several important observations. Among children < 5 years of age the bulk of disease burden is found in those less than 2 years old, with peak onset occurring after 5 months of age, the point at which maternally-derived placental IgG antibodies have waned [1, 11]. Results from age cross-sectional studies of invasive S. Typhimurium infections among infants, toddlers and pre-school children in Malawi documented a direct relationship between disease incidence and serum bactericidal activity (SBA), wherein the drop in SBA titers during the first 8 months of life mirrored the rise in iNTS infections [12, 13]. Remarkably, the decline in disease incidence by 35 months of age was directly correlated with an increase in SBA titers to those present in newborns. A positive correlation was also found between SBA titers and antibody levels against S. Typhimurium lipopolysaccharide (LPS), which is the surface polysaccharide in this serovar [12].
Support for the protective capacity of anti-LPS antibodies also comes from preclinical studies in animal models of iNTS infection, whereby protection has been achieved when anti-O polysaccharide (OPS) antibodies were delivered by passive transfer or induced by active immunization [14–17]. As such, there is marked interest in the development of OPS-based vaccines to prevent iNTS infections. Whereas isolated bacterial polysaccharides generally fail to induce robust and durable antibody responses in children less than 2 years old [18], covalent linkage to protein carriers has improved the immune response to these antigens, and enabled the development of carbohydrate vaccines to prevent invasive bacterial infections in human infants (e.g., for Streptococcus pneumoniae, Haemophilus influenzae type b, Neisseria meningitidis). We previously reported the development of a S. Enteritidis core-OPS (COPS) glycoconjugate that utilized the homologous serovar phase 1 flagellin (FliC) as the carrier protein and documented efficacy of this vaccine against fatal invasive infection in adult mice experimentally challenged with a S. Enteritidis blood isolate from a Malian child [15, 19]. While adult mouse models are advantageous for proof-of-principle studies, they may not optimally approximate the vaccine-induced immune responses in young human infants that are the target population for this vaccine. In order to assess further the immunogenicity of S. Enteritidis COPS:FliC in the context of early life immunization, we evaluated herein immune responses and efficacy when vaccination was initiated either during early murine life or adulthood, and when formulated with two different adjuvants. Additionally, as we previously found that the appearance of anti-FliC IgG preceded detection of IgG anti-COPS antibodies [15, 19], we also assessed whether protection could be achieved with only one or two vaccine doses rather than the full 3-dose regimen.
The strains used in this study are detailed in S1 Table. Growth conditions for wild-type S. Enteritidis R11, genetic mutant R11 ΔfliC, and attenuated derivative CVD 1943 (ΔguaBA ΔclpP ΔfliD) were described previously [20].
S. Enteritidis R11 fliC::kan was engineered by disruption of the fliC gene using the lambda red-mediated mutagenesis system as described previously [21, 22]. Disruption of fliC was confirmed by motility assay as described [20], and Western blot with an S. Enteritidis-specific monoclonal antibody (clone: CA6IE2) as described [23] (S1 Fig).
Purified COPS and FliC monomers for use as vaccine antigens for immunization and antibody measurements by enzyme-linked immunosorbent assays (ELISAs) were generated from S. Enteritidis reagent strain CVD 1943, as described [20, 23]. Endotoxin removal was confirmed by endpoint limulus amebocyte lysate assay with an Endosafe PTS system (Charles River, MA). Nucleic acid removal was assessed by A260 nm for COPS and Quant-IT Sybr Green assay (Life Technologies, CA) for FliC. Removal of host cell protein in the COPS preparation was confirmed with the bicinchoninic acid (BCA) assay (Thermo, MA) and by SDS-PAGE with Coomassie staining for FliC. S. Enteritidis COPS identity and molecular size were confirmed by Dionex HPAEC-PAD and HPLC-SEC respectively as described [16]. Flagellin identification was accomplished by SDS-PAGE/Western blot analysis with monoclonal antibody CA6IE2, and confirmation of monomeric form by HPLC-SEC as described [23].
Bioconjugation of COPS to FliC was performed essentially as described [16, 19]. Briefly, S. Enteritidis FliC monomers were suspended to 5 mg/mL in 100 mM MES / 0.1% Tween 20, pH 6.5. Adipic acid dihydrazide (ADH) was then added to 0.5 M and the reaction brought to 5 mg/mL N-(3-Dimethylaminopropyl)-N′-ethylcarbodiimide with incubation for 12–16 h at 4°C. The ADH-derivatized FliC was then dialyzed with 10 kDa molecular weight cutoff dialysis cassettes (Thermo, MA) against 3 x 2,000 volumes of 50 mM sodium tetraborate, 300 mM sodium chloride, 0.1% Tween 20 pH 9.15, and then concentrated to 15 mg/mL with 10 kDa Amicon Spin filters (Millipore, MA). S. Enteritidis COPS was brought to 10 mg/mL and precooled by incubation on ice. Activation at random polysaccharide hydroxyls was achieved by addition of 150 mg/mL 1-cyano-4-dimethylaminopyridinium tetrafluoroborate (CDAP) in acetonitrile at a ratio of 0.5 mg CDAP per mg polysaccharide. The pH was raised to pH 9.5 using dimethylaminopyridine as the base and maintained using 0.5 M sodium hydroxide as needed. The reaction was incubated for 5 min on ice, and then added to an equal amount of ADH-derivatized flagellin, and mixed by tumbling rotation for 2 h at room temperature and then 18 h at 4°C, at which point the reaction was quenched with 2 M glycine pH 9. Unreacted protein and polysaccharide were removed by size-exclusion chromatography using Superdex 200 resin (GE/Amersham, NJ) with an AKTA Purifier (GE Biosciences, NJ). Fractions that contained material of ≥100 kDa (verified by SDS-PAGE) were pooled and used for subsequent immunization. Polysaccharide and protein concentrations in the final conjugate preparation were assessed by the resorcinol-sulfuric acid method and BCA assay respectively [16].
All animal studies were performed in facilities that are accredited by the Association for Assessment and Accreditation of Laboratory Animal Care and were in compliance with guidelines for animal care established by the US Department of Agriculture Animal Welfare Act, US Public Health Service policies, and US federal law. All animal experiments were in compliance with study protocols (1114008) approved by the University of Maryland School of Medicine Institutional Animal Care and Use Committee.
Experimental CD-1 adult females and male/female breeders were purchased from Charles River Laboratories (MA). Pups were bred and raised in the animal facility at the University of Maryland, Baltimore.
Male and female infant (2 weeks old) or female adult (6–8 weeks old) CD-1 mice were immunized with 5 μg FliC or S. Enteritidis COPS:FliC (5 μg polysaccharide per dose) delivered in a 50 μL (infants) or 100 μL (adults) volume distributed across both gastrocnemii. Adjuvanted S. Enteritidis COPS:FliC formulations were generated as follows: for alum adsorptions, Alhydrogel aluminum hydroxide wet gel suspension (“alum”; Brenntag, Germany) was added to sterile-filtered S. Enteritidis COPS:FliC at a ratio of 28.6 mg Al per mg FliC (95% adsorption), and incubated on ice for 30 min with gentle mixing. The solution was pelleted (5 min at 10,000 x g), and the supernatant discarded to remove un-adsorbed S. Enteritidis COPS:FliC. Formulations of S. Enteritidis COPS:FliC with monophosphoryl lipid A (“MPL”; InvivoGen, CA) were generated by admixture with 1 μg of MPL followed by sterile filtration. In all experiments, control mice received sterile-filtered PBS. Serum was collected from the retro-orbital plexus or facial vein at baseline (for adult mice) and 12–14 days after each immunization unless otherwise indicated. For challenge studies, mice were infected intraperitoneally with 1 x106 CFU of S. Enteritidis R11 or R11 ΔfliC in 500 μL of sterile PBS. Animals were monitored daily for 14 days after infection. Weights were recorded daily, and any animal that appeared moribund (displaying lethargy or non-responsiveness, unkempt fur, hunched posture and/or ≥ 20% weight loss) was euthanized by CO2 inhalation and recorded as having succumbed to challenge.
(i) Antibody titer: Anti-COPS and anti-FliC immunoglobulin were measured by ELISA, as described previously [16]. In summary, medium binding, 96-well microplates (Greiner Bio-One, NC) were coated with 5 μg/mL of S. Enteritidis COPS or S. Enteritidis FliC for 1 hour at 37°C. Plates were blocked with 10% Omniblok (AmericanBio, MA) diluted in 1x PBS and incubated at 37°C for 2–3 hours. Serum samples were diluted in blocking buffer containing 0.05% Tween-20 (Sigma, MO) on a separate non-treated, microplate (Costar, NY) and were then transferred to the blocked ELISA plates and incubated at 37°C for 1 hour. Anti-COPS and anti-FliC immunoglobulin were detected with horseradish-peroxidase (HRP)-conjugated secondary antibodies specific for mouse IgG (SeraCare, MD) or IgG1, or IgG2b (Southern Biotech, AL) diluted 1:1,000 in blocking buffer containing 0.05% Tween-20; plates were incubated at 37°C for 1 hour. Tetramethylbenzidine substrate (“TMB”, SeraCare) was added to each well and incubated at room temperature, in darkness, with mild rocking for 15 minutes. The reaction was quenched with 1M phosphoric acid, and the absorbance at 450 nm was read using a Multiskan FC microplate photometer (Thermo, MA). End-point titers were calculated by interpolation of linear regression curves from a standard 7-parameter curve generated with pooled anti-S. Enteritidis hyperimmune mouse sera generated from S. Enteritidis COPS:FliC-immunized adult CD-1 mice that had convalesced from S. Enteritidis R11 challenge (University of Maryland, Baltimore). Titers were reported as the inverse dilution at which an OD450 of 0.2 over background was observed and represented in ELISA units (EU)/mL.
(ii) Avidity: Changes in anti-COPS and anti-FliC IgG avidity were assessed with urea as a chaotropic agent. For these analyses, the ELISA protocol described above was followed with the following exceptions: before the addition of secondary antibody, plates were washed three times and 200 μL of either 6M urea (Sigma, MO) or PBS were added to each well. Plates were incubated at room temperature with mild rocking for 10 minutes before washing another three times followed by addition of secondary antibody. Antibody titer in the presence or absence of urea was calculated as above. The avidity index was determined as follows: (IgG titer with 6M urea treatment / IgG titer with PBS) x 100. Values that were >100% were adjusted to 100%.
Statistical analyses were performed using GraphPad Prism v6 (GraphPad Software, CA), and no data points were excluded from analysis. For ELISA analyses, the majority of data did not meet the criteria for Gaussian distribution, and as such, statistical comparisons were accomplished using the non-parametric Mann-Whitney U test (two-tailed, α = 0.05). Adjustments for multiple comparisons were not made. Survival curves of challenged mice were compared by the log-rank test, and mean time till death (MTTD) for each group was calculated by averaging the days at which challenged mice succumbed to infection. Vaccine efficacy was calculated based off of the attack rate (AR) in control and vaccinated mice as follows: (ARcontrols−ARvaccinated)/ARcontrols) x 100. A two-tailed Fisher’s exact test (FET, α = 0.05) was utilized to compare the survival proportions between vaccinated mice and controls. P-values of ≤ 0.05 were considered statistically significant.
We previously reported that unadjuvanted S. Enteritidis COPS:FliC conjugates were immunogenic in adult mice, producing robust titers of anti-FliC IgG after one dose, whereas the highest anti-COPS IgG titers appeared after three immunizations [15, 19]. The immunological maturity of IgG antibody responses in 7–28 day old mice has been shown to approximate that of human infants, whereupon transition to an adult immune system occurs after approximately 28–35 days [24]. In order to compare immune responses to S. Enteritidis COPS:FliC in the context of an immature immune system and when formulated with an adjuvant, infant (2-week old) or adult (6–8 week old) mice were vaccinated with 3 intramuscular doses, administered at 2-week intervals, of either PBS or 5 μg S. Enteritidis COPS:FliC formulated alone or with alum or MPL adjuvant. S. Enteritidis COPS:FliC delivered alone or in the presence of either adjuvant was well-tolerated. Control mice administered PBS alone did not exhibit detectable levels of IgG antibody for either antigen throughout the experiment. Overall, the kinetics of appearance of serum anti-FliC IgG were similar between S. Enteritidis COPS:FliC alone and the different adjuvanted formulations. After the 1st immunization with S. Enteritidis COPS:FliC-containing formulations, all adult mice seroconverted to anti-FliC IgG with similar geometric mean titers [GMTs] (Fig 1A). Similarly, 100% of the infant mice immunized with S. Enteritidis COPS:FliC manifested measurable anti-FliC IgG after the priming dose, albeit with lower end-point titers and greater individual variation compared to adults. Anti-FliC IgG titers were essentially maximal after the 2nd dose for all groups. After three doses, the anti-FliC antibody levels for infant mouse groups were ~2- to 7-fold lower than that of the adults, but nevertheless robust (≥ 105 EU/mL) and comparable among all animals within each experimental group (Fig 1B).
With regard to anti-COPS IgG, endpoint titers and kinetics of seroconversion for both adults and infants immunized with S. Enteritidis COPS:FliC were lower relative to those directed against the FliC carrier protein (Fig 2A). Serum anti-COPS IgG was detectable in some infant and adult mice after the 2nd immunization, with higher titers occurring after the 3rd dose. Formulation with MPL significantly improved anti-COPS IgG levels after 3 doses in mice immunized as adults relative to unadjuvanted controls (P ≤ 0.05), with 47.4% of the adult mice receiving MPL formulated vaccine seroconverting to a titer of ≥ 50 EU/mL (4-fold over PBS) compared to 30% of mice in the unadjuvanted and alum groups (Fig 2B). By comparison, none of the adjuvant formulations enhanced the anti-COPS immune response in mice primed during infancy compared with unadjuvanted S. Enteritidis COPS:FliC.
To assess the quality of the serum antibody response elicited by COPS:FliC alone or formulated with an adjuvant, IgG isotype profiles and avidity indices in post-vaccination sera were evaluated. A step-wise increase in the FliC-specific IgG avidity index (AI) was observed after each immunization, and by the completion of the vaccine regimen, most vaccine groups displayed a relatively similar FliC-specific IgG AI (28.0%–38.7%) regardless of age (Fig 3A). The exception to this were infant mice immunized with alum-adsorbed S. Enteritidis COPS:FliC where an AI of 13.9% was found. IgG subclass analysis of FliC serum antibodies from adult- and infant-primed mice after 3 COPS:FliC doses revealed a mixed Th1/Th2 response characterized by the predominance of IgG1 relative to IgG2b (36.2–176.8-fold IgG1 bias), that was similar between mice immunized with unadjuvanted COPS:FliC or formulated with alum (Fig 3B). Adult-primed mice immunized with MPL-adjuvanted COPS:FliC manifested a more balanced Th1/Th2 profile (4-fold IgG1 bias) as evidenced by a greater contribution from the IgG2b subclass. An increase in IgG2b was also seen when MPL-adjuvanted COPS:FliC was administered to mice during infancy, which significantly decreased the IgG1:IgG2b ratio compared to the unadjuvanted vaccine but to a lesser extent than in the adult group (P ≤ 0.0005). Comparable analyses of anti-COPS IgG were limited by the lower IgG titers and number of seroconverting animals; only a subset of adult sera (titer ≥ 103 EU/mL) obtained after the 3rd dose of S. Enteritidis COPS:FliC formulated with MPL (Fig 4A and 4B) could be examined. These samples had a high AI (84.7%) and COPS-specific IgG was predominantly IgG1 with a wider range of variability in the ratio of IgG1:IgG2b compared to anti-FliC antibodies (Fig 4B).
We first confirmed the protective efficacy of the FliC carrier protein against infection with S. Enteritidis R11, a blood isolate from Mali. Protection against mortality was negligible, but there was a significant delay in time till death (TTD) after three doses in adult mice (Fig 5). Next, we determined the protective efficacy of S. Enteritidis COPS:FliC against lethal infection with S. Enteritidis R11 in mice immunized as adults or infants with the different vaccine formulations (Fig 6). Whereas 90%–100% of PBS control mice succumbed to fatal infection, adult and infant mice immunized with three doses of S. Enteritidis COPS:FliC conjugate were significantly protected from death (vaccine efficacy [VE] = 61.1%–76.5%; P ≤ 0.0005). Amongst S. Enteritidis COPS:FliC-immunized adult mice, the alum and MPL formulations induced a moderate increase in VE over unadjuvanted vaccine that was significant for the MPL group (P = 0.04) but not for the alum group (P = 0.13). In infant-primed mice, however, none of the adjuvanted formulations significantly improved VE relative to S. Enteritidis COPS:FliC alone.
As immunization with S. Enteritidis COPS:FliC induced robust titers of anti-FliC IgG but negligible anti-COPS IgG after a single dose, with the highest anti-COPS IgG GMT and seroconversion rate occurring after 3 doses, we wanted to assess whether protection could be achieved with fewer doses than the full 3-dose regimen. We chose to assess the MPL-adjuvanted vaccine for these analyses as this formulation induced the highest level of anti-COPS IgG. Accordingly, infant or adult mice were immunized with 1, 2, or 3 doses of PBS or S. Enteritidis COPS:FliC formulated with MPL, and then challenged with S. Enteritidis R11 (Fig 7). We found that 2 doses of S. Enteritidis COPS:FliC + MPL were required to achieve significant protection from challenge (VE = 65.0%–72.9%, P ≤ 0.0005) and that the level of protection at this time point was similar to the that observed after 3 doses for both adult and infant mouse groups. Additionally, although measurable protection against mortality was not achieved after a single COPS:FliC + MPL immunization, there was an increase in the TTD that was significant for the adult group (P ≤ 0.05). Analyses of the serum anti-COPS IgM and IgG titers from these cohorts revealed a pattern similar to that seen previously where titers increased after sequential doses, with higher titers in adults compared to infant-primed groups (S2 Fig). To determine the relationship between anti-polysaccharide antibody titer and survival after challenge, we combined groups that received either 1, 2 or 3 doses into a single cohort, as this permitted assessment of a wide range of serum antibody levels. This analysis demonstrated a higher IgM and IgG anti-COPS GMT among animals that were protected versus those that succumbed (S3A Fig). Furthermore, stratifying the anti-COPS IgG response by quartiles revealed progressively higher rates of survival after challenge as IgG titer increased (S3B Fig).
In order to further dissect the contribution to protection imparted by the early FliC-specific immune response induced by the conjugate vaccine, we immunized adult mice with one or two doses of S. Enteritidis COPS:FliC formulated with MPL and challenged with S. Enteritidis R11 ∆fliC, a phase 1 flagellin mutant (Fig 8). After receiving a single vaccine dose, the majority of vaccinated mice succumbed to challenge. Administration of two doses of S. Enteritidis COPS:FliC + MPL, however, afforded significant protection against S. Enteritidis R11 ∆fliC challenge (VE = 47.4%, P ≤ 0.005), which was accompanied by a delay in TTD (P ≤ 0.005).
Analyses of the immune response of humans following natural exposure to S. Typhimurium in early life suggest that protection against invasive disease develops in parallel with the appearance of titers of bactericidal antibodies directed against the surface lipopolysaccharide [11, 12]. It is thus presumed that immunization with an OPS-based iNTS vaccine during early infancy may confer protection during and after the latter half of the 1st year of life, the period of highest vulnerability to invasive iNTS disease among infants and toddlers living in sub-Saharan Africa [1]. As part of a program to develop a vaccine to prevent iNTS disease in human infants in Africa, we have previously documented that S. Enteritidis and S. Typhimurium COPS:FliC glycoconjugate vaccines are immunogenic and protective in adult mice [15, 16, 19]. The work described herein demonstrates that our candidate S. Enteritidis COPS:FliC vaccine is also immunogenic and can impart functional protection when immunization is initiated during murine infancy.
The scalable polysaccharide and flagellin purification protocols used herein generate highly pure preparations with essentially undetectable levels of residual host protein, endotoxin and nucleic acid [16, 23]. Additionally, we have documented that conjugation to polysaccharides ablates the propensity of flagellin to activate TLR5 [19]. The presence of residual TLR stimulating impurities has been hypothesized to enhance the immunogenicity of some commercial polysaccharide preparations [25]. To determine whether addition of an immuno-stimulatory adjuvant would enhance vaccine-induced immune responses to S. Enteritidis COPS:FliC in mice, we assessed co-formulation with different compounds that mediate adjuvanticity by different mechanisms. This included aluminum hydroxide, a known activator of the inflammasome that has a long track record of safety in human pediatric vaccines [26], or chemically-detoxified Re LPS from Salmonella Minnesota (MPL), a TLR4 agonist that is used as part of the adjuvant formulation in several licensed adult vaccines [27] as well as for 6 to 12-week old children receiving the RTS,S malaria vaccine [28].
In both infant and adult mice, we demonstrated 100% seroconversion and a similar prime-boost antibody response to the FliC carrier after immunization with S. Enteritidis COPS:FliC. Consistent with prior observations in adult mice, significant rises of anti-COPS IgG in infant or adult mice were generally apparent only after the 2nd or 3rd dose with more than one non-responder (by ELISA) in each group [19]. Formulation with alum did not have a measurable effect on anti-COPS or anti-FliC IgG levels or isotype profiles. Unexpectedly, we found that MPL enhanced the anti-COPS immune response in adult mice and altered the Th1/Th2 profile for anti-FliC IgG favoring the production of IgG2b in adult and infant-primed mice. While anti-COPS IgG avidity analyses were limited due to the lower number of responding animals, we did note robust avidity for anti-COPS antibodies in adult mice that received three doses of the MPL formulation. The reasons for the modest and variable anti-COPS responses observed in this study remain unknown. However, our results are similar to those reported for S. Enteritidis COPS conjugates with CRM197 which showed similarly low and variable serum post-vaccination anti-COPS IgG levels despite the use of a different carrier protein, conjugation chemistry, and mouse strain [29, 30]. One possibility for the variable antibody responses is that the antibody-antigen interactions in a portion of the polyclonal anti-polysaccharide antibody population are disrupted by the standard ELISA conditions and reagents used herein. This possibility will be addressed in future studies.
Given the differential kinetics of antibody induction by the carrier and hapten, we further assessed whether protection could be attained prior to onset of peak anti-polysaccharide immunity. Immunization with three doses of FliC alone conferred poor protection against mortality but significantly delayed the TTD in adult mice challenged with wild-type S. Enteritidis R11. This is consistent with prior studies where we found that passive transfer immunization in adult mice with a monoclonal IgG1 specific for S. Typhimurium FliC imparted partial protection against mortality and a significant increase in the TTD following challenge with a virulent S. Typhimurium Malian blood isolate [31]. Immunization with a single dose of COPS:FliC conjugate similarly did not significantly protect against mortality but produced a slight delay in TTD for mice infected with S. Enteritidis R11. Since immunization with unconjugated FliC alone prolonged survival but did not significantly protect against mortality (in our challenge conditions), poor induction of COPS-specific immunity may account for the lack of discernable protection by a single priming immunization with the conjugate. The similarly robust level of protection seen between adult mice immunized with two COPS:FliC + MPL doses and challenged with wild-type S. Enteritidis R11 or the ΔfliC mutant suggests that despite the low level of anti-polysaccharide serum IgG, protection at this intermediate point may be mediated primarily by anti-COPS antibodies.
As challenge occurred during adulthood for both groups, we sought to establish a putative relationship between serum anti-polysaccharide antibody levels and protection after challenge. We found that the anti-COPS IgM and IgG GMT in sera from mice immunized with COPS:FliC was higher in survivors compared to mice succumbing after S. Enteritidis challenge. Additionally, acquisition of successively higher serum anti-COPS IgG levels was associated with a proportionate drop in mortality post challenge whereby a COPS-specific IgG GMT >200 EU/mL was associated with mortality in less than 1/3rd of mice. This threshold aligns with the range of the anti-COPS IgG GMT in infants and adults after two doses of COPS:FliC where we found 25–35% mortality after challenge. These results are also similar to our prior findings where COPS IgG induced by COPS:CRM197 vaccination in adult mice was associated with protection against lethal S. Typhimurium infection [16]. In this as well as the previous study, several outliers were noted wherein some animals with high anti-COPS IgG levels succumbed while others with low or undetectable titers survived. This could be reconciled by several possibilities. As discussed above, it is possible that some anti-COPS antibodies are not detected with our ELISA method. Additionally, the genetic variability in the outbred CD-1 mice used herein may also influence susceptibility to challenge and the associated protective antibody threshold among vaccinated animals with similar IgG titers. As determination of vaccine efficacy relies on group responses (i.e., measuring the proportionate reduction in the attack rate for the vaccine group versus controls), comparison of anti-COPS GMTs with proportionate survival allows for a better analysis of a possible association between vaccine-induced antibodies and protection despite the presence of outlier individual responses. Confirmation of these findings by passive transfer with vaccine-induced sera should be conducted in future studies.
The FliC-specific IgG GMT in immunized infant mice was lower than in adult mice. Nonetheless, infant anti-FliC IgG displayed an adult-like avidity and Th1/Th2 profile after three immunizations. In both adult and infant mice, the avidity of anti-FliC IgG increased sequentially after each immunization in a manner that was not directly related to the fold change in titer. These data suggest that the magnitude and avidity of serum anti-FliC IgG may not be linearly associated. Changes in avidity may reflect either a B cell intrinsic factor (e.g., affinity maturation, or the nature and diversity of the paratope repertoire), or the influence of T cell subsets important for immune maturation. Our findings for the reduced antibody responses to COPS:FliC in infant CD-1 mice can be reconciled by several possible explanations, related to the immunomodulatory effects of the early life immune system [24, 32]. Studies in neonatal mice (0–7 days old) have demonstrated that a combination of decreased co-stimulatory molecule expression coupled with immature interactions between dendritic cells, T cells and B cells may act in concert to delay the induction and limit the magnitude of germinal center responses [32]. This could result in suboptimal class switch recombination, plasma cell formation and a reduction in IgG responses to T-dependent antigens, such as glycoconjugate vaccines. It is therefore possible that the reduced anti-FliC and anti-COPS humoral immunity observed in infant mice (7–28 days old) may reflect suboptimal priming of naïve B cells or incomplete germinal center formation and feeble B cell memory. Furthermore, differences in the frequency of S. Enteritidis FliC- or COPS-specific B cells within the primary repertoire of infant and adult CD-1 mice may offer another possible explanation for the diminished prime-boost effect observed in infant mice. With glycoconjugate vaccines, helper T cells induced by the carrier protein are thought to promote hapten-specific B cell responses. Using a monovalent pneumococcal conjugate vaccine (Pnc1-TT) in mice, Jakobsen et al. demonstrated that polysaccharide-specific antibody responses increased with age, and that a weak and skewed carrier-specific T cell response generated after immunization of neonates or infants correlated with reductions in anti-polysaccharide IgG titers and protection against pneumococcal infection [33]. While COPS:FliC-induced cell-mediated immunity was not measured in this study, we speculate that the FliC-specific T cell response generated after infant priming, which may display differences in magnitude and/or breadth with respect to adult mice, contributes to the modest COPS-specific antibody response in infant mice.
Antibody responses to bacterial polysaccharides are generally limited in human infants where covalent linkage of polysaccharides to a protein carrier typically improves their immunogenicity. One proposed mechanism accounting for the poor responses of rodent and human infants to isolated polysaccharides is the lack of splenic marginal zone (MZ) B cells, which appear in the first 2–3 weeks and 2 years of life in rodents and humans, respectively [34], and are thought to be important components of the immune response to polysaccharides [18]. Interestingly, in adult mice, TLR4 stimulation has been shown to selectively promote MZ B cell proliferative responses and plasma cell differentiation [35]. By contrast, B cells from murine neonates secrete anti-inflammatory cytokines (e.g., IL-10) in response to TLR4 ligands and can suppress dendritic cell production of IL-12, a pro-inflammatory and Th1-promoting cytokine. This phenomenon is weakly evident, however, in B cells isolated from 6–8 week-old adult mice [36]. In the present study, considering that the first immunization occurred during early murine infancy (i.e., 2 weeks of age), it is conceivable that either the absence of MZ B cells or an anti-inflammatory cytokine milieu created by MPL-activated B cells at the time of priming could account for the differential adjuvant response relative to that seen in adults. We found similarly that the 3rd immunization with S. Enteritidis COPS:FliC with MPL, delivered during adulthood (i.e., 6–8 weeks of age), was unable to drive adult equivalent anti-FliC IgG2b production in the infant-prime group. The bias toward IgG1 among vaccinated murine infants may indicate the presence of IgG1+ memory B cells and/or Th2-commited memory CD4+ T cells established during infant priming and preferentially expanded after subsequent immunizations. The importance of individual FliC-specific IgG subclasses, or combinations thereof, to clearance of Salmonella in vivo is not well-understood however. Using recombinant humanized IgG specific for a peptide-tagged FliC, Goh et al. demonstrated that IgG3 was superior at mediating in-vitro OPA of S. Typhimurium, although all IgG subclasses were opsonic to varying degrees [37].
Our findings demonstrate that S. Enteritidis COPS:FliC can induce protective antibody responses in mice when immunization begins during infancy. An important caveat to our immunization model is that while adult mice were immunized throughout adulthood, the early life group received their 1st dose as infants (2 weeks), the 2nd dose during a theoretical transition between infancy and adulthood (4 weeks) and the final dose during adulthood (6 weeks). We opted to forego an accelerated schedule here (e.g., 3 doses at weekly intervals) as this could conceivably have dampened seroconversion and the magnitude of antibody responses [38–40]. Furthermore, as protection against challenge was assessed 28 days after the final dose in order to permit maturation of adaptive immunity and resolution of the innate immune response, this further precluded assessment of vaccine efficacy during infancy. Future studies should address the role of anti-COPS antibodies in early life protection of neonates and young infants (≤14 days old) and may be accomplished by passive administration of COPS:FliC-induced antibodies. The observation that measurable protection can be achieved after only 2 doses, could also inform possible analogous vaccine schedule strategies in infants. Delaying the 2nd booster dose in a 3-dose regimen (2+1) may in fact be preferable as this could potentially improve the anti-polysaccharide immune response, as has been observed in European infants receiving the pneumococcal conjugate vaccine [41]. In order to assess protection in the context of maximally induced anti-COPS IgG, we utilized COPS:FliC formulated with MPL as this produced higher anti-COPS responses in adult mice relative to the unadjuvanted formulation. It is unknown whether formulation with an adjuvant would be required for induction of immune responses to COPS:FliC in humans as immune responses to vaccination may be different relative to those obtained in rodents. Taken together, these preclinical results provide helpful insight with respect to planned first-in-human phase 1 clinical studies as well as potential immunization strategies for human infants with iNTS COPS:FliC glycoconjugates that would occur later in clinical development.
|
10.1371/journal.pgen.1002744 | Fitness Landscape Transformation through a Single Amino Acid Change in the Rho Terminator | Regulatory networks allow organisms to match adaptive behavior to the complex and dynamic contingencies of their native habitats. Upon a sudden transition to a novel environment, the mismatch between the native behavior and the new niche provides selective pressure for adaptive evolution through mutations in elements that control gene expression. In the case of core components of cellular regulation and metabolism, with broad control over diverse biological processes, such mutations may have substantial pleiotropic consequences. Through extensive phenotypic analyses, we have characterized the systems-level consequences of one such mutation (rho*) in the global transcriptional terminator Rho of Escherichia coli. We find that a single amino acid change in Rho results in a massive change in the fitness landscape of the cell, with widely discrepant fitness consequences of identical single locus perturbations in rho* versus rhoWT backgrounds. Our observations reveal the extent to which a single regulatory mutation can transform the entire fitness landscape of the cell, causing a massive change in the interpretation of individual mutations and altering the evolutionary trajectories which may be accessible to a bacterial population.
| Bacteria rely on complex genetic regulatory networks to respond to hazards or opportunities that they encounter. These networks consist of a series of sensory modules, coupled with various response elements that must be appropriately activated to deal with a given set of environmental conditions; all of these condition-specific elements interact with the cell's core machinery for gene expression. When they encounter a novel environment, populations of bacteria rapidly evolve to adapt to that environment; alterations in gene expression play a major role in this process and, in particular, mutations to the cell's central gene expression machinery are surprisingly common in laboratory evolution experiments. Focusing on one such mutation that had previously been shown to enhance the host cell's ethanol tolerance, we show that the same alteration can in fact aid cellular survival under a wide variety of conditions. In addition, the interactions of this regulatory mutation with other genes throughout the genome cause these mutations to fundamentally reshape the effects of any other genomic changes that occur, and thus alter the overall evolutionary course taken by a population.
| Rho-dependent termination is a crucial component of transcriptional regulation in bacteria, and is estimated to terminate approximately half of the transcripts present in E. coli [1], [2]. Recent studies have shown that this type of transcription termination is particularly prevalent in prophage and other horizontally acquired DNA, thus insulating the cell from the deleterious expression of such elements [3], [4]. Rho has also been shown to safeguard genomic integrity by reducing co-directional collisions between transcriptional and replication machinery [5], [6]. The rho* allele was initially identified in a set of short-term laboratory evolution experiments as a major modifier of ethanol tolerance in E. coli MG1655 [7]. This allele contains a missense mutation (F62L) in the RNA binding domain of Rho, which has been previously shown to cause a 20% higher read-through of the termination site tR1 [8], and raise the dissociation constant for (rC)10 by a factor of four [8]. The ethanol tolerance caused by rho* can be traced to overexpression of a few loci (namely the prpBCDE and cadBA operons [9]), which are also among the transcriptional units strongly affected by chemical inhibition of Rho-dependent termination [3]. Mutations to rho have also been observed in several other laboratory evolution experiments [10]–[13], although the nature of their contribution to fitness in those cases is unclear.
Given the pervasive effects on transcription throughout the genome caused by short term inhibition of Rho-dependent termination [3], [4], we sought to determine the full breadth of effects of rho*, both on cellular phenotype and on secondary mutations at other loci. We found widespread effects from both classes; rho* significantly alters cellular fitness in the presence of a variety of nutrient sources and antibiotics, and shows epistatic interactions with mutations at ∼5% of other loci in the genome. Our results illustrate that mutations to rho*, and presumably other central components of the transcriptional apparatus, facilitate the rapid generation of broad phenotypic diversity in bacteria, with significant consequences for the evolution of populations under stress.
Based on the biological function of Rho, one naturally expects that rho* cells will show increased transcription immediately downstream of Rho-dependent termination sites. Indeed, measurements of transcript abundances [3] and RNA polymerase occupancy [4] have recently shown that after short-term inhibition of Rho-dependent termination using the compound bicyclomycin (BCM), hundreds of transcriptional readthrough events are apparent throughout the E. coli genome, with significant over-representation of recently and horizontally acquired genomic regions. In order to assess the effects of rho* on transcriptional output during balanced growth, we performed transcriptional profiling comparing WT and rho* cells using tiling microarrays (raw data available at the Gene Expression Omnibus; Accession GSE32022). We then identified genomic regions showing significant differences in transcript levels between the two genetic backgrounds (Bonferroni-corrected p<0.01 and greater than twofold change in representation; see Text S1, Section 1.6). We found a total of 2535 probes (out of 92794 positions) showing significant differences, located in 1281 genes and 433 intergenic regions; a few example loci are shown in Figure S1. We identified the most significantly perturbed genes in rho* by flagging all cases for which the median WT:rho* expression ratio for all sense-stranded probes in a given gene indicated a greater than 1.5-fold change in expression level; using this threshold, 155 genes were overexpressed and 44 underexpressed in rho*. The presence of such a substantial underexpressed fraction again illustrates the presence of indirect effects of rho* on the genetic regulatory network, whereas the overexpressed fraction likely represents a combination of genes overexpressed directly due to transcriptional readthrough and those altered due to regulatory perturbations. Consistent with this interpretation, probes which are significantly overexpressed in rho* cells relative to WT show 1.3-fold enriched overlap with a set of prophages, insertion sequences, and K-12 specific elements (the MDS42 deletion sites [14]; p = 0.011 by random permutation of site locations). Probes overexpressed in WT cells, in contrast, show no significant correlation with MDS42 deletion sites (1.2-fold depletion; p = 0.198).
It is also notable that of the probes identified as significantly overexpressed in rho* cells relative to rhoWT, 82% of those overlapping genes were on the antisense strand (cf. 55% for those underexpressed in rho*; see Figure S2A). A recent RNA-seq study showed the presence of pervasive antisense transcription throughout the E. coli genome [15], which the authors presume to be limited in extent primarily by Rho-dependent termination [15]. In addition, Peters and coworkers identified 24 novel antisense transcripts appearing in BCM-treated cells [4], more directly illustrating a role for Rho-dependent termination in at least some cases. In order to assess the effects of rho* on previously identified antisense transcripts, we compared the log-ratios of transcript levels in rho* vs rhoWT cells along a series of windows centered at 50 bp increments downstream of the 1,005 antisense transcription start sites identified by Dornenburg et al. [15]. As seen in Figure S2B, a significant increase in transcription is apparent in rho* cells along the first several hundred bp of these antisense transcripts, illustrating a major mechanism through which rho* likely alters cellular physiology. Furthermore, this analysis does not capture the effects on antisense transcripts which are at undetectable levels in rhoWT cells (and thus would have been missed from the Dornenburg study).
In order to obtain a pathway-level view of the changes in gene expression caused by rho*, we applied iPAGE [16] to identify gene ontology (GO) pathways which share significant mutual information with the log-ratio of rho* vs. WT RNA from microarray experiments (see Text S1, Section 1.7 for details). In all, 19 non-redundant GO terms show significant mutual information with the expression profile for sense-strand RNA and 10 non-redundant GO terms for the antisense profile (out of 1340 present in the annotation set [17], using a threshold of p<0.0001). The changes in expression patterns for a few example pathways of particular interest are shown in Figure 1A, and the full set of significantly perturbed GO terms is shown in Figure S3. These changes in expression affect a variety of cellular pathways including diverse aspects of metabolism and regulation; for example, genes involved in transcriptional attenuation and post-transcriptional regulation were over-expressed in the rho* background, which may represent a regulatory coping strategy for minimizing the deleterious effects of transcriptional read-through. For the most part, however, the fitness consequences (if any) of these broad-reaching expression modifications were not readily identifiable.
In order to measure the extent to which the altered gene expression state of rho* MG1655 cells affects their fitness in different environments, we compared the growth of these cells to that of wild type cells in the presence of a variety of nutrient conditions and antibiotics (see Text S1, Section 1.2 for details). We identified 22 conditions (shown in Figure 1B and Table 1) in which the relative fitness of WT and rho* differed significantly from that in our reference condition (glucose minimal media), with 8 conditions favoring WT and 14 favoring rho* cells (we use steady-state growth rate as a proxy for fitness unless otherwise noted; see Section 2 of Text S1, Table S1, and Figure S4 for a discussion of other relevant quantities). The number and nature of these discovered environments show that the regulatory perturbations caused by the rho* mutation functionally modify a variety of pathways in the cell. In some cases the fitness differences between WT and rho* cells can be directly explained by modified gene expression. For example, the pathway-level analysis in Figure 1A shows that pathways involved in oxidative metabolism are under-expressed in the rho* background, which may explain their increased aminoglycoside resistance [18]. Most conditions showing fitness differences, however, defy such simple explanations.
The varied, pleiotropic effects of rho* on fitness under different growth conditions suggested that the rho* mutation may also result in global changes in the fitness landscape, altering the effects of any additional mutations. To test for such changes, we used fitness profiling of transposon-mutagenized libraries [19] to create coarse-grained representations of the fitness landscape under four conditions (a schematic of the procedure is shown in Figure 2A–2B and detailed methods are provided in Text S1, Section 1.5; raw data are available from the Gene Expression Omnibus, Accession GSE32022). For a given condition, a modified fitness landscape implies that there are loci whose fitness consequences are different in rhoWT and rho* backgrounds. These loci, in turn, provide insight into the specific mechanisms through which rho* alters the cell's regulatory and physiological state (we provide more detailed analysis of several such cases, including follow-up experiments on knockout strains, in Text S1, Section 3; see also Figures S5 and S6 and Table S2). Similar patterns emerged in all four conditions tested: both the WT and rho* fitness profiles show hundreds of sites at which transposon insertions lead to significant changes in fitness, with the majority unique to one genetic background or the other (see Table S3). Comparisons of the distributions of selection scores between WT and rho* cells in each condition are shown in Figure 2C; the low correlations between scores of genes in the two genetic backgrounds under all four conditions indicate that the fitness consequences of secondary mutations are heavily dependent on the genotype at the rho locus, whereas correlations between replicates from the same genetic background under each condition are quite high. The overlaps of loci and pathways with significant fitness effects in the two backgrounds are shown in Figure 3; in all four cases, a common core of loci exists which strongly contributes to fitness in both the rhoWT and rho* backgrounds, but the majority (>70%) are unique to one background or the other. This indicates that the effects of these mutations are in fact strongly altered by the rho* allele. Consistently, in Figure 4 we show examples of several loci where significant epistasis between rho* and a secondary mutation was observed in defined strains during follow-up experiments (details of the epistasis experiments and calculations are given in Text S1, Section 4; see also Tables S4 and S5).
The genetic basis of laboratory-evolved ethanol tolerance provides an example of the reshaping of the fitness landscape by rho*. In the course of the experiments reported here, we found that rho* alone is insufficient to confer the levels of ethanol tolerance observed in the evolved strain from [7]. Instead, using global linkage analysis, we found that an epistatic interaction between rho* and rpsL* (a nonsense mutation in the S12 ribosomal protein RpsL) provides a substantial portion of the increase in ethanol tolerance (see Section 5 of Text S1, Figures S7 and S8, and Table S6 for further details). Relative growth rates for all combinations of wild type and identified mutant alleles of rho* and rpsL* are shown in Figure 5. Whereas the rpsL*/rho* double mutant showed a maximum growth rate of 1.01 doublings/hour in the presence of 5.5% ethanol, rho*/rpsLWT cells grew at 0.85 doublings/hour, and both rhoWT/rpsLWT and rhoWT/rpsL* cells showed no or negligible growth. Conversely, in LB alone the double mutant was less fit than all other allelic combinations (despite the beneficial effects of rho* in isolation.). Thus, rho* shows a positive epistatic interactions with rpsL* in ethanol-containing media and a negative epistatic interaction in the absence of ethanol.
The wholesale reworking of the cell's fitness landscape due to rho* illustrates its potential to open evolutionary paths that would not otherwise be accessible. rho* provides both direct fitness effects and broadly varying (and often positive) epistatic relationships with perturbations at other loci, allowing it to provide benefits early in an evolutionary trajectory while at the same time providing a different, and frequently larger, profile of possible adaptive secondary mutations (see Tables S3 and S7). The interaction between rho* and rpsL* described above represents one such case: rho* itself provides a beneficial fitness effect in the presence of ethanol, and also exhibits positive epistasis with a mutation at the rpsL locus. A more general schematic is shown in Figure 6: the fitness effects of mutations throughout the genome are strongly influenced by the genotype at rho (and presumably other core transcriptional proteins as well), making some secondary mutations more or less beneficial than they would be otherwise (Figure 6, genotype B). Mutations such as rho* can also both provide a fitness benefit relative to the wild type under common growth conditions, and reveal higher fitness genotypes upon exposure to stress conditions (Figure 6, genotype C). rho* is expected to exert its effects simply by altering transcription (in this case primarily by allowing expression of regions which would not otherwise be transcribed); we thus expect that mutations to other core components of the cell's transcriptional machinery, or to other broadly influential regulators, would show similar levels of evolutionary and phenotypic leverage.
In support of this view, mutations to rho [10]–[13], RNA polymerase [11], [13], [20]–[23] and DNA supercoiling proteins [24]–[26] have frequently been observed in a variety of other recent directed evolution experiments. In a few cases, specific epistatic interactions involving these core transcriptional components were found to shape the future adaptive trajectory of populations. For example, Applebee and coworkers [23] found that in a set of E. coli populations evolved to grow efficiently in glycerol minimal media, RNA polymerase mutations arising earlier in the evolutionary trajectories showed positive epistasis with subsequent glpK mutations (and possibly mutations to dapF and murE as well). Similarly, in analyzing populations from an extremely long-term evolution experiment, Woods et al. [26] found the presence of two variant topA alleles in competition; of these, the allele present in the subsequently evolved strain had a less positive direct effect on fitness, but also showed positive epistasis with a secondary mutation at spoT that yielded an overall higher fitness phenotype. In general, these previous studies have not, however, fully explored the full breadth of both direct phenotypic and epistatic effects of the housekeeping mutations that they identified.
Because the primary effect of a hypomorphic rho allele such as rho* is to allow expression of regions of the genome that would not typically be expressed (see above; also [3], [4]), we thus see that the impairment of a system setting baseline boundaries for gene expression can in fact bring forth beneficial, but normally hidden, phenotypes. The concept that robustness to the effects of mutations may facilitate adaptive evolution by allowing the accumulation of genetic diversity that can be subsequently released by a single perturbation, has been proposed repeatedly in the theoretical literature. Wagner [27] discussed the “neutral space" of a biological system – a range of equivalent solutions to a given condition – and notes that the presence of diversity within the neutral space allows variation that may be useful under subsequently encountered conditions. Draghi et al. [28] illustrated precisely this phenomenon more quantitatively using a computational model, showing that intermediate levels of robustness (modeled as the probability of a given mutation being neutral) accelerated the adaptation of populations by providing a reservoir of phenotypically neutral genetic diversity, including variants that could be adaptive under changing conditions. More recently, in modeling tradeoffs involved in the regulation of translational readthrough, Rajon and Masel [29] found bistable solutions which required either global regulation to reduce readthrough rates, or a combination of higher readthrough rates but reduced incidence of deleterious products upon readthrough; the high readthrough rate solution was found to be more evolvable by allowing the accumulation of non-deleterious genetic diversity downstream of translational stop sites, which can subsequently be incorporated through a single mutation to the stop codon.
The behavior of rho* is also reminiscent of two phenomena related to the core translational machinery of yeast. Jarosz et al. [30] recently showed that the chaperone Hsp90 acts to suppress the effects of genetic variation occurring naturally between yeast strains; temperature stress or chemical inhibition of Hsp90 yielded a wide variety of phenotypic changes among ∼100 different yeast strains under 100 low-level stress conditions, frequently with differing signs of effect on fitness for different strains under the same condition. Furthermore, the authors found that Hsp90 in fact shows epistatic interactions with 20% of naturally occurring genetic variations between the strains under consideration. Similar phenomena have been observed for the yeast prion state [PSI+] [31]–[33], where (as with rho*) an alteration in the behavior of a regulatory protein gives rise to a highly pleiotropic phenotype which may be harmful or beneficial under a variety of conditions, interacts strongly with the precise genetic background of the cell in question, and appears to exert its effects by causing ectopic expression of sequences which are generally silent. The comparison between both mechanisms in yeast and rho* must not be taken too far, as there are also substantial differences, most notably in that Hsp90 and [PSI+] act post-transcriptionally, [PSI+] in particular represents an epigenetic rather than genetic mechanism, and both the prion states and Hsp90 relaxation have been shown to be encouraged by environmental stress [30], [34], whereas no similar mechanism would be expected to mutate core housekeeping genes in stressed E. coli cells preferentially. Nevertheless, the effects of both yeast mechanisms, and bacterial rho mutations, illustrate that microorganisms possess the genetic potential to grow under a broader array of conditions than their regulatory logic allows, that some of the hidden potential may be unlocked through perturbations of core regulatory proteins, and that even a single such perturbation may unleash a wide variety of positive or negative effects and interactions with other loci throughout the genome.
Taken together, our findings illustrate that a single amino acid substitution in the global transcriptional terminator Rho leads to a wholly different regulatory and phenotypic state, in which gene expression is globally altered and cellular fitness in a broad variety of environments has changed. The same mutation also dramatically alters the fitness landscape with regard to other genetic variations, making accessible a number of beneficial secondary mutations that are otherwise neutral or deleterious. The set of states reachable through rho* or other point mutations of core regulatory proteins comprise a previously underappreciated reservoir of additional phenotypes accessible to bacterial populations under selective conditions. These findings imply a role for mutations to regulators such as rho both as evolutionary catalysts, by making a variety of secondary mutations more favorable than they would be in the parental strain, and as evolutionary capacitors [35], by allowing silently accumulating genetic diversity to take effect rapidly upon changes in gene regulation. The full extent to which this capacity of core housekeeping and regulatory proteins is used during evolutionary trajectories, and the identity of the complete set of genes showing such broadly influential behavior, are not yet clear. It is also intriguing to speculate that classical global regulators may also show similarly diverse effects, either upon genetic perturbation or as a response to environmental signals, given that the number of genes substantially perturbed by rho* (∼200) is comparable to the number directly or indirectly affected by each global regulator (e.g., CRP, IHF, or FNR) [36].
A complete listing of strains in the present study, including abbreviations used throughout the text, is given in Table S8, and PCR primers are shown in Table S9. The E. coli K12 strain MG1655 [37] (ATCC strain 700926) provides the genetic background for all experiments reported here. For measurement of transcript abundances, cells were grown to mid-log phase in M9t/glucose, and RNA extracted using total RNA purification kit (Norgen Biotek, Cat 17200). After poly-A tailing, the extracted WT and rho* RNA were separately labeled, pooled, and then hybridized to Agilent custom arrays tiling the whole genome at 50 bp intervals, alternating between strands. Transposon mutagenized libraries were prepared as described by Girgis et al. [19]. Selections were carried out for 16 hours in 25 mL of either selective or reference media, and genomic DNA isolated using a DNeasy Blood and Tissue Kit (Qiagen). The transposon footprinting and labeling protocol for quantifying relative fitness under different conditions is described by Girgis et al. [18]. Bacterial growth curves were measured in Costar 96-well clear polystyrene plates, using either a Biotek Synergy MX or Powerwave XS2 plate reader (Biotek; Winooski, VT). Plates were incubated at 37°C with continuous shaking, and optical density (OD) reads at 600 nm taken every 10 minutes. Abbreviations for nutrient sources and antibiotics are given in Table S10. Complete methodological details are provided in Section 1 of Text S1, as well as Figure S9.
|
10.1371/journal.pntd.0005156 | Comparison of Doxycycline, Minocycline, Doxycycline plus Albendazole and Albendazole Alone in Their Efficacy against Onchocerciasis in a Randomized, Open-Label, Pilot Trial | The search for new macrofilaricidal drugs against onchocerciasis that can be administered in shorter regimens than required for doxycycline (DOX, 200mg/d given for 4–6 weeks), identified minocycline (MIN) with superior efficacy to DOX. Further reduction in the treatment regimen may be achieved with co-administration with standard anti-filarial drugs. Therefore a randomized, open-label, pilot trial was carried out in an area in Ghana endemic for onchocerciasis, comprising 5 different regimens: the standard regimen DOX 200mg/d for 4 weeks (DOX 4w, N = 33), the experimental regimens MIN 200mg/d for 3 weeks (MIN 3w; N = 30), DOX 200mg/d for 3 weeks plus albendazole (ALB) 800mg/d for 3 days (DOX 3w + ALB 3d, N = 32), DOX 200mg/d for 3 weeks (DOX 3w, N = 31) and ALB 800mg for 3 days (ALB 3d, N = 30). Out of 158 randomized participants, 116 (74.4%) were present for the follow-up at 6 months of whom 99 participants (63.5%) followed the treatment per protocol and underwent surgery. Histological analysis of the adult worms in the extirpated nodules revealed absence of Wolbachia in 98.8% (DOX 4w), 81.4% (DOX 3w + ALB 3d), 72.7% (MIN 3w), 64.1% (DOX 3w) and 35.2% (ALB 3d) of the female worms. All 4 treatment regimens showed superiority to ALB 3d (p < 0.001, p < 0.001, p = 0.002, p = 0.008, respectively), which was confirmed by real-time PCR. Additionally, DOX 4w showed superiority to all other treatment arms. Furthermore DOX 4w and DOX 3w + ALB 3d showed a higher amount of female worms with degenerated embryogenesis compared to ALB 3d (p = 0.028, p = 0.042, respectively). These results confirm earlier studies that DOX 4w is sufficient for Wolbachia depletion and the desired parasitological effects. The data further suggest that there is an additive effect of ALB (3 days) on top of that of DOX alone, and that MIN shows a trend for stronger potency than DOX. These latter two results are preliminary and need confirmation in a fully randomized controlled phase 2 trial.
Trial Registration: ClinicalTrials.gov #06010453
| Onchocerciasis is a vector borne disease that is still a major health burden in endemic countries, despite twenty years of control that has led to effective control and reduction of blindness and morbidity in West Africa. To reach the goal of onchocerciasis elimination, alternative strategies are needed to overcome existing hurdles. Ivermectin is used for mass drug administration (MDA) and kills the microfilariae, thus preventing uptake by the vector and reducing transmission. Since adult worms live for 10 years or longer, MDA requires many years of treatment, a heavy burden on health care systems. Inadequate population coverage, sub-optimal responses, possible development of resistance and the risk of severe adverse reactions with co-endemic Loiasis are current hurdles for achieving elimination. The search for new drugs that could enhance elimination by permanently sterilizing and killing the adult worms has identified a 4–6 weeks course of doxycycline, a slow-killing drug due to its indirect mode of action by killing endosymbiotic bacteria. The present study aimed to investigate a new anti-Wolbachia drug, minocycline, and doxycycline in combination with one of the standard broad-spectrum anti-helminthic drugs, albendazole, to identify shorter treatment regimens. Our data confirm the efficacy of a 4-weeks doxycycline regimen and show that albendazole has additive effects when used in combination with doxycycline. In addition, comparison of a 3-week course of minocycline with a 3-week course of doxycycline revealed a trend for stronger efficacy of minocycline. Further development of anti-Wolbachia and anti-filarial drug combinations are warranted to improve treatment options to overcome existing hurdles and for use in “end-game” scenarios when switching from MDA to “test & treat” strategies.
| More than 200 million humans are parasitized by filarial nematodes, causing the neglected tropical diseases: lymphatic filariasis, loiasis and onchocerciasis. The lymphatic, ocular and dermatological pathologies have severe economic and social consequences including poor school performance, low productivity, higher health related costs among infected adults and a reduced life span [1–3]. In onchocerciasis, several programs and developments have greatly improved the situation in Africa since the 1970’s when the Onchocerciasis Control Programme (OCP), a programme relying on vector control, in West Africa was initiated. In 1987, a new treatment-based strategy was made possible due to the donation of ivermectin (IVM) by Merck as long as needed and the African Programme for Onchocerciasis Control (APOC), a coordinated community directed distribution of IVM mass drug administration (MDA) in 28 African countries [4] that has now ended in 2015, was launched. Despite the tremendous progress [5], disease elimination will require continuity and new approaches with APOC and onchocerciasis control are being integrated into the Expanded Special Project for Elimination of Neglected Tropical Diseases (ESPEN).
A major problem with the current treatment is that IVM has minimal efficacy against adult worms and does not permanently stop microfilariae (Mf) production. Simulation studies have suggested that administering IVM at shorter intervals of 6 instead of 12 months intervals have a higher likelihood of eliminating the infection, but incur large logistical costs on health infrastructures of the endemic countries [6, 7]. Immigration of infected persons into areas where filariasis is considered eliminated and IVM treatment has ceased may occur with potential re-emergence of the disease. Equally important is the evidence of persistent IVM suboptimal efficacy in some communities [8–11], which is of particular concern as there is currently no alternative treatment suitable for MDA.
In addition, MDA cannot easily be undertaken in regions endemic for onchocerciasis, where patients are co-infected with Loa loa (primarily in Central Africa) and which are estimated to cover at least 20% of onchocerciasis-endemic areas [12]. This is because microfilaricidal drugs also kill Loa loa Mf and in patients with high parasitemia this may elicit loiasis-specific adverse reactions (Loa loa encephalopathy), leading to severe neurological disorders or death [13]. Thus there is an urgent need for a safe macrofilaricide, targeting adult worms that can be used for onchocerciasis and LF, particularly in problem areas, such as those with emerging ivermectin resistance or Loa loa co-endemicity. Further development of anti-Wolbachia macrofilaricides is therefore warranted to improve treatment options to overcome existing hurdles and for use in “end-game” scenarios in areas with hypoendemicity or when switching from MDA to “test & treat” strategies (who.int/apoc/ATS_Report_2015.12).
Targeting Wolbachia endosymbionts by antibiotics is an alternative approach that has been verified by a number of clinical trials in onchocerciasis [14–16] and LF [17–21]. Depletion of Wolbachia leads to long-term sterilization and to a macrofilaricidal effect. This has been shown first in animal models [22–24] and was successfully translated into clinical trials in human onchocerciasis [15] and lymphatic filariasis [21, 25]. A recent meta-analysis on the available data from clinical trials in onchocerciasis has shown that the 3 so far used regimens, doxycycline (DOX) 200 mg/day for 4 weeks, DOX 200 mg/day for 6 weeks [15], and DOX 100 mg/day for 5 weeks [26] are broadly equivalent, in particular with regards to the sterilizing effects of adult female worms [27], and therefore DOX treatment with 200 mg/day for 4 weeks can be considered as a “standard” anti-wolbachial therapy for onchocerciasis.
The advantages of DOX for the depletion of Wolbachia endobacteria have stimulated the development of new drugs. The indirect mode of action with its slow antifilarial activity confers an excellent safety profile by avoiding inflammatory reactions [28] and can safely be used in L. loa coendemic areas, since this filarial parasite does not contain Wolbachia. The A∙WOL Consortium was established to find new anti-wolbachial drugs or drug combinations that are deliverable in a shorter regimen, with a secondary goal to optimize regimens using currently known anti-Wolbachia antibiotics (www.a-wol.com). The tetracycline derivate minocycline (MIN), was identified as priority hit in the cell culture screening assay [29] as well as on adult O. gutturosa male worms [3] and in the L. sigmodontis animal model [30]. Albendazole (ALB) is widely used in combination with IVM for lymphatic filariasis and soil transmitted helminths, yet no clear evidence for efficacy against onchocerciasis has been provided.
Here, we present the results of a randomized open pilot-trial to compare DOX, MIN, as well as the combination of DOX plus ALB to ALB alone in their efficacy against onchocerciasis. The underlying rationale that led to the design of this trial was to provide a shortened treatment that is equally efficient in Wolbachia depletion or blockage of embryogenesis after 6 months as the current regimen of 4 weeks DOX 200mg compared to 3 days ALB 800 mg.
This study was undertaken in 14 villages adjacent to the river Offin in Ghana (Upper and Lower Denkyira Districts in the Central Region and the Amansie Central and Adanse South Districts in the Ashanti Region). These rain forest areas are within the distribution range of the vector (<12 km from the breeding site, the river Offin and affluent creeks) and hyperendemic for onchocerciasis but not for other filarial infections and were not part of either OCP or APOC programs. MDA has been implemented by the Ghanaian Ministry of Health since 2001 in Upper and Lower Denkyira districts and from 2008 in the Amansie and Adanse South areas. However, MDA compliance has not met MDA targets in all areas, especially in the Upper and Lower Denkyira districts, and therefore, at the time of sampling a considerable number of people had not or not frequently taken part in IVM therapy. This study was approved by the Committee on Human Research, Publications and Ethics of the School of Medical Sciences of the Kwame Nkrumah University of Science and Technology (KNUST), Kumasi, Ghana,—Research Ethics Committee of the Liverpool School of Tropical Medicine as well as the Ethical Committee of the University Hospital of Bonn, Germany. The trial was registered at Current Controlled Trials (www.isrctn.com, # 06010453).
Eligible for this study were healthy male and female patients (without clinical conditions requiring long-term medication and normal renal and hepatic laboratory profiles), aged between 18–55 years, > 40 kg of body weight and the presence of at least one palpable onchocercoma. Hepatic and renal functions as well as pregnancy tests were assessed by dipstick chemistry using venous blood and urine. Skin snips (biopsies) were taken to determine skin microfilarial (Mf) loads. Exclusion criteria were: abnormal hepatic and renal enzymes, pregnancy, breast-feeding, intolerance to the study drugs, alcohol or drug abuse, history of TB, glucose in urine, hypertension or history of any condition requiring long term medication (see also ISRCTN registry). Written informed consent was obtained from all individuals. A Data Monitoring and Ethic Committee (DMEC) was established prior to this study.
This study is a randomized, open, clinical phase 2a drug trial. The purpose of the 5 different treatment regimens was to determine if a reduced treatment period could be achieved using doxycycline 200mg alone for 3 weeks (DOX 3w), doxycycline 200mg for 3 weeks plus 3 days albendazole (DOX 3w+ALB 3d), minocycline 200mg for 3 weeks (MIN 3w) or doxycycline 200mg for 4 weeks (DOX 4w) as standard anti-wolbachial regimen compared to albendazole 800mg for 3 days (ALB 3d), which is considered equivalent to an untreated control arm, when given on its own [31].
This study was a pilot trial to gain first experience regarding the primary and secondary endpoints under the intended treatment regimens. Therefore a sample size of 30 participants (20 participants plus 50% drop-out rate) per treatment arm was chosen in line with the sample sizes normally taken for pilot trials. Participants were randomly assigned to one of 5 treatment arms according to their participation in the MDA and their Mf-status as follows:
Participants received the study drugs under daily observation. Participants were given IVM after the nodulectomies. After study completion, DOX 4w was offered to all participants. Even though the study was an open trial and therefore participants as well as the trial clinicians were not blinded for the treatment, the outcome assessors (histology and PCR) were blinded for the treatment allocation.
Primary outcome of this trial was the absence of Wolbachia endobacteria in adult living female worms assessed by immunohistology 6 months after treatment. Secondary outcomes were the quantitative evaluation of Wolbachia endobacteria reduction after 6 months by immunohistology as well as by PCR. Furthermore reduction and presence or absence of Mf in the skin and embryogenesis within the worms as well as the number of live or dead worms was evaluated.
All infected patients presented at least one palpable nodule [32]. For Mf analysis, two skin biopsies of 1–3 mg were taken from the buttocks using a corneoscleral (Holth) punch. Each biopsy was immersed in 100 ml of 0.9% NaCl solution in a well of a microtiter plate. The skin biopsies were incubated overnight at room temperature to allow Mf to emerge. The solution was then transferred onto a slide for microscopic examination [15, 33]. The biopsies were weighed using an electronical balance and Mf load was calculated per mg skin. The presence of further parasitic infections were assessed using standard diagnostic tests on stool and urine samples [34]. Five participants were positive for hookworm, one for Giardia lamblia and one for Strongyloides.
Nodules were fixed in 80% ethanol or 4% phosphate buffered formaldehyde solution. Samples were embedded in paraffin and several sections were stained with hematoxylin and eosin, Gomori’s method for iron, or immunostained using antibodies against Dirofilaria immitis Wolbachia surface protein (Diwsp) or Wolbachia PAL-lipoprotein (wBmPAL) for presence of Wolbachia and cathepsin D-like lysosomal aspartic protease of O. volvulus (APR) for worm vitality [35]. Thereby at least 8 sections across the nodules were histologically assessed as described previously [36, 37]. In brief, characteristics for death of a worm included loss of body wall integrity, loss of nuclei of all organs and absence of APR-staining. Very degenerated worms, still APR-positive, were classified as moribund and grouped in the category “dead”. The designation “living” refers to worms judged as being alive at the time of nodulectomy. Characteristics for differentiation include the size, general organ or tissue structure as well as iron deposition in the intestine of the worm. The age of the living worms was estimated as "newly acquired", “young”, “middle”, “old” (see Fig 2) [37]. Thirteen male worms were defined as newly acquired. They were not subtracted from this analysis because these worms could have been acquired already shortly before treatment onset. The classification of Wolbachia content was: no Wolbachia (negative for Diwsp or wBmPAL staining), few or many. Since Wolbachia are densely packed within the hypodermis, the female worms containing such conglomerates were assigned as many, whereas those female worms containing visible Wolbachia (<50) but not in dense packages were assigned as few (see Fig 2). Developmental stages from the two-cell stage through to stretched Mf in the uterus were classified as “embryos”. When only the stretched Mf were degenerated and the other embryo stages were not, embryogenesis was recorded as “normal” for the purpose of this tetracycline-oriented study, since the previous IVM treatment would have conferred precisely these effects, i.e. degeneration of intra-uterine Mf (in addition to killing of skin Mf). The sections were assessed by two experienced parasitologists (SS, BD).
For DNA extraction 8 nodule paraffin sections of 4 μm were placed in microcentrifuge tubes and DNA was extracted according to the manufacturer’s instructions (Quiamp DNA Mini Kit). The Wolbachia ftsZ and O. volvulus actin gene were quantified from the purified DNA by real-time duplex PCR (qPCR) using the following conditions: 10x HotStar Taq Polymerase buffer (Qiagen), 200 μM dNTP, Primers: OvFtsZ forward primer (aggaatgggtggtggtactg), OvFtsZ reverse primer (ctttaaccgcagctcttgct), OvActin forward primer (gtgctacgttgctttggact), OvActin reverse primer (gtaatcacttggccatcagg), OvFtsZ taqman probe (5’6-FAM ccttgccgctttcgcaatcac 3’DDQ-1), OvActin taqman probe(5’HEX aacaggaaatggcaactgctgc 3’BHQ-1), 2.5 units HotStar Taq, and 2 μl DNA in a 20 μl reaction. Final primer concentrations were 400 nM for OvwFtsZ and OvActin, final taqman probe concentrations were 25 nM for OvwFtsZ and 50 nM for OvActin and 6 mM for the MgCl2 concentration. Genes were amplified in a Rotorgene 3000 (Quiagen, Hilden, Germany) using the following conditions: 1X 15 min at 95°C, 45 cycles of 95°C for 15 sec, 58°C for 30 sec. Fluorescence was acquired on the FAM and JOE channel.
Analyses were done using SPSS (IBM SPSS Statistics 22; Armonk, NY) and SAS version 9.2 (SAS Institute Inc. Cary, NC, USA).
Three data sets (per protocol, PP; intention to treat, ITT; Mf-PP) were used to analyze the data (Fig 1). The intention to treat set (ITT) includes all patients who were randomized to one of the 5 treatment regimens and who took the drugs at least once. The PP set is a subset of the ITT set and includes all patients who completed the treatment without any violations of the protocol and were present for nodulectomies 6 months after treatment onset. Since this is a non-confirmatory pilot-trial all endpoint analyses were primarily carried out using the PP set. The ITT set was used to describe the baseline data, the adverse events and for confirmation of the PP analyses. For the ITT outcome analyses of the histological and PCR results we included all operated individuals, missing values were not replaced. Additionally, we set up a data set, including all patients present for skin-snipping before treatment onset as well as at the 6 months follow-up (Mf-PP).
Alternating regression, as implemented in the SAS-Procedure Genmod, was used to analyze the histological and PCR data. This procedure allows accounting for a potential dependency between the worms in one patient. Odds ratios (with 95% confidence intervals (CIs)) were derived from the regression models to depict the difference between treatment groups with respect to the outcome measurements (Wolbachia depletion, inhibition of embryogenesis, occurrence of dead worms, FtsZ, actin, FTsZ/actin). Except for the variable "treatment groups", no covariates were included as effects in the primary analyses. Quantitative PCR data were log10-transformed before testing. Comparisons between histological outcome and PCR were done using the unpaired t-test for dead vs. live and ANOVA followed by unpaired t-test for no, many, few Wolbachia as well as ANOVA followed by Tukey’s method as post-hoc test for comparison of embryogenesis.
For the analysis of Mf, the Wilcoxon-signed-rank-test was used for comparisons within the patients before and after therapy and the Kruskal-Wallis-test for comparisons between all groups. Baseline data were analyzed using ANOVA for patient age and weight, Fisher’s exact test for gender and the Kruskal-Wallis test for number of nodules and sites.
The trial profile of the study is illustrated in Fig 1. Altogether, 158 volunteers from 14 villages affected by onchocerciasis that met the inclusion criteria were enrolled into the study and subsequently randomized into the following treatment arms: doxycycline 4 weeks (DOX 4w), doxycycline 3 weeks (DOX 3w), doxycycline 3 weeks plus albendazole 3 days (DOX 3w + ALB 3d), minocycline 3 weeks (MIN 3w) or albendazole 3 days (ALB 3d) as equivalent to a negative control (see Materials and Methods). Baseline data for the participants who took the drugs for at least one day (ITT) with regard to gender, age, weight, and rounds of IVM of the volunteers in each treatment group are given in Table 1. On average, participants had two palpable nodules and one round of IVM prior to treatment begin. 39% of the study participants had never taken IVM. Mf positivity and the intensity of microfilaridermia (Mf/mg skin) did not differ between the groups at study onset with 63.6% Mf-positive individuals.
Since already 8 participants dropped out before (N = 2) or during the first 2 treatment days (N = 6), 8 additional participants were randomized to avoid a higher drop-out rate than primarily calculated. From 156 participants who had started treatment, 116 (73.4%) presented for nodulectomy (109 at the 6 months follow-up and 7 two months later); of these, 99 participants could be included in the PP set (Fig 1), as they had completed treatment according to protocol, i.e. had not been absent for more than 3 consecutive days or during the ALB period (treatment days 9–11).
During treatment, adverse reactions (AR) to the study drugs were monitored. The analyses showed that although significant differences occurred between the study groups (S1 and S2 Tables), the adverse reactions reported were minor (Grade 1 or Grade 2, only once Grade 3) and included stomach pain, dizziness, vomiting and nausea. Most of the reported AR occurred in the combination group DOX 3w + ALB 3d, however AR were not associated with the time point of ALB intake. The number and type of AR were similar in the DOX 4w and DOX 3w group, therefore a longer treatment time does not alter AR. It is important to note that the range of AR after DOX is similar to that seen in a previous study [10], in which DOX was used in lower doses and was performed under blinded conditions. MIN associated AR were mainly dizziness, a common side effect of this drug.
Taken together, the study drugs were well tolerated and no severe or unexpected events occurred due to intake of the medications.
Immense efforts have been undertaken in many countries to achieve elimination of onchocerciasis by MDA. However, still 130 million people are at risk of infection [1]. Current approaches to increase the frequency of MDA from annual to biannual treatments are predicted to improve the chances of reaching the 2020/2025 elimination goals at least in some countries. However, its benefits and costs are highly sensitive to systematic noncompliance and it may not always be feasible to implement biannual treatment, particularly in hard-to-reach populations [6, 7, 42]. In addition, given the problem of re-emergence of infections, suboptimal efficacy of IVM [10] and regions that are co-endemic for Loa loa, the development of new drugs or drug regimens is urgently needed to achieve the elimination of onchocerciasis. This would increase cost effectiveness by avoiding unnecessary treatments of uninfected individuals within the MDA schemes, especially in “end-game” scenarios or when switching from MDA to “test & treat” strategies.
In this study, we have compared DOX 4w, MIN 3w, DOX 3w plus ALB 3d, DOX 3w and ALB 3d alone in their efficacy against onchocerciasis in a randomized, open-label pilot trial. Since we have shown earlier that Wolbachia depletion precedes the inhibition of embryogenesis and adult worm death [14, 15, 43], the follow-up at 6 months post treatment was chosen to investigate the primary outcome, i.e. Wolbachia reduction. In the present study, histological analyses of the adult worms within the extirpated nodules 6 months after treatment revealed a Wolbachia reduction of 99% in the DOX 4w group showing superiority to all other treatment arms. Accordingly, the extent of Wolbachia depletion was highest in the DOX 4w group, followed by DOX 3w + ALB 3d, MIN 3w and then DOX 3w compared to ALB 3d. The Wolbachia reduction in the DOX 4w group was comparable to that of a 6-week 200 mg/d regimen with DOX after 6 months [15]. In addition, real-time PCR analysis from DNA isolated from paraffin embedded histological sections revealed that similar gradations occurred between the groups, suggesting the use of real-time PCR as a sensitive method for the quantification of Wolbachia loads within the worms after a given treatment. We further validated the FtsZ and actin signal by subdividing the data according to the biological status of the worm (live, dead, Wolbachia levels and empty uterus and embryogenesis). FtsZ levels increased according to the histologically defined Wolbachia categories, whereas actin levels were elevated, when worms were alive compared to dead worms. In addition, actin levels were associated with increased embryogenesis.
Inhibition of embryogenesis after doxycycline/anti-wolbachial therapy leading to sterile females has been shown to result in the absence of Mf in the skin of infected patients, a requirement for interruption of transmission Therefore, embryogenesis was analyzed as secondary outcome histologically within the nodules as well as by skin MF loads. We found that already after 6 months only 8% (DOX 4w), 14.7% (DOX 3w + ALB 3d), 11.7% (MIN 3w), 17.7% (DOX 3w) of the living female worms with embryogenesis had normal embryonal development compared to 17.7% (ALB 3d). DOX 4w and DOX 3w + ALB 3d showed a higher number of female worms with degenerated embryogenesis compared to ALB 3d and reached statistical significance (OR 4, p = 0.0283; OR 4.8, p = 0.0423, respectively), whereas MIN 3w and DOX 3w did not. An overall low number of nodules contained MF. However, similar to the results on the absence of Wolbachia, the extent of inhibition of normal embryogenesis follows the same order with the highest being in the DOX 4w group, followed by DOX 3w + ALB 3d, MIN 3w and finally DOX 3w. All findings could be confirmed within the ITT analysis.
Other secondary outcomes were microfilaridermia as well as the number of dead worms within the nodules. The 6-months time point is too early for the treatment efficacy to be reflected in the disappearance of peripheral Mf in the skin, since Mf are not killed directly by anti-wolbachial drugs (an advantage e.g. in Loa co-endemic areas). In an earlier study, despite the percentage of females with degenerated embryogenesis being increased to 95.5% at 6 months after 6 weeks doxycycline treatment [15], all individuals were still Mf-positive. Similarly, Mf counts were not significantly different between the groups in the present study and 62.7% of the patients were Mf-positive at 6 months after treatment.
Also for the observation of macrofilaricidal efficacy a comparatively long observation period is needed and we have shown that a distinct macrofilaricidal effect can be seen with female worms at 20 and 27 months following a six week treatment course with 200 mg DOX per day [15]. Therefore, similar to the absence of effects on microfilaridermia after 6 months, expectedly no differences could be found in the number of dead worms histologically analyzed within the nodules. Despite this, our data provide clear evidence that the time point of 6 months after treatment can be used for the analysis of Wolbachia depletion and in utero inhibition of embryogenesis and is appropriate for identification of anti-wolbachial drugs or drug regimens.
In our study, the 6 months analysis shows that reducing the treatment time with DOX from 4 to 3 weeks is not recommended, as DOX 4w is clearly superior to DOX 3w in Wolbachia depletion and inhibition of normal embryogenesis. Therefore a four weeks course of 200 mg DOX should be further recommended to achieve long-term female worm sterility and death [44, 45].
In addition, the combination of 3 weeks of DOX and ALB 3d showed a higher extent of Wolbachia depletion compared to 3 weeks of DOX alone, indicating an additive effect when used in combination. Studies to determine whether drug-drug interactions or other mechanisms of action of combinations of anti-Wolbachia drugs and anti-filarial drugs lead to enhanced anti-Wolbachia and macrofilaricidal effects to further shorten treatment regimens are currently underway as part of the A·WOL programme.
Minocycline, MIN, another broad-spectrum tetracycline antibiotic, was revealed in an in vitro assay exposing O. gutturosa adult males to be more active than DOX [29], as measured by motility of adult worms and viability by reduction of MTT to MTT formazan assay. Since MIN was also among the top hits of the A·WOL screening activities [3], these data were recently confirmed using the Wolbachia containing insect cell line C6/36. On top of that, data generated from a rodent infection model using Litomosoides sigmodontis confirmed the efficacy in vivo, showing a similar superiority of MIN over DOX (Specht et al, in preparation, [30]). Our clinical data reported here support that MIN may actually be more efficacious than DOX, since MIN showed a trend for being better than DOX 3w in the relevant comparisons, i.e. absence of Wolbachia, inhibition of normal embryogenesis. The differences did however not reach statistical significance, as the pilot trial design was not sufficiently powered to detect these differences.
In summary, these results provide further evidence and confirm earlier studies [15] that DOX 4w is sufficient for Wolbachia depletion and the desired parasitological effects, i.e. depletion of Wolbachia, inhibition of embryogenesis, whereas DOX 3w delivers sub-optimal effects. The data presented here further suggest that there is an additive effect of ALB 3d on top of that of DOX 3w alone, since for both outcomes (Wolbachia depletion, inhibition of embryogenesis), this group showed the second strongest efficacy after DOX 4w. Furthermore, MIN 3w albeit not significant showed a trend for being more efficacious compared to DOX 3w. This may suggest that when used for longer treatment times, MIN may actually be more efficacious than DOX. These latter two results are preliminary and need confirmation in a full-randomized controlled phase 2 trial.
|
10.1371/journal.pgen.0030116 | X Chromosome Reactivation Initiates in Nascent Primordial Germ Cells in Mice | During primordial germ cell (PGC) development, epigenetic reprogramming events represented by X chromosome reactivation and erasure of genomic imprinting are known to occur. Although precise timing is not given, X reactivation is thought to take place over a short period of time just before initiation of meiosis. Here, we show that the cessation of Xist expression commences in nascent PGCs, and re-expression of some X-linked genes begins in newly formed PGCs. The X reactivation process was not complete in E14.5 PGCs, indicating that X reactivation in developing PGCs occurs over a prolonged period. These results set the reactivation timing much earlier than previously thought and suggest that X reactivation may involve slow passive steps.
| X chromosome inactivation is a mechanism to compensate gene dosage difference between XY males and XX females in mammals. During early embryogenesis, one of two X chromosomes in every female cell is inactivated, and the inactive X chromosome is stably inherited through cell divisions of somatic cells. Although precise timing is not given, the inactive X chromosome is known to be reactivated during germ cell development. It is generally believed that the dynamics of X chromosome activity is tightly correlated with major genomic reprogramming events occurring during mammalian development. Therefore, elucidation of the X reactivation kinetics is important for understanding the mechanism of X chromosome inactivation/reactivation processes and the epigenetic reprogramming processes as well. Here we investigated when X reactivation is initiated during development of female mouse germ cells. Contrary to the previous suggestions, X reactivation already begins in nascent primordial germ cells in female mice and proceeds gradually requiring a prolonged period. The activity status of the X chromosomes of germ cells appears to vary from cell-to-cell and from gene-to-gene during the reactivation processes. These results indicate that the X reactivation coincides with the formation of germ cells and suggest that this involves slow passive steps.
| In mice, germ cell formation is first observed in postimplantation embryos; primordial germ cells (PGCs) appear at the base of allantois by embryonic day (E) 7.25 [1]. PGCs are unique compared to other cell types; extensive genomic reprogramming events such as erasure of genomic imprinting and reactivation of the inactive X chromosome take place in this cell type [2,3].
In female mammals, one of two X chromosomes in every cell is inactivated during early embryonic development to compensate gene dosage difference between XY males and XX females [4]. This phenomenon, X chromosome inactivation, represents one of the most remarkable examples of epigenetic gene regulation in mammals. In mouse, X inactivation randomly occurs in embryo proper, whereas the paternal X chromosome is preferentially inactivated in extraembryonic tissues [5,6]. X chromosome inactivation is regulated by a noncoding Xist RNA. It was shown that prior to X inactivation, Xist RNA is transcribed from both active X chromosomes. Upon differentiation, Xist RNA expression is upregulated on the future inactive X chromosome [7], and such Xist transcripts eventually coat the entire inactive X chromosome in differentiated cells [8–10]. Although its mechanistic role in gene silencing is not precisely known, the Xist RNA seems to be required for spreading the inactive state along the chromosome. The expression of Xist RNA is negatively regulated by Tsix, the antisense transcript of Xist [11,12].
The inactivated X chromosome is stably inherited through cell divisions of somatic cells. However, it is known that the two X chromosomes are active in oocytes [13–15], indicating that the inactive X chromosome must be reactivated during germ cell development. Previously, it was thought that X reactivation occurs in PGCs following entry into the gonads just before the initiation of meiosis [16–19] and that the reactivation coincides with rapid DNA demethylation of PGC genome, which is completed in one day of development [2]. Recent studies have shown that X reactivation also occurs in preimplantation development. The paternal X chromosome is inactivated at cleavage stage embryos of female mouse [20–22]. The inactive paternal X chromosome is reactivated in epiblast cells of peri-implantation embryos before the onset of the random X inactivation [21,23]. This reactivation proceeds rapidly and appears to be completed in ∼24 h [21]. Reactivation just after fertilization has also been suggested (reviewed in [24]).
These findings indicate that reactivation of the inactive X chromosome occurs at least twice during mammalian development, once in the epiblast cell lineage at the peri-implantation stage and once in the PGCs at the midgestation stage, and that the reactivation of the inactive X chromosome appears to be tightly correlated with major genomic reprogramming events occurring during mammalian development [25]. Therefore, elucidation of the X reactivation kinetics is important for understanding the mechanism of X chromosome inactivation/reactivation processes and the epigenetic reprogramming processes as well. While epigenetic dynamics of X chromosome inactivation and reactivation in pre- and peri-implantation stage embryos have been studied in detail [20–22], the timing of X reactivation in developing female PGCs has not been clearly defined. The results presented in the previous reports were somewhat contradictory, possibly because of the technical difficulties in analyzing the activity of X chromosomes in developing germ cells [17–19,26].
Here, we devised novel sensitive assays to determine the timing of X reactivation and demonstrated that, contrary to the previous suggestions, X chromosome reactivation has already been initiated in newly formed female germ cells as early as E7.0 and that the reactivation requires a prolonged period (≥7 d), suggesting that the reactivation may involve slow passive processes.
The accumulation of Xist RNA is a hallmark of the inactive X chromosome [8–10]. Xist expression is usually detected by RNA fluorescence in situ hybridization (FISH) analysis of dissociated cells adhered to slide glasses. As PGCs comprise a small cell population surrounded by somatic cells throughout the embryogenesis, dissociation of embryos during sample preparation would lead to loss of PGCs or bring a contamination of somatic cells into the specimens. To avoid these difficulties, an RNA FISH method that detects RNA expression in intact embryos is required. For this purpose, we devised a novel whole-mount RNA FISH method that allows the sensitive detection of Xist RNA without sacrificing the integrity of the embryonic structure (Figure 1A–1E). Using an Xist probe and an antibody against PGC-specific markers, Oct4, we were able to accurately locate PGCs in embryos and assess Xist expression. A Cot-1 probe, which hybridizes to nascent RNA, was used to visualize the nuclei of specimens. Although the Cot-1 signal is eliminated from the inactive X chromosome [20], it was not clear whether the “Cot-1 hole” was present on the inactive X chromosome using our whole-mount RNA FISH method (Figures 1F–1I and S1). We first analyzed the genital ridges isolated from female embryos at E12.5, at which time the PGCs were about to enter meiosis. At this stage, all the PGCs were completely negative for Xist, whereas the somatic cells surrounding the PGCs had a large and strong Xist signal, indicating the presence of an inactive X chromosome (Figure S1; Table 1). At E10.5, when the PGCs began to colonize the genital ridges [3], almost all the PGCs were already Xist-negative (Figure 1F; Table 1), suggesting that X reactivation occurs earlier than previously thought.
Xist-negative PGCs were found even at E8.5, comprising 16.5% of the total cells examined. The Xist-negative PGCs increased to 48.2% at E10.0. PGCs with strong Xist expression comprised 36.3% of total PGCs at E8.5 and 10.9% at E10.0. Most of the remaining cells displayed an Xist signal that was smaller than those of the somatic cells (Figure 1G), and a small number of cells had a large but very faint Xist signal. We hereafter describe those cells showing a small or faint Xist signal as Xist (±) cells. Cells with a large and strong Xist signal are described as Xist (+), and cells with no Xist signal as Xist (−). To investigate whether the cessation of Xist expression is initiated at an even earlier stage, we examined PGCs at E7.75, identified by an antibody against one of the earliest PGC markers, Stella (also known as PGC7) [27,28]. About 70% of Stella-positive PGCs were Xist (+), whereas ∼20% were Xist (±) (Figure 1H; Table 1). Unexpectedly, Xist (−) PGCs already existed at this early stage, although their frequency was lower than that of Xist (−) PGCs in later stages, but higher than that of somatic cells (Table 1).
PGCs are first identified in mice as a cluster of alkaline phosphatase-positive cells at E7.25 [3]. It was recently shown that Blimp1 expression marks nascent PGCs as well as precursor of PGCs [29,30]. To explore the X inactivation status of the earliest PGCs (or their precursors), we generated bacterial artificial chromosome (BAC) transgenic mouse lines in which the gene for monomeric red fluorescent protein (mRFP) [31] was inserted into the Blimp1 locus (Figure S2). mRFP expression was detected at the posterior end of the embryonic ectoderm and visceral endoderm in E7.0 embryos, consistent with endogenous Blimp1 expression (Figure S2) [29]. Because the mRFP faded from the visceral endoderm after the embryos were processed for the whole-mount RNA FISH experiment, only nascent PGCs were clearly visualized (Figure 1I). As in the case of E7.75 PGCs, a minor but significant proportion of mRFP-positive cells (4.7%) were Xist (−) (Figure 1I; Table 1). About 18% of the mRFP-positive cells were Xist (±), whereas the rest of the cells were Xist (+). In contrast, all the mRFP-negative cells, i.e., the somatic cells, observed at the same stage were Xist (+), although about 4% of them were Xist (±) (unpublished data). Interestingly, Xist (±) PGCs continued to be present until their arrival at the genital ridges (maximum 47.2% at E8.5) (Figure 1J; Table 1). It is thus likely that X reactivation was in progress in those cells.
To examine the X reactivation process from a different perspective, we next performed single-cell reverse-transcriptase PCR (RT-PCR) analysis of X-linked gene expression in individual PGCs. We used an Oct4-green fluorescent protein (GFP) transgenic mouse line in which the PGCs were specifically marked by reporter expression (Figure S3) [32,33] and crossed transgenic female mice with male MSM/Ms mice, an inbred mouse strain derived from Mus musculus molossinus subspecies [34]. Because the genetic background of the Oct4-GFP mice was M. m. domesticus type, SNPs were readily detectable between the Oct4-GFP and MSM/Ms strains and were used to detect the allele-specific expression of each X-linked gene in the hybrids. GFP-tagged PGCs were picked individually from hybrid female embryos and subjected to the RT-PCR analysis. We analyzed the expression of ten X-linked genes, including Xist and Tsix, and three PGC markers (Stella, Oct4, and Mvh), as shown in Figure 2. Xist expression was examined using two pairs of primers mapped to exon 1 and exon 7 of the Xist gene. Mouse embryonic fibroblasts (MEFs) isolated from (BDF1 × MSM/Ms) F1 female embryos and hybrid female embryonic stem (ES) cells derived from (Oct4-GFP × MSM/Ms) F1 blastocysts were used as controls (Figure S4). In our hands, Xist expression was detected in about 73% of single MEFs (Figure S4; Table S1). In all MEFs analyzed, most of the X-linked genes were monoallelically expressed from either the maternal Oct4-GFP or the paternal MSM/Ms X chromosome, and no Tsix expression was detected. The biallelic expression of Zfp261 was detected in only one of the 26 MEF samples. It is known that both of the X chromosomes in female ES cells are active [6]. Our single-cell RT-PCR analyses successfully detected biallelic expression of all the X-linked genes tested in ES cells. Results obtained with MEFs and ES cells proved that our procedure could effectively distinguish mono- and biallelic expression of the X-linked genes in a single cell (Figure S4). In E7.75 PGCs, the expression of Xist was detected in 15 of 21 cells (71.4%) at a rate similar to that observed in MEFs (Figure 3A; Table S1). Most of the X-linked genes were monoallelically expressed in 13 of 21 samples. Interestingly, however, the biallelic expression of at least one gene was observed in eight of 21 PGC samples (Figures 2 and 3A; Table S1). In particular, in one PGC, all of the X-linked genes showed biallelic expression (Table S1). These findings demonstrate that one of two X chromosomes was inactivated in most of these early PGCs, which in turn suggests that the derepression of some of the X-linked genes begins in the newly formed PGCs or in even earlier PGC precursors. Tsix, the antisense transcript from the Xist locus, is thought to be a regulatory factor for X inactivation that acts by repressing Xist transcription [11,12]. Tsix expression was, however, barely detectable in early PGCs, and this trend persisted into later stages.
At E8.75, Xist-positive PGCs decreased to 42.1% (Figure 3A; Table S1), whereas the derepression of the X-linked genes did not proceed significantly from the E7.75 stage. At E10.5, Xist expression was no longer observed in PGCs, consistent with the RNA FISH results. Derepression of three X-linked genes, i.e., Np15, Fgd1, and Pdha1, was obvious in more than half the samples (Figure 3A; Table S1). At E12.5, the number of biallelically expressed genes increased even further; five of eight X-linked genes (Np15, Hprt, Fmr1, Fgd1, and Pdha1) were expressed biallelically in more than 50% of PGCs. However, there was no single cell in which X reactivation had been completed at this stage (Figures 3A and S4; Table S1). As shown in Figure 3, derepression of the X-linked genes did not appear to proceed linearly, probably due to the limited number of PGCs assayed. However, most of the X-linked genes tested showed gradual reactivation in the course of PGC development (Figure 3B).
The prevailing view suggests that X reactivation begins soon after the entry of PGCs into the genital ridge and before the initiation of meiosis [16–19]. However, even at E14.5, when most of the female PGCs are in meiotic prophase [3], derepression appeared to be incomplete. Two genes, Np15 and Hprt, were reactivated in all the samples, but other genes showed variable results (Figure 3B; Table S1). In particular, G6pd and Rex3 tended to show monoallelic expression. Because all the genes except Rex3 were completely reactivated in oocytes, E14.5 PGCs are probably still in the process of X reactivation.
Unexpectedly, we found that many single PGCs at E14.5 expressed one gene from the paternal X chromosome and the other from the maternal X chromosome. For example, as shown in lane 14–2 of Figure 2, in this particular cell, G6pd showed monoallelic expression from the maternal domesticus type X chromosome, whereas Rex3 was expressed from the paternal molossinus allele. Such “mosaic” patterns of expression were frequently (11/24) observed in E14.5 PGCs but rare in earlier PGCs. Only one each of PGC at E7.75 and E8.75 showed a mosaic pattern (Table S1). It is likely, therefore, that this is a phenomenon specific to E14.5 PGCs or to PGCs undergoing meiosis.
Here, we showed that the cessation of Xist RNA transcription had already been initiated in newly formed PGCs, and some X-linked genes were actually reactivated in these nascent PGCs. Furthermore, Xist accumulation was completely undetectable at E10.5, more than two days before the initiation of meiosis in female PGCs. However, the reactivation of X-linked genes was not complete even at E14.5, when meiotic division is initiated in most of female PGCs.
Despite their biological importance, the timing of both the inactivation and reactivation of the X chromosome has not been precisely determined in PGC. In previous studies of X reactivation, the results were based on the expression of protein products, which lags behind dynamic changes in RNA expression. Somatic contamination of germ cell samples might have disturbed previous studies. In contrast, we assessed the inactivation status in PGCs using whole-mount RNA FISH and measured X chromosome activity directly by detecting the allelic transcripts of ten X-linked genes in single PGCs. The use of multiple appropriate markers for each stage of PGC development further increased the validity of our assay.
Among the earliest PGCs marked by Blimp1 expression, there were some cells showing no Xist expression, and hence undergoing X reactivation. This may imply delayed X inactivation in the germ line as suggested previously by Tam et al. [18], who assessed the X inactivation and reactivation status by examining the X-linked lacZ transgene expression. However, even at E7.0 a large number of cells (77.5%) possessed Xist paint signals, indicating that most of the PGCs and/or their precursor cells are not exempt from X inactivation. Because the numbers of such Xist (+) cells decreased and Xist (−) cells increased progressively in number, the simplest interpretation is that while every epiblast cell in the early egg cylinder undergoes X inactivation once, PGC precursor cells in the epiblast are somehow susceptible to reactivation. Thus, X reactivation is likely to occur slightly before the advent of PGC. It is tempting to speculate that the initiation of X reactivation coincides with the determination of the germ-cell lineage.
We found that a significant number of X-linked genes showed monoallelic expression even though Xist painting was no longer observed. Similar observation made by Csankovszki et al. [35] suggests that continued Xist expression is not essential for maintenance of the inactive state after establishment of the X inactivation in vitro. Here, we showed that Xist expression and accumulation were hardly detected in E10.5 PGCs, but several X-linked genes were still monoallelically expressed even at E14.5. To our knowledge, these results are the first to demonstrate that the maintenance of X inactivation does not require Xist RNA accumulation on the inactive X chromosome in normal embryonic development in vivo.
Furthermore, we observed “mosaic” patterns of mono- and biallelic expression of X-linked genes in E14.5 PGCs in which no Xist signals were detected as described above. This phenomenon could be explained by very low expression levels of these genes at this stage, causing biased PCR amplification resulting in “pseudo” monoallelic expression. However, our microarray data (N. Mise and K. Abe, unpublished data) showed that steady state mRNA levels of these genes were high (and constant) at both E12.5 and E14.5 stages. Therefore, it is likely that the mosaic expression pattern may be caused by intermingling of inactive and active chromosomal regions through meiotic recombination. Further investigations on this mosaicism would provide insight into roles of “chromatin environment” in establishment and maintenance of X inactivation status.
Another important finding presented here is that X reactivation is not completed within a narrow window of time during development. Genome reprogramming events in PGCs represented by imprinting erasure and X reactivation are thought to take place over a short period of time, perhaps during the transition from pregonadal to gonadal PGCs [2]. Our results demonstrate that X reactivation proceeds gradually in a stepwise fashion, requiring a relatively long time (≥7 d). As the first step, the cessation of Xist occurred, and reactivation of a few X-linked genes already commenced during E7.0–E10.5. By E10.5, Xist accumulation on the inactive X chromosome has almost completely gone. After E10.5, reactivation of X-linked genes was accelerated significantly, and most of the X-linked genes tested showed biallelic expression in the next two days. As this second phase of reactivation is more prominent than the first one, previous studies might have overlooked the early commencement of the X reactivation. However, the X reactivation in fact has initiated as soon as PGC differentiation occurs, and this fact may change a general view on X inactivation/reactivation and genomic reprogramming phenomena in the germ cell lineage. During PGC development, their number increases from 50 to 25,000 [36], involving more than nine cell divisions. The activity status of the X chromosomes of PGCs during this period appears to vary from cell-to-cell and from gene-to-gene. There are certain tendencies for some genes to be reactivated earlier than other genes, and one could also argue that genes distant from the Xist locus tend to be reactivated earlier than genes close to the Xist. Further extended analyses on expression of other X-linked genes and their surrounding genomic sequences will be needed to make definitive statements on this point.
The present study suggests that X reactivation in PGCs is not involved in active epigenetic remodeling, but may instead involve slow passive steps that require many cell divisions. In contrast, X reactivation events observed at peri-implantation stages (and zygotic gene activation) are completed in a much shorter period of time [21–23]. Kinetics of X reactivation in PGCs is rather similar to that of the reversal of X inactivation induced by cell fusion between an embryonal carcinoma cell and female somatic cell [37], for which four to five cell divisions are required [38]. These differences in kinetics of each reactivation process seem to reflect differences in mode of epigenetic regulation operating in random and imprinted X inactivation.
For maintenance of random X inactivation, methylation of CpG sites at the 5′ region of X-linked genes is important (reviewed in [6]). In a mutant deficient for maintenance DNA methyltransferase (Dnmt1), random X inactivation in the embryonic lineage is unstable, whereas the imprinted X inactivation in the extraembryonic lineage was unaffected [39], suggesting differences in the epigenetic state of the X chromosome in the two lineages. It is generally thought that global DNA methylation level is relatively low in the extraembryonic lineage and that imprinted X inactivation in this lineage is less stable compared to random X inactivation in embryonic lineage. In the extraembryonic lineage, therefore, the epigenetic gene-silencing mechanism may not rely on DNA methylation [39,40], and erasure of the repressive state may be more readily accomplished. In contrast, DNA methylation is apparently important for establishment and maintenance of random X inactivation in embryo proper [39], and reactivation of the DNA methylation-dependent inactive state may require a longer period. Seki et al. [41] reported that global DNA methylation is substantially removed from migrating PGCs at around E8.0, and we also found that expression of all the DNA methyltransferases is hardly detected at an even earlier stage (S. Kobayakawa and K. Abe, unpublished data), suggesting that genome-wide erasure of DNA methylation commences from a very early stage of PGC development. Seki et al. [41] also demonstrated that global DNA methylation was further reduced between the stages E9.5 and E12.5. Such two-step reduction of global DNA methylation approximately coincides with the stepwise processes of X reactivation described here. Therefore, the unique epigenetic properties of the PGC genome likely constitute a foundation for the prolonged reactivation process.
As described, X reactivation may be a consequence of genome-wide epigenomic remodeling in PGCs, but the reactivation should have its own biological significance. One obvious reason is to avoid producing “inactive X–Y” offspring, in which no active X chromosome exists. On the other hand, female mice carrying only one X chromosome, X0 mice, are fertile, and in X0 female embryos PGCs have a single active X chromosome throughout their development. During maturation, some X0 oocytes release a first polar body containing the X chromosome and become mature oocytes without the X chromosome. When these “0-oocytes” fertilize with “X-sperm,” they develop as phenotypically normal female mice. Therefore, twice the amount of transcripts from the X chromosome is not essential for female germ cell development.
We think that the presence of two active (or euchromatic) X chromosomes may be needed for successful meiotic recombination. The inactive X chromosome is highly heterochromatic (reviewed in [6]), and heterochromatic chromosomal regions are generally underrepresented in the synaptonemal complex [42]. Pairing of active and inactive X is thus difficult to achieve, resulting in strong recombination suppression. Therefore, X reactivation in germ cells may have significance for appropriate segregation of X chromosome into mature oocytes.
In conclusion, we made an unexpected finding that initiation of X reactivation coincides with onset of PGC formation. Germ cell specification appears to be associated with genome-wide epigenetic reprogramming [41], raising the issue of how chromosome-wide X reactivation and genome-wide epigenetic changes are related. The kinetics of the X reactivation indicates that this process requires a prolonged period, suggesting that passive mechanisms may be involved. We found negligible Tsix expression in PGCs during the reactivation process. Recently, it was shown that CpG methylation of the Xist promoter region is mediated by Tsix transcription itself [7,43]. However, results presented in this study suggest that repression of Xist transcription during the reactivation process in PGCs is probably not mediated by Tsix transcription. Therefore, X reactivation is not exactly a reversal process of X inactivation. Apparently X reactivation process holds many important, unanswered questions, and we believe that investigation on this process will lead to better understandings of mechanisms of X inactivation, genomic reprogramming, and germ–soma differentiation in mammals.
To create a transgenic construct for mRFP under the control of the Blimp1 regulatory elements, we used a highly efficient Escherichia coli-based chromosome engineering system [44]. Homologous sequences for Blimp1 were introduced to both ends of the mRFP-BGHpA-FRT-Neo-FRT sequence by PCR. Primers used for this amplification were 5′-CTAGCTCCGGCTCCGTGAAGTTTCAAGGACTGGCAGAGACTGGGATCATGATGGCCTCCTCCGAGGACGT-3′ and 5′-AACTCGGCCTCTGTCCACAAAGTCATATCAGCGTCCTCCATGTCCATTTTTGTGGAATTGTGAGCGGATA-3′. The resulting amplified fragment was recombined immediately after the methionine of exon 3 in the Blimp1 gene in a 203-kb BAC, RP23-1D12, purchased from BACPAC Resources Center (http://bacpac.chori.org), carrying the 146-kb upstream region from the transcription start site and the 35-kb downstream region from the 3′ end of the 3′ untranslated region of Blimp1 (Figure S2). The mRFP cDNA clone was kindly provided by Dr. Roger Tsien (Howard Hughes Medical Institute at the University of California, San Diego, California, United States, http://www.tsienlab.ucsd.edu).
The Oct4-GFP transgenic mouse line has been described previously [33] and is available from the RIKEN BioResource Center (http://www.brc.riken.jp/inf/en) as the TgN(deGFP)18Imeg strain. These mice carry the GFP sequence driven by an 18-kb Oct4 genomic fragment with deletion of the proximal enhancer [32].
The MSM/Ms mouse strain was derived from M. m. molossinus mice [34]. Because the BDF1 (M. m. domesticus) and MSM/Ms mice are quite divergent, this F1 hybrid is rich in SNPs.
All methods were approved by the Institutional Animal Experiment Committee of RIKEN BioResource Center.
A probe to detect Xist RNA was prepared by nick translation with Cy3-dCTP (GE Healthcare, http://www.gehealthcare.com) or SpectrumGreen-dUTP (Vysis, http://www.vysis.com) from an equimolar mixture of a series of Xist cDNA clones encompassing exons 1–7 (kindly supplied by Dr. Takashi Sado) [12]. Cot-1 DNA was also labeled by nick translation with Cy5-dCTP (GE Healthcare).
Isolated embryos were incubated in 0.5% Triton X-100 in PBS for 3–5 min on ice and fixed with 4% paraformaldehyde in PBS for 10 min at room temperature. Staging of the embryos was according to Downs and Davies [45] and Kaufman [46]: Our E7.75 embryos corresponded to either late bud (LB) or early headfold (EHF) stages [45]; E8.5 embryos corresponded to Theiler's stage 12, and E9.5 embryos were at Theiler's stage 14 [46]. Appearances of Oct4-GFP-positive PGCs were shown in Figure S3. Hybridization was carried out at 37 °C overnight. Following stringent washing, the embryos were incubated for 1 h at room temperature with primary antibody diluted with blocking buffer (1% BSA, 0.1% Tween 20, and 4× SSC): either 1:200 anti-Oct4 (sc-8628, Santa Cruz Biotechnology, http://www.scbt.com), 1:1,000 anti-Stella (kindly supplied by Dr. Toru Nakano) [28], or 1:500 anti-RFP (MBL, http://www.mbl.co.jp/e). The embryos were incubated in secondary antibody diluted in blocking buffer (1:500 Alexa-Fluor-488-conjugated rabbit antigoat IgG, Alexa-Fluor-488-conjugated goat antirabbit IgG, or Alexa-Fluor-555-conjugated goat antirabbit IgG, Invitrogen, http://www.invitrogen.com) for 45 min at room temperature. Because of the Triton X-100 treatment before fixation, mRFP expression in visceral endoderm was faded as shown in Figure 1I. Double staining with another PGC marker confirmed that mRFP expression was specific to PGCs. The embryos were mounted in 90% glycerol, 0.1× PBS, and 1% Dabco (Sigma-Aldrich, http://www.sigmaaldrich.com). Fluorescent images were taken with an LSM510 meta confocal laser scanning microscope (Zeiss, http://www.zeiss.com).
Embryos from Oct4-GFP × MSM/Ms matings were dissected at each stage. Following trypsinization, single GFP-positive cells were picked and stored at −80 °C until use. As control cells, single embryonic fibroblast cells prepared from E12.5 female embryos derived from BDF1 × MSM/Ms matings and single (Oct4-GFP × MSM/Ms) F1 hybrid ES cells, which were established previously in our laboratory (N. Mise and K. Abe, unpublished data), were used. The sexes of the E12.5 and E14.5 embryos were determined by the morphology of the genital ridges and by PCR in the earlier stage embryos. A single primer pair (5′-TGGATGGTGTGGCCAATG-3′ and 5′-CACCTGCACGTTGCCCTT-3′) amplifies both the X-linked Ube1x and Y-linked Ube1y sequences but yields products of different sizes (Ube1x, 252 bp and Ube1y, 334 bp).
Single cells were incubated in reverse transcription buffer supplemented with 0.1% NP-40 and 0.5 U of RQ1 RNase-free DNase (Promega, http://www.promega.com) for 15 min at 37 °C, for 3 min at 75 °C, and for 5 min on ice. Reverse transcription was carried out by adding 0.5 μl of 0.5 μg/μl oligo dT18 primer, 0.5 μl of 10 mM dNTP mix (Invitrogen), and 0.5 μl of 200 U/μl SuperScript III reverse transcriptase (Invitrogen). This was followed by incubation at 50 °C for 1 h. The reactions were incubated at 37 °C for 15 min with 1 U of RNase H (Invitrogen). We carried out two rounds of PCR amplification of the cDNA to detect 11 sequences of ten X-linked genes (exons 1 and 7 of Xist, Tsix, Np15, Hprt, Fmr1, G6pd, Zfp261, Rex3, Fgd1, and Pdha1) and the sequences of three PGC markers (Stella, Oct4, and Mvh) in single cells. The X-linked genes, apart from Xist and Tsix, were selected because they showed relatively high expression in PGCs according to our microarray analysis (unpublished data). We designed these primer sequences carefully to avoid amplifications of pseudogenes and to include SNPs in the amplicons for distinction of BDF1 and MSM/Ms alleles. For the first PCR, we used a mixture of all the primers listed in Table S2 to amplify all of the sequences in 100-μl reactions. Aliquots (0.5 μl) of the first PCR products were used as templates for the second PCR in 20-μl reactions. Each primer pair listed in Table S2 was used to amplify specific sequences in each single-cell-derived sample. For each series of experiment, we included negative controls; a single cell was processed in the same way as the experimental group except for the addition of reverse transcriptase. We had no amplification at all from negative controls. The second PCR products were digested with the appropriate restriction enzymes (Table S2). |
10.1371/journal.ppat.1002048 | Viral CTL Escape Mutants Are Generated in Lymph Nodes and Subsequently Become Fixed in Plasma and Rectal Mucosa during Acute SIV Infection of Macaques | SIVmac239 infection of rhesus macaques (RMs) results in AIDS despite the generation of a strong antiviral cytotoxic T lymphocyte (CTL) response, possibly due to the emergence of viral escape mutants that prevent recognition of infected cells by CTLs. To determine the anatomic origin of these SIV mutants, we longitudinally assessed the presence of CTL escape variants in two MamuA*01-restricted immunodominant epitopes (Tat-SL8 and Gag-CM9) in the plasma, PBMCs, lymph nodes (LN), and rectal biopsies (RB) of fifteen SIVmac239-infected RMs. As expected, Gag-CM9 did not exhibit signs of escape before day 84 post infection. In contrast, Tat-SL8 escape mutants were apparent in all tissues by day 14 post infection. Interestingly LNs and plasma exhibited the highest level of escape at day 14 and day 28 post infection, respectively, with the rate of escape in the RB remaining lower throughout the acute infection. The possibility that CTL escape occurs in LNs before RBs is confirmed by the observation that the specific mutants found at high frequency in LNs at day 14 post infection became dominant at day 28 post infection in plasma, PBMC, and RB. Finally, the frequency of escape mutants in plasma at day 28 post infection correlated strongly with the level Tat-SL8-specific CD8 T cells in the LN and PBMC at day 14 post infection. These results indicate that LNs represent the primary source of CTL escape mutants during the acute phase of SIVmac239 infection, suggesting that LNs are the main anatomic sites of virus replication and/or the tissues in which CTL pressure is most effective in selecting SIV escape variants.
| Strong antiviral CD8+ T lymphocytes limit SIV replication by recognizing short pathogen-derived peptide epitopes. The cytotoxic CD8+ T cell responses specific for this highly mutable virus can select for viruses bearing mutations that prevent CD8+ T cell recognition of cells infected with these escape mutants. To determine the anatomic origin of these escape mutants, we tracked a particular escape mutant in multiple tissues (plasma virus, lymph nodes, rectal mucosa, and peripheral blood immune cells) during the early, acute phase of SIVmac239 infection of rhesus macaques. We found that escape mutants first reach high frequency in lymph nodes 2 weeks after infection, and the particular mutants generated in lymph nodes disseminate to other tissues by week 4. Furthermore, we found that epitope-specific CD8+ T lymphocyte responses in the lymph nodes and peripheral blood, but not the gut mucosa, are significantly correlated with the frequency of escape mutants in the plasma virus at week 4. This suggests that lymph nodes, and not the gut, are the primary site of anti-SIV CD8+ T cell responses and/or SIV replication during the acute phase of infection.
| Human immunodeficiency virus (HIV) infection of humans and Simian Immunodeficiency Virus (SIV) infection of rhesus macaques (Macaca mulatta, RM) results in a progressive and irreversible decline of immune function characterized by depletion of CD4 T cells, chronic immune activation, and high susceptibility to opportunistic infections that is commonly referred to as AIDS.
While the host immune system mounts strong cellular and humoral immune responses against HIV and SIV, these responses ultimately fail to control virus replication in the overwhelming majority of infected individuals. A key reason underlying this immune failure is the extreme genetic variability of these primate lentiviruses, which occurs as a result of a high mutation rate caused by the relative infidelity of the HIV and SIV reverse transcriptases [1]. This extreme genetic diversity combined with a large in vivo effective population size [2] practically ensures that the virus will always be able to evade or “escape” from recognition by the host immune system. In clinical terms, these biological features of HIV in the absence of “natural immunity” against the virus are key indicators of the complexity and difficulty that the scientific community faces when trying to design an effective AIDS vaccine.
In the absence of immunogens that are able to predictably elicit the production of broadly reactive HIV- or SIV-specific neutralizing antibodies [3], [4], there has been significant interest in immunogens that elicit strong antiviral CD8+ T cell-mediated cytotoxic T lymphocyte (CTL) responses [5]. A large body of evidence indicates that CD8+ T cells do play a significant role in the control of HIV and SIV replication that involves both cytolytic and non-cytolytic mechanisms. First, CD8+ T cells can inhibit HIV and SIV replication in vitro [6], [7]. Second, there is a temporal association between post-peak decline of acute viremia and emergence of CD8+ T cell responses [8], [9]. Third, antibody-mediated in vivo depletion of CD8+ lymphocytes is consistently associated with increased virus replication in SIV-infected RMs [10], [11], [12]. Fourth, there is a strong association between specific major histocompatibility complex (MHC) alleles and ability to control virus replication during HIV and SIV infection [13]. In this context, the fact that CTL escape mutants consistently arise during both acute and chronic HIV/SIV infections [13] demonstrate the presence of selective CD8+ T cell-mediated immune pressure on the virus population. On the other hand, the fact that escape mutants are consistently observed is also an indicator of the overall inability of these cells to fully suppress virus replication.
SIVmac239 infection of RMs bearing the MamuA*01 MHC class I allele elicits CD8+ T cell responses against two very well characterized immunodominant epitopes: Tat-SL8 [14] and Gag-CM9 [15]. While SL8- and CM9-specific CD8+ T cell responses are both generated during the acute phase of infection [14], [16], the emergence of CTL escape mutants occurs much more rapidly in the Tat-SL8 epitope than it does in Gag-CM9 [14], [17], [18], presumably due to strong functional constraints imposed on the gag gene and the need for extra-epitopic, secondary compensatory mutations to allow effective virus replication [19], [20], [21]. Of note, in these studies the kinetics of the generation and fixation of CTL escape mutations occurring in the SL8 and CM9 regions of SIV was analyzed only in plasma virus, which is thought to provide an overall representation of the virus that are replicating within the host. However, to the best of our knowledge, a comprehensive and comparative longitudinal analysis of CTL escape mutants in the mucosal and lymphoid tissues of SIV-infected RMs has not been yet been conducted. Determining if CTL escape mutants emerge more rapidly in lymphoid or, alternatively, mucosal tissues would provide important information with respect to the predominant sites of CD8+ T cell-mediated immunological pressure in vivo.
In this study, we characterized the appearance, dynamics, and dissemination of CD8+ T cell escape mutants in lymphoid vs mucosal tissues during SIVmac239 infection of RMs. To this end, we conducted an extensive longitudinal assessment of viral sequences derived from plasma viral RNA as well as cell-associated viral DNA in peripheral blood mononuclear cells (PBMC), lymph node biopsies (LN), and biopsies of the rectal mucosa (RB) of 15 SIV-infected RMs. These animals were included in a previously published study designed to investigate the immunogenicity and protection from SIVmac239 challenge conferred by two MVA-base candidate AIDS vaccines expressing SIVmac gag and tat [22]. As expected, we observed that CTL escape mutants occurred early in the Tat-SL8 epitope and much later in the Gag-CM9 epitope. Importantly, we found that in all RMs, Tat-SL8 escape mutants appeared earlier and in higher frequency in LNs then in RBs, with variants found at high frequency in LNs at day 14 post infection becoming dominant in the RBs at day 28 post infection. These results indicate that LNs represent the primary source of CTL escape mutants during the acute phase of SIVmac239 infection, suggesting that LNs are the main anatomic sites of virus replication and/or the tissues in which CTL pressure is most effective in selecting SIV escape variants.
A group of fifteen MamuA*01-positive rhesus macaques (Macaca mulatta; RM) were infected intravenously with 10,000 TCID50 of SIVmac239 as part of a previous study designed to assess the immunogenicity and potential protection from challenge conferred by MVA-based candidate AIDS vaccines [22]. In this study, ten RMs were immunized three times with either MVA “wild-type” (5 animals) or a genetically engineered MVA in which the gene for the Uracil DNA Glycosidase (UDG) was deleted, i.e. ΔUDG (5 animals), with both vectors expressing the Gag and Tat proteins of SIVmac239. Five additional unvaccinated RMs were used as controls. In all animals, SIV-specific CD8+ T cell responses were measured at various time points post-immunization and post-challenge in multiple tissues, including PBMC, RB, and LN by tetramer staining for the Gag-CM9 and Tat-SL8 epitopes of SIVmac239. The main results of this study are summarized in Table S1. Briefly, administration of MVA-SIV immunogens resulted in a partial (∼1 log) and transient (60–120 days) decline in plasma viral load that did not translate into protection from CD4+ T cell depletion and disease progression. Of note, in four RMs the level of virus replication decreased to near undetectable levels during the chronic phase of infection, suggesting that these “controller” animals were able to mount antiviral immune responses that could successfully suppress virus replication in vivo. The large amount of virologic and immunologic data collected in this study combined with an extensive archive of tissue samples presented an ideal opportunity to explore in detail the relationship between SIV-specific CD8+ T cell responses in various tissues, and the appearance and dissemination of CTL escape mutants in a relatively large cohort of SIVmac239-infected RMs. Due to relatively modest protective effect of the used immunization regimen [22], we chose to conduct our analysis in the entire group of 15 SIV-infected animals without dividing them into vaccinated and controls.
To characterize the emergence of CTL escape mutants in our cohort of SIVmac239-infected RMs, we amplified a 435 nucleotide region surrounding the Gag-CM9 epitope and a 390 nucleotide region surrounding the Tat-SL8 epitope from reverse transcribed plasma viral RNA and genomic DNA derived from PBMCs, RB, and LNs collected at multiple time points post infection. These amplicons were then sequenced from one end using Roche's 454 pyrosequencing technology. After eliminating all sequence reads that did not meet the minimum set of quality criteria (see Materials and Methods), source animals for each read were identified via pre-determined barcodes, and the reads were aligned with the corresponding wildtype SIVmac239 sequence. High frequency insertions and deletions (indels) that resembled well-characterized artifacts of the 454 pyrosequencing procedure were repaired with reference to the wildtype SIVmac239 sequence. All resulting sequences that contained no indels and were at least 243 nucleotides long for Gag-CM9 and 220 nucleotides long for Tat-SL8 were included in all subsequent analyses. To ensure that the sequence selection process did not bias our results, all analyses were performed on the full set of reads that had been repaired indiscriminately using the SIVmac239 wildtype sequence. No bias was found in our results (data not shown).
HIV and SIV gag genes are highly conserved and thus escape mutations in CD8+ T cell epitopes in this region are typically associated with either highly unfit viruses or the appearance of compensatory mutations outside of the epitope itself [20]. In order to characterize the rate and mechanisms of escape in the highly conserved Mamu-A*01-restricted Gag-CM9 epitope in our SIVmac239 infected RMs, we amplified and sequenced this epitope from plasma virus and cell-associated viral DNA in PBMCs, LN, and RB. Escape in Gag-CM9 occurred via mutations at the second position (threonine) in the epitope (Figure 1A). The substituted amino acids observed included serine, isoleucine, and cysteine. Interestingly, the threonine to cysteine amino acid substitution is achieved via a nucleotide substitution at the first and second positions of the threonine codon and was observed in the RB but not other tissues of a single animal, RDo8 (data not shown). There was no evidence of a significant increase in the frequency of the intermediate codons, thus suggesting that viruses bearing the resultant amino acid substitutions are at a severe selective disadvantage. However, it is also possible that the frequency of sampling during the chronic phase of SIVmac239 infection was not sufficient to detect the presence of relatively transient intermediate amino acid substitutions. As expected based on previous studies [17], [18], [20], escape in Gag-CM9 was observed in only two RMs (RDo8 and RWi8) and not until day 84 post infection, when escape mutants appeared in plasma virus and, although at lower frequencies, in PBMC-derived cell-associated virus (Figure 1B). In these two animals viruses isolated from RBs at day 168 post infection were almost entirely comprised of Gag-CM9 escape mutants (Figure 1B).
Previous studies of the kinetics of emergence of CTL escape in the Mamu-A*01-restricted Tat-SL8 immunodominant epitope have shown that in plasma virus, Tat-SL8 escape occurs during the early stages of SIV infection and via multiple amino acid substitutions [14], [23], [24]. For this reason, we focused our analysis of the emergence of Tat-SL8 escape mutants in our group of SIVmac239-infected RMs during the acute phase of infection. As described for the Gag-CM9 epitope, we PCR amplified the viral genomic region surrounding the Tat-SL8 epitope from reverse transcribed plasma viral RNA and sequenced it by 454 pyrosequencing.
As expected based on previous studies [14], [23], [24], the emergence of Tat-SL8 escape mutants consistently occurred at high frequency during acute SIV infection. In particular, we observed that CTL escape in Tat-SL8 occurred through one or more of several amino acid substitutions: serine to proline or phenylalanine at position 1, threonine to isoleucine at position 2, serine to leucine at position 5, and alanine to aspartic acid at position 6 (Figure 2A). These amino acid substitutions occur through single nucleotide mutations and have all been characterized as escape mutations in previous studies [14], [23], [24]. As shown in Figure 2B, Tat-SL8 escape mutants began to emerge by day 14 post infection in most RMs, with especially high levels of escape mutants evident in 4 animals (RDo8, ROu8, RWi8, RWu8). By day 28 post infection, the majority of plasma virus in all SIV-infected RMs was comprised of Tat-SL8 escape mutants, with the median frequency of the wildtype Tat-SL8 epitope at 0.08 (range, 0.008–0.261; Figure 2B). The frequency of the viruses bearing the wildtype Tat-SL8 epitope continued to decrease through day 84 post infection (mean, 0.01; range, 0.004–0.027). Overall, these data confirm the early emergence of numerous Tat-SL8 escape mutants, a finding that reflects both the selective pressure exerted by Tat-SL8-specific CD8+ T cell responses and the relative genetic flexibility of Tat in maintaining its function despite the presence of these mutations [19].
To investigate the kinetics of appearance and dissemination of escape mutants in SIVmac239 infection in different anatomic compartments, we next compared both the frequency and the character of Tat-SL8 CTL escape mutants in viral sequences derived from plasma virus and cell-associated viral DNA from peripheral blood mononuclear cells (PBMCs), lymph nodes (LNs) and intestinal mucosa that was sampled by rectal biopsies (RB). This analysis was performed on samples obtained at day 14 and day 28 post infection. While Tat-SL8 escape mutants were evident in all four examined tissues at day 14 post infection, LNs exhibited a significantly higher frequency of escape mutants than either plasma or RB (Figure 3A; p = 0.0022, Kruskal-Wallis with Dunn's multiple comparisons). Interestingly, the discrepancy in the level of escape between LNs and plasma virus may be seen as reflecting the fact that circulating virus at the peak of acute SIV infection originates primarily from other tissues. By day 28 post infection, plasma virus exhibited a greater level of Tat-SL8 escape than cell-associated virus sampled from either PBMCs or RBs (Figure 3B; p = 0.0001, Kruskal-Wallis with Dunn's test for multiple comparisons), but not from LNs. This finding suggests that plasma virus during the phase of post-peak decline in viremia is likely to originate primarily in lymphoid tissues.
Escape in Tat-SL8 can occur through multiple amino acid substitutions [14], [23], [24]. In order to further delineate the sources of actively replicating viruses that contain Tat-SL8 escape mutants during acute SIVmac239 infection of RMs, we next characterized the frequency distribution of distinct intra-epitopic Tat-SL8 escape mutations. We found that, in the majority of RMs, Tat-SL8 escape mutants sampled from LNs at day 14 post infection were dominated by a serine to proline amino acid substitution at the first position of the epitope, although amino acid substitutions at other positions were observed at much lower frequency (Figure S1A). This pattern of amino acid substitution observed in LNs appeared to be different from what we observed in the Tat-SL8 genotypes sampled from the other three examined tissues, in which there was no clearly dominant amino acid substitution (Figure S1A). In particular, the distribution of Tat-SL8 escape mutants in the plasma virus was much more similar to that of cell-associated viral DNA from PBMCs, with small, equivalent frequencies of mutation at the first, second, fifth and eighth positions, than that of either RBs or LNs. Of note, by day 28 post infection, the distribution of Tat-SL8 escape mutants had become more similar across all tissues (Figure S1B). However, while we observed an increase in the relative abundance of mutations at positions 2 through 8 in LNs, the serine to proline mutation at the first amino acid position of Tat-SL8 continued to be dominant (Figure S1B). In fact, this particular escape mutation became the most frequently sampled mutation in all tissues at day 28 post infection (Figure 3C; 2-way ANOVA with Bonferroni multiple comparison test, p<0.001), and continued to dominate viral populations through the latest time points sampled from all tissues (data not shown). The persistent high frequency of this escape mutation at the first position of Tat-SL8 suggests that anti-SIV immune responses during the acute infection in lymphoid tissues rather than at mucosal sites have a more lasting effect on the evolution of the SIV viral population.
Interestingly, three of the four RMs exhibiting relatively large numbers of escape mutants at day 14 post infection had been vaccinated with the MVA-SIV vector (Figure S1A), thus suggesting an early CTL-driven selection of these viral variants. Indeed, vaccinated RMs showed a non-significant trend towards increased frequencies of Tat-SL8 escape mutants at day 14 post infection compared to unvaccinated controls (p>0.05, Kruskal-Wallis with Dunn's test for multiple comparisons). However, vaccination did not preferentially increase the level of Tat-SL8 specific CD8+ T cells in LNs compared to PBMCs and RBs at days −28, 0, and 7 post infection (Figure S2). Taken together these observations argue against the possibility that the expansion of Tat-SL8 escape mutants in LNs was caused primarily by vaccine-induced prior expansion or redistribution of Tat-SL8-specific CD8 T cells to lymphoid tissues.
The large amount of virological and immunological data collected during the course of SIVmac239 infection in the RMs included in this study allowed us to probe whether specific immune responses or virological outcomes (see Table S1) were associated with the levels of CTL escape mutants observed at days 14 and 28 post infection. We did not find any significant correlations between the levels of escape mutants in any tissue at day 14 post infection and the levels of SIV-specific CD8+ T cell-mediated immune responses as measured by tetramer staining at the same time point. Furthermore, none of the virological parameters measured at days 14 and 28 post infection correlated with the level of escape mutants present in any tissue at the same time points. However, we observed a clear correlation between immune responses at day 14 post infection and the level of Tat-SL8 escape present in several tissues at day 28 post infection (Figure 4). First, the frequency of viruses bearing the wildtype Tat-SL8 epitope among plasma viruses at day 28 post infection was significantly inversely correlated with the abundance of anti-Tat-SL8 CD8+ T cells at day 14 post infection in LNs (Figure 4A; Spearman's correlation, r = 0.6592, p = 0.0075) and PBMCs (Figure 4B; Spearman's correlation, r = 0.5836, p = 0.0224), but not RBs (Figure 4C; Spearman's correlation, r = 0.4850, p = 0.787), suggesting that LNs are the major anatomic sites where the anti-SIV CD8 T cell response is exerting immunological pressure at the peak of acute SIVmac239 infection. Second, the frequency of wildtype Tat-SL8 epitope in RBs and PBMCs at day 28 post infection was strongly correlated with levels of Tat-SL8 CD8+ T cells in RBs at day 14 post infection (Figure 4D–E; Spearman's correlation, RB: r = 0.7992, p = 0.001; PBMC: r = 0.6485, p = 0.0121), perhaps suggesting a local effect of mucosal SIV-specific CTL responses in determining the emergence of escape mutants. Third, the levels of Gag-CM9-specific CD8+ T cell responses in RBs at day 14 post infection also correlated strongly with the frequency of tat-SL8 escape mutants in RBs (Figure 4F; Spearman's correlation, r = 0.6731, p = 0.0164), but not PBMCs (data not shown) at day 28 post infection. Interestingly, despite the significant correlations between the frequency of Tat-SL8 escape and the magnitude of anti-SIV CD8 T cell responses in various tissues, the decline in CD4+ T cells in RB and PBMC did not correlate with the frequency of escape at 14 and 28 days post infection in any tissue (data not shown). Taken together these findings indicate that SIV-specific CTL responses at the peak of acute infection have a strong impact on the resultant viral population, with the pattern of observed circulating virus variants at day 28 post infection being shaped predominantly by LN-based immune responses.
Four SIV-infected RMs included in the current study (three vaccinated and one control) were able to spontaneously control virus replication to near undetectable levels shortly after the resolution of acute viremia (Table S1). These RMs exhibited a ∼1 log lower peak of acute viremia, yet similar levels of SIV-specific CD8+ T cell responses as the normal non-controller animals. To understand the contribution of CD8+ T cell escape to the control of SIV replication, we compared the prevalence of Tat-SL8 escape mutants and the diversity of virus sequences between the controller and non-controller groups of SIV-infected RMs. Control of SIVmac239 replication in RMs was associated with increased levels of viruses bearing the wildtype Tat-SL8 epitope among infected cells in PBMCs (Figure 5A; 2-way ANOVA using the Bonferroni Correction for multiple comparisons) while viruses circulating in the plasma were primarily escape mutants in both groups (Figure 5B). Similarly, viruses sampled from RBs in controllers at day 168 post infection were also composed primarily of viruses bearing the wildtype Tat-SL8 epitope (data not shown). It should be noted, however, that virus was amplifiable from day 168 RB samples of only two of the four controller. Finally, Tat-SL8 and the surrounding region more closely resembled wildtype SIVmac239 in viruses sampled at late time points from PBMC, but not plasma, of controllers than non-controllers (Figure 5C–D; 2-way ANOVA using the Bonferroni Correction for multiple comparisons). In the setting of low virus replication, as in our controller SIV-infected RMs, virus isolated from cellular samples may include a large representation of archival sequences as compared to plasma. As such, it is possible that the large frequency of wildtype viruses in the controllers is a consequence, rather than a cause, of the low virus replication.
Despite the presence of a strong antiviral immune response, HIV infection of humans and SIV infection of RMs usually results in a chronic and progressive immunodeficiency. The inability of anti-SIV immune responses to effectively control virus replication is at least partially due to the singular ability of this primate lentivirus to evade immune responses via a high rate of genetic mutation [13]. The epidemic circulation of extremely diverse viral subtypes and the immune-mediated selection of viral escape variants are two of the main problems that must be overcome by a successful candidate HIV vaccine [25]. In the setting of experimental SIV infection of non-human primates, the comparison of virus sequences obtained in different tissues might be used to determine the anatomic origin of these escape variants. We postulate that this analysis will help us define the anatomic compartments where immune responses have their greatest impact on virus replication and fitness. In addition, as plasma virus includes all viral variants produced in the body at any given time, a comparative kinetic analysis of the emergence of viral escape mutants in plasma vs. tissues could provide an indirect but robust measure of the extent of virus replication in the examined tissues. Identifying these anatomic sites of high viral replication and dominant immune pressure during the acute phase of HIV/SIV infection might help the design of a vaccine that can elicit adaptive immune responses capable of durable control of HIV replication.
In order to define these sites of virus replication and immune pressure during acute SIVmac239 infection of RMs, we investigated the kinetics and characteristics of escape in two well characterized Mamu-A*01-restricted CTL epitopes (Gag-CM9 and Tat-SL8). We sequenced the SIV genomic regions containing these epitopes in longitudinal archival samples of plasma virus and cell-associated viral DNA from PBMCs, LNs and RBs collected from a cohort of vaccinated and challenged RMs [22]. Ten of the rhesus macaques in this study had been vaccinated with MVA-derived vectors expressing the Gag-CM9 and Tat-SL8 epitopes, while five of the challenged RMs were unvaccinated. While our data do not rule out the possibility that other tissues (i.e., spleen, liver, etc) are significant contributors to plasma viremia and/or immune pressure, the unique aspect of this study is the direct comparison between plasma, LNs, PBMCs, and RBs.
As expected, Gag-CM9 escape mutants were observed in only two animals and after day 84 post infection, while the Tat-SL8 epitope acquired escape mutations very quickly during the acute phase infection. Interestingly, the comparative analysis of Tat-SL8 escape mutants revealed, at day 14 post infection, a higher frequency in LN as compared to other tissues. By day 28 post infection, Tat-SL8 escape mutants had become dominant in plasma virus, and were found at higher frequency in LNs as compared to RBs. Of note, at day 14 post infection virus escape mutants in LNs, but not other tissues, were overwhelmingly comprised of mutations at the first amino acid of the epitope. These mutants became dominant in all tissues by day 28 post infection. Taken together, these results suggest that, during acute SIVmac239 infection, lymph nodes are the main sites of virus replication and/or cellular immune pressure.
The large amount of virological and immunological data collected during the original vaccination and challenge study of these SIV-infected RMs [22] allowed us to more explicitly test hypotheses regarding the immunological processes and anatomic compartments that ultimately produced these viral escape mutants. By correlating immune responses and changes in viral dynamics with the level of escape in the examined tissues, we can determine not only the origin of escape mutants, but the immune responses responsible for their generation. We found that the level of Tat-SL8-specific immune responses in LNs, and to a lesser extent PBMCs, at day 14 post infection correlated strongly with the level of escape in plasma virus at day 28 post infection. On the other hand, SIV-specific immune responses in RBs at day 14 post infection did not correlate with the level of Tat-SL8 escape at day 28 in plasma virus. Therefore, LNs rather than mucosal tissues appear to be the tissue in which the strongest immune pressure shapes the genetic composition of the currently replicating virus population during the phase of post-peak decline of viremia in SIV-infected RMs.
In addition, we observed that, in controller SIV-infected RMs (but not in non-controller RMs), a large proportion of PBMC-derived proviruses during the chronic phase of infection exhibited the wildtype Tat-SL8 rather than the escaped epitope. This high frequency of wildtype virus in controller RMs may be the result of the early establishment of a wildtype latent viral reservoir, which, due to their low level of replication, appears over-represented as compared with the non-controller animals. Since the actively replicating virus (i.e., in plasma) contains very few, if any, viruses bearing the wildtype Tat-SL8, these wildtype PBMC-associated viruses are likely to be either latent or non-replication competent.
The observation that immune responses in LNs are more important for shaping post acute viral populations is somewhat surprising as the gut-associated lymphoid tissue has been implicated as the major source of virus replication during the acute phase of SIV infection [26], [27], [28], [29], [30]. One possibility is that, during the post-peak decline of viremia, the gut mucosa is a relatively isolated site of SIV replication, with slower migration of virions from the intestinal interstitial fluids into the systemic circulation as compared to LNs. Indeed, there is evidence for compartmentalization of HIV within the gut [31] suggesting that virus does not move freely within this tissue. An alternative possibility is that the level of immunological pressure, rather than the absolute level of virus production, is responsible for the directional appearance of CTL escape mutants from LNs to mucosal tissues. This concept fits with the fact that during acute SIV infection LNs are the site of antigen-specific CD8+ T cell expansion from the naïve pool, with virus-infected cells encountering anti-SIV CD8+ T cells more frequently than in the intestinal lamina propria. These data are also consistent with the possibility that the gut is not the main source of virus replication during acute SIV infection, as recently suggested in a modeling study [32]. However, we could not rule out the possibility that the mutations in Tat simply provide a selective advantage to viruses in lymphoid tissue (or less of a disadvantage relative to viruses replicating in gut mucosal tissues) independently of their effects on immune recognition.
While our study agrees with the numerous published articles that have characterized immune escape in SIVmac239 infected RMs [14], [15], [17], [23], [33], [34], [35], [36], [37], in almost all of these studies, CTL escape mutants were assessed only in plasma virus. Our study is unique in that we examined virus in multiple longitudinal tissue samples (PBMC, LNs, and RBs) in addition to the plasma virus. We are in fact aware of only a single study that examined SIVmac239 escape mutants in multiple tissues [38], but this study focused only on Gag-CM9 escape measured at a single time point during the chronic infection. It will be important to determine whether the pattern of escape seen in Tat-SL8 during the acute infection is a general phenomenon for CTL epitopes that escape during the acute phase of infection with little or no measurable effects on viral fitness. In addition it will be informative to characterize the anatomic distribution of CTL epitopes that escape during the chronic phase of SIV infection (e.g. Gag-CM9) by designing new studies in which tissue samples are collected accordingly. Furthermore, future studies should address whether the cell-associated viral sequences sampled from relatively small biopsies of the rectal mucosa are representative of the gut as a whole. While this caveat applies to most studies of gut-associated lymphoid tissue in the setting of SIV infection, it is possible that our current results are biased by the limited number of samples available for virus sequencing.
It is unclear whether and to what extent the particular pattern of CTL escape observed in our cohort of SIV-infected RMs represents a generalized phenomenon or, alternatively, is the result of this specific experimental design (i.e., intravenous high titer infection), which does not mimic the typical route and circumstances surrounding HIV infection of humans. Furthermore, the route of infection may influence the anatomic pattern of emergence of CTL escape mutants. For example, a rectal challenge model might result in the early seeding of local draining lymph nodes with virus whereas an intra-venous challenge could possibly disseminate virus more efficiently to systemic lymphoid tissues throughout the body. Future studies in which RMs are infected with low dose intra-rectal or intra-vaginal challenge will help determine how the route and dose of infection influences the early kinetics of CTL escape. Additionally, while the MVA vaccination regime in this study did not protect from SIV disease progression, it would be interesting to determine whether more successful vaccines (i.e., rhesus CMV [39] and AdHu26 [40]) alters the dissemination of viral escape variants and/or the pattern of immune responses during the acute phase of SIV infections.
These studies were carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health, and were approved by the Emory University (AWA# A3180-01) and University of Pennsylvania (AWA# A3079-01) Institutional Animal Care and Use Committees. All animals were anesthetized prior to the performance of any procedure, and proper steps were taken to ensure the welfare and to minimize the suffering of all animals in these studies.
The design of the SIV immunization and challenge study in RMs that has been used as a source of samples for the current manuscript are published elsewhere [22]. Briefly, ten RMs were divided in two groups of five animal and immunized three times with either one of two MVA vaccine vectors (i.e., wildtype and a genetically modified MVA in which the udg gene was deleted) expressing the SIVmac239 gag and tat genes. One year after the initial immunization, all ten animals as well as five unvaccinated control RMs were inoculated i.v. with 10,000 TCID50 of SIVmac239. Throughout the immunization and challenge phase plasma, peripheral blood mononuclear cells (PBMC), lymph node biopsies (LN) and biopsies of the rectal mucosa (RB) were collected for virological and immunological analysis (see Table S1). The MHC-class I genotypes of all fifteen RMs were determined courtesy of Dr. David Watkins. Of note, exclusion of the only MamuB*08+ and MamuB*17+ animal did not change the result of our analyses.
PBMC were collected by gradient centrifugation. LN and RB were processed fresh as previously described. The levels of Tat- and Gag- specific CD8 T cells were assessed via staining with Streptavidin-APC conjugated class I MHC tetramers folded with the Tat28-35-SL8 (STPESANL) or Gag181-189-CM9 (CTPYDINQM) peptide epitopes according to standard procedures.
Genomic DNA was purified from thawed PBMC samples using the QIAamp DNA blood mini kit (Qiagen; Valencia, CA), and from whole Streck-fixed, paraffin embedded RBs or from 50 µM slices of formalin-fixed, paraffin embedded LNs using the QIAamp DNA FFPE tissue kit (Qiagen; Valencia, CA). Viral RNA was extracted from plasma samples using the Qiagen's Viral RNA mini kit (Valencia, CA). Viral cDNA was reverse transcribed using Invitrogen's SuperScript III and primers specific for sequences upstream of the tat (Tat-RT3: 5′-TGGGGATAATTTTACACAAGGC-3′) or gag (Gag-RT2: 5′-AGCTTGCAATCTGGGTTAGC-3′) amplicons.
Viral sequences were amplified from purified genomic DNA or viral cDNA in a nested two step PCR. The thermal cycler program for the first round was: 94°C for 2:00; cycle 94°C for 0:30, 55°C for 0:30, and 68°C for 1:00, 35 times; and then a final 68°C for 7:00 before cooling to 4°C. After the tenth cycle, the extension step (68°C) is extended by 5 seconds every cycle to account for the degrading polymerase. The thermal cycler program for the second round was: 94°C for 2:00; cycle 94°C for 0:30, 53°C for 0:30, and 68°C for 1:00, 35 times; and then a final 68°C for 7:00 before cooling to 4°C. The first round primers for tat were Tat-F1 (5′-GATGAATGGGTAGTGGAGGTTCTGG-3′) and Tat-R2 (5′-CCCAAGTATCCCTATTCTTGGTTGCAC-3′), and the first round primers for gag were Gag-F1 (5′-GAGACACCTAGTGGTGGAAACAGG-3′) and Gag-R2 (5′-GCTCTGAAATGGCTCTTTTGGCCC-3′). The design of the second round primers involved the incorporation of Roche's 454 Adaptor sequences and an animal-specific, 4 nucleotide barcode that allows identification of individual animals in the pooled sequencing runs. The barcode key can be found in Table S1, and their position is indicated by a lowercase ‘b’ in the following primer sequences. Second round primers for tat had sequences similar to those previously published [15] and were Tat-F3 (5′-GCCTTGCCAGCCCGCTCAGbbbbTGATCCTCGCTTGCTAACTG-3′) and Tat-R3 (5′-GCCTCCCTCGCGCCATCAGAGCAAGATGGCGATAAGCAG-3′), and the second round primers for gag were Gag-F3 (5′-GCCTTGCCAGCCCGCTCAGbbbbCACCATCTAGCGGCAGAGGAGG-3′) and Gag-R3 (5′-GCCTCCCTCGCGCCATCAGACCCCAGTTGAATCCATCTCCTG-3′). After amplification, each amplicon was gel purified using the QIAquick gel extraction kit (Qiagen, Valencia, CA). The amplicons were then quantified on a NanoDrop (Company), and mixed at equimolar concentrations within tissues and time points. Massively parallel pyrosequencing was performed by the University of Pennsylvania, Department of Genetics Sequencing Facility on a Roche 454 Genome Sequencer FLX (Branford, CT).
Raw 454 sequencing reads will be available in the NCBI Sequence Read Archive under accession number SRA027346.1. Sequences were subjected to several steps of quality control [41], [42]. First, sequences that contained ambiguous nucleotides and those that did not meet the minimum length requirements (tat: 220 nucleotides; gag: 243 nucleotides) were excluded. Next, each read was aligned individually to the SIVmac239 wildtype sequence using ClustalW and barcodes were identified. Reads where the barcode could not be identified were excluded from the analysis. Several common insertions and deletions (indels) that resembled common 454 sequencing artifacts were then identified and repaired using the wildtype SIVmac239 sequence as a template. Sequences containing multiple indels and low frequency sequencing artifacts were excluded. All sequence manipulations and analyses were performed and implemented in a suite of scripts written in Python. Statistical analyses were all performed in GraphPad Prism.
|
10.1371/journal.ppat.1004165 | Growth Factor and Th2 Cytokine Signaling Pathways Converge at STAT6 to Promote Arginase Expression in Progressive Experimental Visceral Leishmaniasis | Host arginase 1 (arg1) expression is a significant contributor to the pathogenesis of progressive visceral leishmaniasis (VL), a neglected tropical disease caused by the intracellular protozoan Leishmania donovani. Previously we found that parasite-induced arg1 expression in macrophages was dependent on STAT6 activation. Arg1 expression was amplified by, but did not require, IL-4, and required de novo synthesis of unknown protein(s). To further explore the mechanisms involved in arg1 regulation in VL, we screened a panel of kinase inhibitors and found that inhibitors of growth factor signaling reduced arg1 expression in splenic macrophages from hamsters with VL. Analysis of growth factors and their signaling pathways revealed that the Fibroblast Growth Factor Receptor 1 (FGFR-1) and Insulin-like Growth Factor 1 Receptor (IGF-1R) and a number of downstream signaling proteins were activated in splenic macrophages isolated from hamsters infected with L. donovani. Recombinant FGF-2 and IGF-1 increased the expression of arg1 in L. donovani infected hamster macrophages, and this induction was augmented by IL-4. Inhibition of FGFR-1 and IGF-1R decreased arg1 expression and restricted L. donovani replication in both in vitro and ex vivo models of infection. Inhibition of the downstream signaling molecules JAK and AKT also reduced the expression of arg1 in infected macrophages. STAT6 was activated in infected macrophages exposed to either FGF-2 or IGF-1, and STAT6 was critical to the FGFR-1- and IGF-1R-mediated expression of arg1. The converse was also true as inhibition of FGFR-1 and IGF-1R reduced the activation of STAT6 in infected macrophages. Collectively, these data indicate that the FGFR/IGF-1R and IL-4 signaling pathways converge at STAT6 to promote pathologic arg1 expression and intracellular parasite survival in VL. Targeted interruption of these pathological processes offers an approach to restrain this relentlessly progressive disease.
| Visceral leishmaniasis (VL), caused by the intracellular protozoan Leishmania donovani, is a progressive infection that is particularly common in impoverished populations of the world. People die from this disease unless it is treated. We used an experimental infection model that mimics the clinical and pathological features of human VL to study how the parasite causes this severe disease. We found that host macrophages infected with Leishmania donovani are activated in a way that leads to the expression of arginase, an enzyme that counteracts the cell's mechanisms that control the infection. This disease-promoting activation pathway was driven by the convergence of growth factor and cytokine signaling pathways and activation of the transcription factor STAT6. Chemical inhibition of signaling through the fibroblast growth factor receptor-1 (FGFR-1) or insulin-like growth factor-1 receptor (IGF-IR), or genetic knockdown of STAT6 led to reduced expression of arginase and enhanced control of the infection by macrophages. This indicates that the growth factor signaling pathways together with the cytokine pathways promote this disease. Interventions designed to disrupt this signaling could help in the treatment of VL.
| Visceral leishmaniasis (VL), caused by the intracellular protozoan Leishmania donovani or L. infantum, is one of the “Neglected Tropical Diseases” that impacts the poor of the world. Active VL is characterized by a relentlessly progressive infection with cachexia, massive splenomegaly, pancytopenia and ultimately death. VL ranks second to malaria in deaths caused by a protozoal pathogen; mortality is reported in up to 10–20% of patients, even with treatment [1]. The determinants of susceptibility and progressive disease are incompletely defined. However, it is clear that ineffective cellular immune function, dictated by the nature of cytokine response and polarization of macrophages [2], plays a critical role. Macrophages, the primary target of intracellular Leishmania infection, may take on distinct phenotypes in response to parasite signals and inflammatory stimuli within the infected microenvironment. Classically activated (M1) macrophages respond to IFN-γ and microbial products by generating antimicrobial molecules that effectively kill Leishmania and other intracellular pathogens [3], [4]. Central to the killing of intracellular parasites is the production of nitric oxide by the action of inducible nitric oxide synthase 2 (NOS2) on the substrate L-arginine. In contrast, alternatively activated or M2 macrophages, which are typically generated by exposure to type 2 cytokines (IL-4, IL-13), fail to produce antimicrobial effector molecules to kill intracellular pathogens and serve to dampen inflammation and promote wound healing [5], [6].
The activation status of macrophages in human VL has not been directly investigated. However, the progressive nature of the infection in the face of strong expression of IFN-γ [7]–[10], suggests that there is ineffective classical activation. The concomitant production of IL-4/IL-13 and IL-10 [7], [8], [11]–[14], which are known to impair macrophage leishmanicidal activity, may polarize macrophages toward a disease-promoting M2 phenotype. Neutralization of IL-10 in ex vivo splenocyte cultures from patients with VL promoted parasite clearance [15], but the importance of IL-4 and/or IL-13 in the pathogenesis of human VL is not clear. Additionally, Leishmania-driven subterfuge of a number of signaling pathways can render the macrophage less responsive to activating stimuli and more permissive to infection [16].
We have used the hamster model of VL, which closely mimics the clinicopathological features of human VL, to dissect the mechanisms by which L. donovani causes progressive disease. We demonstrated, similar to human VL, that progressive, lethal disease occurred in the face of what would be considered a protective type 1 cytokine response [17], [18]. Despite high expression of IFN-γ, it was ineffective in mediating classical activation of M1 macrophages and control of Leishmania infection. In fact we found that splenic macrophages from hamsters with VL were polarized to a M2-like phenotype with dominant expression of host arginase 1 (arg1) [2]. L. donovani triggered arg1 expression through a STAT6-dependent mechanism, but surprisingly it did not require type 2 cytokines [2]. Arginase contributes to intracellular Leishmania replication by competing with NOS2 for the substrate arginine (thereby reducing NO production), and by driving the generation of polyamines, which promote parasite growth [2], [19], [20]. M2-like macrophages and arginase have also been implicated in the pathogenesis of experimental cutaneous leishmaniasis [19]–[23] and infections with other intracellular pathogens [24]–[27]. Furthermore, there is accumulating evidence that arginase has a role in the pathogenesis of human disease. Although, polarization of isolated human macrophages by exposure to IL-4 in vitro did not lead to upregulation of arginase activity or arg1 expression [28], the presence of M2-like monocytes/macrophages and arginase expression has been found in cancer [29], [30], filariasis [25], tuberculosis [31], [32], and traumatic tissue injury [33]. Elevated arginase activity was also recently reported in the lesions of patients with chronic cutaneous leishmaniasis [34] and arginase expression in peripheral blood leukocytes was found to be a marker of active VL [35].
In this work we have investigated the mechanisms of the pathological upregulation of arg1 in the hamster model of progressive VL. We discovered that the expression of arg1 in L. donovani infected macrophages is driven by activation of fibroblast growth factor receptor (FGFR) and insulin-like growth factor-1 receptor (IGF-IR). Inhibition of these growth factor signaling pathways led to reduced arg1 expression and enhanced control of parasite replication. Furthermore, signaling molecules downstream of the growth factor receptors converged with IL-4 signaling to promote STAT6 activation and arg1 expression in VL. The intersection of these pathways leads to subversion of macrophage effector function and impaired host defense against VL.
We previously determined that L. donovani induced STAT6-dependent, host arg1 expression. Host arginase expression promoted parasite replication, so we sought to understand the mechanisms by which it was expressed in VL. Arg1 transcription required the de novo synthesis of protein [2] suggesting that transcription of arg1 involved signaling pathway(s) other than just direct phosphorylation of STAT6. We postulated that the newly synthesized protein could mediate its effect through RTK signaling pathways, which regulate inflammation and wound repair [36], [37]. Both of these processes are important functions of M2 macrophages. Therefore, we screened a library of 80 RTK inhibitors for inhibition of L. donovani-induced arginase transcription in an ex vivo model of infected splenocytes isolated from hamsters with VL [38]. Inhibitors of the Epidermal Growth Factor Receptor and Platelet-derived Growth Factor Receptor signaling pathways reduced arg1 transcription by >50% (Table 1). Because the RTK signaling pathways are overlapping and broad, and inhibitors of some growth factor receptors were not included in the inhibitor library, we used a RTK antibody array to further define the participation of specific RTKs in VL. We found that Fibroblast Growth Factor Receptor (FGFR) 1 and 2 and other molecules known to participate in growth factor signaling (Insulin receptor substrate 1 (IRS-1), v-akt murine thymoma viral oncogene homolog 1 and 2 (AKT 1/2), Mitogen-activated protein kinase (MAPK)-3, and Signal transducer and activator of transcription (STAT)-1, and STAT-3 were activated in splenic macrophages from hamsters infected with L. donovani (Table 2). Collectively, these data indicated that signaling through growth factor receptor pathways could contribute to the parasite-induced expression of host arg1.
A significant increase in arg1 mRNA expression was observed in L. donovani infected hamster bone marrow-derived macrophages (BMDM) exposed to the recombinant growth factors FGF-2, IGF-1, and PDGF (Fig. 1A). Growth factor-induced arg1 was particularly evident in infected compared to uninfected macrophages, and it was equivalent to, or greater than, IL-4-induced arg1. Arginase protein activity was also significantly increased in L. donovani infected BMDM exposed to FGF-2, IGF-1, and PDGF (Fig. 1B). EGF did not consistently induce a significant increase arg1 mRNA or protein. Together, these data suggested that L. donovani infection of macrophages led to enhanced arg1 transcriptional responsiveness to multiple growth factors.
Analysis of the FGF and IGF-1 signaling pathways in splenic macrophages from hamsters with VL by immunoblotting confirmed the finding of the antibody screening array (Figs. 2 and 3). There was no evidence for activation of other growth factor signaling pathways in VL (see Fig. S1 and S2). Our finding that inhibition of EGFR reduced arg1 mRNA expression (Table 1), when neither increased ligand expression nor receptor activation could be demonstrated, suggested that basal activity of EGF/EGFR modulated arg1 expression through an effect on downstream signaling. As we demonstrated previously [2], arg1 protein expression was increased in macrophages isolated from the spleens of hamsters with VL starting at 14 days post-infection (Fig. 2A). Of the growth factor receptor ligands, only FGF-2 expression was increased in splenic macrophages (Fig. 2B, Fig. S1, and Fig. S2) and it was accompanied by increased phosphorylation of Tyr653/654 of the FGFR-1 (Fig. 2C) relative to overall receptor protein expression (Fig. 2D). The increase in both FGF-2 and its phosphorylated receptor paralleled the expression of arg1 in the splenic macrophages. Multiple molecules involved in the signaling cascade downstream of FGFR (shown in the diagram in Fig. 2M) were activated, including members of the PI3K/AKT pathway [GAB (Fig. 2E), PI3K (Fig. 2F)] and the MAPK/ERK pathway [c-RAF (Fig. 2G), ERK1/2 (Fig. 2H)]. Activation of p38 MAPK (Fig. 2I), that leads to activation of the transcription factor ATF-2 (Fig. 2J) and the cyclic AMP response element-binding protein (CREB) (Fig. 2K) was observed at 14 days post-infection but was then down-modulated at 28 days post-infection. This suggested that sustained activation of these signaling molecules was not required for the expression of arg1 throughout the course of VL (Fig. 2A). The mechanism(s) through which these molecules are down regulated is unknown. Activation of STAT3, which was evident throughout the course of VL (Fig. 2L), may be a consequence of increased IL-10 production (Fig. S4A and reference [2]) or growth factor signaling (Fig. S4D) [39].
We were unable to detect increased expression of IGF-1 or IGF-2 in the spleen or plasma of hamsters with VL (Fig. S1; data not shown). However, by immunoblot we found increased expression of the IGF-1R after 14 days post-infection (Fig. 3A), and somewhat unexpectedly the beta (cytoplasmic) domain of the IGF-1 receptor, which mediates intracellular signaling, was phosphorylated at these time points (Fig. 3B). We confirmed these findings in BMDM exposed in vitro to L. donovani where parasite-induced IGF-1R phosphorylation was evident between 20 minutes and 24 hrs of exposure, and enhanced expression of IGF-1R protein was present at 24 hrs after infection (Fig. 3C). A number of the activated signaling molecules downstream of FGFR overlap with the canonical IGF-1R signaling pathway (compare data in Fig. 2 with schematic in Fig. 3L). Additionally, other pathway members, including IRS-1 (Fig. 3D), SHC (Fig. 3E), AKT (Fig. 3F), p70S6K (Fig. 3G), and GSK3β (Fig. 3H) were activated, as were the downstream transcription factors c-FOS (Fig. 3I) and c-Jun (Fig. 3J). When all of the activated signaling molecules were subjected to network analysis (Ingenuity Pathway Analysis) both the FGFR and IGF-1R pathways were found to be significantly upregulated in splenic macrophages during the course of VL (p<10−7; Fig. 3K).
Treatment of L. donovani-infected hamster BMDMs over 24 hrs of infection with an inhibitor of FGFR-1 resulted in a significant dose-dependent reduction of arg1 mRNA expression (Fig. 4A) and parasite burden (Fig. 4B) without affecting cell viability (Fig. 4C). Notably the concentration of FGFR inhibitor required to inhibit parasite replication was higher than the concentration that reduced arg1 expression. This suggests that growth factor signaling supported parasite growth/survival through additional arg1-independent mechanisms, or that residual arginase activity at the lower inhibitor concentration is enough to support parasite growth. The latter possibility is consistent with our previous finding that >90% arg1 knockdown led to approximately 50% reduction of parasite load [2]. The FGFR inhibitor also blocked the expression of arg1 mRNA (Fig. 4D) and protein (Fig. 4E), and reduced the parasite burden (Fig. 4F) without affecting cell viability (Fig. 4G) in ex vivo cultured spleen cells from infected hamsters. Similar effects were found by inhibition of IGF-1R. In the in vitro infection model, IGF-1R inhibition reduced parasite-induced expression of host arg1 mRNA (Fig. 5A) and the intracellular parasite load (Fig. 5B), without decreasing cell viability (Fig. 5C). Similarly, the inhibitor reduced arg1 mRNA (Fig. 5D) and protein (Fig. 5E), and reduced the parasite burden (Fig. 5F) without affecting cell viability (Fig. 5G) in ex vivo cultured spleen cells from infected hamsters. The FGFR and IGF-1R inhibitors did not have a direct effect on the viability of L. donovani cultured promastigotes (Fig. S3), suggesting that the effect of receptor inhibition was through modulation of the host cell. Inhibition of JAK, which plays a key role in the phosphorylation of STAT proteins following cytokine and growth factor signaling, dramatically reduced arg1 transcription in ex vivo cultured splenocytes from infected hamsters (Fig. 5H). To a lesser degree, inhibition of the protein AKT, which is involved in signal transduction downstream of the IGF-1 and FGF receptors, also decreased Arg-1 expression (Fig. 5H). Both the AKT and JAK inhibitors significantly reduced parasite load (Figs. 5I and 5J).
Since cytokines (IL-4 and IL-10) are known to stimulate the expression of arginase [5], [6], and we demonstrated that growth factors also induced arginase (Fig. 1), we investigated the potential for amplification of arg1 expression in macrophages by simultaneous exposure to these stimuli (all of which are expressed in the spleen during VL (reference [2] and Figs. 2 and 3). The L. donovani-induced expression of arg1 in BMDM was modestly amplified by IL-4 but not IL-10 at the mRNA level (Fig. 6A), but neither significantly amplified the arg1 protein (Fig. 6B). However, IL-4 and IL-10 dramatically enhanced the FGF-2-induced arg1 mRNA (Fig. 6C), and IL-4 (but not IL-10) enhanced FGF-2-induced arg1 protein (Fig. 6D) expression in infected macrophages. IL-4 did not amplify IGF-1-induced arg1 mRNA expression in infected BMDMs (Fig. 6E) but augmented arg1 protein expression (Fig. 6F). Similar to IL-10 and FGF-2, IL-10 enhanced IGF-1-induced arg1 mRNA but not protein expression. A trend of an additive effect of IL-4 and the growth factors was also found in splenic macrophages from infected animals exposed to the cytokine and growth factors ex vivo (Fig. 6G). The additive effect of IL-4 and growth factors in the induction of arg1 expression prompted us to consider that there may be cross-regulation of receptor expression. We found that the expression of IL-13Rα1, but not IL-4Rα, was upregulated in splenic macrophage from hamsters with VL (Fig. 6H) and in BMDMs infected with L. donovani (Fig. 6I). Addition of FGF-2 or IGF-1 to infected macrophages did not further increase the expression of either of these receptor components (data not shown and Fig. 6I). IL-10Rα expression (along with IL-10) was also increased in splenic macrophages from infected hamsters (Fig. S4A) and in in vitro infected BMDMs (Fig. S4B), but FGF-2 or IGF-1 did not augment IL-10 or IL-10Rα expression (Fig. S4B). These data, coupled with the data shown in Figs. 2 and 3, suggest that the cytokine-mediated amplification of growth factor driven arg1 could occur by either increased IL-4-mediated signaling through upregulated type II receptor (IL-13Rα1) expression [40] or through activation of signaling proteins (e.g. Jak-1, STAT6, IRS-1, PI3K, AKT) common to the two pathways.
We previously demonstrated that STAT6 was required for L. donovani-induced arg1 expression in fibroblasts [2]. Here we confirmed that siRNA-mediated knockdown of STAT6 mRNA (Fig. 7A) and protein (see Fig. 8F and 8I) in in vitro infected macrophages led to reduced arg1 mRNA (Fig. 7B) expression, and improved control of parasite replication (Fig. 7C). Similarly, knockdown of STAT6 (75% reduction) in ex vivo cultured splenic macrophages from infected hamsters led to significantly reduced arg1 mRNA expression (Fig. 7D). These data confirm the critical importance of STAT6 in the parasite-driven expression of arg1 in macrophages in VL.
Since STAT6 had a critical role in parasite-induced arg1 transcription, activation of growth factor signaling was evident in L. donovani infection, and there was an additive effect of IL-4 and growth factors in the induction of arg1 expression, we wanted to know if the FGF-2- and/or IGF-1-induced arg1 expression was dependent on the activation of STAT6. In a STAT6 reporter assay (hamster fibroblast cell line; reference [2]), we found that recombinant FGF-2 and IGF-1 induced STAT6 activation, which was blocked when cells were pre-treated with an inhibitor of the corresponding growth factor receptor (Figure 8A). In the fibroblast cell line, exposure to parasites had a relatively weak effect on STAT6 activation, probably because at this parasite dose the cells are infected at a very low level. The growth factor-induced activation of STAT6 in macrophages was confirmed by detection of phosphorylated STAT6 in immunoprecipitated lysates of splenic macrophages from L. donovani infected hamsters exposed ex vivo to recombinant FGF-2 or IGF-1 (Fig. 8B). Parasite-induced STAT6 activation was abrogated completely by an IGF-1R inhibitor and partially by an FGFR inhibitor (Fig. 8C). Conversely, siRNA-mediated knockdown of STAT6 mRNA in infected, FGF-2-treated BMDM (Fig. 8D) identified the requirement for STAT6 in the FGF-2-induced expression of arg1 mRNA (Fig. 8E) and protein (Fig. 8F). Similarly, siRNA-mediated knockdown of STAT6 in infected IGF-1-treated BMDM (Fig. 8G) identified the contribution of, but not absolute requirement for, STAT6 in the IGF-1-induced expression of arg1 mRNA (Fig. 8H) and protein (Fig. 8I). Collectively these data identify the critical importance of growth factor signaling in the parasite-induced activation of STAT6, and of STAT6 in the IGF-1 and FGF-2 driven expression of arg1 in L. donovani infected macrophages.
Since simultaneous exposure of infected macrophages to IL-4 and FGF-2 or IGF-1 led to enhanced arginase expression, and the growth factor- and cytokine-induced expression of arg1 was dependent on STAT6, we reasoned that there might be enhanced activation of STAT6 in cells exposed to both IL-4 and growth factors. Stimulation of the reporter cells with either growth factors (see also Fig. 8A) or IL-4 activated STAT6. There was evidence of an additive effect when the growth factor and cytokine were combined (Figs. 9A and 9B). By immunoblotting, STAT6 phosphorylation was amplified when IL-4 was combined with the growth factors (Figs. 9C and 9D). Inhibition of FGFR and IGF-1R activation led to decreased IL-4-induced STAT6 activation (Fig. 9E). Taken together, these data indicate that bi-directional crosstalk between the growth factor and IL-4 signaling pathways converges at STAT6 to drive arg1 expression in VL.
In an experimental model of progressive VL, we demonstrated previously that parasitized macrophages were polarized to an M2-like phenotype [2], characteristic of macrophages at a site of chronic injury and wound healing [5], [6], and were massively expanded in the spleen [2], [38]. These macrophages had dominant expression of arg1, which promoted parasite growth. The L. donovani-induced macrophage arg1 expression did not require, but was amplified by, type 2 cytokines [2]. In this work we focused our attention on the mechanisms through which pathological arg1 expression occurs in VL. We discovered that FGF-2 and IGF-1 signaling pathways were activated in splenic macrophages from animals with progressive VL. These growth factors, which may be produced by macrophages, fibroblasts, or endothelial cells [41]–[43], induced macrophage arg1 expression. Inhibition of FGFR1 and IGF-1R signaling led to both reduced arg1 expression and improved control of intracellular L. donovani infection. Parasite-induced FGFR and IGF-1R signaling converged with the canonical type 2 cytokine signaling pathway through STAT6 activation to induce arg1 expression. Simultaneous exposure of macrophages to growth factors and IL-4, as would occur in the spleen during VL, enhanced the activation of STAT6 and expression of arg1. The interplay of STAT6 and growth factor signaling was confirmed by demonstrating that FGF-2- and IGF-1-induced arg1 expression was abrogated by knockdown of STAT6, and conversely, that inhibition of growth factor signaling reduced parasite- and IL-4-mediated STAT6 activation and arg1 expression.
Arginase expression contributes to the pathogenesis of cutaneous L. major infection in mice [19]–[23] and progressive experimental VL caused by L. donovani [2]. Its expression in blood leukocytes was also found to be a marker of active VL in patients from Ethiopia [35]. In that study the blood leukocytes that produced arginase were found in the mononuclear cell fraction but expressed CD15 so were identified as low-density granulocytes. Those cells were not further characterized, and we have not evaluated expression of arg1 in granulocytes in our model of experimental VL. Therefore, it remains to be determined if there is a fundamental difference in the source of arg1 in experimental and human VL, or if further characterization of the cell populations will resolve the apparent difference. The disease-promoting effect of arg1 may be mediated through several mechanisms. First, arg1 metabolizes arginine such that this substrate is not available for the generation of the antimicrobial effector molecule, nitric oxide, by the action of inducible nitric oxide synthase. Second, arg1 expression leads to the production of polyamines, which promote intracellular Leishmania growth [2], [19], [20]. Lastly, local depletion of arginine leads to impaired anti-leishmanial T cell responses [44]. The relative contributions of each of these effects on the pathogenesis of VL remain to be determined.
The role of growth factors in modulation of arg1 expression and macrophage function in response to Leishmania or other pathogens has received little attention. The induction of arginase expression is classically a type 2 cytokine (IL-4/IL-13)- and STAT6-driven process [5], although some parasites or parasite products have been shown to directly induce an M2-like macrophage phenotype [2], [20], [45]. Since growth factors modulate inflammation and tissue repair [46]–[50], processes in which M2 macrophages have an integral part, it is not surprising that there would be interconnections between growth factors, type 2 cytokines, and M2 polarization. The tissue remodeling [51], [52], accumulation of macrophages [38], [52]–[55] and collagen deposition/fibrosis [38], [54] observed in the spleens in experimental and human VL are processes that suggest growth factors may contribute to VL pathology. Cytosolic IGF-1 was found increased in L. major infected murine macrophages [56], and IGF-1 induced parasite arginase in L. amazonensis infected macrophages [57]. Although we cannot exclude the potential contribution of parasite arginase in the IGF-1 and FGF-2-mediated effects on macrophages, we found previously that L. donovani arginase contributed little to the overall arginase expression at the site of infection in this model of progressive VL [2]. The increased expression of FGF-2 and evidence of signaling through the IGF-1 and FGF receptors to our knowledge had not been described previously in VL. Surprisingly, robust IGF-1R phosphorylation was evident in the infected spleen in the absence of increased IGF-1, suggesting cross-activation by FGF-2 [58] or by an unknown host or parasite-derived factor. We think cross-activation by FGF-2 is unlikely in the case of VL since we did not find IGF-1R phosphorylation in BMDMs infected with L. donovani and treated with FGF-2 for 20 minutes to 48 hours post-infection (data not shown). Of note, it was reported previously that Leishmania expressed an ortholog of FGF-2 [59] so conceivably other parasite-produced growth factor orthologs could be driving the activation of IGF-1R in the absence of host IGF-1. Insulin-like growth factor binding proteins (IGFBPs) or IGFBP proteases [60] could also be modulating the local availability and activity of IGF-1 during the infection.
From this work we have begun to understand the mechanistic basis for the interplay of L. donovani, IL-4 and growth factors in the induction of arg1. IL-4, which is increased in the spleen during VL in humans and hamsters [2], [12], amplifies the parasite- and growth factor-induced expression of arg1. IL-10 appears to have a more limited role in that it upregulates arg1 mRNA, but not protein expression, in infected macrophages, and does not amplify the growth factor effect. Similarly, in L. donovani-infected mice, IL-10 does not directly induce macrophage arginase, but contributes indirectly to its expression by upregulating the type I IL-4 receptor [61]. FGF-2- and IGF-1 enhance expression of arg1 in L. donovani infected macrophages, but have a more modest effect on uninfected macrophages. Thus, the concomitant expression of IL-4 and growth factors in the infected spleen provide an environment highly suited for arg1 expression. IL-4 and IL-13 were shown previously to induce the expression of macrophage IGF-1 [62] and coincident expression of type 2 cytokines and IGF-1 was demonstrated in experimental helminth infection [63]. The amplification of growth factor-induced arg1 by IL-4 in experimental VL is not associated with growth factor-mediated upregulation of the type 1 IL-4 receptor (IL-4rα). Although we found that L. donovani infection increased expression of IL-13Rα1, which partners with IL-4rα to transduce a signal via IL-4 or IL-13 [40], this receptor is thought to be a less-potent driver of M2 macrophage activation than is IL-4 signaling through the type I receptor. Furthermore, IRS-1/2, which we found strongly activated in VL, is activated primarily via IL-4 signaling through the type I rather than the type II receptor [64], [65]. Collectively, these data suggest that the growth factor/IL-4-mediated amplification of arg1 expression results from an effect downstream of the IL-4 receptor. Since IL-13 [2] and its receptor are also increased in the spleen during VL, they may also contribute to the induction of arg1.
The body of work presented here supports the conclusion that the signaling pathways downstream of the growth factor and IL-4 receptors converge at STAT6 to drive pathological arg1 expression. Figure 10 illustrates our current working model for the expression of arg1 in VL. IGF-1 is known to activate STAT6 through an IRS-1/2-dependent pathway. IL-4, which is also an activator of IRS-1/2 [64], can amplify this effect [66], [67]. To our knowledge FGF-2 had not been shown previously to activate STAT6. The pathway through which the growth factors activate STAT6-dependent arg1 transcription in VL remains to be fully elucidated, but for the reasons noted above, IRS-1/2 and JAKs, which are activated in splenic macrophages during VL, are likely key intermediates (see Fig. 10). The downstream activation of other transcription factors (CREB, STAT-3) and signaling molecules, including PI3K/AKT, ERK, p38 MAPK, and GSK3β, are also likely to directly or indirectly contribute to the growth factor induced macrophage polarization and arg1 expression. Notably, p38 MAPK and downstream transcription factors (CREB and ATF-2) are only transiently upregulated so do not account for the sustained increase in arg1 expression over time (reference [2] and this work). Down-modulation of the p38 pathway, however, may contribute to the survival and local expansion of splenic arginase-expressing macrophages [68]. Leishmania infection of macrophages was shown previously to activate the PI3K/AKT pathway, which is a critical regulator of the IL-10 and IL-12 response [69]–[71]. L. donovani-induced production of IL-10 by macrophages involved activation of the PI3K/AKT pathway and downstream phosphorylation-mediated inactivation of GSK-3beta and phosphorylation of CREB [72]. Our data suggest that parasite-induced arg1 is driven at least in part through activation of the same pathway that mediates production of IL-10 by macrophages. However, arg1 and IL-10 expression appear to result from parallel rather than interdependent processes because IL-10 was not a strong inducer of arg1 and did not amplify growth factor-induced arg1 as did IL-4. Also the inhibition of the AKT pathway, which drives IL-10 production, had a less dramatic effect in the down regulation of arg1 than the inhibition of JAKs with the consequent block in STAT activation. Taken together, these data indicate that the expression of arg1 downstream of the growth factor/PI3K/AKT pathway, which is enhanced by IL-4/STAT6 signaling, is an additional mechanism of parasite-mediated subversion of macrophage effector function. Further work is needed to definitively determine the role of IL-10 and STAT3 in this process.
The pathological signaling through the IGF-1R and FGFR that leads to arginase expression in progressive VL is a potential target for adjunctive chemotherapy. Therapies targeting these pathways have recently emerged for a number of proliferative diseases, in particular hematopoietic malignancies and solid tumors (reviewed in [73], [74]). Our ex vivo data suggest that inhibition of FGFR or IGF-1R signaling could have therapeutic potential. Furthermore, it was previously demonstrated in a murine model of L. donovani infection that in vivo administration of a receptor tyrosine kinase inhibitor, when combined with conventional anti-leishmanial chemotherapy had a therapeutic effect [75]. Future pre-clinical studies of FGFR and IGF-1R inhibitors, alone or in combination with current anti-leishmanial therapies, are warranted.
In summary, we determined that the convergence of FGFR/IGFR and IL-4 signaling pathways is responsible for the expression of arg1 in disease-promoting macrophages during chronic progressive VL. FGF-2 and Th2 cytokines [2] are produced in the spleen and lead to activation of the FGFR and STAT6 in infected splenic macrophages. Although the infection does not appear to increase IGF-1 production, the IGF-1R is activated on splenic macrophages through a yet to be identified host or parasite factor. Activation of the FGF and IGF-1 receptors leads to phosphorylation of downstream signaling molecules such as IRS1/2, PI3K, and AKT, which lead to expression of IL-10 [72] and converge with downstream components of the IL-4R pathway to drive arg1 expression. Activation of these pathways, along with the parallel effects of IL-10 in subverting macrophage function [16], [76], [77], plays an important role in the pathogenesis of VL. Targeted interruption of these pathological processes offers an approach to restrain this relentlessly progressive disease.
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 protocol was approved by the Institutional Animal Care and Use Committee of the University of Texas Medical Branch, Galveston, Texas (protocol number 1101004).
6–8 week old Syrian golden hamsters (Mesocricetus auratus) were obtained from Harlan Laboratories.
L. donovani (MHOM/SD/001S-2D) promastigotes were cultured as described previously [78]. Hamsters were infected by intracardial injection of 106 peanut agglutinin purified metacyclic promastigotes [78]. For in vitro infections, stationary phase promastigotes were washed with PBS and used immediately to infect hamster BMDMs. Cells were infected at a promastigote to macrophage ratio of 2∶1 and cultured thereafter in complete medium (CM) composed of DMEM supplemented with 1 mM sodium pyruvate (Gibco), 1× MEM amino acids solution (Sigma), 10 mM HEPES buffer (Cellgro), and 100 IU/mL penicillin/100 mg/mL streptomycin solution (Cellgro), which was supplemented with 2% heat inactivated fetal bovine serum (HIFBS). When infecting BMDMs at this ratio all parasites were internalized so that no extracellular parasites could be observed by light microscopy at 24 hrs post-infection.
Bone marrow cells were flushed from normal hamster femurs and adjusted to 8×106/mL in RPMI with 10% HIFBS, 50 µM β-mercaptoethanol (Sigma), and supplemented with 20 ng/mL recombinant human macrophage-colony stimulating factor (M-CSF) (R&D Systems). After 3 days of culture the medium was changed and at 6–7 days of culture the cell monolayer (>95% macrophages as determined by microscopy) was washed 3 times with PBS and detached with Trypsin/EDTA (Gibco) and cell scraping. The cells were starved of M-CSF or serum in CM with 2% HIFBS overnight before the assays.
Arg1 expression and arginase enzymatic activity in BMDM was determined at 24 hrs or 48 hrs by real-time RT-PCR or by production of urea, respectively, as described previously [2]. For Western blot a goat anti-hamster arg1 polyclonal antibody was used [2]. The antibody used for detection of hamster arg1 did not react with L. donovani parasite lysates. The cells were left unstimulated or exposed to recombinant human Epidermal Growth Factor (EGF), mouse Insulin-like Growth factor-1 (IGF-1), human Platelet-Derived Growth Factor (PDGF) (Cell Signaling), human Fibroblast Growth Factor basic (heparin stabilized) (Sigma), 0.3–2.5% recombinant hamster IL-4 conditioned medium (equivalent to 3–25 IU/mL determined by STAT6 reporter bioassay) [2], or human IL-10 (R&D Systems) and/or infected with L. donovani promastigotes at 1∶2 macrophage∶parasite ratio. The activity of human IL-10 on hamster cells was verified using hamster BMDMs transiently transfected with a STAT3 lentiviral reporter construct (Cignal Lenti-reporter, SA Biosciences).
Real time RT-PCR for arg1 and STAT6 mRNA was performed as described [2].
Spleen cells from 28-day L. donovani infected hamsters were cultured ex vivo as described previously [38] and treated for 24 hrs with each inhibitor from a library of 80 RTK inhibitors (Biomol International, Inc.) at twice the dose reported to cause 50% inhibition. Total RNA was isolated and the level of arg1 transcription determined by real time PCR as described [2].
An RTK antibody array (PathScan Array, Cell Signaling), which contains antibodies against 28 phospho-RTKs and 11 key signaling nodes of the RTK pathways, was used to identify RTKs activated by L. donovani infection. The mean dot-spot chemiluminescent intensity of splenic macrophages (n = 4) from infected hamsters (28 days post-infection) was compared to that of 4 uninfected hamsters by densitometry analysis (GeneTools Analysis Software, Syngene).
BMDMs were seeded in white clear bottom 96-well plates at 20,000 cells per well in CM and pre-treated with Fibroblast Growth Factor Receptor-1 inhibitor (PD166866; CAS 192705-79-6; Calbiochem) or Insulin-like Growth Factor Receptor inhibitor (PPP; CAS 477-47-4; Calbiochem). After 1–2 hrs the medium containing the inhibitor was discarded and the cells infected with L. donovani promastigotes for 20 min. Medium containing fresh inhibitors was then added back to the infected cells and the cells collected at 24 h post-infection for measurement of arg1 expression and parasite burden. Parasite load was determined by measurement of luciferase activity from luciferase-transfected parasites as described previously [38] or by real time RT-PCR using primers and a Taqman probe against the conserved sequence of the 18S gene of Leishmania [79] (forward primer: TTACCACCTTACGTA TCTTTTCTATTCG; reverse primer: AAAACAGAAAACGTGCTGAGG AT; Taqman probe: FAM-CT TTACCGGCCACCCACGGGA-TAMRA). Similar experiments were performed with adherent spleen cells cultured ex vivo from hamsters infected with L. donovani (21 days post-infection) as described [38]. The viability of treated cells was assessed in parallel experiments (20,000 cells/well/100 µL in 96-well white plates) by luminometric measurement of ATP (Cell Titer Glo Assay, Promega).
To confirm the results of the PathScan Array, we immunoprecipitated (IP) selected growth factor receptors from fresh lysates of splenic macrophages isolated from infected hamsters using cross-reacting anti-mouse/human/rat growth factor receptor antibodies (Table S1). Following cell lysis in RIPA buffer supplemented with protease/phosphatase inhibitors (Santa Cruz) the protein concentration of total cell lysates was adjusted to 3 µg/300 µL buffer and the IP procedure was followed according the manufacturer's instructions using protein A/G agarose (Santa Cruz). In brief, pre-cleared samples were incubated with the anti-receptor antibody overnight at 4°C on an Orbital shaker, then 20 µL protein A/G agarose was added to the Ag-antibody complex and incubated for 4 hr at 4°C. The protein A/G/antibody complex was precipitated by centrifugation, washed 3 times with PBS, suspended in 50 µL of 1× LDS running buffer (Invitrogen) and the antibodies released from the agarose beads by heat (100°C, 5 min). After resolving 20 µL of sample by SDS PAGE the separated proteins were transferred to nitrocellulose membranes, blocked with TBS-T 5% milk with 1 mM sodium orthovanadate (Na3VO4) and incubated overnight at 4°C with the anti-phospho RTK in TBS-T with 0.4% BSA or TBS-T 3% milk with 1 mM Na3VO4.
Growth factor receptor ligands were measured in plasma or spleen homogenates from uninfected or infected hamsters. IGF-I and PDGF-β were measured by ELISA using anti-rat/mouse IGF-I and anti-rat/mouse PDGF-β using ELISA kits (R&D Systems). Epidermal Growth Factor, heparin-binding EGF-like growth factor (HB-EGF), Epiregulin and Amphiregulin were measured by immunoprecipitation/western blot using antibodies reactive against the mouse/rat/human proteins (Santa Cruz).
Splenic macrophages isolated by adherence from infected or uninfected hamsters were lysed and suspended in RIPA buffer containing 1× protease/phosphatase inhibitors. Lysates were stored at −80°C and used within 2 months. Ten µg of total protein was suspended in 1× LDS sample buffer and separated by SDS-PAGE in pre-cast gels (NuPage, Bis-Tris 4–12%). The separated proteins were transferred to nitrocellulose membranes using the iBlot system (20 V, 9 min) (Invitrogen). Then membranes were incubated with primary antibody (Table S1) either in TBS-T with 0.4% BSA or TBS-T with 3% milk and 1 mM Na3VO4 followed by the secondary antibody conjugated to HRP. The reaction was detected with enhanced chemiluminescent substrate (West Pico; Thermo Scientific) and captured with a Chemi X T4 camera (G BOX, SynGene) and analyzed with Gene Tools analysis Software (SynGene). The fold change of protein expression was calculated by densitometry analysis of western blot bands of infected samples (at 7, 14 and 28 days post-infection) with reference to uninfected samples.
STAT6 activity was determined in the hamster BHK-21 cell line stably transfected with the luciferase reporter plasmid p(IE-IL4RE)4-LUC as described previously [2]. p-STAT6 was detected by immunoprecipitation of cell lysates (5×106 cells/300 µL RIPA with phosphatase inhibitors) with 1 µg of STAT6 capture antibody (M-20, Santa Cruz Biotechnology) at 4°C. overnight. Protein A/G immunoprecipitated complexes were washed 4 times with PBS, eluted by heat 5 min 100°C in 50 µl of 1× LDS loading buffer and detected by SDS-page using anti-p-STAT6 antibody (# 9361, Cell Signaling, 1∶1000 TBS-T, 0.4% BSA, 4°C, overnight), anti-rabbit HRP conjugate, and West Pico substrate (Thermo Scientific) as above.
Stealth RNAi sequences were designed in silico using the BLOCK-iT RNAi Designer (Life Technologies) and chosen based on the sequences spanning 2 regions that were successfully targeted in knockdown of STAT6 previously [2] as follows: region 1: top, UGGCCACCAUCAGACAAAUACUUCA; bottom, UGAAGUAUUUGUCUGAUGGUGGCCA; region 2: top, CACAGUUCAACAAGGAGAUCCUGUU; bottom, AACAGGAUCUCCUUGUUGAACUGUG (each duplex synthetized and annealed by Life Technologies). Hamster BMDMs were differentiated for 6 days with 20 ng/mL of recombinant human M-CSF (R&D Systems) and plated overnight (250,000 cells per well in 24-well plates and 500 µL CM with 10% HIFBS). For transfection, 25 nM of each stealth duplex (239 and 1451) targeting hamster STAT6 was mixed in a volume of 100 µL Optimem with 0.9 µL of Lipofectamine RNAiMAX (Invitrogen) according the manufacturer's instruction. A non-targeting oligonucleotide (low GC, Invitrogen) was used as a control. Then the culture medium was discarded and 500 µL of Optimem (Invitrogen) with 10% HIFBS without antibiotics was added to the cell monolayer together with 100 µl of the transfection mix to achieve a final concentration of 8.3 nM of siRNAi oligos in 600 µL per well. The next day the transfection medium was changed for fresh Optimem with 10% HIFBS without antibiotics. At 48 hr post-transfection cells were serum starved in CM overnight, and stimulated with either L. donovani promastigotes or growth factors at 72 hr of transfection. Both STAT6 knockdown efficiency and arg1 transcription was measured 24 h later by real time RT-PCR and Western blot.
Comparison between groups was typically performed using ANOVA. A parametric or non-parametric test was selected according the distribution of the raw data, followed by a post-test analysis for multiple groups (e.g. Dunnett's Multiple Comparison Test) as appropriate. Paired t test and Wilcoxon signed rank test were used to identify differences between inhibitors and vehicle controls. All analyses were conducted using GraphPad InStat version 3.00 software for Windows 95 (GraphPad Software, San Diego California USA). P values of <0.05 were considered significant.
|
10.1371/journal.ppat.1002960 | A New Chromosomal Phylogeny Supports the Repeated Origin of Vectorial Capacity in Malaria Mosquitoes of the Anopheles gambiae Complex | Understanding phylogenetic relationships within species complexes of disease vectors is crucial for identifying genomic changes associated with the evolution of epidemiologically important traits. However, the high degree of genetic similarity among sibling species confounds the ability to determine phylogenetic relationships using molecular markers. The goal of this study was to infer the ancestral–descendant relationships among malaria vectors and nonvectors of the Anopheles gambiae species complex by analyzing breakpoints of fixed chromosomal inversions in ingroup and several outgroup species. We identified genes at breakpoints of fixed overlapping chromosomal inversions 2Ro and 2Rp of An. merus using fluorescence in situ hybridization, a whole-genome mate-paired sequencing, and clone sequencing. We also mapped breakpoints of a chromosomal inversion 2La (common to An. merus, An. gambiae, and An. arabiensis) in outgroup species using a bioinformatics approach. We demonstrated that the “standard” 2R+p arrangement and “inverted” 2Ro and 2La arrangements are present in outgroup species Anopheles stephensi, Aedes aegypti, and Culex quinquefasciatus. The data indicate that the ancestral species of the An. gambiae complex had the 2Ro, 2R+p, and 2La chromosomal arrangements. The “inverted” 2Ro arrangement uniquely characterizes a malaria vector An. merus as the basal species in the complex. The rooted chromosomal phylogeny implies that An. merus acquired the 2Rp inversion and that its sister species An. gambiae acquired the 2R+o inversion from the ancestral species. The karyotype of nonvectors An. quadriannulatus A and B was derived from the karyotype of the major malaria vector An. gambiae. We conclude that the ability to effectively transmit human malaria had originated repeatedly in the complex. Our findings also suggest that saltwater tolerance originated first in An. merus and then independently in An. melas. The new chromosomal phylogeny will facilitate identifying the association of evolutionary genomic changes with epidemiologically important phenotypes.
| Malaria causes more than one million deaths every year, mostly among children in Sub-Saharan Africa. Anopheles mosquitoes are exclusive vectors of human malaria. Many malaria vectors belong to species complexes, and members within these complexes can vary significantly in their ecological adaptations and ability to transmit the parasite. To better understand evolution of epidemiologically important traits, we studied relationships among nonvector and vector species of the African Anopheles gambiae complex. We analyzed gene orders at genomic regions where evolutionary breaks of chromosomal inversions occurred in members of the complex and compared them with gene orders in species outside the complex. This approach allowed us to identify ancient and recent gene orders for three chromosomal inversions. Surprisingly, the more ancestral chromosomal arrangements were found in mosquito species that are vectors of human malaria, while the more derived arrangements were found in both nonvectors and vectors. Our finding strongly suggests that the increased ability to transmit human malaria originated repeatedly during the recent evolution of these African mosquitoes. This knowledge can be used to identify specific genetic changes associated with the human blood choice and ecological adaptations.
| Complexes of sibling species are common among arthropod disease vectors [1]–[3]. Members of such complexes are morphologically similar and partially reproductively isolated from each other. The Anopheles gambiae complex consists of seven African malaria mosquito sibling species. Anopheles gambiae and An. arabiensis, the two major vectors of malaria in Africa, are both anthropophilic and can breed in temporal freshwater pools. Anopheles gambiae occupies more humid areas, while An. arabiensis dominates in arid savannas and steppes. Anopheles merus and An. melas breed in saltwater, and the habitat of An. bwambae is restricted to mineral water breeding sites. These three species are relatively minor malaria vectors mainly due to narrow geographic distributions [4]. Anopheles quadriannulatus A and An. quadriannulatus B are freshwater breeders and, although to various degrees susceptible to Plasmodium infections, are not natural vectors of malaria mainly due to zoophilic behavior [5]–[7]. Inferring the evolutionary history of the An. gambiae complex could be crucial for identifying specific genomic changes associated with the human blood choice, breeding site preference, and variations in vector competence. However, the high degree of genetic similarity, caused by the ancestral polymorphism and introgression, complicates the use of molecular markers for the reconstruction of a sibling species phylogeny [8]–[10]. Even the most recent genome-wide transcriptome-based phylogeny reconstruction of multiple Anophelinae species could not unambiguously resolve the relationships among An. gambiae, An. arabiensis, and An. quadriannulatus [11].
An alternative approach to inferring the phylogenetic relationships among species is to analyze the distribution of fixed overlapping inversions [4], [7], [12]. This approach is based on the fact that species-specific inversions do not introgress [13] and that inversions are predominantly monophyletic, despite rare occurrences of breakpoint reuse [14]. In addition, chromosomal inversions are more rare events and more consistent characters as compared with nucleotide substitutions [12], [15]. Phylogenies based on inversion data are highly congruent with phylogenies based on DNA sequence data and are shown to be more information rich than are nucleotide data [15]. Members of the An. gambiae complex carry 10 fixed inversions that can be used for a phylogeny reconstruction [7]. Five fixed inversions are present on the X chromosome, three inversions are found on the 2R arm, and one is found on each of the 2L and 3L arms (Figure S1) [7]. The only nonvectors in the complex, Anopheles quadriannulatus A and B, had been traditionally considered the closest species to the ancestral lineage because they have a large number of hosts, feed on animal blood, tolerate temperate climates, exhibit disjunctive distribution, and possess a “standard” karyotype [4], [7], [16], [17]. More recently, the An. arabiensis karyotype had been assumed ancestral because it has the fixed 2La inversion, which was also found in two outgroup species from the Middle Eastern An. subpictus complex [18]. Both chromosomal phylogenies assumed the most recent speciation of An. merus and an independent origin of the cytologically identical 2La′ inversion in this species [19]. A phylogenetic status of an inversion can be determined more precisely when breakpoints are identified and gene orders across breakpoints are compared between ingroup and multiple outgroup species. The genes found across inversion breakpoints in ingroup and outgroup species are expected to be in their ancestral order [12]. For example, the molecular analysis of the 2La inversion breakpoints and physical mapping of the sequences adjacent to the breakpoints in outgroup species identified the shared 2La inversion in An. gambiae, An. merus, and An. arabiensis and determined the ancestral state of the 2La arrangement [20]–[22].
Based on the X chromosome fixed inversions, three species clades can be identified in the complex: (i) An. bwambae, An. melas, and An. quadriannulatus A and B (X+), (ii) An. arabiensis (Xbcd), and (iii) An. merus and An. gambiae (Xag) (Figure 1). The An. gambiae–An. merus and An. bwambae–An. melas sister taxa relationships have been supported by independent phylogenetic analyses of nuclear genes and mitochondrial DNA sequences [9], [10], [23]. Each clade has unique fixed inversions that can be used to unambiguously determine its phylogenetic status if compared to gene arrangements in outgroup species: X+, 2Rm, 3La in the An. bwambae–An. melas–An. quadriannulatus clade, Xbcd in An. arabiensis, and Xag, 2Ro, 2Rp in the An. gambiae–An. merus clade. However, to efficiently pursue this research was not possible until recently when genome sequences of several outgroup mosquito species became available, including An. stephensi (series Neocellia, subgenus Cellia, subfamily Anophelinae) (this paper), and Aedes aegypti and Culex quinquefasciatus (both from subfamily Culicinae) [24], [25]. In this study, we identified genes at the breakpoints of fixed overlapping inversions 2Ro and 2Rp of An. merus and homologous sequences in An. stephensi, Ae. aegypti, and C. quinquefasciatus. We demonstrated that the “inverted” 2Ro and the “standard” 2R+p arrangements are ancestral in the complex. In addition, we found that the “inverted” 2La arrangement is present in evolutionary distant Culicinae species and, therefore, is ancestral. The inversion data support the basal position of the An. gambiae–An. merus clade and the terminal positions of the An. arabiensis and An. melas lineages. This rooted chromosomal phylogeny could be a means to examine specific genomic changes associated with evolution of traits relevant to vectorial capacity.
To infer the ancestral-descendant relationships among chromosomal arrangements in the An. gambiae complex, we determined gene orders at the breakpoints of the An. merus-specific fixed overlapping inversions 2Ro and 2Rp in ingroup and several outgroup species, including An. stephensi, Ae. aegypti, and C. quinquefasciatus. In our first approach, we used An. gambiae DNA probes, which were identified at breakpoints of “standard” 2R+o and 2R+p arrangements, for the mapping to polytene chromosomes of An. merus and An. stephensi by fluorescence in situ hybridization (FISH). In our second approach, we performed mate-paired sequencing of the An. merus genome and mapped the read pairs to the An. gambiae AgamP3 genome assembly. The inversion breakpoints of 2Ro and 2Rp in the An. gambiae–An. merus clade and their homologous sequences in the outgroup species were obtained and analyzed. This study reconstructed a rooted chromosomal phylogeny and revised evolutionary history of the An. gambiae complex.
We mapped multiple An. gambiae DNA probes derived from the cytological breakpoints to the chromosomes of An. merus by FISH. Anopheles gambiae BAC clone 141A14 that spans the proximal 2R+o breakpoint was identified by comparative mapping with An. merus in our previous study [21]. FISH of the BAC clone to An. merus chromosomes produced two separate signals on 2R indicating an inversion. Reiteration of this procedure with PCR fragments derived from the BAC clone allowed us to localize the breakpoint region within the BAC between genes AGAP002933 and AGAP002935. Further comparative mapping with An. merus demonstrated that the distal 2R+o breakpoint in An. gambiae is located between genes AGAP001759 and AGAP001762 (Figure S2). We also performed FISH with polytene chromosomes of An. merus using multiple probes located near the 2R+p cytological breakpoints of An. gambiae. The proximal 2R+p breakpoint was found between genes AGAP003327 and AGAP003328, and the distal 2R+p breakpoint was localized between AGAP001983 and AGAP001984 in An. gambiae. These gene pairs were neighboring in the genome of An. gambiae, but they were mapped in separate locations in An. merus (Figure S3). To determine gene arrangements in an outgroup species, we mapped genes at the 2R+o and 2R+p breakpoints to polytene chromosomes of An. stephensi (Figure S4 and Figure S5). The FISH results showed that the “inverted” 2Ro and “standard” 2R+p arrangements are present in the outgroup species An. stephensi (Figure 2).
We performed mate-paired sequencing of the An. merus genome and mapped the read pairs to the An. gambiae AgamP3 genome assembly, which has all “standard” arrangements [26], [27]. Mate-paired sequencing is the methodology that enables the generation of libraries with inserts from 2 to 5 kb in size. The 2 kb, 3 kb, and 5 kb DNA fragments were circularized, fragmented, purified, end-repaired, and ligated to Illumina paired-end sequencing adapters. The final libraries consisted of short fragments made up of two DNA segments that were originally separated by several kilobases. These genomic inserts were paired-end sequenced using an Illumina approach. Paired-read sequences that map far apart in the same orientation delineate inversions [28]. We executed a BLASTN search to find read pairs mapped to the putative breakpoint regions in the same orientation on chromosome 2 (Figure 3 and Table S1). Alignment of the read pairs to the genome of An. gambiae identified the 2Ro breakpoints at coordinates ∼9.48 Mb and ∼29.84 Mb. We also identified the 2La breakpoints at coordinates ∼20.52 Mb and ∼42.16 Mb, which confirmed a previous study and, thus, validated the approach [20]. However, the BLASTN search did not find the paired-read sequences that map at the opposite 2Rp breakpoints in the same orientation. This approach could not detect breakpoint regions longer than 5 kb. The 2Rp breakpoint regions in An. merus likely have larger sizes caused by accumulation of repetitive sequences. We also used the Bowtie program [29] to confirm the genomics positions of the 2Ro breakpoints (Table S2). Both BLASTN and Bowtie results supported the position of the proximal 2Ro breakpoint to the region between genes AGAP001762 and AGAP002935, and they refined the position of the distal 2Ro breakpoint to the region between AGAP001760 and AGAP002933.
The genes adjacent to the 2Ro breakpoint were used as probes to screen the genomic phage library of An. merus. Positive An. merus phage clones were confirmed to span inversion breakpoints by FISH to polytene chromosomes of An. gambiae, An. merus, and An. stephensi. For example, hybridization of Phage 6D produced only one signal in the proximal 2Ro breakpoint in An. merus but two signals at both 2Ro breakpoints in An. gambiae (Figure S6). Phage 6D hybridized to only one locus in An. stephensi, confirming the 2Ro arrangement in this species. Confirmed phage clones were sequenced, and the exact breakpoint regions were identified by aligning the An. merus sequences and An. gambiae AgamP3, AgamM1, and AgamS1 genome assemblies available at VectorBase [26], [30], [31]. Thus, distal and proximal breakpoints were identified on polytene chromosome map [7] and in the genome assembly of An. gambiae (Figure 4). In the AgamP3 assembly, the distal and proximal breakpoint regions span coordinates 9,485,167–9,486,712, and 29,838,366–29,839,163, respectively. The 2Ro breakpoint regions were 2.6 and 5.9 times smaller in An. merus as compared with the 2R+o breakpoint regions in An. gambiae due to accumulation of transposable elements (TEs) in the latter species. The presence of TEs is a common signature of inversion breakpoints, as TEs usually mark breakpoints of derived arrangements [20], [32]. Five various DNA transposons were found at the distal 2R+o breakpoint, and one novel miniature inverted-repeat TE (MITE), Aga_m3bp_Ele1, was identified at the proximal 2R+o breakpoint in An. gambiae (Figure 4). Smaller sizes of the breakpoint regions and the lack of TEs at the breakpoints of An. merus strongly suggest the ancestral state of the 2Ro arrangement.
We determined gene orders at the breakpoints of the An. merus-specific fixed overlapping inversions 2Ro and 2Rp in several outgroup species, including An. stephensi, Ae. aegypti, and C. quinquefasciatus. The genes adjacent to the 2Ro and 2Rp breakpoint were used as probes to screen the genomic BAC library of the outgroup species An. stephensi. Sequences homologous to genes from the distal 2Ro breakpoint were found in the BAC clone AST044F8 of An. stephensi. In addition, we performed sequencing of the An. stephensi genome using 454 and Illumina platforms. Sequences homologous to genes from the proximal 2Ro breakpoint were identified in scaffold 03514 of the An. stephensi genome. We also detected homologous sequences in the genome assemblies of Ae. aegypti and C. quinquefasciatus available at VectorBase [27]. The analysis demonstrated that all studied outgroup species had the gene arrangement identical to that of An. merus confirming the ancestral state of the 2Ro inversion (Figure 5). The An. stephensi sequences, which correspond to the 2Ro breakpoints, had sizes more similar to those in An. merus than in An. gambiae, and they did not display any TEs or repetitive elements, further supporting the 2Ro ancestral state. However, we found TEs in sequences corresponding to one of the 2Ro breakpoints in Ae. aegypti. Incidentally, the areas between the homologous breakpoint-flanking genes were 12,055 bp in Culex and 31,352 bp in Aedes, and this probably reflects the repeat-rich nature of the Culicinae genomes. The demonstrated conservation of gene orders between Anophelinae and Culicinae species is remarkable given the ∼145–200 million years of divergence time between these two lineages [33].
Approximate genomic positions of the 2R+p breakpoints were determined between AGAP001983 and AGAP001984 and between AGAP003327 and AGAP003328 by physical mapping of An. merus chromosomes (Figure 2). Using these genes as probes, we obtained a positive Phage 3B of An. merus that was mapped to the proximal 2Rp breakpoint in An. merus (Figure S6). Sequencing and molecular analyses of Phage 3B revealed the presence of AGAP001983 and AGAP013533 in this clone indicating that the actual distal breakpoint is located between AGAP013533 and AGAP001984 in An. gambiae. However, the available Phage 3B sequence ended at gene AGAP013533 and, thus, did not encompass the actual breakpoint sequence in An. merus. We performed the comparative analysis of gene orders at the 2Rp breakpoints in three outgroup species, An. stephensi, C. quinquefasciatus, and Ae. aegypti. The results demonstrated the common organization of the distal 2R+p breakpoint in An. gambiae and outgroup species, indicating that this arrangement is ancestral (Figure 6). Interestingly, a gene similar to AGAP013533 was absent, but genes similar to AGAP001983 and AGAP001984 were present in supercontig 3.153 of C. quinquefasciatus. Genes similar to AGAP003327 and AGAP003328 were found in different scaffolds and supercontigs of the outgroup species. This pattern was expected because AGAP003327 and AGAP003328 were mapped to neighboring but different subdivisions on the An. stephensi chromosome map (Figure 2). Therefore, it is possible that an additional inversion separated these two genes in the An. stephensi lineage. The highly fragmented nature of the C. quinquefasciatus and Ae. aegypti genome assemblies could also explain the observed pattern. No TEs were found in the breakpoint regions of An. stephensi and C. quinquefasciatus. However, multiple TEs were found in the intergenic regions of An. gambiae and Ae. aegypti (Figure 6).
Using sequencing and cytogenetic approaches, the common 2La arrangement was previously found in An. gambiae, An. merus, and An. arabiensis [4], [20], as well as in several outgroup species, including An. subpictus [18], An. nili, and An. stephensi [22]. Here, we used sequences available for breakpoints of the 2La inversion [20] to execute BLAST searches against genomes of more distantly related outgroup species C. quinquefasciatus and Ae. aegypti. BLAST results of genes adjacent to the 2La proximal breakpoint, AGAP007068 and AGAP005778, identified orthologs CPIJ004936 and CPIJ004938 in the Culex genome as well as orthologs AAEL001778 and AAEL001757 in the Aedes genome. These genes were found within supercontig 3.77 in C. quinquefasciatus and within supercontig 1.42 in Ae. aegypti. Similarly, BLAST results of genes neighboring with the 2La proximal breakpoint, AGAP007069 and AGAP005780, identified homologous genes CPIJ005693 and CPIJ005692 in the Culex genome (supercontig 3.99) as well as AAEL011139 and AAEL011140 in the Aedes genome (supercontig 1.543). The obtained data confirmed the identical gene arrangement in distant outgroup species and the ancestry of the 2La inversion.
Physical chromosome mapping and bioinformatic analyses identified the 2Ro and 2R+p arrangements in several outgroup species indicating that these arrangements are ancestral (Figure 5 and Figure 6). Because these two inversions overlap, only certain evolutionary trajectories and inversion combinations are possible (Figure 2). Specifically, the 2Rop–2Ro+p–2R+op order of inversion events is possible, while the 2Rop–2R+op–2R+op evolutionary sequence is not possible, regardless of the direction. Identification of 2Ro and 2R+p as the ancestral arrangements agrees well with this argument. We have also examined three different scenarios in reconstructing chromosomal phylogeny based on the established ancestry of 2Ro, 2R+p, and 2La and on the alternative hypothetical ancestries of X chromosomal arrangements (X+, Xag, or Xbcd) using the Multiple Genome Rearrangements (MGR) program [34]. Three different X chromosome arrangements (X+, Xag, and Xbcd) in an outgroup species were examined (Figure S7). The MGR program calculated the phylogenetic distances among species related to the ancestry of the X chromosome arrangement. Three hypothetical trees were obtained and used for interpretation of phylogenetic relationship and inversion reuse in the complex. Of the three scenarios, only the phylogeny based on the ancestry of 2Ro, 2R+p, 2La, and Xag had all inversions originating only once in the evolution of the An. gambiae complex. The other scenarios (with X+ and Xbcd being ancestral) had multiple origins of one of the inversions implying that they are less parsimonious (Figure S7). Because Xag uniquely characterize the An. gambiae–An. merus clade, these two species have the least chromosomal differences from the ancestral species of the complex as compared with other members (Figure 7). The ancestry of Xag can be tested by mapping of the X chromosome genome sequences from several species of the An. gambiae complex, which soon will be available [10]. Importantly, the new phylogeny is in complete agreement with the previous discoveries of 2La being the ancestral arrangement [18], [20]. Moreover, this is the first phylogeny based on knowledge about the status of a species-specific inversion (2Ro of An. merus). Therefore, the future data on the ancestry of the X chromosome arrangement are expected to support the new phylogeny.
Speciation in the An. gambiae complex has been accompanied by fixation of chromosomal inversions, except for speciation within the An. quadriannulatus lineage [7], [35]. Therefore, the chromosomal phylogeny likely reflects the species' evolutionary history. For a long time, the An. quadriannulatus lineage had been traditionally considered ancestral [4], [7], [16], [17] (Figure 8A). This evolutionary history was reconstructed from an unrooted phylogeny without any knowledge about chromosomal arrangements in outgroup species. Later, the An. arabiensis lineage had been assumed basal because it has the fixed ancestral 2La inversion and based on knowledge about biogeography and ecology of An. arabiensis [18] (Figure 8B). In these two scenarios, saltwater species An. merus and An. melas had been assumed the most recently originated members in the complex. However, the ancestry and the unique origin of the 2La inversion [20] imply that An. arabiensis, An. gambiae, or An. merus could be the closest to the ancestral species. The new chromosomal phylogeny led us to the substantial revision of the evolutionary history of the An. gambiae complex (Figure 8C). Accordingly, the ancestral species with 2Ro, 2R+p, and 2La arrangements might have arisen in East Africa where An. merus and An. gambiae are present in sympatry. The ancestral species may have been polymorphic for the 2Rp and 2R+o inversions and one lineage or population gave rise to An. merus with the 2Rp inversion while the other gave rise to the sister species An. gambiae containing the 2R+o inversion. Otherwise one would have to postulate that An. gambiae and An. merus arose from independent ancestors. At some point in evolutionary history, An. gambiae acquired polymorphic 2La/+ inversion and entered forested regions in central Africa. Later, An. gambiae acquired multiple polymorphic inversions on 2R, which allowed this species to spread to the arid areas of West Africa [4]. A hypothetical karyotype might have originated from the An. gambiae chromosomal arrangements by acquiring X+ag inversions. This karyotype in turn gave rise to the An. arabiensis chromosomes by generating the Xbcd inversions and fixing 2La and to the An. quadriannulatus karyotype by fixing the 2L+a arrangement. The 3La inversion in An. bwambae originated from the An. quadriannulatus karyotype, followed by the origin of the 2Rm inversion in An. melas.
The two major malaria vectors An. arabiensis and An. gambiae are sympatric species in most of their distribution range, allowing for introgressive hybridization between them. Available data support the hypothesis of introgression of the 2La arrangement from An. arabiensis into An. gambiae [9], [36], [37]. According to the new chromosomal phylogeny, introgression of 2La has been happening from the more derived karyotype of An. arabiensis to the more ancestral karyotype of An. gambiae. Therefore, the 2La arrangement in isolated An. gambiae populations must retain alleles that are more distantly related to alleles of the 2La arrangement in An. arabiensis. This hypothesis can be tested by the genomic analysis of An. gambiae island populations that do not have a history of hybridization with An. arabiensis. Because the 2La inversion in An. gambiae mainland populations has been associated with a tolerance to aridity and slightly reduced susceptibility to Plasmodium falciparum [4], [38], [39], the expected differences between the “original” and “introgressed” 2La arrangements could impact our understanding of a role of the inversion polymorphism in mosquito adaptation and malaria transmission.
The results of this study indicate that An. merus is closely related to an ancestral species from which the An. gambiae complex arose. Anopheles merus is a minor vector of human malaria in African mainland. A role of An. merus in malaria transmission in Madagascar has also been documented [40]. Based on the unique origin of fixed inversions and X-linked sequences, An. merus and An. gambiae are considered sister taxa [9], [10]. Therefore, according to the new chromosomal phylogeny, these two species possess the most “primitive” karyotypes in the complex. Our data suggest that the major malaria vector in Africa An. gambiae could be more closely related to the ancestral species than was previously assumed. Unexpectedly, we found that the karyotype of nonvectors An. quadriannulatus A and B was derived from the karyotype of An. gambiae (Figure 7 and Figure 8). Anopheles quadriannulatus is not involved in malaria transmission in nature due to its strong preference for feeding on animals [7]. Anopheles melas has the most recently formed karyotype and is a malaria vector in West Africa [41], [42].
The new chromosomal strongly suggests that vectorial capacity evolved repeatedly in the An. gambiae complex. Increased anthropophily could not have evolved in An. gambiae and An. arabiensis before humans originated and evolved to high enough densities. Therefore, the ability to effectively transmit human malaria must be a relatively recent trait in the complex. If An. quadriannulatus were the ancestral species, as it was assumed earlier [4], [7], then vectorial capacity could have originated only once when all other members split from the An. quadriannulatus lineage (Figure 8A). However, if the An. gambiae–An. merus clade is ancestral, as we demonstrated here, then vectorial capacity must have arisen independently in different lineages after the species were diversified. The available data cannot clearly delineate between the loss of vectorial capacity in An. quadriannultus and its subsequent reappearance in An. bwambae and An. melas with a possible alternative that vectorial capacity in present day An. quadriannulatus was only lost after An. bwambae and An. melas split from the An. quadriannulatus lineage. Depending on when the phenotypic change occurred (before or after An. bwambae/An. melas split from the An. quadriannulatus lineage) different scenarios are possible. However, even if a zoophilic behavior was acquired by An. quadriannulatus after the split from An. bwambae and An. melas, one still has to assume repeated origin of vectorial capacity. In this case, it originated independently in An. gambiae, An. merus, An. arabiensis, and the lineage that led to An. quadriannulatus/An. bwambae/An. melas. This alteration of the phylogeny of the An. gambiae species complex will likely have direct impact on studies aimed at understanding the genetic basis of traits important to vectorial capacity.
The chromosomal phylogeny also supports the idea of multiple origins of similar ecological adaptations in the complex. An early cytogenetic and ecological study postulated the repeated evolution of saltwater tolerance in the complex [4]. Anopheles melas and An. merus breed in saltwater pools in western and eastern Africa, respectively. Our finding revealed that the physiological adaptation to breeding in saltwater originated first in An. merus and then independently in An. melas.
Because of the high degrees of genetic similarities among sibling species, attempts to use molecular markers to reconstruct phylogenetic trees often fail [10]. Our study provides the methodology for rooting chromosomal phylogenies of sibling species complexes, which are common among disease vectors, including blackflies, sandflies, and mosquitoes [1]–[3]. The robustness of this methodology is supported by the agreement between the two alternative approaches to breakpoint mapping (cytogenetics and sequencing) and by the consensus among the three inversions in the phylogenic analysis (2Ro, 2Rp, and 2La). As we demonstrated, inversion breakpoints can be physically mapped on polytene chromosomes by FISH and identified within genomes by mate-pair and clone sequencing. Importantly, the increasing availability of sequenced and assembled genomes provides an opportunity for identification of gene orders in multiple outgroup species for rooting chromosomal phylogenies.
The high genetic similarity among the species of the An. gambiae complex suggests their recent evolution [10], [18]. The identified chromosomal relationships among the species demonstrate rapid gains and losses of traits related to vectorial capacity and ecological adaptations. This study reinforces the previous observations that vectors often do not cluster phylogenetically with nonvectors [1], [10]. The genome sequences for several members of the An. gambiae complex are soon to be released [10], and the new chromosomal phylogeny will provide the basis for proposing hypotheses about the evolution of epidemiologically important phenotypes. An intriguing question is whether or not evolution of independently originated traits, such as anthropophily and salt tolerance, is determined by changes of the same genomic loci in different species. In addition, the revised phylogeny will affect the interpretation of results from population genetics studies such as shared genetic variation and the detection of signatures of selection. Specifically, variations shared with An. merus but not with An. quadriannulatus would be interpreted now as ancestral. Knowledge about how evolutionary changes related to ecological and behavioral adaptation and how susceptibility to a pathogen in arthropod vectors had happened in the past may inform us about the likelihood that similar changes will occur in the future.
The OPHASNI strain of An. merus, the Indian wild-type laboratory strain of An. stephensi, and the SUA2La strain of An. gambiae were used for chromosome preparation. To obtain the polytene chromosomes, ovaries were dissected from half-gravid females and kept in Carnoy's fixative solution (3 ethanol: 1 glacial acetic acid) in room temperature overnight. Follicles of ovaries were separated in 50% propionic acid and were squashed under a cover slip. Slides with good chromosomal preparations were dipped in liquid nitrogen. Then cover slips were removed, and slides were dehydrated in a series of 50%, 70%, 90%, and 100% ethanol.
Multiple An. gambiae DNA probes derived from the cytological breakpoints of An. gambiae were physically mapped to the chromosomes of An. merus and An. stephensi. DNA probes obtained from PCR products were labeled by the Random Primers DNA Labeling System (Invitrogen Corporation, Carlsband, CA), and phage clones were labeled by the Nick Translation Kit (Amersham, Bioscience, Little Chalfont Buckinghamshire, UK). DNA probes were hybridized to chromosome slides overnight at 39°C. Then chromosomes were washed with 1× SSC at 39°C and room temperature. Chromosomes were stained with 1 mM YOYO-1 iodide (491/509) solution in DMSO (Invitrogen Corporation, Carlsbad, CA, USA) and were mounted in DABCO (Invitrogen Corporation, Carlsbad, CA, USA). Images were taken by a laser scanning microscope and by the fluorescent microscope. Location of the signals was determined by using a standard photomap of An. stephensi [43] and An. gambiae [44].
Mate-paired whole genome sequencing was done on genomic DNA isolated from five adult males and females of An. merus. Genomic DNA of An. merus was isolated using the Blood and Cell Culture DNA Mini Kit (Qiagen Science, Germantown, MD, USA). Three libraries of 2 kb, 3 kb, and 5 kb were obtained. These libraries were used for 36 bp paired-end sequencing utilizing the Illumina Genome Analyzer IIx at Ambry Genetics Corporation (Aliso Viejo, CA, USA). The 16× coverage genome assembly for An. stephensi was obtained by sequencing genomic DNA isolated from Indian wild-type laboratory strain. The sequencing was done using Illumina and 454 platforms at the Core Laboratory Facility of the Virginia Bioinformatics Institute, Virginia Tech.
Screening the An. merus Lambda DASH II phage library with genes adjacent to standard 2R+o and 2R+p was performed. To prepare probes for screening phage and BAC libraries, genomic DNA of An. gambiae was prepared using the Qiagen DNeasy Blood and Tissue Kit (Qiagen Science, Germantown, MD, USA). Primers were designed for genes adjacent to breakpoints using the Primer3 program [45]. PCR conditions were the following: 95°C for 4 min; 35 cycles of 94°C for 30 s, 55°C for 30 s, and 72°C for 30 s; and 72°C for 5 min. All PCR products were purified from the agarose gel using GENECLEAN III kit (MP Biomedicals, Solon, OH, USA). DNA probes were labeled based on random primer reaction with DIG-11-dUTP from DIG DNA Labeling Kit (Roche, Indianapolis, IN, USA). Anopheles merus Lambda DASH II phage library and An. stephensi BAC library (Amplicon Express, Pullman, WA, USA) were screened. Library screening was performed using the following kits and reagents (Roche Applied Science, Indianapolis, IN) according to protocols supplied by the manufacturer: Nylon Membranes for Colony and Plaque Hybridization, DIG easy Hyb, DIG Wash and Block Buffer Set, Anti-Dioxigenin-AP, and CDP Star ready to use. Positive phages were isolated with Qiagen Lambda midi Kit (Qiagen Science, Germantown, MD, USA), and positive BAC clones were isolated using the Qiagen Large Construct Kit (Qiagen Science, Germantown, MD, USA).
Primers 1760RCL (5′AGCAACAGGGACGATTTGTT3′) and 2933RCL (5′CTCGCTTTGGTTTGTGCTTT3′) were designed based on AGAP001760 and AGAP002933 sequences, and they were used to obtain the distal 2Ro breakpoint from Phage 7D DNA. The PCR conditions with Platinum PfX DNA polymerase (Invitrogen, Carlsbad, CA, USA) were: 94°C for 2 min; 35 cycles of 94°C for 15 s, 55°C for 30 s, and 68°C for 2 min; and 68°C for 10 min. Sanger sequencing of Phage 7D was performed using an ABI machine at the Core Laboratory Facility of the Virginia Bioinformatics Institute, Virginia Tech. Other positive phage and BAC clones were completely sequenced by the paired-end approach using an Illumina platform. Libraries of phages and BAC clones were made using Multiplex Sample Preparation Oligonucleotide Kit and Paired End DNA Sample Prep Kit (Illumina, Inc., San Diego, CA). Paired-end sequencing was performed on the Illumina Genome Analyzer IIx using 36 bp paired-end processing at Ambry Genetics Corporation (Aliso Viejo, CA, USA).
Phage clone of An. merus, BAC clone of An. stephensi, and genome sequences of An. merus, An. stephensi, An. gambiae, C. quinquefasciatus, and Ae. aegypti were analyzed with BLASTN, TBLASTX, and BLAST2 using the laboratory server and the Geneious 5.1.5 software (www.geneious.com), a bioinformatics desktop software package produced by Biomatters Ltd. (www.biomatters.com). Identification of the accurate breakpoint was performed by aligning the An. merus sequences and An. gambiae AgamP3, AgamM1, and AgamS1 genome assemblies available at VectorBase [27]. The DNA transposons and retroelements were analyzed by using the RepeatMasker program [46] and by comparing to Repbase [47] and TEfam (http://tefam.biochem.vt.edu/tefam/) databases. To characterize novel TEs in the breakpoint, each candidate sequence was used as a query to identify repetitive copies in the genome using BLASTN searches. These copies, plus 1000 bp flanking sequences, were aligned using CLUSTAL 2.1 to define the 5′ and 3′ boundaries. Using this approach, a novel MITE was discovered in the An. gambiae breakpoint. According to the TEfam naming convention, this MITE was named Aga_m3bp_Ele1 because it was associated with a 3 bp target site duplication.
All sequence data have been deposited at the National Center for Biotechnology Information short read archive (www.ncbi.nlm.nih.gov/Traces/sra/sra.cgi) as study no. SRP009814 of submission no. SRA047623 and to the GenBank database (http://www.ncbi.nlm.nih.gov/Genbank/) as accession nos.: JQ042681–JQ042688.
|
10.1371/journal.pntd.0005197 | Mitigating Diseases Transmitted by Aedes Mosquitoes: A Cluster-Randomised Trial of Permethrin-Impregnated School Uniforms | Viral diseases transmitted via Aedes mosquitoes are on the rise, such as Zika, dengue, and chikungunya. Novel tools to mitigate Aedes mosquitoes-transmitted diseases are urgently needed. We tested whether commercially insecticide-impregnated school uniforms can reduce dengue incidence in school children.
We designed a cluster-randomised controlled trial in Thailand. The primary endpoint was laboratory-confirmed dengue infections. Secondary endpoints were school absenteeism; and impregnated uniforms’ 1-hour knock-down and 24 hour mosquito mortality as measured by standardised WHOPES bioassay cone tests at baseline and after repeated washing. Furthermore, entomological assessments inside classrooms and in outside areas of schools were conducted.
We enrolled 1,811 pupils aged 6–17 from 5 intervention and 5 control schools. Paired serum samples were obtained from 1,655 pupils. In the control schools, 24/641 (3.7%) and in the intervention schools 33/1,014 (3.3%) students had evidence of new dengue infections during one school term (5 months). There was no significant difference in proportions of students having incident dengue infections between the intervention and control schools, with adjustment for clustering by school. WHOPES cone tests showed a 100% knock down and mortality of Aedes aegypti mosquitoes exposed to impregnated clothing at baseline and up to 4 washes, but this efficacy rapidly declined to below 20% after 20 washes, corresponding to a weekly reduction in knock-down and mosquito mortality by 4.7% and 4.4% respectively. Results of the entomological assessments showed that the mean number of Aedes aegypti mosquitoes caught inside the classrooms of the intervention schools was significantly reduced in the month following the introduction of the impregnated uniforms, compared to those collected in classrooms of the control schools (p = 0.04)
Entomological assessments showed that the intervention had some impact on the number of Aedes mosquitoes inside treatment schools immediately after impregnation and before insecticidal activity declined. However, there was no serological evidence of protection against dengue infections over the five months school term, best explained by the rapid washing-out of permethrin after 4 washes. If rapid washing-out of permethrin could be overcome by novel technological approaches, insecticide-treated clothes might become a potentially cost-effective and scalable intervention to protect against diseases transmitted by Aedes mosquitoes such as dengue, Zika, and chikungunya.
ClinicalTrials.gov NCT01563640
| Viral diseases transmitted via Aedes mosquitoes are on the rise, such as Zika, dengue, and chikungunya. Novel tools to mitigate Aedes mosquitoes-transmitted diseases are urgently needed. We tested whether commercially available insecticide-impregnated school uniforms can reduce dengue incidence in school children. To test this hypothesis we designed a school based randomized controlled trial where we enrolled 1,811 school children aged 6–17. For study monitoring, we also measured the effect of the impregnated uniforms on the survival of Aedes mosquitoes based on a standard bioassay test called WHOPES cone test. Furthermore, we counted the number of Aedes mosquitoes in classrooms and outside areas of classrooms. In the control schools, 3.7% and in the intervention schools 3.3% of the students had evidence of new dengue infections during the 5 month long school term, which indicates that there was no protection against dengue infections despite the fact that the knockdown effect of the impregnated uniforms was very high in the laboratory. We also showed a significant reduction of Aedes mosquitoes in the classrooms of the intervention schools. So why did this not translate into clinical protection against dengue? We assume the reason was the rapid wash-out effect of permethrin. Despite the company’s claim that impregnated clothing would withstand up to 70 launderings, we found a rapid decline in permethrin efficacy already after 4 washes, with the efficacy to below 20% after 20 washes. If rapid washing-out of permethrin could be overcome by novel technological approaches, insecticide-treated clothes might become a potentially cost-effective and scalable intervention to protect against diseases transmitted by Aedes mosquitoes such as dengue, Zika, and chikungunya.
| Aedes is a genus of mosquitoes originally found in tropical and subtropical zones, but which is now widespread in most continents. The current geographical distribution is the widest ever recorded and there is potential for further spread [1]. Over 50% of the world’s population live in areas where they are at risk of Aedes transmitted infections [2]. Aedes mosquitoes transmit a large array of viral diseases, notably dengue, Zika, yellow fever, West Nile encephalitis, chikungunya viruses. Aedes mosquitoes are predominantly active during daylight hours, well adapted to urban living, and are difficult to control with currently available control strategies [3]. Although novel approaches, such as Wolbachia-infected mosquitoes or genetically modified male mosquitoes, appear promising [4,5], their use remains controversial and they are unlikely to be implemented at a large scale in the near future. The current Zika virus outbreak in Latin America, associated with severe complications such as microcephaly and neurological complications [6], accentuates the urgent need to develop additional novel approaches that are simple and rapidly scalable; in particular personal protection approaches to protect individuals at high risk such as pregnant women. Since Aedes aegypti is a day-biting mosquito, developing technologies that can be applied during the day to offer protection against mosquito bites should be top priority.
Permethrin is a pyrethroid-based insecticide registered by the US Environmental Protection Agency (EPA) since 1977 that has been extensively used as insect repellent and insecticide with a documented safety record [7]. Permethrin can be bound to fabric fibres in clothing via different techniques such as micro-encapsulation and polymer coating [8,9]. Insecticide-treated clothing has been used for many years by the military and in recreational activities as personal protection against bites from a variety of arthropods including ticks, chigger mites, sandflies and mosquitoes and is thought to be safe [7]. Insecticide-treated clothing has been reported to give between 0% and 75% protection against malaria and between 0% and 79% protection against leishmaniasis [7]. No field trials have been conducted to demonstrate the efficacy against Aedes-transmitted diseases. However, one study showed that wearing permethrin-treated clothing resulted in a reduction in the number of Aedes bites by 90% [10] suggesting that it could potentially be a promising intervention for Aedes-transmitted diseases such as Zika, dengue, and chikungunya.
Of all Aedes-transmitted diseases, dengue is the most frequent arboviral disease globally, with an estimated 390 million infections annually [11]. Children carry a significant burden of morbidity and mortality for dengue with a higher rate of more severe disease than adults [12]. Because children spend most of their day at school at a time of peak biting behaviour of Aedes mosquitoes, personal protective measures such as impregnated clothing should be investigated. As school children in most dengue endemic countries wear school uniforms as a social norm, impregnated school-uniforms could be easily scaled up as a national programme, if such a strategy were proven to be impactful.
InsectShield manufactures permethrin-impregnated apparel for recreational and military purposes. InsectShield Repellent Apparel is registered by the US EPA [13]; their approach is a polymer-coating technique which is claimed to withstand up to 70 washings [13]. InsectShield clothing was successfully used for tick bite prevention [14]and also protective against mosquito bites by measuring changes in antibody titers to mosquito salivary gland extracts [15]. We conducted a school based field trial to assess the efficacy of InsectShield school uniforms on reducing dengue infections in children.
We conducted a randomised controlled trial blinded over a school-term (5 months) in Thailand to evaluate the effectiveness of insecticide-treated school uniforms for the prevention of dengue in school children under field conditions. The trial was funded by the European Commission under the 7th framework (Grant No. 282589) and registered at www.clinicaltrial.gov: NCT01563640).
The study area is located about 150 km east of Bangkok, in Plaeng Yao District, Chachoengsao Province, Eastern Thailand. The region covers an area of 237 km2 with a population of 36,607, with a total of 24 schools. Official permission from the Office of Basic Education Commission was obtained. School directors were informed about the trial, and 10 school directors agreed to participate. In these 10 schools, students in grades one to nine, aged 6 to 17 years, were eligible to participate if parental consent was given. Awareness seminars for the school senior management, teachers and parents were held prior to the trial to ensure high recruitment rates and compliance. The rainy season associated with the highest dengue incidence runs from June to October, and the school term (corresponding to the study period) was from mid-June to mid-November.
The protocol of this school-based trial was reviewed and approved by the Mahidol University Institutional Review Board (MU-IRB 2009/357.1512).
The intervention was permethrin-impregnated school uniforms. The impregnation process involved washing and then coating the uniforms in a proprietary process resulting in 0.054 mg/m2 permethrin (InsectShield USEPA2009). As, hypothetically, individuals wearing insecticide-impregnated clothing could also provide indirect protection to others not wearing impregnated clothing, we randomised the intervention by school, rather than by individual. To ensure acceptability by the school senior management and to ensure real-life scenarios, we used locally used school uniforms, comprising the standard school uniform, Scouts uniform, sports uniform and cultural uniform. Uniforms were typically short-sleeved and only covered the legs down to the knees (shorts or skirts).
Computer randomisation into intervention versus control group was by school, randomised into equal groups. Only the investigator who carried out the randomisation in Sweden and the overseas impregnating factory (InsectShield) were aware of the allocation; schools, children and on-site investigators were blinded. To maintain blinding for the on-site investigators and schools, all uniforms (from both intervention and control schools) were collected and sent to the InsectShield impregnating factory, but only the uniforms of the intervention schools were impregnated. To ensure that the correct uniforms were returned to the correct owners, all uniforms received labels indicating the child’s name, class, and school.
We collected blood samples via finger-prick (<0.2ml) from study subjects at the beginning and end of the school term (June and November). Dengue IgG ELISA was first performed for all paired samples. A primary infection was defined as a seroconversion from baseline negative IgG to positive IgG at follow-up. If the baseline sample was IgG positive, we did an additional analysis to identify new dengue infections by using the monoclonal antibody (MAb)-based capture enzyme-linked immunosorbent assays (MAb-ELISA) to measure the increase in IgG, whereby Dengue IgG indirect ELISA was performed using purified 2H2 monoclonal antibody for coating plates on paired samples and cut-off values used as described by Johnson et al [16].
Impregnated standard uniforms were evaluated for their effect on mosquito knock-down and 24-hour mortality at baseline and after repeated washing, measured by standardised WHOPES bioassay cone tests [17]. A WHO plastic cone was secured onto the cloth using rubber bands. Batches of five nulliparous starved female Aedes aegypti mosquitoes (3–5 days old) were placed in the cone via a mouth aspirator, and a small cotton plug was used to close the aperture. Bioassays were carried out at 25 ± 2°C and 65 ± 10% relative humidity. Mosquitoes were exposed to the materials for three minutes and removed using a mouth aspirator. The mosquitoes were then placed in a holding cup inside an insectary (25 ± 2°C and 65 ± 10%) with a net secured over the top with two elastic bands and cotton wool soaked in 10% sugar solution. Knock-down was recorded one hour post exposure, and mortality was recorded after 24 hours. Ten replicates were carried out for each sample of 10 shirts and 10 skirts or trousers. Testing was done at baseline and repeated after weekly laundering where the uniforms were hand-washed and dried in the shade for 24 hours to simulate field conditions.
Portable vacuum aspirators were used to collect adult mosquitoes indoors in the school areas at baseline and monthly following the introduction of treated uniforms. The collectors aspirated mosquitoes in 5 classrooms per school and spent 15 minutes per classroom for aspiration. BG sentinel traps, one trap per school, were placed in the corner of the corridor of each school building outside the classrooms. They were left in both treatment and control schools in the morning and were collected in the evening of the same day. The collectors collected mosquitoes from these traps in all schools on two consecutive days.
Mosquitoes collected were transferred to plastic tubes and were stored on ice during the transport to the laboratory at the Center of Excellence for Vectors and Vector-Borne Diseases in Salaya Campus of Mahidol University for further processing. The collected mosquitoes were then sorted, counted and identified to genus and species according to each school before storage at -80°C.
School class teachers recorded all children who were absent for at least one day. Children or their parents were contacted by phone to obtain the reason of absenteeism. Absenteeism for sick leave (for any cause, with or without fever) for at least one day was recorded, and those who were absent from school for at least 2 days due to a febrile illness were also documented.
The primary outcome was the incidence of laboratory confirmed dengue infections during the school term in individuals wearing impregnated uniforms versus non-impregnated uniforms. Secondary outcomes were (1) the number of Aedes aegypti mosquitoes in intervention and control schools (in classrooms, and school corridors) and (2) the 1 hour knock-down and 24 hour mortality of Aedes aegypti mosquitoes exposed to our impregnated school uniforms as measured by WHOPES cone tests at baseline and after weekly washing. We also recorded school absenteeism for 1 day or more, and also absenteeism for 2 days or more because of a febrile illness.
The sample size calculation based on the primary endpoint has been described in detail in the trial protocol [18]. We originally planned a cross-over design spanning two transmission seasons because dengue transmission can vary greatly between schools and seasons, which can to some extent be controlled for by a cross-over design. The original assumptions underlying the sample size for the trial were an incidence rate (symptomatic plus asymptomatic) averaging 5% during a transmission season [19], i.e. 10% over two seasons; an effect size of halving incidence by using impregnated uniforms, and a drop-out rate (children leaving the school or withdrawing from the trial, etc.) of 20%. Taking into account a conservative cluster design effect of 3, the total sample size was calculated to be 2,012 (i.e. 1,006 in each study arm). Retaining the design effect of 3 (since we did not gain appreciable understanding of the local geographic and temporal variability in dengue incidence), a sample size of 270 x 3 in each arm of the curtailed trial would have given 80% power to detect a significant difference of 7% versus 2% incidence between the two groups (alpha = 0.05). The total documented enrolment of 1,655 approximated to this sample size of 270 x 3 x 2 = 1,620. The 7% versus 2% difference corresponds to our original design assumption that a difference of at least 5% would be necessary in order to have policy implications.
Statistical analyses used STATA 12. The effectiveness of the intervention was determined by comparing proportions of students in intervention and control schools who had confirmed incident dengue infections during the trial, with adjustment for clustering by school as the unit of randomisation, using the clchi2 command. Differences in mean monthly numbers of Aedes mosquitoes trapped at intervention and control schools during the month in which the intervention was efficacious were assessed using the ttest command on ln(n+1) transformed values.
The flow chart (Fig 1) summarizes the numbers of participants in both intervention and control groups at all stages of the trial. The ten participating schools had 2,314 students at the start of the trial; consent was given by 1,811 students’ parents. Out of the 1,811 enrolled students (mean age 10.1 years; range 6–17; 908 males, 903 females), 11 were withdrawn because of skin irritation, which were all mild and transient (7 in intervention; 4 in control schools). Of the 1,800, 1,655 provided complete sets of paired blood samples (Table 1). The IgG baseline dengue seropositivity rate was 53.0% (878/1,655).
Of the 1,655 students with paired samples, 57 had evidence for a dengue infection during the study period, and 16 had equivocal results. Between schools, the number of dengue infections varied considerably from 0 to 9.9% (Table 1). In the control schools, 24/641 (3.7%) and in the intervention schools 33/1,014 (3.3%) students had evidence of new dengue infection. For our primary outcome, there was no significant difference in proportions of students having incident dengue infections between the intervention and control schools, with adjustment for clustering by school (χ2 = 0.02, p = 0.89).
The proportion of absenteeism (for any reason) was relatively high in all schools and ranged from an average of 23.0% (SD = 15.9%) over the school term, with a range of 9.6% to 49.9% in the treatment schools and 0.9% to 32.0% in the control schools. The proportion of pupils not going to school for at least 2 days due to a febrile illness ranged from an average of 1.4% (SD = 1.17%) over the school term from 0.6.% to 4.0% in the treatment schools and from 0.1% to 2.8% in the control schools. The overall proportion of absenteeism due to fever of at least 2 days was relatively stable over the school term. There were no statistical differences between the treatment and control groups for any of the school months.
Fig 2 shows both 1 h knock-down and 24 h mortality of Aedes mosquitoes exposed to the impregnated school uniforms. This started close to 100% and remained at high levels for up to 4 washes. After 4 washes, both knock-down and mortality declined rapidly. After 20 washes, the efficacy was below 20%. Knock-down decreased at an approximately linear rate of 4.7% per week, mortality at 4.4% per week. Due to this unexpectedly rapid waning of intervention efficacy, although the study was originally planned as a cross-over trial covering two dengue transmission seasons, we decided to abandon the second phase of the cross-over trial design.
Results of the entomological assessments showed that the mean number of Aedes aegypti mosquitoes caught inside the classrooms of the intervention schools was significantly lower in the month following the introduction of the impregnated uniforms, when compared to those collected in classrooms of the control schools (back-transformed mean of ln(n+1) transform in control schools 4.9, in intervention schools 1.4; ratio = 3.5, 95% CI 1.1 to 5.5). There was no significant difference in the mean numbers of Aedes aegypti mosquitoes collected at other times (control 4.0, intervention 3.8; ratio = 1.1, 95% CI 0.7 to 1.7).
Given the day-biting behaviour of Aedes mosquitoes, impregnated school uniforms could potentially be a simple novel tool to reduce mosquito-borne diseases and local vector populations. Our results from the WHOPES cone tests at the start of the trial underpin the potential for insecticide-treated uniforms to protect against dengue by reducing the populations of Aedes mosquitoes and hence mosquito-bites: knock-down effect and mortality immediately after impregnation of uniforms with permethrin by the InsectShield proprietary method were close to 100%, consistent with results obtained under laboratory conditions at the London School of Hygiene and Tropical Medicine. [20] Furthermore, we documented a significant reduction in Aedes mosquito numbers in the classrooms of the intervention schools in the first month after the start of the trial at the time when the insecticidal activity of impregnated uniforms was still 100%.
However, our cluster- randomised controlled trial in ten Thai schools involving 1,811 children did not show serological evidence of a protective effect over the 5-month study period of one school term. Given the theoretical support for this strategy, we need to carefully examine plausible reasons for the apparent failure to protect in our trial, so that this intervention is not discarded as an ineffective strategy for control against Aedes-transmitted diseases such as dengue, Zika or chikungunya. The main reason for the negative result was the rapidly waning efficacy of InsectShield permethrin-impregnated clothes under field conditions. We chose InsectShield factory-impregnation over hand-dipping with permethrin because InsectShield impregnation (unlike hand-dipping with permethrin) results in odourless, well tolerated apparel—a fact that is important for a double-blind randomised trial where odour could otherwise had given away the allocation group. We had not anticipated rapid washing-out of permethrin prior to the study as InsectShield have consistently claimed that the insecticidal efficacy of their proprietary method of permethrin impregnation withstands up to 70 washes [13]. However, we documented rapid declines in insecticidal activity after the first 4 washes. After 20 washes, the knock-down and mortality effects on mosquitoes were well below 20%. What might be the reasons for such rapid waning of insecticidal activity? Maybe the quality of already used cotton Thai school uniforms was inferior to clothing materials used by InsectShield for commercial purposes. With suboptimal cloth quality, the coating technique might have been less durable? Or maybe the washing conditions of the tropics, drying in the open air and ironing decreased the insecticidal effect more rapidly? However, we believe it was not just the potentially suboptimal cloth material of local Thai school uniforms, as a very recent study on laundering resistance of five commercially available, factory-treated permethrin-impregnated fabrics also found that permethrin content fell by 58.1 to 98.5% after 100 defined machine launderings, with InsectShield showing the fastest loss [21]. There is an urgent need for a standardised testing and licensing procedure for insecticide-impregnated commercial clothing to avoid misleading information.
We also documented a high heterogeneity in incident infections between the schools and differences in baseline seroprevalence, which may have masked the extent of the efficacy in our trial. Large differences in dengue infection rates between schools within the same year have also been noted by other groups [22,23], and underpin the difficulty in sample size calculations and subject selection of a disease that appears to be not only highly focal but also often exhibits a cyclical pattern with high and low epidemic years.
We need to consider other potential causes for the lack of efficacy in our trial. Although acceptability and compliance with the trial uniforms was high, [24] school uniforms are not worn after school and over weekends. We did some simulation modelling and estimated a reduction of dengue infections by 47% if 60% of all mosquito bites occurred during school hours and 70% of the children wore treated uniforms, assuming that mosquito knock-down and mortality levels remained at baseline (without washing-out effect) [25]. A reduction of dengue infections by 47% would indeed be a major public health victory.
Because we used paired blood samples at baseline and at the end of the study period (5 months), we are unable to tease out the impact of impregnated clothing on dengue incidence in the first month of wearing the intervention uniforms, at a time when the efficacy in terms of knock-down and mortality effect on mosquitoes was still close to 100%. As a proxy marker for symptomatic dengue disease, we recorded absenteeism of 2 days due to a febrile illness, but found no differences between the treatment and control groups, nor did we find statistically significant differences from month to month. Nevertheless, the results of entomological assessments showed that the intervention had an impact on the number of Aedes mosquitoes inside intervention schools in the first month of the intervention before insecticidal activity declined. This is an important finding that encourages continued research on the use of insecticide-treated clothing as a potential strategy for dengue prevention in school children.
Permethrin-treated clothes were also shown to be effective in reducing tick bites and mosquito bites in general in other studies, however, similarly as in our study, this effect waned off as time passed on, possibly also due to laundering effect [14,15]. Long-lasting insecticide-treated bednets were a major breakthrough in the control of malaria. However, bednets do not get washed so frequently. For insecticide-treated clothing to be a viable public health intervention it should withstand regular washing. If the rapid washing-out of permethrin can be overcome by novel technological approaches, insecticide-treated clothes would deserve to be re-evaluated as a potentially cost-effective and scalable intervention. Despite the urgency of the current Zika outbreak associated with serious pregnancy outcomes, we should not rush into recommending factory-impregnated clothing to pregnant women in Zika affected areas, until standardised testing and licensing procedures for insecticide-treated materials are implemented, with defined cut-off values for initial maximum and post-laundering minimum concentrations of permethrin as well as data on toxicity, homogeneity on fabrics, residual activity, and laundering resistance [21]. Failing to do this could generate a dangerous sense of security among Zika-exposed pregnant women using impregnated clothing, since the wearer has no means of judging insecticidal efficacy. Given the increasingly epidemic proportions of Aedes-transmitted viral infections, we hope that the findings from this trial will provide strong impetus to fund research to develop appropriate and safe technologies for long-lasting insecticide-treated clothing materials that can be used for school uniforms, work place uniforms and maternity clothing alike.
|
10.1371/journal.ppat.1005935 | Distinct Effects of p19 RNA Silencing Suppressor on Small RNA Mediated Pathways in Plants | RNA silencing is one of the main defense mechanisms employed by plants to fight viruses. In change, viruses have evolved silencing suppressor proteins to neutralize antiviral silencing. Since the endogenous and antiviral functions of RNA silencing pathway rely on common components, it was suggested that viral suppressors interfere with endogenous silencing pathway contributing to viral symptom development. In this work, we aimed to understand the effects of the tombusviral p19 suppressor on endogenous and antiviral silencing during genuine virus infection. We showed that ectopically expressed p19 sequesters endogenous small RNAs (sRNAs) in the absence, but not in the presence of virus infection. Our presented data question the generalized model in which the sequestration of endogenous sRNAs by the viral suppressor contributes to the viral symptom development. We further showed that p19 preferentially binds the perfectly paired ds-viral small interfering RNAs (vsiRNAs) but does not select based on their sequence or the type of the 5’ nucleotide. Finally, co-immunoprecipitation of sRNAs with AGO1 or AGO2 from virus-infected plants revealed that p19 specifically impairs vsiRNA loading into AGO1 but not AGO2. Our findings, coupled with the fact that p19-expressing wild type Cymbidium ringspot virus (CymRSV) overcomes the Nicotiana benthamiana silencing based defense killing the host, suggest that AGO1 is the main effector of antiviral silencing in this host-virus combination.
| To better understand the specific effect of p19 viral suppressor of RNA silencing (VSR) on antiviral silencing and endogenous small RNA pathways, we generated a N. benthamiana plant (p19syn) capable of sustaining the ectopic expression of the Cymbidium ringspot virus (CymRSV) p19 upon infection with a suppressor-deficient CymRSV (Cym19stop). By using wt and p19syn plants in combination with CymRSV and Cym19stop, we were able to analyze the effects of p19 provided “in trans” and “in cis” during the viral invasion of the plant. We have shown that p19 can efficiently sequester endogenous small RNAs (sRNAs) in mock-inoculated p19syn plants while it does not bind these sRNAs upon Cym19stop infection. Also, the presence of p19 in virus infection did not alter the expression of miRNAs significantly. These findings do not support the widely accepted assumption that viral symptoms are the direct consequence of the impact of VSRs on endogenous silencing pathways. We demonstrated that p19 preferentially sequesters positive:negative viral short interfering RNAs (vsiRNAs) pairs and that the binding by p19 is independent of vsiRNA sequence or the type of the 5’-end nucleotide. We have also found that 3’ truncation is induced on p19 bound sRNAs. Finally using AGO1- and AGO2- immunoprecipitation experiments we observed that p19 specifically compromises vsiRNAs’ loading into AGO1 but not AGO2. Since antiviral silencing is strongly inhibited by p19, this suggests that AGO1 is the main effector protein against CymRSV tombusvirus.
| Viruses are among the most important plant pathogens that cause huge economic losses in many agriculturally important crops worldwide. The invasion of the host by viruses deeply alters the physiology of the plants at cellular and tissue levels due to the interaction of the virus with the cellular pathways, which ultimately leads to viral symptom development. During evolution, plants have developed diverse strategies to combat virus infections. Amongst others, RNA silencing is one of the most important mechanisms that serve to fight against viruses [1,2].
RNA silencing is a conserved eukaryotic pathway involved in almost all cellular processes like development, stress responses and antiviral defense. RNA silencing relays on the 21–24 nt short interfering RNAs (siRNAs) or micro RNAs (miRNAs) the hallmark molecules of silencing [3]. The siRNAs and the miRNAs (collectively named small RNAs, sRNAs) are processed by RNase III-type ribonucleases, the DICER (in plants Dicer-Like, DCL) enzymes [4,5] in collaboration with their partner DOUBLE-STRANDED RNA BINDING (DRB) proteins [6–9]. sRNAs are 2’-O-methylated by HUA ENHANCER1 (HEN1) at their 3’ protruding ends [10], a reaction that serves to protect them against poly-uridylation and subsequent degradation [11]. sRNAs then associate with ARGONAUTE (AGO) proteins [12–14] the central effectors of RNA-induced silencing complex (RISC) [15,16]. Based on the sequence complementarity, sRNAs guide RISC to silence cognate RNAs through cleavage or translational repression (post-transcriptional gene silencing, PTGS) or induce chromatin/DNA modifications of the specific genomic locus (transcriptional gene silencing, TGS) [17–19]. In some specific cases, amplification of silencing occurs through double-stranded RNA synthesis by RNA-dependent RNA polymerases (RDRs) and secondary siRNA production [20–22]. sRNAs are non-cell autonomous, they can move within the plant to transmit gene silencing from cell-to-cell or systemically on long distance as mobile silencing signals [23–25].
Players of the antiviral silencing overlap with those of the endogenous silencing pathway [1,26]. Antiviral RNA silencing is triggered by the presence of viral dsRNA structures such as replication intermediates or intra—molecular fold—back structures of the invading virus. These dsRNA structures are processed by DCL4 or DCL2, to produce viral short interfering RNAs (vsiRNAs) [2,3,27–32]. The vsiRNAs guide self-silencing of the parental viral genomic RNA as part of the antiviral response through the action of AGO effectors [1,13]. Among the AGOs implicated in antiviral defense, AGO1 and AGO2 were identified as the most important players. AGO1 was shown to have antiviral roles against a number of viruses in A. thaliana [33–36], N. benthamiana [26,37,38] and in rice [39]. AGO2 was found to be important in the antiviral silencing response in A. thaliana [35,36,40–45]. In N. benthamiana, the important model organism for plant virology studies, AGO2 was proposed to protect against Potato virus X [46] and the suppressor-deficient Tomato bushy stunt virus (TBSV) [47,48]. However, recent observation suggested that AGO1 constitutes a solid layer of defense against tombusvirus infections [49]. As AGO1 is the negative regulator of AGO2, it is believed that AGO2 represents a second layer of antiviral defense [40].
Viruses, to counteract host defense, have evolved viral suppressors of RNA silencing (VSRs) providing strong evidence for the antiviral nature of silencing [1,50,51]. Most viruses studied so far were found to encode at least one VSR. VSRs were shown to block silencing at multiple steps like initiation, effector complex assembly, silencing amplification but also through transcriptional control of endogenous factors, hormone signal modulation or interaction with protein-based immunity [51,52]. The absence or inactivation of VSRs leads to the recovery of plants from viral infections, demonstrating the effect of plant antiviral silencing response [53–55]. Although several VSRs have been identified in the past, our knowledge about the precise molecular basis of their action and their multifunctional roles have only been resolved in a few cases [1].
The p19 protein of tombusviruses is one of the best-known VSR. Crystallographic studies have shown that p19 tail-to-tail homodimer acts as a molecular caliper to size-select and sequester siRNA duplexes in a sequence-independent manner [56–58] preventing the loading of siRNAs into AGO effector proteins [59,60]. Based on p19 expressing transgenic plants it was proposed that during virus infection p19 efficiently prevents miRNAs loading into RISC deeply compromising the endogenous miRNA pathways of the plants [61–63]. In particular, it was reported that three distinct VSRs (HcPro, p19 and P15) compromised the regulation of the miR167 target AUXIN RESPONSE FACTOR 8 (ARF8) when constitutively expressed in transgenic plants [64]. It was also proposed that misregulation of miR167 is the major cause for the developmental aberrations exhibited by VSR transgenic plants and for the phenotypes induced during viral infections [64]. Contradictory to these, other data suggests that during genuine Cymbidium ringspot virus (CymRSV) infections miRNA sequestration by p19 is very poor and may depend on spatial and temporal co-expression of miRNA duplex and the VSR [65]. vsiRNAs but not miR159 were shown to be sequestered efficiently into p19-homodimer:siRNA nucleoprotein complex, whereas miR159 was efficiently incorporated into RISC complex [66,67]. These findings suggest that, in virus-infected plants, p19 potently affects vsiRNA-pathway and at the much lesser extent the miRNA one. Besides, independently of its siRNA binding capacity, p19 similarly to other VSRs can promote miR168 transcriptional induction that results in miR168-guided AGO1 down-regulation [66,68]. Thus, the interaction of p19 with endogenous silencing pathways and its contribution to viral symptom development is far from being fully uncovered.
To better understand the impact of p19 on silencing and its role in viral symptom development we employed a synthetic p19-expressing transgenic N. benthamiana plant line (p19syn) or wild type plant as control in combination with its wild-type (CymRSV) or suppressor deficient virus (Cym19stop) infection: (i) wt virus infection of wt N. benthamiana (p19 in “cis”), (ii) Cym19stop-infection in wt N. benthamiana (virus present, no suppressor), (iii) Cym19stop-infection in p19syn plants (p19 in “trans”) and (iv) p19syn plants (p19 only). In this way we were able to compare the impact of p19 on its own or in the genuine virus-infected background.
We analyzed p19 ability to sequester vsiRNAs and plant endogenous sRNAs, with and without viral background. Consistently with our previous results, we have found that p19 can bind vsiRNAs when expressed either “in cis” (from the CymRSV wild-type virus, in wild-type N. benthamiana) or “in trans” (transgenically expressed in p19syn plants infected with the suppressor-deficient Cym19stop virus). In line with our previous findings, p19 efficiently bound endogenous sRNA duplexes only in the absence of the virus infection, suggesting that p19 impact on endogenous pathways is restricted. Analyzing the siRNA pool immunoprecipitated by p19 through high-throughput sequencing, we found that p19 changes the bias of positive vsiRNAs towards a more equilibrated positive/negative strand ratio, suggesting a preference for perfect ds-vsiRNAs. We also showed that p19 prevents mainly AGO1 but not AGO2 loading with vsiRNAs. This finding suggests a key role of AGO1 opposed to AGO2 during the antiviral response.
To uncouple p19 effects elicited by virus infection on RNA silencing and host plant symptom development we prepared p19-expressing N. benthamiana plants (Fig 1A–1C). To avoid the interference between the p19 transgene and the challenging virus (p19-deficient, Cym19stop), we modified the p19 transgene introducing all possible silent nucleotide changes. In this way, we reduced the nucleotide sequence similarity between the transgene and the challenging virus to 68% while keeping the amino acid identity at 100% (S1 Fig). These plants were named synthetic-p19 expressing plants (p19syn). The p19syn plants showed strong phenotype characterized with elongated stem internodes and typical leaf distortions (Fig 1A, 1C and 1D and S2A Fig) suggesting that the expressed p19 protein retained it suppressor activity, thus potentially compromising the endogenous silencing pathways. Importantly, this phenotype was clearly different from that of virus-infected stunted dwarf plant (S2A Fig). Transgenic line 1–57 was selected for further studies (Fig 1A and 1B). First we tested the silencing suppressor activity of transgenically expressed p19 in a GFP transient assay (see Materials and Methods). When GFP sense transgene was transiently expressed in wild-type plant leaves, spontaneously triggered silencing almost completely diminished GFP expression at four days post infiltration. In contrary when GFP was expressed in p19syn plants its expression was still strong as visualized under UV light (Fig 1D). The lack of GFP silencing in p19syn plants confirmed the suppressor activity of the p19 transgene. Next we tested p19 suppressor activity in an authentic virus infection context: we challenged the p19syn plants by the infection with Cucumber mosaic virus + yellow satellite RNA (CMV + Y-satRNA). CMV + Y-satRNA was reported to induce bright yellow symptoms on N. benthamiana through targeting the tobacco magnesium protoporphyrin chelatase subunit I (ChlI) gene involved in chlorophyll biosynthesis by Y-satRNA derived siRNA [69]. The CMV-Y-satRNA infected wt N. benthamiana plants developed the bright yellow symptoms while the infected p19syn plants failed to show the typical yellowing (Fig 1E). All these confirmed that the transgenically expressed p19 works as a silencing suppressor in vivo.
It is generally assumed that virus encoded suppressors strongly interfere with the endogenous silencing pathway and are central players in the development of viral symptoms [1,61–64]. However, this notion mostly comes from studies that used VSR-expressing transgenic plants without analyzing the effect of the VSRs in an authentic virus infection background. To reinvestigate this dogma we set up an experiment in which we could compare p19 effects (vsiRNA and endogenous sRNA binding) with or without its parental virus infection background. We compared the sRNA binding capability of p19 both in mock- and Cym19stop-inoculated p19syn plants. This setup allowed us to analyze the impact of p19 when provided “in trans” during virus infection. It is worth noting that the suppressor mutant Cym19stop virus in infected p19syn plant was able to invade whole leaves similarly to the CymRSV in wt plants (S2B Fig). In contrast, in the absence of p19, the Cym19stop virus is restricted to the veins [70] (S2B Fig). Besides this, we also inoculated wt plants with CymRSV to study p19 activity “in cis“. Based on previous studies [61,62] we expected p19 to bind ds-sRNAs of both plant and viral origin. p19 immunoprecipitations (IP) were prepared from pooled systemically-infected leaves of virus-infected plants and the corresponding mock-inoculated leaves of p19syn plants. cDNA libraries of sRNAs were generated using RNA samples isolated from inputs and p19 IPs. After quality control filtering and processing steps (see Materials and Methods), sequences flanked by the 3’ and 5’ Solexa adaptors, and ranging in length from 16 to 28 nt, were aligned to the N. benthamiana and the CymRSV genomes, respectively [71,72].
Analysis of p19-bound sRNAs from mock-inoculated p19syn plants revealed that p19 binds efficiently endogenous sRNAs (Fig 2A and 2B), including members of several miRNA families (S3A Fig) [73]. Surprisingly, the analysis of p19-bound sRNA libraries derived from both CymRSV-infected wt N. benthamiana (“cis“-p19) and Cym19stop-infected p19syn plants (“trans-p19“) have shown a different picture: p19 bound almost exclusively vsiRNAs but not endogenous sRNAs (Fig 2A and 2C–2F). This suggests that the abundantly produced vsiRNAs may outcompete the plant sRNAs from p19 binding during virus infection. Specific enrichment of vsiRNAs versus endogenous miR159, one of the most abundant miRNA, was quantified by Northern blot analysis. p19 had a much weaker affinity for miR159 during virus infection: the IP/input ratio of p19 bound miR159 was 1.2 mock-inoculated samples, while during virus infection (Cym19stop-infected p19syn plants) it dropped to 0.29 (Fig 3A). We also quantified the percentage of enrichments in case of vsiRNAs and endogenous sRNAs within p19 IPs compared to inputs from our deep seq data (Fig 2A). The input library of Cym19stop-infected p19syn plants contained 28% N. benthamiana reads while in the p19-IP they represented only 2% (p19 specifically enriched vsiRNAs from 72% in the input to 98% in the p19-IP). Similarly, in the CymRSV-infected wt plant 12% N. benthamiana reads in the input sample has dropped to 1% (p19 enriched the 88% vsiRNAs of the input to 99% in the p19-IP). We concluded therefore that p19 ability to sequester endogenous sRNAs is strongly decreased by the virus infection and p19 preference to vsiRNAs does not depend on the expressional origin of p19 protein (viral vs transgenic expression).
To better understand the biological relevance of vsiRNA-mediated endogenous sRNA binding and out-competition/release from p19 sequestration we analyzed the behavior of known miRNA-target mRNA pairs [73]. We compared RNAseq data obtained from mock-inoculated p19syn plant samples (when p19 binds to miRNAs) and from Cym19stop virus-infected p19syn plant samples (when p19 binds preferentially vsiRNAs while miRNAs are outcompeted/released). In the absence of the virus, p19 efficiently bound miRNA duplexes (S3A Fig) and this correlated with elevation of most of the miRNA-target mRNAs as the consequence of miRNA duplex sequestration by p19 and inability to program miRISC for cleavage (p19syn compared to wt N. benthamiana, S3B Fig). Upon Cym19stop virus infection however, the levels of most miRNA target RNAs were downregulated (compared to mock-infected p19syn) as the consequence of miRNA out-competition/release from p19 (S3B Fig). We went further and specifically looked to accumulation of trans-acting RNAs derived from TAS3 precursor, the target of miR390 [74] in a Northern blot assay (S3C Fig). In p19syn plants the level of miR390 was slightly elevated while the TAS3-derived D7 tasiRNA dropped below the detection level (compared to wt N. benthamiana). This was likely the consequence of the inhibition of the cleavage of TAS3 transcripts by p19-captured miR390. Indeed, miR390 is efficiently enriched in p19 IP (p19syn mock-infection, S3A Fig). When p19syn plants were infected with Cym19stop virus, miR390 binding by p19 decreased (S3A Fig), and consequently the activity of miR390 was restored that lead to D7 tasiRNA accumulation (to a similar level as detected in wt N. benthamiana, S3C Fig). Altogether our findings support the hypothesis that during virus infection p19 preferentially binds vsiRNAs while endogenous sRNAs are outcompeted/released from binding.
High-throughput sequencing analysis showed that CymRSV-derived vsiRNAs produced during infection have a strong bias towards positive strand polarity (95% positive, 5% negative polarity) (Fig 3B and S4 Fig). These data, which are in line with our and other previous observations, suggest that the majority of vsiRNAs are produced from fold-back structures of the positive strand of the viral RNAs [27,30,31,75]. Hot spots of vsiRNA generation were observed (S4 Fig) as earlier [31,75].
The polarity analysis of p19-immunoprecipitated vsiRNAs revealed a more equilibrated positive/negative strand ratio (significant enrichment in negative strand derived vsiRNAs with 65% positive, 35% negative strands) in the CymRSV-infected plants (Fig 3B). In Cym19stop-infected p19syn plants 79% of vsiRNAs produced were positive-stranded (21% negative), while p19syn—immunoprecipitated the ratio changed to 62% positive, 38% negative (Fig 3B). Based on these we conclude that p19 preferentially enriches positive-negative ds-vsiRNA pairs possessing perfect duplex structure.
To formally test the impact of mismatches within the duplex sRNAs on the affinity of p19 we compared the affinity of p19 protein towards the miR171 duplex miRNA family (Arabidopsis miR171a, miR171b, miR171c all containing mismatches) and a perfect artificial siR171 duplex (for structures see Fig 4A–4D) using in vitro electro-mobility shift assay [57]. The presence of mismatches within the stem of sRNAs strongly reduced p19 binding affinity towards duplex sRNAs (Fig 4A–4F). Consistent with our findings, it has been also shown previously that p19 preferentially binds to perfect sRNAs duplexes but not imperfect miRNAs duplexes [76].
We have also analyzed the 5’-nucleotide preference of p19 binding. No 5’-nucleotide sorting of vsiRNAs in the p19 complex was observed regardless of p19 expressional context (from the virus or the transgene) (Fig 2C and 2E). The relative abundance of sRNAs possessing different 5’-nucleotides closely followed the ratio of the input samples. The distribution of p19 bound vsiRNAs along the viral genome was found to be similar to that in the input (S4 Fig) showing that there is no sequence preference in p19 vsiRNA binding.
To better understand the p19 protein effects on vsiRNAs we analyzed the size distribution of these during infection. Upon CymRSV infection vsiRNAs produced are predominantly of 21nt and 22nt in length (Fig 3C). In addition to these, the 20nt long vsiRNAs are still present although at much lower level. In Cym19stop virus-infected wt plants we observed a shift towards slightly longer forms: most of the vsiRNAs were 22nt long, the abundance of the 21nt and 20nt long forms being reduced (Fig 3C). These results are in line with our previous findings [31]. miRNAs having enhanced electrophoretic mobility were also detected earlier in the presence of p19 [62,77].
To test if shortening is indeed an effect of p19 protein itself we analyzed parental virus- (Cym19stop) derived vsiRNA when p19 was provided “in trans“. The length shift to 1- or 2-nucleotide shorter vsiRNA forms was confirmed (Fig 3A–3C). Shortening of endogenous miRNAs was also observed in the absence of virus infection (Fig 3D). Analysis of selected endogenous miRNAs, where the precise sequence and biogenesis/maturation are known, allowed us to establish that the truncation occurred at the 3’ end but not 5’ end. The truncation of miRNAs happened mainly in p19-sRNA complexes as was observed in p19-IP, however not all p19 bound miRNAs are truncated and the reason for this has not been clarified yet (Fig 3D). Nuclease treatments on in vitro bound p19:siRNA complexes further confirmed that the p19 protein can protect the double-stranded stretch of sRNA duplexes (S5 Fig). The exonuclease (RNaseA)-mediated digestion occurred in discrete 1- and 2-nucleotide steps while the dsRNA region (19nt length) was protected by the p19 protein. Shortening of sequestered sRNAs, therefore, is not dependent on the virus infection, occurs on 3’ end, involves both vsiRNAs and miRNAs and is likely the direct consequence of p19-binding and exonuclease activity.
Multiple AGOs were shown to have antiviral functions. In A. thaliana and N. benthamiana AGO1 and AGO2 were described as the most important effectors while others such as AGO5, 7 and 10 to have minor roles during antiviral silencing [13,34,36,44]. The current model of the inhibitory effect of p19 suggests that sequestering vsiRNAs prevents AGO loading. The inhibitory effect of p19 protein on RNA silencing during infection was quite evident. In fact, wt virus infection showed strong viral symptoms that culminated in complete necrosis and collapse of the plants while the Cym19stop-infected wt plants recovered from viral infection [70], the virus accumulation was restricted to the vascular tissues and a few cell layers around the veins ([70] and S2 Fig).
To get a better insight into the detailed mechanism of p19 actions we analyzed the AGO1- and AGO2-bound vsiRNAs in CymRSV- and Cym19stop-infected wt N. benthamiana by co-immunoprecipitations (Fig 5) followed by deep sequencing analysis (Fig 6 and S6–S10 Figs). Loading of siRNAs into a particular AGO is preferentially directed by their 5’-terminal nucleotide: AGO1 prefers sRNAs having 5’U while AGO2 preferentially binds 5’A sRNAs [78,79]. As expected, the AGO1 co-immunoprecipitated plant sRNAs possessed predominantly 5’U while AGO2 immunoprecipitated sRNAs mainly 5’A, with a relatively high amount of 5’U species (Fig 6). We have also found 5’U endogenous sRNA binding by AGO2 when we processed the raw data obtained from previous report [36].
We expected that p19 would drastically reduce the loading of vsiRNAs into AGO1 and AGO2. Surprisingly, vsiRNA loading into AGO1 was compromised in the presence of p19 (during CymRSV-infection compared to Cym19stop-infection): we observed relatively high “background” of vsiRNAs without 5’ sorting preference in AGO1 (compare Fig 6D with 6F). Conversely, the amount of vsiRNAs and their 5’ sorting into AGO2 was very similar during CymRSV- and Cym19stop- infections (Fig 6D–6F). This suggests that the presence of p19 preferentially impact vsiRNAs’ AGO1 but not AGO2 effector loading. In the same time (the same sample set) endogenous sRNAs were efficiently precipitated as 5'U by AGO1-IP proving that the IP worked correctly (Fig 6C–6E). Note that the reads of endogenous sRNAs in CymRSV and Cym19stop are lower compared to mock-infected sample due to the high amount of vsiRNA presence (that impacts the bias during deep sequencing).
The AGO1 IP derived from CymRSV-infected plants contained a similar miRNA profile as the mock inoculated plants, in contrast to AGO2 IP in which the levels of analyzed miRNAs were reduced (Fig 6C and 6D and S6 and S7 Figs). Importantly, this occurred only in wt CymRSV infection when a high level of p19 is expressed. Efficient incorporation of vsiRNAs into AGO2 but not AGO1 may cause the out-titration of AGO2-bound endogenous sRNA species (during CymRSV infection). In contrast, during Cym19stop-infection AGO1-loading occurred as expected predominantly by 5’U-sorting of vsiRNA and endogenous sRNAs (Fig 6E and 6F). The obtained results were confirmed with a second AGO1 and AGO2 IP that gave very similar results although had slightly higher background of contaminating 24nt species (S9 Fig). These findings suggest that p19 protein itself compromises AGO1- but not AGO2-loading during viral infection.
The specific impact of p19 on vsiRNA AGO1-loading found in the deep sequencing analysis was also confirmed by Northern blot analysis. vsiRNAs loading into AGO1 was less efficient than AGO2 in CymRSV-infected plants when the p19 was provided in “cis” (Fig 5A) or in “trans” when two independent p19syn lines were infected with Cym19stop virus (Fig 5B).
We also analyzed the distribution of AGO1- and AGO2-bound vsiRNA along the viral genome. This generally followed the biogenesis of vsiRNAs and we could not define any sequence preference or specific hotspots of AGO1- or AGO2-loading (S8 Fig). The strong spikes of certain vsiRNAs may arise due to the sequencing bias, therefore, do not necessarily represent vsiRNAs preferred for binding [80].
Most of the identified VSRs are multifunctional, and this nature of VSRs often causes serious difficulties in the exploration of their action during the natural virus infection. The inactivation of VSRs often leads to loss of viability of the given virus, due to their multifunctional nature. It is hard therefore to separate the suppressor’s impact and virus infection effects. Either transgenic or transient expression of VSRs without the parental virus background might not reflect the natural interaction with the host cellular machinery. As a consequence of these difficulties, the molecular mechanisms of the action of VSRs and their impact on the host often remain elusive. To overcome this pitfall, we combined the transgenic expression of the VSR (p19) suppressor with its authentic suppressor mutant virus (Cym19stop)-infection in a novel experimental setup.
During virus infection, high amounts of vsiRNAs are produced. These vsiRNAs are efficiently sequestered by p19 suppressor inhibiting their incorporation into RISC [67]. The consequence of p19 vsiRNA binding is that the strong positive strand bias of vsiRNA biogenesis in the input sample (95:5 positive/negative) is changed to a more equilibrated positive/negative stand ratio (65:35 positive/negative). This result suggests that there are qualitative structural differences between vsiRNAs and that p19 preferentially binds vsiRNAs derived from perfect dsRNA or highly structured RNA species. The preference of p19 towards even structured ds-vsiRNAs is in agreement with p19 crystal structure: p19 homodimer leans on the ds-sRNA backbone. If the backbone structure is distorted by mismatches, the sRNA could become less accessible to p19 sequestration. Indeed, p19 bound siR171a perfect duplex with higher affinity than natural miR171a, miR171b or miR171c duplexes (Fig 4). This may likely be one of the reasons why ds-vsiRNAs are preferred by p19 instead of mismatches containing endogenous miRNA duplexes during viral infection (Figs 2 and 3A). The excess of vsiRNAs over endogenous miRNAs in virus-infected plant may also contribute to the preferential binding of vsiRNAs by p19. Moreover, the difference in the biogenesis of miRNAs versus vsiRNAs could also be a further important factor in the mechanism of the sRNA sequestration by p19.
In addition to the previous findings, we have also shown that sRNA binding by p19 happens without 5’-end nucleotide selection including vsiRNAs or endogenous miRNAs.
Previous studies [62,77] have reported the truncation of p19-bound sRNAs by 1 or 2 nucleotides. In the case of vsiRNAs, the site of truncation (5’ vs. 3’ end) cannot be defined since the generated vsiRNAs started from almost every single nucleotide of viral genome (S4 and S8 Figs). During Cym19stop-infection on wild-type plants truncation does not occur while in p19syn plants, which provide p19 in trans, it can be observed. Shortening also happens on miRNAs in p19syn plants without virus infection. In summary, the truncation is likely induced directly by p19-binding and is not due to the virus infection or restricted to a specific class of sRNAs. Why shortening does not happen in the absence of p19, on the free vsiRNAs, which theoretically would be more accessible? In cells endogenous free sRNA duplexes (e.g., miR168/miR168star) [54] and free vsiRNAs in Cym19stop infection can be observed [32,67]. The stability of these sRNAs (miRNAs and vsiRNAs) is conferred by HEN1-mediated methylation [11]. The crystal structure of p19:siRNA complex shows that the last two single-stranded nucleotides at 3’ terminus of siRNAs are protruding from the complex [57,58]. Furthermore, the p19 bound vsiRNAs are not methylated at 3’ terminus [65] therefore may be sensitive to exonucleases. We propose therefore that p19 binding inhibits sRNAs methylation and as a consequence of this the protruding unprotected two nucleotides at the 3’-end of sRNAs are trimmed by cellular 3’-exonucleases. Whether the truncation of sRNAs is a simple byproduct of binding or has a definite biological importance remains to be seen. Trimming of sRNAs may inactivate and render them incompetent for AGO-loading. Contradictory to this, we find efficient binding of 19nt and 20nt vsiRNAs by AGO2 (Fig 6D, S9D Fig). This observation supports the “catch and release” of vsiRNAs by p19 proposed earlier [81].
It has been long suggested that VSRs interfere with endogenous silencing pathways, and this may contribute to the viral symptom development [1,61–64]. Constitutive expression of p19 in N. benthamiana leads to the development of a strong phenotype that is quite different from symptoms observed during parental viral infection (Fig 1 and S2 Fig). The strong phenotype of p19syn plants may arise, at least in part, due to the sequestration of endogenous sRNAs by p19. Indeed, we could immunoprecipitate endogenous miRNAs with p19 from transgenic plants (Fig 2B). Importantly, however, miRNA sequestration by p19 provided either in cis or in trans was drastically reduced when the virus was present (Fig 2C–2F). Importantly, miRNA out-competition/release correlated with downregulation of miRNA targets (S3B Fig) and reestablishment of tasiRNA biogenesis (S3C Fig). Out-competition/release of p19-bound endogenous sRNAs/miRNAs upon virus infection seems to be biologically relevant and could have an important role in moderating the virus impact on plant. This needs to be further investigated. In conclusion therefore, our findings deny the model in which miRNA binding by p19 is the key step for the development of virus-induced symptoms [1,61–64]. It is more likely that p19 has an indirect effect through the specific inhibition of antiviral plant response and the viral symptoms are the outcome of a complex virus-host interaction during the viral invasion of plant cells.
What could be the criteria for vsiRNA selection by AGO-loading machinery? We previously observed that in wt tombusvirus infection, p19 protein prevents vsiRNA loading to AGO/RISC complexes, however, even in the absence of p19 only a small fraction of the vsiRNAs is loaded into effector complexes (Fig 5) the rest remaining in a free, probably double-stranded form [32,67]. This suggests that a big part of the abundantly produced vsiRNAs is AGO-incompetent, or there is no free AGO protein present to be loaded into. The structure of the ds-sRNA stem could be an important feature for vsiRNA selection into AGOs (as we have shown for p19). A similar analysis of sRNA duplexes as in the case of p19 cannot be done, since in p19 binding both strands of ds-sRNAs are retained, while in AGOs, after loading, one strand is eliminated. The other possibility of the inability of vsiRNAs to load into effector complexes could be the shortage of silencing proteins like DCL/DRB or AGOs during the assembly of these effectors. It was shown that specific regulatory mechanisms are induced by the virus to dampen silencing: translation of AGO1 protein is decreased by the suppressor-mediated miR168 over-accumulation [54,66]. However, the down-regulation of AGO1 protein was not observed in p19syn plants (Fig 1B). The reason for this could be the relatively low level of p19 produced from transgene compared to virus infection (Fig 5).
Another important observation is that vsiRNAs loading is selectively prevented mainly into AGO1 but not AGO2 in the presence of p19 in both CymRSV infected wt or Cym19stop infected p19syn plants (Fig 5). During virus infection, the decrease in the translation of AGO1 protein leads to accumulation of AGO2, due to the absence of AGO1-miR403-mediated posttranscriptional down-regulation of AGO2 [54,66]. One possibility, therefore, is that the vsiRNAs will be loaded into the available AGO2 while AGO1-loading will be decreased. We could not observe a significant increase of AGO2 protein in the presence of p19 (Fig 1B). The other possibility is that AGO1 and p19 compete for the same set of vsiRNAs while AGO2 requirement for vsiRNA features is different. p19 therefore, would selectively impact AGO1 but not AGO2-loading.
Interestingly, we observed relatively high p19-depending “background” of vsiRNAs without 5’ sorting in AGO1 IP, unlikely to be AGO1-incorporated sRNAs. p19 could affect the connection between biogenesis/loading complexes DCLs/DRBs with AGO1 effector [7,9]. In line with this hypothesis, it was demonstrated that p19 can compromise the transfer of siRNA from DICER-R2D2 into RISC complex using Drosophila embryo extracts based in vitro system [59].
Regardless of the reason of how AGO1-loading is compromised by p19, it seems that AGO2 is not enough to fight off the virus and help the plant to recover in the absence of AGO1-loading/activity. It has been suggested previously that AGO2 but not AGO1 plays role in the antiviral response against tombusvirus infections, including Tomato bushy stunt virus (TBSV) [47,48]. We have done TBSV-VIGS (Virus Induced Gene Silencing) experiment, using p19 inactivated virus vector (TBSVp19stop), which carried Nb-PDS and Nb-AGO1 sequence (S11 Fig). When NbAGO1 was silenced by VIGS the virus accumulated at higher level and plants have shown stronger phenotype (S11 Fig). The obtained results further support the idea that AGO1 has a major role in antiviral response against tombusvirus infection. However, the additional role of other plant AGOs in antiviral response remained to be explored and it likely depends on specific features of the highly diverse plant viruses. The availability of CRISPR/Cas9 system for plant research will also help to clarify the specific roles of plant effectors in antiviral silencing response.
The synthetic CymRSV ORF5 (p19) was essentially constructed following the previously described antivirus-induced transgene silencing strategy [82]. As the first step, we introduced, in CymRSV ORF5, all possible silent point mutations by selecting those most compatible with the N. tabacum codon usage. The resulting nucleotide sequence was further modified to avoid the presence of cryptic splicing and polyadenylation signals using Net2gene splicing prediction (http://www.cbs.dtu.dk/services/NetGene2/) and HCpolyA (http://bioinfo4.itb.cnr.it/~webgene/wwwHC_polya_ex.html) software, respectively. The synthetic ORF5 (S1A Fig) was synthesized by Life Technologies and cloned in pJIT61 [83] between the CaMV 35S promoter and 35S terminator. The gene cassette was excised with KpnI and BglII and cloned in KpnI-BamHI of pBin19 (pBinCymRSVp19syn)
N. benthamiana was transformed with the recombinant Agrobacterium tumefaciens strain C58C1 (pGV2260) harboring the plasmid pBinCymRSVp19syn, and kanamycin-resistant plants were regenerated as previously described [83]. The primary transformants were checked for the presence of p19syn transgene by PCR and for the expression of the p19 protein by Western blotting with the anti-CymRSV-p19 antibody as previously described [56].
N. benthamiana plants were grown at 22°C. At six-leaves stage plants were infiltrated with A. tumefaciens C58C1 harboring the appropriate constructs in the pBIN61 plasmid. pBIN61-Cymp19 and pBIN61-GFP were grown on selective media overnight, resuspended in the infiltration buffer (10 mM MES, 0.15 mM acetosyringone, 10 mM MgCl2) kept on 25°C for 4h, and subsequently infiltrated into wild-type or p19syn plant leaves at OD600 = 0.4.
In vitro transcription of CymRSV, Cym19stop, TBSV-PDS-GFP and TBSV-PDS-AGO1-1 RNAs from linearized template plasmids and inoculation of RNA transcripts onto N. benthamiana plants were performed as described previously [84]. CMV Y-sat infection was performed as described earlier [69].
Total RNA was extracted from 100 mg of leaf tissue. The homogenized plant materials were resuspended in 600 μl of extraction buffer (0.1 M glycine-NaOH, pH 9.0, 100 mM NaCl, 10 mM EDTA, 2% SDS) and mixed with an equal volume of phenol. The aqueous phase was treated with equal volumes of phenol-chloroform and chloroform, precipitated with ethanol, and finally resuspended in sterile water. RNA gel blot analysis of higher molecular weight RNAs was performed as described previously [84].
RNA gel blot analysis of 21–24 nt RNAs was performed as follows. Approximately 5 μg of total RNA was separated by 15% PAGE with 8.6 M urea and 1xTris-borate-EDTA. RNA was electro-blotted onto Hybond-NX membranes and fixed by chemical crosslinking at 60°C for 1 hr [85]. Small RNA Northern blot hybridization and analysis were performed using complementary locked nucleic acid (LNA) oligonucleotides (Exiqon, http://www.exiqon.com).
Mock- or virus-infected systemical leaf tissues were homogenized in extraction buffer (150 mM Tris-HCl, pH 7.5, 6 M urea, 2% SDS, and 5% μ-mercaptoethanol). Samples were boiled, and cell debris was removed by centrifugation at 18,000 x g at 4°C for 10 min. The supernatants were resolved on 12% SDS-PAGE, transferred to Hybond PVDF membranes (GE Healthcare) and subjected to Western blot analysis. For detection anti-p19 [70], NbAGO1 [86] and NbAGO2 custom antibody were used. NbAGO2 antibody was generated by immunization of rabbits with the synthetic peptide (CLEDPEGKDPPRDVF)(GenScript, http://www.genscript.com/). The proteins were visualized by chemiluminescence (ECL kit; GE Healthcare) according to the manufacturer’s instructions.
For immunoprecipitation, 1–5 grams of mock-, CymRSV- or Cym19stop-infected N. benthamiana leaves showing systemic symptoms (or leaves at the same stage and positions from mock-inoculated plants) were collected, ground in 1:3 (w/v) amount of immunoprecipitation buffer (40 mM HEPES/KOH 7.4, 100 mM KOAc, 5 mM MgOAc, 5% glycerol, freshly added 4 mM DTT), and cleared by centrifugation (twice at 15,000 x g for 5 min). Cleared lysates were kept on ice until immunoprecipitation with antibody-coated protein A-Sepharose (GE Healthcare). Beads were washed before adding the antibodies (described earlier). For mock immunoprecipitation preimmune serum was used. Antibodies coated beads were incubated with the relevant cleared lysates for 4h at 4°C. After immunoprecipitations the beads were washed five times with ice-cold immunoprecipitation buffer for 2 min each. Input extracts and eluates of immunoprecipitations were used for Western and Northern blot analysis.
The library preparation was described previously [31], shortly: RNA samples were purified by cutting the sRNA region from 8% denaturing polyacrylamide gels (acrylamide:bisacrylamide (19:1) 1xTBE, 8.6 M urea). After gel electrophoresis, the gels were stained by SYBRGold (Thermo Fisher Scientific). Bands at the small RNA range were cut out and crushed. Gel particles were shaken overnight in RNase free water at 4°C, followed by RNA isolation (described above). TruSeq Small RNA Sample preparation kit (Illumina) was used for library preparation; we followed the manufacturer’s protocol. In the case of the AGO1- and AGO2- immunoprecipitation 9 libraries were pooled together. In the case of the p19 immunoprecipitations, 4–4 libraries were pooled together. The libraries were sequenced on Illumina HiScanSQ platform (UD-GenomMed Medical Genomic Technologies Ltd., Debrecen, Hungary) that yielded approximately 100 M reads per lane (50bp, single end) (S10 Fig).
RNA-seq library preparation was done according to the manufacturer protocol (TruSeqStranded mRNA Library Prep Kit). The libraries were sequenced on Illumina HiScanSQ. 135 M 100 bp paired-end reads were produced per lane. 3 samples were pooled together on a lane. The libraries were submitted to GEO and can be accessed through series accession number GSE77070.
One hundred nanograms of synthetic sRNA (5’-UGAUUGAGCCGCGCCAAUAUC-3’) was 5’ end labeled by T4 polynucleotide kinase (Fermentas) with 32P isotope. After the reaction was stopped, 10 ng was saved for further process and the rest was mixed with 500 ng of unlabeled synthetic RNA with the sequence of 5’-UAUUGGCGCGGCUCAAUCAGA-3’. The mixture was heated to 95°C for 2 minutes in a thermocycler and was cooled to 5°C (2°/2 min) to gain 19 nt perfectly matched double strand with 2 nt protruding at the 3’ end. The sufficient amount of DNA loading dye was added to the mixture and to the saved labeled single stranded RNA. Both samples were run on a 8% acrylamide:bis-acrylamide 19:1 1xTBE gel. Gel was directly exposed. The dsRNA region was cut from the gel. The gel piece was shredded using a 0.5 ml tube with several holes in the bottom in a 2 ml tube via centrifugation. The shredded gel pieces were shaken overnight in 500 μl of 300 mM NaCl at 4°C. Gel pieces were filtered out by using Spin-X column (Corning). 400 μl of dsRNA solution was precipitated by adding 20 μg glycogen (Fermentas) and 1 ml ethanol. The precipitated ds-RNA was resolved in IP buffer described before.
Purified p19 described earlier was used for the assay [56]. A dilution series of p19 (~1 μg) was made in 1x IP buffer. An equal amount of gel-purified ds-siRNA was added to each p19 dilution and incubated at room temperature for 10 min. Then 10 ng of RNase-A (Sigma) was added and incubated for 10 and 30 min at room temperature. After incubation the samples were mixed with DNA dye and ran on a 16% acrylamide: bis-acrylamide (19:1) 1xTBE gel. Decade marker (Ambion) and synthetic RNAs were used as size markers. Gels were dried and were directly exposed.
For band shift assays wild-type p19 protein was purified from E. coli as described previously [56,67]. Custom RNAs used were ordered from Dharmacon, (http://dharmacon.gelifesciences.com) for sequence see Fig 4. Labeling and annealing of si/miRNA duplexes was carried out as described previously [67]. Purified p19 protein and labeled si/miRNAs were incubated for 30 min at room temperature in band-shift buffer [67]. Complexes were resolved on 8% polyacrylamide 0.5xTBE gels. Gels were dried and exposed to a storage phosphor screen and bands quantified (Molecular Dynamics Typhoon Phosphorimager, GE Heathcare).
In situ hybridization was performed as previously described in [87]. Detection of viral RNA expression patterns were made by using nonradioactive in situ hybridization on histological sections of leaf tissues. Digoxigenin labeled antisense RNA probe was synthesized by in vitro transcription from the linearized CymRSV construct.
After demultiplexing of the raw data we used the UEA workbench version 3.0 [88] for adapter removal. Quality control consisted of filtering out reads with less than 14 nt (without the adapter sequences) and reads showing low complexity. We used PatMaN v1.2.2. [89] to align reads allowing 0 mismatches. Reads not matching either genome were removed. Reads passing quality control is referred as “total” in this article.
In Fig 3B reads were normalized to 1 million viral reads. In S7 Fig reads were normalized to 1 million N.benthamiana genome matching reads. In all other cases reads were normalized to 1 million total reads.
After demultiplexing we used FastQC 0.10.1 to check general attributes. Trim_galore 0.4.1 and FASTX Toolkit 0.0.13 were used to remove adaptor sequences, low quality bases, reads under 20 nt and unpaired reads. Bowtie2 [90] was used to align reads to Nbv5 [91] transcriptome database. Reads were counted for homologs of known miRNA targets. NCBI-blast+ 2.2.28 [92] was used to validate miRNA targets. Samtools 0.1.19-96b5f2294a was used during alignment evaluation. Read counts were normalised to 1 million total reads.
|
10.1371/journal.ppat.1000323 | A Discontinuous RNA Platform Mediates RNA Virus Replication: Building an Integrated Model for RNA–based Regulation of Viral Processes | Plus-strand RNA viruses contain RNA elements within their genomes that mediate a variety of fundamental viral processes. The traditional view of these elements is that of local RNA structures. This perspective, however, is changing due to increasing discoveries of functional viral RNA elements that are formed by long-range RNA–RNA interactions, often spanning thousands of nucleotides. The plus-strand RNA genomes of tombusviruses exemplify this concept by possessing different long-range RNA–RNA interactions that regulate both viral translation and transcription. Here we report that a third fundamental tombusvirus process, viral genome replication, requires a long-range RNA–based interaction spanning ∼3000 nts. In vivo and in vitro analyses suggest that the discontinuous RNA platform formed by the interaction facilitates efficient assembly of the viral RNA replicase. This finding has allowed us to build an integrated model for the role of global RNA structure in regulating the reproduction of a eukaryotic RNA virus, and the insights gained have extended our understanding of the multifunctional nature of viral RNA genomes.
| Plus-strand (i.e. messenger-sensed) RNA viruses are responsible for significant diseases in plants and animals. The single-stranded RNA genomes of these viruses serve as templates for translation of viral proteins and perform other essential functions that generally involve local RNA structures, such as RNA hairpins. Interestingly, plant tombusviruses utilize a number of long-range intra-genomic RNA–RNA interactions to regulate important events during infection of their hosts, i.e. viral translation and transcription. Here, we report that an additional essential tombusvirus process, viral RNA replication, also requires a long-range RNA–RNA interaction. Our analyses indicate a role for this RNA–based interaction in the assembly of the viral replicase, which is responsible for executing viral RNA synthesis. This information was used to generate a comprehensive higher-order RNA structural model for functional long-range interactions in the genome of this eukaryotic RNA virus. The model highlights a critical role for global RNA structure in multiple viral processes that are necessary for successful infection of hosts.
| Plus-strand RNA viruses infect a wide variety of organisms and are responsible for causing significant diseases in plants, animals, and humans. A key step in the reproduction of these pathogens is replication of their single-stranded RNA genomes. This process occurs in the cytosol of host cells in association with membranes and requires a virally-encoded RNA-dependent RNA polymerase (RdRp) [1]. During infections, the RdRp associates with other viral and host proteins to form the RNA replicase, which is the complex responsible for synthesizing negative-strand RNA intermediates and progeny viral genomes [2]. The mechanism by which replicase assembly occurs is fundamental to understanding genome replication and is currently the focus of intense study [3].
Besides being templates for replication, plus-strand RNA genomes also serve additional functions during infections, including acting as templates for (i) translation of viral proteins, (ii) transcription of viral mRNAs, and (iii) assembly of virus particles. Accordingly, these RNA genomes are multifunctional molecules that possess regulatory mechanisms to ensure that these different processes occur accurately and at the proper time during the infectious cycle. Integral to this control is the presence of different regulatory RNA elements within viral genomes that act as signals for modulating molecular events. Traditionally, such RNA elements have been viewed as localized sequences or structures (e.g. RNA hairpins); however, this perspective is rapidly changing due to increasing discoveries of functional viral RNA elements that are formed by long-range RNA–RNA interactions spanning significant distances [4]–[9]. Consequently, our structural concept of a functional viral RNA genome is shifting from that of a “linear” molecule to one that is three-dimensional [4].
Tombusviruses, such as Tomato bushy stunt virus (TBSV), have been excellent model systems for understanding molecular aspects of virus reproduction and, particularly, the role of both local and long-range RNA elements in regulating and coordinating the multiple processes that occur during viral infections [10]. The plus-strand RNA genome of TBSV is ∼4.8 kb long and encodes five functional proteins [11]. The RNA replication-related proteins, p33 and its read-through product p92, are translated directly from the genome, while 3′-proximally encoded open reading frames (ORFs) are translated from two subgenomic (sg) mRNAs that are transcribed during infections (Figure 1A) [10]. Interestingly, translation of p33 and p92 occurs via a 5′ cap- and 3′ poly(A) tail-independent mechanism that involves a long-distance RNA–RNA interaction between a 3′ cap-independent translational enhancer (3′CITE) in the 3′-untranslated region (UTR) and the 5′UTR of the genome [12],[13]. Sg mRNA transcription also requires long-range RNA–based interactions within the TBSV genome that involve sequences immediately upstream from transcriptional start sites and partner sequences far upstream [14]–[16]. Accordingly, TBSV utilizes RNA–RNA interactions spanning thousands of nucleotides in two different essential processes: translation and transcription.
Viral RNA replication in TBSV has been studied extensively in both plant and yeast cells and the latter system has served as a genetically-tractable surrogate host that supports authentic viral RNA replication [2],[10],[17]. Two TBSV proteins are required for viral RNA replication: p92, the RdRp, and p33, an auxiliary protein that plays multiple critical roles [10],[18]. Both of these viral proteins are part of the RNA replicase, and several host proteins have also been determined to be components of this complex [19]–[21]. p33 contains peroxisomal targeting signals and transmembrane segments in its N-terminus [22]. This protein also binds, as a dimer, to an internal RNA element in the TBSV genome termed region II (RII; Figure 1) and recruits the genome to peroxisomal membranes, where active RNA replicase is formed [23],[24]. Only the central portion of RII, an extended stem-loop (SL) structure called RII(+)-SL, is required for p33-dimer binding (Figure 2A). p33 also interacts with p92, thus p92 is recruited into replicase assembly by associating with the p33 dimer bound to RII [25]. Interestingly, in addition to the internally-located RII segment, replicase assembly also requires a 3′-terminal section of the viral genome, termed RIV (Figure 1) [26]. RIV is composed of a series of three RNA SLs, which form a compact structure that is located immediately downstream of the 3′CITE (Figure 2A) [27],[28]. Consequently, two non-contiguous RNA elements in the TBSV genome (RII and RIV), separated by ∼3,000 nts, are required for viral replicase assembly [26].
In this report we have pursued the hypothesis that the distantly located RII and RIV require some form of communication to facilitate replicase assembly. Our results support this theory by identifying a long-range RNA–RNA interaction in the TBSV genome that unites RII and RIV and is essential for efficient replicase assembly and viral RNA replication. This finding, along with previous results, has allowed us to generate an integrated higher-order RNA structural model for functional long-range interactions in the genome of a eukaryotic RNA virus. Mechanistic and evolutionary insights provided by this model are discussed.
Based on the demonstrated requirement for both RII and RIV for replicase assembly in vivo [26], we hypothesized that these two discontinuous RNA elements need to communicate with each other in order to coordinate this event. Considering the abundance of existing long-range RNA–RNA interactions in other fundamental tombusvirus processes [4], we entertained the possibility that the RII-RIV communication might also be RNA-mediated. To this end, all sequenced tombusvirus genomes were analyzed by the RNA secondary structure-predicting program mfold [29],[30] in an attempt to identify candidate RNA–RNA interactions. This analysis revealed a potential RNA base pairing interaction involving two 11 nt long sequences, one located in RII and the other in RIII. The sequence in RII was located just 3′ to the essential RII(+)-SL core structure and was termed upstream linker (UL), while its complementary partner sequence in RIII was termed the downstream linker (DL) (Figure 2A). Although the DL in RIII is located some 267 nts away from RIV in the linear RNA sequence, formation of the intervening and experimentally-confirmed Y-shaped domain (i.e. R3.5, which includes the 3′CITE) would position the putative interaction close to RIV in the higher-order RNA structure (Figure 2A) [12],[13],[28]. Thus, the UL–DL interaction would situate RII(+)-SL immediately adjacent to RIV, thereby allowing for communication between these two distal RNA elements. It should be noted that the UL–DL interaction is not a direct interaction between RII and RIV; instead, it could function as an RNA–based bridge that juxtaposes the two RNA elements. The functional relevance of the proposed UL–DL interaction was further supported by comparative sequence analysis of tombusvirus genomes that revealed mono- and co-variations in the two sequences that would either maintain or not significantly disrupt the base pairing interaction (Figure 2B). Accordingly, both thermodynamically-based analysis of secondary structures of full-length viral genomes and comparative sequence analysis of the proposed paired segments support the formation of the UL–DL interaction.
To assess whether the UL–DL interaction was functionally relevant to viral infections, the partner sequences were subjected to compensatory mutational analysis in the context of the TBSV genome. In the genomic mutant T-dU, substitutions were introduced into the UL at wobble positions within the p92 ORF so as not to alter the amino acid sequence of the encoded p92 protein. The changes introduced into T-dU, which were predicted to destabilize the interaction (Figure 3A), reduced genome accumulation levels in plant protoplast infections to ∼6% that of wt TBSV (Figure 3B, top panel). Similarly, disruptive substitutions in the DL in mutant T-dD reduced genome accumulation levels to ∼20% that of wt TBSV (Figure 3B). However, when the two sets of disruptive substitutions were combined in mutant T-cUD, so as to restore base pairing potential, genome accumulation levels showed recovery to ∼69% that of wt TBSV (Figure 3B). These results indicate that the UL–DL interaction is functionally important for efficient TBSV genome accumulation, as well as for robust levels of both sg mRNAs. Additional analysis of these mutants for minus-strand viral RNA accumulation indicated that the interaction is required at an early step in genome and sg mRNA production, because the disruptions inhibited minus-strand synthesis for both of these classes of viral RNA (Figure 3B, lower panel). The findings with TBSV were confirmed by carrying out the same compensatory mutational analysis in the UL–DL interaction in another tombusvirus, Carnation Italian ringspot virus (CIRV) (Figure 3C). It was also noted, for both viruses, that the dD mutants containing two central UG wobble pairs were ∼three- to fourfold more active than the corresponding dU mutants with CA mismatches (Figure 3B, C, upper panels). This difference suggests that the UL–DL interactions are functionally relevant in the plus-strands of these viral genomes (refer to Figure 3A legend for additional information).
The above results indicated a defect at an early step in viral RNA synthesis. One possible explanation for this is that the UL–DL interaction acts to facilitate translation of the RNA replication proteins p33 and p92 from the viral genome. Inhibition of this function would lead to lower levels of viral RNA replication and reduced RNA genome accumulation. Notably, the UL–DL interaction is positioned directly adjacent to the 3′CITE, a location that could potentially aid it in regulating 3′CITE activity (Figure 2A). To address this possibility, the same set of genomic mutants that was analyzed in Figure 3C was assessed in a wheat germ translation extract. CIRV was used for this analysis as, unlike TBSV, its 3′CITE is fully active in this plant-derived in vitro system [31], as illustrated by the significant decrease in translation observed in its absence (i.e. mutant CIRVΔTE in Figure 4A). In the translation assay, the level of p36, the homologue of p33 in TBSV, was monitored for the wt CIRV genome and each of its mutants. In general, the viral genomes yielded similar levels of p36, suggesting roughly equivalent efficiencies of translation (Figure 4A). This conclusion was supported by stability analysis of these messages, which showed comparable profiles of RNA decay (Figure 4B). Taken together, these data indicate that the UL–DL interaction does not markedly affect translation of viral proteins.
The lack of involvement of the UL–DL interaction in translation prompted us to refocus our attention to its possible role in directly mediating viral RNA replication. As proposed earlier, the interaction could function simply to juxtapose RII and RIV. However, the potential exists that the RNA helix formed by the UL–DL interaction also contributes to the activity. To investigate this possibility, two mutant TBSV genomes were generated. The first mutant, CorP1, contained the UL and DL sequences, but lacked the intervening sequence between these two partner segments (Figure 5A). To facilitate formation of the UL–DL interaction, the intervening sequence deleted in CorP1 was replaced by a stable GNRA-type tetraloop. In the second mutant, CorP2, the UL and DL sequences, along with their intervening sequence, were removed (Figure 5A). Both CorP mutants were defective for autonomous replication, because they encoded C-terminally truncated p92 ORFs. Consequently, replication of these viral RNAs had to be complemented with a genomic RNA, AS1m1, which provided full-length p92 in trans. AS1m1 does not transcribe sg mRNA1, due to a mutation in its RNA–based transcriptional signal [15], and it was used as the helper genome so as to avoid obscuring the detection of the CorP RNAs, which have lengths that are similar to that of sg mRNA1. The results from co-transfection of AS1m1 with CorP1 into plant protoplasts revealed readily detectable accumulation of the latter; however, the corresponding co-transfection containing CorP2 resulted in significantly higher levels of accumulation of the CorP2 replicon (Figure 5B). The efficient accumulation of the viral replicon lacking the UL–DL interaction indicates that the RNA helix formed by the interaction is not necessary for function and may in fact be somewhat inhibitory to efficient viral RNA replication (though this effect could be related to its stable hairpin context in CorP1). Regardless, these results support the notion that the primary role of the interaction is to mediate the juxtaposition of RII and RIV.
To further investigate the role of the UL–DL interaction in viral RNA accumulation, we employed the use of a small, non-coding, efficiently replicating, TBSV-derived RNA replicon that requires co-infection with helper genome for its replication [10]. Replicon DI-73 was selected for these studies as it contains both RII and a contiguous 3′ end containing RIII-R3.5-RIV, thereby providing a potentially suitable context to study the UL–DL interaction (Figure 1B) [32]. To validate that the UL–DL interaction was functionally important for its accumulation, the same set of compensatory mutations that was tested in the TBSV genome (Figure 3A) was introduced into DI-73. As observed for the full-length TBSV genome, the disruptive mutations (mutants 73dU and 73dD) led to substantially reduced DI-73 levels of accumulation in plant protoplast co-transfections with helper genome, while restoration of the interaction (mutant 73cUD) led to notably recovered levels of accumulation, both at the plus- and minus-strand levels (Figure 6A and 6B). These results show that even though the UL and DL in DI-73 are separated by only 70 nts, the formation of the interaction is still required. This indicates that the UL–DL interaction is needed to mediate a specific RNA arrangement that is required for optimal function. To determine if the differences in replicon accumulation were related to changes in RNA stability, stability assays were carried out (Figure 6C). No major differences in RNA decay rates were observed, suggesting that the UL–DL-related defect was related to RNA replication efficiency.
The analyses of both genomic and DI-73 replicon RNA accumulation indicated that the UL–DL-based defect in replication affects minus-strand synthesis (Figure 3B and Figure 6B). One possible explanation is that the interaction facilitates folding of the plus-strand viral RNA templates into conformations that are suitable for minus-strand synthesis. To test this idea, DI-73 replicons harboring the compensatory mutations in the interaction were tested in vitro for their ability to serve as templates for minus-strand synthesis using a tombusvirus replicase extract derived from plants [17]. All of the RNA templates tested directed the synthesis of terminally-initiated (ti) minus strands with similar efficiencies (Figure 6D). Internally-initiated (ii) products, commonly observed in such in vitro systems, were, for the most part, also produced at similar levels. These results indicate that the UL–DL interaction neither promotes more efficient minus-strand synthesis by assembled replicase nor acts to redirect internal initiation sites to the 3′-terminus. Accordingly, the UL–DL-related function may act at a step in the RNA replication pathway that precedes minus-strand synthesis.
A possible early step in viral RNA replication that could require the UL–DL interaction is replicase assembly. This concept relates back to our original hypothesis that RII and RIV, the only two discontinuous RNA regions required for this process, need to communicate in order to function. Importantly, the juxtaposition of RII and RIV by the UL–DL interaction would satisfy this proposed condition. To address this possibility, a well-defined tombusvirus replicase assembly assay utilizing a yeast-based system was employed [26]. Before carrying out the replicase assembly assay, the DI-73-based compensatory mutants were tested in an in vivo replication assay to determine if the results observed in plant cells could be accurately recapitulated in yeast cells [33]. This replication assay involved expressing p33, p92, and the DI-73 replicon transcript from cotransformed plasmids and monitoring the accumulation of the RNA replicon by northern blot analysis. The relative levels of accumulation of the different viral replicons in yeast were similar to those observed in plant protoplasts, thereby validating use of the yeast system for further investigation (Figure 7A). For the assembly assay, replicase was purified from yeast cells (transformed as described above) and then assayed for its ability to synthesize a complementary strand to an exogenously added viral RNA template, DI-72(−). DI-72(−) is a negative-sense viral RNA and does not contain the UL–DL interaction. Results from this assay showed that the added DI-72(−) RNA template was copied efficiently only in extracts that were isolated from cells containing DI-73 replicons that retained a functional UL–DL interaction (i.e. Y73 and Y73cUD; Figure 7B). Western blot analysis indicated that equal levels of p33 were expressed in the cells used for replicase purification (Figure 7C). Next, as a more stringent test of assembly, the wt and compensatory set of DI-73 mutants were modified at their 5′ ends so as to make them replication-defective; as verified by the replication assay (Figure 7D). Under these conditions, DI-73 levels would be similarly low for both wt and mutant forms, thus reducing the effect of replicon levels on the efficiency of replicase assembly. When the replicase assembly assay was carried out under these more strict conditions, a similar trend was observed, where assembly was more efficient when the UL–DL interaction was predicted to be stable (Figure 7E). Collectively, these data indicate that the UL–DL interaction contributes to the efficiency of replicase assembly.
We have identified a novel long-distance RNA–RNA interaction in the TBSV genome that mediates viral RNA replication by facilitating viral replicase assembly. Our results are consistent with the recognized critical roles for RII and RIV in this process [26] and they indicate that the mere presence of both of these regions in the genome is not sufficient for optimal function. Based on our findings, we propose a model in which efficient formation of the viral RNA replicase requires RII and RIV to be spatially united at some point during the assembly process (Figure 8A).
Different mechanistic variants can be envisioned for how the UL–DL interaction facilitates assembly of the replicase by forming a discontinuous RNA platform. The three schemes presented here, however, are neither exhaustive nor necessarily mutually exclusive; moreover, hybrid versions are also possible. Scheme 1: The UL–DL interaction forms first and allows proteins to associate with either RII or RIV; protein-protein interactions then mediate replicase assembly. Scheme 2: The juxtaposed RII and RIV together form a discontinuous binding site that recruits an important component(s) necessary for replicase assembly. Scheme 3: Protein factors bind to the individual non-united RII and RIV elements; the UL–DL interaction then juxtaposes the bound factors, which mediates replicase assembly. With respect to the latter scheme, studies have shown that viral p33 is able to bind in vitro to RII(+)-SL in the absence of other RNA elements [24]. This suggests that, for at least some factors, formation of the UL–DL interaction may not need to precede protein binding. Also, as p92 associates with RII by interacting with p33 [24],[25] (Figure 8A), a secondary function for the UL–DL interaction may be to reposition the internally-bound p92 RdRp close to the 3′ terminus of the RNA genome, thereby allowing it to efficiently initiate minus-strand synthesis (Figure 8A). This function, however, does not seem necessary for assembled replicase, as such complexes were able to effectively initiate minus-strand synthesis in vitro irrespective of the UL–DL interaction and the interaction did not facilitate more efficient terminal initiation (Figure 6D). Nonetheless, if replicase assembly is tightly coupled to minus-strand synthesis in vivo, it is possible that the UL–DL interaction also mediates 3′-end positioning of the p92 RdRp. Indeed, this type of repositioning mechanism has been implicated in facilitating the function of an RNA hairpin enhancer element present in the minus-strand of a small TBSV defective interfering RNA [34].
Interestingly, RdRp-positioning roles for long-distance RNA–RNA interactions have also been proposed for two different plus-strand RNA viruses that infect animal and bacterial hosts. For Dengue virus (DenV), an RNA–based interaction spanning ∼11,000 nts in its genome has been proposed to reposition the viral RdRp, bound to its 5′-proximal promoter, close to the 3′-end of the genome; thereby allowing it to efficiently initiate minus-strand synthesis [35] (Figure 8B). Similarly for Q-beta bacteriophage, the viral replicase was shown to bind to an internal site in the RNA genome that is juxtaposed to the 3′-terminus by an RNA-mediated interaction spanning ∼1000 nts [36] (Figure 8B). Long-range genomic RNA–RNA interactions involved in viral RNA replication are thus prevalent and include viruses that infect organisms from three different kingdoms (Figure 8B). Interestingly, all three of these viruses belong to the supergroup-II class of RdRps [37], which suggests that in addition to having a common polymerase ancestry, they also share a history of employing long-distance RNA–RNA interactions in their RNA replication processes. Possible selective advantages for separating codependent RNA replication elements throughout a viral genome have been reported [35], e.g. preferentially facilitating the amplification of full-length RNA templates, and such advantages may also be relevant to TBSV. However, TBSV is distinct from both DenV and Q-beta phage in that it produces 3′-coterminal sg mRNAs (Figure 1A). Notably, these sg mRNAs do not include RII in their structure, a feature that would preclude replicase formation on these messages and could act to prevent interference with their primary function of viral protein translation. Accordingly, the underlying principles behind spatially separating functional RNA elements are likely varied and are undoubtedly influenced by viral reproductive strategy.
The discovery of a long-range RNA–RNA interaction in the TBSV genome that is involved in RNA replication is significant, because it designates this virus as the first shown to use distal RNA–based communication in three different fundamental viral processes, the other two being cap-independent translation and sg mRNA transcription (Figure 9A). Importantly, this finding has allowed us to generate a comprehensive higher-order RNA structural model of functional long-range interactions in the genome of this eukaryotic RNA virus (Figure 9B). This model effectively illustrates that the different interactions are highly integrated and thus require a significant level of coordination in order to function properly (Figure 9B). Indeed, the newly identified UL–DL interaction places replication elements RII and RIV close to both translational (i.e. 3′CITE/5′UTR) and transcriptional (i.e. AS1/RS1 and AS2/RS2) RNA elements (Figure 9B). Thus, if all of the long-range interactions were to occur at once, as is depicted in Figure 9B, they would form a core of regulatory RNA elements. Notably, the large intervening segments, which are primarily coding regions, are predicted by mfold to form discrete domains (Figure S1). Based on this proposed configuration, the different regulatory RNA elements (and any associated proteins) could potentially communicate directly with each other in a common zone that could be used for coordinating different viral processes and/or sharing protein factors (Figure 9B).
Temporally, initiation of a viral infection would start with translation, followed by replication (i.e. replicase assembly), and then transcription. Accordingly, each of the different types of functional long-range RNA–based interactions would need to be dynamic and able to form in parallel with this series of events. Mechanistically, some of these interactions are predicted to be inhibitory to others and, accordingly, some processes may be impeded by others. Such relationships could be integral to coordinating different events or may simply act as safe-guards to prevent two processes from occurring simultaneously. For example, the UL–DL interaction would be inhibited by translation that is mediated by the 5′-3′ SL3-SLB interaction, because ribosomes translating the read-through portion of the p92 ORF would disrupt the UL–DL interaction (as well as the RII structure) (Figure 9B). Conversely, minus-strand synthesis of the genome, which is mediated by the UL–DL interaction, would cause the actively copying RdRp to disrupt the translation-related SL3-SLB interaction (as well as the 3′CITE structure), thereby inhibiting translation (Figure 9B), as has been proposed for Barley yellow dwarf virus (BYDV) [38]. Such incompatibilities could serve regulatory functions and aid the virus in switching from one process to another (e.g. translation to replication; [38]). In other cases, interactions, such as the UL–DL replication-related interaction and the AS-RS transcription-related interactions, may be mechanistically compatible, because both of the associated viral processes require the RdRp and initiation of minus-strand synthesis [39]. The higher-order RNA genome structural model in Figure 9 serves to illustrate an extraordinary level of integration of long-range RNA–based contacts and extends our fundamental view of RNA virus genome structure and function. Importantly, it also provides a useful molecular framework for future studies aimed at unraveling the dynamic regulatory interplay between these diverse sets of interactions.
TBSV represents an extreme example of long-range RNA–RNA interactions participating in three distinct viral processes. Similar types of distal RNA networks have also been reported in other positive-strand RNA viruses. For instance, the luteovirus BYDV utilizes two distinct sets of long-range RNA–based interactions for mediating translational initiation [40] and readthrough [38]. The requirement for base pairing between terminal genomic UTRs for efficient translation was shown initially in BYDV [40] and subsequently in TBSV [12],[13]. Interestingly, a recent report on the nepovirus Blackcurrant reversion virus distinguishes it as an additional virus confirmed to have this terminal pairing requirement [41]. In contrast, translational regulation in the bacteriophage MS2 involves internally located RNA–based interactions that modulate internal initiation events [42]. Accordingly, both prokaryotic and eukaryotic ribosome function can be modulated by the activity of this type of RNA interaction.
Sg mRNA transcription in the nodavirus Flock house virus involves an intra-genomic long-distance interaction [43] that likely functions in a manner similar to those that mediate transcription in TBSV [14]–[16]. The dianthovirus Red clover necrotic mosaic virus is also proposed to use an analogous premature termination mechanism; however it represents an extraordinary case in which transcription requires interaction between its two genomic RNA segments [44]. More recently, a distal intra-genomic interaction in the coronavirus Transmissible gastroenteritis virus was discovered that acts as an enhancer of sg mRNA transcription [8]. Collectively, these findings indicate both primary and secondary roles for these interactions in the process of sg mRNA transcription.
For genome replication, the potexvirus Potato virus X (PVX) utilizes an extensive set of long-range RNA–RNA interactions [6] that, interestingly, are also involved in sg mRNA transcription [45]. Unlike the viruses presented in Figure 8B, PVX possesses a supergroup-III RdRp [37]. Additionally, its functional interactions differ from those described for the supergroup-II RdRp viruses in that the primary sequence of the participating RNA segments, in addition to their complementarity, is important for activity [6],[45]. This extra requirement suggests a mechanism that is more complex than the bridging functions proposed for TBSV, DenV and Q-beta phage.
A role for long-range intra-genomic interactions in viral replication has also been demonstrated for different flaviviruses [9], [46]–[50], as well as for related Hepatitis C virus [7], [51]–[53]. Moreover, important long-distance RNA–RNA interactions have been described for both minus-strand RNA viruses [54]–[57] and retroviruses [58],[59]. This ever growing list of diverse RNA viruses illustrates both the prevalence and fundamental importance of long-range RNA–RNA interactions in a wide assortment of reproductive strategies.
Conceptualizing viral RNA genomes as complex higher-order RNA structures will provide valuable models for understanding their many functions. Indeed, recent computationally-based structural studies suggest that viral genome-scale ordered RNA structures (GORS) are more prevalent than previously appreciated [60],[61] and atomic force microscopic analysis has shown that viral RNA genomes are capable of adopting pseudo-globular conformations [61]. These findings further underscore the importance of considering overall RNA structure when investigating the roles of viral RNA genomes. Indeed, increasing our knowledge of global RNA structure and function will improve our understanding of mechanistic features of virus reproduction, facilitate the engineering of effective viral vectors, and help define genome-level structural constraints that influence RNA virus genome evolution.
Constructs described previously that were used in plant-based experiments include: T100, the wt TBSV genome construct [11]; AS1m1, a modified TBSV genome with sg mRNA1 transcription inactivated [15] and; DI-73, a small TBSV-based replicon [32]. Using the above constructs, and the infectious clone of CIRV [62], additional viral mutants were made and the relevant modifications are presented in the accompanying figures. For yeast-based experiments, the p33- and p92-expression constructs pGBK-His33 and pGAD-His92 have been described previously [33]. The newly generated DI-73 containing plasmids (Y73 series; Figure 7) were based on pYES-DI-72(+)Rz [33] and expressed full-length DI-73(+) RNA transcripts in yeast cells. The corresponding replication-defective mY73 series contained substitutions in RI that are described elsewhere (see mutant m2 in Figure 6 in [63]). All modifications were made using PCR-based mutagenesis and standard cloning techniques [64]. The PCR-derived regions in the constructs were sequenced completely to ensure that only the intended modifications were present.
In vitro RNA transcripts of viral RNAs were generated using T7 RNA polymerase as described previously [32]. Preparation and transfection of cucumber protoplasts and extraction of total nucleic acids were carried out as outlined in Choi and White, 2002 [15]. Briefly, isolated cucumber protoplasts (∼300,000) were transfected with RNA transcripts (3 µg for genomic RNA; 1 µg for replicon RNA) and incubated at 22°C for 22 hr. Isolated total nucleic acid preparations were subjected to northern blot analysis to detect plus- and minus-strand viral RNAs as described previously [15]. Nucleic acids were either treated with glyoxal and separated in 1.4% agarose gels or denatured in formamide-containing buffer and separated in 4.5% polyacrylamide-8% urea gels. Equal loading for all samples was confirmed prior to transfer via staining the gels with ethidium bromide. Viral RNAs were detected using strand-specific 32P-labeled probes and relative isotopic levels were determined using PharosFx Plus Molecular Imager (BioRad).
DI RNA stability assays were performed as described previously [63] and RNA secondary structures were predicted using mfold version 3.2 [29],[30].
Plant-derived replicase assays were carried out as described previously [17]. Briefly, extracts prepared from tombusvirus-infected N. benthamiana plants were supplemented with a buffer containing NTPs (UTP being labeled), viral RNA templates and other components. The reactions were then incubated at 25°C for 120 min. After phenol/chloroform extraction and ammonium acetate/isopropanol precipitation, half the amount of the RNA products was analyzed in 5% polyacrylamide-8% urea gels followed by detection by autoradiography.
Yeast replication assays were carried out in whole cells expressing p33, p92, and the RNA replicon transcript from cotransformed plasmids, as described previously [33]. Following induction and an incubation period, isolated viral RNAs were separated in 5% polyacrylamide-8% urea gels and viral RNA was detected by northern blot analysis. For yeast replicase assembly assays, viral replicase was prepared from yeast cells (transformed as described above) by affinity purification as described elsewhere [26]. The level of replicase activity was determined by adding minus-strand DI-72, DI-72(-) and assessing the amount of 32P-labeled complementary-strand product generated by 5% polyacrylamide-8% urea gel electrophoresis and autoradiography. Standard western blot analysis was used to monitor the levels of p33 expressed in the cells used for replicase purification [26].
Translation of sub-saturating amounts (0.5 pmol) of RNA transcript was carried out in nuclease-treated wheat germ extract (Promega) as described previously [31]. Protein products were separated by sodium dodecyl sulfate-polyacrylamide gel electrophoresis and quantified by radioanalytical scanning using a PharosFx Plus Molecular Imager (Bio-Rad) and QuantityOne software (Bio-Rad).
|
10.1371/journal.ppat.1006825 | Phenotypic deficits in the HIV-1 envelope are associated with the maturation of a V2-directed broadly neutralizing antibody lineage | Broadly neutralizing antibodies (bnAbs) to HIV-1 can evolve after years of an iterative process of virus escape and antibody adaptation that HIV-1 vaccine design seeks to mimic. To enable this, properties that render HIV-1 envelopes (Env) capable of eliciting bnAb responses need to be defined. Here, we followed the evolution of the V2 apex directed bnAb lineage VRC26 in the HIV-1 subtype C superinfected donor CAP256 to investigate the phenotypic changes of the virus populations circulating before and during the early phases of bnAb induction. Longitudinal viruses that evolved from the VRC26-resistant primary infecting (PI) virus, the VRC26-sensitive superinfecting (SU) virus and ensuing PI-SU recombinants revealed substantial phenotypic changes in Env, with a switch in Env properties coinciding with early resistance to VRC26. Decreased sensitivity of SU-like viruses to VRC26 was linked with reduced infectivity, altered entry kinetics and lower sensitivity to neutralization after CD4 attachment. VRC26 maintained neutralization activity against cell-associated CAP256 virus, indicating that escape through the cell-cell transmission route is not a dominant escape pathway. Reduced fitness of the early escape variants and sustained sensitivity in cell-cell transmission are both features that limit virus replication, thereby impeding rapid escape. This supports a scenario where VRC26 allowed only partial viral escape for a prolonged period, possibly increasing the time window for bnAb maturation. Collectively, our data highlight the phenotypic plasticity of the HIV-1 Env in evading bnAb pressure and the need to consider phenotypic traits when selecting and designing Env immunogens. Combinations of Env variants with differential phenotypic patterns and bnAb sensitivity, as we describe here for CAP256, may maximize the potential for inducing bnAb responses by vaccination.
| HIV-1 infected individuals rarely develop broadly neutralizing antibodies (bnAbs) that inhibit diverse HIV-1 subtypes. As activity against the majority of HIV-1 isolates is necessary for effective immunization against HIV-1, current vaccine development seeks to generate regimens that evoke bnAb responses. Delineating why bnAbs develop in certain cases of HIV-1 infection is thus of pivotal importance. In all infected individuals, HIV-1 inevitably escapes neutralizing antibody pressure and even the most potent bnAbs cannot clear the infection from the patient where they emerged. Yet, exactly this continuous interplay between virus escape and antibody maturation is believed to be crucial in the evolution of bnAbs, with virus escape variants containing modified HIV-1 envelope (Env) proteins, the target of neutralizing antibodies that in some cases can direct the immune response towards breadth. Identifying features of naturally occurring Env proteins that are involved in evoking bnAb responses is thus of high interest. Here, we analyzed Env features of virus isolates from donor CAP256 who developed the potent V2 apex bnAb VRC26, one of the current lead bnAbs for HIV-1 therapy and vaccine development. We show that the viral escape variants that appeared soon after the onset of the VRC26 response had highly altered Env protein properties that, in addition to reducing sensitivity to VRC26, affected their capacity to infect and altered entry dynamics. This highlights constrained viral escape pathways, but also features of VRC26 that may have prevented rapid escape. We postulate that this may have resulted in a prolonged circulation of partially VRC26 sensitive viruses, hence allowing the bnAb response to mature.
| Broadly neutralizing antibodies (bnAbs) are a focus of HIV-1 vaccine development and passive immunization strategies [1–7]. Owing to the exceptionally potent and broad bnAbs that have been isolated over recent years [8–12], a wealth of information on their function has become available. However, factors that govern bnAb evolution in infection are not fully resolved, nor have current vaccine designs succeeded in eliciting bnAb responses. The expectations for vaccines are high as they will need to do substantially better than in infection where bnAbs only evolve in around 10–25% of HIV-1 infected individuals, with the most potent elite neutralizing antibodies restricted to approximately 1% of infections [13–16].
A number of parameters have been implicated in the development of neutralization breadth including the length of HIV-1 infection, high viral loads, virus diversity, CD4 T cell loss, involvement of regulatory T cell subsets, viral subtype and host factors including ethnicity and HLA genotype [13,16–30]. It has become clear that a tight interplay of antibody and virus escape variants directs antibody adaptation and diversification towards the generation of bnAbs [31–36]. Escape from even the most potent bnAbs appears inevitable, as all bnAbs identified to date have been isolated from individuals who ultimately failed to control viremia [13,18,37,38]. However, it is the continuous exposure to these gradually escaping viruses that appears to be key for the evolution of bnAbs.
Env variants that are capable of stimulating the germline ancestors of bnAbs as well as intermediate Env variants that steer antibody evolution to tolerate viral diversity, conferring breadth, are therefore urgently sought [23–25]. The phenotypic features of Env variants associated with the elicitation and maturation of breadth are largely unknown. Does the initial infecting strain harbor distinct features that enable it to more efficiently initiate bnAb precursors? [39,40] Are the same features relevant for the later evolving Env variants? Does incomplete escape, paired with replication in sites where antibody access is restricted, enable the virus to survive despite the presence of potent neutralization activity?
In the present study, we have addressed these questions by investigating the phenotypic plasticity of HIV-1 Env during bnAb development. As a model, we took the evolution of autologous viruses during the maturation of the potent CAP256-VRC26 bnAb lineage. The V2 specific CAP256-VRC26 bnAb lineage developed in a HIV-1 subtype C superinfected individual[33,41]. Frequent sampling and detailed analysis of this donor over several years has resulted in a large collection of both longitudinal Env variants and VRC26 bnAbs[33,41]. These co-evolved Env and VRC26 bnAbs therefore provide an ideal setting to explore phenotypic shifts in Env that may reveal insights into properties that are important for antibodies like VRC26 to develop. Comparing Env properties of the primary infecting (PI) virus, the superinfecting (SU) virus and their descendants circulating prior to and during the rise of the CAP256-VRC26 bnAb lineage, our study aimed to assess differential functional capacities that may have shaped the evolution of the CAP256-VRC26 bnAb lineage.
The V2 apex targeting CAP256-VRC26 lineage emerged 30–34 weeks post infection (p.i.) after superinfection at week 15 (Fig 1 [34]). Plasma neutralization breadth developed around 48 weeks p.i. [42], and the first broad monoclonal antibodies were isolated from week 59 [33]. We included 17 longitudinal autologous viral envelopes (Env) from bnAb donor CAP256 (Fig 1, S1 Table) which circulated before and during the evolution of the bnAbs when the autologous viruses were still (partially) sensitive to the VRC26 bnAbs [33,34,41–43] and probed them against 12 members of the CAP256-VRC26 bnAb lineage with breadth varying from 7–63% (Fig 1, S2 Table). Fourteen of the CAP256 Env were isolated within the first 48 weeks of the infection and showed at least partial sensitivity to the VRC26 mAbs. These included the primary infecting (PI) virus (week 6) as well as the superinfecting (SU) virus (week 15) to which the bnAb response was mainly directed [33,34,42]. Later CAP256 viral variants were defined as PI-like, SU-like or PI/SU recombinant viruses according to their V1V2 sequence, which is the target of the VRC26 bnAbs [33,34] (S1 Table). In addition to the earlier variants up to week 48, we included three Env variants isolated at week 176 of infection, which had completely escaped the VRC26 lineage [33,42]. These Env were used to assess phenotypic functions that might impact antibody efficacy, namely viral infectivity, mode of transmission, sensitivity to neutralization, and entry kinetics. We specifically focused on defining properties for the SU virus variants, given the clear evidence that they induced the VRC26 lineage [33,41]. Due to the scope of the analyses, certain tests were restricted to the assessment of a smaller panel of Envs in which cases we prioritized the analyses of SU over PI viruses.
To assess the phenotypic properties of the CAP256 viruses, we first assessed their properties in free virus and cell-cell spread. HIV-1 has the ability to spread as free virus or through contacts between infected and uninfected target cells [44,45]. Most neutralizing antibodies (nAbs) including bnAbs, neutralize cell-cell transmission less effectively than free virus spread [46–53]. The loss of neutralizing activity (regardless of specificity) during cell-cell transmission is often substantial, decreasing bnAb activity 10–100 fold compared to free virus transmission [47,49–53]. Importantly, this reduced activity during cell-cell transmission can therefore substantially contribute to neutralization escape [47,52,54]. However, the VRC26 bnAbs [52], similar to MPER bnAbs [47,52], frequently retain comparable activity against heterologous HIV-1 strains in both transmission routes. Understanding how this capacity evolved in VRC26 is of interest as preserved activity in both modes of transmission will limit the emergence of escape variants [54] and would therefore be highly desirable for therapeutic and vaccine induced bnAbs.
To assess whether autologous cell-cell transmission was similarly sensitive to VRC26, and whether this was maintained over the course of infection, we assessed the selected 17 longitudinal CAP256 Env variants for neutralization sensitivity in the A3.01-CCR5 cell-based free virus and cell-cell transmission assays as described [52]. Week 176 viruses were fully resistant to VRC26 bnAbs (S2 Table) and were therefore not included in the comparison of free and cell-cell neutralization activity. As observed in free virus transmission (Fig 2A) [34,42,43], PI-like viruses were resistant to neutralization by most members of the VRC26 bnAb lineage during cell-cell transmission while SU-like viruses were sensitive (Fig 2B). In line with what we have reported previously for heterologous viruses [52], the VRC26 bnAbs showed a 10-fold mean activity loss across all VRC26 variants against the autologous virus strains during cell-cell transmission compared to free virus transmission (Fig 2B and 2C) with occasional virus-VRC26 bnAb combinations showing substantially lower and some higher activity losses in cell-cell than in free virus transmission (Fig 2C). Activity in cell-cell transmission was preserved across all VRC26 bnAb lineage members, regardless of the heterologous breadth of individual mAbs (S1A and S1B Fig) or the extent of their maturation, as measured by their level of somatic hypermutation (S1C and S1D Fig), suggesting that the preserved activity during cell-cell transmission may be a result of the specific mode of action of the VRC26 bnAbs.
Based on their sensitivity to VRC26 bnAbs, SU-like viruses were grouped into SU-like VRC26 sensitive and SU-like VRC26 early escape viruses. Later SU-like viruses (week 42–48) were significantly less sensitive to VRC26 inhibition than early SU-like viruses (week 15–34) in both transmission pathways except for clone 42-wk.24SU which phylogenetically clustered with sensitive Env isolated at week 34 post infection (S2 Fig). Across the longitudinal PI and SU-like viruses, we observed similar antibody activity during free virus and cell-cell transmission (Fig 2A–2D) suggesting that cell-cell transmission may not have been a preferential escape pathway of the autologous virus from VRC26 pressure.
We therefore investigated if the high activity of the VRC26 lineage against cell-cell spread is unique to VRC26 or a feature common to V2 apex-directed bnAbs. We compared inhibition of free virus and cell-cell transmission of the CAP256 virus panel by bnAbs targeting the CD4 binding site (CD4bs; PGV04 [27] and 3BNC117 [55]), the V3 glycan supersite (PGT121 [56]), the V2 apex (PG9 [57] and PGT145 [56]) and the MPER region (10E8 [58]) with VRC26.25, the broadest and most potent VRC26 variant (Fig 3).
Increased resistance of CAP256 viruses to VRC26 was mirrored by a significantly higher resistance of both PI and SU-like viruses to free virus neutralization by PG9 and PGT145 (Mann-Whitney Test, p<0.0001), highlighting the similar modes of action of these V2 apex bnAbs (Fig 3A and 3B; [42]). While SU-like viruses up to week 48 were neutralized by the V2 glycan bnAbs, PI-like Env were highly resistant in free virus infection but even more so in cell-cell transmission where both PG9 and PGT145 completely lacked activity (Fig 3A and 3B). Sensitivity to the CD4bs, V3 glycan and MPER bnAbs did not fluctuate substantially in CAP256 viruses over time in either pathway (S3 Table). While the difference of SU-like virus 42-wk.5SU cell-cell neutralization sensitivity to the other strains did not reach statistically significance, 42-wk.5SU had notably the highest loss in cell-cell neutralization sensitivity for all bnAbs including the V2 glycan bnAbs. This may suggests that 42-wk.5SU either adopted a specific Env conformation that occludes mAb access during cell-cell transmission for a wide spectrum of mAb specificities, or that this virus is particularly well adapted for cell-cell spread. In line with previous observations [47,52], MPER bnAb 10E8 displayed preserved neutralization activity against most CAP256 viruses during cell-cell transmission similar to the V2 glycan bnAbs (Fig 3). CD4bs and V3-directed bnAbs, however, showed a significantly lower activity during cell-cell transmission compared to V2- and MPER-directed bnAbs (Fig 3; Mann-Whitney Test of fold change IC50s of CD4bs- and V3 bnAbs versus V2- and MPER bnAbs, p<0.0001). Collectively, this suggests that the capacity to retain activity in cell-cell transmission is linked to the bnAb’s specificity and its mode of action.
We compared Env functionality of the longitudinal CAP256 virus panel in free virus (Fig 4A) and cell-cell transmission (Fig 4B), and detected substantial variability in Env infectivity of PI and SU-like viruses in both transmission modes. High variability was even observed for Env variants from the same time point, as exemplified by Env clones from week 42 (Fig 4A and 4B). Overall, early escape viruses exhibited a significantly reduced entry fitness compared to VRC26 sensitive SU-like viruses in free virus but not cell-cell transmission. Interestingly, in both modes of transmission, the 176 week viruses showed significantly reduced (Mann-Whitney test, p<0.0001) entry capacity compared to earlier variants suggesting an eventual fitness cost for escape from VRC26 bnAbs. Although we have no formal proof at this stage for a direct causality, it is intriguing to note that the CAP256 donor experienced reduced viral loads at later time points that could be the result of decreased viral fitness [42].
The free virus and cell-cell infection assay systems require distinct culture conditions and therefore do not allow for a direct quantitative comparison of virus infectivity but provide a means for relative measures and comparison of patterns across pathways. Analyzing the relative infectivity of longitudinal CAP256 Env, we found that infectivity in free virus and cell-cell transmission was tightly linked, indicating that enhanced cell-cell transmission did not compensate for substantial defects in replication observed in free virus transmission (Fig 4A, 4B and 4C).
To obtain a more direct comparison of the infectivity patterns, we normalized all results to the PI virus (Fig 4D). With one exception (34-wk.81SU), the relative infectivity of viruses in the cell-cell format was comparable (+/- 10%) or higher than in free virus transmission across all three virus groups (PI-like, SU-like, PI/SU recombinants; Fig 4D). Interestingly, clone 42-wk.5SU, a clone from the onset of VRC26 escape (Fig 1), which displayed higher resistance to neutralization in cell-cell transmission (Fig 2), showed a significantly lower entry fitness than the other Env variants in both transmission pathways (Mann-Whitney test, p<0.0001). Thus, neutralization escape from VRC26 bnAbs coincided with loss in neutralization sensitivity across different bnAb specificities in cell-cell transmission for this Env variant but at the same time at a reduced entry capacity. Overall, cell-cell transmission efficacy was significantly higher amongst later evolving PI and SU variants (Fig 4E) and often coincided with low infectivity (Fig 4A and 4B). This may indicate that cell-cell transmission, while not fully compensating for entry defects, allowed for better replication for the infectivity impaired escape variants that emerged in response to VRC26 and the non-VRC26 autologous nAbs.
To probe the impact of the polyclonal autologous neutralization response on virus evolution, we determined the sensitivity of CAP256 viruses in both transmission pathways to autologous plasma collected at week 145 p.i. (Fig 5A) which had been previously shown to have the maximum neutralization titer against heterologous viruses [33]. Viruses from week 176 were not included in this analysis, as they are resistant to week 145 plasma [42]. While both PI and SU-like viruses were sensitive to the autologous plasma in free virus transmission, titers were significantly higher for the SU-like viruses which were also significantly better neutralized during cell-cell transmission (Mann-Whitney test, p = 0.0426 for free virus and cell-cell), highlighting the dominance of VRC26 bnAbs in the autologous plasma. As observed for VRC26 bnAb neutralization, SU-like VRC26 early escape viruses were less sensitive to week 145 plasma inhibition than early SU-like viruses. However, differences were only significant for the cell-cell but not the free virus transmission pathway. (Fig 5A). In contrast, all but one PI-like virus showed high resistance to plasma antibodies during cell-cell transmission (Fig 5A and 5B). Furthermore, neutralization activity of VRC26 bnAbs and plasma against SU but not PI viruses correlated during free virus and cell-cell transmission (Fig 5A and 5C and S3A Fig), confirming that neutralization of PI viruses by the autologous plasma is mediated by other nAb specificities [42] that are presumably not active in neutralizing cell-cell transmission. A central question in understanding escape from neutralization is the consequence of escape mutations on Env fitness. Interestingly, increased VRC26 resistance was associated with free virus but not cell-cell infectivity loss (Fig 5D, S3B Fig). This is in line with the observation that late SU-like viruses maintain higher cell-cell than free-virus infectivity (Fig 4) paired with the slightly reduced cell-cell neutralization by VRC26 (Fig 2). Importantly, our findings highlight that in the early phase of VRC26 bnAb escape that we investigated here, virus variants with decreased fitness can emerge.
Our analyses thus far highlighted the importance of entry fitness differences during virus escape. To obtain mechanistic insights into which aspects of the entry process are linked with gains and losses in infectivity, we examined the entry kinetics of selected PI-like and SU-like viruses in an inhibitor time-of addition experiment as previously described [59,60] (Fig 6A). As this assay allows only for the analysis of a restricted number of Env at the same time, we focused on the SU-like Env to better study the phenotypic evolution of these autologous viruses. We analyzed the kinetics of CD4 attachment (blocked by the CD4 agent DARPin 55.2 [61]) and fusion (blocked by the fusion inhibitor T-20 [62]). Infection time courses for both entry steps were established by adding the respective inhibitors at distinct time points from 5 to 120 min. The mean half-maximal time (mean t1/2) required for viruses to progress with the entry process beyond these two steps were calculated from fitted curves (Fig 6B, S4 Fig) and time intervals between three different stages of the entry process (CD4 attachment, progress from CD4 attachment to fusion, and fusion) were compared (S5 Fig).
Amongst PI-like viruses, the PI virus (6-wkPI) and PI-like clone 42-wk.16PI engaged CD4 rapidly but required an extended time period to progress to fusion and establish infection. PI-like viruses 34-wk.18PI and 48-wk.17PI showed a different entry phenotype with a trend to less rapid CD4 engagement and faster progression towards fusion narrowing the time window between completion of CD4 binding and fusion (Fig 6B) which, however, did not reach significance. A similar prolonged progression from CD4 binding to fusion was also seen for later SU-like viruses from week 42 onwards, which was significantly slower compared to earlier SU virus variants (Mann-Whitney test, p = 0.0286).
We hypothesized that these differences in entry kinetics could be linked with VRC26 escape and changes in infectivity. To probe this, we compared associations of entry kinetic parameters with neutralization sensitivity to VRC26 and infectivity, both in free virus and cell-cell transmission (Fig 6C, S6 Fig). In line with the high resistance of PI-like viruses to the VRC26 bnAb lineage, overall no relevant association between entry kinetics and sensitivity to the VRC26 bnAbs was observed (S6 Fig). SU-like viruses, however, showed a very interesting evolution in their kinetic properties. Prolonged entry kinetics for attachment and fusion were highly significantly linked with lower infectivity, both in free virus and cell-cell transmission. This pattern was largely driven by the SU-like VRC26 early escape variants. Indeed, prolonged entry kinetics correlated with increased VRC26 resistance in both, free virus and cell-cell transmission, confirming that the acquisition of mutations conferring increasing resistance to the VRC26 bnAb lineage were associated with an altered entry phenotype and fitness loss (Fig 6C, S6 Fig).
Interestingly, the time to CD4 engagement had no impact on free virus VRC26 potency and was also the weakest influence for cell-cell inhibition (Fig 6C), suggesting that VRC26 may not depend on a rapid binding prior to CD4 attachment in order to neutralize. We thus directly explored the efficacy of VRC26 against free virus before and after CD4 engagement for a selection of Env variants, focusing again on the SU-like autologous viruses (Fig 7). To distinguish pre and post CD4 activity, we synchronized infection by spinoculation at 23°C, a temperature that allows for virus binding to CD4 but not for fusion, and added the bnAbs either before or after CD4 attachment [47,52].
Individual virus-VRC26 bnAb combinations differed in pre- and post-attachment activity (Fig 7A and 7B) but overall, sensitivity in free virus and cell-cell transmission correlated with pre- and post-attachment activity for PI-like and SU-like viruses (Fig 8A). Most intriguingly, the SU-like VRC26 sensitive viruses were significantly better neutralized during both pre- and post-attachment (Figs 7B and 8B; Mann-Whitney test, p<0.0001). This ability to access their epitope both on the native trimer and post-CD4 triggering allows for a prolonged window of action that may thus contribute to the bnAbs’ potency and breadth for both viral transmission routes. For SU-like VRC26 early escape viruses, however, post-attachment neutralization decreased to a significantly greater extent (Mann-Whitney test, p<0.0001) compared to SU-like VRC26 sensitive viruses (Figs 7B and 8B), suggesting that resistance conferring mutations may have a more pronounced effect on the CD4 bound conformation of the CAP256 Env.
Considering all phenotypic traits we investigated, our analyses lead us to suggest that an initial pressure on free virus might have resulted in the generation of escape variants with altered entry properties. This first wave of partial escape from VRC26 might have led to a reduced entry capacity that was linked with altered entry kinetics and a more pronounced decrease in post-attachment activity of the VRC26 bnAb lineage.
A continuous, tight interplay between the HIV-1 envelope (Env) antibody response and virus escape is a fundamental component of bnAb development [31,32,34,35,41]. Consequently, elucidation of escape pathways can provide valuable insights for bnAb based vaccine design. Viral escape leads to the formation of new Env variants that affinity mature the bnAb response. Deciphering which Env variants were instrumental in shaping the bnAb response in natural infection will thus help to inform immunogen design. To date, we lack criteria that allow for pinpointing these Env variants amongst the wide spectrum of quasispecies that evolve during virus escape. Knowing which phenotypic features were preserved or alternatively lost during escape evolution will allow to select Env variants not only based on epitope variation, but also on phenotypic traits that may be relevant for epitope exposure, virus survival and antibody efficacy. In the present study, we explore phenotypic traits of the virus Env during bnAb evolution to distinguish Env properties that may have influenced bnAb development or resulted from its action. We do this by studying longitudinal Env and bnAb variants available from donor CAP256 who developed the potent V2 apex VRC26 bnAb lineage [33,34,41,42].
Although a range of host and viral factors that contribute to bnAb development have been identified [13,16–30], we currently do not know to what extent a given Env will steer the antibody response towards bnAb development. That env genes are not equally effective in inducing bnAbs is widely assumed and we have recently shown this in a survey of neutralization breadth in close to 4500 individuals where we observed higher frequencies of CD4bs bnAbs amongst HIV-1 subtype B infections while V2 apex bnAbs proved more frequent in non-subtype B infections [16]. Env immunogens that are capable of triggering a bnAb response may thus harbor distinct phenotypic features that facilitate bnAb precursor binding or mAb maturation. This could include a multitude of features that influence epitope exposure such as the degree of shielding, Env stability, the density of the trimeric spike on virions or specific conformation traits that affect the exposure of the target epitope.
bnAb evolution has been shown to benefit from viral diversity [13,16–28,30,39]. Viral populations which have evolved to high diversity or harbor high diversity due to superinfection, as in the case of patient CAP256, represent a mixture of Env immunogens that undoubtedly will widely differ in phenotype. It is feasible that the necessity to simultaneously react with highly differential epitopes, e.g. open (neutralization sensitive, epitope exposing) and closed (neutralization resistant, epitope largely hidden) immunogens, may steer the immune system into recognizing a broad variety of Env conformations. Intriguingly, such a dichotomy in sensitivity was present in CAP256 during VRC26 development with the highly VRC26 resistant PI strain and the VRC26 sensitive SU strain co-circulating eventually resulting in the evolution of broad VRC26 variants that harbor activity also against the PI strain [33,41] (Figs 1 and 2).
Other phenotypic aspects and characteristics of the virus-Ab interplay are also likely to be important. Slow escape from the bnAb lineage may be beneficial by prolonging virus-bnAb evolution and allowing breadth to develop [63]. Comparatively higher initial resistance to neutralization or a higher tolerance to mutations in Env could thereby prove beneficial.
Likewise, replication properties may be important. As escape from CAP256 highlights, the kinetics of the entry process may influence neutralization efficacy by altering the window of action for neutralization. A differential capacity to replicate as free virus or by cell-cell transmission has been recognized to affect neutralization activity, with cell-cell transmission often less effectively inhibited by nAbs. This results in a higher propensity to select for resistance mutations in cell-cell transmission [47,49–52,54,63,64]. As we show here, VRC26 is one of the few antibodies that retains substantial activity against cell-associated virus. This was particularly evident for virus strains for which VRC26 has a comparatively lower activity in free virus transmission (Fig 2E) [52]. The fact that with increasing VRC26 resistance the transmission mode appears to have less influence on nAb efficacy is intriguing. This could imply that the Env of evolved strains adopts an Env conformation in the unliganded stage that resembles Env conformations that are relevant during cell-cell transmission. By investigating the phenotypic properties of autologous virus strains which circulated prior to and during the evolution of the bnAb CAP256-VRC26 lineage we show that the capacity of VRC26 to maintain activity during cell-cell transmission is shared by other V2 apex bnAbs, suggesting that their common mechanism of action contributes to this. What constitutes the differences for certain bnAbs but not others in the context of free and cell-cell transmission will be important to resolve in forthcoming studies. Multiple scenarios may apply, and dissecting whether a reduced capacity results from an inability to recognize Env on recently budded, immature viruses or the CD4 bound Env will be of particular interest.
The capacity of VRC26 to act before and after CD4 engagement provides a first clue to how VRC26 is equally effective in cell-cell transmission [52] as it allows the bnAb to be active over a longer time window during the entry process. Considering that attachment to CD4 in cell-cell transmission is rapid, the capacity to block entry after CD4 binding is likely a benefit for cell-cell neutralization activity [52].
The first wave of VRC26 escape variants we investigated here thus hints at an intriguing phenotypic plasticity in Env functionality. While the virus managed to maintain its Env replication competence, these changes seem to have come at entry fitness costs in the initial phase of resistance adaptation to VRC26. Both, a loss in replicative fitness of the initial escape variants and the sustained sensitivity to VRC26 during cell-cell-spread may have aided the evolution of VRC26 breadth by reducing virus progeny and thus slowing formation of escape variants. Prolonged exposure until full escape is reached will favor the long-term circulation of virus variants with an intact VRC26 epitope, thereby increasing the chances for the bnAb response to mature. Indeed, slow viral escape has been reported in several bnAb lineages [32,34,63,65] though the mechanism for this has not been clear as passive administration of bnAbs results in rapid viral escape [66–70].
In summary, our findings illustrate the phenotypic plasticity of the HIV-1 Env in evading bnAb pressure. As exemplified by VRC26, alterations in phenotypic traits that emerge in response to a bnAb response can provide insights into functional consequences of viral escape and potentially may highlight differential Env conformations that should be considered when selecting and designing Env immunogens. Sequential immunization protocols to mature bnAb responses by vaccination may be beneficial if combinations of Env variants with differential phenotypic patterns and bnAb sensitivity, as we describe here for CAP256, are included.
A cryo-preserved plasma sample from week 145 p.i. of the CAP256 patient for the analyses of the current study was provided by the repository of the CAPRISA 002 Acute Infection study, Durban, South Africa [71]. The CAPRISA 002 study was reviewed and approved by the research ethics committees of the University of KwaZulu-Natal (E013/04), the University of Cape Town (025/2004) and the University of the Witwatersrand (MM040202).
CAP256 VRC26 bnAbs and expression plasmids have been described previously [33,41]. We thank D. Burton, J. Mascola, M. Nussenzweig and M. Connors for providing antibodies and antibody expression plasmids for this study either directly or via the NIH AIDS Research and Reference Reagent Program (NIH ARP). All antibodies were expressed in FreeStyle 293-F cells and purified on protein G affinity and size exclusion chromatography columns as described [72].
The CD4-directed DARPin 55.2 was produced as described [61], the fusion inhibitor T-20 [62] was purchased from Roche Pharmaceuticals.
293-T cells (obtained from the American Type Culture Collection (ATCC)) and TZM-bl cells ([73]; obtained from the National Institute of Health AIDS Reagent Program) were maintained in DMEM with 10% heat inactivated FCS and 1% Penicillin/Streptomycin. FreeStyle 293-F cells were purchased from Thermo Fisher Scientific and maintained according to the manufacturers instructions. A3.01-CCR5 cells were previously generated [47] and were cultivated in RPMI with 10% heat inactivated FCS and 1% Penicillin/Streptomycin.
Envelope (Env) genes of patient CAP256 were previously cloned as described [33,34,42] and GenBank accession numbers are summarized in S1 Table. We only had the capacity to include a selection of available Env clones in the current study, to allow an assessment of all variants in the same assay runs to restrict assay variability. We primarily focused on multiple Env variants from key time points prior to and during the emergence of the VRC26 lineage, when the autologous viruses were still partially sensitive to VRC26 [34] as the interaction with VRC26 was a main aim of our study. The 34wk and 48wk time points thus allowed a clear focus on VRC26-mediated pressure. We did not include strains beyond wk176 as these viruses are resistant to VRC26. For certain complex and labor intensive analyses (e.g. kinetics assay and pre-post attachment assay), only a restricted virus panel was assessed to still allow processing of all Env variants in the same assay run. As the emphasis of our study was on the SU viruses, these were preferentially included. Amongst PI viruses, we preferentially included later viruses over earlier (23wkPI and 30wkPI) variants as phenotypic evolution in later clones which experienced VRC26 potentially are of higher interest in the context of the current study.
For the production of single-round replicating cell-free pseudovirus stocks, 293-T cells were transfected with the luciferase reporter HIV-1 pseudotyped vector pNLlucAM [74] (a gift from A. Marozsan and J. P. Moore) and the respective Env expression plasmids as described [75]. Infectivity of reporter viruses was quantified by titration of virus containing supernatants on 1*104 TZM-bl or 5*104 A3.01-CCR5 in a 1:4 ratio starting from 100 μl virus solution/well in the presence of 10 μg/ml diethylaminoethyl (DEAE, Amersham Biosciences, Connecticut, USA). Infection of target cells was assessed by measuring the firefly luciferase activity from the lysed cells using firefly luciferase substrate (Promega, Madison Wisconsin, USA). Emitted RLU were quantified on a Dynex MLX luminometer (Dynex Technologies Inc., Chantilly, Virginia, USA).
Infection via the cell-cell-route is described below. As the A3.01-CCR5 cell-cell transmission assay relies on the omission of polycations to exclude free virus spread, all CAP256 Env isolates used in the current study were confirmed to require polycations for free virus entry (S7 Fig).
To assess viral infectivity in free virus cell-cell transmission, we employed recently described assay formats utilizing A3.01-CCR5 cells as target cells and either cell-free virus or 293-T transfected cells as source of virus [52]. In both systems, a reporter virus backbone lacking env and an env expression vector are co-transfected to generate Env pseudoviruses or Env pseudovirus expression cells, respectively. An essential difference in the free virus and the cell-cell transmission assay is the virus backbone used. The free virus set up uses the conventional NL4-3 based pNLlucAM luciferase reporter Env pseudotyping backbone [74]. The cell-cell transmission assay utilizes a NL4-3 derived pseudotyping HIV-1 backbone with an intron-regulated Gaussia luciferase LTR-reporter construct called inGluc (kind gift from Dr. M Johnson [76–78]). The reverse orientation of the reporter and the intron allow luciferase expression only after correct splicing, packaging into viral particles and infection of A3.01-CCR5 target cells. Free virus infection in the cell-cell transmission set up is restricted by the omission of DEAE in the infection medium as described previously [47].
To measure infectivity in free virus and cell-cell transmission, 5*104 293-T cells per 24-well were transfected with Env and NLinGluc backbone for cell-cell transmission or pNLlucAM backbone plasmids for free virus transmission in a 1:3 ratio, using polyethyleneimine (PEI) as transfection reagent. To test cell-cell infectivity, 293-T cells were transfected with the inGluc reporter and an Env plasmid of choice. After 6 h incubation, cells were detached and 5*103 transfected cells /100 μl RPMI medium seeded per 96 well. A3.01-CCR5 target cells (1.5*104 /100 μl RPMI medium) were added to each well. After 65 h of incubation at 37°C, Gaussia luciferase activity in the supernatant was quantified using the Renilla Luciferase Assay System (Promega, Madison Wisconsin, USA) according to the manufacturer’s instructions.
To test free virus infectivity, 293-T cells transfected with the pNLlucAM luciferase reporter and an Env expression plasmid were incubated for a total of 65 h, and the transfection medium exchanged after 8 h. After 65 h, the supernatant was collected, briefly centrifuged at maximum speed and frozen for 24 h to prevent carry-over of transfected cells. 100 μl of a 1:2 virus dilution with DMEM medium were seeded into 96 well plates in duplicates or triplicates and 5*104 A3.01-CCR5 target cells in 100 μl per 96 well in the presence of 10 μg/ml DEAE were added. After 65 h incubation at 37°C, infection was assessed by luciferase production after cell lysis and addition of firefly luciferase substrate (Promega, Madison, Wisconsin, USA).
Free virus inhibition by bnAbs and plasma was assessed on A3.01-CCR5 cells using Env-pseudotyped NLlucAM reporter viruses as described [52]. Briefly, a virus input of around 10,000 relative light units (RLU) per 96 well in absence of inhibitors was pre-incubated with the respective bnAbs or patient plasma for 1 h at 37°C. 5*104 A3.01-CCR5 target cells per 96 well in the presence of 10 μg/ml DEAE were added and incubated for 65 h at 37°C. Infection was assessed by firefly luciferase production as described above. The inhibitor concentrations or plasma dilutions causing 50% reduction in viral infectivity (50% inhibitory concentration; IC50 or 50% neutralizing titers, NT50) were calculated by fitting pooled data from two to four independent experiments to sigmoid dose response curves (variable slope) using GraphPad Prism. If 50% inhibition was not achieved at the highest bnAb concentration or lowest plasma dilution, a greater-than value was recorded.
For measuring neutralization of cell-cell transmission, 293-T cells were transfected with Env and NLinGluc plasmids in a 1:3 ratio. 6 h post transfection, 5*103 transfected cells were seeded in 50 μl per 96 well and serial dilutions of bnAbs or patient plasma in 50 μl per 96 well were added. After 1 h incubation at 37°C, 1.5*104 A3.01-CCR5 target cells in 100 μl RPMI medium were added for 65 h at 37°C. Gaussia luciferase activity in the supernatant was quantified using the Renilla Luciferase Assay System (Promega, Madison Wisconsin, USA) according to the manufacturer’s instructions. Neutralization data were analyzed with GraphPad Prism as described above.
The change in neutralization activity for cell-cell compared to free virus transmission (fold change IC50) was calculated as the ratio of the IC50 of neutralization during cell-cell and free virus transmission. If for only one of the pathways, the virus was resistant to a particular bnAb, the IC50 was nominally set to a value of two times the highest ineffective bnAb-concentration tested for that pathway.
Entry kinetics were assessed using an inhibitor time of addition set up as recently described [59,60] (Fig 6A). The essence of the assay is a synchronized infection of free virus that is blocked at distinct time points by inhibitors interfering with CD4 attachment (the CD4 blocking agent DARPin 55.2 [61]) or fusion (T-20 [62]).
Briefly, 6*103 TZM-bl cells in 60 μl DMEM medium per 384-wells were seeded and incubated for 24 h at 37°C. Cells were then shortly cooled at 4°C before removing the medium and adding HIV-1 pseudovirus stocks yielding 50’000 RLU in 60 μl DMEM medium with 10 μg/ml DEAE-Dextran at 4°C per 384-well. Virus binding to the cells was synchronized by spinoculation of the plates for 30 min at 2095 g and 4°C. Supernatants including unbound viruses were removed, 37°C warm DMEM was added to start the infection and plates were incubated at 37°C. At indicated time points, the infection was stopped by the addition of inhibitors of CD4 binding (1 μM DARPin 55.2) or fusion (50 μg/ml T-20). The chosen concentration of each of the inhibitors exceeded the 100% inhibitory concentrations, which had been determined for the individual virus strains. After 48 h incubation at 37°C, virus infection was quantified by measuring the Gaussia luciferase activity in the supernatant as described. To allow comparison of virus infectivity across independent experiments, infectivity after 120 min was set to 100% and virus infectivity in all other wells were expressed in relation to this. The time to reach 50% of infection (half-maximal entry times, t1/2) was used as a surrogate for timing of CD4 receptor binding or fusion. For each Env, each inhibitor and each replicate, a general kinetic equation
(A-D)/(1+/(x/C)^B) + D
was fitted to the time series of the data points and a t1/2 value was estimated from the fitted equation. If the least-squares approximation used to fit the kinetic equation did not converge, a straight line was instead used to estimate t1/2. To deal with irregularities of the data, this line connects the data point left of the first point with >50% relative infectivity and the point right of the last point with <50% relative infectivity. The reported t1/2 value for each Env and each inhibitor is the mean t1/2 value across all replicates. To compare the time intervals between three different stages of the entry process (synchronized start, CD4 binding, fusion) among different Env, Mann-Whitney tests were performed. Only Env of the same type (PI-like or SU-like) were compared.
Infection time courses for both inhibitors were generated and the mean half-maximal time (mean t1/2) required for viruses to progress with the entry process beyond these two steps were calculated from the fitted curves (Fig 6B, S4 Fig). Time intervals between three different stages of the entry process (CD4 attachment, progress from CD4 to fusion, and fusion) among either the PI-like or SU-like Env were compared (S5 Fig).
The neutralization capacity of bnAbs at pre- and post-attachment of NLlucAM reporter viruses to A3.01-CCR5 target cells was assessed as described [47,52]
Briefly, total neutralization activity at the pre- and post-attachment stage was measured by pre-incubating NLlucAM reporter viruses yielding around 10’000 RLU per 96 well with the respective bnAbs for 1 h at 37°C. The virus-bnAb mix was then spinoculated onto 1*105 A3.01-CCR5 target cells in RPMI with 50 μM Hepes and 10 μg/ml DEAE per 96 well for 2 h at 1200 g and 23°C. The reaction was transferred to 37°C and incubated for 65 h.
To assess pre-attachment neutralization activity, samples were additionally washed after the spinoculation step to remove unbound viruses and inhibitors.
To assess inhibitory capacity at the post-CD4 attachment step, NLlucAM reporter viruses were first spinoculated onto the A3.01-CCR5 target cells and then bnAbs were added for 1 h incubation at 23°C before rising the temperature to 37°C for 65 h.
Infectivity was determined by firefly luciferase reporter production from the lysed cells as described. Samples measuring the total inhibition activity were set to 100% and pre- and post-attachment samples were expressed relative to the total activity.
We reconstructed a phylogeny based on the Env sequences of the 17 CAP256 clones included in the present study. The CAP256 Env sequences [34] were aligned to the HXB2 genome using AliView [79]. The phylogenetic analysis was performed employing the sampled ancestor model [80] in BEAST2 [81]. The maximum credibility tree was constructed with TreeAnnotator and the phylogeny displayed with FigTree. The resulting summary of the posterior distribution of phylogenies is shown in S2A Fig. The entire posterior distribution of phylogenetic trees is displayed with DensiTree [82] in S2B Fig. The xml-file of the BEAST2 analysis is summarized in S1 Dataset.
Correlation analyses according to Spearman using the untransformed data sets were performed in GraphPad Prism. Unmatched groups were compared using the nonparametric Mann-Whitney test in GraphPad Prism.
|
10.1371/journal.ppat.1007520 | GPI-anchor signal sequence influences PrPC sorting, shedding and signalling, and impacts on different pathomechanistic aspects of prion disease in mice | The cellular prion protein (PrPC) is a cell surface glycoprotein attached to the membrane by a glycosylphosphatidylinositol (GPI)-anchor and plays a critical role in transmissible, neurodegenerative and fatal prion diseases. Alterations in membrane attachment influence PrPC-associated signaling, and the development of prion disease, yet our knowledge of the role of the GPI-anchor in localization, processing, and function of PrPC in vivo is limited We exchanged the PrPC GPI-anchor signal sequence of for that of Thy-1 (PrPCGPIThy-1) in cells and mice. We show that this modifies the GPI-anchor composition, which then lacks sialic acid, and that PrPCGPIThy-1 is preferentially localized in axons and is less prone to proteolytic shedding when compared to PrPC. Interestingly, after prion infection, mice expressing PrPCGPIThy-1 show a significant delay to terminal disease, a decrease of microglia/astrocyte activation, and altered MAPK signaling when compared to wild-type mice. Our results are the first to demonstrate in vivo, that the GPI-anchor signal sequence plays a fundamental role in the GPI-anchor composition, dictating the subcellular localization of a given protein and, in the case of PrPC, influencing the development of prion disease.
| The prion protein (PrPC) is a glycoprotein attached to the neuronal surface via a GPI-anchor. When misfolded to PrPSc, it leads to fatal neurodegenerative diseases which propagates from host to host. PrPSc is the principal component of the infectious agent of prion diseases, the “prion”. Misfolding occurs at the plasma membrane, and when PrPC lacks the GPI-anchor, neuropathology and incubation time of prion disease are strongly modified. Moreover, the composition of the PrPC GPI-anchor impacts on the conversion process. To study the role of the GPI-anchor in the pathophysiology of prion diseases in vivo, we have generated transgenic mice where the PrPC GPI-signal sequence (GPI-SS) is replaced for the one of Thy-1, a neuronal protein with a distinct GPI-anchor and membrane localization. We found that the resulting protein, PrPCGPIThy-1, shows a different GPI-anchor composition, increased axonal localization, and reduced enzymatic shedding. After prion infection, disease progression is significantly delayed, and the neuropathology and cellular signaling are changed.
The present work demonstrates that the GPI-SS per se determines the GPI-anchor composition and localization of a given protein and it stresses the importance of PrPC membrane anchorage in prion disease.
| The cellular prion protein (PrPC) is a cell surface GPI-anchored protein (GPI-AP) with two putative N-glycosylation sites [1, 2] targeted to detergent-resistant membranes (DRMs; or lipid rafts)[3]. All GPI-APs share a common GPI-anchor core structure which is highly conserved among species and consists of ethanolamine phosphate linked through an amide group to the carboxyl terminus of the protein, three mannose residues, glucosamine, and a phosphatidylinositol (PI) group. During biosynthesis, the signal sequence for a GPI-anchor (GPI-SS) is rapidly removed after ER translocation, and a common GPI-anchor core is attached to the protein via a GPI-transamidase. Once attached to the protein, this core undergoes several modification steps during ER and Golgi transport involving both elimination/addition of side branching sugars to the glycan moiety together with fatty acid remodeling [4, 5]. How these side chains and lipid moieties are chosen, and how this, in turn, affects the trafficking of GPI-APs is unclear, but it is probably cell- and species-specific and may depend on the functional context [6]. It could also be, that the GPI-SS itself influences GPI-anchor remodeling [7] which could then affect its intracellular sorting [3, 4]. In polarized Madin-Darby canine kidney (MDCK) cells, N-glycosylated GPI-APs are mostly apically sorted, suggesting that the GPI-anchor is an apical sorting signal [8], yet PrPC is an exception in this regard being basolaterally sorted in these cells [9, 10]. When the GPI-SS of PrPC is replaced by the one of the DRM-resident Thy-1, PrPC partially relocates from the basolateral to the apical side in MDCK cells [11]. Likewise, GPI-SS-dependent relocalization of EGFP-tagged proteins occurs in cultured cells [12].
Most recently, Bate et al. demonstrated that sialylation of the PrPC GPI-anchor plays a role in its synaptic targeting [13]. Although it has already been shown that the GPI-SS influences intracellular sorting in vitro, how differences in the GPI-SS impact on the GPI-anchor composition in neuronal cells or in the brain, has not been investigated yet.
A key event in the pathogenesis of prion diseases is a (templated) conformational change of PrPC to its misfolded isoform (PrPSc), the critical component of prion infectivity [14]. Neuropathological characteristics of prion diseases include vacuolization of neuropil and white matter, astro- and microgliosis, neuronal loss and PrPSc deposition [15]. Of outstanding importance for the pathogenesis of prion diseases are the membrane attachment and lipid raft localization of PrPC [16]. Thus, cells expressing GPI-anchorless PrP cannot be infected with prions [17], and cholesterol depletion or expression of a transmembrane PrP-CD4 fusion protein (shifting PrP-CD4 out of DRMs) in cells, interferes with prion propagation [18]. Furthermore, when cell membranes are treated with analogs of GPI-anchors, such as glucosamine-PI, the membrane composition is altered, and the formation of PrPSc is reduced, probably by displacing PrPC from lipid rafts [19]. Along the same line, enzymatic elimination of one GPI-anchor acyl chain in cells removes PrPC from DRMs and reduces PrPSc amounts [20]. Also, mice expressing GPI-anchorless PrPC show delayed clinical onset of prion disease together with altered clinical and neuropathological presentation [21, 22].
Interestingly, the GPI-anchor of PrPC bears a sialic acid, a rare modification for a mammalian GPI-AP [23]. PrPC lacking sialic acid in its GPI-anchor cannot convert to PrPSc in vitro [13, 24]. Sialo and asialo GPI-anchored PrPSc are equally present in infected mice [25]. Therefore, the contribution of the GPI-anchor sialic acid modification to the conversion of PrPC to PrPSc clearly needs further investigations.
In the present study and based on our previous results [11], we assessed how the substitution of the GPI-SS of PrPC for that of Thy-1 influences the biology of the resulting chimeric PrPCGPIThy-1 in vivo, and how this impacts on the pathophysiology of prion disease. We provide evidence that this replacement alters the resulting GPI-anchor composition regarding sialic acid content and leads to relocalization of PrPCGPIThy-1, increasing its presence in axons when compared to wild-type PrPC (WTPrPC). This is accompanied by decreased proteolytic shedding of the protein. After intracerebral challenge with mouse-adapted prions, incubation time to clinical prion disease is extended, correlating with decreased activation of the mitogen-activated protein kinase ERK and reduced glial activation.
We generated four lines of transgenic mice (PrPCGPIThy-1) expressing PrPC with the GPI-SS of Thy-1 (Fig 1A) which were backcrossed into PrP knockout (Prnp0/0 in C57/Bl6 background) mice. Two of the lines, L27 and L16, had the 3F4 tag [26]. However, since this modification can increase prion disease incubation time [27], we also generated two other lines, L150 and L159, without the 3F4 tag, of which L150 was used for prion inoculation experiments. A WTPrPC line was generated with littermates of the founders, which did not contain the transgene but had an identical genetic background. WTPrPC mice were backcrossed with C57/Bl6 mice to avoid differences in prion incubation times depending on the genetic background. All lines described here developed normally for the period observed (>350 days) and did not present any obvious phenotypic alterations.
Analyses of PrPC expression by RT-qPCR and western blot (Fig 1B and 1C) showed similar PrPC-levels for line L27 and a two-fold increase of PrPC levels for line L16 (S1A Fig) when compared to wild-type mice (WTPrPC). Both lines of transgenic mice also showed a similar pattern of expression in different organs as WTPrPC (S1B Fig). L27 was used for the detailed biochemical characterization of PrPCGPIThy1 in mice (Figs 1 and 2).
To assess whether a proper GPI-anchor was added, we performed a Triton X-114 phase separation assay [28]. We found that PrPCGPIThy-1, like WTPrPC, is mainly present in the insoluble pellet, indicating GPI-anchorage (Fig 1D). Thy-1 and PrPC are both DRM residents, but share different lipid domains at the plasma membrane that can, in principle, be isolated by differential detergent solubilization [29]. We hypothesized that the exchange of the GPI-SS of PrPC for the one of Thy-1 would alter the lipid subdomain localization of the protein. We isolated lipid rafts from frontal cortex using Brij 96 (0.5%) and sodium deoxycholate (NaDOC, 0.5%), as these detergents were described to discriminate between lipid subdomains [30]. As shown in S2A Fig, both WTPrPC and PrPCGPIThy-1 were mainly solubilized and present at the bottom of the gradient (fractions 11 and 12), whereas flotillin, a marker of lipid rafts, and Thy-1 were found in the upper fractions of the gradient under these conditions (S2B Fig). We myelin-depleted the sample before detergent incubation, as myelin interferes with proper detergent solubilization. After solubilization with either a Brij 96 (0.5%)/NaDOC (0.5%) mixture or with Brij 98 (1%) as previously described [31], we obtained similar results, with the majority of WTPrPC and PrPCGPIThy-1 solubilized and present at the bottom of the gradient (S2C Fig).
Since the PrPC GPI-anchor contains sialic acid and the Thy-1 GPI-anchor does not, we assessed sialic acid content of both GPI-anchors. WTPrPC and PrPCGPIThy-1 brain homogenates were first deglycosylated (as the N-glycans of PrPC also contain sialic acid), immunoprecipitated with the PrPC-directed POM1 antibody (Fig 1E) and, after proteinase K (PK) digestion, their GPI-anchors were dot-blotted. As shown in Fig 1F, the amounts of phosphatidylinositol (PI) and mannose (man) as controls did not differ substantially between samples, whereas the amount of sialic acid was drastically reduced in PrPCGPIThy-1, showing a GPI-anchor composition similar to Thy-1 (S3 Fig). Thus, although the sublipid domain occupancy was not altered, the composition of the GPI-anchor was indeed changed. This suggests that the GPI-SS dictates the composition of the resulting GPI-anchor.
As we have described [11], PrPCGPIThy-1, in contrast to WTPrPC, is mainly sorted to the apical compartment when expressed in a polarized model of epithelial cells, thus behaving like Thy-1 (Fig 2A). In neurons, PrPC is mainly present in the somatodendritic compartment whereas Thy-1 is more uniformly present in cell bodies and axons [29].
We could observe that, under non-permeabilizing conditions, PrPCGPIThy-1 was found at the plasma membrane similarly to WTPrPC, indicating that GPI-SS does not act on overall plasma membrane localization (Fig 2B). Next, we evaluated the presence of PrPC in axons of primary hippocampal neurons isolated from PrPCGPIThy-1 and WTPrPC mice (Fig 2C). PrPCGPIThy-1 shows a significantly higher degree of localization in tau-positive axons than WTPrPC when assessed in primary neurons (100 ± 6.5% for WTPrPC vs. 176 ± 8.8% for PrPCGPIThy-1; ***p = 0.0001; unpaired t-test).
As stated in material and methods, for prion inoculation experiments we generated new lines of transgenic mice lacking the 3F4 tag (S1 Table). As shown in Fig 3A and S4 Fig, we obtained two lines, PrPCGPIThy-1 L150 and PrPCGPIThy-1 L159, with different expression levels of the transgene. We chose L150 for prion infection as these mice expressed amounts of PrPCGPIThy-1 equal to endogenous PrPC levels in WTPrPC mice (Fig 3A and S4A Fig). As shown in Fig 3B, terminal prion disease in transgenic animals inoculated with mouse-adapted RML prions occurred at 195 ± 2 days post-infection (dpi; SEM; n = 10), thus showing a significant delay compared to WTPrPC mice (155 ± 1.6 dpi (SEM; n = 8); Log Rank (Mantel-Cox) ****p<0.0001). The delay was independent of the prion strain as PrPCGPIThy-1 L150 mice inoculated with another strain (22L; S5 Fig), also presented with a significant delay in incubation time (156 ± 3 dpi; (n = 5) compared to 144 dpi in WTPrPC (n = 5); Log Rank (Mantel-Cox) **p = 0.003). Interestingly, transgenic RML-infected mice, although showing a delay to terminal disease, presented a more rapid disease progression after clinical onset than the WT mice (duration of clinical phase: 20 ± 2 dpi in PrPCGPIThy-1 L150, compared to 40 ± 2 dpi in WT [32]).
Sagittal brain sections of terminally prion-diseased WTPrPC and PrPCGPIThy-1 L150 mice were neuropathologically examined. As shown in Fig 3C, RML prion-infected PrPCGPIThy-1 L150 mice presented with decreased spongiosis and less severe astro- and microgliosis when compared to WTPrPC. By analyzing the lesion profile (as described elsewhere [33], Fig 3D), we found an overall decrease in the severity of prion-associated lesions, namely less spongiosis in all the studied areas, reduced microgliosis (less pronounced in the cerebellum) and astrogliosis (although with similar amounts in pons).
In another set of experiments, we also inoculated PrPCGPIThy-1 L16 mice with RML and 22L mouse-adapted prions. RML-inoculated transgenic mice reached terminal disease after 400 ± 56 dpi (n = 10), whereas WTPrPC mice became terminally sick at 155 ± 1.6 dpi (n = 8), thus also showing a highly significant delay for PrPCGPIThy-1-expressing mice (Log Rank (Mantel-Cox)****p<0.0001; S6A Fig). Upon inoculation with 22L prions, PrPCGPIThy-1 L16 mice again presented with a significant delay (200 ± 23 dpi; n = 5) compared to controls (144 ± 1 dpi; n = 5; Log Rank (Mantel-Cox) **p<0.003; S6A Fig). Although this drastic delay to terminal disease in PrPCGPIThy-1 L16 mice may partially be explained by the presence of the 3F4 tag, decreased lesion severity in these mice was conspicuously similar to the infected PrPCGPIThy-1 L150 mice lacking the 3F4 tag, with an overall decrease in spongiosis and glial activation (S6B Fig).
To investigate the amount and type of PrPSc, we performed biochemical analyses of brain samples from terminally prion-diseased mice. Interestingly, PrPCGPIThy-1 L150 mice infected with RML showed decreased amounts of total PrP (Fig 4A) with significantly less PrPSc compared to WTPrPC mice (100 ± 14.26% for WTPrPC mice vs. 29.4 ± 2% for PrPCGPIThy-1 L150 mice; Fig 4B and 4C, **p = 0.0021; unpaired t-test). Likewise, RML-infected PrPCGPIThy-1 L16 mice also had significantly less PK-resistant PrPSc (S6C Fig). The reduction in total PrP was not due to an age-dependent decrease in PrPC expression as we did not observe significant differences between 20 and 40 weeks-old PrPCGPIThy-1 L150 mice (S7A Fig). Moreover, no significant differences in total PrP levels were observed between terminally prion-diseased (around 30 weeks old) and non-infected PrPCGPIThy-1 L150 mice (40 weeks old; S7B(ii) Fig). Taken together, these findings suggest that the relative decrease of total PrP after infection in PrPCGPIThy-1 is due to less efficient conversion to PrPSc in these mice compared to WTPrPC. In addition, the PrPSc glycopattern was also changed in prion-diseased PrPCGPIThy-1 L150 mice, which showed significantly less monoglycosylated PrPSc (Fig 4D, 42.9 ± 1.9% in WTPrPC vs. 27.5 ± 3.6% in PrPCGPIThy-1 L150; *p = 0.0173, unpaired t-test).
PrPC undergoes several physiological cleavages which are highly conserved through evolution, indicating essential functions [34]. We have previously shown that the metalloprotease ADAM10 is the responsible protease for the shedding of PrPC in vivo [35] and that impaired shedding has significant consequences for prion diseases [32]. Thus, to investigate the possible mechanisms implicated in the clinical delay and the altered neuropathology, we decided to analyze the shedding of PrPCGPIThy-1. We have recently generated an antibody (sPrPG228) that specifically recognizes shed PrP (as the antibody is directed against the carboxy terminus Gly228 only exposed upon ADAM10-mediated cleavage) allowing for the direct detection of shed PrP in mouse brain homogenates [36]. By using this antibody, we found a significantly decreased shedding in PrPCGPIThy-1 L150 and PrPCGPIThy-1 L159 mice compared to WTPrPC (Fig 5A and 5B; 100 ± 72% in WTPrPC vs. 19.37 ± 3.5% in PrPCGPIThy-1 L150; ***p = 0.0005; S8A and S8B Fig, 100 ± 14.8% in WTPrPC vs. 37.5 ± 3.9% in PrPCGPIThy-1 L159; *p = 0.015; unpaired t-test). This indicates that the altered GPI-anchor and/or localization of PrPCGPIThy-1 interferes with the ADAM10-mediated release from the plasma membrane. Remarkably, in terminally diseased RML-infected PrPCGPIThy-1 mice, even though total PrP shedding is decreased compared to WTPrPC (Fig 5C and 5D, 100 ± 6.4% in WTPrPC vs. 37.77 ± 3.9% in PrPCGPIThy-1 L150; ***p<0.0001; unpaired t-test), it does not show differences when it is referred to total PrP amounts (Fig 5E). This indicates a relative increase of PrPCGPIThy-1shedding upon infection, which was further confirmed by comparing levels of shed PrPCGPIThy-1 between non-infected and infected mice (S7B(i) Fig; 100 ± 11.8% in PrPCGPIThy-1 L150 vs. 272 ± 21% in PrPCGPIThy-1 L150 RML infected; ***p = 0.0004; unpaired t-test). We also observed that in brains of both, RML-infected WTPrPC and PrPCGPIThy-1 L150 mice, all PrP glycoforms could be shed (Fig 5C), contrasting with the almost exclusive shedding of diglycosylated PrPC in non-infected samples (Fig 5A and S7 Fig, also characterized in [36]). In PrPCGPIThy-1 L150 mice, the monoglycosylated isoform is significantly less shed compared to WTPrPC (Fig 5F, 25.08 ± 1.6% in WTPrPC vs. 15 ± 2.5% in PrPCGPIThy-1 L150; *p = 0.0158; unpaired t-test), possibly reflecting its decreased presence in these infected brains as shown in Fig 4D. In conclusion, ADAM10-mediated shedding is significantly altered for PrPCGPIThy-1 compared to WTPrPC under both physiological and pathological conditions. In order to properly interpret these results, one should keep in mind that shed PrP (made visible by the highly sensitive sPrPG228 antibody) represents only a minor fraction of the total PrP pool (repeated own observations and [37]).
Activation of ERK1/2, a member of the mitogen-activated protein kinase (MAPK) family, can promote the formation of PrPSc in a cell model of prion infection [38]. In the same study, it was demonstrated that activation of other MAPKs, such as p38 or JNK, has an adverse effect on the formation of PrPSc. We have recently shown that a C-terminal deletion retains PrP in the secretory pathway, leading to p38 activation and neuronal death [39]. Since in the present study RML-infected PrPCGPIThy-1 L150 mice showed decreased PrPSc formation and increased survival, we performed western blot analyses to assess the status of ERK1/2 and p38 MAPK phosphorylation. As shown in Fig 6A and 6B, clinically terminal PrPCGPIThy-1 L150 mice presented a significant decrease of phosphorylated ERK1/2 (*p = 0.031) when compared to WTPrPC. Phosphorylation of p38 instead was not significantly changed at terminal disease.
Since MAPK signaling was changed, and to assess whether PrPCGPIThy-1 is capable of mediating pro-apoptotic signaling induced by PrPSc, we employed a cell culture model [40]. This assay is based on the co-cultivation of PrPC-expressing cells with chronically prion-infected cells that release PrPSc into the cell culture medium. As illustrated in Fig 7A, co-cultivation of SH-SY5Y cells expressing PrPC with prion-infected mouse neuroblastoma (ScN2a) cells increased apoptotic cell death, as determined by activation of caspase-3. In contrast, SH-SY5Y cells expressing a PrP mutant containing a heterologous C-terminal transmembrane domain instead of the GPI-anchor (PrP-CD4) could be co-cultured with ScN2a cells without signs of apoptosis. Notably, PrP-CD4 also inhibits PrPSc-formation in scrapie-infected neuroblastoma cells [18]. However, SH-SY5Y cells expressing PrPCGPIThy-1 transduce PrPSc-mediated toxicity similar to WTPrPC. This supports that, despite an altered GPI-anchor composition and localization at the cell surface, PrPCGPIThy-1 is still able to contribute to the formation of signaling-competent complexes relevant to prion diseases.
The GPI-anchor is a complex structure for the attachment of proteins to the outer leaflet of the plasma membrane that involves more than 20 proteins in its production [5]. It bestows the capacity to localize in lipid rafts (determining compartmentalization), and although it does not reach the intracellular space, it can confer the ability of signal transduction through transmembrane spanning partners [41, 42]. The signal sequence for GPI-anchor attachment has a wide range of sequence diversity and its participation in determining the final composition of the GPI-anchor itself is not known. It is assumed that remodeling of the core GPI-anchor depends on the protein it is attached to, and on the cell type by which it is synthesized [42]. In the present study, we have made several novel observations important not only for prion diseases but the biology of GPI-APs. We could show in a newly generated transgenic mouse model that (i) by changing the amino acids of the GPI-SS, the composition of the GPI-anchor is modified, resulting in a differential sorting and proteolytic processing of the protein. Upon infection with prions, these alterations (ii) are associated with significantly prolonged prion disease incubation time, and (iii) influence the conversion to PrPSc as well as the neuropathological presentation including decreased gliosis and spongiosis. Fittingly, we also found (iv) reduced activation of the ERK signal cascade during prion disease (please refer to Fig 8 for a graphical summary of the principal findings). To our knowledge, this study is the first to demonstrate the importance of the GPI-SS in determining a GPI-anchor composition and sorting of a GPI-AP in a mouse model. In the specific case of PrPC, this has a direct link to the pathophysiology of prion disease.
Our previous observation in a model of polarized epithelial cells already indicated that the GPI-SS plays a role in protein sorting. We observed that, by changing the GPI-SS of PrPC for the one of Thy-1, PrPCGPIThy-1 was partially, yet significantly, relocalized from the basolateral to the apical side [11], confirming previous observations from others [12]. In the present work, we extended our observations to transgenic mice expressing PrPCGPIThy-1 on a Prnp knock-out background. PrPCGPIThy-1 is correctly expressed at the plasma membrane, showing identical N-glycosylation as WTPrPC, and is located in lipid rafts. In primary hippocampal neurons, we showed a 1.8-fold increase of PrPCGPIThy-1 in axons when compared to controls, despite equal PrPC expression levels. Notably, we showed that by changing the GPI-SS, the biochemical composition of the GPI-anchor is altered, presenting with a loss of the sialic acid usually present in the GPI-anchor of PrPC. It has recently been shown that the sialic acid in the GPI-anchor of PrPC has a role in the synaptic targeting of PrPC in cultured cells [13]. Thus, when sialic acid from exogenously administered GPI-anchored PrP is depleted, or the lipid moiety is changed, PrPC still partitions in DRMs but is no longer targeted to synapses. In vivo, we also observed that PrPCGPIThy-1 lacking sialic acid is differently sorted, but we observed an increase in axonal targeting. The fact, that not only the sialic acid is missing but also the GPI-SS was changed for PrPCGPIThy-1, may account for these differences.
The fact that in the GPI-anchor of PrPCGPIThy-1 not only the sialic acid is missing but also the SS-GPI anchor is changed, can account for these differences. Our mouse model provides the opportunity to address this on the molecular level.
In mammals, the sorting of GPI-APs is complex, and in epithelial cell models it appears to depend on (i) the presence of the GPI-anchor itself, (ii) its remodelling in the trans-Golgi network (TGN), (iii) its partitioning in DRMs, and (iv) the capacity to oligomerize in the Golgi, which in turn depends on the clustering of DRMs among others [7, 43]. In neurons, the signals implicated in the targeting of proteins either exclusively to dendrites/axons or to both sides, are not clear. For GPI-APs, such as Thy-1, it has been suggested that early interaction with DRM components (especially with sphingomyelin and cholesterol) is necessary for axonal delivery of Thy-1 in mature neurons [44]. According to our data and recent findings of others, it could be hypothesized that the GPI-SS itself directs the protein to different lipid environments, which then influence how the GPI-anchor will be remodeled in the TGN, thereby affecting its oligomerization capacity and, hence, its further sorting [7, 45]. Actually, PrPC and Thy-1 are both located in DRMs but occupy different subdomains therein [30]. However, even after using different detergents previously reported to allow for discrimination between PrPC- and Thy-1-enriched DRM subdomains [29, 31], we were unable to observe differences in the distribution between PrPCGPIThy-1 and WTPrPC.
Nonetheless, support for a relocalization of PrPCGPIThy-1 to a different subdomain comes from our finding of reduced ADAM10-mediated shedding. This proteolytic cleavage at the cell surface is likely to occur at the interface between raft and non-raft regions, where–supported by the ability of PrPC to leave and re-enter DRMs [46]–protease and substrate are thought to interact [34]. In that regard, the in vivo data presented here complement our recent finding of significantly decreased shedding of PrPCGPIThy-1 in N2a cells [36]. Relocalization of the prion protein from the periphery of rafts towards more central regions (where Thy-1 resides [29]) could, therefore, explain the reduced shedding in PrPCGPIThy-1 mice. It is interesting that, despite reduced shedding compared to WTPrPC under normal conditions, upon prion infection this cleavage is relatively increased in PrPCGPIThy-1 mice. We and others have shown that shedding interferes with PrPSc formation and beneficially influences prion disease incubation times [32, 47, 48]. Thus, upregulation of shedding during prion disease could represent a protective feedback mechanism to lower PrPSc production and prolong survival in our transgenic mice.
Our further aim was to elucidate how the differential localization of PrPCGPIThy-1 impacts on prion disease. We intracerebrally inoculated our transgenic and WTPrPC mice with prions. Upon RML-infection, PrPCGPIThy-1 mice showed (i) delayed onset of terminal disease, (ii) different neuropathological presentation with decreased spongiosis, gliosis, reduced PrPSc amounts, and an altered PrPSc-glycotype pattern, (iii) relatively increased shedding, and (iv) a decrease in ERK phosphorylation. Although, as demonstrated by others [27, 49, 50] and now confirmed by us, the 3F4 tag prolongs prion disease incubation times, we included these mice (PrPCGPIThy-1 L16) as they showed several neuropathological characteristics that were likewise present in the non-3F4-tagged transgenic mice (PrPCGPIThy-1 L150), thus strengthening our overall results.
It is known that not only the presence of PrPC but also its type of membrane anchorage is fundamental for the pathophysiology of prion disease. Briefly, (i) presence of extracellular PrPSc does not lead to prion-induced neurodegeneration in the absence of PrPC on neurons [51, 52], (ii) cells expressing either anchorless PrP or PrP with a transmembrane domain instead of the GPI-anchor are resistant to prion infection [17, 18, 53], and (iii) prion-infected transgenic mice expressing low amounts of anchorless PrP do not show clinical symptoms of prion disease despite high titers of infectivity and high levels of PrPSc [21, 54]. When the expression of anchorless PrP is increased, this leads to delayed onset of disease and generation of a new prion strain [22, 55]. Remarkably, the type of GPI-anchor also affects PrPSc formation. Thus, when amino acids in the C-terminus or within the GPI-SS of the murine sequence are replaced by their cunicular homologs (with rabbits being naturally resistant to prion infection), this modified PrPC cannot be converted to PrPSc in a cell culture model [56].
Moreover, the presence of sialic acid in the GPI-anchor can influence conversion as PrP with a desialylated GPI-anchor cannot be converted to PrPSc and can even stop an ongoing infection in vitro [57]. Thus, a desialylated GPI-anchor may modify the lipid environment leading to inhibition of the signal transduction associated with PrPSc neurotoxicity. Fitting to this, our model with a PrP mutant lacking sialic acid in its GPI-anchor also shows delay to terminal disease as well as altered signal transduction when compared to WTPrPC, highlighting the importance of the sialic acid modification of the GPI-anchor in prion disease. However, in vivo we found conversion of PrPCGPIThy-1 to be reduced, not completely abolished.
As stated above, the sialylation status of the PrPC GPI-anchor modifies the lipid environment in vitro [58]. PrPC and Thy-1 share different lipid environments, and it has been shown that the lipid environment composition is fundamental for conversion of PrPC to PrPSc [59, 60]. PrPC-enriched lipid rafts isolated from rat brain have a significant increase in cholesterol compared to Thy-1-containing lipid rafts [30]. In vitro, desialylated PrPC is found in lipid rafts with increased cholesterol content, probably stabilizing PrPC in lipid rafts and increasing its half-life at the plasma membrane [58]. Although, by performing differential detergent extraction, we could not detect a differential distribution of PrPC/PrPCGPIThy-1 in lipid rafts, we cannot rule out that the lipid environment of PrPCGPIThy-1 is altered in a way that impairs the conversion to PrPSc.
In prion-infected mice both, sialo and asialo GPI-anchored PrPC forms, can be converted to PrPSc and are present in infected brains and spleens [61, 62]. Our results, where delay to clinical disease is associated with a significant decrease in the amount of PK-resistant PrPSc, suggest that asialo forms can be converted in vivo but with comparably low efficiency. Interestingly, Katorcha et al. [63] have found that a decrease in PrP N-glycan sialylation leads to a different glycoform pattern after prion infection. They also observed that when prion-infected desialylated brain homogenates where inoculated to Syrian hamsters, there was a drop in infectivity and PK-resistance. Because in their experiments they used a sialidase that can also eliminate the sialic acid of the GPI-anchor [25], and in view of our present results, it would be interesting to study if the sialic acid of the PrP GPI-anchor also partially contributes to the observations of Katorcha et al.
Several other aspects may also contribute to reduced PrPSc conversion and prolonged survival of our PrPCGPIThy-1 mice. On the one hand, Nemoto et al. [64] have recently described that the behavior of Thy-1 and PrPC at the plasma membrane is different, with PrPC showing a slower membrane diffusion compared to Thy-1. This may increase the probability of homophilic interactions for WTPrPC but not for PrPCGPIThy-1. Since the conversion of PrPC to PrPSc occurs at the plasma membrane [65], this effect could favor interactions between WTPrPC and critical PrPSc seeds, thus enhancing the propagation of PrPSc. In contrast, the lateral diffusion behavior of PrPCGPIThy-1 may be more similar to that of Thy-1, thus decreasing the conversion rate. On the other hand, the GPI-anchor itself influences the structure of a given protein [41]. In the case of PrPC, this could result in structural hindrance and, thus, less efficient PrPC to PrPSc conversion.
PrPCGPIThy-1 mice presented with significantly increased incubation times with two different prion strains, RML and 22L (with longest delay to terminal disease with RML prions). Several factors may account for that. On the one hand, each prion strain has different cellular tropism [66] and 22L is mainly associated to astrocytes where it mediates indirect neuronal damage. On the other hand, the fact that RML and 22L prions use different endocytic pathways to infect the cell [67], could indicate that PrPCGPIThy-1 follows an altered endocytic pathway that facilitates 22L over RML infection.
Intriguingly, after prion infection, we observed a decrease in the glial response in all PrPCGPIThy-1 lines compared to controls. Prion diseases, like other neurodegenerative diseases, are accompanied by an increase in microglia with a phagocytic phenotype and by reactive astrocytes. Microglia, the resident immune cells of the central nervous system, may act beneficial or detrimental in prion diseases [68, 69]. In the present study, a decreased amount of microglia in PrPCGPIThy-1 mice coincides with a delay to the terminal stage of prion disease, but further experiments would be needed to establish a direct correlation. Of note, a recent study has functionally linked ADAM10-mediated shedding of PrPC to inflammatory responses and monocyte recruitment to the brain [70]. So, it is conceivable that reduced levels of shed PrP (as a potential chemoattractant and activating factor) account for the impaired glial response in our mice. With regard to astrogliosis, it has been shown that astrocytes (i) accumulate PrPSc early in the disease [71], (ii) can rapidly internalize, traffic, and spread PrPSc in cell culture [72, 73], and (iii) that astrocytic PrPC supports the development of prion disease in mice [74]. Since astrogliosis is decreased in our transgenic mice, this could contribute to both, delay to terminal disease and reduced PrPSc deposition. Moreover, because destruction of the extracellular matrix due to factors released by microglia and astrocytes may lead to the vacuolization observed in prion diseases [75, 76], decreased gliosis in PrPCGPIThy-1 mice could also explain their low degree of spongiosis. Interestingly, lack of vacuolization of the gray matter was also observed in prion-infected mice expressing GPI-anchorless PrP [22].
The GPI-anchor is necessary for toxic signaling associated to PrP [40] and it seems plausible that an altered GPI-anchor and localization of PrPC affect its association with signaling-competent, membrane-spanning binding partners and, thus, signaling outcome [77]. We observed that, in vitro, PrPCGPIThy-1 is as able as WTPrPC to transduce neurotoxic signals, and, in vivo, we showed decreased ERK1/2 phosphorylation whereas p38 MAPK was unchanged at a terminal stage of disease. Using an in vitro model of prion infection, Fang et al. [78] recently showed that phosphorylation of p38 is an early event in synaptotoxicity and that PrPSc specifically targets post-synaptic PrPC. In our transgenic mice, PrPCGPIThy-1 is partially depleted from the dendritic compartment, and we observed a delay in the clinical onset. We did not detected a decrease in p38 phosphorylation but we cannot rule out the possibility that, at an earlier disease state, p38 activation in PrPCGPIThy-1 mice might be lower than in controls, thus contributing to delayed disease onset.
An increase in ERK1/2 signaling has been consistently observed in cellular [79–82], mouse [83] and hamster models of prion infection [84], and is probably related to neurodegeneration and cell death. Activated ERK1/2 also participates in the conversion of PrPC to PrPSc in vitro [38]. Accordingly, it could be hypothesized that the decrease in ERK phosphorylation in our model delays clinical disease and reduces PrPSc deposition. How exactly PrPCGPIThy-1 affects MAPK signaling deserves further studies, but it is interesting that interfering with this signaling cascade is considered a potential therapeutic option [80, 81]. However, since astrocytes represent the main cell population that increases ERK phosphorylation after prion infection [83], it is also possible that the observed reduction is due to reduced astrogliosis in our transgenic mice.
Although PrPCGPIThy-1 mice present with extended survival and delay to clinical disease, we observed a faster progression to terminal disease compared to controls. Several factors can account for this. On the one hand, the different localization of PrPCGPIThy-1 may induce alternative signaling pathways not investigated in this study and these could follow a kinetic leading to rapid disease development once a critical threshold is reached [85]. On the other hand, glia is less activated (maybe due to decreased levels of PrPSc), which may prolong survival (given less inflammatory brain injury) but could also contribute to toxicity as non-functional cells and PrPSc may not be efficiently eliminated. The final sum up of positive and negative factors may influence the complex findings observed in PrPCGPIThy-1 mice.
In conclusion, we showed that the GPI-SS influences GPI-anchor composition and localization of a GPI-AP in vivo. This study attributes a novel function to the GPI-SS in subcellular trafficking in vivo and sheds light on several molecular events underlying prion-associated neurodegeneration.
Animal experiments were approved by the Behörde für Gesundheit und Verbraucherschutz of the Freie und Hansestadt Hamburg (permit numbers 80/08, 38/07 and 84/13). All the procedures were performed under the guidelines of the animal facility of the University Medical Center Hamburg-Eppendorf and in compliance with the Guide for the Care and Use of Laboratory Animals. Mice used for prion infection were anesthetized with a mixture of Xylazin hydrochloride and Ketamine hydrochloride in 0.9% NaCl prior to intracerebral prion inoculation. To sacrifice the mice, first they were anesthetized with halothane followed by neck dislocation.
The generation of the PrPCGPIThy-1 construct was already described elsewhere [11]. To insert PrPCGPIThy-1 in the half-genomic expression vector (mPrPHGC, a generous gift from M. Groschup, Institute for Novel and Emerging Infectious Diseases at the Friedrich-Loeffler-Institut, Greifswald-Insel Riems, Germany)[86], a PmlI restriction site was inserted after the stop codon of PrPCGPIThy-1 DNA by using the following primers: F:5´-CCCAAGGAGAAACACGTGCCCTCGAGGTCCTTC-3´; R: 5´-GAAGGACCTCGAGGGCACGTGTTTCTCCTTGGG-3´ (PmlI restriction site is in bold), with the QuickChange Lightning mutagenesis kit (Stratagene). PrPC was excised from mPrPHG by AgeI and PmlI (Fast Digest, Fermentas). PrPCGPIThy-1 was cut with AgeI and PmlI and ligated into the mPrPHG. For pronuclear injection, the mPrPHG vector was cut with SalI and NotI and separated in an agarose gel. The pronuclear injection was performed at the Transgenic Mouse Facility (ZMNH, Hamburg). Positive animals for the transgene were selected by PCR with the following primers: F: 5´-ATGTGGACTGATGTCGGCCT-3´; R: 5’-CTTGGAGGAGGGAGAGGGAA-3’. Lines L27 and L16 were established, and animals were backcrossed to PrP0/0 mice (on a C57/Bl6 background). To generate the control line, a littermate not presenting the transgene was backcrossed with C57/Bl6 mice with the same backcrossing scheme as the transgenic mice.
To eliminate the 3F4 tag we used the PrPCGPIThy-1 in mPrPHGC as a template and we change the two methionine residues at position 108 and 111 for leucine and valine respectively by using QuickChange Lightning (Stratagene) and the following primers: F: 5´-CAAACCAAAAACCAACCTCAAGCATGTGGCAGGGGCTGCGGCAGC-3´; R: 5´-GCTGCCGCAGCCCCTGCCACATGCTTGAGGTTGGTTTTTGGTTTG-3´ (mutations are in bold).
RNA was extracted from mouse brain tissue using Precellys Lysing Kit (Bertin Technologies) and precooled QIAzol (Qiagen). Tissue (n = 7 and n = 5 for WTPrPC; n = 6 for PrPCGPIThy-1 L16; n = 5 for PrPCGPIThy-1 L27; n = 5 for PrPCGPIThy-1 L150) was homogenized for 30 s at 2,000 rpm in a dismembrator and subsequently centrifuged at 2,000xg for 2 min at room temperature. The supernatant was mixed with 200 μl chloroform and incubated at room temperature for 3 min following centrifugation for 15 min at 12,000xg. Total RNA was purified from the upper phase using RNeasy Mini Kit (Qiagen) according to the manufacturer’s instructions. RNA concentration and purity were determined using the NanoDrop system (Thermo Fisher Scientific). First strand cDNA was synthesized using 1 μg of total DNase-treated RNA using RevertAid H Minus First Strand cDNA Synthesis kit (Thermo Fisher Scientific). Real-time PCR reactions were performed in a volume of 10 μl, consisting of 10 ng cDNA, 2XSYBR GreenPCRMasterMix (Applied Biosystems), and 0.2 μM of each primer in Rotor Gene Q (Qiagen). For the detection of murine Prnp gene, the following primer pairs were used: 5´GGCCAAGGAGGGGGTACCCATAAT 3´ and 5´TAGTAGCGGTCCTCCCAGTCGTTGC 3´. The RPL gene (F: 5´CGGAATGGCATGATACTGAAGCC 3´; R: 5´TTGGTGTGGTATCTCACTGTAGG 3‘) was used as a reference to calculate relative expression levels of Prnp using ΔCT values.
Briefly, 8–10 weeks old WTPrPC (n = 8), PrPCGPIThy-1 L16 (n = 10) and PrPCGPIThy-1 L150 mice (n = 10) mice were anasthesized and intracerebrally inoculated with 30 μl (corresponding to 3x105 log LD50) of RML 5.0 prion inoculum. Moreover, WTPrPC (n = 5), PrPCGPIThy-1 L16 (n = 5) and PrPCGPIThy-1 L150 (n = 5) mice were inoculated with 30 μl of 10% 22L prion inoculum. Mice were checked two to three times per week for clinical signs of prion disease and daily once clinical symptoms appeared.
Frontal cortex samples were homogenized as a 10% in RIPA buffer (50 mM Tris-HCl pH8, 150 mM NaCl, 1% NP40, 0.5% Na-Deoxycholate, 0.1% SDS) containing a cocktail of protease and phosphatase inhibitors (Roche), left for 10 min on ice, and centrifuged for 5 min at 12,000xg. The protein content of the supernatant was assessed by colorimetric analysis (QuickStart Bradford 1x Dye, Biorad) following the instructions of the supplier. Samples were standardized to 1 μg/μl in 4x loading buffer (250mM Tris-HCl, 8% SDS, 40% glycerol, 20% β-mercaptoethanol, 0.008% Bromophenol Blue, pH 6.8), boiled for 5 min at 95°C and subjected to electrophoresis (20 μg per sample). Proteins were then transferred to nitrocellulose membranes (Biorad) and incubated with mouse monoclonal antibodies against PrPC (POM1, 1:2.500; A. Aguzzi, Zurich, Switzerland) and actin (1:2,000; Millipore), ERK (1:1,000), ERK-P (1:1,000) p38 (1:1,000), p38-P (1:1,000) all from Cell Signaling Technologies; and rabbit polyclonal specific for shed PrP (1:1,000 [36]). Membranes were incubated with appropriate secondary antibodies and then developed either with Pierce ECL Western Blotting Substrate or Pierce Femto (Thermo Scientific) in a CD camera imaging system (BioRad) or with Odyssey Image system (Licor) and quantified with Image Studio software (Licor).
Processing of samples, including cutting and staining of paraffin sections with HE, was performed as published [39]. For immunohistochemistry, all sections were stained using the Ventana Benchmark XT machine (Ventana, Tuscon, Arizona, USA). Deparaffinised sections were boiled for 30–60 min in CC1 solution (Ventana, Tuscon, Arizona, USA) for antigen retrieval. Primary antibodies were diluted in 5% goat serum (Dianova), 45% Tris-buffered saline pH 7.6 (TBS) and 0.1% Triton X-100 in antibody diluent solution (Zytomed, Berlin, Germany). Sections were then incubated with primary antibody POM1 (1:100), Iba1 (Dako, 1:2.000) or GFAP (Dako, 1:400) for one hour. Anti-mouse or anti-rabbit histofine Simple Stain MAX PO Universal immunoperoxidase polymer (Nichirei Biosciences) was used as secondary antibody. Detection of secondary antibodies was performed with an ultraview universal DAB detection kit from Ventana with appropriate counterstaining and sections were cover-slipped using TissueTek glove mounting media (Sakura Finetek).
The neuropathological assessment (n = 3 or n = 4 depending on the genotype) was conducted by three independent investigators in a blinded fashion by grading the samples 1 (mild) to 3 (severe) depending on the staining intensity or spongiosis.
The assay was performed as previously described by Hooper et al. [28] which is a modification of the method described earlier by Pryde and Philips [87]. All the buffers contained a protease inhibitor cocktail (Roche). Briefly, frontal cortex samples (n = 3 for each genotype) were homogenized in 10 vol. of 0.32M sucrose in 50mM HEPES/NaOH pH 7.4, centrifuged for 15 min at 8,000xg and the supernatant further centrifuged at 26,000xg for 2h. The resulting pellet was resuspended in H buffer (10mM HEPES/NaOH pH 7.4) with the addition of 2% of pre-condensed Triton X-114 (Sigma-Aldrich) in a total volume of 200 μl (final concentration of 2 μg/μl). Samples were vortex mixed for 1–2 s, let on ice for 5 min and centrifuged again at 8,880xg at 4°C in a fixed angle rotor. The resulting pellet was washed with 0.2 ml H buffer, centrifuged at 8,800xg for 10 min at 4°C and the resulting pellet was resuspended in 180 μl of H buffer. This was kept as the Insoluble Pellet. The supernatant was layered over a 0.3 ml of 6% sucrose cushion in T buffer (10mM Tris-HCl pH 7.4, 0.15M NaCl) and 0.06% precondensed Triton X-114, incubated at 30°C for 3 min and further centrifuged at 3,000xg for 3 min in a swinging bucket rotor. The sucrose cushion was removed from the pellet (Detergent Phase), and the latter was resuspended in 180 μl of H buffer.
The supernatant (upper aqueous phase) was mixed with 0.5%(v/v) pre-condensed Triton X-114, vortexed 1–2 s, kept 5 min on ice, and further layered over a 0.3 ml 6% sucrose cushion in T buffer, incubated at 30°C for 3 min and centrifuged again at 3,000xg for 3 min in a swinging bucket rotor. The pellet was discarded, and the upper phase mixed again with 2% (v/v) pre-condensed Triton X-114 and processed again as described in the step before but without the sucrose cushion. After the last centrifugation, the supernatant was kept as final aqueous phase.
An equal amount of sample was then mixed with 4x loading buffer, and 30 μl of sample was subjected to gel electrophoresis and western blot as described above.
The procedure was repeated with three times with three different samples for quantification.
GPI anchors were isolated as previously described, with few modifications [88]. Briefly, 10% mouse brain homogenates were prepared with RIPA buffer plus protease inhibitor cocktail (Roche) and protein content was determined as described before. As sialic acids in N-glycans would interfere with the analysis of GPI-anchor modifications, samples containing 100 μg of total protein were first subjected to deglycosylation for 4 h at 37°C using the PNGase F kit (New England Biolabs) according to the manufacturer`s protocol. To only yield PrPC and avoid presence of interfering IgGs in eluates, deglycosylated samples were subsequently used for immunoprecipitation using monoclonal antibody POM1 that had been covalently linked to beads following the instructions of the Pierce Co-Immunoprecipitation Kit (Thermo Scientific) before. Thy-1 was immunoprecipitated with the rat anti-mouse Thy-1 (MCA1474, Serotech). The eluted samples containing purified deglycosylated PrPC were digested with 100 μg/ml proteinase K, at 37°C for 24 hours, resulting in GPI anchors attached to the terminal amino acid. The released GPIs were extracted with water-saturated butanol, washed with water 5 times and loaded onto C18 columns. GPIs were eluted under a gradient of propanol and water. The presence of GPIs was detected by ELISA. Maxisorb immunoplates were coated with 0.5 μg/ml concanavalin A (binds mannose) and blocked with 5% milk powder. Samples were added and any bound GPI was detected by the addition of the phosphatidylinositol-reactive mAb 5AB3-11, followed by a biotinylated anti-mouse IgM (Sigma), extravidin-alkaline phosphatase and 1mg/ml 4-nitrophenyl phosphate.
The presence of phosphatidylinositol in GPI anchors was identified using mAb (5AB3-11) and specific glycans were detected with biotinylated lectins. Isolated GPI anchors were bound to nitrocellulose membranes by dot blot and blocked with 5% milk powder. Samples were incubated with mAb 5AB3-11, biotinylated SNA (detects terminal sialic acid residues bound α-2,6 or α-2,3 to galactose), biotinylated concanavalin A (detects mannose) or biotinylated RCA I (detects terminal galactose) (Vector Labs). Bound lectins were visualised using extravidin peroxidase and enhanced chemiluminescence. The mAb was visualised by incubation with a horseradish peroxidase conjugated anti-murine-IgG and chemiluminescence.
For preparation of primary hippocampal neurons, we used postnatal P0-P2 mice. Briefly, pups were killed by decapitation, and after removing skull and meninges, the hippocampus was dissected and collected in 10mM glucose in PBS containing 0.5 mg/ml papain (Sigma-Aldrich) and 10μg/ml DNAse (Roche). Hippocampi of PrPCGPIThy-1 mice were collected individually, and the tails were used for genotyping. After 30 min incubation at 37°C, samples were washed 4 times with plating medium (MEM 1X (Gibco), 20mM glucose (Sigma), 10% Horse serum (PAA Laboratories) and 3% of NaHCO3 7.5% (Gibco)) and carefully pipetted up and down several times in order to mince the tissue. Cells were plated in 6-well plates containing coverslips previously treated with 0.5 mg/ml of Poly-L-Lys and incubated at 37°C in a 5% CO2 cell culture incubator. After 4 hours, the media was changed to Neurobasal A medium (Gibco) containing 2% of B27 serum, Glutamax (Gibco) and penicillin/streptomycin (PAA Laboratories). Next day, AraC (Sigma-Aldrich) was added to kill proliferating cells. Half of the media was changed every three days.
After seven days in culture, coverslips were washed in PBS (Sigma-Aldrich), fixed (4% paraformaldehyde in PBS, 15 min room temperature (RT)), washed (PBS) and permeabilized with PBS containing 0.1% bovine serum albumin (BSA) and 0.3% Triton-X-100 for 1h at RT. Cells were then incubated with primary antibody (POM1 antibody at 1:250 and anti-tau antibody (Synaptic Systems, 1:500) diluted in PBS/0.1% BSA) for 1h at RT, washed (PBS), incubated with secondary antibody (donkey anti-mouse antibody AlexaFluor 488 and anti-guinea pig AlexaFluor 555 (Invitrogen), diluted in PBS/0.1% BSA, 1 h at RT) and washed again with PBS. DAPI (Roche) was added to the last wash and samples were then mounted with Fluoromount G media (SouthernBiotech). Consecutive Z-stacks (between 30–35 Z-stacks per picture) were taken with Leica Laser Scanner Confocal Microscope TCS SP5 (Leica), and the reconstructed 3D images were further processed with the IMARIS Software to quantify colocalization. For this, 12 pictures for the controls (from 5 animals) and 7 pictures for PrPCGPIThy-1 (from 5 animals) were analyzed. Out of every picture, ten regions of interest (ROI) were selected, and the total intensity of PrP and tau within these ROIs was measured. The ratio of PrP/tau intensity of 10 ROIs was then added to make a total mean. Total mean of all the pictures for WTPrPC and the ones from PrPCGPIThy-1 neurons were then added and subjected to statistical analysis (unpaired Student's t-Test).
DRMs were isolated as previously described [29]. Briefly, about 0.2 mg of frontal cortex was homogenized in 10 vol. of homogenization buffer (10mM Tris-HCl pH 8.2, 0.02% sodium azide and 0.32M sucrose) on ice. After centrifugation at 500xg for 5 min to pellet nuclei and cell debris, the supernatant was further centrifuged for 40 min at 18,000xg to obtain a membrane-enriched fraction. The amount of protein from this fraction was quantified, and 1–2 mg of protein were diluted with 2x detergent buffer (10mM Tris-HCl pH 8.2, 1% Brij 96 (Sigma-Aldrich), 1% sodium deoxycholate (Sigma-Aldrich)) that had previously been mixed at least for 16h. Samples were then incubated for 30 min at 4°C and mixed at a ratio of 1:1 with 80% sucrose in the detergent buffer. 35% of sucrose (8 ml) and 5% of sucrose (1 ml) in detergent buffer were layered on top. After 18 h of centrifugation at 200,000xg, 12 fractions were taken and subjected to electrophoresis and western blot as previously described. Primary antibodies for DMR characterization used in the western blot were Flotillin (1:1,000; Cell Signaling Technologies) and calnexin (1:1,000; BD Transduction). For DRM isolation of samples depleted of myelin, a modification of the protocol from Chen et al. was used [31]. Briefly, samples were homogenized with ISB buffer (10mM HEPES/KOH pH 7.6; 200mM sucrose, 50mM K acetate, 1mM Mg acetate, 1mM EGTA, 1mM DTT and protease inhibitors) and centrifuged at 5000xg for 5 min. The resulting supernatant was further centrifuged at 22.000xg for 60 min to obtain a pellet with a membrane enriched fraction. This pellet was resuspended in ½ of the ISB buffer starting volume and centrifuged at 20.000xg for 60 min again. The resulting pellet was resuspended in 1ml ISB buffer and layered on a top of a 0.83M sucrose cushion in ISB buffer, and further centrifuged at 75.000xg for 35 min. After centrifugation, a myelin sheath band was visible at the border of the 0.83M sucrose. All the sample under the myelin layer was further diluted to 0.2M sucrose and centrifuged 30.000xg for 40 min. The resulting pellet was resuspended in 200 μl of ISB buffer without sucrose, the amount of protein quantified, and equal amounts of protein were further incubated either with 2X detergent (1% Brij 96/1% sodium deoxycholate in ISB buffer without sucrose) for 30 min at 4°C or with Brij 98 (Sigma-Aldrich) for 5 min at 37°C. After incubation, samples were mixed at a ratio of 1:1 with 80% sucrose in ISB buffer and proceeded with the sucrose gradient and ultracentrifugation as described above.
Frontal cortex brain samples (n = 3 for WTPrPC; n = 4 for PrPCGPIThy-1) were homogenized 1:10 in RIPA buffer and digested with 20 μgrs/ml of PK (Roche) at 37°C for 1 h. The reaction was stopped as before. All samples were subjected to electrophoresis and western blot analysis.
Cells were cultivated and transfected as described [89]. The human SH-SY5Y cell line is a human neuroblastoma cell line (DSMZ number ACC 209). Co-cultivation experiments were done as described previously [40, 90]. In brief, SH-SY5Y cells were grown on glass cover slips and transfected with Lipofectamine (Invitrogen). 2 h after transfection cover slips were transferred into dishes containing a 90% confluent cell layer of either N2a (immortalized neuroblastoma cell line, ATCC No. Ccl 131) or chronically infected ScN2a cells (established by infecting N2a cells with an enriched preparation of prions isolated from the brains of mice infected with RML prions [91]). After 16 h of co-cultivation, apoptotic cell death was analyzed (see below).
16 h after co-cultivation, SH-SY5Y cells were fixed on glass cover slips with 3.7% paraformaldehyde for 20 min, washed and permeabilized with 0.2% Triton X-100 in PBS for 10 min at room temperature. Fixed cells were incubated with an anti-active caspase 3 antibodies overnight at 4°C, followed by incubation with the fluorescently labeled secondary antibody Alexa Fluor 555 for one hour at room temperature. Cells were then mounted onto glass slides and examined by fluorescence microscopy using a Zeiss Axiovert 200M microscope (Carl Zeiss). The numbers of cells positive for activated caspase-3 out of at least 1,000 transfected cells were determined in a blinded manner. All quantifications were based on at least three independent experiments.
IBM SPSS Statistics 22 and GraphPad Prism 5 statistic software programs were used in the statistical analysis. To assess differences between Kaplan-Meier survival curves, the Breslow test was used. For comparison between the groups in western blots and neuronal countings, unpaired Student's t-test was used. Statistical significance was considered when p-values were as follows: *p < 0.05, **p < 0.005, ***p < 0.001, ****p<0.0001. The exact p value is also given.
|
10.1371/journal.pcbi.1000234 | A Methodological Framework for the Reconstruction of Contiguous Regions of Ancestral Genomes and Its Application to Mammalian Genomes | The reconstruction of ancestral genome architectures and gene orders from homologies between extant species is a long-standing problem, considered by both cytogeneticists and bioinformaticians. A comparison of the two approaches was recently investigated and discussed in a series of papers, sometimes with diverging points of view regarding the performance of these two approaches. We describe a general methodological framework for reconstructing ancestral genome segments from conserved syntenies in extant genomes. We show that this problem, from a computational point of view, is naturally related to physical mapping of chromosomes and benefits from using combinatorial tools developed in this scope. We develop this framework into a new reconstruction method considering conserved gene clusters with similar gene content, mimicking principles used in most cytogenetic studies, although on a different kind of data. We implement and apply it to datasets of mammalian genomes. We perform intensive theoretical and experimental comparisons with other bioinformatics methods for ancestral genome segments reconstruction. We show that the method that we propose is stable and reliable: it gives convergent results using several kinds of data at different levels of resolution, and all predicted ancestral regions are well supported. The results come eventually very close to cytogenetics studies. It suggests that the comparison of methods for ancestral genome reconstruction should include the algorithmic aspects of the methods as well as the disciplinary differences in data aquisition.
| No DNA molecule is preserved after a few hundred thousand years, so inferring the DNA sequence organization of ancient living organisms beyond several million years can only be achieved by computational estimations, using the similarities and differences between chromosomes of extant species. This is the scope of “paleogenomics”, and it can help to better understand how genomes have evolved until today. We propose here a computational framework to estimate contiguous segments of ancestral chromosomes, based on techniques of physical mapping that are used to infer chromosome maps of extant species when their genome is not sequenced. This framework is not guided by possible evolutionary events such as rearrangements but only proposes ancestral genome architectures. We developed a method following this framework and applied it to mammalian genomes. We inferred ancestral chromosomal regions that are stable and well supported at different levels of resolution. These ancestral chromosomal regions agree with previous cytogenetics studies and were very probably part of the genome of the common ancestor of humans, macaca, mice, dogs, and cows, living 120 million years ago. We illustrate, through comparison with other bioinformatics methods, the importance of a formal methodological background when comparing ancestral genome architecture proposals obtained from different methods.
| The reconstruction of ancestral karyotypes and gene orders from homologies between extant species is a long-standing problem [1]. In the case of mammalian genomes, it has first been approached using cytogenetics methods [2]–[7]. The recent availability of sequenced and assembled genomes has led to the development of bioinformatics methods that address this problem at a much higher resolution, although with fewer available genomes. Such methods propose in general more detailed ancestral genome architectures than cytogenetics methods (see [8]–[12] and reviews in [13]–[15]). The comparison of the two approaches was recently investigated and discussed in a series of papers, sometimes with diverging point of views [16]–[18]. Among the bioinformatics methods that have been applied to mammalian genomes (previous works were limited to small genomes such as organellar genomes [19] or to bacterial genomes [20]), the one based on a parsimony approach in terms of evolutionary events such as reversals, translocations, fusions and fissions [8],[11], leads to results that are sometimes in disagreement with cytogenetics studies [16]. Recent results on this approach point out that the modeling of genome rearrangements probably needs further studies before it can be used for the reconstruction of ancestral genomes (see [21], or [17], where it was suggested that inferring parsimonious rearrangement scenarios is more intended to infer evolutionary dynamics characteristics, such as rearrangement rates, than ancestral genomes). Another type of approach infers ancestral genome segments, called Contiguous Ancestral Regions (CARs), from syntenic features that are conserved in extant species (the terminology is borrowed from [12]). We call this principle model-free, following [22], even if it is based on certain assumptions, like the absence of events inside a conserved synteny, which is a parsimony principle. But this terminology stresses the difference with rearrangement-based methods, which contraint the reconstruction by allowing prescribed operations that define then an evolution model. It is then less ambitious than the rearrangement-based approach as it does not propose evolutionary events, neither does it ensure that proposed CARs are ancestral whole chromosomes. However, when recently applied on mammalian genomes [12] it gave results more in agreement with cytogenetic methods, while exhibiting few other points of divergence [18].
We describe here a very general model-free framework for the reconstruction of CARs, that formalizes and generalizes the principles used in several computational [12],[22] and cytogenetics [5]–[7] studies. This framework takes as input a representation of extant genomes as sequences of homologous genomic markers (synteny blocks or orthologous genes for example). Then it decomposes into two main steps: we first compute a collection of possible ancestral syntenic groups (in general small groups of genomic markers that were possibly contiguous in the ancestral genome), each weighted according to its conservation in the extant species; from this set of possible ancestral syntenies, we group and order the considered genomic markers into one (or several alternative) set(s) of CARs, each of these sets of CARs representing a possible ancestral genome architecture. An important feature of our framework is that we propose the set of all possible genome architectures that agree with the conserved ancestral syntenies. This framework is general in the sense that both steps can be made effective in several ways. For example, during the first phase, the signal for ancestral syntenies can be defined from extant species in terms of conserved adjacencies between homologous markers as in [12] or between chromosome segments as in [5]–[7]. We propose one possible implementation of this framework, choosing as ancestral features both conserved adjacencies and gene teams [23],[24], generalizing the approach of Ma et al. [12] (where only adjacencies were considered), and mimicking the methods employed with cytogenetic data [5]–[7] (conserved chromosome segments may be formalized as gene teams). The second step, that computes CARs and ancestral genome architectures, benefits from a combinatorial framework, centered around the Consecutive Ones Problem and an ubiquitous combinatorial structure called PQ-tree [25], well known and used in physical mapping [26],[27], and recently applied in other comparative genomics problems [28],[29]; in particular, in [22],[30],[31], PQ-trees were already considered to represent ancestral genomes. In our implementation of this second step, we follow the same principle as in [12]: we extract a maximum unambiguous subset of ancestral syntenies.
We apply our method on several datasets. We first consider the case of the ancestral boreoeutherian genome using a dataset obtained from the whole genome alignments available on the UCSC Genome Bioinformatics website [32]; from these alignments, we build sets of synteny blocks at different levels of resolution (we use from 322 to 1675 homologous markers). Our experiments show that the results of our method are quite constant, in the sense that they are very similar, independently of the chosen resolution. This reinforces the impression that algorithmic aspects may have an important impact on the differences in the results of [11],[12] discussed in [16]–[18], together with the differences of data acquisition and interdisciplinarity problems [18]. Moreover, the results we obtain are very close to the ones towards which cytogenetics methods tend to converge. As these are obtained from many more species and much expertise, we take it as a validation of the framework and method we propose. We performed intensive comparisons with other computational methods, and ran our method on several published datasets. Compared to the recently published method of Ma et al. [12], we obtain sets of CARs that are less well defined, as we propose a large set of possible ancestral boreoeutherian genome architectures, instead of only one, but better supported, as any proposed adjacency or segment is supported by at least one syntenic group that is conserved in at least two extant species whose evolutionary path in a phylogenetic tree contains the wished ancestral species. We also reconstruct an ancestral ferungulate genome architecture for the the same data as [11]. On this dataset, our method and the method of Ma et al. obtain similar results. The CARs are comparable to those of the ferungulate chromosomes from e-painting studies [33] that are ancestral boreoeutherian features, while the rearrangement-based method of [11] on the same dataset gives divergent results.
In the next section, we describe the general framework and how we implemented it to design a new method for ancestral genome reconstruction. We then describe the results of our method on the considered mammalian datasets. We use our reconstruction of possible genome architectures for the boreoeutherian ancestor at several levels of resolution to assess both the internal stability of our method and the consistency of its results when compared to other published ancestral genome architectures. We compare our results to the results proposed by cytogenetic methods and by the bioinformatics method of Ma et al. [12], that received some attention recently [18] as it was the first bioinformatics method that tended to agree well with cytogenetics. We conclude by a discussion on our results and methodology and describe several possible extensions of our framework.
We now describe more precisely the two steps of the framework, together with their implementation into an effective method for reconstructing a set of CARs. We separate the general principles from the implementation details to emphasize that there are many possible implementations: the method of Ma et al. [12] is one possibility, and we also propose a variant of our method targeted at analyzing datasets with less well defined outgroups.
In this section, we first report the results of our method in reconstructing the architecture of the boreoeutherian ancestral genome from five datasets, at different levels of resolution, that we computed from whole genome alignments. Next we report results based on the original dataset used in [12] and on the ferungulate ancestral genome from the dataset of [11]. All data and results discussed in this section are available on a companion website: http://lbbe-dmz.univ-lyon1.fr/tannier/ploscb2008_supmat/.
We computed five datasets, with parameters max_gap = 100 kb and min_len = 100 kb (1675 markers), 200 kb (824 markers), 300 kb (510 markers), 400 kb (406 markers) and 500 kb (322 markers). Their coverage of the human genome goes from 2173 Mb (min_len = 100 kb) down to 1487 Mb (min_len = 500 kb).
We also analyzed the dataset of 1338 conserved segments used in [12], downloaded from the website http://www.bx.psu.edu/miller_lab/car/. It has the impressive property that these conserved segments span slightly more than 94% of the human genome based on alignments at a 50 kb resolution level. On the other hand, it considers less species, an unbalanced phylogeny (one of the branch from the ancestral node contains a single species, the dog, while the other branch contains three species, human, mouse and rat) and the segments are less well defined in the outgroups: they can be duplicated (due to ambiguous orthology signal), missing or overlapping. In order to analyze this challenging dataset, we modified our method, to handle the different combinatorial nature of segments in outgroups, and we chose to define ancestral syntenies in terms of conserved adjacencies and approximate common intervals which do not require the exact same markers content and allow for duplicated markers (see Material and Methods). This illustrates the generality of our framework: the way to define ancestral syntenies and the type of dataset is flexible. While we prefer to present the results with our own dataset due to its better proximity to the C1P property, we performed our method on this dataset for the method comparison to be as exhaustive as possible.
The set of possible ancestral syntenies contains 2515 subsets of segments, and 208 needed to be discarded in order to clear all ambiguities and get the C1P property. This shows that by relaxing the definition of ancestral synteny by allowing inexact content, we introduced a large number (at least 10%) of false positives (i.e. groups of segments which were not consecutive in the ancestral genome). We obtained an ancestral genome with 35 CARs, 1281 adjacencies and the following human chromosomal associations: 3-21, 4-8, 12-22, 12-22, 14-15, to compare to 29 CARs and 1309 adjacencies and the same human chromosomal associations in [12]. Among our 1281 adjacencies, 1077 are present in the 1309 adjacencies obtained with the method of Ma et al.. As before, we define a weak adjacency as an adjacency obtained by the method of Ma et al. whose segments are not included in any of our ancestral syntenies: 8 of the 1309 adjacencies obtained in [12] are weak. Among these adjacencies are several human or rodent or dog specific adjacencies. The fact that we have significantly fewer common adjacencies while the adjacencies of Ma et al. are still well supported can be explained by the fact that some adjacencies inferred in [12] are supported by false positive ancestral syntenies, which are much more frequent with this dataset than when using or own datasets of universal markers, where we used several filters to eliminate them. For example, by assessing the support of the adjacencies in the 29 CARs obtained by Ma et al. in terms of the ancestral syntenies conserved after our second phase, which produces a C1P matrix, 21 are not supported, and the general level of support of adjacencies decreases in general.
We also tested our framework on the ferungulate ancestor based on the dataset of Murphy et al. [11]. This dataset contains seven genomes, which are represented by 307 synteny blocks that cover 1343 Mb of the human genome [11]. It is hazardous to reconstruct boreoeutherian ancestors with this dataset, because there is no outgroup for the boreoeutherian clade here, but it is interesting to use this dataset to compare several methods on a dataset we did not construct. We ran both our method and the one of Ma et al. [12] on this dataset and compared the inferred genome architectures. We include in the comparison the results obtained by Murphy et al. [11] on the same dataset, and those of Kemkemer et al. [33] obtained independently by a computational method called e-painting, see Table 3. The ancestral genome architecture we propose is based on 457 ancestral syntenies from an initial number of 461, and here again the dataset seems to contain very little ambiguity.
Some syntenies obtained belong to the boreoeutherian ancestor, and others are ferungulate specific. The synteny between human chromosomes 5 and 19 is inferred only by Murphy et al. (where it is not marked as weak, which means that it was found in all alternative genome architectures) but not by our method. However, it is due to an adjacency between two synteny blocks that is not found in any of the ancestral syntenies we detected in the first step of our method, and is found only in the pig genome. The synteny between human chromosomes 1 and 22 is inferred only by Murphy et al., where it is marked as weak. It is due to an adjacency that is not found in any genome, nor supported by any of our ancestral syntenies. The same holds for the synteny between human chromosomes 2 and 20 (which is not weak according to Murphy et al.), and seems to be more rodent-specific. The synteny between human chromosomes 1 and 10 was inferred by MGR and our method, and considered weak by Murphy et al., and is supported by three of our ancestral syntenies that have significant weights. The synteny between human chromosomes 2 and 7, which is found only by the method of Ma et al. is due to an adjacency that is found only in the pig and is not supported by any of our ancestral syntenies. We can also note that among the 250 adjacencies inferred by our method, only 196 are common with the results obtained with the methods of Ma et al. and Murphy et al., while 240 are common with the ancestor obtained with the method of Ma et al. and 204 are common with the ancestor proposed by Murphy et al. We have only the boreoeutherian syntenies in common with Kemkemer et al. [33], and those that are supposed to be ferungulate specific all disagree (we don't recover the giant chromosome 1-19-3-21, and recover 1-10 instead).
We proposed a general model-free framework for reconstructing ancestral genome architectures from current genomic marker orders. We implemented this framework in a method that considers adjacencies and common intervals in extant genomes and applied our method on two ancestral genome reconstruction problems: the boreoeutherian ancestor, from a set of homologous markers we computed from UCSC whole genome alignments [32] and a dataset proposed in [12], and the ferungulate ancestor from the synteny blocks defined in [11]. We believe that our experimental results mark a progress as compared to previous bioinformatics studies, and that the framework we propose is a useful tool to compare methods.
We perform here a comparative analysis of different methods for the reconstruction of ancestral genomes, independently of the type of data used for these reconstructions. For the boreoeutherian ancestor, Ma et al. [12], with their own set of markers called conserved segments, recovered 29 CARs, with 8 “weak adjacencies”. Those adjacencies correspond to features that are only present in human and mouse for example, which would more account for an euarchontoglire feature, or even only in human (as the junction of both parts of human chromosomes 10 or 16 for example). In contrast, at a resolution of 200 kb and with universal synteny blocks, we infer 26 CARs, which is comparable, but no such weakly supported adjacency is inferred. At the resolution of 50 kb, with Ma et al. data, we infer 35 CARs, which compares to 29 CARs plus 8 weak adjacencies. Moreover, all our chromosomal syntenies, at several resolution levels, are also supported by cytogenetic studies, but the fusion of a synteny block of human chromosome 4 with a segment of human chromosome 1 that is found only at high resolution (min_len = 100 kb). The method of Ma et al. gives 31 to 37 CARs on our datasets, with a significant number of weak adjacencies, as well as some variations in terms of human chromosomal associations. The most likely explanation for the difference between the two methods lies in methodological reasons, primarily the way ancestral syntenies are defined (adjacencies computed through a Fitch-like approach in [12], see below for a discussion on that topic), rather than to the dataset itself as the way we compute synteny blocks are very similar, even if we conserve only blocks that are present in all genomes. Nevertheless, the results obtained both by our method and Ma et al. method, which both rely on model-free algorithmic principles, like cytogenetics methods but on other kind of data, strongly agree with cytogenetics results.
We also tested our method on the ferungulate ancestor and compared our results with the ancestor inferred through a rearrangement-based method in Murphy et al. [11]. With the method Murphy et al., based on a genome rearrangement model and MGR [8], the results diverged from the cytogenetics data and provoked the discussion in [16]–[18]. Using the same synteny blocks as Murphy et al., we found 24 CARs, all of which are chromosomes of the boreoeutherian ancestor, except a fusion of the homologs of human chromosomes 1 and 10, which seem to be ferungulate-specific, and was also inferred by MGR. None of the other chromosomal syntenies proposed by [11] were recovered by our method, or the Ma et al. method. However, the number of common inferred ancestral adjacencies points out that our method and the method of Ma et al. compute similar ancestral genome architectures, which are different from the one proposed by MGR, despite the fact that this last one has 24 CARs, as with our method. We believe that this three-way comparison indicates that the differences discussed in [16],[17] are partly due to the methods themselves, and more precisely to the fact that MGR is a rearrangement-based method, whereas all the others are model-free.
We now summarize the main methodological features of the framework we propose, and discuss them, as well as some possible extensions. We propose to decompose the process of ancestral genome architecture inference into three steps: detection and weighting of ancestral syntenies, representation as a 0/1 matrix and a generalized PQ-tree, clearing ambiguities and representation of a set of alternative genome architectures as a PQ-tree. Although these three steps are performed independently, the implementation choices for each of them can have important consequences on the other ones, as we discuss below. We implemented this method using (1) unique and universal synteny blocks, which appear once in each genome, (2) ancestral syntenies defined as unambiguous adjacencies and maximal common intervals (or gene teams) which are present in at least two genomes whose evolutionary path along their phylogeny meets the considered ancestral species and (3) a combinatorial optimization approach, based on the Consecutive Ones Submatrix Problem, to clear ambiguities. The comparison of our method and the one of Ma et al. [12] through the prism of this framework highlights the important effects of some methodological choices on ancestral genome proposals. We discuss below these choices on the combinatorial nature of the considered sets of genomic markers, the definition and computation of ancestral syntenies, and the method to clear ambiguities.
We construct several datasets, by a unique method depending on two parameters, max_gap and min_len. This method, or very similar ones, are often used to construct synteny blocks from genomic alignments [40],[41],[59].
We first use the notion of “teams of markers” [24]. This notion relies on a parameter δ, a positive integer. In a genome, the position of a marker m, denoted by p(m), is its relative rank on the its chromosome. That is, the first marker on a chromosome has rank 1, the second has rank 2, and so on. Two markers m1 and m2 are said to be close to each other in a genome, for the parameter δ, if they lie on the same chromosome, and |p(m1)−p(m2)|≤δ. A subset of markers M is said to be a team for a genome if for any two markers a,b from M, there exists a sequence S = a,a1,…ak,b of markers from M, such that any two consecutive markers in S are close to each other. Given two genomes X and Y, a team S common to X and Y is a set of markers labels (a subset of Σ the alphabet of markers) that is a team in both genomes X and Y. Such a team S is maximal if no other team is common to X and Y and contains S. Maximal common intervals are maximal common teams for δ = 1. Maximal common teams can be computed efficiently thanks to an algorithm by Beal et al. [23] and a software described in [24]. We collect a set of teams, representing possible ancestral syntenies, by computing all maximal common teams of pairs of species which evolutionary path contains the wished ancestor.
In order to analyze the dataset of [12], due to less defined markers in the two outgroup genomes, we used maximal approximate common intervals defined as follows: a subset M of markers is an approximate common interval between two genomes if there exists a genome segment in each of the two genomes whose 80% of the gene content is equal to S. An approximate common interval S is maximal if no other approximate common intervals is common to X and Y and contains the two occurrences of S in X and Y.
As teams rely only on similarity in markers content, and do not involve any marker order constraints, we added to this set of ancestral syntenies the set of putative ancestral adjacencies, defined as pairs of markers that are consecutive in at least two genomes whose evolutionary path contains this ancestor and do not belong to a conflict. A conflict is defined as follows (Figure 7 in [12]): an adjacency {i, j} belongs to a conflict if, in the graph G whose vertices are the markers (V(G) = Σ) and the edges are the conserved adjacencies, either i or j has degree more than 2, or the edge {i, j} belongs to a cycle.
Each of these ancestral syntenies was weighted following the same principle as in [12]. Let S be a subset of Σ that represents a possible ancestral synteny. In any leaf X of the species tree, if S is a team in X, the weight of S in X is wX(S) = 1, otherwise, wX(S) = 0. Then, in any internal node N of T (other than the ancestral node A) having two children R and L, wN(S) is defined recursively by the formulawhere dL and dR are respectively the length of the branch between N and L and N and R. The weight of S in A is then defined bywhere A1, A2 and A3 are the three neighbors of the ancestral node A in T, and , and are the respective length of the branch between A and A1, A and A2 and A and A3.
Recall is the set of homologous markers, is the set of subsets of that represent possible ancestral syntenies and the corresponding 0/1 matrix.
We say that two elements Si and Sj of overlap if their intersection is not empty, but none is included in the other. Let be the family of all subsets of that do not overlap with any member of ; in other words, given X an element of , any Si of either contains all elements of X or contains no element of X. Among the subsets of , call strong the elements that do not overlap any other elements of . The inclusion tree of the strong elements of , denoted , is a tree where each strong element of corresponds to a single node and the node corresponding to a strong subset X is an ancestor of the node corresponding to a strong subset Y if and only if X contains Y as a subset.
Given a node N of , we associate to it the subset of the elements of defined as all Si's that are included in N but in none of its children. The PQ-tree is defined from as follows: an internal node N such that s(N) = Ø is a P-node, while an internal node N such that s(N)≠Ø is a Q-node if s(N) can be partitioned by a partition refinement process [68] and a R-node otherwise. The construction of can be achieved in optimal O(n+m) time where and , as described in [45].
In the last step, we want to remove the minimal amount (in terms of weight) of ancestral syntenies from in order that the resulting matrix is C1P. This problem, which is known as the Consecutive Ones Submatrix Problem generalizes the Minimum Path Partition (or Path Cover) problem used in [12] and is known to be NP-hard [69] even for sparse matrices [70], which is the case of the matrices we obtain. However, using the structural information given by the PQ-tree , it is possible to design an efficient branch-and-bound algorithm.
More precisely, it follows immediately from the definition of that ambiguous information that prevent a matrix to be C1P can only be located in the submatrices defined by the subsets s(N) of for the degenerate nodes of . Hence each of these subsets of can be processed independently of the remaining of . For such a subset, say , we first compute an upper bound on the maximum subset S of s(N) that defines a matrix that is C1P, using the same approach than in [12]: start with S = Ø and, for each element of s(N), taken in decreasing order of weight, if adding to S defines a matrix that is not C1P (which can be tested using the efficient algorithms described in [46],[68]), then discard it, else leave it in S. From that upper bound, using the same principle, we use a classical branch-and-bound algorithm that looks for a better subset of s(N) that defines a C1P matrix.
Let an adjacency in an ancestral genome architecture be defined by two markers X and Y that are adjacent in a CAR of this ancestral genome, for a given resolution (say 100 kb). We say that it is conserved at a lower resolution (say 200 kb) if either the synteny blocks corresponding to X and Y in the human genome are both included in a single synteny block in the human genome at the lower resolution or if X and Y are contained in two blocks X′ and Y′ at the lower resolution level whose corresponding markers are adjacent in the ancestral genome inferred at this resolution. The adjacency is weakly conserved if the markers X′ and Y′ are not adjacent but present on the same CAR (weakly conserved adjacencies point at local rearrangements resulting from changing the resolution of the considered data). Otherwise, if the two markers X′ and Y′ are not on the same CAR, we say that the adjacency between X and Y is not conserved. Note that we do not consider this adjacency is not conserved if at least one of the two synteny blocks corresponding to X or Y is not included in a lower resolution synteny block.
Part of this work was done while CC visited the LRI (Université Paris-Sud, Orsay, France) and LaBRI (Université Bordeaux I, Talence, France).
|
10.1371/journal.pbio.0050062 | The Gene vitellogenin Has Multiple Coordinating Effects on Social Organization | Temporal division of labor and foraging specialization are key characteristics of honeybee social organization. Worker honeybees (Apis mellifera) initiate foraging for food around their third week of life and often specialize in collecting pollen or nectar before they die. Variation in these fundamental social traits correlates with variation in worker reproductive physiology. However, the genetic and hormonal mechanisms that mediate the control of social organization are not understood and remain a central question in social insect biology. Here we demonstrate that a yolk precursor gene, vitellogenin, affects a complex suite of social traits. Vitellogenin is a major reproductive protein in insects in general and a proposed endocrine factor in honeybees. We show by use of RNA interference (RNAi) that vitellogenin gene activity paces onset of foraging behavior, primes bees for specialized foraging tasks, and influences worker longevity. These findings support the view that the worker specializations that characterize hymenopteran sociality evolved through co-option of reproductive regulatory pathways. Further, they demonstrate for the first time how coordinated control of multiple social life-history traits can originate via the pleiotropic effects of a single gene that affects multiple physiological processes.
| Animals that live in groups often specialize in different tasks, creating a division of labor. One extreme example can be seen in honeybees, in which most tasks are performed by thousands of worker females that are essentially sterile helpers. Workers start out as nurse bees that care for larvae in the nest. Later they embark on foraging trips, specializing in either pollen or nectar collection, and continue to forage until they die. The age when workers initiate foraging and the tendency to collect pollen or nectar have been linked to a rudimentary reproductive physiology in which the protein vitellogenin appears to play a central role. Vitellogenin is normally used to produce egg yolk, but it may affect behavior and lifespan in workers. We tested this hypothesis by knocking down the vitellogenin gene of worker bees. Workers with suppressed vitellogenin levels foraged earlier, preferred nectar, and lived shorter lives. Thus, vitellogenin has multiple effects on honeybee social organization. By using gene knockdown to understand insect social behavior, our study supports the view that social life in bees evolved by co-opting genes involved in reproduction.
| Vitellogenin has versatile regulatory functions in honeybees, suggesting that this glycolipoprotein may be involved in the control of social life-history traits [1–5]. Vitellogenin is the common yolk precursor protein of oviparous taxa [6]. However, it appears to have evolved pleiotropic functions in the advanced eusocial honeybee that have not as yet been given attention in other species that rely on vitellogenin for oocyte development [1,4].
Honeybee vitellogenin has been hypothesized to work together with juvenile hormone in a double repressor network to coordinate behavior [7]. In this network, vitellogenine suppresses juvenile hormone and inhibits the worker honeybees' age-associated shift from nest tasks to foraging duties [3]. This shift is a complex behavioral transition characterized by decreasing vitellogenin and increasing juvenile hormone titer [8]. It has been proposed also that variation in vitellogenin gene expression early in life is associated with subsequent behavioral specialization that gives rise to a division of labor between nectar and pollen foraging workers [9]. Finally, honeybee vitellogenin can reduce oxidative stress by scavenging free radicals, thereby prolonging lifespan in the facultatively sterile worker castes and the reproductive queen castes [4]. Similar antioxidant function has been suggested for vitellogenin molecules in the nematode Caenorhabditis elegans [10] and in eggs of the eel Anguilla japonica [11], but positive effects of vitellogenin on adult longevity have not been demonstrated in these species.
In summary, the proposed pleiotropic effects of honeybee vitellogenin suggest that the vitellogenin gene is a central element in the life-history regulation of this social insect. Here we test these proposed functions by using RNA interference (RNAi) to knock down expression of the honeybee vitellogenin gene.
The vitellogenin RNAi tool [12] was used in combination with observations of the behavior and lifespan of worker honeybees living in otherwise unmanipulated colonies. RNAi-mediated knockdown of the vitellogenin protein has been confirmed repeatedly in 5–7-d-old worker bees [3,5,12]. In this first RNAi study of honeybee social life history, however, we aimed to monitor workers over several weeks. Therefore, RNAi was validated in cohorts of 10-d-old (n = 31), 15-d-old (n = 27) and 20-d-old (n = 27) bees (Figure 1).
This initial test demonstrated that workers with a vitellogenin RNAi phenotype (n = 30) were characterized by persistent suppression of vitellogenin protein levels compared to controls (n = 24), which received injections of double-stranded RNA (dsRNA) derived from green fluorescent protein (GFP) encoding sequence (p < 0.005; see the legend to Figure 1 for details on statistics). The GFP dsRNA control represents a handling disturbance control [5], which is necessary because honeybees respond to many kinds of handling stress with changes in endocrines, neuromodulators, and behavior [13–15]. This control was monitored relative to a non-injected reference group (n = 31; see Figure 1 for details). The vitellogenin levels of the GFP dsRNA control and the non-injected reference group were not significantly different (p = 0.27).
Next, we found that vitellogenin knockdowns (n = 122) initiated foraging flights earlier in life than GFP dsRNA controls (n = 179, p < 0.003; see Figure 2 for details). These results confirm the hypothesis that honeybee vitellogenin gene activity influences worker division of labor via an inhibitory effect on the shift from nest tasks to foraging [3,7].
In addition, down-regulation of vitellogenin gene activity (n = 160) resulted in foragers collecting larger loads of nectar relative to GFP dsRNA controls (n = 159, p < 0.010; see Figure 3 for details). Overall, loads were within the range normally collected by honeybees (up to 60-mg nectar and 30-mg pollen [16,17]), and thus interfering with vitellogenin expression did not change the maximum load size collected by workers. The observed bias towards nectar collection in vitellogenin knockdowns is consistent with earlier studies showing low hemolymph (blood) levels of vitellogenin in young worker bees from genetic stocks that preferentially collect nectar [9]. Genetic stocks with bias for collecting pollen are characterized by high levels of vitellogenin prior to foraging onset [9]. Our data, however, go beyond these correlations and demonstrate that the vitellogenin gene influences social foraging specialization.
Survival data showed that vitellogenin also is involved in the regulation of honeybee lifespan. Lifespan was reduced in vitellogenin knockdowns (n = 122) compared with GFP dsRNA controls (n = 179, p < 0.036, see Figure 4 for details). The effect was not due simply to bees initiating foraging behavior earlier in life, because these traits were not correlated in the knockdown phenotype (r = 0.121 [Colony 1]; r = −0.003 [Colony 2], p > 0.05). Our finding is supported by the previous results showing that worker bees with reduced vitellogenin activity levels are more susceptible to oxidative stress [4], a physiological state that is an established indicator of aging [18,19].
Our results suggest that honeybee vitellogenin has an integrative function in regulating social organization through its pleiotropic effects on division of labor and foraging specialization. Vitellogenin inhibits the onset of foraging (our study) but declines with age in workers [8,20], thereby serving as a pacemaker for age polyethism and lifespan, as first hypothesized by Omholt and Amdam [7,21]. Higher titers early in life [9] prime bees for pollen collection, whereas low titers prime bees for collecting nectar (our study). vitellogenin RNAi established at adult emergence triggers persistent suppression of vitellogenin activity [3,5,12] (Figure 1), and therefore, the knockdown phenotype was expected to initiate foraging early, collect nectar, and live a short life (Figure 5).
Life-history pleiotropy demonstrated by the effects of vitellogenin is similar in principle to trait associations that are controlled by the systemic endocrine factors juvenile hormone and ecdysone in the solitary insect Drosophila (reviewed by Flatt et al. [22]). In Drosophila, yolk precursor peptides [23] are downstream components of these hormonal signaling cascades [24], whereas in honeybees, vitellogenin is part of a regulatory feedback loop that enables vitellogenin and juvenile hormone to mutually suppress each other [3,4,7]. As a consequence, vitellogenin and juvenile hormone should be considered joint effectors of division of labor and foraging specialization, at least until methods can be developed to separate their individual effects. This feedback relationship is uncommon in insects [3], suggesting in combination with our findings that evolutionary co-option and remodeling of vitellogenin and juvenile hormone action [1,4–6] have been important steps in honeybee social evolution [3].
Previous studies have identified genes that affect honeybee foraging onset (Amfor [25] and malvolio [26]), and multiple genes with mRNA levels that correlate with foraging behavior [27] or lifespan [28]. Yet our work represents the first successful RNAi approach to decipher gene and protein function in honeybee social behavior. Our data demonstrate for the first time that several key characteristics of a social phenotype can be coordinated by a single reproductive gene. This pleiotropy lends support to studies showing that complex social behavior in insects can evolve from ancestral reproductive traits [9,29–31].
The non-injected reference (noREF), GFP dsRNA control (injGFP), and vitellogenin knockdown (vgRNAi) phenotypes were obtained as previously described by Amdam et al. [5,12]. In short, primers were designed from the sequence of the A. mellifera vitellogenin cDNA clone AP4a5, and the GFP encoding sequence of the pGFP vector (Clontech, Palo Alto, California, United States). GFP dsRNA does not affect vitellogenin [5], but was used in planned comparisons with vitellogenin dsRNA to control for laboratory handling that affects sensory and physiological correlates of worker foraging behavior [5,13,14].
Primers were fused with T7 promoter sequence (underlined): for clone AP4a5: 5′-TAATACGACTCACTATAGGGCGAACGACTCGACCAACGACTT-3′ and 5′-TAATACGACTCACTATAGGGCGAAACGAAAGGAACGGTCAATTCC-3′; and for pGFP: 5′-TAATACGACTCACTATAGGGCGATTCCATGGCCAACACTTGTCC-3′ and 5′-TAATACGACTCACTATAGGGCGATCAAGAAGGACCATGTGGTC-3′. PCR reactions were performed according to standard procedures using AP4a5 and the pGFP vector as templates. Resulting products excluding the fused T7 promoters were 504 base pairs (bp) (vitellogenin) and 503 bp (GFP derived). Products were purified using the QIAquick PCR purification kit (Qiagen, Valencia, California, United States), and RNA was prepared using the Promega RiboMax T7 system (Promega, Madison, Wisconsin, United States). RNA was extracted by TRIzol LS reagent (GIBCO-BRL, San Diego, California, United States), resuspended in nuclease-free water, heated at 96 °C for 2 min in an Eppendorf Thermomixer (Brinkmann Instruments, Westbury, New York, United States), and left to cool at room temperature for 20 min. The integrity of the dsRNA was tested using 1.5% agarose gels, and the products were diluted with nuclease-free water (Qiagen) to the final concentration of 5 μg/μl.
Newly emerged workers were randomly assigned to one of three treatments and marked with paint (Testors Enamel; Testor Corporation, Rockford, Illinois, United States) to indicate treatment identity. The noREF group was set aside. The two remaining groups were injected with vitellogenin-derived or GFP-derived dsRNA (to make up the vgRNAi or injGFP treatment, respectively) between the fifth and sixth tergite using Hamilton syringes with G30 disposable needles (BD, Palo Alto, California, United States). Injection volume was 2 μl. Injections were performed by trained personnel that were blind to treatment identities. Workers were introduced into two colonies kept in commercial hive boxes (n = 50 for each treatment group and in each colony, respectively) and collected after 10, 15, and 20 d during non-foraging hours.
Bees were anesthetized on ice, and hemolymph was extracted with Drummond Scientific Company (Broomall, Pennsylvania, United States) micropipettes after puncturing the abdomen between the third and fourth tergite with a sterile G30 needle (BD). Care was taken to avoid contaminating the samples with tissue fragments and foregut content. Hemolymph (1 μl) was dissolved in 10-μl Tris buffer: 20 mM Tris, 150 mM NaCl, 5 mM EDTA (pH 7.5), 1 mM phenylmethylsulfonyl fluoride, 5 mM benzamidin, 0.7 μM pepstatin, 8 μM chymostatin, 10 μM leupeptin, 0.8 μM aprotinin (Sigma-Aldrich, St. Louis, Missouri, United States), before samples were separated by 8% SDS-polyacrylamide gel electrophoresis using standard methods [32]. A β-galactosidase standard (Sigma-Aldrich) was included to allow densitometrically quantification by the method of Lin et al. [33], in which vitellogenin is detected as a single band of 180 kDa [5,33,34]. The densitometrical analysis was performed by the Quantity One imaging software (Bio-Rad, Hercules, California, United States) after staining the gels with Commassie Brilliant Blue (Sigma-Aldrich). Gel-to-gel variation in staining intensity was controlled by background correction and the β-galactosidase standards as we have described previously [5].
Newly emerged workers were randomly assigned to one of three treatments. noREF bees were uniquely tagged and set aside. Groups of vgRNAi and injGFP workers were obtained as described above. About equal numbers of bees from each treatment group were introduced into two host colonies (n = 288 noREF, 338 injGFP, and 293 vgRNAi into Colony 1; and n = 288 noREF, 338 injGFP, and 299 vgRNAi into Colony 2). Each colony was set up in an observation hive with about 5,000 unmarked adult bees of diverse ages. Observers were blind to treatment identity. Colonies were surveyed daily for tagged bees during non-foraging hours to establish survivorship (the combs of each colony were scanned twice daily), and for two 40-min periods daily during peak foraging hours to establish foraging activity. Ramps from the outside port leading to the bottom comb were observed for incoming tagged forager bees during these 40-min sessions. A bee's age of foraging onset was the number of days since adult emergence that she was first seen returning from a foraging trip. A bee's day of death was established as the day after the last day she was observed.
Set-ups were separate, but identical, to the one described for age of foraging onset and lifespan, except that host colonies were kept in commercial hive boxes. During peak foraging hours, incoming foragers were collected at the hive entrances, brought to the laboratory, and sampled once for type of substance collected. Bees were not returned to the colonies after sampling. At the time of collection, the workers were 10–16 d old (Colony 1) or 7–13 d old (Colony 2). Pollen and/or nectar loads were removed and quantified for each individual worker as described before [35]. In all, n = 80 noREF, 84 injGFP, and 65 vgRNAi bees were taken from Colony 1, and n = 78 noREF, 75 injGFP, and 95 vgRNAi bees were taken from Colony 2. Workers that returned empty (with no measurable nectar or pollen) were not used in our analyses.
The dataset on hemolymph vitellogenin levels was not normally distributed, as determined by Bartlett's test of sphericity. Therefore, the non-parametric Kruskal-Wallis test was used (n = 85) to examine the data for effects of treatment (H = 14.32, p = 0.001) and age (H = 0.79, p = 0.68). A Mann-Whitney U test was used to test for colony effects (Z = −3.54, p < 0.001). Vitellogenin titers were higher overall in Colony 2, but the relative expression pattern between treatments was the same in the two colonies (unpublished data). Therefore, the dataset was not split by colony, but used in full to increase the statistical power of the post hoc analysis. A Mann-Whitney U test was used as a post hoc test of the planned comparisons of vitellogenin levels between noREF and injGFP, and between injGFP and vgRNAi. Treatment differences in age of foraging onset and lifespan were analyzed using the Cox proportional hazards regression model with the likelihood ratio chi-square tests (LRT) for bees that foraged at least twice (n = 149 noREF, 69 injGFP, and 48 vgRNAi for Colony 1; and n = 137 noREF, 110 injGFP, and 74 vgRNAi for Colony 2). The proportional hazards assumption was verified by Schoenfeld residuals [36]. Colony effects were controlled for by using the colony as a stratifying variable in the analyses. Effect of treatment was detected in the dataset overall (foraging onset LRT = 95.1, df = 2, p < 0.0001; lifespan LRT = 82.2, df = 2, p < 0.0001). Thus, planned pair-wise comparisons were made between noREF and injGFP, and between injGFP and vgRNAi. The data on foraging loads passed Bartlett's test of sphericity. Two-way analysis of variance (ANOVA) was therefore used to test for effects of treatment on foraging loads while controlling for the colony factor. A significant effect of treatment was detected (ANOVA, F2,471 p < 0.0001) and planned pair-wise comparisons were made using a two-way ANOVA. Pearson analysis was used to correlate age at foraging onset and lifespan. Statistical software was SPLUS version 6.1 (http://www.insightful.com), Systat 6.0 (http://www.systat.com), and Statistica 6.0 (http://www.statsoft.com).
The GenBank (http://www.ncbi.nlm.nih.gov/Genbank) accession numbers for the genes discussed in this paper are vitellogenin (AJ517411) and the GFP encoding sequence of the pGFP vector (AF324407).
|
10.1371/journal.pcbi.1001111 | Minimal Models of Multidimensional Computations | The multidimensional computations performed by many biological systems are often characterized with limited information about the correlations between inputs and outputs. Given this limitation, our approach is to construct the maximum noise entropy response function of the system, leading to a closed-form and minimally biased model consistent with a given set of constraints on the input/output moments; the result is equivalent to conditional random field models from machine learning. For systems with binary outputs, such as neurons encoding sensory stimuli, the maximum noise entropy models are logistic functions whose arguments depend on the constraints. A constraint on the average output turns the binary maximum noise entropy models into minimum mutual information models, allowing for the calculation of the information content of the constraints and an information theoretic characterization of the system's computations. We use this approach to analyze the nonlinear input/output functions in macaque retina and thalamus; although these systems have been previously shown to be responsive to two input dimensions, the functional form of the response function in this reduced space had not been unambiguously identified. A second order model based on the logistic function is found to be both necessary and sufficient to accurately describe the neural responses to naturalistic stimuli, accounting for an average of 93% of the mutual information with a small number of parameters. Thus, despite the fact that the stimulus is highly non-Gaussian, the vast majority of the information in the neural responses is related to first and second order correlations. Our results suggest a principled and unbiased way to model multidimensional computations and determine the statistics of the inputs that are being encoded in the outputs.
| Biological systems across many scales, from molecules to ecosystems, can all be considered information processors, detecting important events in their environment and transforming them into actions. Detecting events of interest in the presence of noise and other overlapping events often necessitates the use of nonlinear transformations of inputs. The nonlinear nature of the relationships between inputs and outputs makes it difficult to characterize them experimentally given the limitations imposed by data collection. Here we discuss how minimal models of the nonlinear input/output relationships of information processing systems can be constructed by maximizing a quantity called the noise entropy. The proposed approach can be used to “focus” the available data by determining which input/output correlations are important and creating the least-biased model consistent with those correlations. We hope that this method will aid the exploration of the computations carried out by complex biological systems and expand our understanding of basic phenomena in the biological world.
| There is an emerging view that the primary function of many biological systems, from the molecular level to ecosystems, is to process information [1]–[4]. The nature of the computations these systems perform can be quite complex [5], often due to large numbers of components interacting over wide spatial and temporal scales, and to the amount of data necessary to fully characterize those interactions. Constructing a model of the system using limited knowledge of the correlations between inputs and outputs can impose implicit assumptions and biases leading to a mischaracterization of the computations. To minimize this type of bias, we maximize the noise entropy of the system subject to constraints on the input/output moments, resulting in the response function that agrees with our limited knowledge and is maximally uncommitted toward everything else. An equivalent approach in machine learning is known as conditional random fields [6]. We apply this idea to study neural coding, showing that logistic functions not only maximize the noise entropy for binary outputs, but are also special closed-form cases of the minimum mutual information (MinMI) solutions [7] when the average firing rate of a neuron is fixed. Recently, MinMI was used to assess the information content in constraints on the interactions between neurons in a network [8]. We use this idea to study single neuron coding to discover what statistics of the inputs are encoded in the outputs. In macaque retina and lateral geniculate nucleus, we find that the single neuron responses to naturalistic stimuli are well described with only first and second order moments constrained. Thus, the vast majority of the information encoded in the spiking of these cells is related only to the first and second order statistics of the inputs.
To begin, consider a system which at each moment in time receives a -dimensional input from a known distribution , such as a neuron receiving a sensory stimulus or post-synaptic potentials. The system then performs some computation to determine the output according to its response function . The complete input/output correlation structure, i.e. all moments involving and , can be calculated from this function through the joint distribution , e.g. . Alternatively, the full list of such moments contains the same information about the computation as the response function itself, although such a list is infinite and experimentally impossible to obtain. However, a partial list is usually obtainable, and as a first step we can force the input/output correlations from the model to match those which are known from the data. The problem is then choosing from the infinite number of models that agree with those constraints. Following the argument of Jaynes [9], [10], we seek the model which avails the most uncertainty about how the system will respond.
Information about the identity of the input can be obtained by observing the output, or vice versa, quantified by the mutual information [11], [12]. The first term is the response entropy, , which captures the overall uncertainty in the output. The second term is the so-called noise entropy [13],(1)representing the uncertainty in that remains if is known. If the inputs completely determine the outputs, there is no noise and the mutual information reaches its highest possible value, . In many realistic situations however, repeated presentations of the same inputs produce variable outputs producing a nonzero noise entropy [14] and lowering the information transmitted.
By maximizing the noise entropy, the model is forced to be consistent with the known stimulus/response relationships but is as uncertain as possible with respect to everything else. We show that this maximum noise entropy (MNE) response function for binary output systems with fixed average outputs is also a minimally informative one. This approach is a special closed-form case of the mutual information minimization technique [8], which has been used to address the information content of constraints on the interactions between neurons. Here we use the minimization of the mutual information to characterize the computations of single neurons and discover what about the stimulus is being encoded in their spiking behavior.
The starting point for constructing any maximum noise entropy model is the specification of a set of constraints , where indicates an average over the joint distribution . These constraints reflect what is known about the system from experimental measurements, or a hypothesis about what is relevant for the information processing of the system. For neural coding, the constraints could be quantities such as the spike-triggered average [15]–[18] or covariance [19]–[22], equivalent to and , respectively. With each additional constraint, our knowledge of the true input/output relationship increases and the correlation structure of the model becomes more similar to that of the actual system.
Given the constraints, the general MNE response function is given by (see Methods)(2)where the -dependent partition function ensures that the MNE response function is consistent with normalization, i.e. . The MNE response function in Eq. (2) has the form of a Boltzmann distribution [23] with a Lagrange multiplier for each constraint. The values of these parameters are found by matching the experimentally observed averages with the analytical averages obtained by from derivatives of [23].
Many systems in biological settings produce binary outputs. For instance, the neural state can be thought of as binary, with for the silent state and for the “spiking” state, during which an action potential is fired [13]. The inputs themselves could be a sensory stimulus or all of the synaptic activity impinging upon a neuron, both of which are typically high-dimensional [24]. Another example is gene regulation [25], where the inputs could be the concentrations of transcription factors and the binary output represents an on/off transcription state of the gene. For these systems, the constraints of interest are proportional to . This is because any moments independent of will cancel due to the partition function and any moments with higher powers are redundant, e.g. if or 1. In this case, the set of constraints may be written more specifically as and the MNE response function becomes the well-known logistic function(3)with . Thus for all binary MNE models, the effect of the constraints is to perform a nonlinear transformation of the input variables, , to a space where the spike probability is given by the logistic function (inset, Fig. 1).
For neural coding, one of the most fundamental and easily measured quantities is the total number of spikes produced by a neuron over the course of an experiment, equivalent to the mean firing rate. By constraining this quantity, or more specifically its normalized version , the MNE model is turned into a minimum information model. This holds because the response entropy is completely determined by the distribution , which is in turn constrained by if the response is binary. With the response entropy constrained to match the experimentally observed system, maximizing the noise entropy is equivalent to minimizing information. Therefore, as was proposed in [7], any model that satisfies a given set of constraints will convey the information that is due only to those constraints. With each additional constraint our knowledge of the correlation structure increases along with the minimum possible information given that knowledge, which approaches the true value as illustrated schematically in Fig. 1.
The simplest choice is a first order model () where the spikes are correlated with each input dimension separately. This model requires knowledge of the set of moments , the spike-triggered average stimulus. For , the transformation on the inputs is linear, , where the constant is the Lagrange multiplier for the spike probability constraint. With knowledge of only first order correlations, we see that the model neuron is effectively one-dimensional, choosing a single dimension in the -dimensional input space and disregarding all information about any other directions.
With higher order constraints, the transformation is nonlinear and the model neuron is truly multidimensional. For instance, the next level of complexity is a second order model (), in which spikes may also interact with pairs of inputs. This model is obtained by constraining , equivalent to knowing the spike-triggered covariance of the stimulus, resulting in the input transformation . Any other MNE model can be constructed in the same fashion by choosing a different set of constraints, reflecting different amounts of knowledge.
The mutual information of the MNE model is the information content of the constraints. The ratio of to the empirical estimate of the true mutual information of the system is the percent of the information captured by the constraints. This quantity is always less than or equal to one, with equality being reached if and only if all of the relevant moments have been constrained. This suggests a procedure to identify the relevant constraints, described in Fig. 2A. First, a hypothesis is made about which constraints are important. Then the corresponding MNE model is constructed and the information calculated. If the information captured is too small, the constraints are modified until a sufficiently large percentage is reached. Any constraints beyond that are relatively unimportant for describing the computation of the neuron.
As an illustrative example of the MNE method, consider a binary neuron which itself receives binary inputs (i.e. a logic gate). If the neuron in question receives binary inputs, we are guaranteed to capture 100% of the information with -order statistics because all moments involving powers greater than one of either or any are redundant. However, different coding schemes may encode different statistics of the inputs. For instance, if the neuron receives only two inputs (Fig. 2B), the well-known AND and OR logic gate behaviors are completely described with only first order moments [26]. Correspondingly, the first order model captures 100% of the information. Such a neuron can be said to encode only first order statistics of the inputs, and the spike-triggered average stimulus contains all of the information necessary to fully understand the computation. On the other hand, the XOR gate (Fig. 2C, left) requires second order interactions. This is reflected by and accounting for 0% and 100% of the information, respectively. More complicated coding schemes may involve both first and second order interactions, such as for the gate shown in the right panel of Fig. 2C. Here, and account for 10% and 100% of the information, respectively, and correctly quantify the degree to which each order of interaction is relevant to this neuron.
Similar situations show up for neurons that receive three binary inputs. The top panel of Fig. 2D shows an example of a neuron which only requires second order interactions. The parameters of are exactly the same as , with the third order coefficient . The bottom panel shows an example of a situation in which third order interactions are necessary. Correspondingly, increases the information explained over from 71% to 100%. These simulations demonstrate that despite the different coding schemes used by neurons, the information content of each order of interaction can be correctly identified using logistic MNE models.
In their natural environment, neurons commonly encode high-dimensional analog inputs, such as a visual or auditory stimulus as a function of time. It is important to note that the non-binary nature of the inputs means that the ability to capture 100% of the information between and the inputs with -order statistics is not guaranteed anymore. Often, the dimensionality of the inputs may be reduced because the neurons are driven by a smaller subspace of relevant dimensions (e.g. [27]–[33]). However, even in those cases we are often forced to use qualitative terms such as ‘ring’ or ‘crescent’ to describe the experimentally observed response functions. With no principled way of fitting empirical response functions, the details of the interactions between neural responses and reduced inputs have been difficult to quantify.
The MNE method provides a quantitative framework for characterizing neural response functions, which we now apply to 9 retinal ganglion cells (RGCs) and 9 cells in the lateral geniculate nucleus (LGN) of macaque monkeys, recorded in vivo (see Methods). The visual input was a time dependent sequence of luminance values synthesized to mimic the non-Gaussian statistics of light intensity fluctuations in the natural visual environment [34]–[36].
A 1s segment of the normalized light intensity is shown in Fig. 3A. A previous study has shown that the responses of these neurons are correlated with the stimulus over an approximately 200 ms window preceding the response. When binned at 4 ms resolution, which ensures binary responses, the input is a vector in a 50 dimensional space. However, spikes are well predicted by using a 2 dimensional subspace [29] identified through the Maximally Informative Dimensions (MID) technique [37].
These two relevant dimensions, shown for a RGC in Fig. 3B, form a two dimensional receptive field which preserves the most information about the spikes in going from 50 to 2 dimensions. The two linear filters are convolved with the stimulus to produce reduced inputs and , shown in Fig. 3A. The resulting input probability distribution in the reduced space is shown in Fig. 3C. The measured responses of the neuron then form a two-dimensional response function shown in Fig. 3D, where the color scale indicates the probability of a spike as a function of the two relevant input components.
To gain insight into the nature of this neuron's computational function and find the important interactions, we apply the MNE method starting with the first order MNE model shown in Fig. 3E. The first order model produces a response function which bears little resemblance to the empirical one and accounts for only of the information. The next step is a second order MNE model (Fig. 3F), which produces a response function quite similar to the empirical one in both shape and amplitude, while accounting for of the information. Thus, for this neuron, knowledge of second order moments is both necessary and sufficient to generate a highly accurate model of the neural responses.
This result was typical across the population of cells, as illustrated in Fig. 4A by comparing the information captured by the first order versus second order models. The majority of the cells were well described by the second order model, accounting for over of the information. When averaged across the population, the first order model captured and the second order model captured of . These results suggest that the inclusion of second order interactions are both necessary and sufficient to describe the responses of these neurons to naturalistic stimuli.
Since the MNE response function is a distribution of outputs given inputs, another way to check the effectiveness of any MNE model is to compare its moments with those obtained from experiments. The moments constrained to obtain the model will be identical to the experimental values by construction; it is the higher order moments, left unconstrained, that should be compared. In Fig. 4B we show two such comparisons for the correlation functions and , which involve moments unconstrained in the and models. In both cases, the first order model predictions show more scatter than those of the second order model; the latter does a reasonable job of predicting the experimentally observed correlations. This result broadly demonstrates the sufficiency of second order interactions to model these neural responses, and shows that higher-order moments carry little to no additional information.
The two-dimensional second order MNE response functions have contours of constant probability which are conic sections. The parameter which governs the interaction between the two input dimensions, , is related to the degree to which the axes of symmetry of the conic sections are aligned with the two-dimensional basis. For example, if the contours are ellipses, then if the semi-major and semi-minor axes are parallel to the axes chosen to describe the input space, and otherwise (see inset, Fig. 5). To assess the importance of this cross term, we compared the performance of second order MNE models with and without . This additional term can only improve the performance of the model; however, as shown in Fig. 5, the improvements across the population are small. Thus, the dimensions found using the MID method are naturally parallel to the axes of symmetry of the response functions; however, this does not imply that the response function is separable due to the dependence of the normalization term .
For neural coding of naturalistic visual stimuli in early visual processing, we see that the bulk of what is being encoded is first order stimulus statistics. While the information gained by measuring the spike-triggered average is substantial, it is insufficient to accurately describe the neural responses. A second order model, which takes into account the spike-triggered input covariance, adds a sufficient amount of information. Thus the firing rates of these neurons have encoded the first and second order statistics of the inputs. Due to the fact that the natural inputs are non-binary and non-Gaussian, there exists a potential for very high-order interactions to be represented in the neural firing rate. It is known that higher order parameters of textures are perceptually salient [38]–[40], but it is unknown whether high order temporal statistics are also perceptually salient. Our results suggest that such temporal statistics are not encoded in the time-dependent firing rate, although they could be represented through populations of neurons or specific temporal sequences of spikes [41], [42].
Jaynes' principle of maximum entropy [9], [10] has a long and diverse history, with example applications in image restoration in astrophysics [43], extension of Wiener analysis to nonlinear stochastic transducers [44] and more recently in neuroscience [45]–[47]. In the latter studies, was maximized subject to constraints on the first and second order moments of the neural states and for a set of neurons in a network. The resulting pairwise Ising model was shown to accurately describe the distribution of network states of real neurons under various conditions. Since then the application of the Ising model to neuroscience has received much attention [48], [49], and it is still a subject of debate if and how these results extrapolate to larger populations of neurons [50]. Temporal correlations have also been shown to be important in both cortical slices and networks of cultured neurons [47].
In contrast to maximum entropy models that deal with stationary or averaged distributions of states, the goal of maximizing the noise entropy is to find unbiased response functions. This approach is equivalent to conditional random field (CRF) models [6] in machine learning. The parameters of a CRF are fit by maximizing the likelihood using iterative or gradient ascent algorithms [51] and have been used, for example, in classification and segmentation tasks [52]. The parameters of MNE models may also be found using maximum likelihood, or as was done here, by solving a set of simultaneous constraint equations numerically. Another example of a maximum noise entropy distribution is the Fermi-Dirac distribution [23] from statistical physics, which is a logistic function governing the binary occupation of fermion energy levels. Thus, in the same way that the Boltzmann distribution was interpreted by Jaynes as the most random one consistent with measurements of the energy, the Fermi-Dirac distribution can be interpreted as the least biased binary response function consistent with an average energy. However, to our knowledge, this method has never been used in the context of neural coding to determine the input statistics which are being encoded by a neuron and create the corresponding unbiased models.
Previous work has applied the principle of minimum mutual information (MinMI) [7] to neural coding, thus identifying the relevant interactions between neurons [8]. We have shown that the closed-form MNE solutions for binary neurons constitute a special case of MinMI, since the response entropy is fixed if the average firing rate is constrained. In general, the MinMI principle results in a self-consistent solution that must be solved iteratively to obtain the response function. The reason why MNE models are closed-form is that the constraints are formulated in terms of moments of the output distribution instead of the output distribution itself. In addition to the case of binary responses, MNE models can become closed-form MinMI models for any input/output systems where the response entropy can be fixed in terms of the moments of the output variable. Examples include Poisson processes with fixed average response rate or Gaussian processes with fixed mean and variance of the response rate. The framework for analyzing the interactions between inputs and outputs that we present here can thus be extended to a broad and diverse set of computational systems.
Our approach can be compared to other optimization techniques commonly used to study information processing. For example, rate-distortion theory [11], [12], [53], [54] seeks minimum information transmission rate over a channel with a fixed level of signal distortion, e.g. lossy image or video compression. In that case, the best solution is the one which transmits minimal information because this determines the average length of the codewords. In our method, we also obtain minimally informative solutions, not because they are optimal for signal transmission, but because they are the most unbiased guess at a solution given limited knowledge of a complex system.
At the other end of the optimization spectrum is maximization of information [1], [13], [55]. The goal in that case is to study not how the neuron does compute, but how it should compute to get the most information, perhaps with limited resources. This strategy has been used to find neural response functions for single neurons [56], [57], as well as networks [58], [59]. When confronted with incomplete knowledge of the correlation structure, a maximum information approach would choose the values of the unconstrained moments such that they convey the most information possible, whereas the minimum information approach provides a lower bound to the true mutual information, and allows us to investigate how this lower bound increases as more moments are included. If the goal is to study the limits of neural coding, then maximizing the information may be the best procedure. If, however, the goal is to dissect the computational function of an observed neuron, we argue that the more agnostic approaches of maximizing the noise entropy or minimizing the mutual information are better-suited.
Experimental data were collected as part of the previous study using procedures approved by the UCSF Institutional Animal Care and Use Committee, and in accordance with National Institutes of Health guidelines.
A maximum noise entropy model is a response function which agrees with a set of constraints and is maximally unbiased toward everything else. The constraints are experimentally observed moments involving the response and stimulus , , where , which must be reproduced by the model. The set of constraints, including the normalization of , are then added to the noise entropy to form the functional(4)with a Lagrange multiplier for each constraint. Setting and enforcing normalization yields Eq. 1. For a binary system, or 1, all the constraints take the form , and the partition function is , where
The values of the Lagrange multipliers are found such that the set of equations(5)is satisfied, with the analytical averages on the right-hand side obtained from derivatives of the free energy [23]. Simultaneously solving this set of equations has previously been shown to be equivalent to maximizing the log-likelihood [51].
The neural data analyzed here were collected in a previous study [29] and the details are found therein. Briefly, the stimulus was a spot of light covering a cell's receptive field center, flickering with non-Gaussian statistics that mimic those of light intensity fluctuations found in natural environments [35], [36]. The values of light intensities were updated every (update rate ). The spikes were recorded extracellularly in the LGN with high signal-to-noise, allowing for excitatory post-synaptic potentials generated by the RGC inputs to be recorded. From such data, the complete spike trains of both RGCs and LGN neurons could be reconstructed [60].
The neural spike trains were binned at 4 ms resolution, ensuring that the response was binary. The stimulus was re-binned at 250 Hz to match the bin size of the spike analysis. The neurons were uncorrelated with light fluctuations beyond 200 ms before a spike, and the stimulus vector was taken to be the 200 ms window (50 time points) of the stimulus preceding . Just two projections of this 50-dimensional input are sufficient to capture a large fraction of the information between the light intensity fluctuations and the neural responses ( for the example neuron mn122R4_3_RGC, and on average across the population). The two most relevant features of each neuron were found by searching the space of all linear combinations of two input dimensions for those which accounted for maximal information in the measured neural responses [37], subject to cross-validation to avoid overfitting. Each of the two features, and , is a 50-dimensional vector which converts the input into a reduced input, calculated by taking the dot product, i.e. . The algorithm for searching for maximally informative dimensions is available online at http://cnl-t.salk.edu.
|
10.1371/journal.ppat.1003285 | Hepatitis C Virus Induces the Mitochondrial Translocation of Parkin and Subsequent Mitophagy | Hepatitis C Virus (HCV) induces intracellular events that trigger mitochondrial dysfunction and promote host metabolic alterations. Here, we investigated selective autophagic degradation of mitochondria (mitophagy) in HCV-infected cells. HCV infection stimulated Parkin and PINK1 gene expression, induced perinuclear clustering of mitochondria, and promoted mitochondrial translocation of Parkin, an initial event in mitophagy. Liver tissues from chronic HCV patients also exhibited notable levels of Parkin induction. Using multiple strategies involving confocal and electron microscopy, we demonstrated that HCV-infected cells display greater number of mitophagosomes and mitophagolysosomes compared to uninfected cells. HCV-induced mitophagy was evidenced by the colocalization of LC3 puncta with Parkin-associated mitochondria and lysosomes. Ultrastructural analysis by electron microscopy and immunoelectron microscopy also displayed engulfment of damaged mitochondria in double membrane vesicles in HCV-infected cells. The HCV-induced mitophagy occurred irrespective of genotypic differences. Silencing Parkin and PINK1 hindered HCV replication suggesting the functional relevance of mitophagy in HCV propagation. HCV-mediated decline of mitochondrial complex I enzyme activity was rescued by chemical inhibition of mitophagy or by Parkin silencing. Overall our results suggest that HCV induces Parkin-dependent mitophagy, which may have significant contribution in mitochondrial liver injury associated with chronic hepatitis C.
| Hepatitis C virus (HCV) infection alters host lipid metabolism. HCV-induced mitochondrial dysfunction may promote the metabolic alterations by affecting mitochondrial β-oxidation and oxidative phosphorylation. Dysfunctional mitochondria are detrimental to cell survival and require rapid clearance to sustain cell viability. Here, we investigated the effect of HCV gene expression in promoting selective autophagy of dysfunctional mitochondria, also termed mitophagy. HCV infection stimulated the gene expression of Parkin and PINK1, the two key mediators of mitophagy. Parkin stimulation was also observed in liver biopsies of chronic hepatitis C patients. HCV infection induced the perinuclear clustering of mitochondria and triggered Parkin translocation to mitochondria, a hallmark of mitophagy. Concomitant with the mitochondrial translocation of Parkin, we observed ubiquitination of Parkin and its substrates in HCV-infected cells. We also demonstrate the formation of mitophagosomes and their subsequent delivery to lysosomes in HCV-infected cells. Silencing both Parkin and PINK1 hindered HCV replication, suggesting the functional significance of mitophagy in HCV life cycle. Furthermore, we demonstrate that Parkin-dependent mitophagy is directly associated with HCV-mediated decline in oxidative phosphorylation. Our results implicate the functional significance of Parkin and mitophagy in the persistence of HCV infection and mitochondrial injury commonly seen in patients with chronic hepatitis C.
| Hepatitis C virus (HCV) infection most often leads to chronic hepatitis, which can progress to steatosis, fibrosis, cirrhosis, and hepatocellular carcinoma [1]. HCV RNA genome encodes a polyprotein, which includes; core, E1, E2, p7, NS2, NS3, NS4A, NS4B, NS5A, and NS5B [2]. Viral RNA replicates in the endoplasmic reticulum (ER)-derived modified membranous structures and is subsequently assembled on lipid droplets [3], [4]. Most of the viral proteins are either associated with or tethered to the ER [3]. These associations and relevant activities overburden the ER and induce an ER stress response exhibited by the unfolded protein response (UPR) [5].
ER stress response is a potent inducer of autophagy [6]. Several reports have described that HCV gene expression perturbs the autophagic pathway to induce bulk autophagy [7]–[17]. Several reports have highlighted the functional role of autophagic machinery in various steps of HCV life cycle (viral replication, translation, and propagation) [7]–[17]. HCV-induced UPR and autophagy have also been functionally linked to inactivation of innate antiviral response [10], [18]. Interestingly, recent reports hint that HCV-induced autophagosomal membrane may serve as platform for HCV replication [11], [19]. However, further evidence is required in support of this notion. Overall these results are consistent with the notion that viruses in general either induce or suppress autophagy to benefit the infectious process [20], [21]. Although recent studies describe the involvement of bulk autophagy in various steps of HCV lifecycle, our understanding of the biological significance and the precise role of autophagy in HCV lifecycle are still rudimentary.
HCV infection is associated with physiological insults like ER stress, oxidative stress, ROS accumulation, and mitochondrial Ca2+ overload that can trigger collapse of mitochondrial transmembrane potential (ΔΨm) and subsequent mitochondrial dysfunction [5], [22]–[26]. Quality control of the dysfunctional mitochondrial is essential to sustain the bioenergetic efficiency and to prevent the initiation of mitochondria-mediated intrinsic cell death signaling cascade [27].
In humans, loss of function and mutations in genes encoding Parkin and PINK1 are linked to autosomal recessive form of Parkinson's disease [28], [29]. Recent advances have unraveled their functional role in selective autophagic elimination of damaged mitochondria, termed as mitophagy [30]. Parkin is an E3 ubiquitin ligase, which is normally localized in the cytoplasm. However, upon mitochondrial stress, it is rapidly recruited to the damaged mitochondria [28], [31], [32]. PINK1, a mitochondrial Ser/Thr kinase, recruits Parkin on depolarized mitochondria and interacts with mitochondrial outer membrane complex to regulate Parkin translocation and activation [30], [33].
In this study, we investigated the involvement of mitophagy in HCV-infected cells. Our results demonstrate that HCV induces Parkin-dependent mitophagy. HCV infection dramatically triggered Parkin translocation to mitochondria, which was convincingly demonstrated both by confocal microscopy and subcellular fractionation of highly pure mitochondria. HCV stimulated Parkin and PINK1 gene expression at transcriptional level. Electron microscopy of HCV-infected cells displayed engulfment of damaged mitochondria in double membrane vesicles and subsequent fusion of these mitochondria-containing vesicles with the lysosome. These results were further strengthened by immunoelectron microscopy. In HCV-infected cells, Parkin associated with mitochondria was ubiquitinated and promoted the degradation of its mitochondrial substrates. Further, our results showed that silencing Parkin and PINK1 hindered HCV replication suggesting the functional relevance of mitophagy in HCV replication. HCV-induced mitophagy was also directly associated with HCV-mediated decline in mitochondrial complex I enzyme activity. Overall our results implicate that HCV-induced mitophagy is physiologically relevant in maintenance of cellular homeostasis, persistence of HCV infection, and liver disease pathogenesis.
Parkin translocation to mitochondria is a well-characterized event that triggers the induction of mitophagy [30]. To investigate HCV-induced mitophagy, we assessed mitochondrial translocation of Parkin in cell culture-derived HCV Jc1 strain (hereafter referred to as HCVcc) infected human hepatoma Huh7 cells by immunofluorescence imaging [34]. In these images, a significant perinuclear clustering of mitochondria in HCV-infected cells was observed (Figure 1A). To further strengthen this observation, we conducted electron microscopy (EM) using Huh7 cells harboring HCV full-length replicon (FLR-JFH1) [35] (Figure 1B) or those infected with HCVcc (Figure S2). EM analysis of these cells revealed prominent clustering of mitochondria in the perinuclear regions of HCV-infected cells, displaying a dramatic loss of mitochondrial cristae (Figures 1B, 1C, S1A and S1C). In contrast, in the uninfected cells typical cytoplasmic distribution of mitochondria with intact cristae was observed (Figures 1A and B). Oxidative stress has been shown to induce the spheroid formation of damaged mitochondria [36]. Using immunoelectron microscopy, we noted similar perinuclear clustering of impaired mitochondria as well as mitochondrial spheroid formation (Figures S1C and S1D).
Cyanide m-cholorophenyl hydrazone (CCCP) triggers mitochondrial translocation of Parkin [32]. Mitochondrial translocation and aggregation of Parkin was observed in CCCP-treated Huh-7 cells (Figure S2) as well was in HCV-infected (Figure 1A). Quantitative analysis of mitochondrial translocation of Parkin is presented in Figure 1D. HCV-induced Parkin translocation was further investigated by immunoelectron microscopy as described in Fig. S11A. HCV infected cells were stained with antibodies specific to TOM20, HCV E2, and Parkin, respectively and treated with secondary antibodies conjugated with 18-nm, 12-nm and 6-nm gold particles, respectively. As can be seen, Parkin is localized to damaged mitochondria with notable loss of cristae (Figure S11A).
Next, we utilized purified mitochondrial and cytosolic fractions of HCV-infected and uninfected Huh7 cells to examine Parkin translocation. Western blot analysis of these fractions is presented in Figure 1E. Parkin was highly enriched and predominantly ubiquitinated in the pure mitochondrial fraction (pMito) of HCV-infected cells (Figure 1E). This result was further substantiated by immunoprecipitation of Parkin in pMito fraction with anti-Parkin antibody followed by subsequent immunoblotting with anti-ubiquitin antibody (Figure 2A). Further, we carried out immunoprecipitation with anti-ubiquitin antibody using whole cell lysates of HCV-infected cells followed by subsequent immunoblotting with anti-Parkin antibody. As shown in Figures 2A and 2B, significant levels of ubiquitinated endogenous Parkin were detected during HCV infection. Substantial levels of Parkin ubiquitination on mitochondria were also observed in HCV-infected cells by immunofluorescence analysis (Figure 2C). Quantitative analysis of endogenous Parkin ubiquitination on mitochondria is presented in Figure 2D. Parkin is an E3 ubiquitin ligase, which exerts its E3 ubiquitin ligase activity by ubiquitinating itself and other mitochondrial target proteins [30], [31]. A mitochondrial outer membrane protein mitofusin 2 (Mfn2) and voltage-dependent anion-selective channel 1 (VDAC1) are known substrates of Parkin [30], [37]. Indeed, both Mfn2 and VDAC1 levels were reduced concomitant to increased Parkin levels during HCV infection (Figure 2E). To examine whether HCV infection induces Parkin-mediated ubiquitination of Mfn2 and VDAC1, we performed immunoprecipitation of HCV infected cells with anti-Mfn2 and VDAC1 antibodies followed by subsequent immunoblotting with anti-ubiquitin antibody (Figure 2G). This analysis revealed that both Mfn2 and VDAC1 were significantly ubiquitinated (Figures S4 and S5). This is further strengthened by confocal images of HCV infected cells showing enhanced ubiquitination of Mfn2 and VDAC1 (Figures S4 and S5). Furthermore, these increased ubiquitination of Mfn2 and VDAC1 were attenuated by Parkin silencing. Thus, these results indicate that HCV infection enhances Parkin-mediated ubiquitination of its substrates (Figure 2G).
p62/Sequestosome 1 (p62) has been previously shown to cooperate with Parkin during perinuclear clustering of mitochondria and required for Parkin-mediated mitophagy [38], [39]. Thus, we investigated the degradation of p62 during HCV infection. As shown in Figure 2F, Western blot analysis of HCV infected cells exhibited a decrease in p62 level. We also observed that HCV infection enhances the interaction of p62 and Parkin on mitochondria and results in subsequent increase of p62 ubiquitination in HCV-infected cells (Figures S6 and S7).
Interestingly, a majority of LC3B protein enriched in the mitochondrial fraction of HCV-infected cells was lipidated, which indicates the formation of phagophore involving mitochondria, an early event of mitophagy [30] (Figure 1E). Consistent with previous reports, HCV core is also enriched in mitochondria (Figure 1E) [23], [40]–[43]. We also examined mitochondrial translocation of Parkin in Huh7 cells harboring HCV 2a full-length (FLR-JFH1) or subgenomic replicon (SGR-JFH1) and Huh7.5.1 cells harboring HCV 1b subgenomic replicon (BM4–5 Feo) [35], [44], [45] (Figure S3). Parkin recruitment to the perinuclear mitochondrial clusters was uniformly observed in full-length and subgenomic replicons of genotype 2a and 1b subgenomic replicon cells, suggesting that this phenomenon is independent of the differences in HCV genotypes.
We then examined if HCV infection stimulates the expression of Parkin and PINK1, two key regulators of mitophagy. As shown in Figures 3A and B, HCV infection stimulated both the mRNA and protein levels of Parkin and PINK1. CCCP-treatment also stimulated Parkin and PINK1 gene expression, in agreement with previous reports [46], [47] (Figure 3B). Further, confocal microscopy also revealed that Parkin protein expression is stimulated in HCV-infected cells compared to uninfected cells (Figure 3D). ATF4 is the transcription factor that is activated during UPR, which promotes Parkin upregulation both by ER and mitochondrial stress [46]. Increased ATF4 expression has also been reported in HCV chronic liver tissues [48]. In the present analysis, we observed ATF4 transcriptional stimulation in HCV infected cells, in agreement with previous reports [49], [50] (Figure 3A). Parkin expression was also analyzed in liver biopsies of chronic hepatitis C patients. We observed a consistent increase in Parkin protein level in the liver biopsies of all the chronic hepatitis C patients compared to non-hepatitis C donors (Figure 3C).
To analyze HCV-induced mitophagosome formation, Huh7 cells ectopically expressing GFP-LC3 were infected with HCVcc and analyzed by immunofluorescence imaging. We observed numerous GFP-LC3 puncta in HCV-infected cells (Figure 4A). Analysis of mitophagosome formation by merge of GFP-LC3 puncta with TOM20, a mitochondrial marker, revealed significantly higher number of mitophagosomes in HCV-infected cells compared to uninfected cells (Figures 4A and B). We further observed the colocalization of Parkin with mitophagosomes, which signifies the role of Parkin in mediating mitophagosome formation (Figure 4A). To strengthen these observations, we established stable cell line expressing Parkin-specific shRNA (P-KD). Using P-KD cells ectopically expressing GFP-LC3, we observed only a few GFP-LC3 puncta in HCV infected cells that do not localize to mitochondria, signifying the role of Parkin in mitophagy (Figure S8). Whereas, in non-targeting shRNA (NT-KD) stable cells infected with HCV, mitophagosome formation was observed (Figure S8). 3-methlyadenine (3-MA) inhibits the phagophore formation, whereas Bafilomycin A1 (BafA1), a specific inhibitor of vacuolar-type H+-ATPase, blocks the downstream step of fusion between the autophagosomes and lysosomes, resulting in the accumulation of undegraded autophagosomes (Figure S9) [51], [52]. 3-MA treatment of HCV-infected cells significantly reduced not only the number of total GFP-LC3 puncta but also mitophagosome-specific GFP-LC3 puncta, whereas BafA1 treatment resulted in their accumulation (Figures 4A and B). Quantitative analysis of GFP-LC3 puncta associated with TOM20 and Parkin is summarized in Figure 4C. The proteolytic cleavage and transient lipidation of LC3 (conversion from LC3-I to LC3-II) with phosphatidylethanolamine is essential for phagophore expansion [53]. In correlation with earlier studies, conversion ratio of LC3-II/LC3-I with anti-LC3B antibody is increased in HCV-infected cells compared to uninfected cells (Figure 1E, see the short exposure panel). To confirm mitophagosome formation in HCV-infected cells, we performed electron microscopy using Huh7 cells harboring HCV full-length replicon (FLR-JFH1) or those infected with HCVcc. Ultrastructural analysis of HCV-infected cells revealed damaged mitochondria, displaying traces of cristae or no cristae (empty) (Figures 1B, S1, and 4D). The empty mitochondria or mitochondria with traces of cristae are engulfed by double membrane vesicles in HCV-infected cells (Figure 4D). In contrast, uninfected Huh7 cells displayed normal mitochondrial morphology with typical cristae (Figure 4D). Formation of mitophagososmes was further analyzed by immunoelectron microscopy. HCV infected cells were labeled with antibodies specific to GFP-LC3, HCV E2, and TOM20, respectively and subsequently treated with secondary antibodies conjugated with 18-nm, 12-nm and 6-nm gold particles, respectively. As shown in Figures 4E and S11B, GFP-LC3 colocalizes on the damaged mitochondria in HCV-infected cells, thus confirming HCV-induced formation of mitophagosome. Together, these results strongly demonstrate that HCV-infected cells display higher incidence of Parkin-mediated mitophagosome formation.
Next, we determined the formation of mitophagolysosome or the fusion of mitophagosome with lysosomes. HCV-infected cells transiently expressing GFP-LC3 were stained with LysoTracker and/or lysosomal-associated membrane protein 2 (LAMP2). Immunofluorescence analysis revealed that HCV-infected cells displayed several distinct lysosomes containing mitochondria-associated with GFP-LC3 in comparison to uninfected cells (Figure 5A). The analysis of colocalization between lysosomes (immunostained with LAMP2) and Parkin-associated GFP-LC3 puncta in HCV-infected cells revealed several distinct lysosomes that were positive for Parkin and GFP-LC3 (Figure S10). In the presence of either 3-MA or BafA1, such colocalizations were severely reduced (Figures 5A and S10). In support of these observations, we performed electron microscopy using HCV full-length replicon-bearing Huh7 cells. HCV-infected cells revealed fusion of lysosomes with double membrane vesicles containing damaged mitochondria (Figure 5B). Next, we conducted immunoelectron microscopy of HCVcc infected Huh7 cells. Cells were labeled with antibodies specific to LAMP1, HCV E2, and Parkin, respectively and treated with secondary antibodies conjugated with immunogold 18-nm, 12-nm, and 6-nm gold particles, respectively. As shown in Figure S11C, fusion of LAMP1-positive lysosome and damaged mitochondria containing Parkin was observed in HCV-infected cells. These observations confirm and support results of confocal immunofluorescence imaging that HCV infection induces the formation of mitophagolysosome. Quantitative analyses of colocalization between lysosomes and GFP-LC3 puncta-associated with mitochondria or Parkin are represented in Figures 5C and D. Quantitative analysis of mitochondria-specific fluorescence intensity revealed significant reduction of mitochondria in HCV-infected cells that was restored to normal levels by treatment with both 3-MA and BafA1 (Figure 5E). As previously described, CCCP treatment also resulted in a decline in the number of mitochondria [54] (Figure 5E). Together, these results clearly suggest that mitophagosomes formed in HCV-infected cells subsequently fuse with lysosomes leading to the formation of mitophagolysosomes.
To evaluate the functional role of Parkin and PINK1 in the HCV infectious process, we employed the siRNA strategy to silence the gene expression of Parkin and PINKI. We first verified silencing of Parkin and PINK1 gene expression by qRT-PCR analysis (Figure 6B). Analysis of HCV RNA replication in the presence of respective gene-specific siRNAs shows that Parkin and PINK1 silencing effectively inhibited HCV replication (Figure 6A). Atg5, a key regulator of phagophore elongation, has been previously shown to be involved in HCV replication [10], [55]. Atg5 silencing by siRNA inhibited HCV RNA replication in the cells infected with HCVcc (Figures 6A and B).
To further study silencing effect of Parkin on HCV replication, we analyzed HCV replication using P-KD cells infected with HCVcc. Consequently, HCV RNA replication levels declined in P-KD cells compared to NT-KD cells (Figure 6C). Parkin-specific shRNA did not affect cellular viability as determined by the resazurin-based cytotoxicity (Figures S12). We then performed a Parkin rescue experiment in HCV-infected P-KD cells. Dose-dependent ectopic expression of shRNA-resistant Parkin rescued the inhibition effect of Parkin silencing on HCV RNA replication (Figure 6C). In correlation with the rescue of HCV RNA replication by ectopic expression of shRNA-resistant Parkin, we observed corresponding stimulation of HCV core expression (Figure 6D). These results directly implicate a functional role of Parkin and PINK1 in HCV replication.
Previous reports have shown that HCV infection affects mitochondrial oxidative phosphorylation [23], [24], [56]. Parkin is also shown to influence mitochondrial oxidative phosphorylation by promoting the degradation of mitochondrial outer membrane proteins [30]. Hence, we speculated that HCV triggered decline in oxidative phosphorylation is a consequence of Parkin-mediated mitophagy triggered by HCV infection. We measured the mitochondrial respiratory chain complex I enzyme (complex 1) activity in HCV-infected cells in the presence and absence of 3-MA and BafA1. HCV infection resulted in the reduction of mitochondrial complex I activity, which was restored by treatment of both 3-MA and BafA1 (Figure 7A). Consistent with previous reports [23], [24], we observed that HCV infection leads to about 20–25% decline in mitochondrial complex 1 activity (Figure 7A). To determine the effect of Parkin on HCV-mediated decline of mitochondrial complex I activity, P-KD and NT-KD cells were infected with HCVcc and mitochondrial complex I activity determined. HCV infection decreased mitochondrial complex I activity in NT-KD cells, whereas P-KD cells failed to show any decrease in mitochondrial complex I activity (Figure 7B). HCV has been previously shown to downregulate the expression of mitochondrial complex I and IV enzymes [57]. To further strengthen these results, we pursued the analysis of expression levels of complex I and IV enzymes in HCV infected cells in the context of Parkin. Consistent with previous reports, HCV infection affected the expression levels of both mitochondrial complex I and IV enzymes in NT-KD cells [57]. In contrast, the expression levels of mitochondrial complex I and IV enzymes in P-KD cells were unaffected by HCV infection (Figure 7C). Changes in the expression of mitochondrial complex I and IV enzymes were also detected in CCCP-treated Huh7 cells. It has been previously shown that chronic HCV infection results in the depletion of mitochondrial DNA (mtDNA) in the liver [58]. Here, we examined the effect of Parkin silencing on HCV-induced mtDNA depletion. Analysis of mitochondrial NADH dehydrogenase-2 (ND-2) and cytochrome c oxidase-2 (COX-2) in the presence of specific siRNA for Parkin shows that Parkin silencing effectively blocked HCV-induced decline of mtDNA levels (Figure 7D). Taken together, these results indicate that HCV-induced mitophagy is functionally associated with HCV-mediated impairment of oxidative phosphorylation and depletion of mitochondria.
Autophagy involves clearance of protein aggregates, damaged mitochondria, peroxisosmes and bacteria and viruses [53]. A growing body of literature on autophagy implicates its role in cellular homeostasis, innate immunity, defense mechanisms against lethal entities and in maintenance of persistent viral infections [21]. Pathogens have developed strategies to usurp autophagic pathways and hijack the machinery to favor viral proliferation and persistent infection [20]. Several reports described the HCV-induced events of bulk autophagy and have linked this pathway to aiding viral proliferation [9], [10], [12], [13], [17], [55]. The requirement of autophagy was emphatically shown by silencing key components of autophagy (Atg5, Atg7, Beclin-1, Atg8) respectively, and observing a decline in viral translation and replication process [9], [10], [12], [13], [17], [55]. Here, for the first time, we demonstrate that HCV induces organelle selective autophagic degradation of mitochondria (mitophagy). This was shown by marked translocation of Parkin to mitochondria (Figures 1, S3, and S11) in HCV-infected cells. Parkin translocation to mitochondria is considered a hallmark of mitophagy [30]. We also observed a significant stimulation of Parkin expression in both HCV-infected Huh7 cells and liver tissues samples obtained from chronic HCV patients (Figure 3). Similarly, HCV also stimulated the expression of PINK1, another key component of mitophagy involved in recruitment of Parkin to damaged mitochondria and its subsequent activation.
Diverse independent reports suggest the importance of autophagy machinery in multiple steps of HCV life cycle such as, translation, replication and secretion raising considerable controversy in depicting the precise role of autophagy in HCV infection [7]–[17]. In addition, some previous reports also claim that HCV induces incomplete bulk autophagy and prevents the fusion of autophagosomes with lysosomes [13]. Ke and Chen showed that HCV promotes complete autophagy, which culminates in the formation of autophagolysosome and that this event is crucial for viral replication [10]. Here, we also show the formation of mitophagosomes and mitophagolysosomes as evidenced by the presence of lipidated LC3B in pure mitochondrial fraction and the colocalization of GFP-LC3 puncta with Parkin loaded mitochondria and the subsequent fusion/colocalization of these mitophagosomes with lysosomes in HCV-infected cells (Figures 1, 4, 5, and S10, S11). To further reinforce our observation we performed ultrastructural and immunogold-labeled electron microscopy of HCV infected cells. In correlation to our earlier observations the electron microscopy data clearly suggests that HCV infection induces the perinuclear accumulation of damaged mitochondria followed by subsequent formation of mitophagosomes and mitophagolysosomes in HCV-infected cells (Figures 1B, 4D, 4E, 5B, and S1, S11). It should be noted that in HCV-infected cells, we observed a few mitophagolysosome at a given time point, because of the rapid turnover rate of the mitophagic process. Further, unlike any chemical treatment, extensive autophagic puncta were not seen in HCV-infected cell, suggesting that HCV tightly regulates the autophagic turnover. This observation is consistent with previous reports on HCV-associated autophagy [7]–[17]. Most importantly, in agreement with previous reports on bulk autophagy, specific inhibition of mitophagy was detrimental to viral replication (Figure 6). Parkin and PINK1 silencing affected viral RNA synthesis, thus implicating a functional role of mitophagy during HCV RNA replication process. HCV RNA synthesis is believed to occur on ER-derived membranous web like structures [3], but a detailed characterization of such structures is still lacking. Recent reports suggest that there is segregation of ATP pools at sites of HCV replication near cellular perinuclear region [59] and that autophagosomal membrane may serve as platforms for HCV replication [19]. In our study, HCV-infected cells displayed peculiar Parkin-dependent clustering of mitochondria near the perinuclear region as shown by confocal and electron microscopy. We surmise if such Parkin-dependent mitochondrial clustering serves to segregate ATP pools at HCV replication sties.
Previous studies have shown that HCV infection triggers ROS generation, which leads to loss of mitochondrial transmembrane potential (ΔΨm) and decline in mitochondrial complex I activity [23], [24], [56]. A decrease in mitochondrial complex I activity in HCV infection has been shown to be a direct consequence of ROS generation and oxidation of mitochondrial glutathione pools and glutathionylation of mitochondrial complex I subunits [23]. Our results suggest that HCV induced Parkin-mediated mitophagy also causes the reduction of mitochondrial complex I activity in HCV infection. Knockdown of Parkin in HCV-infected cells restored the levels of mitochondrial complex I activity to the levels observed in uninfected cells (Figure 7). This return to normalcy may also be attributed to reduction of viral replication and protein synthesis in Parkin knockdown cells.
Chronic hepatitis B and C is associated with high mitochondrial ROS levels, ER ballooning and mitochondrial swelling [56], [60]. This observation lends support to our studies described here on mitophagy which marks damaged mitochondria for degradation. A recent report demonstrates a decrease in the number of mitochondria in the HCV-infected cells [8], [59]. Similarly, in our study, a decrease in mitochondria number in HCV infected cells was observed that was restored by treatment with both 3-MA and BafA1 (Figure 5E).
Autophagy and/or mitophagy appear to play an essential role during HCV infectious process. The process both removes damaged organelles and also promotes the survival and maintenance of persistently infected hepatocytes. Thus, these studies on selective mitophagy provide unique insight into the HCV associated liver disease pathogenesis and offer new avenues for the design of antiviral strategies.
Human hepatoma cell lines Huh7 and Huh7.5.1 used in this study were grown in high-glucose DMEM (Gibco) supplemented with 10% fetal bovine serum (Hyclone), 1% MEM non-essential amino acids (Gibco), 100 units/ml penicillin (Gibco), and 100 µg/ml streptomycin (Gibco). Cell culture-derived HCV Jc1 genotype 2a (HCVcc) used in this study was propagated and prepared, as described previously [35]. HCVcc infection in this study was carried out at multiplicity of infection (MOI) of 1. Stable HCV replicon cells harboring full-length FLR-JFH1 (genotype 2a) and two subgenomic SGR-JFH1 (genotype 2a) and BM4–5 Feo (genotype 1b), respectively, were maintained in the presence of 0.4 mg/ml G418 (Invitrogen).
The cells were grown on glass cover slips, fixed in 4% paraformaldehyde, washed, and then permeabilized with 0.25% TritonX-100. MitoTracker CMXRos Red (Invitrogen) and LysoTracker (Invitrogen) were used to stain mitochondria and lysosomes in live cells before fixation. The cells were stained with the indicated antibodies. Wherever indicated, nuclei are stained with DAPI (Invitrogen). Images were visualized under a 100× oil objective using an Olympus FluoView 1000 confocal microscope. Quantification of images was conducted with ImageJ or MBF ImageJ softwares.
The pEGFP-LC3 plasmid DNA used in this study was a kind gift from Dr. Tamotsu Yoshimori (National Institute of Genetics, Japan). The pRK5-HA-Parkin plasmid DNA (plasmid ID: 17613) was purchased from Addgene.
Chemical reagents used in this study were Bafilomycin A1 (Enzo Life Sciences), cyanide m-cholorophenyl hydrazone and 3-Methyladenine (Sigma). Primary antibodies used in this study include the following: rabbit monoclonal anti-LC3B (Cell Signaling); rabbit polyclonal anti-Parkin (Abcam); rabbit polyclonal anti-PINK1 (Abcam); rabbit polyclonal anti-VDAC1 (Cell Signaling); rabbit polyclonal anti-Calreticulin (Cell Signaling); rabbit polyclonal anti-GAPDH (SantaCruz); goat polyclonal anti-β-actin (SantaCruz); mouse monoclonal anti-TOM20 (BD); goat polyclonal anti-TOM20 (SantaCruz); rabbit polyclonal anti-TOM20 (Abcam); mouse monoclonal anti-Mitofusin2 (Mfn2) (Abcam); mouse monoclonal anti-SQSTM1/P62 (Abcam); mouse monoclonal anti-Ubiquitin (Cell Signaling); rabbit polyclonal anti-Ubiquitin (SantaCruz); mouse monoclonal anti-LAMP2 (Abcam); rabbit monoclonal anti-LAMP1 (Abcam); mouse monoclonal anti-HA (Roche); normal rabbit IgG (Cell Signaling); normal mouse IgG (SantaCruz); mouse monoclonal anti-human total OxPhos complex (Invitrogen); mouse monoclonal anti-HCV core (Thermo Scientific); human monoclonal anti-HCV E2 [61]; mouse monoclonal anti-HCV NS5A (clone 9E10). The secondary antibodies used for immunofluorescence were Alexa Fluor 350, 488, 594, or 647 donkey anti-mouse, rabbit, or goat IgG (Molecular Probe), Alexa Fluor 555 goat anti-human IgG (Molecular Probe), and Alexa Fluor 568 goat anti-mouse IgG (Molecular Probe). The secondary antibodies used for immunoblot analysis were HRP-conjugated anti-mouse IgG (Cell Signaling), HRP-conjugated anti-rabbit IgG (Cell Signaling), and HRP-conjugated anti-goat IgG (Jackson Laboratories).
siGENOME SMARTpool small interfering RNA (siRNA) for Parkin (NM_004562), PINK1 (NM_032409), ATG5 (NM_004849), and non-targeting #1 control (NT) were used in this study (Dharmacon). Huh7 cells were transfected with siRNAs (50 nM) using DharmaFECT 4 transfection reagent according to the manufacturer's instructions (Dharmacon), prior to HCVcc infection.
To establish stable human hepatoma cell line expressing Parkin-specific shRNA (P-KD), Huh7 cells were transfected with shRNA construct (pLKO.1-puro/Parkin, Sigma) encoding siRNA targeting Parkin using TransIT-LT1 transfection reagent (Mirus, Madison, WI) according to the manufacturer's instructions and subsequently, selected in the presence of 5 µg/ml of puromycin for 3 weeks. Parkin shRNA sequence is as follows: TRCN0000000281; CCGGCGTGAACATAACTGAGGGCATCTCGAGATGCCCTCAGTTATGTTCACGTTTTT. Two stable M-KD and NT-KD cell lines expressing empty vector (pLKO.1-puro, Sigma) and nontargeting shRNA (pLKO.1-puro/non-targeting, Sigma), respectively, were also established as a negative controls. All cells were maintained in the presence of 2.5 µg/ml of puromycin. The knockdown level of Parkin gene was analyzed by immunoblotting with a specific antibody against Parkin.
For Western blot analysis, cells were resuspended in RIPA buffer (20 mM Tris-HCl [pH 7.5], 150 mM NaCl, 50 mM NaF, 1 mM Na3VO4, 0.1% SDS, and 0.5% TritonX-100) supplemented with a Halt protease inhibitor cocktail (Thermo Scientific). The whole cell lysates (WCL) were subjected to SDS-PAGE, transferred to nitrocellulose membrane (Thermo Scientific), and Western blot analyzed with antibodies against the indicated proteins.
For immunoprecipitation of the ubiquitinated Parkin in the pure mitochondria fraction, 100 µg of the purified cytosolic and mitochondrial fractions were resuspended in RIPA buffer without SDS. The resuspended mixtures were immunoprecipitated with anti-Parkin antibody and protein-G Sepharose (GE Healthcare) followed by Western blotting with anti-ubiquitin antibody.
For immunoprecipitation of the ubiquitinated Parkin, Mfn2, and VDAC1 in WCL, Huh7 cells infected with HCVcc were suspended in 0.1 ml of RIPA buffer. The suspended cells were incubated for 20 min on ice and clarified by centrifugation at 15,000×g at 4°C for 20 min. The supernatant was mixed with 1.9 ml of RIPA buffer without SDS and immunoprecipitated with anti-ubiquitin, Mfn2, and VDAC1 antibodies, respectively, and protein-G Sepharose followed by Western blotting with anti-Parkin and ubiquitin antibodies, respectively.
To analyze the expression levels of Parkin, PINK1, and ATG5 genes, total cellular RNA was extracted from cells using RNAeasy Mini kit (Qiagen) and subsequently, complementary DNAs (cDNAs) was synthesized by using SuperScript III First-Strand Synthesis System with oligo(dT)20 primer (Invitrogen) according to the manufacturer's instructions, respectively. DyNAmo HS SYBR Green qPCR kit (Finnzymes) was used to quantify the cellular RNA levels. The following primer sets were used for RT-PCR: ATF4 forward, 5′-AGTCCCTCCAACAACAGCAA; ATF4 reverse, 5′-GAAGGTCATCTGGCATGGTT; Parkin forward, 5′-TACGTGCACAGACGTCAGGAG; Parkin reverse, 5′-GACAGCCAGCCACACAAGGC; PINK1 forward, 5′-GGGGAGTATGGAGCAGTCAC; PINK1 reverse, 5′-CATCAGGGTAGTCGACCAGG; ATG5 forward, 5′-GCCATCAATCGGAAACTCAT; ATG5 reverse, 5′-ACTGTCCATCTGCAGCCAC; GAPDH forward, 5′-GCCATCAATGACCCCTTCATT; and GAPDH reverse, 5′-TTGACGGTGCCATGGAATTT. HCV RNA levels were quantified by qRT-PCR, as described previously [35].
To analyze the expression levels of mitochondrial DNA, total cellular DNA was extracted from the cells using AllPrep DNA kit (Qiagen) and subsequently quantified by qPCR using DyNAmo HS SYBR Green qPCR kit according to the manufacturer's instructions. The following primer sets were used for qPCR: ND-2 forward, 5′-TAGCCCCCTTTCACTTCTGA; ND-2 reverse, 5′-GCGTAGCTGGGTTTGGTTTA; COX-2 forward, 5′-GGCCACCAATGGTACTGAAC; COX-2 reverse, 5′-CGGGAATTGCATCTGTTTTT. Real-time qPCR was conducted by using an ABI PRISM 7000 Sequence Detection System (Applied Biosystems).
To isolate pure mitochondrial fraction from HCV-infected cells, Huh7 cells were infected with HCVcc. At 5 days post-infection, cells were homogenized and then pure cytosolic and mitochondrial fractions were isolated by Percoll gradient fractionation, as described previously [62]. Equivalent amounts of protein from each fraction were analyzed by Western blotting with the indicated antibodies.
The activity of mitochondrial oxidative phosphorylation respiratory chain complex I (NADH dehydrogenase) in HCV-infected cells was measured by using mitochondrial complex I activity assay kit according to the manufacturer's instructions (Novagen). Briefly, Huh7, NT-KD, and P-KD cells were infected with HCVcc for 3 days. Huh7 cells were treated with 3-MA (10 mM) and BafA1 (100 nM) for 12 h before harvest. 500 µg of the detergent-soluble WCL were used for this assay.
To assess cytotoxic effects of Parkin silencing during HCV infection, NT-KD and P-KD cells infected with HCVcc for 3 days were incubated with 10% resazurin solution (TOX-8, Sigma) for 4 hours at 37°C and then cell viability was measured according to the manufacturer's instructions.
Briefly, Huh7 and HCV-infected cells grown in 10 cm dishes were washed and fixed with fixative containing 2% glutaraldehyde in 0.1 M sodium cacodylate buffer [pH 7.4]. Cell pellets were embedded in 2% agarose, post-fixed with 1% osmium tetroxide, and dehydrated with an acetone series. Samples were infiltrated, embedded in Durcupan and polymerized at 60°C for 48 h. Ultrathin sections were prepared and examined using a JEOL 1200 EX II transmission electron microscope at 80 kV. HCV infected cells were stained with indicated antibodies and treated with secondary antibodies conjugated with indicated immunogold particles for immunoelectron microscopy.
The frozen human liver biopsy specimens (n = 7) were a kind gift from Dr. Tarek Hassanein, UCSD hepatology clinic. The liver biopsy specimens used in this study include the following: anti-HCV negative normal patient, n = 1; anti-HCV negative patient with hepatocarcinoma, n = 1; anti-HCV positive patients with chronic HCV including 3 females and 2 male, n = 5.
Statistical analyses using unpaired Student's t-test were performed by using Sigma Plot software (Systat Software Inc., San Jose, CA, USA).
|
10.1371/journal.pntd.0003385 | Bartonella henselae Endocarditis in Laos – ‘The Unsought Will Go Undetected’ | Both endocarditis and Bartonella infections are neglected public health problems, especially in rural Asia. Bartonella endocarditis has been described from wealthier countries in Asia, Japan, Korea, Thailand and India but there are no reports from poorer countries, such as the Lao PDR (Laos), probably because people have neglected to look.
We conducted a retrospective (2006–2012), and subsequent prospective study (2012–2013), at Mahosot Hospital, Vientiane, Laos, through liaison between the microbiology laboratory and the wards. Patients aged >1 year admitted with definite or possible endocarditis according to modified Duke criteria were included. In view of the strong suspicion of infective endocarditis, acute and convalescent sera from 30 patients with culture negative endocarditis were tested for antibodies to Brucella melitensis, Mycoplasma pneumoniae, Bartonella quintana, B. henselae, Coxiella burnetii and Legionella pneumophila. Western blot analysis using Bartonella species antigens enabled us to describe the first two Lao patients with known Bartonella henselae endocarditis.
We argue that it is likely that Bartonella endocarditis is neglected and more widespread than appreciated, as there are few laboratories in Asia able to make the diagnosis. Considering the high prevalence of rheumatic heart disease in Asia, there is remarkably little evidence on the bacterial etiology of endocarditis. Most evidence is derived from wealthy countries and investigation of the aetiology and optimal management of endocarditis in low income countries has been neglected. Interest in Bartonella as neglected pathogens is emerging, and improved methods for the rapid diagnosis of Bartonella endocarditis are needed, as it is likely that proven Bartonella endocarditis can be treated with simpler and less expensive regimens than “conventional” endocarditis and multicenter trials to optimize treatment are required. More understanding is needed on the risk factors for Bartonella endocarditis and the importance of vectors and vector control.
| Infection of heart valves (endocarditis) with bacteria is an important condition, especially afflicting those with rheumatic heart disease, and has a high mortality if untreated. Most of the evidence for optimal antibiotic and surgical management comes from wealthy countries. There are no published data from poorer countries in SE Asia despite a high burden of rheumatic heart disease. We investigated the bacterial infections of heart valves in the Lao PDR (Laos) through heart ultrasound scans and analysis of patients' blood. We provide evidence of infection with the poorly understood bacteria Bartonella henselae (the cause of cat scratch disease) in two patients from Laos. We argue that it is likely that Bartonella endocarditis is more widespread than appreciated, as there are few laboratories in Asia able to make the diagnosis. This is important as it is likely that proven Bartonella endocarditis can be treated with simpler and less expensive regimens than “conventional” endocarditis. There have been great advances in the wealthy world in the diagnosis and treatment of endocarditis but these have not been assessed or implemented in poorer countries. More evidence on the causes and optimal management of endocarditis in low income countries is needed.
| There has been emerging interest in the importance of Bartonella henselae endocarditis in the wealthier countries of Asia [1]–[3] but few data from poorer countries. We describe two patients admitted with this condition at Mahosot Hospital, Vientiane, Lao PDR (Laos), and discuss the public health implications.
As part of a blood culture and infectious disease liaison service at Mahosot Hospital we identified thirty patients with culture negative endocarditis 2006–2012 [4]. Mahosot Hospital (17.960 N, 102.612 E) is a primary-tertiary care teaching hospital of ∼400 beds including cardiology and infectious disease wards. Patients were identified through liaison between the microbiology laboratory and the wards, especially with those performing cardiac ultrasound. Those aged >1 year admitted to Mahosot Hospital with definite or possible endocarditis according to modified Duke criteria were included in a retrospective study 2006–2012 and, since then, in a prospective study. The hospital has trans-thoracic echocardiography, with occasional trans-oesophageal echocardiography. Blood cultures were performed as described [5]. The clinical significance of positive blood cultures was determined by physicians at the time of the result. Acute and convalescent (when available) sera from those with culture negative endocarditis were tested for antibodies to Brucella melitensis, Mycoplasma pneumoniae, Bartonella quintana, B. henselae, Coxiella burnetii and Legionella pneumophila were tested for by indirect immunofluorescence assay (IFA) [6]. Sera exhibiting phase 1 immunoglobulin G (IgG) titers >1∶800 for C. burnetii or IgG titers >1∶800 for Bartonella quintana and B. henselae were considered as highly predictive of endocarditis caused by these microrganisms. We also considered total antibody titers≥1∶256 for L. pneumophila as positive. Specific antibodies to Brucella melitensis and Mycoplasma pneumoniae were detected with an immunoenzymatic antibody test and the Platellia M. pneumoniae IgM kit (Bio-Rad, Marnes-la-Coquette, France), respectively. Titers of >1∶200 were considered positive. When the results of the tests described above were negative, or when IgG titers to Bartonella were ≤1∶800, we performed Western blot using B. henselae and B. quintana antigens, followed by cross-adsorption as described [6]. Patients with Bartonella endocarditis exhibit specific Western blot profiles, different from other types of Bartonella infections [7].
Patients gave informed written consent for a prospective description of the causes of infection approved by the Oxford Tropical Research Ethics Committee, UK, and the National Ethics Committee for Health Research, Laos.
Of the 30 patients tested, two were Western Blot positive for B. henselae. They had definite and possible endocarditis, respectively, by modified Duke criteria.
In 2012 a previously healthy 57-year-old army officer from Pakse, southern Laos, was admitted with one month of fever, headache, myalgia, back pain, productive cough and 4 days of chills and dyspnea. On examination he was afebrile, normotensive but with a pansystolic (3/6) murmur at the mitral and tricuspid areas with clear lungs and no peripheral signs of endocarditis. His admission peripheral blood count was white blood count (WBC) 7.2×109/L, haemoglobin 7.4 g/dL, mean cell volume 79 fL, mean cell haemoglobin 24.6 pg and platelets 159×109/L. Transthoracic echocardiogram showed a vegetation on the mitral valve (maximum length 1.9 cm), with mild mitral regurgitation, mild aortic and tricuspid valve regurgitation. Three sets of blood cultures incubated for 7 days showed no growth. The patient exhibited IgG titers of 1∶400 to both B. henselae and B. quintana in acute serum, and then 1∶200 in the convalescent serum. He also had a specific Western blot profile for B. henselae endocarditis (Fig. 1). He was treated with intravenous ceftriaxone 2 g once a day for 6 weeks and gentamicin 240 mg/d for 2 weeks and was well at one year follow up.
A 69-year old housewife from Xaysettha district, Vientiane Capital, was admitted in 2008 with two months of fever, headache, arthralgia, back pain, myalgia, jaundice, diarrhea, productive cough and dyspnea, with a history of hypertension. She had been treated with ceftriaxone before admission. On examination she was afebrile and normotensive but with a systolic heart murmur. Her admission peripheral blood count showed hematocrit 32%, WBC 4.5×109/L (lymphocytes 62%) and platelets 85×109/L. Trans-thoracic echocardiogram demonstrated significant mitral and aortic valve disease but no vegetation: thickening of mitral valve (D∼5–6 mm) with moderate to severe mitral regurgitation (regurgitant volume = 57 ml/s) and thickening of aortic valve with aortic regurgitation grade 1–2/4, and elevated estimated pulmonary artery systolic pressure of 70 mmHg. Five pairs of blood cultures, incubated for 7 days, showed no growth. The patient had negative IFA results to Bartonella species but her Western blot profile was typical of B. henselae endocarditis. She was treated with ceftriaxone 2 g/day for 14 days with resolution of symptoms but died of gastric perforation in 2011.
In Asia, Bartonella endocarditis has been described from Japan (B. henselae [1]), Korea (B. quintana [2]), Thailand (B. henselae [3]) and India (B. quintana [8]), but we found no reports from Cambodia, Lao PDR, Vietnam, Burma/Myanmar, China, Indonesia, Taiwan or Malaysia (Pubmed using these country names and ‘Bartonella’ ‘endocarditis’). Two additional Bartonella species, resembling those reported as causing endocarditis elsewhere, have been described from Thailand, B. vinsonii from humans and B. elizabethae from rodents [9]. B. henselae occurs in cat fleas, and possibly ticks [10], and this pathogen has been recorded in fleas from Malaysia and Thailand but not elsewhere in mainland Asia [11], [12].
It is likely that Bartonella endocarditis is more widespread than appreciated as few have looked and laboratories in Asia able to make the diagnosis are infrequent. The diagnosis of culture-negative endocarditis is most effectively performed from valve tissue, but there are few cardiac surgeons in rural Asia. Considering the high prevalence of rheumatic heart disease in Asia there is remarkably little evidence on the bacterial etiology of endocarditis. Prospective investigations, especially at centers able to perform valvular surgery, would inform prevention and treatment policies for these very large populations devoid of evidence. Improved methods for the rapid diagnosis of Bartonella endocarditis are needed, as it is likely that proven Bartonella endocarditis can be treated with simpler and less expensive regimens (e.g. gentamicin for 2 weeks plus oral doxycycline for 6 weeks) than ‘conventional’ endocarditis [13], but multicenter trials to optimize treatment are also needed.
|
10.1371/journal.ppat.1005584 | HTLV-1 Tax Functions as a Ubiquitin E3 Ligase for Direct IKK Activation via Synthesis of Mixed-Linkage Polyubiquitin Chains | The HTLV-1 oncoprotein Tax plays a key role in CD4+ T cell transformation by promoting cell proliferation and survival, mainly through permanent activation of the NK-κB pathway and induction of many NF-κB target genes. Elucidating the underlying molecular mechanism is therefore critical in understanding HTLV-1-mediated transformation. Current studies have suggested multiple but controversial mechanisms regarding Tax-induced IKK activation mainly due to blending of primary Tax-induced IKK activation events and secondary IKK activation events induced by cytokines secreted by the primary Tax-induced IKK-NF-κB activation events. We reconstituted Tax-stimulated IKK activation in a cell-free system to dissect the essential cellular components for primary IKK activation by Tax and studied the underlying biochemical mechanism. We found that Tax is a putative E3 ubiquitin ligase, which, together with UbcH2, UhcH5c, or UbcH7, catalyzes the assembly of free mixed-linkage polyubiquitin chains. These free mixed-linkage polyubiquitin chains are then responsible for direct IKK activation by binding to the NEMO subunit of IKK. Our studies revealed the biochemical function of Tax in the process of IKK activation, which utilizes the minimal cellular ubiquitination components for NF-κB activation.
| Human T-cell leukemia virus type 1 (HTLV-1) is the etiologic agent of tropical spastic paraparesis/HTLV-1-associated myelopathy (TSP/HAM), a distinct neurological disorder with inflammatory symptoms and incomplete paralysis of the limbs, and adult T-cell leukemia/lymphoma (ATL), a highly aggressive malignant proliferation of CD4+ T lymphocytes. Both TSP/HAM and ATL are mainly driven by the activation of IκB kinase (IKK)-NF-κB stimulated by HTLV-1 oncoprotein Tax. The molecular mechanism by which Tax activates IKK remains unclear. Here, we found that Tax is an E3 ubiquitin ligase, which, together with its cognate ubiquitin-conjugating enzymes (E2s) UbcH2, UhcH5c, or UbcH7, catalyzes the assembly of unanchored free mixed-linkage polyubiquitin chains. The polyubiquitin chains can activate IKK complex directly by binding to the NEMO subunit. Our studies uncovered the essential cellular factors hijacked by HTLV-1 for infection and pathogenesis, as well as the biochemical function and the underlying mechanism of Tax in the process of IKK activation. Our work might shed light on potential development of therapeutics for TSP/HAM and ATL.
| Human T-cell leukemia virus type 1 (HTLV-1), the first human oncogenic retrovirus that was originally described in 1980 [1], has been received much scientific attention due to its ability to transform primary T-lymphocytes in cell culture and its association with adult T-cell leukemia/lymphoma (ATL), a highly aggressive malignant proliferation of CD4+ T lymphocytes, and with tropical spastic paraparesis/HTLV-1-associated myelopathy (TSP/HAM), a distinct neurological disorder with inflammatory symptoms and incomplete paralysis of the limbs [2, 3].
HTLV-1 possesses an open reading frame (ORF) encoding a transactivator Tax that is critical to the viral life cycle for proviral transcription from the viral long terminal repeat (LTR) promoter [4]. Numerous studies have shown that it is also this Tax protein that is essential for mediating malignant T cell transformation by HTLV-1. Tax can transform rodent fibroblasts and human primary T cells in the presence of IL-2 [5]. In addition, Tax-transformed lymphoid cells and fibroblasts form tumors when inoculated into immunodeficient nude mice [6, 7], and Tax-based transgene expression induces an ATL-like syndrome in the T cell compartment in mice [8] and plasmatocyte proliferation in Drosophila [9]. Moreover, HTLV-1 genome without Tax loses its transformation ability [10]. Tax mutant that is defective in NF-κB activation loses the ability to transform T cells [11] and shows defect in cutaneous disease development in transgenic mice [12]. Additionally, increased studies have shown the minus strand of HTLV-1 encodes a bZIP protein HBZ that is critical for promoting proliferation of ATL cells [13].
Tax exerts a variety of activities in cells and undergoes heavy post-translational modifications such as phosphorylation, ubiquitination, sumoylation, and acetylation to control or modulate its cellular activities [14]. Tax interacts with more than 100 host cell proteins [15] and engages multiple signaling pathways such as activation of cAMP response element-binding protein (CREB), NF-κB, serum response factor (SRF) and inactivation of the tumor suppressor gene p53 [16]. Activation of such cellular proliferation-promoting pathways in turn induces a diverse array of genes encoding proliferative cytokines, cytokine receptors, co-stimulatory molecules as well as survival proteins [17]. Among these pathways, activation of NF-κB is arguably the most critical for Tax-associated cellular transformation and human diseases [18].
The NF-κB transcription factors include five members: RelA/p65, c-Rel, RelB, p105(p50), and p100(p52). These five members can form dimers with one another and bind to target DNA sequences called κB sites to modulate gene expression. In most un-stimulated cells, the NF-κB complexes are retained in the cytoplasm and inactive due to their binding by inhibitory IκB proteins (IκBα, IκBβ, IκBε, etc.) [19]. Upon activation, IκB proteins are phosphorylated, ubiquitinated and then degraded by the proteasome leading to release and translocation of NF-κB into the nucleus. Phosphorylation of IκBs is mediated by the IKK kinase complex, which consists of two active kinase subunits, IKKα and IKKβ, and the regulatory scaffolding subunit IKKγ (also called NEMO) [20]. In the TNFR and IL-1R/TLR activated NF-κB pathways, IKK activation requires an upstream kinase TGF-β-activating kinase 1 (TAK1) and adaptor proteins TRAFs such as TRAF6 [21]. In the case of TRAF6, it functions as an E3 ubiquitin ligase, together with Ubc13/Uev1, to catalyze assembly of K63-linked polyubiquitin (polyUb) chains to mediate TAK1 activation [22].
Activation of NF-κB by Tax also depends on IKK. Tax was shown to interact directly with IKKγ and induce its oligomerization [23–26]. Overexpression of Tax fusion protein to either IKKα or IKKβ was shown to be sufficient for IKK activation [27]. Tax was also found to localize to the lipid rafts, to where it recruited IKK for its persistent activation, which mechanism was further strengthened by cell adhesion molecule 1 (CADM1) [28, 29]. These studies suggest Tax and IKKγ interaction is important for Tax-mediated IKK activation [17]. Tax was also shown to interact with TAK1, the upstream kinase of IKK in the IL-1R/TLR signaling pathways, and this interaction was shown to mediate TAK1 interaction with IKK facilitating its activation [30]. In addition, MEKK1, NIK and Tpl2 were reported to be the putative IKK kinases for Tax-induced IKK activation [31–34]. However, by using various knockout MEF and siRNA-based knockdown cells, other studies illustrated those kinases were not required for Tax-induced IKK activation [35, 36]. Nonetheless, Tax alone was not sufficient for direct IKK activation and required other cellular factors [37].
It is generally recognized that Tax activation of IKK-NF-κB requires ubiquitination events, although there are debates as to what the ubiquitin-conjugating enzyme (E2s) and ubiquitin ligase (E3) are, and what the ubiquitination targets and the underlying mechanisms are. It was initially demonstrated that the E2 enzyme Ubc13, together with TRAF2, 5, or 6, promotes lysine-63 (K63) polyubiquitination of Tax, which targets IKK to centrosome thus promoting IKK activation [38, 39]. Tax also induces K63 polyubiquitination of NEMO for IKK activation [40]. A more recent study suggested that Tax stimulates E3 ligase RNF8, together with Ubc13/Uev1 and 2, to assemble K63 polyUb chains leading to activation of TAK1-IKK-NF-κB cascade [34]. In further support of this, the deubiquitinating enzyme (DUB) USP20 deubiquitinates Tax to negatively regulate activation of IKK [41]. But studies using RNAi-based knockdown and knockout MEF cells illustrated Ubc13 and TRAF6 are not required for Tax-induced IKK activation [35, 37]. Although Tax induced K63-linked polyubiquitination of NEMO, this event was not required for IKK activation [37]. In agreement with this, the K63-specific DUB CYLD effectively removed polyUb chains from NEMO; but this didn’t affect Tax-induced IKK activation [35]. In addition to ubiquitination, sumoylation of Tax was also reported to be important for IKK-NF-κB activation [42, 43]. However, Tax mutant defective in sumoylation, was still able to activate IKK without difference from Tax wild-type (WT) [44].
Therefore, a general but convincing mechanism about Tax-induced IKK activation awaits further studies. Like other IKK-NF-κB activation events, Tax-dependent IKK-NF-κB activation induces a variety of cytokines such as TNFα and IL-1β [45]. These cytokines in turn activate IKK-NF-κB through an autocrine and paracrine fashion. So the primary IKK-NF-κB activation events triggered by Tax and the secondary IKK-NF-κB signaling events triggered by secreted cytokines are blended together. However, all the cell- and animal-based studies cannot dissect the primary events from the secondary events. Therefore, these complications make it difficult to clearly define essential cellular factors and events for Tax-induced IKK activation.
To circumvent the complications brought about by the secondary IKK-NF-κB activation events, we developed a cell-free system to reconstitute IKK activation using recombinant Tax. This system allowed us to define critical cellular factors that are involved in primary events of Tax-induced IKK activation and to explore the underlying biochemical mechanism. We found that Tax is a putative E3 ubiquitin ligase, which, together with UbcH2, UhcH5c, or UbcH7, catalyzes the assembly of free unanchored mixed-linkage polyUb chains. These free mixed-linkage polyUb chains are then responsible for direct IKK activation by binding to the NEMO subunit of IKK. Our studies revealed the biochemical function and the underlying mechanism of Tax in the process of IKK activation, which utilizes the minimal cellular ubiquitination components for IKK-NF-κB activation.
To identify essential cellular components required for the primary IKK activation by Tax, we tried to reconstitute the activation process in a cell-free system. We generated recombinant Tax proteins by using baculoviral recombinant protein expression system (S1 Fig, left panel) and tested them in cellular cytosolic extracts S100 generated from Jurkat T cells. Incubation of Tax with S100 in the presence of ATP resulted in efficient IKK activation, as evidenced by the phosphorylation of IKKα, IKKβ and its physiological substrate IκBα (Fig 1A, lane 2). The mutant form of Tax, M22 (a double-site mutant originally reported to be defective in NF-κB activation) [46] [47], didn’t show detectable activity under the same conditions. The activation was dependent on the presence of NEMO, since there was no detectable IKK activity if we used S100 from NEMO-deficient Jurkat T cells (Fig 1A, lanes 4–6). But the activation was restored by adding recombinant NEMO back into the system (Fig 1A, lanes 7–9). Consistent with the IKK activation assay results in vitro, Tax WT but not the M22 stimulated the 3xκB-Luciferase (3xκB-Luc) reporter activity in a dose-dependent manner in 293T cells (Fig 1B). This set of experiments demonstrates that we have established a cell-free system to recapitulate Tax-dependent IKK activation in intact cells.
We next fractionated the S100 into three fractions by using HiTRAP Q-sepharose column (GE Healthcare) through step elution with increasing concentration of NaCl (Fig 2A) and tested which fractions were required to restore IKK activation by Tax. Combination of Q/I and Q/III were sufficient for IKK activation by Tax (Fig 2A, lane 4; S2A Fig, lane 10). Q/I was also required for the activity, without which there was no IKK activation by Tax (Fig 2A, lane 6; S2A Fig, lane 12). These results suggest there is factor(s) in Q/I that supports Tax-dependent IKK activation.
We then focused on further fractionation for purification of the factor in Q/I. After five steps of conventional chromatograph (Fig 2B and S2B Fig), we achieved purification of the factor responsible for supporting Tax-dependent IKK activation. Fractions from the last Superdex-75 step were used for silver staining and IKK stimulatory activity testing (Fig 2B). By correlating the band intensities (Fig 2B, silver staining panel) and the IKK-stimulatory activity (Fig 2B, activity assay panel), the band present in fractions 2, 3 and 4 on the silver staining gel with an apparent molecular size of about 15 kDa is the most likely candidate. This band was excised and subjected to mass spectrometric analysis. The result suggested it is human UbcH7 (UBE2L3), a ub-conjugating enzyme E2 (Fig 2C).
To verify that UbcH7 is indeed the factor in Q/I that supports IKK activation by Tax, we expressed it in E.coli as a Hexahistidine (His6)-tagged recombinant protein and purified it to apparent homogeneity. We also expressed and purified its corresponding enzymatic activity dead mutant Cysteine-86 to Alanine (C86A) (S2C and S2D Fig). Similar to Q/I, recombinant UbcH7 activated IKK in the presence of Tax (Fig 2D). The C86A mutant was not able to stimulate Tax-dependent IKK activation, suggesting the enzymatic E2 activity is required to support Tax-dependent IKK activation (Fig 2D).
Together these data demonstrate that Tax-induced IKK activation is mediated by UbcH7 and depends on ubiquitination.
Considering that many ubiquitination systems use several cognate E2s, we tested a panel of other E2s (S3A and S3B Fig) to check whether any of them could also support Tax-dependent IKK activation. As shown in Fig 3A and S3B Fig, in addition to UbcH7, two other E2s, UbcH2 (UBE2H) and UbcH5c (UBE2D3) were also capable of stimulating IKK activation by Tax. It is worth to note that the heterodimeric E2 complex, Ubc13/Uev2, specializing in the synthesis of K63 polyUb chains, was not able to activate IKK together with Tax (Fig 3A and S3B Fig). It has been well established that Ubc13/Uev1 (or Ubc13/Uev2) is a key E2 for IKK activation in the IL-1R and TLR pathways. There are also reports suggesting Ubc13 is involved in Tax-dependent IKK-NF-κB activation, although this requirement is currently under debate [35, 38, 39].
To determine whether the above-identified E2s are also required for IKK-NF-κB activation in living cells, we tested the effect on Tax-dependent NF-κB activation in the 3xκB-Luc reporter assay by transfecting their respective active site mutants into 293T cells. Transfection of Tax activated the 3xκB-Luc reporter. Remarkably, the activation was dramatically reduced by overexpression of these E2 Cysteine-to-Alanine (C-to-A) mutants, which presumably function as dominant-negative mutants (Fig 3B). In contrast, the three C-to-A mutants didn’t show any detectable effect on the reporter activity stimulated by TRAF6 overexpression (Fig 3B), which has been known to function together with Ubc13/Uev1 for IKK-NF-κB activation [48].
To further exclude the possibility that Ubc13 is able to support Tax-dependent IKK activation, we treated S100 with OspI or its active site mutant C62A and then tested it in the in vitro IKK activation assay. OspI has recently been shown to deamidate Gln100 of Ubc13 into Glu thus inactivating its E2 activity [49, 50]. As shown in S3C Fig, although both Tax and TRAF6 activated IKK as determined by phosphorylation of IKK and IκBα in the S100, only Tax but not TRAF6 still activated IKK in the S100 pre-treated with OspI. As a control, C62A mutant didn’t destroy the activity. Notably, only TRAF6 but not Tax treatment led to detectable phosphorylation and thus activation of TAK1 (S3C and S3E Fig). In agreement with the in vitro assay, in the 3xκB-Luc reporter assay in 293T cells, OspI only inhibited TRAF6 but not Tax-dependent reporter activity (S3D Fig).
Based on our previous experiences, Q/III fraction contains TAK1, TRAF6, IKK and its substrate IκBα. It has been reported that both TRAF6 and TAK1 are involved in Tax-dependent IKK activation [30]. To test this, we generated TAK1 and TRAF6 knockout (KO) cells by using CRISPR/Cas9-based genome editing technology in 293T cells and used them in the 3xκB-Luc reporter assays. KO of either TAK1 or TRAF6 didn’t have any effect on reporter activity by Tax overexpression, which displayed comparable activity to that in parental 293T cells (S3F Fig). As a positive control, KO of TAK1 abolished TRAF6-stimulated reporter activity. We also tested this directly by preparing S100 from 293T and TAK1 KO cells. Although S100 from 293T cells supported IKK activation by both Tax and TRAF6, S100 from TAK1 KO cells only supported IKK activation by Tax but not TRAF6 (S3E Fig). These results strongly demonstrate that activation of IKK by Tax doesn’t require TAK1 and TRAF6.
After exclusion of TAK1 and TRAF6 as the putative factors for Tax-dependent IKK activation in Q/III, we then tested if IKK complex itself is the only factor in Q/III for its activation by Tax. For this purpose, we generated a stable cell line expressing NEMO-FLAG by using CRISPR/Cas9-based knock-in in 293T cells, which was then used for IKK complex purification. Silver staining shows that the highly purified IKK complex contains IKKα, IKKβ and NEMO-FLAG (S4A Fig). We then tested if the IKK complex could be directly activated by Tax and UbcH7. As shown in Fig 4A, incubation of IKK, UbcH7 and Tax, together with E1, ubiquitin (Ub) and ATP, resulted in activation of IKK as determined by phosphorylation of both IKKα and IKKβ (Lane 4). Lack of either UbcH7 or Tax didn’t lead to significant IKK activation (Lanes 1–3). The activation depends on the E2 activity of UbcH7, since UbcH7 C86A didn’t support the activation. Similarly, replacement of UbcH7 by UbcH5c also activated IKK (Fig 4B). But the Ubc13/Uev2 was not able to facilitate IKK activation by Tax (S4B Fig). Interestingly, immunoblotting of the reaction products revealed presence of polyUb chains, which correlated very well with IKK activation in the case of UbcH7 and UbcH5c (Fig 4A and 4B, top panels). Combination of Tax and Ubc13/Uev2 also generated significant amount of polyUb chains, although there was no IKK activation (S4B Fig). Together, these data suggest that IKK itself is the only minimal factor in Q/III, which can be activated by Tax directly in a polyubiquitination-dependent manner.
The demonstration that highly purified IKK can be activated by Tax directly in a polyubiquitination-dependent manner prompted us to consider the possibility that Tax itself could be an E3 ubiquitin ligase. We tested this directly in two ways. Firstly, we carried out a typical ubiquitination assay in vitro. Incubation of E1, UbcH7, Ub and ATP didn’t lead to detectable polyUb signals. However, inclusion of Tax in the reaction led to significant amount of polyUb signals in a dose-dependent manner. Similar results were obtained when UbcH7 was replaced by UbcH5c, UbcH2 or Ubc13/Uev2. (Fig 5A, upper panels). It has been known that a few E2s such as UbcH5c can ubiquitinate a protein without the presence of any E3s. Immunoblotting of the reaction products using anti-Tax antibody didn’t show detectable ubiquitination of Tax (Fig 5A, bottom panels), demonstrating the polyUb signals were not due to polyUb conjugation of Tax protein itself and Tax didn’t simply function as a substrate but as an E3 ubiquitin ligase. These data also imply the polyUb chains are probably free unanchored polyUb chains. In further support of synthesis of free polyUb chains by Tax, we did affinity-depletion using Nickel beads against the ubiquitination reaction products to remove all the His-tagged proteins including E1, E2 and Tax that we added into the ubiquitination reaction. Immunoblotting of the supernatant after the depletion demonstrates that the His-tagged ubiquitination reaction components, especially Tax, have been removed to under detection level (Fig 5B, bottom panel). Under this condition, immunoblotting by using anti-Ub antibody showed comparable polyUb signals before and after the depletion, in agreement with our assumption that most, if not all, of the polyUb chains are free chains that are not conjugated to any proteins (Fig 5B, upper panel). Secondly, one feature of a typical E3 is its ability to promote turnover (discharge) of Ub from its cognate E2s when these E2s are loaded with Ub. So, we tested if Tax could do the same and found that Tax indeed promoted discharge of Ub from Ub~UbcH7, Ub~UbcH5c, and Ub~Ubc13 after incubation together (Fig 5C).
These two sets of experiments strongly suggest that Tax is a bona-fide E3 ubiquitin ligase, and some of its cognate E2s are UbcH7, UbcH5c, UbcH2 and Ubc13/Uev2.
We next tried to characterize what kinds of Ub linkages are formed by Tax. We carried out the ubiquitination reactions by using UbcH7 and Tax or UbcH5c and Tax. The Ub linkages were determined by using mass spectrometric (MS) analysis. As shown in Fig 6A and S5A Fig, all the possible Ub linkages, K6, K11, K27, K29, K33, K48 and K63, were detected, suggesting the polyUb chains are mixed linkage chains. Notably, there was no detectable M1 linkage. To further verify the linkage specificity, we tested a panel of single lysine (K)-only Ub mutants for their ability in polyUb chain synthesis (Fig 6B and S5B–S5D Fig). A single K-only Ub mutant contains only one K with the rest six K residues mutated to arginine (R). Although Ub WT supported the synthesis of polyUb chains, all the other 7 single K-only mutants didn’t show detectable polyUb signals. On the other hand, mutants R48 and R63 (S5B and S5C Fig), each of which has 6 intact K residues, were able to support Tax-dependent synthesis of polyUb chains (Fig 6C). These polyUb chains could be partially digested by CYLD WT but not its enzymatic-dead mutant C601A, suggesting the presence of K63 and/or M1 linkage as well as other linkages in the polyUb chains, which was digested by vOTU WT but not its mutant C40A (Fig 6D). CYLD is a K63- and M1-specific DUB [51]; and vOTU is an ovarian tumor (OTU) domain-containing DUB from Crimean-Congo hemorrhagic fever virus (CCHFV) large (L) protein [52] that can cleave all kinds of Ub linkages except M1 linkage [53]. Since the M1-specific DUB OTULIN didn’t show detectable cleavage on the polyUb chains, which is consistent with the absence of M1 linkage in the polyUb chains analyzed by mass spectrometry (Fig 6A), the linkage cleavage by CYLD was therefore K63 linked. These results further suggest mixed-linkage polyUb chains are synthesized and any single K residue is not sufficient for polyUb chain synthesis by Tax and its cognate E2s.
Since Tax catalyzes the synthesis of free mixed-linkage polyUb chains, we wondered if the formation of this kind of mixed-linkage polyUb chains is required for IKK activation by Tax. Again we examined the panel of single K-only Ub mutants for their ability to interfere IKK activation in our S100-based assay system. Addition of Ub WT had no effect on IKK activation by Tax, similar to the one without exogenous Ub. In contrast, inclusion of all the single K-only mutants and the lysine-null (KO) mutant in the reaction didn’t show IKK activation, presumably due to their dominant-negative effect on mixed-linkage polyUb chain assembly and thus inability to support IKK activation by Tax (Fig 7A). Furthermore, addition of viral OTU (vOTU) WT but not the mutant C40A, or CYLD WT but not the mutant C601A, also blocked IKK activation by Tax (Fig 7B and 7C). Taken together, these data further demonstrate that Tax-mediated IKK activation in vitro involves the assembly of mixed-linkage polyUb chains.
The NEMO subunit of IKK complex is important in mediating its activation by polyubiquitination events [54]. We therefore tested a panel of NEMO mutants, which have been shown to be differentially defective in binding to various kinds of polyUb chains [55], to examine if they could be defective in IKK activation by Tax. We generated S100 from NEMO-deficient Jurkat T cells and used them for in vitro IKK activation assay. As shown in Fig 7D, although NEMO WT restored IKK activation by Tax, the three mutants were not able to achieve it, implying polyUb binding of NEMO is important for IKK activation.
Finally, we tested if the mixed-linkage polyUb chains can activate IKK directly. So we synthesized mixed-linkage polyUb chains first, treated the products with NEM to inactivate all the enzymes (E1, E2s such as UbcH7 or UhcH5c) [53]. Nickel beads were then used to deplete all the ubiquitination components to undetectable level. The remaining free polyUb chains were then concentrated and mixed with purified IKK complex for activation assay. Following the same protocols, we also produced K63, K48 and M1 polyUb chains. As displayed in Fig 7E, the polyUb chains synthesized by Tax-UbcH7 or Tax-UbcH5c activated IKK in a dose-dependent manner. But the K63, K48 and M1 polyUb chains didn’t show any detectable activity on IKK activation. Together, the purified system demonstrates that the mixed-linkage polyUb chains are required and sufficient for direct IKK activation.
We noted that Tax catalyzed assembly of mixed linkage polyUb chains using Ub mutants R48 and R63 (Fig 6C). So we tested them in the S100-based IKK activation assay. As shown in Fig 8A, both Ub WT and R48 didn’t interfere with Tax-induced IKK activation, but the R63 displayed inhibitory activity, implying the K63 linkage might be a requirement in the mixed linkage polyUb chains for IKK activation. Similar observation was also reported by Shibata et al. [37]. However, polyUb chains assembled from R63 by Tax was able to activate highly purified IKK directly without detectable differences from polyUb chains by either Ub WT or R48 (Fig 8B). Detailed time-course analysis of polyUb chain assembly by Tax revealed the difference of efficiency of polyUb chain assembly among Ub WT, R48 and R63 and could explain the seemingly contradictory results shown in Fig 8A and 8B. Although Ub WT, R48 and R63 all supported polyUb chain assembly, with R63 it displayed slower kinetics and accumulated much less amount of polyUb chains, especially those high molecular size polyUb species. Mixing of Ub WT and R63 together didn’t improve the polyUb chain assembly efficiency (Fig 8C). Together, these data illustrate that K63 linkage in the polyUb chains is not required for IKK activation.
HTLV-1 Tax is the first pathogenic agent that has been shown to be able to activate NF-κB and studies on it have advanced greatly our current understanding of pathophysiological activation of NF-κB. However, despite extensive studies spanning almost 30 years, our understanding of the molecular mechanism by which Tax activates IKK-NF-κB is still far from clear. Like other IKK-NF-κB activation events, Tax-dependent IKK-NF-κB activation induces a variety of cytokines such as TNFα and IL-1β [45], which in turn activate IKK-NF-κB through an autocrine and paracrine fashion, complicating the dissection of signaling events of Tax-dependent IKK activation. We reconstituted the IKK activation events by Tax in an in vitro cell-free system. This system captured the primary events of IKK activation induced by Tax. With this system, we identified the essential cellular components required by Tax for IKK activation and revealed the biochemical function of Tax in this process. We found Tax itself is an E3 ubiquitin ligase, which utilizes UbcH7, UbcH2 or UbcH5 as the cognate E2 enzymes to catalyze the assembly of mixed-linkage polyUb chains. This kind of polyUb chains then stimulates activation of IKK directly. This kind of Tax-induced IKK activation is probably the simplest system, which is to a great extent of advantage to HTLV-1 infection by minimizing the events required for IKK-NF-κB activation, an essential event for successful infection of HTLV-1. In other words, HTLV-1 hijacks the minimal host cell components for its own use, minimizing the possible events cells can use for negative regulation of IKK-NF-κB activation and maximizing success of HTLV-1 infection. In the IL-1R/TLR and TNFR signaling pathways, there are multiple steps and key components leading to IKK activation are subject to negative regulation, mainly by DUBs such as A20 [56] and CYLD [57, 58] and protein phosphatases such as PP1 [59] and DUSP14 [60], to subside NF-κB activation. Since Tax-dependent IKK activation is almost at the IKK level and so those negative regulation mechanisms are not effective on Tax-induced IKK activation. This might contribute to Tax-dependent persistent activation of IKK and is beneficial to HTLV-1 infection. Exactly how much this kind of IKK activation contributes to constitutive IKK activation by Tax needs further investigation. Shibata et al. has reported a similar in vitro cell-free system to study Tax-dependent IKK activation, in which it reported K63 polyubiquitination is involved in this process. But they didn’t go further to identify essential factors for IKK activation by Tax [37].
There have been reports suggesting both TAK1 and TRAF6 are important for Tax-induced IKK activation. Our in vitro fractionation studies unequivocally show that neither TAK1 nor TRAF6 are required for Tax-induced IKK activation. In further support of TAK1- and TRAF6-independent activation of IKK by Tax, we generated TAK1 and TRAF6 KO cells using CRISPR/Cas9 technology. In vitro assay with S100 from these KO cells and luciferase reporter assays also demonstrate TAK1 and TRAF6 are not involved in Tax-dependent IKK activation. By using S100 generated from TRAF6 KO MEF cells, Shibata et al. also show TRAF6 is not required for IKK activation by Tax, consistent with our findings [37]. Having said that, however, in cell-based studies, it is highly possible to conclude that TAK1 and TRAF6 are involved in Tax-dependent IKK activation. Cytokines such as IL-1β induced by primary Tax-stimulated IKK-NF-κB activation depend on TAK1 and TRAF6 for IKK activation and there are difficulties to distinguish signaling events by Tax from those autocrine or paracrine signaling events by cytokines.
It is generally accepted that ubiquitination is involved in Tax-induced IKK activation. However, detailed biochemical mechanisms regarding its involvement are not clear, especially about K63 polyubiquitination events. Our studies clarified the confusion and demonstrated that K63 linkage is not a requirement for IKK activation by Tax. We provided several lines of evidence to support this conclusion. In our fractionation experiments, which are an unbiased means for factor identification, Ubc13 was not co-purified with IKK activation activity. Also, inactivation of Ubc13 by OspI in cells and in our S100-based cell-free system didn’t prevent IKK activation from Tax, making it unlikely that K63 polyubiquitination is required for Tax-dependent IKK activation. Together with UbcH7, UbcH2 or UbcH5c, Tax catalyzes assembly of mixed-linkage but not K63 polyUb chain. Although Tax can assemble K63 polyUb chains when the E2 is Ubc13/Uev2, in which the linkage specificity is determined by Ubc13 and Uev2 [61, 62], the K63 polyUb chains are not able to activate IKK. What the function is for K63 polyUb chains catalyzed by Tax is not known yet; it probably functions in DNA damage repair pathway, a reflection of nuclear Tax function [43].
Ub mutant R63 in the S100 inhibited IKK activation by Tax, which would argue for the requirement of K63 linkage for Tax-regulated IKK activation. However, polyUb chains synthesized by Tax using R63 displayed activity on IKK activation, clearly demonstrating that K63 linkage is not a requirement. We were puzzled initially by those seemingly conflicting results. However, our time-course analysis comparing the efficiency of polyUb chain synthesis between Ub WT and R63 provided the answer to the puzzle. Although R63 can support polyUb chain synthesis, its efficiency is low when compared to Ub WT. Mixing of Ub WT and R63 didn’t improve the efficiency of polyUb chain assembly, implying inclusion of R63 in S100 would interfere with polyUb chain assembly leading to insufficiency of polyUb accumulation for IKK activation.
Tax interacts with IKK through the NEMO subunit, which has been suggested to be required for IKK activation by Tax. Tax also undergoes K63 polyubiquitination and induces K63 polyubiquitination of NEMO and these events are also important for IKK activation by Tax [16]. Here, using purified IKK complex and purified polyUb chains synthesized by Tax, we show Tax as well as its ubiquitination is not required in the IKK activation process. Since there are no ubiquitination components in our purified polyUb chains, it is unlikely that NEMO ubiquitination happens during the activation assay in our system. In agreement with our results, Shibata et al. detected Tax-induced NEMO ubiquitination in their system but this event was not required for IKK activation [37]. We conclude Tax interaction with NEMO and NEMO polyubiquitination, if there is any, are not required or at least not critical for Tax-induced IKK activation, clarifying the confusing reports in the literature regarding the molecular mechanism of Tax-dependent IKK activation. However, we cannot exclude that Tax interaction with NEMO and even Tax-induced NEMO polyubiquitination would contribute to IKK activation in cells. We would also stress in cells Tax ubiquitination is still possible which, together with its other PTM modifications such as sumoylation, phosphorylation and acetylation, could regulate some of its other functions such as balancing its cytosol-nucleus translocation [63].
Our studies assigned a novel biochemical function to Tax, that is, as an E3 ubiquitin ligase. It catalyzes assembly of mixed linkage polyUb chains with some of its cognate E2s. Our previous studies have demonstrated that mixed linkage polyUb chains are not able to activate TAK1 complex, which can only be effectively activated by K63 polyUb chains [53]. In the same study we have also shown that mixed-linkage polyUb chains assembled by TRAF6 and UbcH5c were able to activate IKK [53]. In this study, we showed purified free mixed linkage polyUb chains are potent activators towards IKK complex, but other kinds of polyUb chains such as K48, K63 and M1 are poor activators for direct IKK activation, providing another example of direct IKK activation by free mixed polyUb chains. Therefore, we would propose the following model by which Tax activates IKK (Fig 8D): Tax, as a novel E3 ubiquitin ligase, together with its cognate E2s such UbcH7, assembles unanchored free polyUb chains with various linkages, the latter in turn function as potent activators for direct IKK activation. Here mixed linkage is a requirement for efficient IKK activation.
In conclusion, by taking advantage of the in vitro cell-free system, we investigated the biochemical mechanism of IKK activation by HTLV-1 encoded Tax protein and the biochemical function of Tax in this process. Tax itself is an E3 ubiquitin ligase, which works together with its cognate E2s, catalyzes free polyUb chains assembly for direct IKK activation. Understanding the biochemical function of Tax for IKK activation and the underlying molecular mechanism clarifies the confusion on HTLV-1-induced IKK-NF-κB activation, reveals the essential cellular factors hijacked by HTLV-1, and might shed light on potential development of therapeutics for ATL and TSP/HAM.
Rabbit antibodies against IKKα (sc-7218, dilution 1:1,000), IKKβ (sc-7607, dilution 1:1,000), NEMO (sc-8330, dilution 1:1,000) and mouse antibody against ubiquitin (sc-8017, dilution 1:1,000) were obtained from Santa Cruz Biotechnology; antibodies against TAK1 (ab109526, dilution 1:1,000), UbcH7 (ab108936, dilution 1:2,000), Ubc13 (ab109286, dilution 1:5,000), RNF8 (ab131221, dilution 1:1,000), p-IKKα S176 (ab138426, dilution 1:500) were from Abcam; antibodies against CYLD (#8462, dilution 1:1,000), p-TAK1 Thr187 (#4536, dilution 1:1,000), p-IKKα/β Ser176/180 (#2697, dilution 1:1,000) and p-IκBα Ser32/36 (#9246, dilution 1:1,000) were from Cell Signaling; antibodies against FLAG-tag (M20008, dilution 1:5,000) and His-tag (M20001, dilution 1:1,000) were from Abmart; antibodies against IκBα and UbcH5c were homemade; and antibody against Tax (mAB, clone Lt-4) was described in Lee et al. [64].
HEK293T and Jurkat T cells were originally from the American Type Culture Collection (ATCC, Manassas, VA). The 293F cells were originally from Thermo Fisher Scientific (New York, NY). The NEMO-deficient Jurkat T cells were kindly provided by Shao-Cong Sun (M. D. Anderson Cancer Center, Houston, TX).
HEK293T and 293F cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% (v/v) fetal bovine serum (Gibco), penicillin (100 U/ml), and streptomycin (100 mg/ml). 293F cells were suspended in SMM 293-T1 medium (Sino Biological Inc.) with antibiotics. Jurkat T and NEMO-deficient Jurkat T cells were cultured in RPMI 1640 supplemented with 10% fetal bovine serum, 2 mM β-mercaptoethanol, penicillin (100 U/ml), and streptomycin (100 mg/ml). Plasmid DNA transfection of 293T and 293F cells was performed using Polyethylenimine (#24765, Polysciences Inc.).
Guide RNA sequences (gRNAs) for each gene (TRAF6: 5’-GTCTCCACCCGCTTTGACAT-3’; TAK1: 5’-GCAATGCAAAAAACAACTAG-3’) were cloned into a CRISPR/Cas9-based vector modified from pX330 [65] with a puromycin resistance selection marker. This vector was transfected into 293T cells. After selection by puromycin (0.5 μg/ml), single colonies were picked and verified by western blotting.
The targeting vector for knock-in contains 1Kb homology arms on each side around the stop codon of human NEMO gene, and a FLAG-IRES-Puromycin segment just before the stop codon. This vector was co-transfected with the CRISPR/Cas9-based vector pX335 [65] (gRNA sequence: 5’-GTCATGGAGTGCATTGAGTA-3’) into 293F cells. After selection by puromycin (0.5 μg/mL), single clones were picked and verified by western blotting.
cDNAs encoding human UbcH5c (UBE2D3) wild-type (WT) and dominant-negative (DN) mutant (C85A), UbcH7 (UBE2L3) WT and DN mutant (C86A) and Ubc13 (UBE2N) were inserted into plasmids pET15b, pET16b and pET14b, respectively. cDNAs encoding human UEV2 (UBE2V2), UBE2G2, Cdc34 (UBE2R1), UBE2T and CCHFV OTU (1–169) were inserted into plasmid pGEX-4T-1. cDNAs encoding human UBE2A, UBE2B, UbcH10 (UBE2C), UbcH5a (UBE2D1), UbcH5b (UBE2D2), UbcH5d (UBE2D4), UbcH6 (UBE2E1), UBE2G1, UbcH2 (UBE2H), UBE2R2 and human HOIP697-1072 (HOIP-RBRC) were inserted into a modified pGEX-4T-1 vector in which the GST tag was replaced by a MBP-His10 tag. cDNAs encoding Shigella OspI and IpaH265-568 were inserted into plasmid pGEX-6P-2. cDNA encoding ubiquitin was inserted into a modified pET-14b vector in which no tag was expressed. cDNAs encoding human E1 and mouse TRAF6 were inserted into pFastBac-HTB vector, cDNAs encoding HTLV-1 Tax WT and M22 (T130AL131S) mutant were inserted into a modified pFastBac-HTB vector in which a MBP tag was inserted into the N-terminus, and a FLAG-His10 (10xHis) tag the C-terminus. His6-tagged E1, TRAF6 and MBP-Tax-FLAG-His10 were expressed in Sf9 cells. cDNAs encoding HTLV-1 Tax WT and M22 mutant were also inserted into a modified pcDNA3.1 vector for mammalian cell expression in which a MBP tag was inserted into the N-terminus, and a FLAG-His10 (10xHis) tag the C-terminus. cDNAs encoding human IκBα, NEMO, CYLD, OTULIN, UbcH5c DN, UbcH7 DN and UbcH2 DN (C87A) were inserted into pcDNA3.1-FLAG-His10 vector. cDNA for mouse TRAF6 was cloned into pcDNA3.1 with an N-terminal FLAG tag.
The E.coli strain BL21(DE3/pLys) harboring plasmids of E2s and ubiquitin were induced with 0.5 mM IPTG at 37°C for 4 h, and those harboring the other plasmids were induced with 0.1 mM IPTG at 16°C overnight. His-tagged proteins were purified using nickel agarose beads according to the manufacture’s protocol (Thermo, #88223). FLAG-tagged proteins were transiently expressed in 293T cells and purified using anti-FLAG M2 magnetic beads according to the manufacturer's instructions (Sigma, M8823).
Protein purity was shown in S1, S3A, S3C, S4A and S5C Figs.
Cell pellet was collected and resuspended in equal volume of hypotonic buffer [20 mM HEPES-KOH, pH 7.4, 10 mM KCl, 1.5 mM MgCl2, 1 mM EDTA, 1 mM EGTA, 1 mM DTT, 1 mM PMSF], then homogenized using a Dounce homogenizer. After the cell debris was removed by centrifugation at 20,000×g for 30 min, the supernatant (S20) was collected and stored at -80°C, or further undergone ultracentrifugation at 100,000×g for 1 h and the cleared supernatant (S100) was collected and stored at -80°C.
Plasmids encoding indicated inhibitory proteins (OspI or dominant-negative mutants of E2s) were transfected into 293T cells first. After 6 hours, cells were transfected again with plasmids encoding Tax or TRAF6, together with a NF-κB firefly-luciferase reporter (3xκB-Luc) and a renilla-luciferase as the internal reference. A pcDNA3.1 empty vector was used to bring total DNA in each transfection group equal. After 24 hours, cells were harvested and luciferase activities were measured.
To measure the activation of IKK by Tax in vitro, cell extracts (S100) of Jurkat T or NEMO-deficient Jurkat T cells were incubated with recombinant Tax, M22 or TRAF6 (100 nM each), and NEMO (100 nM), DUBs or OspI (100 nM) (as shown in figures) in ATP buffer [50 mM Tris-Cl (pH 7.5), 5 mM MgCl2, 2 mM ATP, 0.1 μM okadaic acid, 0.5 mM DTT]. After incubation at 30°C for 1 h, the reaction products were immunoblotted using antibodies as specified in each figure. To follow the activity of IKK activation during fractionation, the same IKK assay was performed except that S100 was replaced by Q/III, E1 (20 nM), Ub (10 μM), and aliquots from column fractions. For E2 screening, the same assay was used except that column fractions were replaced by recombinant E2s (0.5 μM).
All procedures were carried out at 4°C. Jurkat T S100 from 20 L of suspension culture were prepared and applied to HiTrap Q column with buffer Q/A [20 mM HEPES-KOH, pH 7.4, 10% Glycerol, 1 mM EDTA, 1 mM EGTA, 1 mM DTT, 1 mM PMSF], which was eluted with Q/B [Q/A with 1 M NaCl] to generate fractions Q/I, Q/II and Q/III. Q/I was subjected to ammonium sulfate precipitation (40%–80%), followed by dialysis against Buffer SP/A [20 mM HEPES-KOH, pH 6.5, 10% Glycerol, 1 mM EDTA, 1 mM EGTA, 1 mM DTT, 1 mM PMSF]. The dialyzed proteins were applied to HiTrap SP column with buffer SP/A and SP/B [SP/Q with 1 M NaCl]. Fractions containing the activity were pooled and buffer exchanged by ultrafiltration into buffer Q/A, and applied to HiTrap Heparin column with buffer Q/A and Q/B. Fractions that contained the activity were pooled and concentrated before loading onto a Superdex75 column, which was pre-equilibrated with buffer Q/C [Q/A with 100 mM NaCl].
To load ubiquitin to recombinant E2s, each E2 (5 μM) was incubated with E1 (100 nM) and Ub (50 μM) in ATP buffer at 30°C for 10 min. The reaction products were resolved by non-reducing SDS-PAGE and E2~Ub intermediate was detected by Coomassie brilliant blue staining. To discharge Ub from E2s by Tax, the same Ub loading assay was used to prepare E2~Ub intermediate. After desalting by G-25 (GE Healthcare) to remove ATP, the intermediates were incubated with recombinant Tax (1 μM) at 30°C for 15 min before immunoblotting.
To synthesize polyUb chains by Tax, recombinant E1 (20 nM), Ub (50 μM), E2 (500 nM) and Tax (1 μM) were incubated in ATP buffer in a final volume of 10 μl at 37°C for 2 h, and the reaction products were used for immunoblotting.
For further purification of polyUb chains by Tax used in Figs 7E and 8B, the reaction volume was scaled up and the final products were mixed with equal volume of Ni-NTA beads for 2 hours at 4°C to deplete His-tagged E1, E2 and Tax. The supernatant was then treated with 10 mM N-ethylmaleimide (NEM) to inactivate any possible remaining E1 and E2, and then NEM was quenched by DTT and removed by ultrafiltration. PolyUb chains in the supernatant were concentrated by ultrafiltration (Millipore, UFC500396).
To purify K48-linkage polyUb chains, similar procedure was carried out except that E3 was replaced with IpaH (10 nM). For K63-linkage polyUb chains, similar procedure was carried out except that E2 and E3 were replaced with Ubc13/Uev2 (1 μM) and TRAF6 (100 nM), respectively. For Linear-linkage (M1) polyUb chains, similar procedure was carried out except that E2 and E3 were replaced with UbcH7 (1 μM) and HOIP-RBRC (250 nM), respectively.
To reconstitute IKK activation by E2s and Tax, 0.5 ng/μL IKK complex purified from 293F/NEMO-FLAG knock-in cells was incubated with recombinant E1 (20 nM), Ub (10 μM), Tax (300 nM) and the indicated E2 (50 nM) in ATP buffer in a final volume of 10 μl at 37°C for 2 h. To reconstitute IKK activation by polyUb chains, IKK complex (0.5 ng/μL, final volume 10 μl) was incubated with purified polyUb chains in ATP buffer at 37°C for 2 h. Activation of IKK was determined by immunoblotting using phospho-IKKα/β antibody.
The proteins were precipitated by TCA and then tryptically digested following the procedure described previously [66]. Briefly, the protein precipitate were resolved by 8 M Urea, and then sequentially treated with 5 mM TCEP and 10 mM NEM to reduce the di-sulfide bond and alkylate the resulting thiol group. The mixture was digested for 16 h at 37°C by trypsin at an enzyme-to-substrate ratio of 1:50 (w/w).
The trypsin-digested peptides were desalted with C18 Zip-Tips and then loaded onto an in-house packed capillary reverse-phase C18 column (15 cm length, 100 μM ID x 360 μM OD, 3 μM particle size, 100 Å pore diameter) connected to a Thermo Easy-nLC1000 HPLC system. The samples were analyzed with a 180 min-HPLC gradient from 0% to 100% of buffer B (buffer A: 0.1% formic acid in Water; buffer B: 0.1% formic acid in 20/80 water/acetonitrile) at 300 nL/min. The eluted peptides were ionized and directly introduced into a Q-Exactive mass spectrometer using a nano-spray source. Survey full-scan MS spectra (from m/z 300–1800) were acquired in the Orbitrap analyzer with resolution r = 70,000 at m/z 400.
Protein identification and post-translational modification analysis were done with Integrated Proteomics Pipeline—IP2 (Integrated Proteomics Applications, Inc., http://www.integratedproteomics.com) using ProLuCID/Sequest [67, 68] and DTASelect2 [69, 70]. Spectrum raw files were extracted into ms2 files from raw files using RawExtract [71], and the tandem mass spectra were searched against the Uniprot human protein database; plus sequences of known contaminants such as keratin and porcine trypsin concatenated to a decoy database in which the sequence for each entry in the original database was reversed using ProLuCID/Sequest. Alkylation (+125.0477) of cysteine and oxidation (+15.9949) of methionine were considered as static modifications, and ubiquitination (+114.0429) was considered as a variable modification. We require 2 peptides per protein and at least one tryptic terminus for each peptide identification. Search space included all fully- and half-tryptic peptide candidates with missed cleavage restrictions.
The Uniprot entry IDs for each genes described in the text are: P03409 (Tax), P22314 (E1), P0CG48 (Ubiquitin), P68036 (UbcH7), P62256 (UbcH2), P61077 (UbcH5c), P61088 (Ubc13), Q15819 (Uev2), O15111 (IKKα), O14920 (IKKβ), Q9Y6K9 (NEMO), P25963 (IκBα), O43318 (TAK1), Q8VSD5 (OspI), Q6TQR6 (vOTU), Q96BN8 (OTULIN), Q9NQC7 (CYLD), Q9Y4K3 (TRAF6).
|
10.1371/journal.pntd.0002296 | The Use of NanoTrap Particles as a Sample Enrichment Method to Enhance the Detection of Rift Valley Fever Virus | Rift Valley Fever Virus (RVFV) is a zoonotic virus that is not only an emerging pathogen but is also considered a biodefense pathogen due to the threat it may cause to public health and national security. The current state of diagnosis has led to misdiagnosis early on in infection. Here we describe the use of a novel sample preparation technology, NanoTrap particles, to enhance the detection of RVFV. Previous studies demonstrated that NanoTrap particles lead to both 100 percent capture of protein analytes as well as an improvement of more than 100-fold in sensitivity compared to existing methods. Here we extend these findings by demonstrating the capture and enrichment of viruses.
Screening of NanoTrap particles indicated that one particle, NT53, was the most efficient at RVFV capture as demonstrated by both qRT-PCR and plaque assays. Importantly, NT53 capture of RVFV resulted in greater than 100-fold enrichment from low viral titers when other diagnostics assays may produce false negatives. NT53 was also capable of capturing and enhancing RVFV detection from serum samples. RVFV that was inactivated through either detergent or heat treatment was still found bound to NT53, indicating the ability to use NanoTrap particles for viral capture prior to transport to a BSL-2 environment. Furthermore, both NP-40-lysed virus and purified RVFV RNA were bound by NT53. Importantly, NT53 protected viral RNA from RNase A degradation, which was not observed with other commercially available beads. Incubation of RVFV samples with NT53 also resulted in increased viral stability as demonstrated through preservation of infectivity at elevated temperatures. Finally, NanoTrap particles were capable of capturing VEEV and HIV, demonstrating the broad applicability of NanoTrap particles for viral diagnostics.
This study demonstrates NanoTrap particles are capable of capturing, enriching, and protecting RVFV virions. Furthermore, the use of NanoTrap particles can be extended to a variety of viruses, including VEEV and HIV.
| There is a dire need for fast and efficient diagnosis of many viral diseases. Our research specifically looked at RVFV, a virus that can only be worked with in biosafety level 3 (BSL-3) laboratories, and its capture with NanoTrap particles. NanoTrap particles are hydrogel particles that contain internal affinity baits. They have previously been used in the capture of several analytes, but never in the capture of whole virus particles. We were not only able to capture and detect RVFV at very low titers from both media and serum, but we were also able to inactivate the virus, which allows for its safe transport to BSL-2 laboratories. While there are other commercially available beads that can also capture virus, NanoTrap particles are the only beads that can protect the viral RNA from enzymatic degradation. Furthermore, we demonstrated that whole virus detection with NanoTrap particles is not limited to only RVFV, but that NanoTrap particles can be used to detect other viruses such as Human Immunodeficiency Virus (HIV) and Venezuelan Equine Encephalitis Virus (VEEV).
| Rift Valley fever virus (RVFV) belongs to the genus Phlebovirus and family Bunyaviridae. RVFV is composed of a tripartite single-stranded RNA genome with large (L), medium (M), and small (S) segments [1], [2], [3], [4]. RVFV particles have icosahedral symmetry and are 90–110 nm in diameter [4]. The envelope is made up of a lipid bilayer that is embedded with the Gn and Gc glycoproteins. These glycoproteins, which are the most exposed components of the virus during infection, play a crucial role in the entry of the virus into the host cell.
RVFV is a highly pathogenic arthropod-borne virus that is primarily transmitted by mosquitoes, particularly after heavy rainfall. Although it can infect a wide range of vertebrate hosts, RVFV primarily affects livestock and humans [2]. Animals are infected through mosquito bites and other arthropod vectors. Humans are typically affected when they come in close contact with infected bodily fluids or tissues, but transmission via mosquito bites, as well as aerosolization may also occur. However, humans are dead-end hosts [1], [5].
Since being first identified in 1930 in the Rift Valley of Kenya, outbreaks have led to high mortality rates as well as significant economic loss [3]. RVFV has remained endemic in sub-Saharan Africa, causing major outbreaks throughout the continent over the last century [6]. In 1976, 200,000 individuals were infected and 600 fatal cases were reported in Egypt [5]. Most likely due to international livestock trade, it has since crossed the Arabian Peninsula into Saudi Arabia and Yemen. Over 30 mosquito species, mostly Aedes and Culex are vectors for RVFV [5]. Of particular concern is that the Aedes species is widely distributed in the EU countries and many of those countries (Turkey, Greece, Italy, Spain, Portugal, and France) have high-risk vector habitat areas that may serve as emergent sites. Moreover, in the Unites States, this species has been found in 23 states [7]. Since RVFV is capable of utilizing a wide range of mosquito vectors, the virus has the potential to spread further into non-endemic areas [5], [8].
Mortality rates are dependent on species and age. In livestock, mortality rates are as high as 30%. Mortality rates can reach as high as 95% in newborns and the young, while abortion rates are as high as 100% [5]. Symptoms in humans are usually mild and include febrile illness resembling the flu, with a small percentage developing serious clinical manifestations such as retinal lesions, meningoencephalitis, hepatitis, severe hemorrhagic fever, coma and death. In recent years an increase in mortality amongst humans from 2% to 45% has been reported, suggesting evolving mechanisms of virulence and mutations [3].
Due to its transmission via aerosolization, high pathogenicity, and classification as a Group III (bioterrorism potential) Category A emerging infectious disease by the NIAID, work with RVFV requires BSL-3 containment. It is highly suggested that laboratory staff working with RVFV be vaccinated. Therefore, diagnosis of RVFV is restricted to a small number of laboratories. This limitation has led to some delay in diagnostics associated with virus isolation and identification techniques that may pose a problem for healthcare authorities in the event of an RVFV epidemic. There is a crucial need for rapid detection and identification of the virus [3], [5].
NanoTrap particles are a novel technology that can address all the critical analytical challenges for pathogen identification and measurement. They are homogenous hydrogel particles of about 800 nanometers in size that have a shell made of polymers of N-isopropylacrylamide (NIPAm) and co-monomers such as acrylic acid (AAc) and allylamine (AA) with cross links of N,N′-methylenebisacrylamide (BIS). This shell can be modified to alter permeability or porosity by increasing or decreasing the percentage of BIS [9], [10]. Charge-based affinity baits are incorporated into the NanoTrap particles by copolymerization and covalent binding to the shell [10]. The NanoTrap particles are temperature- and pH-sensitive, decreasing in size with increased temperature and low pH. The molecular sieving properties of the particles depend on several aspects. The degree of cross-linking within the particles provides inclusion and exclusion of high abundance large molecules (e.g. albumin). Affinity baits further facilitate the capture and concentration of the target protein, and prevent it from exiting the particle. They may be of negative or positive charge, therefore attracting analytes of opposite charge. This was seen in an early experiment performed by Luchini et al. where the incubation of particles containing anionic affinity baits captured myoglobin, a protein with a positive charge [9]. Some NanoTrap particles are composed of NIPAm shells, and a few of these shelled NanoTrap particles are also coated with vinyl sulfonic acid (VSA) [11]. NanoTrap particles are able to perform three functions in one step: molecular size sieving, target analyte affinity sequestration, and complete protection of captured analytes from degradation. Furthermore, NanoTrap particles help to bridge the gap between detection and the limits of sensitivity. Mass spectrometry (specifically liquid chromatography coupled with tandem mass spectrometry) is a favored technique for the discovery of candidate biomarkers in biological fluids. However, this technique only accepts a small input volume and complex solutions often lead to decreased sensitivity. The NanoTrap particles concentrate protein analytes in small volumes to effectively amplify the sensitivity of mass spectrometry. In addition, their promiscuity allows for multiple analytes to be harvested from a single sample [9]. Experiments conducted by Luchini et al. demonstrated the capture and enrichment of small molecules spiked in complex solutions such as whole blood and serum [9], [10]. A 2011 study by Douglas et al. on the detection of Lyme disease demonstrated that NanoTrap particles can improve sensitivity more than 100-fold (over existing methods) as well as lead to 100 percent capture and 100 percent elution yield of low abundance antigens in biofluids. Lyme disease antigens at low abundance were detected in both urine samples as well as from a single infected tick [12].
The current library of commercially available NanoTrap particles has been designed to specifically harvest proteins, peptides, metabolites and small molecules. We hypothesized that NanoTrap particles would be able to capture whole virus through the interaction of the NanoTrap particle with the positively charged residues on the surface of RVFV. Our study demonstrates that NanoTrap particles are capable of capturing whole virus, and can be assayed with both qRT-PCR and plaque assays. Importantly, serial dilution studies and studies in serum indicate that NanoTrap particles increase detection sensitivity at lower viral titers. Furthermore, the virus can be inactivated with either heat or detergent, while the virus captured can still be detected with qRT-PCR. Importantly, the NanoTrap particles protect purified viral RNA as well as stabilize the infectivity of RVFV. The studies described here expand upon the NanoTrap particles repertoire to characterize the capture of viruses.
The NIPAm/AA NanoTrap particles were provided by Ceres Nanoscience, Manassas, VA.
The Vero cell line (kidney epithelial cells) was grown in Dulbecco's Modified Eagle Medium (DMEM) supplemented with 10% FBS, 1% penicillin/streptomycin, and 1% glutamax (DMEM+++). The J1.1 cell line, which are Jurkat E6.1 suspension cells chronically infected with the LA1 strain of HIV-1, were grown in medium containing advanced RPMI-1640, 10% fetal bovine serum, 1% penicillin/streptomycin, and 1% L-glutamine. All cell lines were cultured in a humidified environment containing 5% CO2 at 37°C.
The experiments used a live attenuated vaccine derived from the RVFV ZH548 strain, known as MP-12, which had been isolated in 1977 from a patient with uncomplicated RVFV. The virus was generated by 12 serial passages in MRC5 cells, inducing 25 nucleotide changes across the viral genome [13]. Both RVFV ZH548 and MP12 strains were anonymized. MP12 was propagated by infecting Vero cells at 80–90% confluency at an MOI of 0.1 in DMEM+++. Cell culture medium was collected from the cells when ∼75% cytopathic effect was observed (typically 72 hours post-infection (hpi)). Cell culture medium was centrifuged at 10,000 rpm for 10 minutes to pellet the cellular debris. Cell free-viral supernatants were then filtered using a 0.22 µM filter and viral titer determined by plaque assays. Screening experiments for Venezuelan Equine Encephalitis Virus (VEEV) used the live attenuated vaccine TC-83, which had been derived from the Trinidad donkey (TrD) strain by 83 serial passages in fetal guinea pig hearts. This induced changes at 12 nucleotide positions across the viral genome [14]. The viral supernatant of chronically infected J1.1 cells was used in the HIV-1 screening experiments. The LAV strain of HIV-1 had previously been used to infected Jurkat E6 cells at a multiplicity of infection of 0.1 to 0.01 for 2 hours at 27°C and cultured for two weeks. The cells that survived the cytopathic effects of virus infection were cloned and the supernatant from growth-positive wells were screened for Reverse Transcriptase (RT) activity [15]. The J1.1 cells express viral RNA and proteins at low levels.
Six commercially available beads - DEAE-Sephadex (Sigma-Aldrich), Dynabeads M-280 Streptavidin (Invitrogen), Sephacryl S-200 beads (GE Healthcare), Biorex 70 Resin (Bio-Rad Laboratories), SP Sephadex C-25 (GE Healthcare), and Bio-gel HTP Hydroxyapatite (Bio-Rad Laboratories) were used to compare their capture to NT53. Each bead was washed four times with water and a 33% percent slurry with water was prepared.
According to a protocol standardized by Ceres Nanoscience, 100 microliters (µL) of sample was incubated with 75 µL of NanoTrap particles for 30 minutes at room temperature. The sample was centrifuged at 10,000 rpm for 5 minutes and the supernatant was discarded. The pellet was washed with 100 µL of RNase- and DNase-free water four times. The pellet was then resuspended in the appropriate buffer. For lysis of MP12 with NP-40, 1% NP-40 was added to 100 µl of MP12 and incubated at room temperature for 30 minutes. A standard NanoTrap particle incubation was performed followed by a qRT-PCR assay. For the RNase treatment, purified MP12 RNA was treated with RNase A and incubated for one hour at 37°C.
Vero cells were plated in 6 well plates at 1.0E+06 cells/ml in order to achieve 100% confluency. After NanoTrap particle incubation and subsequent washes, the pellet was resuspended in 100 µL of supplemented DMEM and serial dilutions performed. Four hundred µL of the serial dilution was added to each well in duplicate and incubated for 1 hour. Three hundred milliliters (mL) of a primary overlay known as the CV mixture containing equal parts 0.6% agarose in distilled water and media containing 2X EMEM, 5% FBS, 1% Minimum Essential Amino Acids, 1% Sodium Pyruvate, and 1% Glutamax was added directly to each well. The cells were fixed with 10% formaldehyde in water after 72 hpi. The cells were stained with 1% Crystal Violet in 20% ethanol and water. After two hours, the crystal violet stain was washed off and the plaques formed were counted to determine the plaque forming units per milliliter (pfu/ml).
After NanoTrap particle incubation and subsequent washes, the pellet was resuspended in 180 µL of lysis/binding solution (Life Technologies) containing guanidinium thiocyanate and incubated on ice for thirty minutes. The samples were spun at 13,000 rpm for 5 minutes at room temperature. The supernatant was transferred to a 96-well plate and RNA extraction was performed with Ambion's MagMax 96-well Viral RNA extraction kit according to manufacturer's instructions. In order to determine the number of viral genomic copies produced, qRT-PCR with viral specific primers was performed using RNA UltraSense One-Step Quantitative RT-PCR System (Life Technologies). The experiment was performed according to a standardized protocol using fifteen µL of master mix containing enzyme mix, 5X reaction mix, 50 mM magnesium sulfate (excluded for VEEV qRT PCR), ROX reference dye, 10 µM TaqMan fluorogenic probe, 10 µM forward primer (AAAGGAACAATGGACTCTGGTCA), and 10 uM reverse primer (CACTTCTTACTACCATGTCCTCCAAT) added to five µL of extracted RNA. The samples were heated at 50°C for 15 minutes, 95°C for 2 minutes, and at 95°C and 60°C for 40 cycles.
RVFV was spiked into 100% bovine, sheep, and donkey serum (purchased from Innovative Research) at 1.0E+05 pfu/ml. One mL of spiked serum was used in a standard NanoTrap particle incubation with NT53.
After a standard NanoTrap incubation was performed with NT53 and RVFV, the samples were incubated at room temperature with 0.1, 0.5, or 1% NP-40 for 1 hour or heated at 57°C for 0.5, 1, or 2 hours. The samples were then analyzed by plaque assays and qRT PCR.
A standard NanoTrap particle incubation with one mL of HIV-1 supernatant from infected J1.1 cells and NanoTrap particles was performed. An RNA extraction was performed (as described above). A master mix was then prepared with the following components for a 25 µL reaction: SuperScript III RT/Platinum Taq Mix - 0.5 µL, 2X Reaction Mix with ROX - 12.5 µL, and 0.5 µL of Forward and Reverse Primer Mix (LTR forward primer CGAGCTTGCTACAAGGGACT and LTR reverse primer GAGATTTTCCACACTGACTAAAAG) at 10 µM. Five µL of sample, 6.5 µL of water, and 13.5 µL of master mix were aliquoted out into PCR tubes. The PCR was conducted with the following cycles: 15 min at 50°C (cDNA synthesis), 2 min at 95°C (prime reaction), and 35 cycles at 30 seconds at 95°C (denature), 30 seconds at 51°C (annealing), 30 seconds at 72°C (extend), 72°C for 10 min, and a hold at 4°C. A DNA gel was prepared using 1% agarose powder in 1X TAE buffer with the addition of ethidium bromide for a final concentration of 0.5 µg/µl. The gel was visualized on an ultraviolet transilluminator and the volume of each band was quantified.
NanoTrap particles have previously been shown to capture proteins. We hypothesized that the RVFV glycoproteins would be capable of interacting with the NanoTrap particles, facilitating capture in a fashion similar to the way in which protein biomarkers interacted with the NanoTrap particles. To test this hypothesis, seven different NanoTrap particles were tested with RVFV. Four NanoTrap particles possessed shells (NT53, NT55, NT69, and NT71) whereas three did not (NT45, NT46, and NT75) (Table 1). We incubated culture supernatants from RVFV-infected Veros with each NanoTrap particle. All seven NanoTrap particles successfully captured virus, averaging 6.8E+07 genomic copies per reaction. Specifically, NT46, NT53, and NT69 captured higher genomic copies than the other NanoTrap particles, with each capturing approximately 1.0E+08 genomic copies per reaction (Figure 1A). This corresponds to 78–83% capture (Figure 1B) of a sample containing a high titer of virus.
In order to determine if the amplification observed in the qRT-PCR assay was due to the NanoTrap particle capturing intact viral particles or association of viral RNA (presumably due to lysed virus) with the particles, plaque assays were performed. If the particles captured viral RNA or lysed virus, no plaques should be observed. Plaque assays were performed on the three best candidates from the qRT-PCR screening. Captured viruses were not eluted off of the NanoTrap particles, but rather the samples were diluted and added directly to the Vero cells during the plaque assay procedure. We hypothesized that the viral glycoproteins would have a greater affinity for the cellular receptor than the NanoTrap particles and thus would enter the cells. NT53, which contains a cibacron blue bait with a shell, captured infectious RVFV virion six- and four-fold more than NT46 or NT69, respectively (Figure 1C). These plaques were not due to cell death induced by the NanoTrap particles themselves, as NanoTrap particles alone did not produce plaques. Therefore NT53 was chosen for all future experiments with RVFV.
Experiments were performed to determine potential elution methods that would release the virus without affecting the viral particle integrity. It had previously been found that sodium chloride (NaCl) concentrations between 0.5M and 2M could effectively elute various analytes from cibacron blue dyes by disrupting electrostatic interactions between cibacron blue dyes (the bait molecule found within NT53; [16]). We hypothesized that incubating the RVFV-bound NanoTrap particles on ice would allow the particles to swell and, with the aid of vortexing, the virus would disassociate from the NanoTrap particles. Therefore, we tested a NaCl based elution method coupled with an ice-swelling method. Plaque assays were performed to determine the amount of virus eluted from the NanoTrap particles and the amount that remained bound to the particles after a high salt elution. After NT53 incubation with RVFV, the pellets were resuspended in 2.0 M NaCl in DMEM and placed on ice for 30 minutes with vortexing every ten minutes. Both the eluates and pellets were analyzed by plaque assay (Figure S1). The results showed that 5.5% of RVFV was detected after elution with 2.0M NaCl. The addition of NaCl coupled with incubation on ice only slightly released RVFV virions, demonstrating the virus' strong affinity for the NanoTrap particles. However, as seen in Figure 1B, RVFV-bound NanoTrap pellets directly added to Vero cells during the plaque assays procedure were capable of producing plaques. Based on these results, we opted not to elute RVFV from the NanoTrap particles, but rather to add the RVFV bound to the NanoTrap particles directly during the plaque assay procedure.
We next wanted to determine the limit of detection of RVFV in plaque assays with NT53. NT53 was incubated with RVFV at decreasing titers, from 2.5E+6 to 2.5E+1 pfu/ml, and plaque assays performed (Figure 2A). Captured virus was detected down to 2.5E+1 pfu/ml for RVFV. These results show that NanoTrap particles are capable of capturing whole virus even at low viral titers.
We next determined the percentage of RVFV captured by NT53 in comparison to the total input amount. RVFV at 1.0E+6 and 1.0E+3 pfu/ml were added to NT53. At 1.0E+6 pfu/ml, 99.35% of the virus was bound to the NanoTrap particles whereas at 1.0E+03 pfu/ml, ∼100% of the virus appeared bound to the NanoTrap particles (Figure 2B). The results confirm that the NanoTrap particles are efficient at capturing RVFV, especially at a lower titer. Interestingly, the results also suggest that a small volume of RVFV can be captured with NanoTrap particles and then recultured to grow more virus.
In clinical instances of infection, the viral titers in circulation during very early stages after exposure are expected to be low and therefore, hard to detect [17]. We wanted to determine if viral enrichment by the NanoTrap particles (NT53) would enhance detection of RVFV when compared to detection in the absence of enrichment afforded by the NanoTrap. We specifically wanted to see the enrichment potential at lower viral titers when detection would be most difficult. For these assays, we chose to utilize qRT-PCR based detection due to its increased sensitivity over plaque assays. To this end, we spiked RVFV into cell culture media that contained 10% FBS at various concentrations from 1.0E+5 to 1.0E+1 pfu/ml. NT53 was then added to 1 ml of the spiked media. Viral capture with and without NanoTrap particles gave similar yields at higher viral titers (Figure 3A). However, at lower viral titers, there was a significant increase in viral capture with the use of the NanoTrap particles compared to samples without NanoTrap particle capture. There was greater than a 100-fold increase of viral detection with the use of NT53 at 1.0E+1 pfu/ml.
We next wanted to determine if we could capture and enrich virus from a clinically relevant matrix. RVFV was spiked into 100% bovine, donkey, and sheep sera at 1.0E+05 pfu/ml. Incubation of RVFV spiked sera with NT53 resulted in enrichment by 13-, 3-, and 52-fold for bovine, sheep, and donkey sera, respectively (Figure 3B). These results demonstrate that NT53 not only captures but also enriches virus found in complicated matrices such as animal sera. The complex analytes (e.g. albumin) found in the sera are likely excluded by the NanoTrap particles and do not interfere with whole virus capture. However, we speculate that since the serum from each of the three animals contains different analytes, there may be interfering proteins that would lead to the observed enrichment differences.
Viral inactivation is crucial for its transport from the field or a BSL-3 facility to a BSL-2 environment for downstream analysis. However, after inactivation the virus may be susceptible to degradation. Therefore, we wanted to determine if RVFV would remain bound to the NanoTrap particles in an inactivation scenario. After NanoTrap particle incubation with RVFV and subsequent washes, NP-40 detergent was used to inactivate the virus (Figure 4). Plaque assays were performed to confirm viral inactivation. Plaque assays demonstrated that 0.1% NP-40 did not fully inactivate the virus incubated with or without NT53 (Figure 4A). Higher concentrations of NP-40 (0.5% and 1%) fully inactivated RVFV in the presence or absence of NT53. While the plaque assays confirmed inactivity of RVFV, qRT-PCR data demonstrated that RVFV was still captured following NP-40 addition (Figures 4B). In the presence of NP-40 the levels of capture with NT53 were decreased as compared to the controls. This is likely due to the interference of the NanoTrap particle binding to RVFV due to the presence of detergent.
We next tested the ability of NT53 to function in another commonly employed viral inactivation procedure. The samples were heat inactivated at 57°C for three different time points - thirty minutes, one hour, and two hours - and plaque assays were performed to confirm viral inactivation. At thirty minutes approximately 1.0E+4 and 1.0E+2 pfu/ml of RVFV with and without NT53, respectively, were still detectable. Interestingly, RVFV was more resistant to heat inactivation in the presence of NT53, suggesting the NanoTrap particles may have a slight protective effect on the virus. However, complete inactivation was achieved at one hour (Figure 4C). While the plaque assays confirmed inactivation of RVFV, qRT-PCR data demonstrated the ability of NT53 to detect RVFV nucleic acids after heat inactivation (Figure 4D). These data demonstrate that it is possible to fully inactivate the virus before applying downstream assays such as qRT-PCR. Furthermore, these two experiments demonstrate the ability to inactivate a sample and transport it as a non-infectious sample, while still retaining capture.
As we observed viral capture in the presence of NP-40, we hypothesized that the virus was being lysed and the released viral RNA recaptured with the NanoTrap particles. If the NanoTrap particles were not providing protection of the viral RNA, there will be no possibility for any downstream assays using inactivated material, which is a critical step in diagnostics. To test this hypothesis, we first lysed the virus with 1% NP-40 and followed by adding NT53 to the lysed material. Results in Figure 5A indicate that NT53 was able to capture the lysed virus. However, there was a 100-fold decrease in lysed virus with the addition of NT53 compared to the control (no NT53) with and without NP-40. These results mirrored what was observed in Figure 4B, where NT53 was capable of capturing virus to a lesser extent in the presence of NP-40. These results demonstrated that RVFV could be initially inactivated by traditional inactivation methods and then captured with NanoTrap particles. By incubating with the NanoTrap particles, the viral RNA will be protected and hence, can be used for downstream RNA detection.
In order to directly show that NT53 was able to capture and protect viral RNA, we performed a NanoTrap experiment with purified viral RNA. We first incubated NanoTrap particles with purified RNA, and performed qRT-PCR assays. Results in Figure 5B demonstrate that NT53 is capable of capturing purified viral RNA, albeit with less affinity than whole virus capture. NT53 was able to capture 0.01% of the input viral RNA. While we screened the NanoTrap particles for whole virus capture, we did not screen the NanoTrap particles for RVFV viral RNA capture. There is likely a NanoTrap particle that captures viral RNA with greater efficiency than NT53.
As previous studies have shown that proteins captured by NanoTrap particles were protected from trypsin degradation, we next aimed to determine if NanoTrap particle capture could protect viral RNA from RNase degradation [9], [10]. Samples with and without NT53 were treated with RNase A at 140 or 1400 Units/ml and incubated for one hour at 37°C. Interestingly, the RNA incubated with NT53 was protected from RNase A degradation, whereas the RNA controls were subject to complete RNase A degradation (Figure 5B). At 140 Units/ml of RNase A, the captured RNA was detected at the same level as the RNase-untreated sample. Even at a substantially higher RNase concentration (1400 Units/ml), 1% of the viral RNA input was still detected. Our results demonstrate that the NanoTrap particles are capable of capturing and protecting viral RNA from enzymatic degradation.
In some situations, it may be important to retain the infectivity of the captured virus to enable the virus to be propagated for further characterization. Therefore, we evaluated the ability of the NanoTrap particles to capture and preserve the infectivity of RVFV following capture. RVFV was spiked into bovine serum and incubated with or without NT53 at 25°C for 48 or 72 h. In the absence of NT53, the infectivity of RVFV was decreased by ∼3 logs (Figure 5C). In contrast, samples incubated with NT53 displayed only ∼1.5 log decrease by 48 h. Although ∼3.5 log decrease was observed with NT53 at 72 h, this still resulted in increased virus detected as compared to the control samples due to the enrichment afforded by NT53. The infectivity of RVFV was also assessed for samples that were incubated at 37°C, which would likely result in a more rapid decline in viral infectivity and thus 24, 48, and 72 h time points were examined. As suspected, a ∼4 log decrease was observed in samples incubated at 37°C for 24 h without NT53. In contrast, samples captured by NT53 only displayed ∼2 log decrease as compared to the control NT53 sample. At extended time points a further decrease in infectivity was observed with and without NT53, but in all cases a higher amount of infectious virus could be rescued from samples incubated with NT53. Collectively these results demonstrate that NanoTrap particles can capture and preserve viral infectivity up to 72 h at elevated temperatures.
NanoTrap particles have unique properties not demonstrated in other beads that are used for protein purification and albumin exclusion such as dye baits that make them an ideal candidate in virus capture. Therefore, we wanted to directly compare the ability of other beads to capture RVFV with NT53's RVFV capture capability. The capture of RVFV was tested with NT53 and six commercially available beads used in various assays. DEAE-Sephadex beads are used in ion exchange chromatography for purifying and isolating proteins; Dynabeads M-280 Streptavidin are used for isolating nucleic acids and antibodies; Sephacryl S-200 beads are used to purify protein and macromolecules; Biorex 70 Resin beads are used for purification and fractionation of peptides, proteins, and other cationic molecules; SP Sephadex C-25 beads are used in chromatography to separate and purify protein, polypeptides, and other charged molecules; and Bio-gel HTP Hydroxyapatite beads are used in chromatography to separate and purify proteins, nucleic acids, viruses, and other macromolecules. These calcium phosphate beads work by the cationic interaction of the Ca2+ functional groups with the carboxylate residues located on the protein surface and the anionic interaction of the PO42− functional groups with the basic protein residues. RVFV was incubated with each of these beads and plaque assays were performed to determine whole virus capture. Bio-gel HTP Hydroxyapatite (HTP) captured RVFV the most efficiently, averaging 1.0E+7 pfu/ml, while NT53 performed the second best averaging 2.5E+6 pfu/ml (Figure 6A). The other four beads captured RVFV around or below 1.0E+5 pfu/ml.
As we have demonstrated that NT53 not only captures intact RVFV, but can also capture and protect viral RNA from RNase A degradation, we tested the ability of HTP to act in a similar capacity. NT53 was able to fully protect the viral RNA against RNase A degradation and genomic copies for NT53 with and without NT53 treatment were similar. However, HTP beads were unable to provide protection against RNase degradation, and no viral RNA was detected (Figure 6B). A control experiment with RNA alone demonstrated that our RNase treatment was effective. We next compared the ability of NT53 and HTP to capture RVFV during an inactivation scenario. For these experiments NT53 or HTP were added to the samples followed by viral inactivation through treatment with 1% NP-40 or heating at 57°C. The amount of virus captured was quantitated by qRT-PCR (Figure 6C). As was observed in previous experiments, viral inactivation with either NP-40 or heat treatment resulted in some loss of RVFV binding to the NanoTrap (1.2 and 0.8 log, respectively). However, HTP RVFV capture was more dramatically affected, resulting in a 2.5 log decrease with the NP-40 treated samples and a 2.7 log decrease in the heat inactivated samples. Collectively, our experiments demonstrate that while HTP is capable of capturing whole virus, it cannot protect viral RNA against RNase degradation and it displays a reduced ability to capture RVFV during an inactivation scenario. In contrast, NT53 is capable of capturing and protecting RVFV as well as capturing RVFV in samples that have been inactivated by heat or detergent treatment.
We next asked the question if NanoTrap particles were capable of capturing other viruses. For this, we selected VEEV and HIV-1. VEEV, which at approximately 70 nm in diameter is a smaller virus than RVFV, which is approximately 100 nm in diameter. VEEV viral supernatants were incubated with various NanoTrap particles shown in Table 1 and capture was measured by qRT-PCR. Our data indicated that all six NanoTrap particles successfully captured VEEV, averaging 9.9E+06 genomic copies per reaction (Figure 7A), with a slight preference observed with NT45, NT46, and NT55 capturing 1.3E+07, 1.1E+07, and 1.1E+07 genomic copies per reaction, respectively. NanoTrap particles capture was also tested using HIV-1. HIV-1 supernatants from infected J1.1 cells were incubated with NanoTrap particles. RNA extraction was performed, cDNA was synthesized, and RT-PCR was performed. A semi-quantitative analysis shown in Figure 7B, demonstrated that all seven NanoTrap particles were able to capture HIV-1 with NT46 and NT53 demonstrating the best capture (Figure 7B). These results indicate that NanoTrap particles are capable of capturing multiple viruses.
Viral infections in nature do not occur in isolation and are often accompanied by other co-infections (bacterial and/or viral); therefore we sought to determine if the NanoTrap particles could capture RVFV in a “mixed” infection setting. To this end, bovine serum was spiked with RVFV only or with both RVFV and HIV, followed by NT53 viral capture and quantification as measured by qRT-PCR. Results indicated that NT53 was capable of capturing and enriching RVFV from samples that contained only RVFV or both RVFV and HIV (Figure 7C). Seven-fold enrichment was observed in samples containing RVFV only and 5-fold enrichment from samples containing both RVFV and HIV. These data provide evidence that the NanoTrap particles could be used with clinical samples.
Therefore, in conclusion, the results demonstrate that NanoTrap particles can capture and enrich RVFV from both cell culture media and clinically relevant matrices. The captured virus can then be inactivated and viral RNA protected from enzymatic degradation. The bound RVFV can be eluted off the NanoTrap particles, and used in downstream assays such as plaque assays and qRT-PCR. Furthermore, NanoTrap capture can be extended to other viruses as well, including VEEV and HIV.
Rift Valley Fever Virus is a zoonotic virus that primarily affects livestock but has the potential to cause severe disease in humans. RVFV has led to outbreaks in Egypt and the Arabian Peninsula with the potential to spread to the United States and Europe. Changes in climate, travel, and trade have made RVFV an emerging disease that can have deadly economic and social consequences. Furthermore, RVFV is of biodefense interest due to its potential spread via aerosolization. There are currently no FDA-approved vaccines, so there is a reliance on sensitive and specific diagnostics early on in infection.
The current state of RVFV diagnostics includes virus isolation, nucleic acid techniques, and antibody detection. Current RT-PCR-based assays require a critical amount of the virus circulating in the system. This can lead to misdiagnosis, especially false-negative results, of the disease early on in infection. In contrast, our results demonstrate the ability of NanoTrap particles to enrich for RVFV from both cell culture supernatants as well as more complex matrices such as animal serum. The capability of NanoTrap particles to enrich virus is crucial early on in infection during which the virus can go undetected using other diagnostic methods. In our serum sample studies we noted different levels of enrichment depending on the source of the serum. For example, NanoTrap particles incubated in donkey serum resulted in a 52-fold increase in RVFV detection sensitivity, whereas incubation in sheep serum only displayed a 3-fold increase. The presence of other analytes found in serum may be competing for capture with NanoTrap particles, which likely will differ between species as well as between individual animals. Importantly we have also demonstrated the ability to capture RVFV in samples that also contained HIV. This is an important area of investigation, as clinical samples will likely contain multiple pathogens, providing further competition for NanoTrap binding. We hope to extend our spiked serum sample studies to experiments with serum samples taken from animals exposed to RVFV and human clinical samples. These studies will allow further optimization of the NanoTrap particle collection. In our current study we used very stringent wash conditions to ensure that the virus captured was tightly bound to the NanoTrap particles. In clinical samples, it may be necessary to decrease the number of wash steps to ensure that the maximum amount of virus is being captured from more complex samples. Alternatively, different wash buffers (altering salt and detergent concentrations) could be utilized to allow more selective binding of analytes. Due to the complex nature of clinical samples, it may also be necessary to increase the amount of NanoTrap particles added to prevent saturation. Nonetheless, our studies provide an important first step in the application of NanoTrap particles as a sample preparation and enrichment process to improve diagnostics from serum samples.
One important advantage of utilizing NanoTrap particles is their ability to protect analytes from degradation. Previous studies have indicated that protein captured by NanoTrap particles are protected from trypsin degradation [9], [10]. In these experiments PDGF was incubated with an excess of trypsin. Following incubation the majority of the trypsin was found outside of the NanoTrap particles. However, even though some of the trypsin entered the particles, PDGF was completely protected from degradation. In the current study we extend these findings to demonstrate that viral RNA was protected from degradation in the presence of RNase A. Sample preservation is critical for stabilization of sample integrity both during field collection and during transported to diagnostic facilities. Bio-gel HTP Hydroxyapitite, while able to capture RVFV was unable to protect viral RNA from degradation, further demonstrating the advantage of using NanoTrap particles over other commercially available chromatography beads.
Another critical aspect of the NanoTrap particles is the ability to collect viral samples and inactivate them to render them non-infectious, while still retaining the ability to detect the analyte of interest. This was demonstrated by captured of RVFV by NT53 followed by inactivation of RVFV with NP-40 (determined by plaque assays). Following inactivation, viral genetic material was still detected with qRT PCR and to a higher level than that observed with HTP beads. This is of particular importance in the transport of RVFV from a BSL-3 environment or field sample collection setting to a BSL-2 laboratory for diagnostic testing. Given that BSL-3 laboratories are both difficult to access and work in a BSL-3 environment is time-consuming and expensive, the inactivation method will allow for fewer lapses in time between obtaining the samples and the determining the results. In addition, as the NanoTrap particles allow the capture of the whole virus, the samples can then be analyzed with a variety of downstream analysis methods such as ELISA for the nucleoprotein of RVFV, western blotting, plaque assays, and qRT-PCR.
The exact mechanism of NanoTrap binding to RVFV is unclear at this point. We hypothesize the NanoTrap particle capture is occurring through interactions with RVFV's glycoproteins, Gn and Gc. Gn and Gc are the only viral proteins available for capture by virtue of being exposed on the outside of the virion. The fact that the binding observed with HTP beads slightly exceeded NT53 in binding RVFV may provide insight into the mechanism of binding. HTP is known to bind primarily through electrostatic interactions and similarly, Cibacron blue, the affinity bait component of NT53 binds through electrostatic, hydrophobic or a combination of surface and electrostatic interactions. Based on these results, we expect that electrostatic interactions may provide the dominant mode of NT53 binding to RVFV. Due to the size of the virus (90–100 nm) as compared to the size of the NanoTrap particles (800 nm), it is unclear if the viruses are entering inside the core of the NanoTrap particle or binding to the outside of the NanoTrap particles. We have observed preferentially binding of RVFV with NanoTrap particles containing Cibracon blue baits, suggesting that the bait plays at least a partial role in the binding. Even if RVFV binding is partially or primarily found on the surface of the NanoTrap particles, the particles provide a unique advantage over other commercially available beads, which is sequestration of analytes within the NanoTrap particles. This is important as many enzymes (proteases, RNase,etc.) found in serum can rapidly digest protein and RNA. However, the NanoTrap particles can bind to small molecular weight proteins (such as trypsin) rendering them inactive [9], thereby providing protection for other proteins and/or viruses captured by the NanoTrap particles.
NanoTrap particles can be engineered with increased pore sizes to facilitate capture of RVFV inside the NanoTrap particles. This approach has the added advantage of ensuring capture within the particles themselves, which would be predicted to further increase the stability of the virus as well as increase the viral binding capacity of the NanoTrap particles. One potential disadvantage of larger pores sizes would be the loss of some of the sieve sieving capabilities of the particles. The size sieving is important for more complex samples (whole blood, sera, etc.), which would benefit from the ability of the NanoTrap particles to enrich for certain analytes (i.e. viruses) while excluding high abundant proteins such as BSA, which may interfere with downstream assays or mask lower abundant molecules such as low levels of viruses. However, it may be possible to obtain a balance of larger pores with efficient size sieving if the appropriate level of cross-linking could be achieved. This is an active area of research and warrants further investigation.
Our results demonstrated that NanoTrap particle capture was not limited to RVFV, but could be extended to other viruses including VEEV and HIV. All three of these viruses are enveloped RNA viruses. Future studies will focus on the capture and enrichment of different viral classes, including DNA viruses and non-enveloped viruses. That capture of a wide range of viruses is especially important when multiple viruses cause the same type of disease and/or the symptoms of infection are very general. For example, respiratory infections can be attributed to multiple pathogens, including Influenza A and B viruses, Coronaviruses, and Adenoviruses. For this type of application, the promiscuity of the NanoTrap particles is particularly important, as it will allow the capture and enrichment of multiple viruses from the same sample. Therefore, we believe that increasing sensitivity for respiratory viral infections is an important diagnostic issue that the NanoTrap particles could address.
In conclusion, our results have demonstrated that NanoTrap particles are able to capture and enrich whole virus. While other commercially available beads can also capture virus, only NanoTrap particles are capable of protecting the integrity of the virus after inactivation with detergent or exposure to RNase A. However, further research is needed to determine the exact mechanism by which the NanoTrap particles capture and protect the virus.
|
10.1371/journal.ppat.1003975 | Implication of PMLIV in Both Intrinsic and Innate Immunity | PML/TRIM19, the organizer of nuclear bodies (NBs), has been implicated in the antiviral response to diverse RNA and DNA viruses. Several PML isoforms generated from a single PML gene by alternative splicing, share the same N-terminal region containing the RBCC/tripartite motif but differ in their C-terminal sequences. Recent studies of all the PML isoforms reveal the specific functions of each. The knockout of PML renders mice more sensitive to vesicular stomatitis virus (VSV). Here we report that among PML isoforms (PMLI to PMLVIIb), only PMLIII and PMLIV confer resistance to VSV. Unlike PMLIII, whose anti-VSV activity is IFN-independent, PMLIV can act at two stages: it confers viral resistance directly in an IFN-independent manner and also specifically enhances IFN-β production via a higher activation of IRF3, thus protecting yet uninfected cells from oncoming infection. PMLIV SUMOylation is required for both activities. This demonstrates for the first time that PMLIV is implicated in innate immune response through enhanced IFN-β synthesis. Depletion of IRF3 further demonstrates the dual activity of PMLIV, since it abrogated PMLIV-induced IFN synthesis but not PMLIV-induced inhibition of viral proteins. Mechanistically, PMLIV enhances IFN-β synthesis by regulating the cellular distribution of Pin1 (peptidyl-prolyl cis/trans isomerase), inducing its recruitment to PML NBs where both proteins colocalize. The interaction of SUMOylated PMLIV with endogenous Pin1 and its recruitment within PML NBs prevents the degradation of activated IRF3, and thus potentiates IRF3-dependent production of IFN-β. Whereas the intrinsic antiviral activity of PMLIV is specific to VSV, its effect on IFN-β synthesis is much broader, since it affects a key actor of innate immune pathways. Our results show that, in addition to its intrinsic anti-VSV activity, PMLIV positively regulates IFN-β synthesis in response to different inducers, thus adding PML/TRIM19 to the growing list of TRIM proteins implicated in both intrinsic and innate immunity.
| PML is expressed as seven isoforms, designated PMLI to PMLVIIb, each with specific functions conferred by their C-terminal regions. PML isoforms are implicated in several cell processes, including antiviral defense. Very few studies have been performed with all PML isoforms to investigate their individual antiviral properties. Our comparative study with all PML isoforms on VSV replication revealed that only PMLIII and PMLIV are able to inhibit this virus in an IFN-independent way. Importantly, PMLIV is also able to enhance IRF3 phosphorylation resulting in a dramatic IFN-β production in response to viral infection, thus protecting yet uninfected cells. At the opposite, the specific suppression of PMLIV expression in human cells reduced virus-induced IRF3 activation and subsequent IFN-β production. We found that PMLIV affects the endogenous distribution of Pin1, by recruiting this protein in PML NBs where both proteins colocalize. This prevents activated IRF3 degradation, thus enhancing the production of IFN-β. Here we show for the first time that, in addition to its intrinsic anti-VSV activity, PMLIV is implicated in antiviral innate immune response. Thus, PML is an IFN stimulated gene whose products have a broad intrinsic antiviral activity and one of them, PMLIV, is also a potent regulator of innate immune pathways.
| The establishment of an antiviral state in cells is the defining property of interferons (IFNs), as well as the activity that led to their discovery. IFNs are the first line of defense against viral infections. IFN-regulatory factor 3 (IRF3), a ubiquitously expressed transcription factor, is responsible for the primary induction of IFN and is a crucial player in the establishment of innate immunity in response to viral infection [1]. IFNs bind to their receptors and activate the canonical JAK/STAT pathway, leading to the induction of IFN-stimulated genes (ISGs), whose products mediate their biological effects [2], [3].
Vesicular Stomatitis Virus (VSV) belongs to the Rhabdoviridae family. Its single stranded negative sense RNA genome (about 12 kb), that encodes 5 viral proteins, is encapsidated by the nucleoprotein N to form the nucleocapsid that is associated with the RNA-dependent RNA polymerase L and its cofactor the phosphoprotein P. Inside the viral particle, this nucleocapsid is associated with the matrix protein M and surrounded by a membrane containing a unique glycoprotein G. The virus enters the host cell through the endosomal transport pathway via a low-pH-induced membrane fusion process catalyzed by the glycoprotein G [4]. The nucleocapsid released into the cytoplasm serves as a template for transcription and replication processes that are catalyzed by the L-P polymerase complex [5], [6]. During transcription, a positive–stranded leader RNA and five capped and polyadenylated mRNAs are synthesized. These viral mRNA are translated by the cellular machinery to give the viral proteins N, P, M, G and L, then the replication process yields full-length antigenome-sense RNA, which in turn serve as templates for the synthesis of genome-sense RNA. During their synthesis, both the nascent antigenome and the genome are encapsidated by N proteins. The neo-synthesized genome either serves as a template for secondary transcription or is assembled with M proteins to allow budding of the neosynthesized virion at a cellular membrane.
VSV replication is highly sensitive to the inhibitory action of IFN and is routinely used to assay the antiviral activity of IFN in vitro [7]. Although IFN treatment induces the expression of hundreds of ISGs, only a few of them have been demonstrated to be responsible for the inhibition of VSV replication. Indeed, ISG products such as double-stranded RNA-activated protein kinase (PKR) [8], myxovirus resistance protein (Mx) [9], p53 [10], ISG20 [11], Ifit2/ISG54 [12] and ProMyelocytic Leukemia (PML) [13], [14] have been reported to confer resistance to VSV infection. In addition, 34 ISG products including PML have been shown to elicit an antiviral effect on VSV replication [15].
PML (also named TRIM19 for TRIpartite Motif protein 19) is the organizer of small nuclear-matrix structures named nuclear bodies (NBs) [16]. In response to diverse stimuli, PML NBs recruit a growing number of proteins implicated in different cellular processes such as DNA damage response, apoptosis, senescence, protein degradation and antiviral defense [17]–[25]. PML is covalently conjugated to small ubiquitin modifier (SUMO) on three major lysine residues (K65, K160, K490) [26]. This modification, which affects PML localization, stability and ability to interact with other partners, is critical for NB functions [16], [27].
Several PML isoforms generated by alternative splicing from a single gene are designated PMLI to PMLVIIb [28], [29]. They share the same N-terminal region, which encodes the RBCC/TRIM (RING finger, B-box, and Coiled-Coil) motif, but differ in their C-terminal region due to alternative splicing. The variability of the C-terminal part of PML isoforms is important for the recruitment of specific interacting partners and for the specific function of each [29]. The implication of PML in antiviral defense against RNA and DNA viruses from different families has been demonstrated in cells stably expressing individual PML isoform or in cells depleted for PML by RNA interference (reviewed in [17], [21]). We have previously shown that PMLIII confers resistance to VSV [13]. The antiviral effect of PML has been observed in vivo, as PML deficiency renders mice more susceptible to VSV infection [14]. The role of the other PML isoforms in VSV-infected cells is so far unknown. Therefore, we studied the implication of all PML isoforms during VSV infection. We show that only stable expression of PMLIII or PMLIV conferred resistance to VSV infection. Whereas the activity of PMLIII was saturated at a high multiplicity of infection (MOI), PMLIV conferred a higher protective effect towards VSV infection, even at high MOIs. Finally, unlike PMLIII whose anti-VSV activity is strictly IFN-independent, we show that PMLIV confers a strong resistance to VSV infection via two independent mechanisms. Indeed, PMLIV is able to block viral replication in an IFN-independent manner, and also to trigger innate immunity pathways, leading to higher activation of IRF3 and specific enhancement of IFN-β production. Both activities of PMLIV required its SUMOylation. The peptidyl-prolyl isomerase (Pin1) that is known to interact with phosphorylated IRF3 and to promote its degradation via the ubiquitin-proteasome pathway [30], was recruited within PML NBs in human cells expressing PMLIV. Therefore, the interaction of endogenous Pin1 with SUMOylated PMLIV and its recruitment within NBs resulted in an enhanced IRF3-dependent production of IFN-β in response to VSV infection and also to other inducers such as Sendai virus (SeV), encephalomyocarditis virus (EMCV), human T-lymphotropic virus type 1 (HTLV-1), influenza virus, vaccinia virus or poly(I:C). Remarkably, specific depletion of PMLIV in human cells reduced both SeV-induced IRF3 activation and IFN-β production. Our results demonstrate for the first time PMLIV's involvement in both intrinsic and innate immunity.
As PML knockout mice are more sensitive to VSV infection than parental mice [14], we studied viral protein expression and viral production in MEFs derived from these mice. The five structural proteins of VSV are the nucleocapsid, N, the matrix protein, M, the glycoprotein, G, and two minor proteins, the phosphoprotein, P, and the polymerase protein, L, but only the G, N, and M proteins were revealed with our rabbit anti-VSV antibodies. VSV proteins and VSV production were lower in MEFs WT compared to MEFs PML-/- (Figure 1A). In addition, down-regulation of PML expression by RNA interference in human U373MG cells boosted VSV protein expression (Figure 1B).
To determine which PML isoforms were capable of inhibiting VSV replication, we infected at an MOI of 0.1 U373MG cells transfected with an empty vector (EV) or stably expressing each PML isoform (PMLI, PMLII, PMLIII, PMLIV, PMLV, PMLVI, or PMLVIIb). The expression level of the different PML isoforms is shown in Figure S1A. Double immunofluorescence staining for PML and VSV antigens revealed that only PMLIII and PMLIV inhibited expression of viral proteins (Figure 2). Western blot of extracts from all these cells infected at different MOIs confirmed these results (Figure S1B). As previously shown [13], the capacity of PMLIII to inhibit viral protein expression was observed at low MOIs but decreased at higher MOI, whereas PMLIV inhibited viral protein synthesis even at an MOI of 2 (Figure S1B). To further confirm this result, extracts from U373MG-EV, U373MG-PMLIII and U373MG-PMLIV cells infected at MOIs of 0.2 or 1 were analyzed in the same Western blot and their supernatants were used for the determination of virus yield by standard plaque assay. As seen in Figure 3, compared to PMLIII, PMLIV had a higher capacity to inhibit viral protein synthesis and VSV multiplication. Compared to control cells, PMLIII had a slight effect on VSV production (2-fold inhibition) when cells were infected at an MOI of 1 for 8 h, whereas PMLIV exhibited a 500-fold inhibition (Figure 3B). Furthermore, compared to control EV cells, U373MG-PMLIV cells were protected against cell lysis 12 h post-infection at different MOIs (Figure 3C), whereas U373MG-PMLIII cells were protected only at a low MOI of 0.2 (data not shown). To further confirm the effect of PMLIV on VSV infection, we tested another cell clone stably expressing PMLIVa, a PMLIV variant, lacking exon 5 but having the same C-terminal region. This PML isoform was also able to inhibit VSV protein expression, as observed with PMLIV (data not shown).
Taken together, these results demonstrate that among all PML isoforms, only PMLIII, PMLIV and PMLIVa conferred resistance to VSV. However, PMLIV and PMLIVa mediated a much higher protection, suggesting the crucial role of their specific C-terminal portion.
To investigate which viral step is targeted by PMLIV, we first investigated whether VSV entry was affected. Thus, we used a MLV virus encoding GFP pseudotyped with the receptor-binding G protein (VSV-G) [31]. This pseudovirus can undergo VSV-G-mediated entry but cannot produce its own VSV-G envelope, and hence is only capable of a single-round of infection. MLV-G-GFP was used to transduce U373MG-EV and U373MG-PMLIV cells and GFP expression was readily detected 48 h later by immunofluorescence (Figure 4A, left panel) and flow cytometry (Figure 4A, right panel). These analyses revealed that U373MG-EV and U373MG-PMLIV cells had similar GFP staining reaching 80.1% and 84.4% GFP+ cells, respectively. These results demonstrate that VSV entry was not affected in PMLIV expressing cells.
Next, to determine whether viral transcription was altered by PMLIV expression, U373MG-EV and U373MG-PMLIV cells were left uninfected or infected with VSV at an MOI of 3 for 4 h. Total RNA from cell extracts was analyzed by Northern blot for VSV-N mRNA (Figure 4B, left panel) and VSV-L mRNA quantified by RT-qPCR (Figure 4B, right panel). Compared to U373MG-EV cells, the amount of N and L mRNAs was highly reduced in U373MG-PMLIV cells (Figure 4B). It is known that treatment of cells with the protein synthesis inhibitor cycloheximide (CHX) results exclusively in primary mRNA synthesis, as viral genome replication requires the ongoing synthesis of N protein [32]. Interestingly, the N mRNA level was comparable to the level synthesized in the presence of CHX, indicating that secondary transcription was inhibited by the PMLIV expression (Figure 4B, left panel) whereas primary transcription was not. These data suggest that PMLIV had no effect on steps preceding transcription but restricts a post-transcriptional step involving protein synthesis and replication. It is therefore possible that PMLIV induces IFN synthesis during VSV infection, which may in turn inhibit viral mRNA and protein expression.
To determine whether PMLIV affects the expression of IFNs and/or pro-inflammatory cytokines, we analyzed IFN-α, IFN-β, TNF-α and IL8 mRNAs by RT-qPCR in extracts from U373MG-EV and U373MG-PMLIV cells infected with VSV at an MOI of 0.2 for various lengths of time. PMLIV did not significantly alter the mRNA expression of IFN-α, TNF-α or IL8 following VSV infection (Figure 5A). Thus, PMLIV had no effect on the induction of TNF-α and IL-8 mRNAs that are regulated by NF-κB, or on the induction of IFN-α mRNA that is regulated by IRF7 [33]. In contrast, as early as 8 h post-infection, IFN-β mRNA expression that is known to occur through activated IRF3, was enhanced by PMLIV leading to as high as a 2-log increase 12 h post-infection. Like PMLIV, PMLIVa increased IFN-β mRNA synthesis upon VSV infection (data not shown). Thus, both PMLIV and its variant PMLIVa sharing the same C-terminal region unique to this isoform, boosted IFN-β induction upon VSV infection. Next, we asked whether PMLIV-dependent boost of IFN-β transcription was specific to VSV infection or whether it was a broader mechanism triggered by pattern recognition receptor (PRR) activation. To do this, U373MG-EV or U373MG-PMLIV cells were either infected with SeV (40HAU/ml) or transfected with poly(I:C) (1 µg/ml). After 8 h, mRNAs were extracted and IFN-βtranscripts quantified by RT-qPCR. As shown in Figure 5B, we found that PMLIV also enhanced IFN-β mRNA synthesis in U373MG cells infected with SeV, EMCV or transfected with poly(I:C) (Figure 5B). Also, HeLa cells transduced with PMLIV-expressing or with noncoding parental (EV) lentiviral vector and infected with HTLV-1, influenza or vaccinia virus revealed that PMLIV positively regulated IFN-β mRNA production (Figure 5B). This demonstrates that the ability of PMLIV to potentiate IFN-β synthesis is not a specific feature of VSV infection but a more general mechanism following RNA or DNA detection by PRRs.
We investigated the role of SUMOylation in the antiviral property of PMLIV by using PMLIV-3KR mutant in which the three major SUMO-target lysines were substituted with arginines. Double immunofluorescence analysis of PML and viral proteins revealed that PMLIV inhibited viral protein synthesis, whereas PMLIV-3KR did not (Figure S2A, left panel). PML and its SUMOylated forms were indeed produced in cells stably expressing PMLIV but, as expected, only the unmodified form was detected in cells stably expressing the SUMO-deficient PML mutant, PMLIV-3KR (Figure S2A, right panel). Western blot analysis of extracts from infected U373MG-EV, U373MG-PMLIV and U373MG-PMLIV-3KR cells confirmed that VSV protein synthesis was inhibited by PMLIV but was not affected by PMLIV-3KR (Figure S2B, left panel). In addition, VSV growth was inhibited in cells expressing PMLIV, but not in cells expressing PMLIV-3KR (Figure S2B, right panel). To determine the role of PMLIV SUMOylation on IFN-β synthesis, we quantified by RT-qPCR IFN-β mRNA in extracts from U373MG-EV, U373MG-PMLIV and U373MG-PMLIV-3KR cells infected with VSV (Figure 5C). As PMLIII also inhibits VSV multiplication, we also quantified IFN-β mRNA in extracts from U373MG-PMLIII and U373MG-PMLIII-3KR cells. Again, we observed a nearly 2 log increase of IFN-β mRNA expression in VSV infected cells expressing PMLIV. Strikingly, the induction was completely lost when the PMLIV-3KR was expressed, demonstrating that SUMOylation of PMLIV is also required for the enhancement of IFN-β mRNA expression. In contrast, following VSV infection, PMLIII did not increase the mRNA synthesis of IFN-β, IFN-α, TNF-α or IL8 (Figure 5C and data not shown). Taken together, these results demonstrate that SUMO modification of PMLIV is required to confer resistance towards VSV infection and also to increase IFN-β synthesis.
To determine the effect of other PML isoforms on IFN-β synthesis, U373MG-EV cells and cells stably expressing each PML isoform (PMLI, PMLII, PMLIII, PMLIV, PMLV, PMLVI or PMLVIIb) were infected with VSV at an MOI of 0.2 and IFN-β mRNA was quantified in their extracts by RT-qPCR (Figure 6A). Among all PML isoforms, only PMLIV expression resulted to a dramatic increase of IFN-β mRNA synthesis following VSV infection, reaching a nearly 2 log rise. To rule out the possibility that this increase could be due to the stabilization of IFN-β encoding mRNA by PMLIV, we performed a transcriptional assay. To do this, we transfected control or PMLIV-expressing cells with a reporter plasmid containing the firefly luciferase gene under the control of the human IFN-β promoter and infected the cells with VSV. Using this reporter assay, we were able to confirm that PMLIV greatly enhances the transcription driven by the IFN-β promoter (data not shown), confirming that this increase is at the transcriptional level.
Among members of the IFN regulatory factor family, IRF3 plays an essential role in virus-induced IFN-β gene expression [34]. Interestingly, IFN produced by PMLIV-expressing cells upon VSV infection was due to activated IRF3, since a higher amount of phosphorylated IRF3 (P-IRF3) was detected 6 h post-infection of cells expressing PMLIV compared to control cells (Figure 6B). Interestingly, PMLIV also enhanced IRF3 phosphorylation in cells transfected with poly(I:C) (Figure 6B). In contrast, the level of IRF3 was not altered following VSV infection or poly(I:C) transfection. To further determine the role of IRF3 in anti-VSV activity and enhanced IFN synthesis, U373MG-EV and U373MG-PMLIV cells depleted for IRF3 were infected at an MOI of 1 for 8 h and their extracts were analyzed by RT-qPCR for IFN-β mRNA and by Western blot for IRF3 and viral protein expression. As seen in Figure 6C, in PMLIV expressing cells infected for 8 h, depletion of IRF3 abrogated IFN-β mRNA synthesis (Figure 6C, left panel) without affecting the inhibition of VSV protein synthesis (Figure 6C, right panel) and viral production (data not shown). This demonstrates that the intrinsic anti-VSV activity of PMLIV is independent of IRF3.
Thus, taken together, our results demonstrate that PMLIV is the only PML isoform able to inhibit VSV at a high MOI independently of IRF3 and also to stimulate IFN-β synthesis via an increase of IRF3 activation.
PMLIV conferred viral resistance in cells 8 h or 12 h post-infection (Figure 3). This resistance was correlated with an induction of IFN-β mRNA synthesis observed as early as 8 h and increased as high as 2 log at 12 h post-infection (Figure 5). To determine whether or not the observed antiviral effect of PML at these times of infection was a secondary response to IFN synthesis, we tested the capacity of PMLIV to inhibit VSV in cells treated with an anti-IFNAR1 mAb targeting the extracellular domain of the IFNAR1 chain of the human IFN-α/β receptor. The anti-IFNAR1 mAb inhibits the binding and biological activity of type I IFN [35] as well as the IFN-β-induced STAT1 expression and anti-VSV activity (Figure 7A, left panel). In contrast, this antibody did not alter the inhibition of VSV protein synthesis by PMLIV in cells infected for either 8 h or 12 h (Figure 7A, right panel). Since STAT1 is the central transcription factor required for the biological responses of all types of IFN, we determined the effect of its downregulation on PMLIV-induced VSV resistance (Figure S2C). The capacity of PMLIV to inhibit VSV protein synthesis was still maintained in cells depleted for STAT1 (Figure S2C), further demonstrating the intrinsic anti-VSV effect of PMLIV at 12 h post-infection. Collectively, these results demonstrate that PMLIV exerts an early intrinsic anti-VSV activity that is independent of IFN.
To determine whether the IFN produced and secreted after a longer period of infection was active, culture supernatants from U373MG-EV cells (SEV) and U373MG-PMLIV cells (SPMLIV) infected with VSV for 20 h were tested, in comparison with IFN-β, for their capacities to induce ISG products and to inhibit viral protein synthesis. Therefore, HeLa cells treated with medium (Ctrl), IFN-β, SEV or SPMLIV supernatant for 24 h were uninfected (Figure 7B, left panel) or infected with VSV for 8 h (Figure 7B, right panel). As seen in Figure 7B (left panel), STAT1 and PKR expression was increased only in cells treated with IFN-β and SPMLIV supernatant. In addition, VSV proteins were inhibited only in extracts from cells pretreated with IFN-β or SPMLIV supernatant (Figure 7B, right panel). It should be noted that the capacity of SPMLIV supernatant to induce STAT1 and PKR expression as well as to inhibit VSV was higher than that observed with 100 units/ml of IFN-β. In addition, SPMLIV but not SEV supernatant was able to induce STAT1 phosphorylation in HeLa cells (Figure 7C, left panel). Taken together, these results show that the IFN produced and secreted by U373MG-PMLIV cells 20 h post-VSV infection, activates STAT1, induces ISG products and protects HeLa cells from viral infection.
Next, we quantified the amount of type I IFN in supernatants from VSV-infected U373MG-EV and U373MG-PMLIV cells using the human HL116 cell line carrying the luciferase gene under control of the IFN-inducible 6–16 promoter. This experiment showed that PMLIV boosted the amount of type I IFN synthesized in infected cells by up to 200 to 300 international units/ml (IU/ml), depending on the experiment. A typical experiment is presented in Figure 7C (right panel). In addition, SEV and SPMLIV supernatants were also titrated on HeLa cells. The IFN titer of SEV was below the detection limit (less than 2 IU/ml) and that of SPMLIV was 250 IU/ml.
Taken together, these results demonstrate that PMLIV has a dual effect on viral infection: (i) an early intrinsic anti-VSV activity that was not eradicated by treatment with anti-IFNAR1 mAb, knockdown of STAT1 or IRF3 and (ii) an activation of innate immune signaling that occurs later and leads to the production and the secretion of type I IFN, which can protect other cells from viral infection.
Pin1 is known to interact with and to promote phosphorylated IRF3 degradation [30]. Since PMLIV increased IRF3 activation upon VSV infection, we asked whether PMLIV can specifically recruit endogenous Pin1 within PML NBs. Double immunofluorescence studies were performed on endogenous Pin1 and PML in cells expressing PMLIII, PMLIV or PMLIV-3KR. In both uninfected and infected cells endogenous Pin1 was found both in the cytoplasm and the nucleus (Figure 8A and data not shown). Importantly, Pin1 was found colocalizing with PMLIV within the NBs in uninfected or VSV-infected cells (Figure 8A and data not shown). Indeed, Pin1 was found diffuse in the nucleus of EV, PMLIII and PMLIV3KR cells, whereas in PMLIV-expressing cells, it formed speckles colocalizing with PML NBs. Interestingly, such colocalization was not observed in cells expressing PMLIII or PMLIV-3KR (Figure 8A, left panel). Fluorescence intensities were quantified using Image-J software and revealed that the portion of Pin1 associated to PML NBs highly increased in PMLIV-expressing cells (Figure 8A, right panel). Thus, PMLIV induced a relocalization of Pin1 from the nucleoplasm to the NBs.
The recruitment of Pin1 to PML NBs by PMLIV was further demonstrated by Western blot analysis of the RIPA soluble and insoluble fractions (Figure 8B). In EV cells, most of the Pin1 was found in the RIPA-soluble fraction that included both the cytoplasm and the nucleoplasm, whereas a small fraction was associated to the nuclear matrix (RIPA-insoluble fraction). As observed by immunofluorescence, the expression of PMLIV resulted in a shift of Pin1 to the nuclear matrix, resulting in an enrichment of Pin1 in the RIPA insoluble fraction (Figure 8B). Co-immunoprecipitation assays revealed that PMLIV and PMLIV-3KR interacted with endogenous Pin1 whereas a very slight interaction was detected with PMLIII (Figure 8C). The recruitment of Pin1 within PML NBs was observed in various human cell lines including HeLa cells transduced with a lentiviral vector expressing PMLIV (Figure 9A, and data not shown). The recruitment of Pin1 within PML NBs by PMLIV in HeLa cells was also associated with a positive regulation of both IRF3 activation (Figure 9B and data not shown) and IFN-β synthesis upon VSV or SeV infection (Figure 9C). Thus, PMLIV expression in different infected human cells resulted in enhanced IRF3 phosphorylation and IFN-β mRNA production (Figures 5, 6 and 9). At the opposite, the expression of PMLIV in MEF cells neither induced the recruitment of endogenous mouse Pin1 within NBs (Figure 9A) nor enhanced IRF3 activation or IFN-β synthesis upon viral infection (Figure 9C and data not shown). Our results suggest that the recruitment of Pin1 by PMLIV within PML NBs is required for PMLIV-induced enhancement of IFN-β synthesis.
Next, we asked whether the intrinsic anti-VSV activity of PMLIV was observed in PML-/-MEF cells. To do this, PML-/- MEFs were transduced with EV- or PMLIV-encoding lentivector before infection with VSV. As seen in Figure 9D, PMLIV was still able to inhibit VSV protein synthesis when expressed in PML-/- cells, thus revealing that the intrinsic anti-VSV activity of PMLIV did not require the expression of endogenous PML.
To further confirm the specific positive regulation of PMLIV on IFN-β transcription, U373MG cells were infected with SeV 48 h after transfection with scramble siRNA (Sc), siRNA targeting IRF3, siRNA common to all PML isoforms (siRNA PMLc), or with the siRNA specific to PMLIII or to PMLIV that we have previously validated [23]. As expected, the siRNA IRF3 completely abolished IFN-β mRNA synthesis upon SeV infection (Figure 9E). SeV-induced IFN-β expression was not altered by PMLIII depletion but was highly reduced following the depletion of all PML isoforms or the specific depletion of PMLIV (Figure 9E). Furthermore, specific suppression of PMLIV expression reduced SeV-induced IRF3 activation (Figure 9F). This demonstrates that endogenous PMLIV is required for the efficient synthesis of IFN-β transcription upon viral infection and validates our data obtained in cells overexpressing PMLIV.
Taken together, these results show that PMLIV and PMLIV-3KR interacted with Pin1 but only PMLIV was able to recruit it within PML NBs where both proteins colocalized. Therefore, the interaction of endogenous Pin1 with SUMOylated PMLIV and its recruitment in PML NBs could alter Pin1-induced downregulation of activated IRF3 thus resulting in a higher amount of phosphorylated IRF3 during viral infection. In addition, the intrinsic anti-VSV activity of PMLIV is independent of the expression of other PML isoforms.
Many reports implicate PML and PML NBs in antiviral responses targeting diverse cytoplasmic replicating RNA viruses through different mechanisms [13], [17], [19], [21], [23]–[25], [36]. An antiviral effect of PML against rhabdoviridae has been observed in vivo, as PML deficiency renders mice more susceptible to VSV infection [14]. In this report, we show that cells derived from these mice or human cells depleted of PML produced a higher level of viral proteins. Among the various PML isoforms tested, only stable expression of PMLIII and PMLIV conferred resistance to VSV. This inhibitory effect did not alter VSV entry, but was observed at the level of viral mRNA and protein synthesis, resulting in a reduction of VSV yields and in cell lysis protection. The protective property of PMLIV was found to be higher than that of PMLIII, as PMLIV was able to inhibit virus growth up to an MOI of 2, resulting in a 500-fold reduction of VSV yields. Whereas PMLIII confers viral resistance in an IFN-independent way, PMLIV displays two antiviral activities during VSV infection: an early IFN-independent activity targeting VSV replication followed by the activation of innate immunity pathways, leading to an enhanced type I IFN synthesis, which protects yet uninfected cells from viral infection (Figure 10). Interestingly, PMLIV-3KR failed to confer resistance to VSV and also to induce IFN-β synthesis, demonstrating that SUMOylation of PMLIV is required for both intrinsic antiviral activity and innate immune property.
The induction of IFN-β expression is the key event in the initiation of the innate antiviral response. Central to this process is the activation of IRF3 via its phosphorylation [34]. Here we demonstrate for the first time that one particular isoform of PML (PMLIV) is implicated in innate immunity, triggering a dramatic increase of IFN-β synthesis via IRF3 phosphorylation upon VSV infection. Depletion of IRF3 further demonstrates the dual activity of PMLIV, as it abrogated PMLIV-induced IFN synthesis but not PMLIV-induced inhibition of viral replication.
Pin1 was shown to interact with phosphorylated IRF3 and to promote its ubiquitin-mediated proteasomal degradation [30]. We report here that endogenous Pin1 interacts with PMLIV and that both proteins colocalize in PML NBs. This results in sustained IRF3 activation and higher IFN-β induction during VSV infection. Thus, the recruitment of endogenous Pin1 in PML NBs might antagonize Pin1-induced P-IRF3 degradation [30] and could be a novel mechanism for inhibiting Pin1 function. Consistently, host antiviral responses are boosted in the presence of PMLIV. Collectively, our report characterizes PMLIV as a positive regulator of antiviral innate immune responses, which maintains stability of IRF3 phosphorylation through the interaction of endogenous Pin1 with SUMOylated PMLIV and its recruitment in PML NBs.
Thus, our results suggest that, whereas the direct antiviral activity of PMLIV is specifically targeting VSV, the positive regulation of innate immunity should also be observed using other stimuli than VSV infection since PMLIV also recruited Pin1 within the NBs in uninfected cells. Indeed, we observed that IRF3 activation and IFN-β production were also drastically enhanced in PMLIV-overexpressing cells following transfection with poly(I:C) or infection with viruses from different families such as SeV, EMCV, HTLV-1, influenza virus or vaccinia virus. These observations confirm that PMLIV has a specific anti-VSV activity and a much wider positive effect on innate immunity pathways. Remarkably, specific depletion of endogenous PMLIV in human U373MG or HeLa cells reduced VSV- as well as SeV-induced activation of IRF-3 and consequent production of IFN-β (this paper and data not shown).
This is the first demonstration of the implication of PML in the enhancement of IFN-β production upon viral infection. In contrast, the intrinsic antiviral activity of PML has been documented to act on viruses from different families [17], [21]. In the case of EMCV, PMLIV is the only PML isoform that confers viral resistance, by interacting with the viral polymerase 3D, and sequestering it within PML NBs where both proteins colocalize [23]. Interestingly, PMLIV is also able to sequester the ORF23 capsid protein within PML NBs, leading to the inhibition of varicella-zoster virus (VZV) an alphaherpesvirus [37]. PML has also been shown to interfere with retrovirus replication, since PMLIII interacts with Tas, the transcriptional transactivator of human foamy virus, resulting in viral restriction [24]. The mechanism by which the nuclear PMLIV directly inhibits VSV, whose replication takes place entirely in the cytoplasm, remains to be elucidated. We demonstrated that the intrinsic anti-VSV property of PMLIV did not require the expression of other PML isoforms and was independent of IFN since it was maintained in cells treated with antibodies against type I IFN receptors, depleted for IRF3 or depleted for STAT1. However, further investigations are needed to demonstrate how PMLIV and PMLIII exert their intrinsic anti-VSV activity by interacting with a viral or a cellular protein required for VSV replication. Understanding the intrinsic antiviral activity of PML may introduce new ways for targeted antiviral therapy that would bypass the need for IFN treatment.
Thus, PML confers viral resistance in two ways (Figure 10). It can exert an intrinsic anti-VSV activity independent of IRF3, STAT1 and IFN. In addition, we show here that PML enhanced both IRF3 activation and IFN-β synthesis upon viral infection. This produced IFN-β is secreted and protects yet uninfected cells from oncoming infection.
Our results demonstrate that PML, an ISG product with a broad intrinsic antiviral activity, is also able to trigger IFN-β synthesis upon viral infection. There are an increasing number of ISG products that are also implicated in innate immunity processes. PKR for example, long known to mediate the antiviral activities of IFNs, also plays an important role in the induction of type I IFN, particularly IFN-β during measles virus infection [38] or double-stranded RNA treatment [39]. Indeed, activation of PKR increases IRF3 activation, and knockdown of PKR reduces both activated IRF3 level and IFN-β induction [39]. The role of the endoribonuclease RNase L in the innate antiviral immune response has been demonstrated in vivo. Indeed, injection of 2′5′-linked oligoadenylates leads to IFN-β synthesis in wild-type mice but not in RNase L deficient mice. In addition, EMCV- or Sendai virus-induced IFN is highly reduced in mice lacking RNase L [40].
Interestingly, many members of the TRIM protein family, which PML belongs to, (i) are products of ISGs [41], (ii) display a direct antiviral activity [42] and/or have been identified as important players of innate immunity [43]. TRIM5α for instance is induced by type I IFN [44], [45], inhibits retroviral infections [46] and is also able to promote innate immune signaling [47]. TRIM25 protein plays a key role in innate immunity, since it is essential for RIG-I-mediated antiviral activity [48]. Strikingly, TRIM21 is also able to enhance IRF3-mediated antiviral response. Indeed, TRIM21 was shown to interact directly with IRF3 upon viral infection and to interfere with its interaction with Pin1 [49]. Recent studies performed on the entire TRIM protein family allowed the identification of several other members of this family as key components of inflammation and innate immunity signaling pathways [50], [51]. Our results show that TRIM19/PML, another member of the TRIM protein family, acts, through Pin1 recruitment in PML NBs, as a positive regulator of IRF3 phosphorylation, enhancing the strength and duration of IFN-β-induced antiviral response. Thus, PML can therefore be added to the list of TRIM proteins implicated in both intrinsic and innate immunity. Initially considered as two independent arms of the immune system, our results further suggest a closer crosstalk between intrinsic and innate immunity.
Human recombinant IFN-β was purchased from Biogen Inc. Rabbit polyclonal (sc-5621) and mouse anti-PML (sc-966), rabbit anti-IRF3 (sc-9082), rabbit anti-STAT1 (sc-345), rabbit anti-phospho-STAT1 (Tyr701, sc-7988) and rabbit anti-PKR antibodies (sc-707) were obtained from Santa-Cruz Biotechnology. The rabbit anti-phospho-IRF3 (Ser 396) and rabbit anti-Pin1 antibodies were obtained from Cell Signaling, and HRP-conjugate monoclonal anti-Actin antibody from Sigma. The 64G12 monoclonal antibody against human IFN-α/β receptor (anti-IFNAR1 mAb) was a gift from P Eid (INSERM UMR1014) [35]. The rabbit anti-VSV polyclonal antibodies (home-made) were obtained by repeated injection of purified virus. Reactivities against the viral proteins N, M or G were different depending on the batch used in western-blot experiments.
Human glioblastoma astrocytoma U373MG, epithelial HeLa and fibrosarcoma HL116 cells as well as mouse embryonic fibroblasts (MEFs) from wild-type (WT) or knockout PML (PML-/-) mice [52], were grown at 37°C in DMEM supplemented with 10% FCS. U373MG cells transfected with empty vector or stably expressing individual PML isoforms (PMLI to VIIb), PMLIII-3KR or PMLIV-3KR were kept in medium supplemented with 0.5 mg/ml of neomycin. HL116 cells were grown in medium supplemented with HAT (Hypoxanthine: 20 µg/ml, Aminopterin: 0.2 µg/ml, Thymidine: 20 µg/ml).
The accession number (GenBank) for PML isoforms are AF230401 (PMLI), AF230403 (PMLII), S50913 (PMLIII), AF230406 (PMLIV), AF230411 (PMLIVa), AF230402 (PMLV), AF230405 (PMLVI), AF230408 (PMLVIIb). In PMLIII-3KR and PMLIV-3KR mutants, the three SUMO-target lysines (at positions 65, 160, and 490) were replaced with arginines. Stable U373MG cells expressing each of the PML isoforms (PMLI to VIIb), PMLIII-3KR or PMLIV-3KR were obtained via transfection with constructs corresponding to each cloned in pcDNA3.1 and subsequent neomycin selection at a final concentration of 0.5 mg/ml [25]. Control U373MG cells were generated in the same way using the empty vector (U373MG-EV). For expressing PMLIV in MEFs wild-type, MEFs PML-/- and HeLa cells, we constructed an HIV-derived lentiviral vector (pTRIP-PMLIV), which was used to transduce the cells. The pTRIP plasmid was provided by P Charneau (Institut Pasteur, Paris, France) [53].
VSV (Mudd-Summer strain, Indiana serotype) was grown in BSR cells. BSR cells were infected at an MOI of 0.1. After 24 h, supernatants were collected and cellular debris removed by low-speed centrifugation. Virus titers (109 PFU/ml) were determined by standard plaque assay onto BSR cells. EMCV was produced as described [20] and has a titer value of 2.108 PFU/ml. Vaccinia virus has a titer of 2.109 PFU/ml. Sendai virus was kindly provided by E Meurs (Institut Pasteur, Paris, France). It was used at 40 HAU/ml to activate RIG-I, as described [54]. HTLV-1, and influenza A virus (strain A/PR 8/34) were provided by JP Herbeuval (CNRS UMR 8601). HTLV-1 and influenza virus were used at 600 ng/ml of p19-equivalent and at 40 HAU/ml, respectively. MLV-derived vectors encoding GFP pseudotyped with VSV-G, provided by FL Cosset (ENS Lyon), had a titer value of 107 IU/ml and were generated as described [31].
Cells were seeded in six-well plates and transfected with siRNA using Lipofectamine RNAiMax transfection reagents (Invitrogen). The mRNA sequence targeted by the siRNA PMLc (common to all PML isoforms) is 5′-AUGGCUUCGACGAGUUCAATT-3′, by siRNA PMLIII is 5′-AGUGCAUGGAGCCCAUGGATT-3′ and by siRNA PMLIV is 5′UGAAAGUGGGUUCUCCUGGTT-3′. The siRNA scramble sequence is the following: 5′-GCAUGAACCGGAGGCCCAUUU-3. STAT1 or IRF3 expression was silenced using ON-TARGETplus SMARTpool siRNAs purchased from Thermo Scientific.
Cells were infected with VSV at an MOI of 3 in the absence or the presence of cycloheximide (100 µg/ml). After adsorption for 1 h, cells were washed and fresh medium with or without cycloheximide (100 µg/ml) was added. After 4 h, total RNA was isolated from cells with the RNA NOW Kit (Ozyme). Total RNA was separated on 1.5% agarose gel under denaturing conditions and blotted onto Nylon membranes (Roche Molecular Biochemicals). Hybridizations were performed with digoxigenin (DIG)-labeled oligonucleotides recognizing the VSV-N gene sequence and by incubation with anti-DIG antibody conjugated to alkaline phosphatase followed by CDP Star.
Total RNA was extracted using RNeasy Mini Kit (Qiagen) and cDNAs were prepared using Oligo(dT) primer and SuperScript II Reverse Transcriptase (Invitrogen). Real-time PCR reactions were performed in duplicates using Platinum SYBR Green qPCR SuperMix-UDG (Invitrogen) following manufacturer's instructions. GAPDH, IFN-α1, IFN-β, TNF-α, IL8 and VSV-L encoding cDNAs were amplified on a Mastercycler ep realplex (Eppendorf) with a denaturation step of 5 min at 95°C followed by thirty-five cycles of 10 s at 95°C, 10 s at 60°C and 20 s at 72°C. Threshold cycle (Ct) values were converted to 2−Ct in order to be proportional to the amount of transcripts in the samples. To compare samples, 2−ΔCt were calculated by normalizing the data by the expression of GAPDH: 2−ΔCt = 2−Ct(sample)/2−Ct(GAPDH). Primers used for quantification of transcripts by real time quantitative PCR are the following: VSV L (Forward: TGATACAGTACAATTATTTTGGGAC and Reverse: GAGACTTTCTGTTACGGGATCTGG), GAPDH (Forward: ACTTCAACAGCGACACCCACT and Reverse: GTGGTCCAGGGGTCTTACTCC), IFN-α1 (Forward: CCAGTTCCAGAAGGCTCCAG and Reverse: TCCTCCTGCATCACACAGGC), IFN-β (Forward: TGCATTACCTGAAGGCCAAGG and Reverse: AGCAATTGTCCAGTCCCAGTG), TNF-α (Forward: GGCGTGGAGCTGAGAGATAAC, Reverse: GGTGTGGGTGAGGAGCACAT) and IL-8 (Forward: AAGGGCCAAGAGAATATCCGAA and Reverse: ACTAGGGTTGCCAGATTTAACA).
Immunofluorescence analyses were performed as described [23]. PML was detected with mouse anti-PML antibody and the corresponding anti-IgG antibody conjugated to Alexa 594. The VSV antigens or Pin1 protein were detected with rabbit anti-VSV or anti-Pin1 antibodies followed by Alexa 488. The cells were mounted onto glass slides by using Immu-Mount (Shandon) containing 4,6-diamidino-2-phenylindole (DAPI) to stain nuclei. Confocal laser microscopy was performed on a Zeiss LSM 510 microscope.
Quantitative analysis of immunofluorescence data was carried out by histogram analysis of the fluorescence intensity at each pixel across the images using Image J software. The results of the analysis of 20 images acquired in each experimental condition were then combined to allow quantitative estimates of changes in Pin1 localization.
Cells were washed and re-suspended in PBS, lysed in hot Laemmli sample buffer and boiled for 5 min. For cell fractionation, cells were dissociated and washed twice in PBS. The cytoplasmic and nucleoplasm fractions were extracted by incubating the cell pellet in 50 µl of RIPA buffer for 20 min on ice followed by centrifugation at 15000 g for 15 min to separate the RIPA soluble fraction (R) from the pellet (P). This RIPA insoluble fraction (P) was suspended in 50 µl of PBS. Fifty µl of 2X Laemmli buffer were added to each fraction, and the samples were boiled for 5 min before Western blot analysis. Protein extracts were analyzed on a 10% SDS-PAGE gel, and transferred onto a nitrocellulose membrane. The proteins were blocked on the membranes with 5% skimmed milk in PBS for 2 h and incubated overnight at 4°C with rabbit polyclonal anti-PML (clone H-238), anti-VSV, anti-IRF3, anti-phospho-IRF3 or anti-Actin antibodies. The blots were then washed extensively in PBS-Tween and incubated for 1 h with the appropriate peroxidase-coupled secondary antibodies (Cell Signaling Technology). All of the blots were revealed by chemiluminescence (ECL, Bio-Rad).
Cells (107) were incubated for 30 min at 4°C in 0. 5 ml of buffer containing 20 mM Tris-HCl pH 7.4, 1 M NaCl, 5 mM MgCl2, 1% triton, and 1 mM phenylmethylsulfonyl fluoride (PMSF). After cell lysis, 1.25 ml of immuno-precipitation buffer (IB) (20 mM Tris-HCl pH 7.4, 150 mM NaCl, 0.5% DOC, 1% Triton X-100, 0.1% SDS and 1 mM EDTA) were added. Rabbit anti-Pin1 antibodies were added and the samples incubated overnight at 4°C. Protein G beads (Sigma) in IB were then added and the samples mixed 2 h at room temperature. The beads were collected, washed four times with IB buffer and bound proteins were subjected to Western blotting.
To inactivate the virus, culture supernatants from infected cells were brought to pH 2 for 24 h and neutralized before titration. IFN secretion was quantified using the reporter cell line HL116 that carries the luciferase gene under the control of the IFN-inducible 6–16 promoter [55]. HL116 cells (2×104) were plated in 96-well plate and incubated for 8 h with the desired culture supernatants or a standard of human IFN-β reference (Gb-23-902-531). Cells were then lysed (Luciferase Cell Culture Lysis Reagent, Promega) and luciferase activity measured using Perkin Elmer Wallac 1420. IFN titers were also quantified on HeLa cells challenged with VSV. IFN titers, defined as the amount of IFN leading to 50% inhibition of the cytopathic effect, were expressed in international unit/ml relative to the human IFN-β reference (Gb-23-902-531) of the NIH.
PMLI (AF230401), PMLII (AF230403), PMLIII (S50913), PMLIV (AF230406), PMLIVa (AF230411), PMLV (AF230402), PMLVI (AF230405), PMLVIIb (AF230408), STAT1
|
10.1371/journal.ppat.1003665 | Toxoplasma gondii Relies on Both Host and Parasite Isoprenoids and Can Be Rendered Sensitive to Atorvastatin | Intracellular pathogens have complex metabolic interactions with their host cells to ensure a steady supply of energy and anabolic building blocks for rapid growth. Here we use the obligate intracellular parasite Toxoplasma gondii to probe this interaction for isoprenoids, abundant lipidic compounds essential to many cellular processes including signaling, trafficking, energy metabolism, and protein translation. Synthesis of precursors for isoprenoids in Apicomplexa occurs in the apicoplast and is essential. To synthesize longer isoprenoids from these precursors, T. gondii expresses a bifunctional farnesyl diphosphate/geranylgeranyl diphosphate synthase (TgFPPS). In this work we construct and characterize T. gondii null mutants for this enzyme. Surprisingly, these mutants have only a mild growth phenotype and an isoprenoid composition similar to wild type parasites. However, when extracellular, the loss of the enzyme becomes phenotypically apparent. This strongly suggests that intracellular parasite salvage FPP and/or geranylgeranyl diphosphate (GGPP) from the host. We test this hypothesis using inhibitors of host cell isoprenoid synthesis. Mammals use the mevalonate pathway, which is susceptible to statins. We document strong synergy between statin treatment and pharmacological or genetic interference with the parasite isoprenoid pathway. Mice can be cured with atorvastatin (Lipitor) from a lethal infection with the TgFPPs mutant. We propose a double-hit strategy combining inhibitors of host and parasite pathways as a novel therapeutic approach against Apicomplexan parasites.
| Toxoplasma gondii is an obligate intracellular parasite and is not able to replicate outside the host cell. The parasite lives in a specialized parasitophorous vacuole in contact with the host cytoplasm through the parasitophorous vacuole membrane. It is highly likely that a very active exchange of metabolites occurs between parasite and host cell. We present evidence for this exchange for isoprenoids, abundant lipidic compounds essential to many cellular processes including signaling, trafficking, energy metabolism, and protein translation. Our work shows that intracellular T. gondii tachyzoites are able to salvage farnesyl diphosphate (FPP) and/or geranylgeranyl diphosphate (GGPP) from the host, and the parasite is able to grow even when its endogenous production is shut down. However, when extracellular, the parasite depends entirely on its own production of isoprenoids. We propose to use a combination of inhibitors that would hit both the host and the parasite pathways as a novel therapeutic approach against Toxoplasma gondii that could also work against other Apicomplexan parasites.
| Toxoplasma gondii is an important intracellular pathogen causing disease in humans and animals. Most human infections are uncomplicated but the parasite persists and the chronic infection can be reactivated upon immunosuppression in patients undergoing organ transplants, cancer chemotherapy [1], or AIDS due to HIV infection [2]. During pregnancy, infection causes congenital toxoplasmosis with serious consequences to the fetus [3]. There is also growing concern about outbreaks of severe ocular disease due to T. gondii in immunocompetent patients [4]. The parasite masterfully manipulates its host cell to insure favorable conditions for its survival and replication. T. gondii infection results in differential regulation of a variety of host signaling and metabolic pathways [5]. Many of these host changes are still not completely understood but it is quite likely that such modification of host pathways is essential for parasite growth and survival.
Isoprenoids are lipid compounds with many important functions. The enzymes that synthesize and use isoprenoids are among the most important drug targets for the treatment of cardiovascular disease, osteoporosis and bone metastases and have shown promise as antimicrobials in a number of systems [6]. T. gondii lacks the mevalonate pathway for the synthesis of isoprenoid precursors that is used by mammals but harbors a prokaryotic-type 1-deoxy-D-xylulose-5-phosphate (DOXP) pathway in the apicoplast. This pathway generates isopentenyl diphosphate (IPP) and dimethyallyl diphosphate (DMAPP). We recently demonstrated that the DOXP pathway is essential in T. gondii [7]. Knockout of 1-hydroxy-2-methyl-2-(E)-butenyl 4-diphosphate reductase (LytB), which catalyzes the generation of IPP and DMAPP in the final step of the DOXP pathway, or of DOXP reductoisomerase (DOXPRI), which catalyzes the second step of the DOXP pathway, were both lethal [7]. We also characterized the key enzyme of downstream isoprenoid synthesis in T. gondii, farnesyl diphosphate synthase (TgFPPS) [8]. Interestingly, we found it to be a bifunctional enzyme that can catalyze the condensation of IPP with three allylic substrates: DMAPP, geranyl diphosphate (GPP), and farnesyl disphosphate (FPP). The enzyme thus generates not only 15-carbon FPP but also 20-carbon GGPP [8]. A bifunctional FPPS has also been described in Plasmodium falciparum [9]. TgFPPS is inhibited by long alkyl chain (lipophilic) bisphosphonates, which are among the most active inhibitors of human GGPPS [10], as well as by short chain bisphosphonates like risedronate (aminobisphosphonates), which preferentially inhibit human FPPS. T. gondii engineered to overexpress TgFPPS requires considerably higher levels of bisphosphonates to achieve growth inhibition supporting the idea that the T. gondii enzyme is a target of bisphosphonates [11].
In this work we report that drugs acting on the mevalonate pathway, like statins, are active in vitro and in vivo against T. gondii. This is surprising as the parasite lacks this pathway. With the use of null mutants for the TgFPPS (Δfpps) we demonstrate why the parasite is sensitive to these inhibitors. We also show that the parasite is able to salvage some isoprenoid intermediates from the host while depending on its own synthetic machinery for others. Our results reveal a metabolic exchange between host and parasite that quite likely also occurs in other intracellular pathogens like Plasmodium or Cryptosporidium. To take advantage of these findings we propose a double-hit strategy combining inhibitors of both host (statins) and parasite (bisphosphonates) pathways. This strategy will allow leveraging the extensive clinical experience gained with statins towards the treatment of infections and potentially adapt it to other intracellular parasites.
FPPS is an essential component of the isoprenoid biosynthesis pathway in all cells studied so far. This enzyme synthesizes both FPP and GGPP in T. gondii and localizes to the mitochondria [8]. Previous work from our laboratory has shown that the T. gondii FPPS is inhibited by bisphosphonates, which also inhibit parasite growth. Considering the central role of this enzyme in the isoprenoid pathway we wanted to validate the entire pathway as potential target for chemotherapy. We approached this by creating a null mutant for the TgFPPS gene. We used the T. gondii Δku80 strain, which favors homologous recombination [12], [13]. Our targeting construct was a large genomic cosmid recombineered to replace the gene with a drug resistance marker (Fig. 1A) [14]. After initial unsuccessful attempts, we were able to obtain null mutants when supplementing the medium with geranylgeraniol during the selection process. This requirement for geranylgeraniol for growth of mutant parasites is possibly because of their specific metabolite need during the stress of the transfection. We analyzed these mutant clones (Δfpps) by Southern (Fig. 1B) and western (Fig. 1C) blot and demonstrated the lack of both TgFPPS gene and protein. We isolated complemented clones by re-introducing the TgFPPS gene into the T. gondii genome (Δfpps-cm1 and Δfpps-cm2; Figs. 1B, and 1C). We next introduced tandem tomato red fluorescent protein constructs into all strains (Δku80-rfp, Δfpps-rfp, Δfpps-cm-rfp) to measure parasite growth following the intensity of red fluorescence [15]. To our surprise, Δfpps mutant parasites were able to grow at a similar rate in fibroblasts (the cells we routinely use for parasite culture, Fig. 1D), and formed plaques of similar number and size when compared to the parental and complemented strains (Figure S1, top row).
Previous work has shown that T. gondii can enter macrophages by active invasion. However weakened or stressed parasites can be actively phagocytized by macrophages, resulting in parasite death making macrophages a more challenging host cell than fibroblasts [16]. We tested our mutants for their ability to grow in macrophages. Interestingly, Δfpps parasites showed a significant growth defect in these cells (Fig. 1E, red line). We compared growth of our mutants in fibroblast vs. macrophages using a competition growth assay. For this, we mixed unlabeled Δfpps mutants with fluorescent parental (Δku80-rfp) or complemented strains (Δfpps-cm-rfp) at a 20∶1 starting ratio. Fig. 1F and 1G show that parental and complemented clones are able to rapidly outgrow mutant parasite in macrophages (Mac, blue lines) while they grow at a similar rate in fibroblasts (Fib, black lines). Our interpretation of these results is that whether the enzyme is required or dispensable for growth of the parasite depends on the specific host cell and host-parasite interaction. In this context we note that growth of Δfpps mutants was well supported in low passage primary fibroblasts, as measured by plaque assay (Fig. S1, top row), but limited in aging fibroblast cultures (fibroblast with ∼40 passages; Fig. S1, bottom row). This suggests that the parasite isoprenoid metabolism may not only be sensitive to the cell type infected but also to its physiological and/or metabolic state.
We next addressed whether Δfpps mutants would be less virulent in vivo. The RH strain is hypervirulent, which can make it difficult to appreciate modest attenuation. We observed a difference in virulence when infecting mice with low parasite numbers (5–10) while higher doses (15–100) lead to death at 9–10 days (data not shown). We wondered whether the use of a less virulent strain would better highlight the difference in virulence between mutant and parental cell lines. We created conditional mutants using the described TATi cell lines [17]. There are two advantages for using these cells. First, the reduced virulence of the parental cell line allows the use of 104–105 parasites to infect mice. Second, these mutants are maintained in culture expressing an extra copy of TgFPPS, prior to suppression with anhydrotetracycline (ATc) thus avoiding preadaptation. We first expressed a regulatable copy of the TgFPPS in the TATi parental cell line (Fig. S2A) and created the cell line FPPS/FPPSi. In a clonal line derived from these cells we disrupted the endogenous TgFPPS gene as detailed before (see legend for Fig. S2) and generated ΔFPPS/FPPSi mutants. We established ATc regulation and gene deletion by western and Southern blot analyses, respectively (Fig. S2B and S2C). Plaque assays in fibroblasts in the presence of ATc showed no difference in the number and size of plaques (Fig. S2D). In contrast, a highly significant difference in growth was observed when parasites were allowed to infect macrophages (Fig. S2E), as seen before with the Δfpps mutants (Fig. 1E) indicating a fitness defect only evident under stressful conditions.
With the purpose to define a dependable inoculum to use for virulence studies, we first established a protocol in which we passed our Tati-derived strains through mice and performed in vivo titration experiments (see Materials and Methods and Fig. S3). This treatment increased consistency dramatically and we found that using an inoculum defined in this way resulted in reproducible virulence outcomes. To establish whether FPPS knockdown affects the ability of these parasites to cause disease, mice were infected with 10,000 FPPS/FPPSi or ΔFPPS/FPPSi tachyzoites (Fig. S3) (a reproducible number found after our in vivo titration experiments). Mice challenged with the FPPS/FPPSi strain succumbed to the infection even if they were given ATc in the water (Fig. S3, black lines). In contrast, mice infected with the ΔFPPS/FPPSi and receiving ATc survived the infection while mice infected with the same parasites but given a placebo were susceptible to infection (Fig. S3, compare red lines).
With the aim of understanding how the Δfpps mutant parasites manage to survive without the production of essential isoprenoids we measured growth of mutant Δfpps parasites and their parental strain after being deprived of host cells for a determined length of time. We exposed mutant, parental and complemented parasites for 30 min to an extracellular buffer and for a more accurate readout we switched to a plaquing efficiency protocol as described [18] in which there is only a 30-min contact interval between parasite and host (Figure 2A). Plaques were counted after 7 days of incubation. We observed that the number of plaques was significantly lower for the Δfpps parasites after being exposed to these stress conditions. We also measured ATP levels of parasites incubated in extracellular media for one hour. No difference in the ATP levels was observed in recently egressed parasites but there was a significant decrease in the Δfpps mutants after incubating them for one hour in extracellular buffer with glucose (Figure 2B). These results indicate that the Δfpps parasites do have a defect in energy generation, which is not evident under the protected intracellular environment. However, this defect becomes relevant when the parasite is outside the host and we were able to increase it by incubating them for an extended length of time before letting them continue with its lytic cycle (Figure 2A). A possible cause of this defect could be the synthesis of ubiquinone, isoprenylated cofactor of the mitochondrial respiratory chain, which may be more important when the parasites are extracellular. We measured mitochondrial membrane potential of Δfpps and the parental Δku80 parasites using JC1, a lipophilic, cationic dye that accumulates in mitochondria in a membrane potential dependent fashion and that changes color from green to red as it accumulates [14] (Figure 2C, upper panels). The lower panels show that for the mutant parasites JC1 stays mostly green indicative of a partially depolarized mitochondrial membrane (Figure 2C, lower panels). We also used flow cytometry to quantify the effect of knocking out the TgFPPS gene (Δfpps) and compare it with the parental strain Δku80 (Figure S4). We observed a dramatic drop of mitochondrial fluorescence in Δfpps parasites (56.2%, compared to 85.2%). This indicates a mitochondrial defect that is not important for intracellular life but becomes accentuated when the parasites are deprived of host cells.
Our surprising findings could imply that intracellular parasites salvage isoprenoids from their host and that we impinge on this ability by cultivation in different cells. To test this hypothesis, we performed two labeling experiments testing different conditions. We first labeled infected fibroblasts with 14C-glucose (Figure 3A). Under these conditions, radioactive glucose is available to both host and parasite to label isoprenoids generated by host and parasite specific de novo synthesis pathways. In the second experiment the strategy was to first label the fibroblasts with 14C-glucose, remove unincorporated label by washing the cells with fresh medium and only then infect with parasites (Figure 3B). In both settings we compared parental (Δku80), mutant (Δfpps), or complemented (Δfpps-cm) T. gondii. Parasites were purified through several filtration steps before isoprenoid extraction.
Parasite isoprenoids were isolated by solvent extraction of purified tachyzoites and analyzed by thin layer chromatography and autoradiography (TLC). When infected cells were labeled (Figure 3A) the most abundant isoprenoids were FPP, GGPP, an intermediate co-migrating with a 25C standard (25 C), and a longer unidentified derivative co-migrating with a 35C standard (long prenyl diphosphate; LPP, 35C). The results indicate that mutant parasites (Δfpps) have levels of intermediates similar to the parental strain despite the lack of FPPS (differences between labeled compounds were not statistically significant, n = 3, data not shown). Figure 3B shows the isoprenoids obtained from the parasite after labeling only the host cells followed by infection with unlabeled parasites. FPP and GGPP were still present and labeled in the extracts obtained from mutant parasites despite the fact that they lack the enzyme required for their synthesis (Figure 3B, Δfpps). However, labeling of the longer chain product was stronger in extracts from Δfpps mutants. This likely indicates that the parasite synthesizes these longer chain products using both host and its own precursors, and that labeling via the host pathway becomes more evident in the absence of parasite synthesis (Figure 3B, Δfpps). The parental and complemented cells did not show this labeling arguing that it is generated in the parasite using unlabeled precursors. The results were quantified by calculating the ratio of labeling for this long chain product to that of GGOH and is displayed with bar graphs in Figs. 3C and 3D. This analysis shows no difference between mutant and wild type when parasite and host cells are simultaneously labeled with 14C-glucose (Figure 3C). However, the ratio was significantly higher for the mutant parasite when only the host cells were prelabeled (Fig. 3D). Taken together these results suggest that mutant parasites lacking their own production of FPP and GGPP import these intermediates from the host (pre-labeled with 14C in our experimental set-up) and convert them into the long chain isoprenoid. Under similar experimental conditions, when analyzing the isoprenoid products made by the parental strain, the labeling of this long chain isoprenoid product becomes diluted as a consequence of the endogenous production of unlabeled FPP and GGPP by the TgFPPS.
If the parasite is taking up FPP and/or GGPP from the host, then inhibiting the synthesis of these host compounds may affect parasite growth. We directly tested this idea using an inhibitor of hydroxymethyl glutaryl-CoA reductase (HMG-CoA reductase), the rate-limiting enzyme of the host mevalonate pathway (this pathway is absent in T. gondii). We tested atorvastatin (Lipitor) in tissue cultures (Fig. 4A and 4B, black lines) and found that atorvastatin is able to inhibit growth of the parental strains with an IC50 of ∼40 µM. We thought that this modest level of efficacy points to other sources of FPP and GGPP for the parasite, in particular its own synthesis. We hypothesized that the Δfpps mutants, unable to produce FPP and GGPP should be more sensitive to the inhibition of the host by atorvastatin. This is indeed what we observed when testing the drug against the mutant parasites (Fig. 4A and B, red lines) and we calculated an IC50 of 2 µM (20 times lower than the efficacy against the parental cell lines). This effect of atorvastatin is specific to its inhibition on the production of isoprenoid metabolites because it was possible to rescue parasite growth by adding geranylgeraniol to the medium (Fig. 4B, red lines: Δfpps and Δfpps+GGOH).
To investigate whether atorvastatin inhibits parasite growth mainly as a result of impaired cholesterol synthesis we tested WC-9, a known inhibitor of squalene synthase (SQS) [19]. We found WC-9 to inhibit parasite growth with an IC50 of 4–5 µM (Fig. 4C). T. gondii does not encode SQS and acquires cholesterol from its host [20], [21], [22]. We therefore attributed the effect of WC-9 to its action against the host pathway. Importantly, we found no difference in the WC-9 susceptibility of Δku80, Δfpps and Δfpps-cm parasites (Fig. 4C). This suggests that WC-9 acts downstream of the formation of FPP and GGPP, and that inhibition of cholesterol synthesis is not the most important anti-parasitic effect of statin treatment. This is in agreement with previous findings that suggested that the parasite relies on LDL trafficking rather than de novo synthesis by the host cell to satisfy its cholesterol requirement [20], [21], [22].
We next tested another statin (mevastatin) on FPPS/FPPSi or ΔFPPS/FPPSi parasites grown in the presence or absence of ATc (Figs. 4D). FPPS/FPPSi tachyzoites express an extra copy of TgFPPS (Fig. S2B) and possesses higher FPPS activity (not shown). There was a reverse correlation between mevastatin inhibition and the expression level of TgFPPS (Fig. 4D). ΔFPPS/FPPSi parasites were the most susceptible in the presence of ATc (IC50∼4 µM mevastatin) while FPPS/FPPSi cells with an extra copy of the TgFPPS gene were resistant to concentrations up to 18 µM (Fig. 4D). The effect of mevastatin was rescued by supplementation of the medium with 1 µM geranylgeranyol (Fig. 4E, compare red and blue lines), again supporting its direct effect on the production of FPP and GGPP.
We also tested the efficacy of atorvastatin treatment against T. gondii infection of mice using wild type RH strain. Fig. 5A shows a summary of 3 experiments using groups of 5 mice treated with different doses of atorvastatin. While 100% of control mice died between 9–13 days post-infection, 80% of mice treated with the higher 40 mg/kg/day dose, survived more than 30 days. Note that this is not an excessive drug dose but the standard concentration of atorvastatin commonly used and well tolerated in mouse experiments [23], [24]. An atorvastatin ED50 of 32.3 mg/kg per day was calculated (Fig. 5A). We also were interested in comparing the efficacy of atorvastatin against the infection of mice with Δku80, and Δfpps cells. We infected mice with a lethal dose of parasites (parental and mutants) to highlight the effect of atorvastatin against infection with the Δfpps clone. Fig. 5B shows that atorvastatin is highly effective at treating mice infected with Δfpps parasites: 9 of 10 mice survived the infection when treated with atorvastatin, while 8 of 10 mice died in the absence of atorvastatin. To establish further that knockdown of TgFPPS make T. gondii infection more amenable to treatment with atorvastatin we infected mice with a lethal dose of 100,000 ΔFPPS/FPPSi or FPPS/FPPSi tachyzoites and treated with ATc in their drinking water (Fig. 5C). This high parasite dose was lethal even when infecting with ΔFPPS/FPPSi (compare with Fig. S3 for which we used 10,000 parasites, ten fold difference in dose) [25]. Most mice infected with FPPS/FPPSi and treated with atorvastatin succumbed to this high infection (Fig. 5C, black lines). In contrasts most mice infected with ΔFPPS/FPPSi were cured by atorvastatin when the mutation was induced by ATc treatment (Fig. 5C, red lines).
T. gondii appears to be able to rely on both synthesis and salvage of isoprenoids. Δfpps mutants are more dependent on salvage. Could this be exploited pharmacologically by combining inhibitors of TgFPPS with atorvastatin? Bisphosphonates are known inhibitors of FPPS and have shown antiparasitic activity [11]. We chose to test zoledronic acid [26] because our previous work had identified this compound as the bisphosphonate with the highest specificity against TgFPPS, and its activity decreased significantly when we overexpressed the parasite enzyme [11]. To evaluate interaction between atorvastatin and zoledronate, we mixed both drugs at different concentrations following a protocol designed for testing synergy [27]. This protocol measures and calculates the IC50 of one drug in the presence of subtherapeutic concentrations of the second drug [27] . The results were plotted in an inhibition isobologram using IC50s of individual drugs and of five different drug combinations (Fig. 6A). The resulting curve is concave for atorvastatin and zoledronic acid and thus indicative of synergistic drug interaction (Fig. 6A).
FPP and GGPP production in the parasite requires the isoprenoid precursors IPP and DMAPP. We therefore next wanted to test whether atorvastatin would interact with fosmidomycin, a specific inhibitor of the DOXP pathway. T. gondii is insensitive to fosmidomycin because the drug is not able to cross the parasite membrane [7]. However a T. gondii transgenic parasite that expresses the bacterial transporter glycerol-3-phosphate transporter (GlpT) capable of importing fosmidomycin, is sensitive to fosmidomycin [7] (Fig. 6B). We assessed the growth of these parasites in the presence of 50 µM atorvastatin and 0.78 µM of fosmidomycin. This represents the IC10 for fosmidomycin and this low concentration was deliberately chosen to be able to detect drug interaction. Individually these drugs affected parasite growth as expected, approximately 50% inhibition with 50 µM atorvastatin and very little inhibition with 0.78 µM fosmidomycin. Interestingly, combining both drugs abolished parasite growth, indicating strong interaction also between atorvastatin and fosmidomycin. We tested the interaction between atorvastatin and fosmidomycin in these transgenic parasites by a simplified checkerboard technique [27] and calculated the fractional inhibitory concentration (FIC) index to be 0.36, confirming synergistic interaction (FIC<0.5) [28], [29]. This assay provided additional strong evidence that the parasite, although capable of generating its own isoprenoids, also depends on the host isoprenoids for continuous growth and successful infection. Our results show that therapeutic strategies aimed at interfering with both parasite and host isoprenoid synthesis could provide a higher rate of success in curing T. gondii infections.
Our work reveals a crucial metabolic interaction between the intracellular pathogen T. gondii and its host cell to secure the parasite's access to isoprenoids. Isoprenoids are essential for all cells and in most Apicomplexans their five carbon precursors are produced by the apicoplast [30], [31]. The synthesis of these precursors is now viewed as the most important function of the apicoplast and the reason the organelle was maintained long after the loss of photosynthesis [32]. Genetic analysis in T. gondii demonstrates that loss of the apicoplast isoprenoid pathway is lethal and mimics complete loss of apicoplast metabolism [7], [14]. Inhibiting this pathway with the antibiotic fosmidomycin is effective against Plasmodium, Babesia, and against Toxoplasma (once parasites are engineered to take up the drug) [7], [33], [34]. Most intriguingly, in Plasmodium falciparum cells cured of their apicoplasts by antibiotic treatment targeting plastid translation can nonetheless be continuously maintained in culture when the media are supplemented with high concentrations of IPP [35]. Overall these studies suggest that the synthesis of IPP and DMAPP by the parasite is essential and cannot be circumvented by salvage from the host under physiological conditions.
This makes our observation that the parasite enzyme catalyzing the next step in the isoprenoid pathway – the synthesis of FPP and GGPP from IPP and DMAPP is dispensable for T. gondii in fibroblasts all the more surprising. FPPS-catalyzed reactions are essential in most organisms studied so far and are important drug targets. T. gondii is not only able to make its own isoprenoids but can also import from the host cell (Fig. 7). We note that this ability to salvage appears not to be universal but restricted to certain compounds (Fig. 7). At the moment it is not fully understood whether this difference is due to difference in transport capability of the parasite or availability and abundance of the metabolites in the host cell. However, in our experiments we measured a strong impact of the host cell environment for FPP and GGPP. Extracellular parasites or parasites infecting macrophages rather than fibroblasts show more pronounced defects upon loss of the synthesis capacity. The intracellular survival of T. gondii depends on its unique ability to invade cells actively. Active invasion is fundamentally different from phagocytosis and requires parasite motility [16]. When extracellular parasites are incubated in PBS, their ability to invade cells actively rapidly declines, and they are mostly internalized by phagocytosis with parasites engulfed in phagosomes, which fuse with endosome/lysosomes and are further digested [16]. Parasite fitness is essential for its ability to actively invade host cells and/or escape from the phagosome. Lack of endogenous production of FPP and GGPP by T. gondii renders them less able to grow in macrophages. This could be the result of a fitness defect or because of a shortage of metabolites in macrophages or a different mechanism of transport of isoprenoids in these cells.
How dependent is the parasite on isoprenoid salvage under normal conditions with its synthesis capability intact? Our labeling studies show robust import of host cell-synthesized isoprenoids even in wild type parasites. Import is also supported indirectly by microarray studies of T. gondii-infected fibroblasts that revealed a significant induction of genes encoding enzymes of the mevalonate pathway following infection [5], [36] including the rate-limiting enzyme HMG-CoA reductase, and FPPS [5]. Previous work has shown that T. gondii does not synthesize cholesterol and imports it from the host low–density lipoprotein (LDL) [20], [21], [22]. It is possible that the inhibition of cholesterol synthesis by statins results in reduced parasite invasion or reduced parasite growth. Interestingly, a recent study has shown that atorvastatin treatment of endothelial cells reduced cytoadherence of Plasmodium falciparum [37].
We consider that inhibition of host cholesterol synthesis is unlikely as the reason for the effect of statins because of three reasons. First, the isoprenoid intermediate geranylgeraniol was able to rescue almost completely the growth inhibition by two statins, atorvastatin and mevastatin. Second, growth inhibition by an SQS inhibitor that blocks the pathway downstream to the production of FPP and GGPP was not enhanced in the Δfpps mutants, as observed with atorvastatin and mevastatin. And third, statins do not reduce overall plasma cholesterol levels in mice as they do in humans (due to low levels of low density lipoproteins in rodents) [38], [39]. In addition, it has been demonstrated that host cell cholesterol production has no significant effect on parasite replication and that the bulk of parasite cholesterol requirement can be satisfied by exogenous cholesterol from low-density-lipoprotein delivered to the parasitophorous vacuole [20]. Our results with the squalene synthase inhibitor strongly support that conclusion. It is interesting to note that the growth in macrophages of Salmonella enterica serovar Typhimurium, which also lacks a mevalonate pathway, is inhibited by statins and this inhibition is not due to a deficient production of sterols but of intermediates of the pathway between mevalonate and squalene 2,3-oxide [38]. It would be very interesting to investigate whether these intermediates are also FPP and GGPP as in the case of T. gondii.
Our labeling results indicate that T. gondii may use its own enzymes to make specific long chain isoprenoids. We previously reported that TgFPPS localizes to the mitochondria [8]. Our results showed that loss of TgFPPS resulted in alteration of the mitochondrial membrane potential, and rapid decrease in the ATP levels of extracellular parasites. These results suggest that TgFPPS functions to make FPP and GGPP in the mitochondrion as precursors for long chain isoprenoids and ubiquinone synthesis. Therefore we could deduce that TgFPPS plays an important role in maintaining mitochondrial function. This appears to be crucial for the parasite during its extracellular stage. Our results suggest a requirement for oxidative phosphorylation for generation of ATP in extracellular parasites. This is consistent with our previous results showing active oxidative phosphorylation in extracellular parasites [40] and a recent report showing that oxidative phosphorylation is responsible for >90% ATP synthesis in extracellular tachyzoites [41]. The deficient synthesis of ubiquinone precursors would not affect tachyzoites when they are intracellular while seriously impeding extracellular parasites as a consequence of rapid depletion of ATP, which is needed for gliding motility and invasion. Taken together, we show that T. gondii has a versatile system for its isoprenoid needs. During its replicative stage with needs for large quantities of isoprenoids, the parasite is able to manipulate the host and salvage isoprenoids. However, under stressful situations the parasite is able to provide by itself and this was emphasized when exposing Δfpps or conditional knockout mutants to challenging conditions like infection in vivo, growth in macrophages, growth in metabolic inactive host cells, or during extended extracellular life. These findings show that the endogenous activity of TgFPPS, while low compared to other FPPs (Table S1) is needed under stress or for other functions.
This ability of the parasite to use not only its own metabolites but also to manipulate the host cell metabolism and salvage its products makes it a challenge for drug therapy. However, in the case of isoprenoid metabolism this split reliance may also prove to be an opportunity as it can build on the massive investment made into controlling this pathway pharmacologically in the host. Our work demonstrates that inhibition of the host mevalonate pathway enhances the impact of blocking the parasite isoprenoid pathway and we propose a double hit strategy that combines inhibitors of the parasite enzyme with host isoprenoid pathway inhibitors. We tested combinations of two approved and widely used drugs, zoledronic acid (Zometa) and atorvastatin (Lipitor) and showed synergism in the inhibition of T. gondii growth. We demonstrated that impinging on host or parasite isoprenoid synthesis reduces parasite virulence but that blocking both produces stronger effects and affords considerable protection. This strategy could prove even more promising when tested in other parasites. For example our experiments combining fosmidomycin with atorvastatin suggest that atorvastatin may boost the efficacy of fosmidomycin as an antimalarial. This combination may also make it more difficult for the parasite to develop resistance extending the useful life of the drug. Our strategy will benefit from the extensive clinical knowledge on both statins and bisphosphonates and this knowledge will facilitate their use for the treatment of other infections.
Mice experiments in this work followed a reviewed and approved protocol by the Institutional Animal Care and Use Committee (IACUC). Animal protocols followed the US Government principles for the Utilization and Care of Vertebrate animals. The protocol was reviewed and approved by the University of Georgia IACUC (Protocol number A2012-3-010).
Oligonucleotide primers were obtained from Integrated DNA Technologies (Coralville, IA). Taq DNA polymerase, and restriction enzymes were from Invitrogen or New England Biolabs. Plasmid miniprep and maxiprep and gel kits were from Qiagen Inc. (Chatsworth, CA). IPP, DMAPP, GPP, FPP, GGPP were from Isoprenoids, LC (FL, USA). [4-14C] Isopentenyl diphosphate triammonium salt (55.0 mCi/mmol) and [14C(U)]-glucose (319 mCi/mmol) were from PerkinElmer Life Sciences. Atorvastatin (Lipitor) was from Pfizer. Zoledronic acid and WC-9 were a gift from Dr. Juan B. Rodriguez, University of Buenos Aires. Fosmidomycin was a gift from Dr. Yongcheng Song (Baylor College of Medicine). All other reagents were analytical grade.
Tachyzoites of T. gondii RH strain were cultured in human fibroblasts or hTERT cells in Dulbecco's modified Eagle's medium (DMEM) supplemented with 1% fetal bovine serum, 2 mM glutamine, and 1 mM pyruvate, and purified as described before [8]. A basic electroporation protocol was used for transfection. Briefly, 107 recently released parasites and 20 µg of sterilized cosmid (see below) or plasmid DNA were mixed in a 2-mm gap electroporation cuvette. Electroporation was performed using a Genepulser Xcell from BioRad and after 15 min of recovery the parasites were allowed to infect fibroblasts. Stable transfectants were selected with 20 µM chloramphenicol and cloned by limited dilution. For the TgFPPS cDNA complemented cell line, the Δfpps parasites were transfected with a TgFPPS cDNA construct [8], then cultured in 15 µM atorvastatin for 4 passages. These parasites were sub-cloned by limited dilution medium containing 5 µM atorvastatin. Subcloned parasites were analyzed by PCR, Southern and western blot. All the clones that survived to atorvastatin selection have the TgFPPS cDNA stably integrated. The complemented clone used for the experiments was named Δfpps-cm.
The protocol for creating TgFPPS conditional deletion mutants is described in the supporting information (Fig. S2 legend).
The cosmid PSBLI36TV was obtained from L. David Sibley (Washington University). The knockout cassette from pH3CG was amplified by using the primers (5′-GCGGCCACCGTCCATAATTGCAAAAATGGAGCGGCTGTGTTTCCGTCTCCTCGACTACGGCTTCCATTGGCAAC-3′ and 5′-CTATTTCTGCCGTTT GTGGAGCCTCCCGAGGACGAGGCCGAAGAAGGCCTATACGACTCACTATAGGGCGAATTGG-3′). The TgFPPS gene targeting cosmid construct was obtained by recombineering in E. coli EL250 as described previously [14] (Fig. 1A).
Plaque assays and growth assay of tagged parasites were performed as described before [7]. Plaquing efficiency was measured infecting hTERT monolayers with 1,000 parasites per well and allowing contact with host cells for 30 min. At this point, wells are washed with PBS, fresh media added and parasites allowed to grow for 4–7 days, fixed and stained with crystal violet [18]. Parental and mutant strains of T. gondii were transfected with a plasmid containing a tandem tomato RFP gene and red fluorescent parasites were sorted and subcloned by FACS analysis. Δku80-rfp, Δfpps-rfp and Δfpps-cm-rfp cell clones were obtained. Growth competition assays were performed by mixing strains: Δfpps parasites with Δku80-rfp and Δfpps-cm-rfp cell lines at 20∶1 ratio (5% of red cells in the mixture). These parasite mixtures were used to infect fibroblasts or macrophages. 1×106 parasites were inoculated in each passage. Percentage of red cells at each passage was calculated using a standard curve generated by measuring the fluorescence intensity for a fixed number of cells.
T. gondii genomic DNA was digested with SalI, separated in a 0.8% agarose gel, and transferred to a nylon membrane. The DNA probe was generated by PCR with primers (5′-TGACGCGCTGAGCAGTGGTGAGCA-3′ and 5′-AGCCATTTCAACTTCAAACCGCA-3′). The purified PCR product was 32P labeled by random priming.
Western blots were done using an affinity purified rabbit polyclonal antibody raised against TgFPPS at 1∶1500 in PBS-T. The secondary antibody was horseradish peroxidase-conjugated goat anti-rabbit and immunoblots were visualized on blue-sensitive x-ray film by using an ECL detection kit.
Purified parasites were washed in Ringer (155 mM NaCl, 3 mM KCl, 1 mM MgCl2, 3 mM NaH2PO4-H2O, 10 mM Hepes, pH 7.3, 10 mM glucose) and resuspended in Buffer A with glucose (BAG, 116 mM NaCl, 5.4 mM KCl, 0.8 mM MgSO4, 5.5 mM D-glucose and 50 mM Hepes, pH 7.2) at 1–5×108 cells/ml. The parasite suspension was incubated at 37°C for 1 hour. The suspension was extracted with perchloric acid as described previously [42]. Briefly, the parasite suspension was centrifuged and resuspended in 100 µl of BAG and mixed with 300 µl of 0.5 M HClO4 and incubated in ice for 30 min. The supernatant was neutralized with 0.72 M KOH/0.6 M KHCO3 and used immediately or stored at −20C for ATP measurement. The kit A22066 from Invitrogen was used to measure ATP.
Two labeling strategies were used. The first one consisted on labeling both host cells and parasites. Briefly, hTERT-fibroblasts were first infected with fresh tachyzoites and grown in DMEM medium containing 14C-glucose until the natural release of parasites. The second strategy consisted of labeling only host cells and infecting afterwards. Host cells were grown in medium containing 14C-glucose and before infection the monolayer was thoroughly washed and fresh medium containing glucose was added. For both experiments, released tachyzoites were purified by several filtration steps (8, 5, and 3 µM membranes), to ensure the absence of host cells, and lipids extracted in chloroform/methanol at 4°C overnight. After filtration followed a saponification step (with KOH and ethanol) and the radioactive prenyl products in the mixture were hydrolyzed to the corresponding alcohols with alkaline phosphatase at room temperature, overnight. The resulting alcohols were extracted with hexane and separated on a HP-TLC-RP18 plate using acetone∶H2O (10∶1; v/v) as the moving phase. Standard prenyl alcohols were run in parallel and were visualized by iodine vapor. Radioactivity of the products was also measured by autoradiography or phosphorImaginer analysis.
All parasite clones were grown in fibroblasts using similar conditions to those used for the RH strain. For in vitro drug testing, confluent hTERT monolayers in 96-well plates were first prepared with phenol-free medium containing the drugs serially diluted and infected with 4,000 parasites per well. The plates were incubated at 37°C and the fluorescence measured every day. Regression analysis and IC50 calculations were performed using SigmaPlot 10.0.
Isobolograms were constructed by plotting the IC50 of one drug against the IC50 of the other for each of five drug ratios, with a concave curve indicating synergy, a straight line indicating addition, and a convex curve indicating antagonism. For simplified checkboard studies, drugs were mixed in fixed ratios of their respective IC50s and dose-response curves generated from serial dilutions carried out in triplicate. Results were expressed as the sums of the fractional inhibitory concentration (sum FIC = IC50 of drug A in mixture/IC50 of drug A alone)+(IC50 of drug B in mixture/IC50 of drug B alone), as described by Berenbaum [43]. Sum FIC values indicate the kinds of interactions as follows: <0.5, synergy; 1, addition; >2, antagonism.
For in vivo infection with T. gondii, fresh tachyzoites were harvested, washed with PBS twice, and resuspended in PBS before inoculation. Female Swiss Webster or BALBc mice were injected with 5–20 tachyzoites of the RH strain i.p. in a 200 µl PBS final volume or 10,000–100,000 tachyzoites of the TATi strain in a similar volume. When using low parasite numbers, plaque assays were performed with the parasite suspensions used to inoculate mice to ensure that the number represented the number of viable and infectious parasites.
Because initial in vivo experiments with cultured TATi-derived strains gave inconsistent virulence results, we developed a protocol by which we first infected mice with a high dose (106 parasites) of parasites (parental, knock-in and knock outs) and collected the peritoneal fluid (containing tachyzoites) five days p.i. This suspension was used to infect a confluent flask of fibroblasts and allowed to grow and lyse. The supernatant from these flasks containing tachyzoites was collected, centrifuged and parasites resuspended in the appropriate media to prepare aliquots for freezing in liquid nitrogen. For each experiment, one vial was thawed and passed once through tissue culture before used for infection. Tati-derived strains showed a remarkable recovery in virulence with this treatment. We performed titration experiments and determined that 10,000 parasites of Tati-derived strain (no ATc) are lethal to mice 9–10 days p.i.. Results shown in Figs. S3 and 5C were obtained with parasites previously treated as described.
Drugs were dissolved in phosphate-buffered saline (PBS) containing approximately 2% DMSO, at pH 6.8, and were also inoculated i.p. Treatment was initiated 6 hours after infection and administered daily i.p. for 10 days.
|
10.1371/journal.pcbi.1006677 | Chemogenomic model identifies synergistic drug combinations robust to the pathogen microenvironment | Antibiotics need to be effective in diverse environments in vivo. However, the pathogen microenvironment can have a significant impact on antibiotic potency. Further, antibiotics are increasingly used in combinations to combat resistance, yet, the effect of microenvironments on drug-combination efficacy is unknown. To exhaustively explore the impact of diverse microenvironments on drug-combinations, here we develop a computational framework—Metabolism And GENomics-based Tailoring of Antibiotic regimens (MAGENTA). MAGENTA uses chemogenomic profiles of individual drugs and metabolic perturbations to predict synergistic or antagonistic drug-interactions in different microenvironments. We uncovered antibiotic combinations with robust synergy across nine distinct environments against both E. coli and A. baumannii by searching through 2556 drug-combinations of 72 drugs. MAGENTA also accurately predicted the change in efficacy of bacteriostatic and bactericidal drug-combinations during growth in glycerol media, which we confirmed experimentally in both microbes. Our approach identified genes in glycolysis and glyoxylate pathway as top predictors of synergy and antagonism respectively. Our systems approach enables tailoring of antibiotic therapies based on the pathogen microenvironment.
| The antibiotic resistance epidemic has created a pressing need to understand factors that influence antibiotic efficacy. An often-overlooked factor in the search for new treatments is the pathogen environment. Understanding the differences in pathogen sensitivity to antibiotics in lab conditions versus inside the host is necessary for translating new discoveries into the clinic. Hence, we experimentally measured the sensitivity of E. coli to drugs and drug combinations in different metabolic conditions. Our data revealed that the environment dramatically changes treatment potency. Each antibiotic class was affected uniquely by each metabolic condition. The large number of metabolic conditions inside the host greatly complicates the identification of effective therapies. To address this challenge, we present a computational approach called MAGENTA that accurately predicted efficacy of antibiotic regimens in different conditions, which we confirmed experimentally. Furthermore, we show that MAGENTA can be applied to other bacterial pathogens such as A. baumannii and M. tuberculosis without the need for generating expensive data in each organism. MAGENTA accurately predicted efficacy in the pathogen A. baumannii using data from E. coli by identifying genes that are common between the two bacteria. Our study revealed the significant yet predictable impact of environment on drug combination potency.
| The threat of antibiotic resistance coupled with a diminishing pipeline of new drugs has created a pressing need to enhance efficacy of existing antibiotics [1]. Combinations of antibiotics are now being increasingly used to enhance the efficacy of treatment regimens and concurrently reduce resistance [2]. A key factor influencing the efficacy of antibiotic therapies is the environmental context [3,4]. Antibiotics need to act in diverse and complex metabolic environments in vivo, in contrast to well controlled lab conditions. Environmental factors such as the availability of oxygen and extracellular metabolites impact cell killing by antibiotics [5]. The strong impact of metabolic state on drug efficacy has been observed across diverse microbial pathogens [3,6,7].
In addition to individual drugs, pathogen metabolism can also influence the efficacy of drug combinations. Drugs in a combination can enhance or interfere with other drugs’ actions, leading to synergistic and antagonistic interactions [8]. For example, combinations involving bacteriostatic antibiotics, which inhibit cell growth, and bactericidal antibiotics, which induce cell death, are typically avoided in the clinic due to their antagonistic interaction [9]. This antagonism is hypothesized to occur due to their opposing effect on cellular respiration [10]. While recent studies have focused on the influence of metabolic environment (i.e. availability of nutrients, oxygen, extracellular metabolites) on individual drugs, a systematic analysis of the impact of metabolic environment on drug interactions is lacking. It is unclear if drug interactions are sensitive or robust to the environment.
Understanding the impact of metabolic environments on antibiotic efficacy is essential for clinical translation of antibiotic therapies discovered from in vitro screens [11]. This can ultimately help predict the impact of gut and tissue microenvironment on antibiotic susceptibility. Further, knowledge of the pathogen metabolic environment is critical for treating slow-growing pathogens like M. tuberculosis and targeting pathogen biofilms.
Yet traditional in vitro testing is typically done in a single metabolic condition [11]. Given the large space of possible metabolic environments in vivo [12], in silico algorithms are needed to predict the impact of various metabolic environments on drug combinations. Existing approaches to infer drug-drug interactions in both microbes and cancer cells, including the INferring Drug Interactions using chemo-Genomics and Orthology (INDIGO) approach that we previously developed [13], assume interaction outcomes are fixed for a drug combination.
To address this challenge of predicting the impact of metabolic environment on drug combination efficacy, here we develop the Metabolism And GENomics-based Tailoring of Antibiotic regimens (MAGENTA) approach. This systems-biology approach comprehensively captures the cellular processes involved in drug action and drug interactions such as stress response, drug transport and metabolism. MAGENTA achieves this by harnessing chemogenomic profiles associated with both distinct metabolic environments and drugs. Chemogenomic screens measure fitness of gene knockout strains treated with drugs and stress agents of interest [14]. MAGENTA identifies genes that significantly impact fitness when exposed to drugs or metabolic stressors from chemogenomics data resulting in a set of drug-gene and metabolic environment-gene interactions. It then uses these chemical-genetic interactions to predict drug-drug interactions in a new environment. MAGENTA applies a machine learning algorithm called Random Forests to identify a core group of genes in the chemogenomic profiles that are predictive of drug synergy and antagonism across metabolic environments.
MAGENTA enables, for the first time, the prediction of impact of metabolic conditions on antibiotic combination efficacy. We then experimentally validate this approach by testing predictions involving several drug combinations across distinct metabolic conditions in E. coli. In addition, we apply the E. coli MAGENTA model to predict interactions in the pathogen—Acinetobacter baumannii, by identifying genes that are conserved between the two species. A. baumannii is frequently associated with multi-drug resistance and is ranked as one of the most dangerous pathogens in hospitals worldwide by the Infectious Diseases Society of America [15]. Here we identified drug combinations that are synergistic across multiple conditions in A. baumannii by overlaying orthologous genes on to the E. coli MAGENTA model. This orthology mapping approach can enable the application of data from expensive chemogenomics and drug-interaction screens in model organisms for identifying synergistic combinations in several related bacterial pathogens with genome sequence information.
To understand the impact of metabolic environment on drug interactions, we experimentally measured all pairwise interaction outcomes of 8 antibiotics against E. coli cells grown in LB (rich) and M9 glucose (minimal) media (Fig 1). We used the popular Loewe-additivity model and the Fractional Inhibitory Concentration (FIC) metric to quantify the drug interactions [16,17] (Methods). The FIC for a drug combination is obtained by adding the fractional MIC (i.e. Minimum Inhibitory Concentration) of each drug in a combination. The FIC scores were log2-transformed for ease of interpretation. Negative log-FIC scores (i.e. < 0) imply synergy, i.e. the same amount of growth inhibition is achieved with a lower dose when both the drugs are combined. Antagonistic interactions have a positive log-FIC score.
Analysis of our experimental drug interaction FIC scores revealed that change in metabolic state strongly influenced sensitivity to drug combinations. The drug interaction FIC scores changed considerably between the two conditions, with only 42% of the combinations showing the same direction of interaction (i.e. antagonism–log-FIC > 0.2, synergy–log-FIC < -0.2). Interestingly, interactions were significantly more synergistic in glucose media (mean log-FIC = -0.13) compared to LB (mean log-FIC = +0.28; p-value = 0.001, paired t-test). Interactions involving combinations of bactericidal and bacteriostatic drugs showed the strongest difference between the two conditions compared to other drug combinations (Fig 1). These combinations became strongly synergistic in glucose media from being weakly antagonistic in LB media (mean log-FIC = -0.37 in glucose media, mean log-FIC = +0.14 in LB; p-value = 0.08, paired t-test).
These results suggest that drug interaction outcomes are not fixed for a drug combination and can change significantly depending on the metabolic environment. This further complicates the challenge of predicting drug interactions–in addition to the large space of drugs and dosage, the metabolic environment of the pathogen should also be considered.
To account for this variability in drug interactions, we developed the MAGENTA framework to predict the impact of metabolic environment. MAGENTA takes as input the chemogenomics data of individual drugs and known drug-drug interaction training data, and outputs the predicted interaction score for a list of novel drug combinations (Methods). To predict the impact of metabolic state on drug sensitivity, we first identify genes that impact fitness during growth in distinct metabolic conditions from chemogenomic profiling. We then model metabolic perturbations using a similar framework for modeling drugs, and subsequently predict the impact of metabolic state on drug combination efficacy.
We introduce two key features in MAGENTA to enable prediction of metabolic impact on drug sensitivity, which is not possible using existing drug interaction tools such as INDIGO or Overlap2 Method (O2M) [18]. Firstly, MAGENTA uses chemogenomic profile of the metabolic condition as input, in addition to using chemogenomic profiles of the drugs in a combination. This involves the integration of three different chemogenomic profiles. While existing chemogenomics frameworks can predict interactions between pairwise combination of drugs, here we show that MAGENTA can make predictions of combinations of multiple stress agents (i.e. > 2). This allowed us to simulate the impact of metabolic conditions on drug interactions. The mathematical framework used by MAGENTA for quantifying drug chemogenomic profile similarity and uniqueness respectively are directly scalable to multiple combinations (S1 Fig). To account for dosage, we re-scale the scores by normalizing them by the number of drugs in a combination in order to achieve the same units as the model defined for two drugs.
Further, we found that chemogenomic profiles of metabolic conditions differ distinctly from chemogenomic profiles of drugs and other stress agents (S2 Fig). This is because, while drugs have significant number of chemical-genetic interaction with genes that confer both resistance and sensitivity, nutrients such as glucose predominantly contain genes that confer sensitivity. Hence, as a second addition to the MAGENTA framework, we also used the genes that confer resistance as input. Data on genes conferring resistance is relevant here for differentiating media conditions from drugs.
We trained MAGENTA using our experimental pairwise drug-drug interaction data in LB and glucose media (Fig 1) along with data for 171 drug pairs generated in our prior study [13]. We used chemogenomic profiles for these drugs and metabolic perturbations from the Nichols et al. study [19], which screened 3979 gene-deletion strains of E. coli with 72 drugs and 8 metabolic perturbations.
Since the MAGENTA framework integrates multiple chemogenomic profiles, we first confirmed MAGENTA’s ability to predict the outcome of multi-drug combinations from individual drug chemogenomic profiles. To test the predictions by MAGENTA, we experimentally measured 56 three-way combinations involving 8 antibiotics in LB media. These 8 antibiotics included drugs with distinct targets and mechanism of action (Fig 2; Table 1). Comparison of MAGENTA predictions with experimental measurement revealed that it accurately predicted three-way drug interaction outcomes with significant correlation (Rank correlation R = 0.57, p-value = 5 x 10−6; Fig 2). We observed a similar accuracy for predicting pairwise drug combinations in our prior study [13]. This suggests that the MAGENTA framework can be seamlessly extended to multi-drug combinations. Further, we observed that the majority of the three-way interaction outcomes were surprisingly antagonistic and only 2 out of the 56 combinations showed synergy (interaction score < -0.2). Given the uneven distribution of synergy and antagonism, we also assessed MAGENTA predictions using the Anova statistic and found that the predicted interaction scores differed significantly for synergy and antagonism respectively (p-value = 0.0003; S3 Fig). The high predominance of antagonism underscores the need for a computational approach to discover synergistic combinations.
Having confirmed that MAGENTA can accurately predict three-way interactions, we then applied MAGENTA to predict the impact of metabolic environment on drug interactions. Out of the 8 metabolic conditions for which chemogenomic data was available from the Nichols et al study [19], we used MAGENTA to predict interactions in minimal media with glycerol as carbon source. We chose glycerol condition as E. coli is predicted to use a different metabolic state than glucose [20]. We hypothesized that this major shift in metabolism will have a significant impact on drug combination efficacy.
To validate the model predictions, we experimentally measured all 55 pairwise drug interactions of 11 antibiotics in glycerol media, in duplicate. In addition to the 28 drug combinations previously tested, we also tested 27 new drug combinations involving three new antibiotics that were not part of the glucose media training data. These three antibiotics—cefoxitin, nalidixic acid and spectinomycin, use distinct mechanisms of action compared to those drugs used in the training data set. Hence this validation data set would test the limits of the algorithm with interactions involving both new metabolic conditions and new drugs with distinct mechanism of action.
In contrast to the glucose condition, MAGENTA did not predict strong synergy between bacteriostatic and bactericidal drugs in glycerol media. For example, MAGENTA predicted that combination of ampicillin, a bactericidal drug, and tetracycline, a bacteriostatic drug, were additive in glycerol media. Similarly, combination of aztreonam and azithromycin was also predicted to be additive in glycerol media. Experimental measurement of 55 drug combinations in glycerol media validated the MAGENTA predictions. Overall, comparison with experimental data revealed that MAGENTA accurately predicted interaction outcomes across all 55 drug combinations (Rank correlation R = 0.69, p-value = 1 x 10−8; Fig 3). Thus, the interaction scores predicted by MAGENTA accurately represent whether the metabolic perturbation can enhance or impede the efficacy of a drug combination.
Notably, a large subset of the test set involved new drug combinations for which we do not have training data. For the drug combination subset that was shared across all conditions (28 pairs), the overall consistency between conditions was lower than the predictions from MAGENTA. We split the interaction data as synergy (Interaction score < -0.2), neutral or antagonistic (> 0.2). The consistency was 32% (p-value = 0.5) for LB and glycerol, 53% (p-value– 6 x 10−4) for glucose and glycerol, 42% (p-value = 0.03) for LB and glucose. The consistency for MAGENTA was 65% (p-value = 1 x 10−6) for this subset and 62% overall for 55 combinations (p-value = 4 x 10−8). P-values were estimated by comparison of observed consistency with 1,000 random permutations from the training data set using a t-test.
The extent of change in interaction outcomes in glycerol media was influenced by the bacteriostatic and bactericidal nature of the drugs in the combination. Interactions involving combinations of bactericidal and bacteriostatic drugs did not show strong synergy in glycerol growth condition (mean log-FIC = -0.04), in contrast to our observation in glucose minimal media (mean log-FIC = -0.37). Surprisingly combinations involving two bacteriostatic drugs showed the strongest synergy (mean log-FIC = -0.28). Hence while the correlation was high between conditions for the subset of shared combinations, (R = 0.54 between LB and glucose, 0.64 between LB glycerol, and 0.77 for glucose and LB), it doesn’t represent the shift towards synergy or antagonism. The mean of the glucose interaction scores was -0.13, while for glycerol it was -0.27 for the combinations that overlapped and +0.26 for those that didn’t (S6 Fig). MAGENTA had a similarly high correlation of 0.78 for this subset, but also captured both the shift toward synergy and the relative ordering (S7 Fig, S8 Fig). To make predictions for a new condition using MAGENTA, instead of measuring all pairwise combinations of n drugs (n choose 2) across different media, only n chemogenomic profiles are needed along with a small training data set.
Growth in glycerol had a unique impact on drug interactions compared to LB and glucose media. Drugs that depend on facilitated transport like the aminoglycoside antibiotics—spectinomycin and amikacin, were more synergistic in glycerol media, possibly due to increased active uptake due to the higher activity of TCA cycle in glycerol condition [21]. The difference in interaction outcomes might also occur due to osmotic stress induced by glucose [22]. Drugs such as ampicillin, aztreonam and triclosan that disrupt bacterial cell wall were more synergistic in glucose media. The synergistic effect of ampicillin and triclosan is hypothesized to be due to the disruption of the cell wall by these drugs resulting in enhanced cellular penetration of a second drug [23]. The osmotic stress in glucose media relative to other conditions will further enhance the synergy of cell wall disrupting drugs by increasing membrane permeability. The differences in drug interaction outcome in glycerol media highlight the impact of metabolic environment on drug combinations.
The drug interaction validation data in glycerol media was used as additional training data for MAGENTA to further improve its accuracy. The accuracy of the updated MAGENTA model was then assessed through ten-fold cross validation analysis. Through cross validation analysis we found that even higher accuracies (Rank correlation R = 0.7; Methods) can be obtained if some training data involving the corresponding condition or drug is provided as input for MAGENTA (S5 Fig). We use this final MAGENTA model for predicting drug interactions in different environments and to identify genes predictive of drug interactions.
While the chemogenomic profiling data used as input to MAGENTA encompasses 3979 genes, analysis of MAGENTA model revealed that a small subset of genes was sufficient to explain most of the model’s predictive ability. The top 60, 319 and 867 genes are sufficient to predict 50, 75, and 95% of the model’s predictive ability.
Several metabolic pathways were over-represented among the top predictive genes (Table 2; S1 Table). Genes in the oxidative phosphorylation pathway showed the highest extent of over-representation among the top predictive genes (Table 2). The very high enrichment of the oxidative phosphorylation pathway relative to other cellular processes is consistent with the fact that this pathway is related to both drug sensitivity, drug-drug interactions and drug-media interaction [10,24,25]. In addition to this pathway, other top pathways were related to target processes of antibiotics like cell wall synthesis, DNA recombination and DNA mis-match repair. In addition, pathways related to drug transport and resistance were also over-represented among the top predictive genes. Analysis of top predictive genes for making three-way drug interaction predictions also revealed a significant enrichment for metabolic genes (S2 Table, S3 Table). Notably, the gene glmS, which was identified as the most predictive gene in our INDIGO model, was among the top ten predictors in the MAGENTA model as well for predicting three-way interactions and interactions across media conditions. glmS is a metabolic enzyme that catalyzes the first step in hexosamine pathway which produces precursors for cell wall synthesis and biofilm formation.
The interaction outcomes for each drug combination in a metabolic condition depends on complex interplay between many genes. Nevertheless, the presence of genes associated with specific pathways in the chemogenomic profiles of the drugs or metabolic conditions can be a strong predictor of synergy or antagonism. When the chemical-genetic interaction with the genes in these pathways change in a new condition, it influences MAGENTA predictions of drug synergy. For example, we found that presence of genes in the glycolysis and TCA cycle pathway in the drug chemogenomic profile were strongly associated with synergy (p-value = 0.001 and 0.01 for glycolysis and TCA cycle respectively, hypergeometric test). Surprisingly, we found that genes in the galactose and glyoxylate metabolism pathway were the top predictors of antagonism (p-value = 0.008 and 0.01 respectively, hypergeometric test; Table 2). Increased activity of the glyoxylate pathway has been previously found to reduce sensitivity to antibiotics in both M. tuberculosis and P. aeruginosa by decreasing the activity of the TCA cycle [6,7]. Our results suggest that drugs or metabolic conditions that increase the activity of the glyoxylate pathway may result in antagonistic interactions. The presence of the glyoxylate- and TCA cycle- pathways on opposing sides of the drug interaction outcomes supports the validity of MAGENTA in inferring the underlying mechanisms influencing drug interaction outcomes.
To further understand the underlying mechanism behind the differences in interaction outcomes between the media conditions, we compared their corresponding chemogenomic profiles and identified genes unique to each condition. If the chemical genetic interactions of the top predictive genes change significantly in a new condition, it impacts the predicted interaction scores by MAGENTA. The gene sensitivity profile in glucose condition contained significantly higher number of genes encoding transporters, two-component sensors, efflux pumps (drug resistance genes) and sugar metabolism enzymes compared to glycerol condition (S4 Table). The predominance of transporters is consistent with the fact that glucose requires active transport while glycerol can enter the cell by passive diffusion through the membrane. The differential use of transporters and efflux pumps may further alter sensitivity to drugs that require active uptake.
Since drug interactions changes across metabolic conditions depended on pathways such as glycolysis that are highly conserved across evolution, we next tested the conservation of metabolism-related drug interaction changes in clinically-relevant organisms. In our prior study we discovered that the extent of conservation of drug interaction related genes identified by INDIGO were predictive of drug-drug interaction conservation between species. This enabled us to use widely available chemogenomic data in E. coli to make predictions for pathogens such as S. aureus and M. tuberculosis that are difficult to study and lack chemogenomics data. In this study, we tested the conservation of metabolism-related drug interaction changes in the pathogen A. baumannii.
Due to rising resistance, drug combination therapy is being explored for treating A. baumannii infections [26]. A. baumannii is an opportunistic pathogen causing pneumonia, skin-, wound, urinary-tract, brain- and bloodstream infections [27]. The metabolic flexibility of A. baumannii is predicted to contribute to its persistence and colonization [27,28]. Hence, given the importance of drug combinations to treat this pathogen and its metabolic flexibility, we focused on the impact of pathogen metabolic environment on drug interactions in A. baumannii.
Genes that were orthologous between E. coli and A. baumannii were obtained from OrtholugeDB and mapped onto the MAGENTA model [29]. Overall, we found 1180 genes in A. baumannii that were orthologous with genes in the E. coli MAGENTA model. These orthologous genes were highly enriched among the top 319 drug interaction predictive genes (p-value = 7 x 10−5, hypergeometric test). Among the top predictive genes, pathways in central metabolism were conserved between the two species, while genes in lipopolysaccharide synthesis & DNA mis-match repair were not conserved (S5 Table). Thus, we hypothesized that many of the drug interaction outcomes across metabolic conditions will be conserved between the two species.
Using this MAGENTA A. baumannii model, we predicted all 15 pairwise interaction outcomes of 6 antibiotics in three media conditions we previously studied using E. coli–LB, glucose minimal and glycerol minimal media. These 6 drugs were chosen based on their efficacy in both E. coli and A. baumannii, and the availability of chemogenomics data. In addition, two drugs–amikacin and nalidixic acid, were included as they were not measured in the glucose media training data for MAGENTA. The 45 drug interactions across three media conditions, including 13 novel interactions involving the two new drugs will allow us to assess the accuracy of MAGENTA for predicting interactions in a new organism involving drug-media combinations that it was not trained on.
Overall, the predicted drug-drug interaction scores by MAGENTA across the three media conditions significantly correlated with the measured interaction scores (Rank correlation R = 0.57, p-value = 0.0001; Fig 4). The interaction scores for E. coli and A. baumannii showed significant correlation for the 32 drug interactions that were experimentally measured in both species (Rank correlation R = 0.57, p-value = 5 x 10−5). This suggests that drug interactions are relatively conserved between the two species and is consistent with our observation that top drug interaction predictive genes in MAGENTA were significantly conserved between the two species. However, even for cases where there is no E. coli interaction data available in that specific metabolic condition, MAGENTA can accurately predict interaction outcomes with equally high correlation (R = 0.57) as demonstrated here using the drugs amikacin and nalidixic acid.
While majority of antibiotic combinations that were synergistic in E. coli were also synergistic in A. baumannii, combinations of amikacin and tetracyline, showed synergy in A. baumannii but were antagonistic in E. coli in LB media. In contrast, combination of amikacin and ampicillin were synergistic in E. coli in glycerol media but not in A. baumannii. Importantly, MAGENTA correctly identified combinations that differed between the two species. The predicted extent of difference in interaction outcome correlated significantly with the observed extent of interaction change (rank correlation R = 0.59, p-value = 2 x 10−5).
We also observed significant difference in interaction outcome for ampicillin and aztreonam, which target cell wall synthesis. For instance, combination of aztreonam with ampicillin was synergistic in A. baumannii but antagonistic in E. coli in glucose media. This observation is in agreement with the lack of conservation of the cell wall lipopolysaccharide (LPS) synthesis pathway between the two species. While LPS synthesis is essential in E. coli, it is dispensable in A. baumannii [30].
Some drug combinations, such as amikacin-tetracycline and ampicillin-aztreonam showed robust synergy across all three conditions in A. baumannii. These combinations might serve as promising leads for treating A. baumannii. To discover other combinations in both A. baumannii and E. coli with broad spectrum metabolic synergy, we predicted interaction outcomes for 2556 pairwise drug combinations involving 72 drugs across 9 growth conditions for which we had chemogenomics data (Fig 5). These 9 metabolic conditions—namely growth in glucose, glucosamine, glycerol, acetate, maltose, succinate, ethanol minimal media, aerobic and anaerobic growth in LB, represent a wide spectrum of potential metabolic states for the bacteria. In vivo metabolic conditions span growth in diverse substrates such as sugars, nucleotides, glycerol, lipids and hypoxic conditions [12] and these 9 metabolic conditions studied here are representative of the conditions in vivo.
The nine metabolic conditions uniquely impacted drug interaction outcomes. Growth in anaerobic condition had a very strong and distinct impact on drug interactions compared to other conditions (Fig 5). Given that oxidative phosphorylation was the top predictive pathway for drug interactions across media conditions, growth conditions that change the activity of this pathway exert strong influence on interaction outcomes. In E. coli, we identified 119 combinations out of the 2556 combinations screened that were synergistic across all metabolic conditions. For example, we found that combinations of azithromycin and rifampicin, and ampicillin and chloramphenicol were synergistic across all 9 metabolic conditions (S6 Table). Several combinations used clinically also showed robust synergy across these conditions. The list of 119 combinations included combinations of rifampicin with tetracycline, ampicillin, azithromycin and clarithromycin. These combinations with rifampicin are frequently used for treating biofilm associated infections [31]. In addition to rifampicin, the antibiotics—ampicillin, vancomycin and fusidic-acid were also over-represented in the list of 119 combinations with robust synergy.
Analysis of drug interaction landscape in A. baumannii revealed 19 combinations that showed synergy across all the growth conditions (Interaction score < -0.5, S6 Table). This represents just 0.74% (1 out of 134) of the total combinations screened. Thus, MAGENTA can potentially reduce the search space by 100-fold compared to a trial and error approach or a blind screen.
The accurate estimation of drug interactions in A. baumannii was possible because of the conservation of top predictive genes in the E. coli MAGENTA model in A. baumannii. To assess the generalizability of this approach, we compared the conservation of the top predictive genes in the pathogens S. aureus and M. tuberculosis. Top genes predicted by MAGENTA were enriched for those that are conserved between the two species. We found a significant enrichment for orthologs of S. aureus (1 x 10−5) and M. tuberculosis (2 x 10−8). This suggests that we can apply MAGENTA model to make accurate predictions across metabolic environments in these systems as well.
We have predicted interaction outcomes for 2556 pairwise drug combinations involving 72 drugs across 9 growth conditions for these two organisms (S9 Fig). Analysis of drug interaction landscape in these two pathogens revealed 113 and 108 combinations that showed strong synergy across all the growth conditions in M. tuberculosis and S. aureus respectively (Interaction score < -0.5, S7 Table). Of note, we identified robust synergistic interactions involving the frontline drugs used for treating Tuberculosis. We found that the antibiotics azithromycin and fusidic acid had robust synergy with the Tuberculosis drugs rifampicin and isoniazid respectively. Overall, this dataset can be used to prioritize combinations effective in specific growth conditions and potentially used to identify metabolically-robust drug combinations.
In this study we found that the pathogen metabolic environment significantly modulates drug combination efficacy. This trend was observed across nine different metabolic conditions and several antibiotics spanning various target processes; each metabolic state had a distinct and unique impact on each drug. This observation greatly complicates the search for finding effective therapies, given the wide range of metabolic conditions that pathogens encounter in vivo or in biofilms. The differences in sensitivity between metabolic conditions may also explain the differences in efficacy in vitro and in vivo observed for drugs [11].
To address this challenge, in this study we developed a computational approach (MAGENTA) to predict how metabolic environments can impact drug combination efficacy. An important finding from this study is that interactions between drugs across metabolic conditions can be predicted based on chemogenomic profiles of the individual drugs and the metabolic perturbation. This suggests that metabolic stress can be modeled using a similar framework as drug induced stress. Our approach can be potentially extended to other stressors including antimicrobial proteases and toxins [32,33].
Multi-drug combinations can greatly reduce the rise of resistance and enhance potency compared to single agents [34]. However, the number of possible permutations increases exponentially with the number of drugs in a combination regimen; this greatly underscores the need for computational tools like MAGENTA to identify most synergistic combinations. Prior studies on predicting drug interactions using chemogenomics focus only on pairwise drug combinations; here we demonstrate that the MAGENTA approach accurately predicted interaction outcomes involving multi-drug combinations. Multi-drug interaction predictions are especially relevant for diseases like tuberculosis, where combinations of 4 antibiotics are commonly used for treatment.
Theoretical models suggest that pairwise combinations of underlying drugs can be used to predict triplet combinations [35,36]. To make predictions for all 3-way combinations of n drugs, only n chemogenomic profiles are needed along with a small training data set for MAGENTA, but (n choose 2) pairwise interactions are needed for the pairwise approach. MAGENTA is hence more effective in exploring large number of combinations. Furthermore, MAGENTA can make predictions for combinations with new drugs that it was not trained on using chemogenomics data. In our case, three out of the 8 tested drugs are not part of the training set. Out of the 56 triplet combinations, the pairwise approach can be used to make predictions for only 10 triplets using pairwise data in the training set.
Notably, MAGENTA was able to predict interaction outcomes in a new metabolic condition (glycerol) based on training data in glucose and LB media. This result corroborates the ability of MAGENTA to extract mechanistic features that influence drug interactions from chemogenomic profiles, such as the role of metabolic pathways. Our unbiased data-driven approach confirmed the importance of oxidative phosphorylation and cellular respiration pathway on antibiotic efficacy. Reducing respiration is known to inhibit bactericidal drug lethality [10]. The oxidative phosphorylation pathway likely affects antibiotic efficacy in multifarious ways including oxidative stress, redox homeostasis and facilitating drug import. Our analysis also revealed the opposing association of TCA cycle and glyoxylate pathway with drug synergy and antagonism respectively, supporting previous studies [6,7,10,37].
Interactions involving combinations of bactericidal and bacteriostatic drugs showed striking differences between different metabolic conditions. This is consistent with the sensitivity of this drug combination to cellular metabolic state [10]. While this combination is avoided in the clinic due to antagonism [9], our results suggest that this combination is not antagonistic during growth in glycerol and becomes strongly synergistic during growth in minimal glucose media. Antagonism between combinations of bactericidal and bacteriostatic drugs is predicted due to their opposing effects on cellular respiration and TCA cycle. Hence shifting E. coli to a growth condition with higher activity of TCA cycle will have a significant impact on respiration and efficacy of these drugs. This change in nutrient source will shift the balance in favor of one class of drugs over the other. Our data revealed that metabolic state not only influences combination involving both bacteriostatic and bactericidal drugs, as previously believed, but it also strongly influences combinations involving only bactericidal or bacteriostatic drugs.
The nine metabolic conditions studied here are representative of in vivo metabolic conditions such as the gut environment and biofilms. Our study takes the first step towards rational design of combination therapies that are robust to the in vivo environment. While we have focused on the impact of a single metabolic perturbation on antibiotics in this study, the in vivo environment is complex and dynamic. Future modeling efforts that expand the capability of MAGENTA to dynamic conditions can enable prediction of effective therapies.
Our analysis of the drug sensitivity landscape revealed that metabolic environments had a significant impact on the efficacy of drug combinations (Fig 5). Nevertheless, by searching through 2556 combinations, we identified a small subset that were synergistic across all metabolic conditions in both E. coli and A. baumannii. Such synergistic drug combinations are urgently needed for treating A. baumannii infections. This opportunistic pathogen is a frequent cause of drug resistant wound-, urinary tract- and pneumonia infections. It is responsible for 2–10% of all Gram-negative hospital infections [38]. A. baumannii infections display resistance to most antibiotics used in the clinic and new treatments are desperately needed [26].
A key limitation of our approach is the need for chemogenomic profiling data. However, with the development of single gene knockout libraries, chemogenomic profiling data is increasing in number. For instance, the Resistome database has a compendium of chemogenomic profiles for 230 different perturbations in E. coli [39]. Similar large compendiums exist for S. cerevisae [40]. This approach could hence be potentially applied to a wide range of drugs in both prokaryotic and eukaryotic systems. Furthermore, our theoretical framework can be extended for discovering effective anti-cancer drug combinations. Drug combinations are frequently used in cancer chemotherapy to reduce resistance [41–44]. Our approach can enable the identification of robust combination therapies targeting the tumor microenvironment.
While our validation data sets tested the algorithm’s predictive ability involving novel drugs and conditions, through cross validation we found that higher accuracies (Rank correlation R = 0.7) can be obtained if some training data is provided as input for the corresponding condition or drug. An optimal experimental design should involve sparse sampling of multiple drug combinations and metabolic conditions rather than exhaustive combinations of a few drugs in one condition. The use of defined media rather than complex undefined media containing yeast extract or serum can also greatly improve modeling efforts. Knowledge of in vivo metabolic environments can enable direct prediction of effective therapies.
In sum, our study demonstrates that metabolic environment can elicit significant effects on antibiotic combination efficacy. Our new approach MAGENTA goes beyond existing drug combination discovery platforms by predicting the impact of metabolic state on combination therapies. Further, we were able to leverage existing chemogenomic and drug interaction data in E. coli to infer antibiotic interactions across metabolic conditions in the pathogen A. baumannii. Our approach can enable identification of robust therapies tailored to the pathogen and the metabolic environment.
All drug interaction experiments were conducted using the diagonal method [36,45]. For each drug, we constructed a linearly increasing concentration range of 14 concentrations from 0 drug to a Minimum Inhibitory Concentration (MIC), which results in almost complete inhibition. For pairwise or 3-way interactions, we constructed a similar dose series, with the top concentration MIC/2 or MIC/3 of each constituent drug. For each pairwise drug interaction assay, sensitivity to linearly increasing doses (dose-response) were collected for two single drugs and a 1:1 mixture of two drugs. For each three-way drug interaction assay, dose-responses were collected for three single drugs and a 1:1:1 mixture of three drugs.
Cell growth in these two or three drug mixtures were compared to growth in single drug components to calculate FIC values. For this, we located the dose fraction that gives the same level of inhibition in the single drug or combination dose-responses. The dose that results in a defined inhibition level is divided by the “expected dose” that would give the same inhibition if the drugs in the combination were same drugs, resulting in the Loewe Additivity based FIC drug interaction measure. For example, when considering a 3-way interaction, if the 50%, 60% and 70% of the top dose of drugs A, B and C result in IC50, then the expected IC50 was defined as ~0.6. If the dose fraction of the combination that gives IC50 is 60%, this 3-way combination is additive. If it is smaller or larger than 0.6, it is synergistic or antagonistic, respectively. The FIC for a drug combination is obtained by adding the fractional MIC of each drug in a combination. The fractional MIC is calculated by dividing the dose of a drug when used in combination by the MIC of that drug when used individually. FIC is equal to 1 if drugs are additive, less or more than 1 if drugs are synergistic or antagonistic, respectively. In this study, we used the log2 of FIC values as drug combination interaction values. We have used a quantitative metric rather than discrete classification of interactions for two key reasons. Firstly, use of quantitative interaction scores allows for a quantitative validation of model predictions. Secondly, regression algorithms perform better with a continuous range of values.
For FIC calculation, we used the drug or mixture dose that resulted in 70% inhibition (IC70) throughout the analysis. This choice was guided by our initial analysis which showed this inhibition level results in the highest reproducibility among replicates (S10 Fig, S11 Fig, S12 Fig). We also show that the scores are robust to the IC choice, as scores obtained using different ICs in the range of IC60 and IC80 highly correlate (r > 0.8) (S11 Fig, S12 Fig).
Escherichia coli MG1655 and Acinetobacter baumannii Bouvet and Grimont ATCC 17978 were used as bacterial strains. All drugs were purchased from Sigma. MICs for each drug are provided in Table 1. We defined drugs as bacteriostatic or bactericidal based on annotation from Nichols et al [19]. All pairwise drug interaction experiments were done in duplicate, with correlation of 0.8 and 0.86 for E. coli and A. baumannii experiments, respectively (S10 Fig). We used the arithmetic average of two replicates as the drug interaction score for each pair.
LB media was prepared by dissolving 20g LB powder (Sigma) in 1l of water and autoclaving. Minimal media was prepared as final concentration of 1X M9 salts, 2uM MgSO4, 0.1uM CaCl2 and carbon source in water, and filtered for sterilization. A final concentration of 0.04% Glucose or 0.08% Glycerol was made, which makes the carbon resource levels equivalent in two media. 5ml bacteria cultures were grown in 15ml breathable culture tubes for 16 hours, by mixing 20ul of 25% glycerol stock of cells at OD600 = 1 and 5ml respective growth media, and shaking at 250RPM at 37C. After overnight growth, cells were diluted to OD = 0.01 and used as inoculums for the drug interaction experiments.
Drugs were dissolved in dimethyl sulfoxide and stored at −20°C. Nanoliter volumes of drugs and their combinations were printed on 384-well plates using a digital drug dispenser (D300e Digital Dispenser, HP). All drug sensitivity quantifications were done by measuring the optical density (OD600) of 50ul bacteria grown for 16 hours at 37C without shaking in 384-well plates (Synergy HT, BioTek). Dispense locations were randomized within each plate to minimize plate position effects. Plates were sealed with aluminum plate seals and incubated without shaking at 37°C. After data collection, plate data were reconstructed from randomized positions for further analysis.
The entire series of steps to predict drug interactions using MAGENTA is described in S4 Fig. The inputs for MAGENTA were chemogenomic profiles of drugs and media conditions from Nichols et al [19], and log2 transformed drug interaction FIC scores for the training set. Chemogenomic data was quantile-normalized using the quantilenorm function in MATLAB. Interactions with chemogenomics fitness score less than -2 (two standard deviations below the mean) or greater than +2 were chosen to be significant and used as input to MAGENTA.
MAGENTA represents each drug in silico as a function of its corresponding drug-gene interactions inferred from chemogenomic profiling. MAGENTA assumes that cellular response to a combination of stressors can be represented as a linear combination of cellular response to individual stressors, as observed in prior studies [46,47]. This enables it to predict drug-drug interaction outcomes across metabolic conditions from individual chemogenomic profiles using the machine learning algorithm–Random forests. The random forest algorithm creates an ensemble of decision trees and outputs the mean prediction of the individual trees [48,49]. We used the RandomForest toolbox in MATLAB. The regression random forest algorithm was used with default parameters—500 trees (default) and number of variables sampled (default value–N/3, where N is the number of variables).
In addition to the test-set predictions, MAGENTA’s predictive ability was also assessed through tenfold cross-validation. In tenfold cross-validation, 10% of the interactions were randomly blinded and predicted by the model based on information from the remaining 90% of the interactions (S5 Fig). Through tenfold cross validation, we found that MAGENTA could accurately predict interactions with compounds that belong to novel chemical classes or with distinct mechanisms of action. Nevertheless, we found that the prediction accuracy could be further improved by choosing drugs and metabolic conditions representative of different classes in the training set.
For predicting interactions in A. baumannii using the orthology mapping approach, orthologous genes in E. coli were obtained from OrtholugeDB. 1633 genes were predicted to be orthologs of A. baumannii among the E. coli genes based on the reciprocal-best-BLAST-hit procedure.
The top genes predicted by MAGENTA to account for 75% of the model’s predictive ability were used for pathway enrichment analysis. KEGG annotations for E. coli were downloaded using the R Bioconductor GAGE Package. All statistical tests of correlation and overlap, and Multi-dimensional scaling analysis were performed in MATLAB. The MATLAB implementation of MAGENTA along with associated experimental data are provided as supplementary materials.
|
10.1371/journal.pcbi.1001009 | Simultaneous Genome-Wide Inference of Physical, Genetic, Regulatory, and Functional Pathway Components | Biomolecular pathways are built from diverse types of pairwise interactions, ranging from physical protein-protein interactions and modifications to indirect regulatory relationships. One goal of systems biology is to bridge three aspects of this complexity: the growing body of high-throughput data assaying these interactions; the specific interactions in which individual genes participate; and the genome-wide patterns of interactions in a system of interest. Here, we describe methodology for simultaneously predicting specific types of biomolecular interactions using high-throughput genomic data. This results in a comprehensive compendium of whole-genome networks for yeast, derived from ∼3,500 experimental conditions and describing 30 interaction types, which range from general (e.g. physical or regulatory) to specific (e.g. phosphorylation or transcriptional regulation). We used these networks to investigate molecular pathways in carbon metabolism and cellular transport, proposing a novel connection between glycogen breakdown and glucose utilization supported by recent publications. Additionally, 14 specific predicted interactions in DNA topological change and protein biosynthesis were experimentally validated. We analyzed the systems-level network features within all interactomes, verifying the presence of small-world properties and enrichment for recurring network motifs. This compendium of physical, synthetic, regulatory, and functional interaction networks has been made publicly available through an interactive web interface for investigators to utilize in future research at http://function.princeton.edu/bioweaver/.
| To maintain the complexity of living biological systems, many proteins must interact in a coordinated manner to integrate their unique functions into a cooperative system. Pathways are typically constructed to capture modular subsets of this dynamic network, each made up of a collection of biomolecular interactions of diverse types that together carry out a specific cellular function. Deciphering these pathways at a global level is a crucial step for unraveling systems biology, aiding at every level from basic biological understanding to translational biomarker and drug target discovery. The combination of high-throughput genomic data with advanced computational methods has enabled us to infer the first genome-wide compendium of bimolecular pathway networks, comprising 30 distinct bimolecular interaction types. We demonstrate that this interaction network compendium, derived from ∼3,500 experimental conditions, can be used to direct a range of biomedical hypothesis generation and testing. We show that our results can be used to predict novel protein interactions and new pathway components, and also that they enable system-level analysis to investigate the network characteristics of cell-wide regulatory circuits. The resulting compendium of biological networks is made publicly available through an interactive web interface to enable future research in other biological systems of interest.
| The complexity of cellular activity is driven not only by interactions among genes and gene products, but also by the timing and dynamics of these interactions, the conditions under which they occur, and the many forms that they can take. Proteins interact in many different functional manners with multiple partners - physically in complexes[1] and through modifications[2], [3], synthetically when employed in parallel pathways[4], and in regulatory roles as activators or repressors[5] - and these interaction types combine to form complete molecular pathways. Functional assays such as gene expression, localization, and binding each capture individual aspects of this molecular activity at a global level, but translating the vast amount of resulting genomic data into specific hypotheses at the molecular pathway level has proven challenging. The heterogeneity of gene interactions within each pathway has compounded this difficulty by preventing any one assay from providing a complete biological picture. It is thus critical to integrate large genomic data collections to describe not only the membership of gene products within pathways, but also their construction from the building blocks of individual types of biomolecular interactions.
In this work, we provide the means for investigators to study complete molecular pathways at a whole-genome level as generated from integrated functional genomic data. First, we relate 30 general and specific biomolecular interaction types, such as transcriptional regulation, ubiquitination (and other post-translational modifications), or protein complex formation, in an ontology of interaction types. This ontology is hierarchical, in that a phosphate transfer is perforce a covalent post-translational modification, which is in turn by definition a transient physical interaction, and so forth. Next, we combine this ontology with Bayesian hierarchical classification methodology [6], enabling the simultaneous prediction of genome-wide interaction networks of all of these 30 types from integrated heterogeneous experimental data. Finally, we apply this method to a compendium of ∼3,500 Saccharomyces cerevisiae experimental conditions, experimentally validating several of the resulting predictions in glucose utilization, DNA topological maintenance, and protein biosynthesis as described below. This methodology ensures that investigators can take advantage of all available data to accurately identify the entire range of functional interaction types within specific pathways and across an organism's genome.
It is important to contrast this genome-wide system for predicting diverse biomolecular interaction types with previous work predicting specific individual interaction networks. A variety of methodologies have been proposed for inferring regulatory networks [7]–[10], physical interaction networks [11], [12], synthetic interaction networks [13], [14], and other interaction types [15], generally from their respective primary data types (ChIP-chip and -seq, proteomics, double knockouts/knockdowns, etc.) Likewise, other methods have been proposed for heterogeneous genomic data integration [16]–[24], but these almost uniformly focus on either general functional interactions or on specific bimolecular interaction types. This work combines the strengths of these two bioinformatic areas, providing a simultaneous platform with which all data available for a system can be integrated and focused onto specific interaction types, genome-wide and for individual gene products.
We first applied our yeast network compendium to explore two cellular processes, carbon metabolism and cellular transport. This generated many promising interactions involving Snf1, Cmk2, Glc7, Adr1 and Gph1 supported by recent published work. We also suggest several novel pathway connections, such as the interplay between the glycogen breakdown and glucose utilization pathways, by systematically layering multiple different interaction types. To experimentally validate a collection of our predicted yeast interactions, we focused on the synthetic lethal interactions, where double knockouts result in lethality, predicted among proteins involved in DNA topological change and regulation of protein biosynthesis. Highly ranked 20 protein pairs, 10 pairs from each pathway, were hypothesized to be synthetically lethal, and we experimentally confirmed 14 of these pairs (70%). Furthermore, we evaluated our posttranslational modification predictions based on recent experimental results on 173 protein pairs, resulting in a prediction AUC over 0.8. In an analysis of the systems-level global and local network structures of our interactomes, we observed differential usage of several recurring subgraphs, providing insight into the functional design principles of pathway components. Finally, we provide a web-based interface to explore all 30 yeast interaction networks at http://function.princeton.edu/bioweaver. This will allow investigators to interactively survey and generate hypotheses from the diverse interaction types comprising the S. cerevisiae cellular circuitry.
We present a general methodology for integrating large, diverse genomic data compendia to simultaneously predict multiple biomolecular interaction network types (physical, genetic, regulatory, etc.; Figure 1). We applied this methodology to ∼3,500 S. cerevisiae experimental conditions to generate 30 whole-genome networks describing predicted gene and gene product interactions in yeast. We first evaluated these predictions quantitatively using cross-validation, achieving AUCs over 0.7 for most interaction types. More qualitatively, we examined a set of predicted molecular linkages of diverse types between glycogen breakdown and glucose utilization genes, which were validated by recent literature. Finally, we experimentally confirmed 14 of 20 predicted novel synthetic lethal interactions in the DNA topological change pathway.
We predicted a compendium of biomolecular interaction networks by integrating diverse yeast genomic data using a multi-label hierarchical classification system ([6], Figure 2A). As briefly outlined in Figure 1, we first independently predict each interaction type using specifically trained SVM classifiers. Next, it is desirable to avoid making inconsistent interactome predictions due to noisy data, e.g. predicting that two genes share a regulatory relationship without occurring within the same pathway. In order to share information among classifiers for related interaction types in a principled manner, each SVM's predictions are treated as noisy observations. The final set of labels for each gene pair is then derived by finding the maximum likelihood assignment of interaction labels by integrating these observations in a Bayesian graphical model.
Based on ∼30% heldout test data, our average AUC over all 30 interaction types was 0.79, with minimal variations in performance across the interaction ontology (Figure 2A, Figure 1 in Text S1). The most general interaction type, functional relationship, also incurred the lowest AUC of 0.63, which remains comparable to state-of-the-art functional interaction prediction systems [25]. In order to quantify the contribution of our hierarchical Bayesian system relative to predicting disparate biomolecular interaction types in isolation, we compared the accuracy of each individual SVM classifier to that of the complete system. For all 30 predicted interactomes, the Bayesian hierarchy showed increased AUC scores, averaging +0.076 and ranging from a minimum of +0.011 to a maximum of +0.166. For example, posttranslational regulation improved from 0.61 to 0.77, while phosphorylation increased from 0.67 to 0.79. (full ROC curves for all interaction networks can be found in Text S1). In combination, these two evaluations suggest that this methodology can accurately leverage large genomic data collections to simultaneously infer a diversity of genome-wide interaction networks.
Many gene interactions are directional and thus asymmetric, e.g. phosphorylation or ubiquitination, in which the two interactors take on distinct source and target roles. It is thus important to correctly infer not only the presence or absence of these directed interactions, but also the correct directionality. Specifically, for each directed interaction type, we constructed a list of all edges ranked by predicted probability; we then compared the rank of the correct interaction direction relative to the incorrectly flipped interaction between the same two genes (Figure 2 in Text S1). Using this as a true- and false-positive rate criterion, we were able to predict the correct direction of gene interactions with average AUC of 0.85 over the 12 directed networks (maximum 0.94, minimum 0.70). This indicates that this methodology can accurately recover not only overall pathway structure, but also the upstream and downstream effects of individual gene products within molecular pathways.
Simultaneous inference of biomolecular networks for many different interaction types allows the generation of very specific novel hypotheses. As a first example, we detail a combination of transcriptional, genetic, post-translational, and metabolic interactions among gene products coordinating glycogen breakdown and glucose utilization in yeast.
As shown in Figure 3, Adr1 is an important transcription factor involved in carbon metabolism in Saccharomyces cerevisiae. It has many known regulatory inputs [26], one of which is the glucose-responsive kinase Snf1, and what proteins transmit this regulatory information has been under investigation for some time. By examining different classes of predicted interactions with Adr1 and other proteins not in our gold standard (Figure 3A), we first identified regulatory and genetic interactions between the protein phosphatase Glc7 and Adr1. Specifically, our prediction of a synthetic alleviating interaction between Glc7 and adr1 mutants places it upstream of Adr1 in this pathway. This combination of interactions is almost always associated with an upstream inhibitory regulator, consistent with the known biological role of Glc7 as a protein phosphatase that removes activating phosphorylations [27].
The predicted yeast networks also hypothesized post-translational regulatory interactions between Cmk2 and both Adr1 and Gkc7 (Figure 3A). This three-protein network creates a feed-forward regulatory motif in which Cmk2 simultaneously activates Adr1 as well as its inhibitor Gkc7, creating a time-delayed inactivation of Adr1. These interactions are supported by a recently published paper [26] linking the calmodulin- and Snf1-dependent pathways to Adr1 regulation. Our predicted pathway takes these results a step further by identifying which of the three calmodulin-dependent kinases (Cmk2) is responsible [28]. Finally, a novel metabolic interaction was predicted between Adr1 and Gph1, the only high scoring interaction of its type for Adr1. Like Adr1, Gph1 is involved in glucose metabolism by glycogen breakdown, and both are regulated by the metabolites glucose and cAMP [29]. A metabolic interaction between Adr1 and Gph1, combined with the known regulation of these genes by glucose and cAMP, suggests that coordinated regulation is occurring between the glycogen breakdown and glucose utilization pathways and is transcriptionally controlled by Adr1.
Protein sorting and trafficking is an essential function of eukaryotes and requires numerous multi-subunit complexes to ensure the proper localization and secretion of proteins (Figure 3C, [30]). At the early stages of this process, the two major transport pathways from the endoplasmic reticulum (ER) to the Golgi are governed by the SNARE and COPI complexes [30]. We predicted synthetic interactions between these pathways (e.g. synthetic aggravation for Arf1-Sec18 and synthetic alleviation for Sec27-Uso1) that are supported by known genetic interactions[31], [32]; Arf1 and Arf2 are a representative example, as they are considered functionally redundant GTPases, and each COPI complex contains either Arf1 or Arf2 [33].
Later in the pathway, Bch1 is a member of the ChAP family of proteins, which direct cargo bound to COPI complexes in the Golgi to their destinations such as the plasma membrane [34]. We predict a physical interaction between Bch1 and the COPI complex that is well established in the literature but was not part of our gold standard. Likewise, Vps1 serves a similar function for vacuole targeting [35], and our predictions of its physical and shared pathway interactions with COPI are supported by the literature [34].
Novel hypotheses in Figure 3C include the predicted physical interaction between Bch1 and Vps1, suggesting competition between the Sec27-Arf1 and Vps1 complexes for the Bch1 sorting function (also supported by a metabolic interaction between Sec27-Arf1 and Vps1). Both Vps1 and Arf1 are GTPases that must hydrolyze GTP to perform their roles in protein sorting [33]. Thus, this predicted pathway hypothesizes a competition between the Arf1 GTPase and Vsp1 GTPase for Bch1 that is likely regulated by GTP availability. Similarly, the uncharacterized membrane-bound protein YDL012c is placed in the same pathway as Vps1, suggesting that the former may be involved in regulating Vps1 activity. By highlighting just a few of our predicted interactions in the protein sorting pathway, we demonstrate the potential for generating hypotheses used to drive novel biological discoveries.
To experimentally evaluate the accuracy of a subset of our predicted interactions in a directed manner, we focused on the DNA topological change and protein biosynthesis regulation processes in S. cerevisiae [36]. 20 synthetic lethality interactions predicted with high probability were experimentally tested using SGA technology [4], [13], with the results summarized in Figure 4. 14 gene pairs (70%) were validated, 8 involved in DNA topological change and 6 in the regulation of protein biosynthesis. Several of the remaining 6 unconfirmed interactions may be synthetic lethal under different conditions. For example, GCS1 and SLT2 deletions both individually decreased resistance to ethanol stress [37], and similar conditions might elicit synthetic lethality. Based on a total of ∼100,000 pairs estimated to have been synthetically lethal in yeast of a possible ∼18 million (0.05%) [13], our predictions are a clear improvement over the baseline rate for novel discovery.
As an additional evaluation, we collected 24 recent publications containing a total of 173 experimentally confirmed post-translationally regulated protein pairs (see Text S2 for the list of publications). None of these interactions was present in our training standard. Evaluating on this set, our Bayesian hierarchical system achieved an AUC of 0.802, demonstrating its ability to accurately predict novel, experimentally verifiable post-translational regulation interactions on a whole-genome scale. This accuracy is comparable to our initial cross-validation AUC of 0.778, indicating that our evaluation provides a reasonable estimate of the expected experimental verification rate for novel predictions.
This rich compendium of inferred interaction types provided an opportunity to analyze systems-level network features genome-wide at multiple levels of biomolecular activity. In particular, we examined the network structural characteristics that potentially help to define the functional roles of each interactome. Biological networks have been proposed to exhibit a scale free topology [38], implying a power-law degree distribution. Previous studies have detected such distributions based on partial networks and single interactomes [39]. Here (Figure 5A), we observe a scale-free degree distribution very robustly in all 30 interaction types. However, the high-degree hubs in each interactome do differ, reflecting the distinct functional activities carried out by each network type. To verify this, we analyzed the extent of the overlap of high-connectivity genes (in the top 5% of the degree distribution) between the networks for each pair of interactomes (Figure 5B; directed interactomes were divided into separate in- and out-degree comparisons). The major clusters show distinct functional similarity, correctly reflecting the similarities captured by our interaction ontology: transient and nontransient physical interactions each group together, synthetic interactions cluster, and so forth. Beyond confirming the structure of the ontology, this also captures relationships such as the sharp divide between regulatory in- and out-degree (the most regulated genes are not themselves high-level regulators with many targets) and a tendency for regulatory hubs to incur more synthetic interactions than expected.
Degree distribution captures a global description of each network, while analysis of small recurring subgraphs has been proposed to describe local network properties [40], [41]. We investigated the enrichment of two types of subgraphs, network motifs and graphlets, in our interactomes. First, network motifs are small directed subgraphs that have been found to recur in a growing number of organisms [42]–[44]. In our 12 directed interaction networks, the feed forward loop motif showed significant enrichment (relative to a random background; see Text S1) consistent with previous studies on the yeast transcription factor network [41]. Feed forward loops are known to accelerate or delay the response of a input signal [45], suggesting in this context a much wider usage of dynamic information processing than has been previously reported in regulatory networks[46]–[48].
A second approach to exploring the local structure of biological networks is to examine graphlet degree distributions [40]. Graphlets are small non-isomorphic subgraphs, and a graphlet's degree for a given node is defined as the number copies of that graphlet to which it is incident. For example, the number of triangle motifs touching a particular node represents its 3-node graphlet degree. Compared to network motifs, for which enrichment can be difficult to detect due to the complexity of the baseline null distribution[49], graphlet analysis may have a higher sensitivity towards infrequent subgraphs. Thus, as a complementary analysis, we computed the graphlet degree distributions for all two to five node graphlets for the 13 specific leaf node interactomes in our interaction ontology (Figure 5C). We compared the graphlet degree distributions between these interactomes, demonstrating a clear divergence in the local network structure between subclasses of metabolic, regulatory and synthetic interactions. Unlike the comparison of high-degree genes, this also captures unexpected similarities between disparate interaction types: phosphorylation and ubiquitination, for example, are siblings in the interaction ontology and represent comparable mechanisms of post-translational modification. The former's local network topology is more similar to that of synthetic interactions, however, while ubiquitination is more strongly regulatory. This differentiating pattern between ubiquitination and phosphorylation provides a base for intriguing network hypotheses for further investigation. One potential explanation could be due to the differing mechanistic activities where ubiquitination is most often employed exclusively as a regulatory mechanism to degrade active proteins, whereas phosphorylation serves both regulatory and dynamic information processing roles [50].
The increasing abundance of genomic data has opened up countless new possibilities for systems-level biological perspectives, but its increasing complexity impedes the understanding of specific cellular circuitry at a mechanistic level. Here, we provide a method with which very large experimental data compendia can be integrated to predict 30 specific biomolecular interaction types at a genome-wide scale. By applying this to more than ∼3,500 experimental conditions in yeast, we have evaluated these predictions at an average AUC of 0.79, validated 70% of experimentally tested synthetic lethal interactions, and proposed novel transcriptional, genetic, post-translational, and metabolic interactions in the yeast carbon metabolism and cellular transport pathways.
As described above, the investigation of specific S. cerevisiae biology in the processes of glucose utilization and protein trafficking demonstrates the use of these interactomes to reconstruct complete pathways. In many instances, experimental biologists are faced with the task of designing experiments to target a specific set of genes. By simultaneously hypothesizing all types of biomolecular interactions in which a group of gene products may be involved, this methodology can be used to select both the proteins to be assayed and the assays that may be most informative. Prior approaches inferring these interaction types in isolation mask this information and may even be inconsistent; how might a biologist interpret predictions that two proteins physically interact, but that they are not part of the same pathway? Such inconsistencies are avoided by simultaneous ontology-based inference, allowing underlying experimental data to be integrated into a consistent description of a cellular system.
To our knowledge, there has been no other method that simultaneously enables researchers to leverage high-throughput data in an interaction-type-specific manner within an ensemble setting. Successful focused attempts to predict specific interaction types have shown comparable AUCs to our results [51], [52], which could be incorporated into a framework like this as base classifiers during future work (instead of the SVMs utilized in this study). Recent “functional coupling” predictions [20] are also related, but fall short of pathway-level interaction predictions, mainly due to a lack of the crucial directional information needed to infer bimolecular pathways. These frameworks typically also do not resolve inconsistencies among predicted interaction type labels that can hinder pathway reconstruction and experimental follow up.
Ultimately, compendia of inferred interaction networks can be used to explicitly construct and understand distinct cellular pathways. By investigating and confirming different interaction types suggested by our system, investigators can stitch together both new pathways and new interconnections between existing ones. This process can be applied in any organism for which diverse genome-scale data is available - a situation that is only becoming more common. We believe that our work can leverage this diversity of experimental results that might otherwise be underutilized, helping to spur new functional discoveries in organisms beyond yeast. Finally, all of our predicted networks are made publicly available through an interactive tool at http://function.princeton.edu/bioweaver for investigators to explore their own biological areas of interest.
We developed an integrated method for concurrently predicting multiple protein interaction types. This method integrates large and diverse collections of functional genomic data in the context of a biomolecular interaction ontology. Each gene interaction type in the ontology is first predicted using an SVM classifier by integrating ∼3,500 experimental conditions from expression, colocalization, regulatory, and other yeast experimental data (withholding data types directly related to the interaction type being predicted; see below). These isolated interactomes are then reconciled using a hierarchical Bayesian framework to obtain the most probable set of consistent labels for each gene pair within the hierarchy of our interaction ontology. Using this system, we generated 30 S. cerevisiae interactomes, with which we validated several mechanistic interaction predictions in carbon metabolism, cellular transport, and 14 new synthetic lethal interactions in DNA topological change and protein biosynthesis.
We constructed an interaction ontology focused on categorizing gene pair relationships. This is similar in spirit to the Gene Ontology (GO) [36], which curates individual proteins' molecular functions, biological roles, and subcellular localizations. Our interaction ontology contains a total of 124 terms and integrates information from existing interaction catalogs [53], [54], the EBI [55], and SGD [56]. The ontology's three major branches are metabolic, interaction pathway, and physical interactions. Metabolic interactions describe protein pairs linked in metabolic pathways, such as isoenzymes or enzymes that catalyze adjacent reactions. Physical interactions include covalent or non-covalent binding, e.g. stable complexes or transient post-translational modifications. Pathway interactions include more conceptual relationships between genes in a pathway, such as regulation or synthetic interactions. We selected the 30 nodes in our interaction ontology with more than 70 annotations (as described below) to include in this evaluation, and the complete ontology with descriptions of each term is provided in Text S1 and Text S3.
There exists no comprehensive curated gold standard repository for all types of gene pair interactions. For the 30 interactomes evaluated here, we assembled a gold standard for each type from various sources. SGD interaction labels were used for all terms under the physical and pathway interaction terms [56]. Additional transcriptional regulation annotations were obtained from the high confidence set from [57]. Co-complex annotations were obtained from gene pairs in the GO Slim term PROTEIN_COMPLEX [58]. Pairs included in terms under metabolic interaction were obtained from reactions in the KEGG database [59]. For the topmost node, functional relationships, we used positive examples from the biological process branch of GO [60]. When possible, we further manually curated gene pairs to more specific terms based on literature examination. Manual curation was performed to annotate ubiquitination interactions based on SGD curated interaction publications and also to cross annotate experimentally validated covalent modification branch examples to regulatory interaction branch terms. The directionality of the gold standards was derived directly from the inherent high throughput experiments (e.g. kinases to targets). All gene pairs annotated to a term were propagated such that they were included as positive interactions for all ancestor terms. This resulted in a total of 1,333,014 unique positive labels across 30 terms (individual terms are detailed in Text S1).
This process established positive interactions for each term in our interaction ontology. For supervised machine learning (such as our SVM-based method described below), negative examples are also required. As protein interactions are sparse, we randomly selected a number of negative gene pairs for each term's gold standard equal to the number of positive interactions [61]. Additionally, to assess the accuracy of our directed interaction predictions, we used negative gene pairs identical to the positive examples but with inverted directionality. Finally, for evaluating predictions on new post-translational regulation completely unrelated to our training gold standard, we selected 173 additional gene pairs from 24 recent publications (see Text S1).
Evaluation was performed by randomly excluding ∼%30 of the genes for each interaction type during training. That leads to a group of genes that are not in the training set and established a test set of interactions containing at least one gene from this exclusive gene set. The remaining pairs were used for SVM training and for parameter estimation in the Bayesian network. We used area under the receiver operator characteristic (ROC) curve (AUC) for evaluation as detailed in Text S1.
As training data for each interaction type, we used subsets of a data compendium consisting in total of microarray, colocalization, protein domains, transcription factor binding sites, and sequence similarity. For each interaction type to be predicted, experimental data closely related to the output was excluded (e.g. TF binding sites for regulatory relationships). 78 yeast microarray datasets were included, comprising 3,516 conditions (see Text S2). Missing values in these datasets were imputed using KNNImpute [62] with k = 10, and genes with more than 30% missing values were removed.
For machine learning, one feature was constructed per expression condition as follows. For directional gene pair interaction types such as phosphorylation, we evaluated various methods and found xi-xj to provide optimal performance, where xi and xj are the expression values of gene i and j in condition x. When predicting non-directional interaction types such as physical interaction, we used |xi-xj|, the absolute value of the subtracted expression values.
Colocalization data for 22 different cell compartments [63] and automatically determined protein family information from Pfam B [64] were both included as binary features (true if both genes in a pairs shared localization or a protein family). TRANSFAC data [65] was incorporated using the Euclidian distance between the two gene's binding site profiles across 211 transcription factors. Sequence similarity between the two genes in each pair's 1,000 bp upstream and 1,000 bp downstream was scored as the sequence alignment E-values from all-against-all BLAST outputs.
We developed an integrated method for predicting diverse protein interactions, based on a multi-label hierarchical classification formulation we have previously applied to gene function prediction in both yeast and mouse [6], [66]. First, for each interaction type, we trained 10 separate SVM classifiers. We use bagging (bootstrap aggregation, [67]) to combine these and improve generalization, training each individual SVM classifier on a bootstrapped subsample of its interaction type's complete gold standard. We thus begin with a total of 300 SVM classifiers for our 30 interaction types in yeast, and each interaction type's group of 10 SVM outputs were averaged (bagged) to produce a non-hierarchically-resolved predicted interactome.
Next, a Bayesian network was constructed based on the structure of the interaction ontology. First, we modeled each interaction type's bagged SVM output i as a random event Yi taking discrete values binned by five standard deviations above or five below the training set mean. Each SVM's predictions in isolation were treated as a noisy observation of a latent event Xi representing the true, binary interactions and non-interactions of each type i. Each Yi was considered to be dependent only on its corresponding Xi, and each Xi was dependent only on its set of children {Xj, ..., Xk} in the interaction ontology, resulting in the “decorated tree” Bayesian network structure seen in Figure 1 and in [6]. Given this structure, conditional probability table parameters for P(Yi|Xi) were learned using maximum likelihood from interaction type i's training data. Finally, parameters for P(Xi|Xj, ..., Xk) were fixed to constrain the hierarchical semantics of the ontology. If a pair is annotated to any child in {Xj, ..., Xk}, it must also be of interaction type i, making P(Xi = 1|Xj = 1) = ... = P(Xi = 1|Xk = 1) = 1. The remaining parameters P(Xi = 1|Xj = 0, ..., Xk = 0) were inferred using maximum likelihood by counting the corresponding training labels. Finally, Laplace smoothing was used to improve parameter robustness.
All 30 interactomes were converted into binary interaction networks by setting a threshold of 5 standard deviations above the mean edge probability, retaining ∼1% of all edges. The degree of each gene was counted in this binarized network. The overlap between each pair of interactomes' high-connectivity genes was computed as the probability of a gene g being in the top 5% of interactome N1's degree distribution(Qi(Nj), defined as genes in the top i percent degree distribution of interactome Nj) given that it was in N2's: P[g in Q0.05(N1)|g in Q0.05(N2)]. For each of the 30 interactomes N2, we generated a sorted gene list by edge degree; for directed interactomes, separate lists were generated for in- and out-degree. Next, we counted the number of shared genes in the top 5% of edge degree in the target interactome N1. Finally, hierarchical clustering was used to generate clusters of shared high degree genes.
Network motif enrichment analysis was carried out using FANMOD [68]. Searches were conducted for 3-node motifs using a sampling method with probability parameters of 0.6, 0.5, 0.4 and compared to 500 random networks generated using an edge swapping process preserving each gene's degree. Computational complexity precluded analysis of 4-node motifs. Graphlet degree distributions were calculated using GraphCrunch [69]. For each interactome, 73 graphlet degree distributions were generated, each representing a unique distribution of 2-5 node graphlets. Comparison between graphlet distributions was performed using the GDD agreement metric, defined as the average normalized distance to provide robust comparisons [40], [69].
All software was implemented using the Sleipnir library [70], which interfaces with the SVMperf software [71] for linear kernel SVM classifiers (the error parameter C was set to 20 for these experiments). Bayesian network inference used the Lauritzen algorithm [72] as implemented in the University of Pittsburgh SMILE library [73].
20 gene pairs predicted to synthetically interact [56] with high probability were selected from the DNA topological change and regulation of protein biosynthesis pathways in yeast (as defined by GO [36]). Synthetic Genetic Array (SGA) technology [4], [13] was applied to these pairs by combining either non-essential gene deletion mutants or conditional alleles of essential genes in haploid yeast double mutants. The query mutant strain for each pair of genes (harboring SGA-specific reporters and markers) was crossed to the complementary single mutant strain. Mating to the non-essential gene deletion collection was followed by meiotic recombination and selection of haploid meiotic progeny, resulting in an output array of double mutants grown in rich medium. Fitness was assessed by comparing this double mutant colony size to the sizes of single mutant colonies, which were assessed for significance as described in [4], [13]. A p-value threshold of 0.05 was used to determine the final confirmed synthetic lethal pairs (the full table of p-values can found in Text S1).
|
10.1371/journal.ppat.1003108 | Structural Basis for Feed-Forward Transcriptional Regulation of Membrane Lipid Homeostasis in Staphylococcus aureus | The biosynthesis of membrane lipids is an essential pathway for virtually all bacteria. Despite its potential importance for the development of novel antibiotics, little is known about the underlying signaling mechanisms that allow bacteria to control their membrane lipid composition within narrow limits. Recent studies disclosed an elaborate feed-forward system that senses the levels of malonyl-CoA and modulates the transcription of genes that mediate fatty acid and phospholipid synthesis in many Gram-positive bacteria including several human pathogens. A key component of this network is FapR, a transcriptional regulator that binds malonyl-CoA, but whose mode of action remains enigmatic. We report here the crystal structures of FapR from Staphylococcus aureus (SaFapR) in three relevant states of its regulation cycle. The repressor-DNA complex reveals that the operator binds two SaFapR homodimers with different affinities, involving sequence-specific contacts from the helix-turn-helix motifs to the major and minor grooves of DNA. In contrast with the elongated conformation observed for the DNA-bound FapR homodimer, binding of malonyl-CoA stabilizes a different, more compact, quaternary arrangement of the repressor, in which the two DNA-binding domains are attached to either side of the central thioesterase-like domain, resulting in a non-productive overall conformation that precludes DNA binding. The structural transition between the DNA-bound and malonyl-CoA-bound states of SaFapR involves substantial changes and large (>30 Å) inter-domain movements; however, both conformational states can be populated by the ligand-free repressor species, as confirmed by the structure of SaFapR in two distinct crystal forms. Disruption of the ability of SaFapR to monitor malonyl-CoA compromises cell growth, revealing the essentiality of membrane lipid homeostasis for S. aureus survival and uncovering novel opportunities for the development of antibiotics against this major human pathogen.
| An opportunistic Gram-positive pathogen, Staphylococcus aureus is a major threat to humans and animals, being responsible for a variety of infections ranging from mild superficial to severe infections such as infective endocarditis, septic arthritis, osteomyelitis and sepsis. The increasing resistance of S. aureus against most current antibiotics emphasizes the need to develop new approaches to control this important pathogen. The lipid biosynthetic pathway is one appealing target actively pursued to develop anti-Staphylococcal agents. Despite its potential biomedical importance, however, little is known about the signaling mechanisms that allow S. aureus to control its phospholipid content. In order to shed light on this fundamental mechanism, we studied S. aureus FapR (SaFapR) a transcription factor that senses the levels of malonyl-CoA, a key intermediate in fatty acid biosynthesis, and modulates the expression of genes involved in fatty acid and phospholipid biosynthesis. Our studies of SaFapR uncovered the mechanistic basis of a complex biological switch that controls membrane lipid homeostasis in S. aureus. We also discovered that disruption of the ability of SaFapR to recognize malonyl-CoA, the ligand that controls SaFapR binding to DNA, compromises S. aureus viability, thus revealing new opportunities for the development of antibiotics against this major human pathogen.
| The cell membrane is an essential structure to bacteria. It primarily consists of a fluid phospholipid bilayer in which a variety of proteins are embedded. Most steps involved in phospholipid biosynthesis are therefore explored as targets for designing new antibacterial drugs [1]. The central events in the building of phospholipids enclose the biosynthesis of fatty acids, which are the most energetically expensive lipid components, by the type II fatty acid synthase (FASII) on the cytoplasmic side of the membrane, and subsequent delivery of the latter to the membrane-bound glycerol-phosphate acyltransferases (Fig. S1). Due to the vital role of the membrane lipid bilayer, bacteria have evolved sophisticated mechanisms to finely control the expression of the genes responsible for the metabolism of phospholipids [2].
Transcriptional regulation of bacterial lipid biosynthetic genes is poorly understood at the molecular level. Indeed, the only two well-documented examples are probably those of the transcription factors FadR and DesT, which regulate the biosynthesis of unsaturated fatty acids (UFA) in Escherichia coli and Pseudomonas aeruginosa, respectively. FadR was discovered as a repressor of the β-oxidation regulon [3], [4] and subsequently found to also activate the transcription of fabA and fabB, two essential genes for the biosynthesis of UFA [5]–[7]. Binding of FadR to its DNA operator is antagonized by different long-chain acyl- CoAs [8]–[11], in agreement with the proposed structural model [12]–[14]. DesT is a TetR-like repressor that primarily controls the expression of the desCB operon, which encodes the components of an oxidative fatty acid desaturase, and secondarily controls the expression of the fabAB operon of Pseudomonas aeruginosa [15]. DesT binds either saturated or unsaturated acyl-CoAs, which respectively prevent or enhance its DNA-binding properties [16], thus allowing the repressor to differentially respond to alternate ligand shapes [17].
Among the known regulatory mechanisms that control lipid synthesis in bacteria, the Bacillus subtilis Fap system is unique in that the regulated final products, fatty acids and phospholipids, are controlled by a metabolite required at the beginning of the fatty acid biosynthetic pathway, malonyl-CoA [18], [19]. To monitor the levels of this metabolite, bacteria employ the FapR protein [20], which is highly conserved among several Gram-positive pathogens. FapR has been shown to globally repress the expression of the genes from the fap regulon (Fig. S1) encoding the soluble FASII system as well as two key enzymes that interface this pathway with the synthesis of phospholipid molecules [18].
Like most transcriptional regulators in bacteria, FapR is a homodimeric repressor [20]. Each protomer consists of a N-terminal DNA-binding domain (DBD) harboring a classical helix-turn-helix motif connected through a linker α-helix (αL) to a C-terminal effector-binding domain (EBD). We have previously determined the structure of a truncated form of FapR from Bacillus subtilis (BsFapR), which included the linker α-helix and the EBD but lacked the DBD [20]. The EBD folds into a symmetric dimer displaying a ‘hot-dog’ architecture, with two central α-helices surrounded by an extended twelve-stranded β-sheet. A similar fold has been found in many homodimeric acyl-CoA-binding enzymes [21], [22] involved in fatty acid biosynthesis and metabolism [23], [24]. However, the bacterial transcriptional regulator FapR appears to be so far the only well-characterized protein family to have recruited the ‘hot-dog’ fold for a non-enzymatic function.
The structure of truncated FapR bound to malonyl-CoA revealed structural changes in some ligand-binding loops of the EBD, and it was suggested that these changes could propagate to the missing DNA-binding domains to impair their productive association for DNA binding [20]. However, the actual mechanisms involved remain largely unknown in the absence of detailed structural information of the full-length repressor and its complex with DNA. Here, we report the structural characterization of full-length FapR from Staphylococcus aureus (SaFapR), a major Gram-positive pathogen causing severe human infections [25]–[27]. The crystal structures of SaFapR have been obtained for the protein alone and for the complexes with the cognate DNA operator and the effector molecule, malonyl-CoA, providing important mechanistic insights into the mode of action of this transcriptional regulator. We further demonstrate that structure-based SaFapR mutants interfering with malonyl-CoA binding are lethal for S. aureus. These data show that membrane lipid homeostasis is essential for S. aureus survival and highlights this regulatory mechanism as an attractive target to develop new antibiotics.
The S. aureus fap regulon is organized as in B. subtilis [18], except for two missing genes (yhfC and fabHB) (Table S1). Electrophoretic mobility shift assays revealed that SaFapR binds to its own promoter (PfapR) and that malonyl-CoA specifically disrupts the repressor-operator complex (Fig. S2). Furthermore, the unlinked genes plsX and fabH from the fap regulon are upregulated in two distinct S. aureus strains lacking the repressor (Fig. S3A) and SaFapR is able to complement a B. subtilis fapR mutant strain (Fig. S3B). These results clearly demonstrate that SaFapR conserves the same regulatory function originally described for the B. subtilis orthologue [20].
The crystal structure of SaFapR in complex with a 40-bp oligonucleotide comprising the PfapR promoter was determined at 3.15 Å resolution (Table 1). Two SaFapR homodimers were observed to bind to a single DNA molecule in the crystal (Fig. 1A). This observation is in agreement with isothermal titration calorimetry (ITC) studies of protein-DNA interactions, which showed a complex isotherm with a large exothermic component and two distinct binding reactions (Fig. 1B). Sequence analysis of the promoters of the fap regulon from B. subtilis [18] and S. aureus (Fig. S4) revealed the presence of a conserved inverted repeat. Our previous DNAse I footprinting analyses performed with BsFapR on both strands of the B. subtilis fapR promoter demonstrated that this symmetric element covers half of the DNAse protected region [20]. Interestingly, this region corresponds to the recognition site of one SaFapR homodimer in the crystal, suggesting a sequential mechanism of binding. Indeed, a sequential binding model fits well the observed isotherm (Fig. 1B), indicating that the two SaFapR homodimers bind the operator with nanomolar dissociation constants of 0.5±0.1 nM for the first and 51±8 nM for the second binding reactions. In the crystal structure of the complex, however, the two SaFapR homodimers are related by a local two-fold symmetry axis, and the DNA molecule is bound in a 50∶50 orientation, as confirmed by determining the crystal structure of SaFapR in complex with an asymmetrically Br-labelled oligonucleotide using the Br anomalous scattering signal (data not shown).
Each protein homodimer exhibits an elongated asymmetric conformation in the crystal structure, in which the two DNA-bound DBDs are structurally detached from the central dimeric ‘hot-dog’ EBD. Indeed, the positioning of the EBD (stabilized by crystal packing contacts) relative to the DBD-DNA complex requires partial unwinding of the C-terminal end of αL, as confirmed by comparing the structures of the ligand-free and DNA-bound forms of SaFapR (see below). The amphipatic linker helix αL plays a major role in stabilizing the molecular architecture of the SaFapR-DNA complex. Helix αL interacts with αL′ from the second protomer mainly through its exposed hydrophobic face (including residues Ile59, Val62 and Ala63; Fig. S5A). This dimerization region is further stabilized by hydrophobic contacts (Phe26) and hydrophilic interactions between αL and the α1–α2 connecting loop from both DBDs. In addition, the guanidinium group of Arg59 gets engaged in electrostatic interactions with the main-chain carbonyl groups of Pro25 and Ile27, and the hydroxyl group of Tyr67 is within hydrogen bonding distance to the main-chain of Ser23 at the end of α1′ on the second protomer (Fig. S5A). Altogether this contact network largely contributes to the correct relative positioning of the two helix-turn-helix motifs on the DNA major groove.
The two DBDs from the asymmetric homodimer interact in a similar manner with DNA. The recognition helix α3 from the helix-turn-helix motif penetrates the double-helix major groove and makes extensive interactions with the two DNA chains (Fig. 2A). The SaFapR-DNA interface mainly involves the DNA backbone phosphates (Fig. 2B), except for base-specific interactions of Gln41 (from the recognition helix α3) in the major groove and Arg56 (from the linker helix αL) in the minor groove. Importantly, the insertion of the guanidinium groups of this arginine from both protomers promotes the opening of the minor groove, inducing a pronounced local bending of DNA (Fig. 2C). The two phosphate-contacting residues from helix α1 (Lys10 and Arg13), as well as residues from the recognition helix α3 and the C-terminal half of the linker helix αL (which include most other DNA-contacting residues) are highly conserved in the entire FapR protein family (Fig. S6), indicating a conserved mode of DNA-binding. The N-terminus of the protein (preceding helix α1) faces the adjacent minor groove in the crystal structure (Fig. 2A), suggesting that it could be engaged in additional DNA interactions. However, this region is poorly conserved in the FapR protein family and is disordered in the crystal structure.
The crystal structure of full-length SaFapR in complex with malonyl-CoA, determined at 1.9 Å resolution (Table 1), revealed a different, more compact conformation of the repressor. Most notably, the two amphipatic linker helices αL, instead of interacting with each other as in the DNA-bound repressor, now bind to either side of the central dimeric EBD (Fig. 3A). This results in a protein conformation with the two helix-turn-helix domains far apart from each other, corresponding to an incompetent DNA-binding state. Interactions of helix αL with the lateral face of the EBD play a major role in stabilizing the observed quaternary organization of the protein, mainly through extended hydrophobic contacts between aliphatic and aromatic side chains (Fig. 3B). In addition, several hydrogen bonding and salt bridge interactions (Fig. S5B) lock a well-defined structure for the loop connecting helix αL to the first β-strand of the EBD (residues 70–82), which further restrict the mobility of αL and contrasts with the loose asymmetric conformation observed for this loop in the DNA-bound SaFapR homodimer (Fig. 1A).
The phosphopantetheine and malonyl groups of malonyl-CoA are well defined in the electron density map (Fig. 3C). The phosphopantetheine group binds within a tunnel at the interface between the two protomers in the SaFapR homodimer, and adopts the same conformation as observed in many acyl-CoA-binding proteins displaying the hot-dog fold [20], [24]. This binding mode results in the complete occlusion of the ligand malonate from the bulk solvent. The charged carboxylate group of malonate is neutralized through a salt bridge with the guanidinium side-chain of Arg110, and makes additional hydrogen bonding interactions with the main-chain nitrogen from Gly111 and the side-chain of Asn119′ from the second protomer (Fig. 3C). The engagement of Arg110 in malonyl-CoA-binding triggers a local reorganization of hydrogen-bonding interactions (Fig. S5B) and surface reshaping that further stabilize the loop connecting αL to the first β-strand of the hot-dog fold in the non-productive conformation (Fig. 3A), thus preventing DNA binding. On the other hand, the adenosine 3′-phosphate moiety of malonyl-CoA is largely exposed to the solvent (it is partially disordered in one of the two protomers) and makes no specific contacts with the protein. This implies that SaFapR specifically recognizes the malonyl-phosphopantetheine moiety of the ligand, in agreement with the observation that either malonyl-CoA or malonyl-acyl carrier protein (malonyl-ACP) can indistinctly function as effector molecules [28].
A detailed comparison of the malonyl-CoA complexes between full-length SaFapR and the truncated form of BsFapR (lacking the DBDs) revealed a conserved structural arrangement of the EBD core, and ligand binding promoted the same conformation of the connecting loop αL-β1 (Fig. S7). On the other hand, helix αL displays a different organization, due in part to the absence of the DBDs. This helix protrudes away from the EBD core in BsFapR to get involved in crystal packing contacts [20]. Interestingly, hydrophobic residues engaged in these inter-domain interactions are largely invariant in the whole FapR family (Fig. S6), as are also key residues involved in electrostatic interactions (Fig. S5). Altogether, the structural alignment indicates not only an identical mode of malonyl-CoA binding but also the conservation of the DBD – αL – EBD interactions required to stabilize the FapR-malonyl-CoA complex as visualized in the SaFapR model.
The crystal structure of full-length SaFapR in the absence of ligands has been obtained in two different crystal forms (Table 1). Interestingly, in three out of four crystallographically independent protomers the repressor has the same non-productive quaternary arrangement as observed in the structure of malonyl-CoA-bound SaFapR, with helix αL bound to the lateral face of the EBD (Fig. 4A), strongly suggesting that this compact structure likely represents a conformation of the protein in solution. However, the helix-turn-helix motifs display high temperature factors or even partial disorder, and are engaged in extensive crystal contacts, suggesting the coexistence of alternative conformational states in solution characterized by flexible DBDs. In that sense, in one protomer of the crystal form 2 (Table 1) both helix αL and its associated DBD are highly flexible and could not be modeled, suggesting a marginal stability of the observed quaternary arrangement. Moreover, the first visible residues of this same monomer (positions 72–77, connecting helix αL with the first β-strand of the EBD) adopt a conformation that differs from that observed in the other ligand-free or malonyl-CoA-bound protomers, but resembles that found for one subunit of the asymmetric DNA-bound form of the repressor (Fig. 4B).
The structure of SaFapR in complex with the DNA operator revealed a strikingly different overall conformation of the protein, compared to those of the ligand-free and the effector-bound repressor. As highlighted by the crystal structures described above, the transitional switch between the relaxed (DNA-bound) and tense (malonyl-CoA-bound) forms of the repressor involves a large-scale structural rearrangement (Fig. 5). The amphipathic helix αL, whose hydrophobic face binds to the protein core in the tense state (Fig. 5A), structurally dissociates and moves >30 Å to interact with the same helix from the other protomer and with DNA in the relaxed state (Fig. 5B). This DNA-driven process requires partial unwinding of the linker helix αL and the solvent-exposure of hydrophobic side-chains from the EBD (Leu82, Ile83, Val85, Ile94, Val123) and from the loop immediately following helix αL (Ile70, Phe78) in both protomers. Such a substantial conformational rearrangement between the tense and relaxed states (Fig. 5C) contrasts with the more subtle structural changes observed for other well-studied bacterial classes of allosteric transcriptional regulators such as the tetracycline [29] or lactose [30] repressors, illustrating that specific physiological responses can be achieved by a variety of mechanisms.
Bacterial FASII has been identified as a promising target for antibacterial drug discovery. Nevertheless, Brinster et al [31] have questioned the feasibility of this approach in Gram-positive pathogens based on the finding that FASII is not essential in Streptococcus agalactiae if the bacterium is supplemented with fatty acids or human serum. This controversy was recently clarified by Parsons et al [32], who showed that externally added fatty acids downregulate the activity of the acetyl-CoA carboxylase in some Gram-positive pathogens, such as Streptococcus pneumoniae. This biochemical mechanism suppresses the malonyl-CoA levels allowing fatty acid supplements to replace endogenous fatty acids completely, thus rescuing bacterial FASII inhibition. In S. aureus, this feed-back regulatory mechanism is not present [32] and thus external fatty acids do not circumvent the treatment of this bacterium with FASII inhibitors. These results and the regulatory properties of the acc genes led the authors to propose that pathogens containing FapR would likely be sensitive to FASII inhibitors, regardless of the addition of external fatty acids [32]. These considerations prompted us to test if structure-based mutations predicted to disrupt the SaFapR-malonyl-CoA interaction are lethal for S. aureus even in the presence of extracellular fatty acids. To this end, based on the structural information and on our previous work [20], we substituted Arg110 by alanine to disrupt the key interaction with the malonyl carboxylate, and introduced a double substitution (Gly111Val, Leu132Trp) to block the ligand-binding tunnel that accommodates the phosphopantetheine moiety in the repressor-effector complex (Fig. 3C). These mutants were expected to retain their DNA-binding capacity and to permanently repress the expression of the fap regulon, independently of the metabolic conditions. To test this prediction we engineered S. aureus fapR null mutants to produce the protein variants in response to an inducer (IPTG). IPTG-induced expression of SaFapRR110A caused a small drop in cell viability (data not shown), while cells expressing SaFapRG111V,L132W failed to grow (Fig. 6). A variety of fatty acids, including oleic acid, a common mammalian fatty acid, and anteiso saturated fatty acids, the most abundant acyl chains found in S. aureus phospholipids were unable to overcome the growth inhibition caused by the expression of the superrepressor SaFapRG111V,L132W in S. aureus (Fig. 6). These results clearly show that disruption of the membrane lipid homeostasis mimics the effects of FASII inhibitors in S. aureus strongly supporting the notion that exogenous fatty acids cannot replace the endogenously produced acyl chain in this bacterium [32].
The structures of SaFapR, in three relevant states of its regulation cycle, uncover a complex biological switch, characterized by completely different protein-protein interactions involving in all cases the linker helix αL and so leading to a distinct quaternary arrangement.for the tense and relaxed states of the repressor (Fig. 5). Similarly to other homodimeric proteins studied by single-molecule or NMR relaxation approaches [33], [34], our crystallographic studies suggest that the two conformational states of SaFapR (i.e. with EBD-bound or EBD-detached DBDs) can be populated in the ligand-free repressor species. Thus, a higher cellular concentration of malonyl-CoA would not only trigger the conformational changes that disrupt the SaFapR-operator complex [20], but would also promote the dynamic shift of the ligand-free repressor population towards the tense state.
Our results highlight the ability of FapR to monitor the levels of malonyl-CoA and appropriately tune gene expression to control lipid metabolism, ensuring that the phospholipid biosynthetic pathway will be supplied with appropriate levels of fatty acids either synthesized endogenously or incorporated from the environment. Bacterial FASII is a target actively pursued by several research groups to control bacterial pathogens [35]. A controversy surrounding FASII as a suitable antibiotic target for S. aureus was based on the ability of this pathogen to incorporate extracellular fatty acids [36]. This apparent discrepancy was recently clarified by showing that although exogenous fatty acids indeed are incorporated by S. aureus, following its conversion to acyl-ACP, they cannot deplete the generation of malonyl-CoA and malonyl-ACP [32]. Thus, these fatty acid intermediates release FapR from its binding sites [28] and are used by FASII to initiate new acyl chains. Thus, when a FASII inhibitor is deployed against S. aureus, the initiation of new acyl chains continues leading to depletion of ACP, which is correlated with diminished exogenous fatty acids incorporation into phospholipids [32]. Strikingly, SaFapR, not only control the expression of the FASII pathway, but also regulate the expression of key enzymes required for phospholipid synthesis such as PlsX and PlsC [18], [37]. Most of the Gram-positive pathogens rely on the PlsX/PlsY system to initiate phospholipid synthesis by converting acyl-ACP to acyl-P04 by PlsX followed by the transfer of the fatty acid to the 1 position of glycerol-PO4 by the PlsY acyltransferase [37], [38]. A strength of targeting these steps in lipid synthesis is that acyltransferases inhibition cannot be circumvented by supplementation with extracellular fatty acids. Thus, targeting of lipid synthesis with compounds that block the expression of PlsX in S. aureus cannot be ignored, specially taking into account that it has been reported that extracellular fatty acids increase the MIC for FASII inhibitors in this important pathogen. [32], [39]. In this sense, the unique mode of action of FapR and our encouraging in vivo results validate lipid homeostasis as a promising target for new antibacterial drug discovery in Gram-positive bacteria.
Conserved in many Gram-positive bacteria, FapR is a paradigm of a feed-forward-controlled lipid transcriptional regulator. In other characterized bacterial lipid regulators it is the long acyl-chain end products of the FASII pathway that act as feedback ligands of lipid transcriptional regulators [40]. The effector-binding domains of these proteins, such as E. coli FadR [12] or the TetR-like P. aeruginosa DesT [17], frequently display an α-helical structure with a loose specificity for long-chain acyl-CoA molecules, possibly because the permissive nature of helix-helix interactions provide a suitable platform to evolve a binding site for fatty acid molecules of varying lengths. In contrast, a high effector-binding specificity is required for the feed-forward regulation mechanism of the FapR repressor family [19], which entails the recognition of an upstream biosynthetic intermediate, namely the product of the first committed step in fatty acid biosynthesis. This high specificity is achieved in SaFapR by caging the charged malonyl group inside a relatively stiff internal binding pocket, and may be the reason why the hot-dog fold was recruited for this function. Nevertheless, it could be expected that, in organisms using the FapR pathway, a complementary feed-back regulatory loop should also operate at a biochemical level, for instance by controlling the synthesis of malonyl-CoA [37]. This would imply that lipid homeostasis in FapR-containing bacteria would be exquisitely regulated by feed-back and feed-forward mechanisms, as it is indeed the case in higher organisms ranging from Caenorhabditis elegans to humans [41].
In frame fapR deletion mutants of S. aureus strains RN4220 and HG001 [42], RN4220ΔfapR and HG001ΔfapR respectively, were constructed and the expression level of the plsX and fabH genes, belonging to the fap regulon, was analyzed by real time PCR as described in Text S1. To evaluate the activity of the S. aureus repressor in B. subtilis, we used the ΔfapR mutant strain GS416 that contains a PfabHB-lacZ reporter fusion and expressed S. aureus fapR from the xylose inducible PxylA promoter (for further details see Text S1).
Mutants SaFapRR110A and SaFapRG111V,L132W, impaired in malonyl-CoA binding were obtained by site-directed mutagenesis using S. aureus RN4220 genomic DNA as template, mutagenic oligonucleotides and overlap-extension PCR. The replicative vector pOS1 [43] and the tight IPTG-regulated PspacOid promoter from pMUTIN4 were used for ectopical expression of fapR alleles in RN4220ΔfapR (see Text S1). S. aureus transformants were selected in THA plates supplemented with cloramphenicol (5 µg/ml) and expression of fapR alleles was achieved by addition of 10 mM IPTG. To evaluate the effect of an external source of fatty acids THA plates were supplemented with 500 µM of the stated fatty acids or 0.1% Tween80 (Fig. 6).
The S. aureus fapR gene was cloned into the pET15b vector (Novagen) as described in Text S1 and expressed in E. coli BL21/pLysS. Bacterial cultures were grown in LB supplemented with ampiciline 100 µg/ml and chloramphenicol 10 µg/ml at 37°C until OD600 0.6. Expression was induced with 0.5 mM IPTG at 20°C for 17 hours. Cells were harvested by centrifugation at 4°C, and protein-containing fractions from a Ni2+-affinity chromatography on a HisTrap column (GE Healthcare) were dialyzed overnight against an excess volume of 50 mM Tris-HCl pH 7.6, 300 mM NaCl and 1 mM DTT at room temperature, in the presence of a 1/40 (w/w) ratio of His-tagged TEV protease. After dialysis, a second affinity chromatography step was performed in the presence of 20 mM imidazole, the flow through was concentrated and injected into a HiLoad 16/60 Superdex 75 prep grade column (GE Healthcare) equilibrated with 50 mM Tris-HCl, pH 7.6, and 300 mM NaCl. After gel filtration SaFapR was dialyzed at room temperature against 20 mM Tris-HCl pH 7.6 and 50 mM NaCl, concentrated to 15 mg/ml and stored in aliquots at −80°C for further use. Selenomethionine (SeMet)-labeled SaFapR was obtained using the E. coli strain B834 (Novagen) grown in M9 medium containing 0.2 g/l selenomethionine and purified as above.
Crystals of the repressor and its complexes with the effector and operator molecules were obtained as described in Text S1. All diffraction datasets were collected from single crystals at 100 K using synchrotron radiation at beamlines ID14.4 and ID29 (European Synchrotron Radiation Facility, Grenoble, France) or Proxima 1 (SOLEIL, Saint-Aubin, France). Data were processed with either XDS [44] or iMosflm [45] and scaled with XSCALE from the XDS package or SCALA from the CCP4 suite [46]. The crystal structures of SaFapR alone and its complex with malonyl-CoA were solved by molecular replacement methods using the program AMoRe [47] and the effector-binding domain of B. subtilis FapR (PDB entry 2F41) as the search model. In all cases, the missing DNA-binding domains were manually traced from sigma A-weighted Fourier difference maps. The structure of the repressor-operator complex was solved by a combination of molecular replacement and single-wavelength anomalous diffraction (SAD) techniques using SeMet-labeled SaFapR. The selenium substructure was determined with the program SHELXD [48] and further refined with the program SHARP [49]. Structures were refined with REFMAC5 [50] or BUSTER [51] (for the SeMet-labeled protein-DNA complex) using a TLS model and non-crystallographic symmetry restraints when present, alternated with manual rebuilding with COOT [52]. Models were validated through the MolProbity server [53]. Data collection and refinement statistics are reported in Table 1. Graphic figures were generated and rendered with programs Pymol [54] and 3DNA [55].
ITC experiments were performed using the high precision VP-ITC system (MicroCal Inc., MA) and quantified with the Origin7 software provided by the manufacturer. All molecules were dissolved in 50 mM TrisHCl, pH 8, 150 mM NaCl and the binding enthalpies were measured by injecting the SaFapR solution into the calorimetric cell containing the 40 bp DNA solution. Heat signals were corrected for the heats of dilution and normalized to the amount of compound injected. Complementary DNA strands were heated to 90°C and annealed by a stepwise decrease to 25°C followed by 30 min on ice prior to use; DNA concentration was determined by absorption at 260 nm.
Crystallographic coordinates and structure factors were deposited in the Protein Data Bank, with accession codes 4a0x, 4a0y, 4a0z and 4a12.
|
10.1371/journal.ppat.1005255 | Kaposi’s Sarcoma-Associated Herpesvirus (KSHV) Induces the Oncogenic miR-17-92 Cluster and Down-Regulates TGF-β Signaling | KSHV is a DNA tumor virus that causes Kaposi’s sarcoma. Upon KSHV infection, only a limited number of latent genes are expressed. We know that KSHV infection regulates host gene expression, and hypothesized that latent genes also modulate the expression of host miRNAs. Aberrant miRNA expression contributes to the development of many types of cancer. Array-based miRNA profiling revealed that all six miRNAs of the oncogenic miR-17-92 cluster are up-regulated in KSHV infected endothelial cells. Among candidate KSHV latent genes, we found that vFLIP and vCyclin were shown to activate the miR-17-92 promoter, using luciferase assay and western blot analysis. The miR-17-92 cluster was previously shown to target TGF-β signaling. We demonstrate that vFLIP and vCyclin induce the expression of the miR-17-92 cluster to strongly inhibit the TGF-β signaling pathway by down-regulating SMAD2. Moreover, TGF-β activity and SMAD2 expression were fully restored when antagomirs (inhibitors) of miR-17-92 cluster were transfected into cells expressing either vFLIP or vCyclin. In addition, we utilized viral genetics to produce vFLIP or vCyclin knock-out viruses, and studied the effects in infected TIVE cells. Infection with wildtype KSHV abolished expression of SMAD2 protein in these endothelial cells. While single-knockout mutants still showed a marked reduction in SMAD2 expression, TIVE cells infected by a double-knockout mutant virus were fully restored for SMAD2 expression, compared to non-infected TIVE cells. Expression of either vFLIP or vCycIin was sufficient to downregulate SMAD2. In summary, our data demonstrate that vFLIP and vCyclin induce the oncogenic miR-17-92 cluster in endothelial cells and thereby interfere with the TGF-β signaling pathway. Manipulation of the TGF-β pathway via host miRNAs represents a novel mechanism that may be important for KSHV tumorigenesis and angiogenesis, a hallmark of KS.
| MiRNAs are small non-coding RNAs which decrease gene expression and function as oncogenes or tumor suppressors. Dysregulation of miRNAs is a hallmark of many human cancers. Recently, it was revealed that the miR-17-92 cluster, up-regulated in many cancers, plays a central role in down-regulation of the TGF-β signaling pathway. Kaposi’s sarcoma-associated herpesvirus (KSHV) is a gamma-herpesvirus associated with Kaposi’s sarcoma and two lymphoproliferative diseases. KSHV is known to target the TGF-β pathway. Here, we found that two viral latent genes, vFLIP and vCyclin, blunt TGF-β signaling by inducing the host miR-17-92 cluster. Moreover, we confirmed that endothelial cells infected with wt KSHV gave no expression of SMAD2, a key component in the TGF-β pathway. Using a vFLIP vCyclin double knock-out mutant virus gave complete restoration of SMAD2 expression in endothelial cells. This finding reveals a new pathway that KSHV utilizes to promote tumorigenesis and angiogenesis in Kaposi’s sarcoma.
| Kaposi’s sarcoma-associated herpes virus (KSHV) is a member of the gammaherpesvirus family and is associated with Kaposi’s sarcoma (KS), a subset of multicentric Castleman’s disease (MCD), and primary effusion lymphoma (PEL) [1,2,3]. KSHV has two distinct phases of infection, latent and lytic. During latency, a small subset of KSHV genes is expressed from the KSHV latency associated region (KLAR). Latent gene products including kaposin, viral Fas-associated death domain IL-1β-converting enzyme inhibitory protein (vFLIP), viral cyclin (vCyc), latency-associated nuclear antigen (LANA), and viral micro RNAs (miRNAs), contribute to the survival and proliferation of KSHV- infected tumor cells.
Viral FLIP is a homolog of cellular FLIP, which can protect cells from Fas-mediated apoptosis. Furthermore, vFLIP does not just block the extrinsic signal but also induces NF-κB signaling, which is important for viral latency and tumorigenesis [4]. Rat-1 cells expressing vFLIP promote the tumor formation in nude mouse, which is associated with NF-κB activation [5]. While vFLIP is a potent inducer of apoptosis and required for PEL cell survival its expression levels are tightly regulated in part due to usage of rare codons [6]. Inducible expression of vFLIP alone in mouse endothelial cells leads to induction of serum proinflammatory cytokines in vivo, and alterations in myeloid differentiation [7]. In addition to regulating apoptosis and autophagy [8], and activating NF-κB, vFLIP proteins also regulate intrinsic innate immunity by targeting IRF-3 (reviewed in [9]). The viral cyclin (vCyc) is a homolog of cellular cyclin D, which functions at the G1/S cell cycle transition by activation of cyclin dependent kinase 6 (cdk6). Unlike cellular cyclin D, vCyc is resistant to p27 cdk inhibitor [10,11]. Despite resistance to cdk inhibitors, the activity of KSHV cyclin is blocked by p53. However, it is reported that vCyc is sufficient to promote proliferation of latently infected cells thereby potentially contributing to cell transformation and tumorigenesis in the presence of p53 inhibiting factors, such as LANA [12].
MicroRNAs are short (21~25 nucleotides) noncoding RNAs that down-regulate gene expression post-transcriptionally by binding to the 3’-untranslated region (3’-UTR) of target messenger RNA (mRNA) with partial complementarity [13]. miRNAs play a central role in central biological processes, such as development, differentiation, apoptosis, and proliferation [14]. Dysregulation of miRNAs is not only a hallmark of many human malignancies but is also involved in the development and progression of cancer. MiRNAs either target tumor suppressor genes or by themselves can have oncomir or tumor promoting activity (for review see [15]).
The miR-17-92 cluster is one of the well-characterized oncogenic miRNA clusters, for which aberrant expression was found in various types of cancers and which has been shown to play a critical role in development. miR-17, 18a, 19a, 20a, 19b-1, and 92a-1, the members of the miR-17-92 cluster, are derived from a single polycistronic transcript located at chromosome 13q31. Amplification of this region is found in several types of lymphomas and lung cancer and overexpression in transgenic mice causes B cell lymphomas [16,17]. The miR-17-92 cluster is regulated by the transcription factor, c-Myc, that is frequently hyperactive in many types of cancers. With respect to cell cycle control, the miR-17-92 cluster miRNAs target the E2F transcription factor family but are also activated by E2F, thereby establishing a negative feedback loop [18,19,20]. In neuroblastoma cells, miR-17-92 has been shown to target components of the TGF-β signaling pathway. TGF-β receptor 2 is targeted by miR-17 and miR-20, and SMAD2 and SMAD4 are inhibited by miR-18a [21]. In addition, two homologous clusters, miR106a-363 and miR106b-25, are located at the X and chromosome 7, respectively. The 15 miRNAs expressed from these three clusters represent 4 different seed sequence families which are miR-17, 18, 19, and 92 [22].
The TGF-β signaling pathway is involved in many cellular events, but in the context of KSHV pathogenesis, the most relevant functions of TGF-β are suppression of cell growth and promotion of apoptosis [23]. The importance of TGF-β during latent KSHV infection is supported by the fact that KSHV negatively modulates TGF-β signaling by different mechanisms. Multiple KSHV miRNAs inhibit TGF-β maturation by targeting thrombospondin1 (THBS1) [24]. KSHV LANA targets the TGF-β pathway by reducing the expression of TGF-β receptor type 2 (TGBR2) [25]. TGF-β signaling in developing or progressing cancers is highly context-dependent, since TGF-β signaling has both pro-apoptotic and proliferative properties [26]. It is widely thought that inhibition of TGF-β signaling is important during the early stages of tumorigenesis, even though the pathway is subsequently reactivated in the later stages of cancers associated with metastasis. TGF-β can also block cell cycle progression by inducing the expression of cdk inhibitors [27,28].
We performed miRNA profiling and observed increased expression of the tumorigenic miR-17-92 cluster in KSHV latently infected cells. Testing of KSHV latency associated genes for their ability to up-regulate the miR-17-92 cluster revealed that vFLIP and vCyclin augment transcription from the miR-17-92 promoter which in turn strongly down-regulates TGF-β signaling.
KSHV encodes 12 miRNA genes and in latently infected cells viral miRNAs can represent a significant percentage of all miRNAs within active RISCs [29]. Additionally, previous reports have analyzed host cellular miRNA expression patterns in KSHV-infected tumor cells [30]. Hence, we hypothesized that infection with KSHV not only modulates host cellular gene expression through viral miRNA targeting but also by directly perturbing cellular miRNA expression. To test this hypothesis, we performed microarray-based miRNA profiling of host miRNAs in mock or stably KSHV-infected SLK and long-term infected TIVE-LTC cells. A custom made array was designed that contained probes for a total of 1667 human miRNAs and several control genes for normalization (for details see Materials and Methods). In KSHV-infected SLK cells, which are of epithelial origin, 80 miRNAs were up-regulated 2-fold or more, compared to mock. Among the highest induced were the oncomir miR-155, miR-27, and several members of the let-7 family (Fig 1A and Table 1). Comparison of TIVE to TIVE/LTC, which are of endothelial origin, revealed 103 miRNAs that were up-regulated 2-fold or more and 8 miRNAs that were down-regulated including miR-125 and miR-100. The top eight induced host miRNAs represented members of the oncogenic miR-17-92 cluster and its orthologs. Comparison of both data sets revealed 59 miRNAs including the miR-17-92 cluster that were commonly up-regulated in both KSHV-infected TIVE and SLK cells. Additional miRNAs that are known to be aberrantly expressed in various cancers and were induced in both cell lines included the let-7 family, miR-16, miR-21, and miR-34 (Fig 1B and Table 1).
Since the miR-17-92 cluster has known oncomir activity and like the miR-155 pathway is targeted in many human malignancies including PEL [31,32], we focused on the regulation of these miRNAs (Fig 1 and Table 1). We note that in TIVE cells, the miR-17-92 cluster miRNAs are not detectable, while they are highly expressed in TIVE-LTC. To confirm the array analysis the expression levels of seven miRNAs were analyzed by stem-loop qRT-PCR assays. As shown in Fig 1 and Table 1, up-regulation of three miR-17-92 cluster miRNAs (mir-17-5p, mir-19b, and miR-92) were confirmed for both KSHV-infected SLK and TIVE cells.
To determine whether the observed up-regulation of the miR-17-92 cluster is due to transcriptional regulation a reporter assay was performed. A vector, containing the promoter of the miR-17-92 cluster upstream of the luciferase gene, was transfected into mock or latently KSHV-infected SLK cells. The miR-17-92 promoter showed approximately 4-fold increased activity in KSHV-infected SLK cells compared to uninfected cells (Fig 2A). Since KSHV infected SLK cells express only a limited subset of genes during latency these results suggested that latent KSHV genes contribute to augmenting expression of the miR-17-92 cluster. Thus, we investigated which latent genes are responsible for increasing miR-17-92 expression by testing LANA, the miRNA cluster, vFLIP and vCyclin. LANA and the KSHV miRNA cluster expression plasmids have previously been described [24,33]. Tagged vFLIP and vCyclin were cloned into the Gateway vector pLenti6-V5 (pLenti6/vFLIP and pLenti6/vCyc) and expression was confirmed by Western blot analysis (Fig 2C). Next, expression vectors were co-transfected with the miR-17-92 promoter luciferase reporter. While LANA and viral miRNA expression did not affect luciferase expression from the miR-17-92 promoter (Fig 2B), vFLIP and vCyclin each increased luciferase expression by a factor of 8 (Fig 2D). Hence, these data show that two latency-associated genes vFLIP and vCyclin are responsible for the up-regulation of the miR-17-92 cluster. Next we investigated the consequences of up-regulation of the miR-17-92 cluster.
It was previously demonstrated that three miRNAs of the miR-17-92 cluster target SMAD2 in neuroblastoma cells. TGF-β is a cytokine with anti-proliferative and pro-apoptotic effects, and this important pathway is known to be targeted by KSHV miRNAs and LANA [24,25]. Thus, we asked if the miR-17-92 cluster up-regulated by vFLIP and vCyclin caused decreased SMAD2 expression. We performed Western blot assay with lysates of SLK cells ectopically expressing vFLIP or vCyclin. Surprisingly, the expression of SMAD2 protein, which is readily detectable in SLK cells, was completely abolished by expression of either vFLIP or vCyclin (Fig 3A). In order to check if TGF-β signaling via SMAD2 is disrupted by vFLIP or vCyclin expression, we performed a luciferase reporter assay with a plasmid containing 4 SMAD binding elements upstream from luciferase. Luciferase activity was induced more than 10-fold in the presence of TGF-β ligand, demonstrating that SLK cells are sensitive to TGF-β treatment. However, transfection of vFLIP or vCyclin diminished the response to the TGF-β measured by luciferase activity (Fig 3B and 3C). Although TGF-β signaling is disrupted by vFLIP and vCyclin expression, it was unclear if vFLIP and vCyclin down-regulate SMAD2 by stimulating miR-17-92 expression. To address this question we co-transfected vFLIP or vCyc expression vectors with sequence-specific antagomirs against miR-17-5p, 18a, and 20 to inhibit the miR-17-92 function, and performed a SMAD2 Western blot. In Fig 4A, SMAD2 expression inhibited by vFLIP and vCyclin was restored in the presence of miR-17-92 antagomirs. Moreover, the responsiveness to TGF-β ligand was fully restored as measured by the SMAD-responsive luciferase assay in vFLIP-transfected cells (Fig 4B). However, in vCyclin-transfected cells, SMAD2 protein level was fully restored while TGF-β responsiveness was partially restored (Fig 4C). Together, these results are consistent with reduction of SMAD2 protein levels by vFLIP and vCyclin being mediated via increased expression of miRNAs in the miR-17-92 cluster.
We utilized mutant viruses, which do not express either vFLIP or vCyclin, in order to study the contribution of vFLIP or vCyclin to SMAD2 regulation in the context of a latent virus infection. To generate single or double knock-out mutant viruses of vFLIP or vCyclin, KSHV BAC16 was used [34]. LANA, vFLIP, vCyclin and viral miRNAs are all expressed from a single promoter, which gives rise to polycistronic multiply spliced mRNAs. To mutate vFLIP or vCyclin without affecting the complex RNA expression pattern in this locus, we mutated start codons rather than deleting open reading frames. After mutant bacmids were confirmed by sequencing, recombinant virus was recovered by first transfecting bacmid DNA into 293 cells followed by co-cultivation with iSLK cells.
Next we monitored SMAD2 expression in mock, vFLIP and vCyc single knockout, and WT-infected iSLK cells (Fig 5A). As previously seen in SLK cells, a high level of SMAD2 protein was detected in mock-infected iSLK cells, but was not detectable in WT-infected iSLK cells. Infection of iSLK cells with the ΔvFLIP mutant resulted in a detectable but significantly lower expression level of SMAD2. In contrast, infection with the ΔvCyclin mutant virus restored SMAD2 protein levels similar to that seen in uninfected iSLK cells. This indicates that in iSLK cells inhibition of SMAD2 by vCyclin is stronger than by vFLIP.
Since it is was recently demonstrated that SLK cells, long thought to be of endothelial origin, are actually are derived from an adenocarcinoma of epithelial origin [35], we also wanted to test vFLIP and vCyc-dependent regulation of SMAD2 in TIVE cells, an endothelial cell model in which to study KSHV pathogenesis [36]. Wt or mutant virus-infected iSLK cells, which express the RTA gene as an inducible transgene [37], were used to generate high titer virus that after quantification was used to stably infect TIVE cells. Cells were infected with 200 genome equivalents per cell and complete infection was confirmed by monitoring GFP expression, and subsequently, lysates and total RNA was collected for Western blot and RT-qPCR to monitor SMAD2 expression.
While the expression of SMAD2 in TIVE cells is lower than in SLK cells, WT KSHV infection decreased SMAD2 levels as observed in SLK cells. However, infection with either ΔvFLIP or ΔvCyc mutant viruses did not restore SMAD2 expression in TIVE cells suggesting differences by which both proteins contribute to the up-regulation of the miR-17-92 cluster between both cell types. To address this, a ΔvFLIP/ΔvCyc double knock-out mutant was generated as described above. Infection of TIVE cells with the ΔvFLIP/ΔvCyc mutant fully restored the SMAD2 expression to levels observed in uninfected TIVE cells (Fig 5B). Together, these data confirm that both KSHV vFLIP and vCyc regulate SMAD2 in the context of viral infection in cells of epithelial and endothelial origin, albeit at different efficiencies. While vFLIP or vCyclin expression alone is sufficient for down-regulation of SMAD2 in TIVE cells, both genes are required for inhibition of SMAD2 expression in SLK cells where base level SMAD expression is higher. Most miRNAs moderately modulate protein levels [38,39]. However, the observed down-regulation of SMAD2 by the miR-17-92 cluster was surprisingly strong.
To test the effects of miR-17-92 dependent targeting on SMAD2 mRNA turnover, real-time PCR was performed. Steady-state SMAD2 mRNA levels were not significantly changed in KSHV infected TIVE cells, compared to mock infected cells (Fig 5C). We observed a slight increase in SMAD2 mRNA in cells infected with single knock-out mutant. However, overall the significant decrease in SMAD2 protein, as detected by Western blotting, cannot be attributed to increased mRNA turnover. This indicates that the miR-17-92 cluster led to down-regulation of SMAD2 by mainly inhibiting translation. In summary, vFLIP and vCyclin contribute to inhibition of TGF-β signaling by transcriptionally activating the miR-17-92 cluster, which results in decreased translation of SMAD2 mRNA.
A number of KSHV latency-associated genes modulate the TGF-β signaling pathway. Firstly, KSHV-encoded miRNAs modulate directly or indirectly TGF-β signaling by targeting TGFBR2, SMAD5, and THBS-1 [24,40,41]. Secondly, LANA directly down-modulates the TGF-β receptor in PEL and endothelial cells [25]. Here, we report a third way to target TGF-β signaling by inducing host miRNA expression. vFLIP and vCyc promote increased expression of the miR-17-92 cluster, which targets SMAD2 protein synthesis via miR-17-5p, miR-18a, and miR-20. Complete loss of SMAD2 was not only observed in SLK cells over-expressing vFLIP and vCyc by transfection, but also in WT KSHV infected cells of both epithelial (SLK) and endothelial (TIVE) origin (Figs 3A, 5A and 5B). Interestingly, the individual contributions of vFLIP and vCyc to this regulatory loop were different in the two cell types. In latently infected SLK cells, deletion of vCyc was sufficient to restore SMAD2 expression to nearly wt levels whereas deletion of vFLIP did not restore SMAD2. Moreover, the rescued response to TGF-β in the presence of antagomirs was lower in vCyc-, compared to vFLIP-transfected cells (Figs 4B, 4C and 5A). In TIVE cells, which express lower levels of SMAD2, expression of either vFLIP or vCyc alone was sufficient to cause down regulation of SMAD2, and only a double knockout restored expression to wt levels (Fig 5B).
Using reporter assays and RT-PCR we demonstrate transcriptional upregulation of the miR-17-92 cluster, but how these viral genes induce this promoter has not been fully resolved. Overexpression of the miR-17-92 cluster in a Myc transgenic mouse model accelerated malignant lymphoma growth and provided the first evidence of miRNA oncogene activity [42]. We note that KSHV also de-regulates the oncomir miR-155 by either inducing miR-155 or expressing a viral ortholog miR-K12-11 (reviewed in [43]). In addition, c-Myc, a gene dys-regulated in many cancers, can transcriptionally activate the miR-17-92 promoter [18,42]. Activation of Myc has been observed in latently infected PEL cells, where it contributes to maintenance of latency [44]; however, Myc was not detectable in SLK cells by Western blot analysis (S1A Fig), indicating that in these cells miR-17-92 is not induced via Myc activation.
Sylvestre et al. reported that the transcription factor E2F1 augments transcription from the miR-17-92 promoter. Furthermore, miR-20a, a member of the miR-17-92 cluster, transcriptionally down-regulated E2F2 and E2F3 but not E2F1 [19]. Therefore, it is plausible that vCyc, an ortholog of cellular cyclinD, activates miR-17-92 expression in part through this auto-regulatory feedback loop, by inducing the E2F family transcription factors [11,45].
vFLIP is a potent activator of NF-κB signaling which is required for PEL cell survival [4]. Moreover, vFLIP down-regulates CXCR4 by miR-146a in a NF-κB dependent manner [46]. Epstein-Barr virus (EBV), a γ-herpesvirus associated with multiple malignancies, also induces several host miRNAs by LMP1 in an NF-κB dependent manner [47,48,49,50,51]. Moreover, vFLIP-induced NF-kB, also drives miR-146a expression in PEL cells. Based on these similarities, we tested if the activation of the miR-17-92 cluster by vFLIP was NF-kB-dependent. Even under conditions where NF-kB signaling was blocked by the inhibitor, Bay-11 (S1B Fig), the expression of the miR-17-92 cluster was still induced by vFLIP (S1C Fig) indicating the involvement of another signaling pathway. Interestingly, not all viral FLIP proteins activate NF-kB. Molluscum Contagiosum Virus (MCV) is a poxvirus that encodes two FLIP proteins termed MC159 and MC160, which encode two DED domains but contrary to KSHV vFLIP do not induce NFkB but rather inhibit it [52]. We therefore asked whether the novel activity of vFLIP to augment transcription from the 17/92 promoter is conserved in the MCV FLIP proteins. In co-transfection experiments the 17/92 promoter was induced between 2.5- and 5.5-fold by MC159 and MC160, respectively (S2 Fig). Hence, in addition to inhibiting apoptosis, modulating NFkB positively or negatively, and inhibiting IRF-3, viral FLIP proteins also induce the oncogenic miR-17-92 cluster. MC159 and MC160 activate the 17/92 promoter without inducing NF-kB, suggesting that KSHV vFLIP may not require NF-kB activation to induce miR-17-92 expression.
Targeting TGF-β signaling in endothelial cells via a number of different mechanisms underlines the importance of this pathway for KS sarcomagenesis. With respect to different epithelial cell tumors many reports have described a so called “TGF-β paradox” in that during the early stages of tumorigenesis TGF-β is blunted to protect cells from apoptosis, while at later stages this pathway is severely activated [53,54]. Moreover, it has recently become clear that TGF-β in endothelial cells regulates angiogenesis, a hallmark of KS tumors, either positively or negatively (reviewed in [55]). At high levels of TGF-β, signaling occurs through phosphorylation of SMAD2, 3, and 4, which in endothelial cells is antiangiogenic and in fully transformed cells is associated with NF-kB activation. Conversely, at low concentrations of TGF-β, signaling occurs through phosphorylation of SMAD 1,5, and 8 which is proangiogenic and induces proliferation and migration associated with high levels of ID1 known to be activated in KSHV latently infected cells [56,57]. Interestingly, similar opposing activities with respect to angiogenesis have recently been identified for components of the miR-17-92 cluster, although collectively expression of the miR-17-92 miRNA cluster is proangiogenic. Two 17/92 cluster miRs, miR18a/19 target the antiangiogenic secreted factor Tsp-1, thereby promoting angiogenesis in the tumor environment. Conversely miR17/20 and miR-92a target Janus kinase 1 (Jak1) and integrin a5 (Itga5) which negatively regulates vascular morphogenesis (reviewed in [55,58,59]).
Based on our new data and previously published data on LANA and viral miRNAs [24,25], we propose a model whereby KSHV latently infected cells target the TGF-β pathway and the miR17/92 cluster to protect cells from apoptosis and at the same time regulate angiogenesis through autocrine and paracrine mechanisms in both the tumor and its microenvironment. Interestingly, viral and virally-induced host miRNAs can reinforce TGF-β signal outcome to support angiogenesis by targeting SMAD2 and Tsp-1. Finally, LANA’s ability to reduce TGFBR expression at the cell surface further reduces the sensitivity of latently infected cells to TGF-β. In summary, we identified new activities for vCyc and vFLIP which via the induction of the miR17/92 cluster integrate three different viral mechanisms to promote angiogenesis via TGF-β signaling.
293T (American Type Culture Collection), SLK, and iSLK cells (NIH AIDS Research and Reagent Program), were cultured in DMEM with 10% FBS and 1% penicillin and streptomycin. Telomerase immortalized vein endothelial cells (TIVE) and long-term cultured KSHV infected cells (TIVE-LTC) have been generated in our laboratory and have been previously described [36]. TIVE cells were cultured in Medium199 with 60 μg/mL of endothelial cell growth supplement (Sigma), 20% FBS and 1% penicillin and streptomycin. iSLK or TIVE cells, infected with KSHV BAC16 [34] wild type or mutant viruses, were treated with 50 μg/mL Hygromycin for maintaining latently infected cells. TGF-β ligand was purchased from AbCam (Cat#ab50036).
Total RNA was extracted from cells using RNA-Bee reagent (Tel-Test, Inc. TX), and quality and yields analyzed using Agilent Bioanalyzer and Nanodrop. RNAs were labeled using the miRCURY LNA microRNA Array Labeling kit (Exiqon). 3’-ends of the total RNA were enzymatically labeled with the Hy3 fluorescent dye (Exiqon) using T4 RNA ligase. Labeled RNA was hybridized to the LMT_miRNA_v2 microarray, which was designed using the Sanger miR9.0 database (http://microrna.sanger.ac.uk) and custom manufactured by Agilent Technologies as 8 x 15k microarrays. 1667 unique mature miRNA sequences across all species were incorporated into 60-mer oligonucleotide probes with a 3’ linker sequence to allow separation from the glass slide surface. The Agilent linker sequence has minimal homology to any GenBank sequence. Each mature miRNA is represented by + and–(reverse complement) strand sequences, and each probe has 4 replicates within each microarray, giving 8 probes per unique mature miRNA. 10 sets of random 22mer sequences served as negative controls. Positive (normalization) controls were designed using U1, U2, U4, U5 and U6 sequences. 22mer sequences corresponding to 5’ and 3’ ends of the small nuclear RNAs were incorporated into 60 mer probes (a total of 8 x 5 x 2 x 2 = 160 probes). Additional controls such as probes to Actin, GAPDH, HSP70 and Line elements are present on the microarray. In total 3556 unique LMT seq ids (miRNA, positive and negative controls, +/- strand) were on the microarray.
A 2x Hybridization Buffer and 10x blocking buffer (Agilent) were added to the fluorescently labeled miRNAs. The samples were heated to 99°C for 3 minutes and snap-cooled before being added to the microarray printed on glass slides and hybridized for 16 h at 47°C. The glass slides were washed with the Agilent wash buffer 1 (room temperature) and 2 (at 37°C), dried with the Agilent stabilization and drying solution, and scanned using the Agilent scanner (model G2505B). The Agilent Scan Control software (version A.7.0.3) was used to produce a high resolution tif image file.
The Agilent Feature Extraction Program (FEP), version 9.5.3.1, was used to identify feature spots and extract signal intensity values. Two types of signal intensity data were used in subsequent analysis: the raw mean signal intensity of the green channel pixels in each feature spot (gMeanSignal) and the average local background signal intensity of the pixels relative to the feature spot (gBGMeanSignal). To compensate for artifacts introduced by outlier background signal intensity values due to the features position on the array, a perl script calculated the average gBGSignal for all features on the chip at the same position on the array (eight arrays per chip). The script then subtracted the average gBGSignal from each replicate feature spot signal intensity (gMeanSignal) and then calculated the average of all background subtracted replicate features per array.
The expression constructs for LANA (pcDNA3.1/LANA) and KSHV cluster miRNAs (pcDNA3.1/cluster) were described in previous reports [24,33]. For construction of vFLIP and vCyclin expression vectors, Gateway Cloning method was used. After ORFs of vFLIP and vCyclin were amplified by PCR and cloned into entry vector (pDONR222, Invitrogen) using BP recombination, ORFs were cloned into pLenti6/V5-DEST (Invitrogen) using LR recombination following the manufacturer’s procedure to create pLenti6/vFLIP and pLenti6/vCyclin. Reporter vector containing the promoter of miR-17-92 cluster upstream of luciferase vector was kindly provided from Dr. De Guire [19]. pGL3-SBE4 (Promega) contains four SMAD binding elements upstream of luciferase gene, activated by TGF-β ligand. pCMV-Renilla (Promega), expressing renilla luciferase, was used for normalization of firefly luciferase activity.
For transfection of latently infected SLK cells, electroporation (NucleofactorII, Amexa) was used with Kit V following the manufacturer’s protocol. 106 of SLK or KSHV infected SLK cells were used for each transfection. Antagomirs of miRNAs, miR-17, 18a, and 20, (Dharmacon, CO) were co-transfected with plasmids using the same methods. For luciferase assays and Western blot analysis, SLK cells were seeded 24 hours prior to transfection and transfected using TransIT-293 reagent (Mirus, WI) at 2.5 x 105 cells per well for 6-well plates or 0.5 x 105 cells per well for 24-well plates, according to the manufacturer’s protocol. Firefly luciferase activity was quantified using the Dual Luciferase Reporter kit (Promega, WI) according to the manufacturer’s protocol. 0.5 μg of reporter vectors (pGL3/PmiR-17-92 and pGL3-SBE4) were co-transfected with expression vectors of latent genes (LANA, KSHV miRNA, vFLIP and vCyclin), and harvested 48 hrs after transfection. Antagomirs were transfected together with reporter vectors and expression vectors. 2 ng/mL of TGF-β ligand was added at 24 hours post-transfection and the cells were harvested at 72 hours post-transfection. 2 ng of the pCMV-Renilla (Promega, WI) was used for luciferase analysis. FLUOstar OPTIMA reader (BMG Labtech) was utilized for measuring firefly luciferase activity, which was normalized to Renilla luciferase activity. All assays were performed as three independent experiments and standard deviation was calculated for triplicates and displayed as error bars.
Cells were harvested 48 hours after transfection or treatment with TGF-β ligand and lysed in lysis buffer (20mM HEPES, 100mM KCl, 0.2mM EDTA, 0.5mM DTT, 2.5% Glycerol, and protease inhibitor (Roche)). 5–10 μg of total protein were loaded in each lane of 10% SDS PAGE gels and transferred to PVDF membranes. Anti-V5-HRP antibody (Invitrogen,CA, 46–0708) was utilized to detect and confirm V5-tagged vFLIP and vCyclin expression from plasmid constructs. Primary antibody for detecting SMAD2 was purchased from Cell Signaling Technology Inc. (Beverly, MA, Cat #3103).
RNA-Bee (Tel-Test, TX) was utilized to extract RNAs from TIVE cells according to the manufacturer’s instructions. SuperScript III (Invitrogen, CA) was used to synthesize cDNA according to the manufacturer’s procedure. Quantitative RT-PCR (qRT-PCR) analysis was carried out using an ABI StepOne Plus system (Applied Biosystems, CA). GAPDH was used as internal control to normalize the expression of all genes. Student t-tests were performed for statistical significance compared to non-infected cells.
Mutants were constructed as described by Brulois et al. using the BAC16 backbone [34]. In brief, the individual start codon ATG was mutated to TCG for vCyclin and vFLIP, whereas in the double mutant both start codons were changed to TCG by using two step red recombineering. The Kan/I-sceI cassette was generated by using primers containing flanking regions and Kan/I-sceI amplification primers. The gel purified linear cassette was electroporated into freshly red activated (42°C) electrocompetent E.coli GS1783 cells harboring BAC16 to carry out intermolecular recombination. The transformants were grown on Kan containing LB media overnight at 30°C. After the verification of Kan insert and bacmid integrity using colony PCR and PFGE respectively, the second Red intramolecular recombination was carried out by activating arabinose inducible SceI system and temperature sensitive Red recombination system. The marker-less mutants were verified for bacmid integrity using PFGE and were sequenced using Sanger sequencing for the confirmation of replacements in all three mutants.
After quality control, wt and mutant bacmids were initially transfected into 293T cells and selected with hygromycin. Two weeks later 293T cells were induced with TPA and co-cultured with iSLK cells that contain the RTA transcactivator as an inducible transgene [37]. After 72 hours co-cultures were treated with puromycin and hygromycin to select KSHV-infected iSLK cells and to kill off residual 293T cells. Two weeks after cultivation iSLK cells were 100% GFP-positive. Subsequently, iSLK cells infected with wild type or delta vFLIP, vCyc or double mutant were induced with 1 μg/mL doxycycline and 1 mM NaB for 72 hours. The media supernatant was passed through a 0.45 μM filter virus particles were centrifuged at 100,000 x g for 1 hour. The number of virus particles was quantified by qPCR assay after viral DNA extraction using DNAzol (Molecular Research Center, Inc.). Serially diluted LANA expression plasmid was used as standard curve. Resulting virus was used to infect TIVE cells using 100 genome equivalents per cell, which yields 100% GFP positive cells 48 hrs post infection.
|
10.1371/journal.ppat.1002875 | The Irish Potato Famine Pathogen Phytophthora infestans Translocates the CRN8 Kinase into Host Plant Cells | Phytopathogenic oomycetes, such as Phytophthora infestans, secrete an arsenal of effector proteins that modulate plant innate immunity to enable infection. We describe CRN8, a host-translocated effector of P. infestans that has kinase activity in planta. CRN8 is a modular protein of the CRN effector family. The C-terminus of CRN8 localizes to the host nucleus and triggers cell death when the protein is expressed in planta. Cell death induction by CRN8 is dependent on its localization to the plant nucleus, which requires a functional nuclear localization signal (NLS). The C-terminal sequence of CRN8 has similarity to a serine/threonine RD kinase domain. We demonstrated that CRN8 is a functional RD kinase and that its auto-phosphorylation is dependent on an intact catalytic site. Co-immunoprecipitation experiments revealed that CRN8 forms a dimer or multimer. Heterologous expression of CRN8 in planta resulted in enhanced virulence by P. infestans. In contrast, in planta expression of the dominant-negative CRN8R469A;D470A resulted in reduced P. infestans infection, further implicating CRN8 in virulence. Overall, our results indicate that similar to animal parasites, plant pathogens also translocate biochemically active kinase effectors inside host cells.
| Phytophthora infestans is the causal agent of late blight on potato and tomato. It is now well established that oomycetes, such as P. infestans, secrete an arsenal of effector proteins that modulate plant innate immunity to enable infection. The Crinkler (CRN) effector family containing the LFLAK translocation motif represents a new class of cytoplasmic effectors from oomycetes. The N-terminal domain, which includes this motif, is dispensable for cell death induction in planta, indicating the presence of modular domains that are involved in distinct processes. The C-terminal/effector domain of CRN8 localizes to the host nucleus and nuclear localization is required for triggering cell death. The CRN8 effector domain has similarity to plant Serine/Threonine RD kinases. This intriguing finding raises the possibility that CRN8 mimics a specific class of plant enzymes, a feature that has been observed for several other pathogen effectors. Using biochemical assays, we confirmed that the C-terminal domain of CRN8 has kinase activity, and have identified by mass spectrometry five serines shown to be auto-phosphorylated. To the best of our knowledge, CRN8 is the first reported secreted active kinase from a microbial plant pathogen.
| Phytopathogenic oomycetes, such as Phytophthora spp., cause some of the most destructive plant diseases in the world [1]. Phytophthora spp. are hemibiotrophic pathogens, meaning that they have a two-step infection process: an early biotrophic phase (the first 2–3 days after infection), followed by a second phase characterized by extensive host tissue necrosis which enables additional growth and sporulation of the pathogen [2]. To achieve this level of host colonization, plant and animal pathogens secrete molecules, termed effectors, that interfere with host immune pathways and enable host colonization [3]–[5]. Oomycetes secrete hundreds of effector proteins that are either translocated inside host cells or act in the apoplast [6], [7]. Although some of the apoplastic effectors are inhibitors of plant hydrolases, the majority of oomycete cytoplasmic (host-translocated) effectors, namely members of the RXLR and Crinkler (CRN) families, lack similarity to known proteins and their biochemical activities remain largely unknown [6], [8]–[10]. Therefore, elucidating the molecular mechanisms underlying effector activity remains challenging and is dependent on identifying host targets of these effectors [10], [11].
In the case of bacterial plant pathogens, much has been learned from studying the role of effectors that target various host processes once delivered inside host cells. Several of these bacterial effectors, such as the protein tyrosine phosphatase HopPtoD2/HopAO1 [12], [13] and the phosphothreonine lyases OspF and HopAI [14], [15], are enzymes that alter host immunity by targeting signaling components such as mitogen-activated protein kinases (MAPKs) [3], [16]. However, most effectors of filamentous eukaryotic pathogens (oomycetes and fungi) lack similarity to known enzymes or proteins [4], [8], making functional predictions nearly impossible. One notable exception is the host-translocated metalloprotease AvrPita of the rice blast pathogen Magnaporthe oryzae [17]. More recently, Avr3b, a Phytophthora sojae RXLR effector, was shown to carry a domain with similarity to Nudix hydrolases [18]. The Nudix motif is important for AVR3b virulence function but is not required for activation of the resistance protein Rps3b [18], [19]. This report focuses on CRN8, a host-translocated effector of Phytophthora infestans that has similarity to serine/threonine kinases and is a candidate enzyme effector.
Even though functional secreted kinases have been described in animal pathogens, this class of effectors has not been reported for plant pathogens to date. Several animal pathogen kinases translocate inside host cells and perturb various host cell processes [20], [21]. The bacterium Yersinia pestis, the causal agent of the bubonic plague, secretes the virulence determinant kinase YpkA into eukaryotic host cells, which eventually leads to alterations in cell morphology [22], [23]. Toxoplasma gondii, an obligate intracellular parasite that causes toxoplasmosis, secretes kinases that are injected from secretory parasitic organelles (rhoptries) into the host cell [24]. Recent studies have shown that certain rhoptry proteins (ROPs), specifically the tandem cluster of polymorphic ROP5 pseudokinases, are essential for pathogen virulence [25], [26]. Other ROPs, for instance the kinases ROP16 and ROP18, have been previously shown to be indispensable for full virulence [27], [28]. ROP16 is secreted into the host and is redirected to the nucleus [29] where it phosphorylates the host proteins Signal Transducer and Activator of Transcription-3 (STAT3) [30] and STAT6 [31], leading to altered cytokine profiles and repression of IL-12 signaling [29]. ROP18, an active kinase that trans-phosphorylates immunity related GTPases that plays a major role in T. gondii proliferation [32]–[34]. The kinase activity of ROP18 also is essential for proteasome-dependent degradation of activating transcription factor 6 (ATF6β) [35].
The CRNs form a major class of oomycete cytoplasmic effectors that are known to alter host responses [4], [36], [37]. They were originally identified following an in planta expression screen of candidate secreted proteins of P. infestans [36]. Intracellular expression of several CRN C-termini in plants results in plant cell death and induction of defense-related genes [9], [36]. Therefore, the CRNs appear to perturb host cellular processes similar to many plant pathogen effectors, causing macroscopic phenotypes such as cell death, chlorosis, and tissue browning when expressed in host cells [36], [38]–[40]. More recently, CRN N-termini were shown to be functionally interchangeable with the N-termini of RXLR effectors, and to successfully deliver the C-terminal portion of the RXLR effector protein AVR3a inside plant cells [41]. Therefore, CRN effectors are modular proteins consisting of conserved N-termini required for translocation inside host cells and highly diverse C-termini responsible for the effector's biochemical activity [9], [33], [41], [42]. Interestingly, the CRNs appear to be chimeric, with strong evidence of recombination after the HVLVXXP motif that occurs at the end of the DWL domain just prior to the C-terminus [9]. The C-termini of CRN proteins are diverse and typically localize to plant nucleus [42]. Specifically, CRN8 requires a functional nuclear localization signal (NLS) for nuclear accumulation and cell death induction [42].
The D2 domain of CRN8 has significant similarity to protein serine/threonine kinases of the RD class [9]. RD kinases are defined as those kinases in which the conserved catalytic aspartate is preceded by an arginine residue in the kinase subdomain VI [43]. RD kinases are regulated by activation loop phosphorylation [9], [43]. Serine/threonine kinases phosphorylate the OH group of serine or threonine, while other kinases act on tyrosine, and a number (dual-specificity kinases) act on all three [44]. Protein kinases are one of the largest groups of kinases. By adding phosphate groups to substrate proteins, kinases direct the activity, localization, and overall function of many proteins, ultimately regulating almost all cellular processes. Dimerization has a crucial role in the regulation of many kinases and, as is the case with MAP kinases, can have a profound impact on their regulation and substrate selectivity [45]. Kinase activation by dimerization forms another layer of regulation. It can be exploited to interfere with normal kinase activity using dominant-negative kinase inactive mutants. For example, in cancer research the mammalian serine/threonine kinase Akt is an important regulator of cell survival and cell proliferation. However, co-expression of a kinase-inactive mutant of Akt in cells interfered with active Akt and inhibited cell proliferation and apoptotic response in tumor cells [46]. Another example is the expression of the catalytically inactive tyrosine phosphatase, HopAO1, that has a dominant-negative effect on the function of the wild type HopAO1 in infected plant cells [13].
In this study, we characterized CRN8, a secreted serine/threonine kinase of P. infestans. The predicted catalytic domain of CRN8 includes amino acid sequence 454 to 573 with a conserved catalytic aspartate (D 470) adjacent to an arginine (R 469). Here, we demonstrated that CRN8 is an active P. infestans kinase that displays catalytic activity and is auto-phosphorylated inside plant cells. Transient in planta over-expression of CRN8 resulted in cell death, and at least five phosphorylated serines were identified in CRN8 by mass spectrometry. Substitution of all five serines into alanines resulted in loss of cell death activity. Transient co-expression of wild type CRN8 and a kinase-inactive CRN8 mutant in Nicotiana benthamiana resulted in a large reduction of CRN8 cell death activity consistent with dominant negative effects. We also discovered that CRN8 is able to form dimers in planta. Finally, we showed that in planta expression of CRN8 enhances P. infestans virulence, whereas expression of the dominant-negative mutant resulted in reduced virulence.
CRN8 consists of four different conserved domains: the signal peptide (SP), the LFLAK motif, the DWL domain, and the D2 domain, which shows homology to the RD class of protein kinases (Figure 1A) [9]. Based on this domain compilation, ten functional CRN8 paralogs were predicted from the sequenced genome of P. infestans isolate T30-4, compared to three in P. ramorum, and none in P. sojae [9]. Five of the ten P. infestans CRN8 paralogs contain a predicted nuclear localization signal (NLS: amino acid sequence KGVRKKHRRA) which occurs after the D2 domain. The three P. ramorum CRN8 paralogs also have a predicted NLS (amino acid sequence: KRKRK), however, unlike the P. infestans CRN8 paralogs, their NLS sequence occurs prior to the D2 domain (Figure 1A). ClustalW alignments of the D2 domain sequences, the original CRN8 allele from P. infestans isolate 88069 [36], the 10 CRN8 paralogs, and the three P. ramorum paralogs revealed conservation of the RD kinase catalytic site (marked by the two asterisks) [9]. Nevertheless, the P. ramorum CRN8 paralog sequences share only 33% amino acid identity to CRN8 compared to 97–99% identity among the P. infestans kinase domains (Figure S1). Analysis of the Phytophthora genome sequences also revealed 38 CRN8 pseudogenes in P. infestans, four in P. ramorum and three from P. sojae (Figure 1A). The high number of P. infestans CRN8 paralogs, the presence of many pseudogenes, and the high amino acid conservation is an indication that the CRN8 family in P. infestans arose from a recent expansion.
The modular domain structure of the CRNs [9] implies that each domain may evolve separately. Taking this into account, we examined the phylogenetic relationship among CRN8 paralogs by analyzing the C-terminal amino acid sequence (the D2 domain) separately from the N-terminal sequence (including the SP, LFLAK and DWL domains). Figure 1 illustrates the two separate phylogenetic trees of the P. infestans and P. ramorum CRN8 paralogs. The N-terminal amino acid sequences in the left tree are connected to their corresponding C-terminal amino acid sequence in the right tree by a dashed line. In the N-terminal phylogenetic tree, the P. ramorum CRN8 sequences are present in different clades, yet analysis of the C-terminus shows P. ramorum D2 sequences to be grouped into a single clade when compared to P. infestans D2 sequences. From these phylogenetic analyses, we conclude that both termini evolved separately. The separate clustering of P. infestans and P. ramorum D2 domains suggests that the expansion of the CRN8 family in P. infestans occurred after the divergence of P. infestans and P. ramorum. This feature of CRN8, together with the phylogenetic studies on all P. infestans CRNs [9], is indicative of recombination between different CRN domains and points to a mechanism responsible for increasing CRN diversity. Expansion of the P. infestans CRN8 family and conservation of its D2 kinase domain suggests that CRN8 kinase activity was preserved in the P. infestans lineage.
The C-terminus of CRN8 was previously shown to induce cell death when expressed in planta [9]. To further define the domain of CRN8 that is required for cell death induction, we used A. tumefaciens-mediated transient expression to assay five N-terminal and three C-terminal deletion mutants for cell death activation in N. benthamiana (Figure 2). The only N-terminal mutant that retained the ability to give a strong cell death induction was the smallest deletion, consisting of amino acid residues 118 to 599 (Figure 2). For C-terminal deletion mutants, cell death induction was partially lost in the 118–582 deletion mutant, and completely lost in any of the larger deletions, indicating that an intact C-terminus containing amino acids 118 to 599 is required for full activity (Figure 2). These results indicate that the region required for cell death induction includes the entire kinase domain, as well as the C-terminal NLS; a finding consistent with previous reports [42]. The full-length CRN8 protein with signal peptide, caused cell death but at weaker levels than the C-terminal domain alone (Figure 2). Previous studies with oomycete cytoplasmic effectors have demonstrated stronger phenotypes when effectors are expressed in planta without secretion signals [41], [47], [48]. The presence of the signal peptide could result in miss-targeting the protein from the endoplasmic reticulum to the cytoplasm or re-entry following secretion [41], [47], [48]. From the deletion mutant analysis, we conclude that the whole D2 domain and the C-terminal NLS are required for cell death induction.
To determine if CRN8 is a functional kinase, we tested the phosphorylation ability of the D2 kinase domain in vitro. Figure 3A illustrates the used FLAG fused protein structure of the CRN8 effector, mutations were generated at the indicated RD site. Large amounts of FLAG epitope-tagged CRN8 fusion protein (amino acids 118–599) and kinase-inactive mutant proteins, FLAG:CRN8D470N and FLAG:CRN8R469A;D470A, were expressed in planta and purified by FLAG immuno-precipitation. To verify that the proteins were present, a fraction of each purified protein was subjected to SDS-PAGE and Western blot analysis with FLAG antibody (Figure 3B, upper panel). When these samples were subjected to an in vitro kinase assay with γ-P32-ATP, an auto-phosphorylated band corresponding to the FLAG:CRN8 (WT) protein was detected, but no signal was detectable for the FLAG:CRN8D470N and FLAG:CRN8R469A;D470A mutant proteins on the autoradiogram (Figure 3B, lower panel). In Figure 3C, we further examined the phosphorylation status of CRN8 and its mutants by using phosphorylation-specific staining (ProQ Diamond [49]) of crude and immuno precipitated material. Only CRN8 (WT) protein was found to be phosphorylated in both crude and immuno-purified plant protein extract (Figure 3C, top panel). While all fusion proteins were present at detectable levels (Figure 3C, middle panel), only the FLAG:CRN8 (WT) protein produced a signal when incubated with the ProQ Diamond stain [49] (Figure 3C, bottom panel). We conclude that CRN8 is an active kinase capable of auto-phosphorylation.
To test if the kinase activity of the CRN8 D2 domain is required for cell death induction, we expressed FLAG:CRN8 (WT), FLAG:CRN8D470N and FLAG:CRN8R469A;D470A in N. benthamiana by Agrobacterium tumefaciens-mediated transient expression. Macroscopic cell death was visible five days post expression for FLAG:CRN8 (WT) and FLAG:CRN8D470N proteins, but not for the FLAG:CRN8R469A;D470A protein (Figure 3D). The observation that the CRN8D470N inactive-kinase mutant still causes cell death, suggests that kinase activity is not required for CRN8-induced cell death.
To identify phosphorylation sites in CRN8, we purified FLAG and GFP epitope-tagged CRN8 wild type (amino acids 118–599) expressed in planta. Appropriate bands from immuno-purified samples were excised after SDS-PAGE separation and trypsin-digested before submitting for LC-MS/MS (LTQ-Orbitrap). From the analyzed samples we obtained 99% coverage of the CRN8 protein (Figure 4A, light yellow box is an area not observed in the MS data), and identified five phosphorylated serines in the CRN8 protein. Serines at amino acid position 249, 281, 385, 474 and 587 (indicated by the green bars in Figure 4A displayed in the FLAG tagged construct) were phosphorylated. Annotated spectra for each identified phosphorylated serine are provided in Dataset S1 and their corresponding Mascot scores in Table S1.
To determine the relevance of the phosphorylated serines for both kinase activity and cell death induction we substituted three (Figure 4A, indicated by an asterisk) or all five of the phosphorylated serines in alanine, thereby generating triple and quintuple serine to alanine FLAG:CRN8 mutant proteins. These mutants, along with FLAG:CRN8 (WT), FLAG:CRN8D470N, and FLAG:CRN8R469A;D470A, were expressed in planta and purified as described above. Phosphorylation levels were determined using the ProQ Diamond stain [49], and revealed phosphorylation of FLAG:CRN8(WT), FLAG:CRN8S248,385,585A, and FLAG:CRN8S249,281,385,474,587A proteins (Figure 4B, middle panel). As demonstrated previously (Figure 3C), no phosphorylation was detected in the FLAG:CRN8D470N and FLAG:CRN8R469A;D470A protein samples (middle panel Figure 4B). The presence of all proteins was confirmed by Western blot (Figure 4B, lower panel). The finding that our quintuple serine mutant protein was still phosphorylated (middle panel of figure 4B), indicates that other serines or threonines were also targeted for phosphorylation. Indeed, upon analysis of purified FLAG:CRN8S249,385,587A by LC-MS/MS, we detected a phosphorylated serine residue in the linker sequence between the epitope tag and the CRN8 protein sequence (data not shown).
We also determined that, in addition to the kinase-inactive FLAG:CRN8R469A;D470A mutant described above (Figure 3D), FLAG:CRN8S249,385,587A and FLAG:CRN8S249,281,385,474,587A mutants were also altered in cell death induction when expressed in leaves of N. benthamiana (Figure 4B, top panel) and Solanum lycopersicum (tomato) cultivar Money Maker Cf-0 (Figure S2). Macroscopic cell death was greatly reduced in the CRN8S249,385,587A mutant, whereas no cell death was detectable in the FLAG:CRN8S249,281,385,474,587A mutant (top panel Figure 4B, Figure S2). Our data indicate that the cell death induced by CRN8 is not a direct result of its kinase activity, but rather a consequence of the phosphorylated state of the five identified serine residues in the CRN8 protein.
To further explore the cell death induced by CRN8, we tested whether the kinase-inactive FLAG:CRN8R469A;D470A mutant could function in a dominant negative manner to suppress CRN8-induced cell death, as is the case for the HopAO1 bacterial effector [13]. Mixtures of A. tumefaciens containing pGR106-CRN8 (amino acids 118–599) and pGR106-CRN8R469A;D470A were infiltrated in various ratios into N. benthamiana leaves and scored for their ability to cause cell death. We observed that CRN8-induced cell death was greatly reduced when co-expressed with CRN8R469A;D470A, especially in wild type to mutant ratios of 1∶5, 2∶3, and 1∶1 (Figure 5). Samples of CRN8 on the left half of the leaf were co-inoculated with GFP as a negative control and show no effect on CRN8-induced cell death (Figure 5). Cell death suppression by the CRN8R469A;D470A mutant suggests that the inactive kinase has a dominant negative effect on CRN8 cell death induction, and that the inactive kinase mutant acts antagonistically on wild type CRN8.
Because co-expression of CRN8R469A;D470A with CRN8 resulted in the suppression of cell death, we tested whether changes in CRN8 protein levels due to CRN8R469A;D470A played a role in this altered phenotype. Combinations of A. tumefaciens strains containing FLAG:CRN8, GFP:CRN8, GFP:CRN8R469A;D470A, or empty vector GFP (EV:GFP) constructs were infiltrated into N. benthamiana leaves and monitored for both cell death induction and protein levels (Figure 6A–E). The first panel of figure 6B shows strong induction of cell death by FLAG:CRN8, indicating that co-expression with the EV:GFP control did not suppress the cell death response. In the second panel, co-expression of FLAG:CRN8 with GFP:CRN8R469A;D470A resulted in a great reduction of CRN8-induced cell death. The third panel shows that cell death occurred when FLAG:CRN8 and GFP:CRN8 were co-expressed, although to a lesser degree than in the FLAG:CRN8 and EV:GFP treatment. To determine if these changes were due to protein accumulation, we extracted proteins 2 days after infiltration and examined protein levels by Western blot (Figure 6C, E). Loading controls were visualized by Coomassie stain (Figures 6D, F) and indicated equal loading of all protein samples. Figure 6C shows a great reduction in FLAG:CRN8 protein levels when co-expressed with GFP:CRN8R469A;D470A, relative to the FLAG:CRN8 levels present when co-expressed with EV:GFP. We attribute this reduction to the presence of the GFP:CRN8R469A;D470A mutant protein because when FLAG:CRN8 is co-expressed with wild type GFP:CRN8, the protein levels remain unaltered (Figure 6C, E). In addition, despite substantially lower protein levels of GFP:CRN8R469A;D470A compared to GFP:CRN8 (Figure 6E), the destabilization impact of the GFP:CRN8R469A;D470A protein on FLAG:CRN8 is greater. Our data indicate that suppression of CRN8-induced cell death is probably due to a reduction in FLAG:CRN8 protein levels, suggesting that the GFP:CRN8R469A;D470A protein destabilizes the FLAG:CRN8 protein.
Kinases are known to form dimers [45], and the destabilization CRN8 by CRN8R469A;D470A suggests that these proteins may interact with one another. To test if these proteins dimerize, we co-expressed GFP- and FLAG-tagged CRN8 with FLAG-tagged CRN8R469A;D470A fusion proteins in planta and performed co-immunoprecipitation experiments. Figure 7A describes the four different co-expressed protein combinations, including two negative controls. Protein expression of all constructs was verified by Western blot prior to immunoprecipitation for FLAG-tagged proteins and post immunoprecipitation for GFP-tagged proteins (Figure 7B, C). After immunoprecipitation with α-GFP, we observed that only samples containing the GFP:CRN8 fusion (lanes 1 and 3) were able to pull down FLAG:CRN8 or FLAG: CRN8R469A;D470A proteins (Figure 7D), indicating that CRN8 proteins associate in planta in a specific manner. The PVDF membrane used for detection of FLAG-tagged protein input was stained with Coomassie blue to verify that proteins were present in relatively equal amounts in all four lanes (Figure 7E). From this experimental data, we conclude that CRN8 occurs as a dimer in planta and that dimerization is not impaired in the CRN8R469A;D470A mutant.
Effectors are thought to play an important role in pathogenicity and plant immunity. To test the extent to which the CRN8 effector increases virulence, we transiently over-expressed both active and inactive kinases, CRN8 and CRN8R469A;D470A, in P. infestans-challenged N. benthamiana leaves. Two days after inoculation with P. infestans zoospores, infected N. benthamiana leaf panels were infiltrated with A. tumefaciens expressing either GFP:CRN8 or EV:GFP. The lesion diameter indicative of pathogen spread was measured 4 and 5 days after P. infestans infection, prior to the onset of CRN8-induced cell death (5 dpi). The graph in figure 8A shows an increased rate of P. infestans lesion size in the presence of the GFP:CRN8 fusion protein relative to lesions occurring with the negative EV:GFP control. From this pathogenicity assay, we conclude that CRN8 enhances virulence of P. infestans.
We also expressed the dominant-negative mutant GFP:CRN8R469A;D470A in N. benthamiana and challenged the leaves with P. infestans 24 hours after A. tumefaciens infiltration. The lesion size was measured 5 and 10 days after P. infestans inoculation. Figure 8B shows that the kinase-inactive GFP: CRN8R469A;D470A caused a decrease in P. infestans lesion growth rate when compared to the negative control EV:GFP. Given that GFP:CRN8R469A;D470A destabilizes CRN8, it is possible that the decreased virulence we observed is caused by a reduction of the P. infestans secreted wild type CRN8 driven by the expression of the dominant-negative CRN8R469A;D470A.
In this study, we functionally characterize the CRN8 effector, a secreted kinase from P. infestans previously shown to be delivered into plant cells via its N-terminal targeting motif [42]. The P. infestans genome consists of gene-dense regions composed mainly of core ortholog genes, and gene-sparse regions which contain many fast-evolving pathogenicity effector families such as the CRNs [9]. The location of the CRN genes in these gene-sparse dynamic regions has likely allowed for the accumulation of rapid evolutionary changes and the considerable expansion of subsets of the CRNs, as evidenced by CRN8 and its homologs. The occurrence of 10 predicted CRN8 paralogs and many additional pseudogenes, all of which carry a highly conserved D2 kinase domain, suggests that the kinase activity of CRN8 is important for P. infestans CRN proteins are modular, and recombination of different domains generates an array of chimeric proteins [9]. Our phylogenetic analyses revealed that the CRN8 C-terminus, the kinase domain, has evolved separately from its N-terminal domains. Additionally, this D2 kinase domain segregates into two separate phylogenetic clusters which correspond to the different Phytophthora species, P. infestans and P. ramorum (Figure 1B). This suggests that the expansion of the CRN8 family in the P. infestans lineage occurred after its divergence from P. ramorum. Indeed, the P. infestans paralogs carry only 11 non-synonymous polymorphisms in the D2 kinase domain, a level of polymorphism consistent with recent duplication.
We discovered that CRN8 is an active kinase capable of triggering cell death even in the absence of an intact catalytic site. With the exception of the CRN8R469A;D470A double mutant, mutations that disrupted kinase activity did not alter cell death induction. Previously, we found that cell death induction by the CRN8 protein depends on its nuclear localization [42]. The absence of CRN8 cell death induction by the CRN8R469A;D470A is not due to its mislocalisation. In Figure S3 we show that YFP tagged CRN8R469A;D470A still localizes to the plant nucleus. Therefore, disruption of the kinase activity does not interfere with CRN8 nuclear localization. Interestingly, we were able to show that CRN8R469A;D470A displayed a dominant negative phenotype. Co-expression of CRN8R469A;D470A with CRN8 resulted in a significant reduction of CRN8-induced cell death (Figure 5) and P. infestans virulence (Figure 8B). The dominant-negative effect attributed to CRN8R469A;D470A can be accounted for by the destabilization of wild type CRN8 we observed when CRN8R469A;D470A and CRN8 were co-expressed (Figure 6C). Similar antagonistic interference has been noted for other kinases [50], [51], such as the mammalian Ser/Thr kinase, Akt, an important regulator of cell survival and cell proliferation [52], [53]. Co-expression of an inactive Akt kinase mutant inhibited both cell proliferation and apoptopic response in tumor cells expressing the active Akt kinase [46]. Dominant-negative effects have also been documented for plant pathogen effectors, e.g. the effect of truncated derivatives HopM1 to the HopM1 effector [54]. In planta expression of the tyrosine phosphatase HopAO1, a Type III secreted effector of the bacterium P. syringae, increases bacterial virulence. Expression of the catalytically inactive form, HopAO1C378S, interferes specifically with the function of wild type HopAO1 delivered by P. syringae by reducing its virulence [13].
We observed that in planta expression of CRN8, prior to the onset of cell death, increased virulence of P. infestans in N. benthamiana (Figure 8A). P. infestans is a hemibiotroph that colonizes living plant cells before proliferating on dead plant tissue. Our experiment suggests that CRN8 may contribute to virulence during the biotrophic phase. In addition to CRN8 and other CRNs, several bacterial Type III secretion system effectors and RXLR effectors are known to trigger tissue necrosis, browning, and chlorosis when ectopically expressed in plant cells [9], [39], [55]. The biological relevance of nonspecific cell death promotion by these effectors remains unclear as discussed by Cunnac et al., 2009 and Oh et al. 2009 [39], [55]. The promotion of cell death was proposed to reflect an excessive virulence activity on one or more effector targets [39], [55]. This view is consistent with the finding that CRN8 can enhance virulence prior to the appearance of visible cell death symptoms.
The finding that heterologous expression of CRN8R469A;D470A reduces P. infestans virulence raises the possibility of using this mutant to engineer enhanced resistance to late blight in potato and tomato. Heterologous expression of CRN8 and its various mutants in Solanum lycopersicum, tomato (Figure S2) a host of P. infestans showed similar results of cell death induction as was shown by heterologous expression of CRN8 and the CRN8 mutants in N. benthamiana. By creating plants that constitutively express the dominant negative form of CRN8, invading P. infestans strains that rely on the delivery of active CRN8 may be unable to effectively colonize their host. Our P. infestans virulence assays, show a promising restriction in pathogen spread when CRN8R469A;D470A is expressed in planta. Nonetheless, it remains to be determined whether these results will translate into enhanced resistance under field conditions.
Many questions remain about to the mode of action of CRN8 during infection. Does CRN8 act as a kinase, as a substrate, or as part of a complex, or possibly all of the above? CRN8 is likely to auto-phosphorylate, but which plant proteins are targeted and trans-phosphorylated by CRN8? In addition, CRN8 could target and activate plant kinases by forming a complex independent of CRN8 trans-phosphorylation. In the future, identification of CRN8 plant targets will result in a better understanding of how this effector contributes to virulence.
We cannot formally rule out that plant kinases are present in our CRN8 kinase detection assays (Figure 3, Figure 4 and Figure S2). However, this is unlikely because mutations in the active sites R469 and D470 abolished the detected kinase activity. We have also subjected the CRN8 immunoprecipitates, both complete mixture and excised gel fragments, to LC-MS/MS for protein composition and detection of secondary modifications and failed to identify additional proteins. Given that we have not identified other phosphorylated proteins by LC-MS/MS we conclude that the measured kinase activity is mainly due to CRN8.
We identified kinase-like proteins with similarity to CRN8 in the genomes of Arabidopsis, grape, poplar, and several other plant species (MvD and SK, unpublished). Given that kinases, including CRN8, often form dimers and heterodimers, it is possible that CRN8 heterodimerizes with plant kinases [45], influencing their proper activation and subsequent cellular activities. CRN8 might also mimic these plant kinases and target the substrates of the CRN8-like plant kinases. Kinases with closely-related catalytic domains tend to be similar in overall structural topology, have similar modes of regulation, and have similar substrate specificities [45]. Plant pathogen effectors that mimic host proteins are well-known [5]. One example of pathogen mimicry of a plant protein is the bacterial effector AvrPtoB, which is thought to mimic a substrate of a conserved plant kinase, leading to the enhanced activity of host plant kinases [56]. Another example is the secreted kinase from the animal parasitic bacterium Yersinia pestis, YpkA [22], [57]. YpkA possesses a domain that binds to the small GTPases RhoA and Rac1. This YpkA domain mimics host guanidine nucleotide dissociation inhibitors (GDIs) for the Rho family of small GTPases and as such, Yersinia pestis utilize the Rho GTPases for unique activities during their interaction [23].
To our knowledge, we describe the first plant pathogen effector, CRN8, that encodes a functional kinase domain. The next challenge is to determine the mechanism by which CRN8 functions inside plant cells, which plant process does it ‘hijack’ to interfere with plant immunity. Our study is the first step towards understanding how this secreted kinase effector perturbs the complex kinase signaling network that modulates plant immunity.
The different C-and N-terminal deletion mutants were generated by PCR amplification using full length CRN8 as a template with various forward primers that included a ClaI restriction site and reverse primers with a NotI restriction site. PCR products were digested with ClaI and NotI and were ligated into the A. tumefaciens binary potato virus X (PVX) vector pGR106 [58]. The TMV-based expression constructs were generated by amplifying CRN8 variants with the PacI restriction site and FLAG sequence embedded in the forward primer and the NotI restriction site included in the reverser primer. These amplicons were then site-directionally cloned into the pTRBO vector [59]. Ligation reactions of pGR106 and pTRBO constructs were directly transformed into A. tumefaciens GV3101 or 1D1249 by electroporation. The GFP fused clones were constructed by cloning CRN8 amplicons into the pENTR/D-TOPO (Invitrogen) entry vector followed by Gateway LR recombination (Invitrogen) into pK7WGF2 [60], resulting in GFP fused CRN8 clones. The different amino acid mutants were generated using primers that introduced the mutation by site directed PCR. For all generated constructs, primers sequences are in Table S2 and all generated sequences were verified to exclude errors.
NLS sequence prediction of the CRN8 proteins was done by NLStradamus [61] with a prediction cut off value of 0.6. This approach uses hidden Markov models (HMMs) to predict novel NLSs in proteins (http://www.moseslab.csb.utoronto.ca/NLStradamus/).
In planta transient expression by Agro-infiltration (A. tumefaciens T-DNA 35S promoter based binary constructs) or Agro-infection (PVX or TMV-based binary constructs) was performed according to methods described elsewhere [36], [59], [62] A. tumefaciens GV3101 [63] was used to deliver T-DNA constructs into 3-week-old N. benthamiana plants. Overnight A. tumefaciens cultures were harvested by centrifugation at 10,000 g, resuspended in infiltration medium [10 mM MgCl2, 5 mM 2-(N-morpholine)-ethanesulfonic acid (MES), pH 5.3, and 150 mM acetosyringone] to an OD600 = 0.3 prior to syringe infiltration into either the entire leaf or leaf sections. For experiments in which co-expression of two constructs was performed in equal ratios, each construct had an OD600 = 0.6. In experiments where different co-inoculation ratios were tested, the total amount of each PVX expression combination was OD600 = 0.6 of A. tumefaciens. The tested PVX ratios included 1∶5; 2∶3 and 1∶1 of CRN8 with CRN8R469A;D470A or ratios of 1∶5; 2∶3 and 1∶1 of CRN8 with GFP.
Proteins were transiently expressed by A. tumefaciens in N. benthamiana leaves and harvested two days post infiltration. Immunoblot analyses were performed on protein extracts prepared by grinding leaf samples in liquid nitrogen and extracting in protein extraction buffer [1 gram in 3 ml extraction buffer (150 mM Tris-HCl pH 7.5; 150 mM NaCl; 10% glycerol; 10 mM EDTA; and freshly added 20 mM NaF: 10 mM DTT; 0.5% (w/v) PVPP; 1% (v/v) protease inhibitor cocktail (Sigma); 1% (v/v) NP-40)]. Suspensions were mixed and centrifuged at 5000 rpm for 15 minutes at 4°C. The supernatant was passed through a through 0.45 µm filter before loading.
Protein samples (25 µl) were separated by SDS-PAGE (12%) and analyzed by Western blot. PVDF membranes were incubated and washed between different incubation steps with TBS-T (20 mM Tris-HCl, pH 7,5 150 mM NaCl+0.1% Tween). Monoclonal α-FLAG M2 antibody (Sigma-Aldrich) was used as a primary antibody at 1∶8000 (in 5% milk), and anti-mouse antibody conjugated to horseradish peroxidase (HRP, Sigma-Aldrich) was used as a secondary antibody at a 1∶20,000 dilution. For GFP immunoblots, monoclonal α-GFP (Invitrogen) was used as a primary antibody at 1∶4000, and anti-rabbit polyclonal antibody conjugated to horseradish peroxidase (HRP, Sigma-Aldrich) was used as a secondary antibody (1∶12,000 dilution). Blots were developed using the Pierce Horseradish Peroxidase detection kit (Thermo Scientific) and exposed for 2 min on Amersham Hyperfilm ECL (GE Healthcare). Blots were stained with Coomassie (Instant Blue, Expedeon) to visualize protein loading.
Proteins were extracted from plant material as described above and immuno-purified by FLAG or GFP affinity chromatography. For FLAG immuno-purification: 2.0 ml of extracted protein was incubated with 50 µl anti-FLAG M2 affinity matrix (Sigma) and rotated for 1.5 hr at 4°C followed by 5× wash (centrifuge 30 seconds at 800× g) with 1 ml 50 mM Tris/HCl. Proteins were eluted with 100 µl IP buffer containing 3 µl of 3xFLAG peptide (150 ng/µl), in 97 µl 50 mM Tris/HCl for 30 minutes, shaking gently at 4°C. Samples were centrifuged for 1 minute at 16,000× g and supernatants were saved for either kinase assays or in gel analysis. For the GFP immuno-purification: 1.5 ml of extracted protein were incubated with 20 µl GFP affinity matrix (Chromotek) and rotated for 4 hr at 4°C followed by 5× wash (1 ml TBS+0.5% NP-40) and centrifugation (0.5× g) to pellet beads. 40 µl of 1× Laemmli sample buffer was added and samples were denatured for 5 minutes at a 95°C boil.
45 µl of immunoprecipitated protein sample was shaken at 900 rpm for 30 minutes at 27°C with kinase assay buffer [∼185 kBq γ-P32-ATP; 20 µM ATP; 50 mM Tris/HCl (pH 7.5); 10 mM MgCl2; 10 mM MnCL2; and 1 mM DTT]. The reaction was stopped by the addition of 4× Laemmli sample buffer (15 µl) supplemented with 70 mM DTT, and was heated for 5 minutes at 95°C prior to protein separation (25 µl) on an SDS denaturing gel (12%; at 100 V). Similar steps were followed when Pro-Q Diamond phosphoprotein gel stain [49] was used for phosphorylation detection, however in those instances γ-P32-ATP was left out of the kinase assay buffer.
Detection was done directly on the SDS denaturing gel by Pro-Q Diamond phosphoprotein gel stain [49], or directly after transfer of the separated proteins onto PVDF membrane according to standard Western blotting procedures. Detection of the radio-active phosphorylation signal on the membrane and in the gel was done using the Fuji FLA 5100 phosphor imaging system.
Preparation of peptides for liquid chromatography–tandem mass spectrometry (LC-MS/MS) was performed as follows. Proteins were separated with SDS/PAGE. CRN8 fused protein bands were excised from the gel. Gel slices were prepared for LC-MS/MS as described previously [64]. Mass spectrometry LC-MS/MS analysis was performed using a LTQ-Orbitrap mass spectrometer (Thermo Scientific) and a nanoflow-HPLC system (nanoAcquity, Waters Corp.) as described previously [64]. The MS data were searched with Mascot v2.3 (Matrix Science) with the following differences: (i) The database was a custom collection of translated sequences from transcript assemblies (TIGR Plant Transcript Assemblies; http://plantta.jcvi.org) of solanaceous plants (Solanum lycopersicum, N. benthamiana, Nicotiana tabacum, Solanum tuberosum, Capsicum annuum, and Petunia hybrida) and Phytophthora infestans sequences containing 1,000,691 sequences (98,308,278 amino acid residues) with the inclusion of sequences of common contaminants such as keratins and trypsin [11]. Carbiodomethylation of cysteine residues was specified as a fixed modification and oxidized methionine and phosphorylation of serine or threonine residues were allowed as variable modifications. Other Mascot (version 2.3) parameters: mass values were monoisotopic and the protein mass unrestricted, the peptide mass tolerance was 5 ppm, and the fragment mass tolerance: ±0.6 Da, two missed cleavages were allowed with trypsin. A second Mascot search was performed allowing ‘error tolerant’ modification of all robustly identified proteins from the previous round. All Mascot searches were collated and verified using Scaffold (Proteome Software) and the subset database was searched using X! Tandem (The Global Proteome Machine Organization Proteomics Database and Open Source Software; www.thegpm.org). Accepted proteins passed the following threshold in Scaffold: 95% protein confidence, with minimum of two unique peptides at 95% confidence.
The C-termini of CRN8 (WT) and CRN8R469A;D470A were expressed via the A. tumefaciens binary Potato virus X (PVX) vector pGR106 in N. benthamiana [36]. A. tumefaciens solutions were mixed in 5 different ratios before infiltration into N. benthamiana leaves. On the left half of the leaf we co-infiltrated CRN8 (WT) with a truncated GFP construct as a negative control, and on the right half of the leaf we co-infiltrated the A. tumefaciens strains containing CRN8 (WT) and the CRN8R469A;D470A. To rule out possible concentration effects of the expressed proteins, we tested five different ratios of A. tumefaciens concentrations on multiple leaves and varied the position of the expression zones on each leaf. In addition, we used untagged proteins to exclude the possible interference of tags. Cell death phenotypes were scored five days post infiltration.
P. infestans infection assays were performed by droplet inoculations of zoospore solutions of the P. infestans isolate 88069 (10 µl of a 50,000 zoospores per mL solution) on detached N. benthamiana leaves. At least ten independent N. benthamiana leaves (4 weeks old) were tested per construct combination. For the GFP:CRN8 and the GFP (negative control) comparison, the leaves were challenged with P. infestans two days prior to A. tumefaciens-mediated expression of these constructs. For the GFP: CRN8R469A;D470A versus the GFP (negative control) comparison, proteins were first expressed by A. tumefaciens in N. benthamiana 24 hours prior to P. infestans infection. P. infestans growth efficiency was quantified by measuring the lesion size (mm) at 4 and 5 days post infection for the GFP:CRN8 versus the GFP (negative control) expressing leaves. And for GFP:CRN8R469A;D470A versus GFP-expressing leaves, P. infestans lesion size (mm) was measured 5 and 10 days after inoculation. Only successful infections at each Phytophthora inoculated spot were used to analyze the growth efficiency. The assay was repeated at least three times, and a representative dataset is used.
|
10.1371/journal.pbio.0050123 | Cyclin B1–Cdk1 Activation Continues after Centrosome Separation to Control Mitotic Progression | Activation of cyclin B1–cyclin-dependent kinase 1 (Cdk1), triggered by a positive feedback loop at the end of G2, is the key event that initiates mitotic entry. In metaphase, anaphase-promoting complex/cyclosome–dependent destruction of cyclin B1 inactivates Cdk1 again, allowing mitotic exit and cell division. Several models describe Cdk1 activation kinetics in mitosis, but experimental data on how the activation proceeds in mitotic cells have largely been lacking. We use a novel approach to determine the temporal development of cyclin B1–Cdk1 activity in single cells. By quantifying both dephosphorylation of Cdk1 and phosphorylation of the Cdk1 target anaphase-promoting complex/cyclosome 3, we disclose how cyclin B1–Cdk1 continues to be activated after centrosome separation. Importantly, we discovered that cytoplasmic cyclin B1–Cdk1 activity can be maintained even when cyclin B1 translocates to the nucleus in prophase. These experimental data are fitted into a model describing cyclin B1–Cdk1 activation in human cells, revealing a striking resemblance to a bistable circuit. In line with the observed kinetics, cyclin B1–Cdk1 levels required to enter mitosis are lower than the amount of cyclin B1–Cdk1 needed for mitotic progression. We propose that gradually increasing cyclin B1–Cdk1 activity after centrosome separation is critical to coordinate mitotic progression.
| When active, the enzyme cyclin B1–cyclin-dependent kinase 1 (Cdk1) commits a growing cell to the process of mitotic cell division and chromosome separation. Cyclin B1–Cdk1 activation is controlled in many ways, but once its activity rises above a certain level, further activation of cyclin B1–Cdk1 is catalyzed by a positive-feedback loop. This generates highly active cyclin B1–Cdk1 and triggers the start of mitosis, which can only be completed when cyclin B1–Cdk1 activity is properly shut off again. However, it is not clear how cyclin B1–Cdk1 activity develops in human cells or how the switch between its inactive and active states is controlled. Our work combines activation measurements with a kinetic model to study how cyclin B1–Cdk1 activity accumulates just before and during mitosis. We show that cyclin B1–Cdk1 activity develops gradually in early mitosis and that different activity levels are required for initiation of, and progression through, mitosis. We also demonstrate that once cyclin B1–Cdk1 activation is truly launched, it is bound to continue and will not lightly drop back again. We propose that the successive cyclin B1–Cdk1 activity levels by themselves may coordinate the progression through the distinct phases of mitosis.
| Mitotic entry is catalyzed by the kinase activity of cyclin-dependent kinase 1 (Cdk1) in complex with cyclin B1 [1]. Cyclin B1 levels first rise during G2 phase, which allows the accumulation of cyclin B1–Cdk1 complexes [2]. At this stage, cyclin B1–bound Cdk1 is kept inactive by phosphorylation of T14 and Y15 by Wee and Myt kinases [3,4]. Subsequently, cyclin B1–Cdk1 can be activated by Cdc25 phosphatases [5]. They dephosphorylate Cdk1 in two separate binding steps, in which T14 is dephosphorylated before Y15 [6,7]. In interphase Xenopus egg extracts, cyclin B1–Cdk1 complexes have low activity. Cdk1 activity in cyclin B1 immunoprecipitates rises gradually through G2, eventually reaching approximately 30% of its maximum activity at the end of G2 [8–10]. At this point, a threshold concentration is reached that autocatalyzes rapid further activation of cyclin B1–Cdk1 and triggers entry into mitosis [10].
Cyclin B1–Cdk1 activation contributes to the separation of centrosomes in late G2 [11,12]. In human cells, active cyclin B1–Cdk1 is initially detected on centrosomes, shortly before they start to migrate apart at the end of G2 or in prophase [13,14]. Active cyclin B1–Cdk1 is then phosphorylated on its cytoplasmic retention sequence, leading to nuclear translocation of cyclin B1–Cdk1 and, subsequently, to enhanced chromosome condensation and nuclear envelope breakdown [15,16]. When cells are in mitosis, the paired sister chromatids need to be correctly connected to the mitotic spindle, a process that is governed by the spindle checkpoint. This checkpoint inhibits ubiquitin ligase activity of the anaphase-promoting complex/cyclosome (APC/C), thereby preventing premature cyclin B1 destruction and Cdk1 inactivation. The sustained high cyclin B1–Cdk1 activity allows cells to stay in mitosis as long as is required for all chromosomes to attach to the mitotic spindle in a bioriented fashion [17,18]. After requirements for the checkpoint have been met, progressive loss of cyclin B1–Cdk1 activity is essential for successful chromosome segregation and completion of cell division [17].
It has been proposed that such a cyclin B1–Cdk1–APC/C module could function as an autonomous oscillator governing cell cycle progression [19–21]. Based on the ability of cyclin B1–Cdk1 to inhibit its inhibitor Wee1 [22–25], to activate its activators of the Cdc25 phosphatase family [26,27], and to stimulate cyclin B1 destruction [28], two major theories of mitotic cyclin–cyclin-dependent kinase (Cdk) activation have been proposed (see Text S1 and Figure S1 for additional information on these models). Some models suggest a limit cycle behavior, in which the levels of Cdk activity will never be stable but oscillate between an inactive and an active state [19,29,30]. Other models describe a bistable system, in which thresholds for activation and inactivation differ and Cdk1 is either inactive, active, or approaching one of these stable states [31–33]. One of the differences between these systems is that, in a bistable system, once a threshold level of cyclin B1–Cdk1 complexes is reached and activation has proceeded, cyclin B1–Cdk1 will not be inactivated even though the cyclin B1–Cdk1 levels are reduced to levels below the activation threshold. This feature of a bistable system is referred to as hysteresis. A commitment of cells to start and complete mitosis once they have activated cyclin B1–Cdk1, is strikingly reminiscent of certain features of bistability. Various models and some experimental data indeed point to bistability governing mitotic Cdk1 activation [34–36] (Figure S1). It has also been argued, however, that the experimental set-ups used so far may have created an artificial bistable system [29]. Furthermore, recently, Csikasz-Nagy and coworkers [37] suggested that both bistability and limit cycle oscillations occur, depending on the cellular context.
Although it is clear that cyclin B1–Cdk1 is mostly inactive in G2 and highly active in mitosis, methods to follow the development of cyclin B1–Cdk1 activity during the rapidly successive mitotic stages have been lacking. As a result, experimental data of Cdk1 activation kinetics in human cells are very limited. Such data could, for instance, reveal whether cyclin B1–Cdk1 activation, once initiated, is sufficiently robust to drive mitotic progression even in the cytoplasm, when a large fraction of cyclin B1–Cdk1 translocates to the nucleus in prophase. This would be indicative of a bistable Cdk1 activation response and further indicate that precise regulation of cyclin B1–Cdk1 activity could control mitotic events per se.
We use novel assays to determine how cytoplasmic cyclin B1–Cdk1 activation proceeds in human cells. We find that cyclin B1–Cdk1 is activated in late G2 and that the activity gradually increases between centrosome separation and prometaphase. Moreover, we investigated the effect of partially reducing Cdk1 levels on mitotic progression. We show that whereas low levels of Cdk1 are sufficient for mitotic entry, higher levels are needed for normal mitotic progression and initiation of anaphase. Altogether, we show that the development of cyclin B1–Cdk1 activity in vivo proceeds as if it were governed by bistability. Our results explain how cyclin B1–Cdk1 can be responsible for cytoplasmic rearrangements even during nuclear translocation. We propose a model in which different thresholds of cyclin B1–Cdk1 activity, in combination with a gradual increase in activity, help to coordinate early and late mitotic events.
Cyclin B1–Cdk1 is initially activated on centrosomes in late G2, shortly before they start to migrate apart [13,14]. To determine how cyclin B1–Cdk1 activation proceeds, we developed an assay to follow the pattern of cyclin B1–Cdk1 activity in single human cells. Because Cdk1 levels are in excess of cyclin levels in mammalian cells [38] (unpublished data) and because Cdk1 is phosphorylated only when in complex with a cyclin [10], we reasoned that the ratio between phosphorylated Cdk1 and cyclin B1 could provide an estimate of the relative activity of the cyclin B1–Cdk1 complex. Therefore, we used an antibody recognizing Cdk1 when phosphorylated on Y15 (Cdk1 phosphorylation [Cdk1-P], which represents inactive Cdk1) combined with an antibody recognizing cyclin B1, in immunofluorescence experiments.
We first wanted to test the specificity of these antibodies, and therefore we used short hairpin RNA (shRNA) to reduce the levels of cyclin B1 or Cdk1 in cells. The cytoplasmic staining of the Cdk1-P antibody in G2 almost completely disappeared by microinjecting cells with an shRNA to cyclin B1 or Cdk1, showing the cytoplasmic signal represents phosphorylated Cdk1 (Figure S2). However, Cdk1 RNA interference (RNAi) could not totally eliminate the staining in interphase nuclei and on centrosomes, indicating some cross-reactivity in these locations, most probably with phosphorylated Cdk2. Furthermore, the cyclin B1 and Cdk1-P signal colocalized linearly in the cytoplasm, indicating that the signal represents cyclin B1–Cdk1 complexes (Figure S2). Therefore, in this study we exclusively focused on cytoplasmic cyclin B1–Cdk1 activation, which we could monitor specifically.
Figure 1 shows maximum intensity projections of cells in various stages of G2 and mitosis, costained for chromosomes, phospho-Cdk1, and cyclin B1. In G2 and mitotic cells, cyclin B1, shown in the second panel, colocalized with cytoplasmic phospho-Cdk1, apparently decreasing around the time when cyclin B1 (third panel) translocated to the nucleus. In metaphase, cyclin B1 levels declined, whereas Cdk1-P appeared to remain similar to the level in prometaphase cells. In anaphase, both cyclin B1 and phospho-Cdk1 levels had largely disappeared.
Subsequently, we analyzed numerous undeconconvolved fluorescent images of G2 and mitotic cells, to carefully quantify the cytoplasmic cyclin B1 and Cdk1-P staining in various phases of G2 and mitosis (see Materials and Methods and Figure S3 for details on image acquisition and data analyses). To control for changes in morphology during mitotic progression, we performed quantitations of nuclear factor κB (NF-κB), a protein that is cytoplasmic in normal cells and has no obvious function in mitosis (Figure S4). We exclusively analyzed endogenous proteins in unsynchronized HeLa cells under normal growth conditions.
After correction for changes in morphology, we subsequently plotted cytoplasmic Cdk1-P levels as a function of cytoplasmic cyclin B1 levels. We observed that both Cdk1-P and cyclin B1 increased in a roughly linear fashion when cells proceeded through G2, with little change in the Cdk1 phosphorylation (P) state, in agreement with cyclin B1–Cdk1 complexes displaying low activity in early G2 (Figures 1 and 2A; blue diamonds). As we showed previously [14], a subset of cells with nonseparated centrosomes and high cyclin B1 levels contained reduced levels of phosphorylated Cdk1, indicating a pool of cyclin B1–Cdk1 is partially active. In cells with high cyclin B1 levels, we saw Cdk1 dephosphorylation further proceeding when centrosomes started to migrate apart. The dephosphorylation gradually continued until chromosomes were aligning in prometaphase (Figures 1 and 2A; top left, red squares and light-blue Xs), whereas total Cdk1 levels did not decline (Figure S5). This strongly indicates cyclin B1–Cdk1 activation progressively develops during mitosis.
Subsequently, we focused on the development of cyclin B1–Cdk1 dephosphorylation when cyclin B1 enters the nucleus due to phosphorylation of its cytoplasmic retention signal in prophase [15,39]. It should be noted that the effect of nuclear translocation on development of cyclin B1–Cdk1 activity is puzzling. Translocation means a drop in the cytoplasmic cyclin B1 concentration, which could cause cytoplasmic Cdk1 activity to plummet. However, cytoplasmic cyclin B1–Cdk1 activity seems required to govern mitotic onset in the cytoplasm, e.g., to induce rearrangements such as microtubule polymerization and spindle assembly, involved in promoting nuclear envelope breakdown at the end of prophase [40,41]. In a bistable model, though, a sudden reduction of cyclin B1 from the cytoplasm at the end of prophase would not necessarily prevent cytoplasmic activation of cyclin B1–Cdk1 to proceed (see Text S1 and Figure S1).
When we compared cyclin B1 and phospho-Cdk1 levels before and after the onset of translocation (Figure 2A, red squares and green triangles), we observed that during cyclin B1 nuclear translocation, both the cytoplasmic cyclin B1– and Cdk1-P levels continued to decrease (Figures 1 and 2A; also see statistical analyses of the data in Figures 2C and S7). To further visualize the details of cyclin B1–Cdk1 dephosphorylation during nuclear translocation, we plotted the nuclear-to-cytoplasmic cyclin B1 ratio as a function of the ratio between cytoplasmic Cdk1-P and cyclin B1 in Figure 2B. When the ratio between phosphorylated Cdk1 and cyclin B1 was approaching 50% of that observed in early G2 cells, cyclin B1–Cdk1 translocated to the nucleus. Strikingly, in cells with translocated cyclin B1, the ratio between cytoplasmic Cdk1-P and cyclin B1 indeed continued to decrease, demonstrating the activation did proceed in the cytoplasm, too (Figure 2B). Thus, our results show that a pool of active cyclin B1–Cdk1 remains in the cytoplasm even throughout the nuclear translocation, potentially explaining how cyclin B1–Cdk1 activity can contribute to mitotic rearrangements taking place outside the nucleus at this time. Furthermore, these results indicate that once a positive feedback has been initiated, the activation of cyclin B1–Cdk1 is remarkably robust, and even proceeds upon a sudden drop in cyclin B1 concentration of approximately 50% (Figure 2C).
Total cyclin B1–Cdk1 activity in the cytoplasm is dependent on the concentration of the complex as well as on the fraction of active complexes present. We therefore reasoned that by multiplying the “relative activity” with the cyclin B1 levels, we would acquire a rough estimate of the “total cyclin B1–Cdk1 activity.” To estimate the relative activity, we assumed the ratio between P-Cdk1 and cyclin B1 in early G2 cells represented inactive complexes (Figures 2D and 2F). We then continued to plot the total cytoplasmic cyclin B1–Cdk1 activity as a function of cytoplasmic cyclin B1 levels, obtaining a loop representing activation and deactivation in G2 and mitosis (Figure 2D and 2E). In this representation, we see that the total cyclin B1–Cdk1 activity starts to increase in late G2, continues to rise to prometaphase, and remains high until cyclin B1 levels start to decline by inactivation of the spindle checkpoint and activation of Cdc20-APC/C in metaphase. The plot generated by our experimental data markedly resembled the relationship between in vitro Cdk1 activity and cyclin B concentration, observed in Xenopus egg extracts [9], with the important difference that human cytoplasmic cyclin B1 levels decreased as a result of nuclear translocation in prophase (Figure 2D, 2E, and 2F). Our results show that the principle of the cell-cycle oscillations between cyclin-Cdk activation and inactivation thus appears to be conserved in Xenopus and human cells.
Next, as an independent approach to determine cyclin B1–Cdk1 activity in human cells, we aimed to extend our analyses to a direct target of cyclin B1–Cdk1. As an additional benefit, this would allow us to confirm the kinetics of cyclin B1–Cdk1 activation when quantifying a positive readout signal. To this end, we used purified antibodies specifically recognizing phosphorylated APC3, a cyclin B1–Cdk1 substrate, in mitosis [42,43]. Figure S6A clearly shows the accumulation of APC3–Thr244-P on centrosomes after cyclin B1 translocation and during prometaphase, while the signal stabilized in metaphase and was greatly reduced in anaphase. Because cytoplasmic staining of this epitope was weak, next we analyzed the appearance of APC3–Thr446-P. Phosphorylation of this site was first identified in mitotic cells and has been reported to be induced by Cdk1, but not by Polo-like kinase 1, in vitro [44]. We saw the emergence of APC3–Thr446-P when centrosomes separated and cyclin B1 started to translocate to the nucleus in prophase (Figure S6B). For comparison, in the same manner as for Cdk1 dephosphorylation, we next quantitated the increase in APC3–Thr446-P exclusively in the cytoplasm. Figure S6C (gray bars) shows that cytoplasmic APC3–Thr446-P first rose slightly in G2 cells, slowly increasing after centrosome separation, and, interestingly, continued to accumulate while cyclin B1–Cdk1 translocated to the nucleus. Figure 3 shows a similar but more detailed quantification throughout G2 and mitosis of APC3–Ser426-P, a well-established Cdk1 target in vitro and in vivo, which is phosphorylated independently of Polo-like kinase 1 in vivo [44]. Accumulation of cytoplasmic APC3–Thr446-P and APC3–Ser426-P was highly comparable (Figure S6C). Cdk1-specific APC3 phosphorylation continued throughout prometaphase, indicating a need for enhanced cyclin B1–Cdk1 activity in prometaphase.
It should be noted that whereas the direct measurements in Figure 2 denote the activity of the cyclin B1–Cdk1 complex when the cell is fixed, the measurements in Figure 3 represent the accumulated sum of kinase as well as phosphatase activities. Moreover, APC3 is situated in both the nucleus and the cytoplasm, leading to an increased accessibility to cyclin B1–Cdk1 after nuclear envelope breakdown. Therefore, the detected APC3 phosphorylation lags slightly behind the detected Cdk1 dephosphorylation. However, both measurements showed a similar trend in gradually increasing cyclin B1–Cdk1 activity, in which the direct activity precedes the accumulated resulting phosphorylation.
To correlate the duration of Cdk1 activation with mitotic progression, we determined the timing of specific G2 and mitotic events: centrosome movement, DNA condensation, chromosome congression and sister chromatid separation in HeLa cells (Figure S8). In late G2, the centrosomes of most cells migrated from a position next to the nucleus to below the nucleus, where they separated. The activation of cyclin B1–Cdk1 preceded centrosome separation, but during the 27 min (SD = 13 min) between centrosome separation and DNA condensation, almost 50% of the cytoplasmic cyclin B1–Cdk1 complexes were activated (Figure 4, black line). At the same time, the amount of Ser426-phosphorylated APC3 started to increase (Figure 4, dotted line). Due to the nuclear translocation of cyclin B1–Cdk1, the total cytoplasmic activity was moderately decreased when DNA condensation became visible (Figure 4, gray line). We showed in Figure 2 that even during nuclear translocation, activation of cytoplasmic cyclin B1–Cdk1 complexes could continue. We also showed that accumulation of APC3–Ser426-P did not decrease during nuclear translocation. The chromosomes started to align on the metaphase plate about 20 min (SD = 9 min) after DNA condensation was initiated. During this process, cyclin B1–Cdk1 gained full activity, and as a consequence, the most dramatic increase in APC phosphorylation occurred.
In conclusion, cyclin B1–Cdk1 activity continued to accumulate between centrosome separation and chromosome congression in HeLa cells during a time span of approximately 45 min. The dephosphorylation of cyclin B1–Cdk1 proceeded in a gradual fashion, and the development of total cytoplasmic activity was modestly influenced by localization changes of the cyclin B1–Cdk1 complex (Figure 4).
If bistability and hysteresis govern cyclin B1–Cdk1 activity, Cdk1 should not be phosphorylated and inactivated when the cyclin B1 concentration drops slightly below the activation threshold [45]. This occurs twice in an unperturbed cell cycle: during nuclear translocation of cyclin B1–Cdk1 in late G2, when cytoplasmic cyclin B1 levels decrease to the level of G2 cells in which cyclin B1–Cdk1 is inactive, and when cyclin B1 starts to be degraded in metaphase (Figures 1, 2D, 2E, and 2F). Indeed, we do not observe any increase of Cdk1-P during these events. Although we need to stress that this does not prove that mitosis is controlled by a bistable cyclin B1–Cdk1 circuit (since the system may not reach a steady state), we show that within the time-frame in which concentration changes occur in vivo, the activation curve for human cyclin B1–Cdk1 proceeds as if it were governed by bistability. It is probable that this contributes to the robustness and unidirectionality of mitosis, especially during the dramatic decrease in cytoplasmic cyclin B1–Cdk1 levels during nuclear translocation.
In a recent spatial theoretical simulation, Yang and coworkers [30] found that removing cyclin B1–Cdk1-dependent destruction of cyclin B1 turns the activation dynamics from limit cycle oscillations to bistability. Interestingly, in this model the cyclin B1–Cdk1 activity continues to rise in the cytoplasm during nuclear translocation. It should be noted that the negative feedback loop, leading to cyclin B1 inactivation, is not directly set off by the active state of cyclin B1–Cdk1. Rather, active cyclin B1–Cdk1 triggers mitotic entry, and the mitotic state is maintained by the spindle checkpoint, which prevents early activation of the APC/C-dependent negative feedback, even after the APC/C has already been phosphorylated by cyclin B1–Cdk1. This means that once cells are in the mitotic state with high cyclin B1–Cdk1 levels, they can remain in this state even when the negative feedback has been “preloaded” (i.e., the APC/C is maximally phosphorylated by cyclin B1–Cdk1). During a prolonged arrest, cyclin B1–Cdk1 activity remains high for hours in the checkpoint-arrested, mitotic state. Eventual slippage through the checkpoint is caused only by progressive, but very slow, degradation of cyclin B1 [46]. In normal mitotic cells, cyclin B1 is destroyed only when spindle checkpoint silencing releases inhibition of the APC/C, and Cdk1 rapidly returns to its low, interphase-state, activity.
Abrupt and efficient Cdk1 activation at the beginning of mitosis might be required to eventually trigger normal Cdk1 inactivation, which is necessary to build an uncompromised Cdk1-APC/C cell cycle oscillator in cycling extracts [9]. Rudner, Hardwick, and Murray reported that a slight decrease in specific Saccharomyces cerevisiae Cdk1 activity (by a mutation rendering active Cdc28 inefficient in ATP binding, or by mutation of the Cdc28 subunit Cks1) prevented Cdk-dependent hyperphosphorylation of the APC/C, resulting in stabilization of cyclins and Pds1 (S. cerevisiae Securin) and mitotic arrest [47]. These findings indicate that specific cyclin-Cdk1 kinase activity needs to accumulate during mitosis to prepare for efficient mitotic exit.
Next, to investigate the requirement for cyclin B1–Cdk1 at different phases of mitosis in human cells, we reduced the expression levels of Cdk1 by vector-driven shRNA (Figure 5A). As previously reported [48], we observed that reducing Cdk1 levels caused an increase in the population of G2 cells and a reduced fraction of cells entering mitosis after release from a G1/S block (unpublished data; see also Figure S9). However, approximately 20% of the cells with reduced Cdk1 levels initiated mitosis. When we compared the levels of Cdk1 in G2 and mitotic cells (collected by shake-off) after Cdk1 RNAi treatment, Cdk1 was selectively more abundant in the mitotic fraction (Figure 5A). This indicates that only cells with modestly reduced Cdk1 levels progressed from G2 to mitosis and suggests the existence of a threshold for Cdk1 at mitotic entry. These Cdk1-attenuated cells, which initiated mitosis, clearly showed reduced phosphorylation of the Cdk1 substrates APC3 and Cdc25C when arrested in nocodazole (Figure 5B), revealing that although the cyclin B1–Cdk1 activity was sufficient for entry into mitosis, phosphorylation of cyclin B1–Cdk1 targets remained incomplete. Coexpression of an RNAi-resistant Cdk1 construct together with the shRNA vector restored APC3 phosphorylation (Figure 5C), confirming the specificity of the RNAi. Biochemical analysis of cyclin B1–Cdk1 activity in cells that entered mitosis with reduced cyclin B1–Cdk1 levels revealed that the activity per complex remained largely unchanged, showing that the remaining cyclin B1–Cdk1 can still be activated even when total cyclin B1–Cdk1 levels are reduced, a result in agreement with a model of bistability (Figure S10). Release from nocodazole arrest revealed that Cdk1-attenuated cells retained the ability to exit from mitosis after 3–4 h (Figure 5B). However, when analyzed by time-lapse microscopy, cells that entered mitosis with reduced Cdk1 levels were strongly delayed before initiation of anaphase, with their chromosomes aligned at the metaphase plate (Figures 5D, 5E, and S9; average of 15 cells: 88 ± 65 min compared with 5 ± 3 min in 15 control cells, cotransfected with Cdk1 shRNA and RNAi-resistant Cdk1–yellow fluorescent protein [YFP]). Similar results were obtained in HeLa cells (unpublished data). Given the large number of Cdk1 substrates [49], it is likely these aberrancies can be attributed to impaired function of many cyclin B1–Cdk1 targets, which prohibited us from pinpointing a single mitotic process being affected. Nevertheless, taken together these results indicate that mitotic cyclin B1–Cdk1 activity needs to further increase after mitotic entry and pass a threshold before metaphase, necessary to allow activation of the machinery required for efficient mitotic progression and initiation of mitotic exit.
An important consequence of our findings is that distinct thresholds for cyclin B1–Cdk1 activity exist that control successive mitotic events. This concept has recently been demonstrated in mitotic exit, which requires the passage of different thresholds of decreasing cyclin B1–Cdk1 activity for the metaphase-to-anaphase and the anaphase-to-telophase transition [50,51]. Here, we propose a model where the cyclin B1–Cdk1 activity threshold for mitotic exit exceeds the threshold for mitotic entry, in line with observations in yeast [47] (Figure 6). It is tempting to speculate that the gradual activation of cyclin B1–Cdk1 during early mitosis, and the passage of distinct thresholds, critically assists in coordinating early and late mitotic events.
HeLa and U2OS cells were cultured in DMEM supplemented with Glutamax, 10% fetal calf serum, and antibiotics (GIBCO, http://www.invitrogen.com). Double-stranded oligos encoding shRNAs targeting human Cdk1 or cyclin B1 mRNA were cloned into the pSuper RNAi vector, under control of the H1 RNA promotor. Specificity and robustness of knock-down by various targeting vectors were tested by FACS analyses and Western blots using mouse anti-Cdk1 for detection, according to described procedures [52]. In this study, we used pSuper constructs driving short hairpins to target GGGGATTCAGAAATTGATC of human Cdk1 or a mixture of two plasmids to target GAACAGCTCTTGGGGACAT or GATGCTGCAGCTGGTTGGT of human cyclin B1. Specificity and stringency of the shRNA-induced gene knock-down were rigorously tested for the indicated RNAi experiments by Western blots, in comparison with unrelated shRNAs and rescue experiments of Cdk1 shRNA-induced phenotypes.
HeLa or U2OS cells were cotransfected using calcium phosphate and HBS with 5 or 10 μg of the indicated pSuper constructs, in the presence of 1 μg of α-tubulin–pEYFP as a reporter or 1 μg of pBABE-Puro as a selection marker, per 8-ml culture medium in 9-cm dishes of subconfluent cells or the equivalent for smaller dishes [52]. At 12 h after transfection, cells were selected with puromycin where indicated and synchronized in the presence of thymidine. Cells were released from thymidine block and puromycin selection after 20 h. Mitotic shake-offs were performed by a single gentle wash with warm PBS at 12 h after release from thymidine and selection or 14 h after release and 5 h after the addition of nocodazole (250 ng/ml). For shRNA-rescue experiments, a YFP-tagged cDNA construct of Cdk1 was generated harboring three silent mutations within the shRNA targeting regions and cotransfected with the Cdk1 shRNA vector at a 1:10 ratio. At 12 h after thymidine release, cell extracts were prepared using ELB buffer supplemented with protease and phosphatase inhibitors. Equalized protein samples were analyzed by Western blots using mouse anti–cyclin B1 (GNS1), mouse anti-Cdk1 (BD Transduction Laboratories, http://www.bdbiosciences.com), rabbit anti-Cdc20 (Santa Cruz Biotechnology, http://www.scbt.com), mouse anti-APC3 (BD Transduction Laboratories), or rabbit anti-Cdc25C (Santa Cruz).
For three-dimensional time-lapse microscopy in Figure S8, see figure legend. For the analyses of mitotic progression after depletion of Cdk1, U2OS cells were plated onto glass-bottomed dishes (Willco Wells, http://www.willcowells.com) and transfected with 500 ng of pSuper-Cdk1 or control (either empty pSuper or pSuper targeting human Cks2, which is redundant in mitosis; R. M. F. Wolthuis, unpublished observations, and [53]), mixed with 40 ng of α-tubulin–YFP. Cells were transfected with indicated shRNAs and maintained in a climate-controlled culture chamber at the stage of a Zeiss Axiovert 200M microscope equipped with a 1.30 NA ×40 AxioPlan objective, specific dual band-pass filters, a 100-W xenon fast-shutter excitation device (DG4), and a Roper HQ Coolsnap CCD camera as shown previously [52]. Acquisition of DIC and fluorescence images started 40 h after transfection, at 100-ms exposure times. Images were captured and analyzed using MetaMorph and PhotoShop software.
Cells were grown on hexametaphosphate/metasilicate-coated coverslips and fixed with PBS containing 3% paraformaldehyde and 2% sucrose for 10 min followed by permeabilization for 2 min in −20 °C methanol. Cells were incubated for 2 × 20 min in PBS containing 50 mM ammonium chloride before blocking in 2% bovine serum albumin and labeled with mouse anti–cyclin B1 (Santa Cruz, GNS1), rabbit anti-Y15P Cdk1 (Cdk1-P) (Cell Signaling, http://www.cellsignal.com), mouse anti-Cdk1 (Cell Signaling, POH-1); rabbit anti-phospho APC3-S426, rabbit anti-phospho APC3-S446 (kindly provided by Jan Michael Peters, IMP, Vienna, Austria), or rabbit anti-phospho APC3–Thr-244 (Abcam, http://www.abcam.com). After labeling with Alexa 488 anti-rabbit (Molecular Probes, http://www.probes.invitrogen.com), Cy3 anti-mouse (Jackson Immunoresearch, http://www.jacksonimmuno.com), and Hoechst 33342, the coverslip was mounted in Vectashield (Boehringer Mannheim, http://www.roche.com) or Mowiol 4–88 solution with antioxidants. z-stacks with 0.2-μm spacing were acquired using a Deltavision Spectris imaging system equipped with a ×60 objective, NA 1.4. A maximum intensity projection of the z-levels (SoftWorx; Applied Precision, http://www.api.com) is shown in Figure 1.
Fifteen image z-stacks with 1-μm spacing were acquired in multiple locations of a single coverslip, using a Deltavision Spectris imaging system equipped with a ×20 objective, NA 0.7. To avoid bleaching of antibody fluorescence, cells were focused using the Hoechst labeling. The images were corrected for variation in illumination as determined by a photosensor (Applied Precision). For each z-stack, the background signal of an area without cells was subtracted. z-stacks where the background Cdk1-P signal differed more than 5% from the average background signal were not used. The average intensity of a square with five-pixel (3.3-μm) side was measured at the z-level that gave the highest average intensity in the cytoplasm, at a place where no obvious structures (e.g., centrosome, spindle, strong dotted staining) were present, of 318 cyclin B1–expressing cells using SoftWorx (Applied Precision). For examples of quantification, see Figure S3. The signal from prometaphase, metaphase, and anaphase cells was corrected for morphological difference by division of the signal by 1.38, as assessed by the difference in cytoplasmic NF-κB staining between G2 and mitotic cells Figure S4. The average background signal of cells not expressing cyclin B1 was quantified and subtracted from the measurements, with morphology corrections for prometaphase, metaphase, and anaphase cells.
|
10.1371/journal.pcbi.1003071 | Constraint and Contingency in Multifunctional Gene Regulatory Circuits | Gene regulatory circuits drive the development, physiology, and behavior of organisms from bacteria to humans. The phenotypes or functions of such circuits are embodied in the gene expression patterns they form. Regulatory circuits are typically multifunctional, forming distinct gene expression patterns in different embryonic stages, tissues, or physiological states. Any one circuit with a single function can be realized by many different regulatory genotypes. Multifunctionality presumably constrains this number, but we do not know to what extent. We here exhaustively characterize a genotype space harboring millions of model regulatory circuits and all their possible functions. As a circuit's number of functions increases, the number of genotypes with a given number of functions decreases exponentially but can remain very large for a modest number of functions. However, the sets of circuits that can form any one set of functions becomes increasingly fragmented. As a result, historical contingency becomes widespread in circuits with many functions. Whether a circuit can acquire an additional function in the course of its evolution becomes increasingly dependent on the function it already has. Circuits with many functions also become increasingly brittle and sensitive to mutation. These observations are generic properties of a broad class of circuits and independent of any one circuit genotype or phenotype.
| Many essential biological processes, ranging from embryonic patterning to circadian rhythms, are driven by gene regulatory circuits, which comprise small sets of genes that turn each other on or off to form a distinct pattern of gene expression. Gene regulatory circuits often have multiple functions. This means that they can form different gene expression patterns at different times or in different tissues. We know little about multifunctional gene regulatory circuits. For example, we do not know how multifunctionality constrains the evolution of such circuits, how many circuits exist that have a given number of functions, and whether tradeoffs exist between multifunctionality and the robustness of a circuit to mutation. Because it is not currently possible to answer these questions experimentally, we use a computational model to exhaustively enumerate millions of regulatory circuits and all their possible functions, thereby providing the first comprehensive study of multifunctionality in model regulatory circuits. Our results highlight limits of circuit designability that are relevant to both systems biologists and synthetic biologists.
| Gene regulatory circuits are at the heart of many fundamental biological processes, ranging from developmental patterning in multicellular organisms [1] to chemotaxis in bacteria [2]. Regulatory circuits are usually multifunctional. This means that they can form different metastable gene expression states under different physiological conditions, in different tissues, or in different stages of embryonic development. The segment polarity network of Drosophila melanogaster offers an example, where the same regulatory circuit affects several developmental processes, including embryonic segmentation and the development of the fly's wing [3]. Similarly, in the vertebrate neural tube, a single circuit is responsible for interpreting a morphogen gradient to produce three spatially distinct ventral progenitor domains [4]. Other notable examples include the bistable competence control circuit of Bacillus subtilis [5] and the lysis-lysogeny switch of bacteriophage lambda [6]. Multifunctional regulatory circuits are also relevant to synthetic biology, where artificial oscillators [7], toggle switches [8], and logic gates [9] are engineered to control biological processes.
The functions of gene regulatory circuits are embodied in their gene expression patterns. An important property of natural circuits, and a design goal of synthetic circuits, is that these patterns should be robust to perturbations. Such perturbations include nongenetic perturbations, such as stochastic fluctuations in protein concentrations and environmental change. Much attention has focused on understanding [1], [2], [4], [10], [11] and engineering [12]–[14] circuits that are robust to nongenetic perturbations. Equally important is the robustness of circuit functions to genetic perturbations, such as those caused by point mutation or recombination. Multiple studies have asked what renders biological circuitry robust to such genetic changes [15]–[20]. With few exceptions [21], [22], these studies have focused on circuits with one function, embodied in their gene expression pattern. Such monofunctional circuits tend to have several properties. First, many circuits exist that have the same gene expression pattern [17]–[19], [23]–[28]. Second, these circuits can vary greatly in their robustness [16], [18], [29]. And third, they can often be reached from one another via a series of function-preserving mutational events [18], [19], [30]. Taken together, these observations suggest that the robustness of the many circuits with a given regulatory function can be tuned via incremental mutational change.
Most circuits have multiple functions, but how these observations translate to such multifunctional circuits is largely unknown. In a given space of possible circuits, how many circuits exist that have a given number of k specific functions (expression patterns)? What is the relationship between this number of functions and the robustness of each function? Do circuits with any combination of functions exist, or are some combinations “prohibited?” Pertinent earlier work showed that there are indeed fewer multifunctional circuits than monofunctional circuits [21], but this investigation had two main limitations. First, it considered circuits so large that the space of circuits and their functions could not be exhaustively explored, and restricted itself to mostly bifunctional circuits. Second, it included only topological circuit variants (i.e., who interacts with whom), and ignored variations in the signal-integration logic of cis-regulatory regions. These regions encode regulatory programs, which specify the input-output mapping of regulatory signals (input) to gene expression pattern (output) [31]–[33]. Variations in cis-regulatory regions [34], such as mutations that change the spacing between transcription factor binding sites [35], are known to impact circuit function [36], [37], and their inclusion in a computational model of regulatory circuits is thus important.
Here, we overcome these limitations by focusing on regulatory circuits that are sufficiently small that an entire space of circuits can be exhaustively explored. Specifically, we focus on circuits that comprise only three genes and all possible regulatory interactions between them. Small circuits like this play an important role in some biological processes. Examples include the kaiABC gene cluster in Cyanobacteria, which is responsible for circadian oscillations [38], the gap gene system in Dropsophila, which is responsible for the interpretation of morphogen gradients during embryogenesis [19], and the krox-otx-gatae feedback loop in starfish, which is necessary for endoderm specification [39]. Additionally, theoretical studies of small regulatory circuits have provided several general insights into the features of circuit design and function. Examples include biochemical adaptation in feedback loops [40] and response delays in feed-forward loops [41], among others [16], [19], [23], [42]–[45]. Lastly, there is a substantial body of evidence suggesting that small regulatory circuits form the building blocks of larger regulatory networks [34], [46]–[48], further warranting their study.
For two reasons, we chose Boolean logic circuits [49] as our modeling framework. First, they allow us not only to vary circuit topology [45], but also a circuit's all-important signal-integration logic [44]. Second, Boolean circuits have been successful in explaining properties of biological circuits. For example, they have been used to explain the dynamics of gene expression in the segment polarity genes of Drosophila melanogaster [50], the development of primordial floral organ cells of Arabidopsis thaliana [51], gene expression cascades after gene knockout in Saccharomyces cerevisiae [52], and the temporal and spatial expression dynamics of the genes responsible for endomesoderm specification in the sea urchin embryo [53]. We consider a specific gene expression pattern as the function of a circuit like this, because it is this pattern that ultimately drives embryonic pattern formation and physiological processes. Multifunctional circuits are circuits with multiple gene expression patterns, and here we study the constraints that multifunctionality imposes on the robustness and other properties of regulatory circuits. The questions we ask include the following: (i) How many circuits have a given number k of functions? (ii) What is the relationship between multifunctionality and robustness to genetic perturbation? (iii) Are some multifunctional circuits more robust than others? (iv) Is it possible to change one multifunctional circuit into another through a series of small genetic changes that do not jeopardize circuit function?
We consider circuits of genes (Fig. 1A). We choose a compact representation of a circuit's genotype G that allows us to represent both a circuit's signal-integration logic and its architecture by a single binary vector of length (Fig. 1B). Changes to this vector can be caused by mutations in the cis-regulatory regions of DNA. Such mutations may alter the binding affinity of a transcription factor to its binding site, thereby creating or removing a regulatory interaction [34]. Alternatively, they may affect the distance of a transcription factor binding site from the transcription start site, changing its rotational position on the DNA helix. In turn, this may alter the regulatory effect of the transcription factor [54], and change the downstream gene's signal-integration logic. Lastly, such mutations may change the distance between adjacent transcription factor binding sites, enabling or disabling a functional interaction between proximally bound transcription factors [35]. We note that mutations in G could also be conceptualized as changes in the DNA binding domain of a transcription factor. However, evolutionary evidence from microbes suggest that alterations in the structure and logic of regulatory circuits occurs preferentially via changes in cis-regulatory regions, rather than via changes in the transcription factors that bind these regions [55].
The dynamics of the expression states of a circuit's N genes begin with a prespecified initial state , which represents regulatory influences outside or upstream of the circuit, such as transcription factors that are not part of the circuit but can influence its expression state. The initial state reflects the fact that small circuits are typically embedded in larger regulatory networks [34], [46]–[48], which provide the circuit with different regulatory inputs under different environmental or tissue-specific conditions. Through the regulatory interactions specified in the circuit's genotype, the circuit's gene expression state changes from this initial state, until it may reach a stable (i.e., fixed-point) equilibrium state . We consider a circuit's function to be a mapping from an initial expression state to an equilibrium expression state (Fig. 1C). In the main text, we consider only circuit functions that involve fixed point equilibria, but we consider periodic equilibrium states in the Supporting Online Material. A circuit could in principle have as many as functions , as long as the initial expression states are all different from one another, and the equilibrium expression states are all different from one another (Material and Methods). The circuits we study may map multiple initial states to the same equilibrium state, but our definition of function ignores all but one of these initial states. While a definition of function that includes many-to-one mappings between initial and equilibrium states can be biologically sensible, our intent is to investigate specific pairs of inputs (i.e., ) and outputs (i.e., ), as is typical for circuits in development and physiology [56]–[58]. We emphasize that a circuit can express its k functions individually, or in various combinations, such that the same circuit could be said to have between one and k functions. For brevity, we refer to a specific set of k functions as a multifunction or a k-function and to circuits that have at least one function as viable.
The space of circuits we explore here contains possible genotypes. We exhaustively determine the equilibrium expression states of each genotype for all initial states, thereby providing a complete genotype-to-phenotype(function) map. We use this map to partition the space of genotypes into genotype networks [17]–[19], [21]. A genotype network consists of a single connected set of genotypes (circuits) that have identical functions , and where two circuits are connected neighbors if their corresponding genotypes differ by a single element (Fig. 1D). Note that such single mutations may correspond to larger mutational changes in the cis-regulatory regions of DNA. For example, mutations that change the distance between binding sites, or between a binding site and a transcription start site, may involve the addition or deletion of large segments of DNA [26], [59]–[62].
We first asked how the number of genotypes that have k functions depends on k. Fig. 2 shows that this number decreases exponentially, implying that multifunctionality constrains the number of viable genotypes severely. For instance, increasing k from 1 to 2 decreases the number of viable genotypes by 34%; further increasing k from 2 to 3 leads to an additional 39% decrease. However, there is always at least one genotype with a given number k of functions, for any . In other words, even in these small circuits, multiple genotypes exist that have many functions.
Thus far, we have determined the number of genotypes with a given number k of functions, but we did not distinguish between the actual functions that these genotypes can have. For example, there are 64 variants of function, since there are potential initial states and potential equilibrium states (). Analogously, simple combinatorics (Text S1) shows that there are 1204 variants of functions, and the number of variants increases dramatically with greater k, up to a maximum of variants of functions. This is possible because individual functions can occur in different possible combinations in multifunctional circuits (Material and Methods). The solid line in the inset of Fig. 2 indicates how this number of possible different functions scales with k. We next asked whether there exist circuits (genotypes) for each of these possible combinations of functions, or whether some multifunctions are prohibited. The open circles in the inset of Fig. 2 show the answer: These circles lie exactly on the solid line that indicates the number of possible combinations of functions for each value of k (Text S1). This means that no multifunction is prohibited. In other words, even though multifunctionality constrains the number of viable genotypes, there is always at least one genotype with k functions, and in any possible combination.
As gene regulatory circuits are often involved in crucial biological processes, their functions should be robust to perturbation. We therefore asked whether the constraints imposed by multifunctionality also impact the robustness of circuits and their functions. In studying robustness, we differentiate between the robustness of a genotype (circuit) and the robustness of a k-function. We assess the robustness of a genotype as the proportion of all possible single-mutants that have the same k-function, and the robustness of a k-function as the average robustness of all genotypes with that k-function [17], [18], [51], [63] (Material and Methods). We refer to the collection of genotypes with a given k-function as a genotype set, which may comprise one or more genotype networks. We emphasize that a genotype may be part of several different genotype sets, because genotypes typically have more than one k-function.
Fig. 3A shows that the robustness of a k-function decreases approximately linearly as k increases, indicating a trade-off between multifunctionality and robustness. However, some degree of robustness is maintained so long as . For larger k, some functions exist that have zero robustness (Text S1), that is, none of the circuits with these functions can tolerate a change in their regulatory genotype. The inset of Fig. 3A reveals a similar inverse relationship between the size of a genotype set and the number of functions k, implying that multifunctions become increasingly less “designable” [64] — fewer circuits have them — as k increases (Text S1). For example, for as few as functions, the genotype set may comprise a single genotype, reducing the corresponding robustness of the k-function to zero. For each value of k, the maximum proportion of genotypes with a given k-function is equal to the square of the maximum proportion of genotypes with a function, explaining the triangular shape of the data in the inset. This triangular shape indicates that the genotype set of a given k-function is always smaller than the union of the k constituent genotypes sets. Additionally, we find that the robustness of a k-function and the size of its genotype set are strongly correlated (Fig. S1), indicating that the genotypes of larger genotype sets are, on average, more robust than those of smaller genotype sets. This result is not trivial because the structure of a genotype set may change with its size. For example, large genotype sets may comprise many isolated genotypes, or their genotype networks might be structured as long linear chains. In either case, the robustness of a k-function would decrease as the size of its genotype set increased.
We have so far focused on the properties of the genotype sets of k-functions, but have not considered the properties of the genotype networks that make up these sets. Therefore, we next asked how genotypic robustness varies across the genotype networks of k-functions. In Figs. 3B–D, we show the distributions of genotypic robustness for representative genotype networks with functions. These distributions highlight the inherent variability in genotypic robustness that is present in the genotype networks of multifunctions, indicating that genotypic robustness is an evolvable property of multifunctional circuits. Indeed, in Fig. S2, we show the results of random walks on these genotype networks, which confirm that it is almost always possible to increase genotypic robustness through a series of mutational steps that preserve the k-function. In Fig. S3, we show in which dynamic regimes (Material and Methods) the circuits in these same genotype networks lie.
We have shown that the genotype set of any k-function is non-empty (Fig. 2), meaning that there are no “prohibited” k-functions. We now ask how the genotypes with a given k-function are organized in genotype space. More specifically, is it possible to connect any two circuits with the same k-function through a sequence of small genotypic changes where each change in the sequence preserves this k-function? In other words, are all genotypes with a given k-function part of the same genotype network, or do such genotypes occur on multiple disconnected genotype networks?
Fig. 4 shows the relationship between the number of genotype networks in a genotype set and the number of circuit functions k. For monofunctional circuits (), the genotype set always consists of a single, connected genotype network. This implies that any genotype in the genotype set can be reached from any other via a series of function-preserving mutational events. In contrast, for circuits with functions, the genotype set often fragments into several isolated genotype networks, indicating that some regions of the genotype set cannot be reached from some others without jeopardizing circuit function. The most extreme fragmentation occurs for functions, where some genotype sets break up into more than 20 isolated genotype networks. Fig. S4 provides a schematic illustration of how fragmentation can occur in a k-function's genotype set, despite the fact that the genotype sets of the k constituent monofunctions consist of genotype networks that are themselves connected. Fig. S5 provides a concrete example of fragmentation, depicting one genotype from each of the several genotype networks of a bifunction's genotype set.
The proportion of k-functions with genotype sets that comprise a single genotype network is shown in the inset of Fig. 4. This proportion decreases dramatically as the number of functions increases from to , such that only 16% of genotype sets comprise a single genotype network when . Figs. 4B–D show that the distributions of the number of genotype networks per genotype set are typically left-skewed. This implies that when fragmentation occurs, the genotype set usually fragments into only a few genotype networks. However, the distribution of genotype network sizes across all genotype sets is heavy-tailed and often spans several orders of magnitude (Fig. S6). This means that the number of genotypes per genotype network is highly variable.
We next ask whether the number of genotypes in the genotype set of a k-function can be predicted from the number of genotypes in the genotype sets of the k constituent monofunctions. To address this question, we define the fractional size of a genotype set as the number of genotypes in the set, divided by the number of genotypes in genotype space. We first observe that the maximum fractional size of a genotype set of a k-function is equal to (Fig. S6), which is the maximum fractional size of a genotype set for monofunctional circuits [44] raised to the kth power. In general, we find that the fractional size of a genotype set of a k-function can be approximated with reasonable accuracy by the product of the fractional sizes of the genotype sets of the k constituent monofunctions, but that the accuracy of this approximation decreases as k increases (Fig. S7). While these fractional genotype set sizes may be quite small, we note that their absolute sizes are still fairly large, even in the tiny circuits considered here. For example, for functions the maximum genotype set size is 262,144. For functions, the maximum is 32,768.
In evolution, a circuit may acquire a new regulatory function while preserving its pre-existing functions. An example is the highly-conserved hedgehog regulatory circuit, which patterns the insect wing blade. In butterflies, this regulatory circuit has acquired a new function. It helps form the wing's eyespots, an antipredatory adaptation that arose after the insect body plan [65]. This example illustrates that a regulatory circuit may acquire additional functions incrementally via gradual genetic change. The order in which the mutations leading to a new function arise and go to fixation can have a profound impact upon the evolution of such phenotypes [66]. In particular, early mutations have the potential to influence the phenotypic effects of later mutations, which can lead to a phenomenon known as historical contingency.
We next ask whether it is possible for a circuit to incrementally evolve regulatory functions in any order, or whether this evolutionary process is susceptible to historical contingency. In other words, is it possible that some sequence of genetic changes that lead a circuit to have k functions also preclude it from gaining an additional function? The genotype space framework allows us to address this question in a systematic way, because it permits us to see contingency as a result of genotype set fragmentation. Specifically, contingency means that, as a result of fragmentation, the genotype network of a new function may become inaccessible from at least one of the genotype networks of a k-function's genotype set. To ask whether this occurs in our model regulatory circuits, we considered all permutations of every k-function. These permutations reflect every possible order in which a circuit may acquire a specific combination of k functions through a sequence of genetic changes. To determine the frequency with which historical contingency occurs, we calculate the number of genotype networks per genotype set, as the k functions are incrementally added. This procedure is outlined in Fig. S4 and detailed in the Material and Methods section. We note that historical contingency is not possible when because all monofunctions comprise genotype sets with a single connected genotype network. Historical contingency is also not possible when , because there is only one genotype that yields this combination (Fig. 2).
In Fig. 5, we show the relationship between the proportion of k-functions that exhibit historical contingency and the number of functions k. For as few as functions, 43% of all k-functions exhibit historical contingency. This percentage is highest for , where 94% of combinations are contingent. The inset of Fig. 5 shows the proportion of the permutations of a k-function in which genotype set fragmentation may preclude the evolution of the k-function. Again, this proportion is highest for functions. These results highlight an additional constraint of multifunctionality. Not only does the number of genotypes with k functions decrease as k increases, but the dependence upon the temporal order in which these functions evolve tends to increase.
In the Supporting Online Material, we repeat the above calculations to show how our results scale to equilibrium expression states with period (For the sake of computational tractability, we restrict our attention to the case where all equilibrium expression states have the same period P). We show that the exponential decrease in the number of circuits with k functions also holds for periodic equilibrium expression states, but that the maximum number of functions per circuit decreases with increasing (Fig. S8). So long as , it is possible for a circuit to have more than one function. In this case, the inverse relationship between robustness to genetic perturbation and the number of functions k also holds (Fig. S9). Similarly, the results pertaining to genotype set fragmentation hold so long as (Fig. S10). Lastly, the results pertaining to historical contingency only hold when . This is because it is not possible for a circuit with an equilibrium expression pattern of period to have more than functions, which is a prerequisite for historical contingency (Material and Methods). Taken together, these additional observations show that the results obtained for fixed-point equilibrium expression states can also apply to periodic equilibrium expression states, so long as is not too large.
We have used a Boolean model of gene regulatory circuits to exhaustively characterize the functions of all possible combinations of circuit topologies and signal-integration functions in three-gene circuits. The most basic question we have addressed is whether multifunctionality is easy or difficult to attain in regulatory circuits. Our results show that while the number of circuits with k functions decreases sharply as k increases, there are generally thousands of circuits with k functions, so long as k is not exceedingly large. Thus, multifunctionality is relatively easy to attain, even in the tiny circuits examined here.
It is worth considering how this result might translate to larger circuits. In a related model of gene regulatory circuits with genes, the genotype sets of bifunctions comprised an average of circuits [21], which is over an order of magnitude more circuits per bifunction than observed here (Fig. 3, inset). For a greater number of functions k, we expect the number of circuits per k-function to increase as the number of genes N in the regulatory circuit increases. This is because the maximum number of circuits with a given k-function is , which is the total number of circuits with N genes () multiplied by the maximum proportion of circuits per multifunction (). For a given number of functions k, this quotient will increase hyper-exponentially as N increases, indicating a dramatic increase in the maximum number of circuits per k-function. More generally, because the fractional size of a k-function's genotype set can be approximated as the product of the fractional sizes of the genotype sets of its k constituent monofunctions (Fig. S7) and because the total number of circuits increases exponentially with N, our observation that there are many circuits with k functions is expected to scale to larger circuits.
The next question we asked is whether there is a tradeoff between the robustness of a k-function and the number of functions k. We found that the robustness of a k-function decreases as k increases. However, some degree of robustness is generally maintained, so long as k is not too large. These observations suggest that the number of circuit functions generally does not impose severe constraints on the evolution of circuit genotypes, unless the number of functions is very large. Our current knowledge of biological circuits is too limited to allow us to count the number of functions per circuit. However, we can ask whether the functional “burden” on biological circuits is very high. If so, we would expect that the genes that form these circuits and their regulatory regions cannot tolerate genetic perturbations, and that they have thus accumulated few or no genetic changes in their evolutionary history. However, this is not the case. The biochemical activities and regulatory regions of circuit genes can diverge extensively without affecting circuit function [55], [59], [61], [67], and the very different circuit architectures of distantly related species can have identical function [24], [28]. Further, circuits are highly robust to the experimental perturbation of their architecture, such as the rewiring of regulatory interactions [20]. More indirect evidence comes from the study of genes with multiple functions, identified through gene ontology annotations. The rate of evolution of these genes is significantly but only weakly correlated with the number of known functions [68]. Thus, the functional burden on biological genes and circuits is not sufficiently high to preclude evolutionary change.
Previous studies of monofunctional regulatory circuits have revealed broad distributions of circuit robustness to genetic perturbation [16], [18], [29]. We therefore asked if this is also the case for multifunctional circuits. We found that circuit robustness was indeed variable, but that the mean and variance of the distributions of circuit robustness decreased as the number of functions k increased. Thus, variation in circuit robustness persists in multifunctional circuits, so long as k is not too large. This provides further evidence that robustness to mutational change may be considered the rule, rather than the exception, in biological networks [1], [18], [20], [29]. However, to make the claim that robustness to genetic perturbation is an evolvable property in multifunctional regulatory circuits requires not only variability in circuit robustness, but also the ability to change one circuit into another via a series of mutations that do not affect any of the circuit's functions.
We therefore asked whether it is possible to interconvert any two circuits with the same function via a series of function-preserving mutational changes. We showed that this is always possible for monofunctions, but not necessarily for multifunctions, because these often comprise fragmented genotype sets. Genotype set fragmentation has also been observed at lower levels of biological organization, such as the mapping from RNA sequence to secondary structure [69]. Such fragmentation has two evolutionary implications, as has recently been discussed for RNA phenotypes [70]. First, the mutational robustness of a phenotype (function) depends upon which genotype network its sequences inhabit, as we have also shown for regulatory circuits (Fig. S11). Second, it can lead to historical contingency, where the phenotypic effects of future mutations depend upon the current genetic background. Such contingency indeed occurs in our circuits, because the specific genotype network that a circuit (genotype) occupies may be influenced by the temporal order in which a circuit's functions (phenotypes) have evolved. This order in turn may affect a circuit's ability to evolve new functions.
These observations hinge on the assumption that the space between two (disconnected) parts of a fragmented genotype set is not easily traversed. For example, in RNA it is well known that pairs of so-called compensatory mutations can allow transitions between genotype networks [71], thus alleviating the historical contingency caused by fragmentation. To assess whether an analogous phenomenon might exist for regulatory circuits, we calculated the average distance between all pairs of genotypes on distinct genotype networks for circuits with the same k-function. We found that this distance decreases as the number of functions k increases, indicating an increased proximity between genotype networks (Fig. S12). However, those pairs of genotypes in any two different genotype networks that had the minimal distance of two mutations never exceeded 1% of all pairs of genotypes on these networks, and was as low as 0.03% for functions (Fig. S12A, inset). This means that transitions between genotype networks through few mutations are not usually possible in these model regulatory circuits. Thus, the multiple genotype networks of a genotype set can indeed be considered separate from one another.
Using a Boolean model of gene regulatory circuits comes with several caveats that are worth highlighting. First, the mutational distance between certain logical functions may not correspond to their distance in a biological context. For example, the signal-integration logic of a gene can mutate from an OR function to an XOR function by changing only a single bit. In contrast, research in synthetic biology suggests that these logical functions are separated by greater mutational distances. While the OR function can be encoded as a simple two-input circuit [37], the XOR function has necessitated cascading signals between distinct circuits [37] or cells [72], [73], or chemically-induced DNA inversions [74]. In some biological circuits, such as the lac operon in E. coli, it may not be possible to transform an OR function into an XOR function at all [32]. However, experimental investigations of the cis-regulatory codes of synthetic and natural circuits are far from exhaustive, and it is therefore possible that there exist alternative implementations of these logical functions that more closely resemble their Boolean representations [31]. Second, the model makes the simplifying assumptions that gene expression states are binary and that regulatory interactions are static. In biological circuits, gene expression is continuous and regulatory interactions are dynamic, varying in both time and space. Despite these limitations, the assumption of binary expression often provides a reasonable approximation [32] and numerous studies have demonstrated the model's ability to precisely replicate the expression dynamics of biological circuits, even under the assumption of static regulatory interactions [50]–[53]. Third, we assume that gene states are updated synchronously [49], which is clearly not the case in biological circuitry. Asynchronous updating can affect the transient dynamics of a circuit [75] and its equilibrium expression patterns [76], and may therefore impact circuit function. This becomes especially problematic when the equilibrium expression pattern is periodic [77]. However, the fixed-point equilibrium expression states of Boolean circuits do not vary between asynchronous and synchronous updating schemes [78], so we did not consider asynchronous updating. While it is possible that some of our results depend upon this assumption, we stress that this study could not have been performed without it. The exhaustive enumeration of genotype space is not computationally feasible under asynchronous updating because all possible orderings of updates have to be considered for each genotype. Fourth, we did not explicitly consider gene expression noise. While this is an important aspect of genetic regulation [79], robustness to gene expression noise is correlated with robustness to genetic perturbation in model regulatory circuits [18]. Thus, we used the latter as a proxy for the former. Lastly, we only considered small, three-gene circuits. This allows for the exhaustive enumeration of all possible circuit topologies and signal-integration functions, but limits the direct applicability of our results to similarly sized circuits. However, we expect our results to also apply to larger circuits, as we have discussed. We emphasize that our observations are not derived from one circuit and its functions, but from an enormous circuit space, comprising a class of circuits that capture biological phenomena in diverse organisms.
We consider fully connected Boolean circuits with genes. The binary state of a gene i at time t is a function of the states of all genes at time :(1)The function maps all of the possible combinations of input expression states to an output expression state. This function represents the gene's signal-integration logic and can be represented as a look-up table (Fig. 1A). The circuit is initialized with an initial expression state and all genes are updated synchronously according to their individual functions f until a steady-state expression pattern is reached. The expression pattern can be a fixed-point () or a cycle ().
The update functions f of all N genes can be represented as a single vector of length (Fig. 1B). We measure the equilibrium expression states for all possible vectors for each of the possible initial expression states . In doing so, we not only enumerate all signal-integration functions, but also all circuit topologies. This is because some functions f make a gene independent of one or more of its N regulatory inputs. For example, in Fig. 1A, the regulatory interaction is inactive because for any combination of regulatory inputs, the expression state of gene a is unaffected by the expression state of gene b.
Boolean circuits exhibit three dynamic regimes that have been called ordered, critical, and chaotic [49]. The ordered regime is characterized by a general insensitivity to perturbation that results from having few equilibrium states, each with large basins of attraction, whereas the chaotic regime is characterized by extreme sensitivity to perturbation that results from having many equilibrium states with small basins of attraction. The critical regime lies at the interface of these two extremes. Several studies have focused on characterizing the dynamic regimes of biological circuits [80]–[83] and on understanding how these regimes influence circuit dynamics in silico [49], [84].
The dynamic regime of a circuit can be determined by calculating its sensitivity , where z is the average number of regulators per gene and is the average probability of gene expression per gene (i.e., the proportion of the genotype G that is nonzero) [85], [86]. The ordered regime corresponds to , the critical regime to , and the chaotic regime to . Since for all circuits considered here, the dynamic regime is determined solely by .
The maximum number of functions a circuit can produce is because we require the equilibrium expression states of any multifunction to be unique (i.e., ). We also require that the initial expression states are unique (i.e., ). While the deterministic nature of the model makes this latter requirement superfluous — different equilibrium states require different initial states — we specify it to highlight the fact that each function pertains to a specific input signal, which may differ between environments or tissue-specific conditions.
A circuit may produce various combinations of k functions, as shown in Fig. 1. We note that some combinations of functions are not feasible. As an example, consider a hypothetical combination where , . This combination is not feasible because the equilibrium expression state of is a transient state of .
Our usage of the word function differs from existing terminology for describing the mapping of initial to equilibrium states in Boolean circuits. For a given circuit, an attractor is an equilibrium state (fixed-point or periodic) that can be reached from at least one initial state. An attractor's basin of attraction is the set of initial states that lead to that attractor. The attractor landscape is the set of all attractors and their basins of attraction. These terms are distinct from our use of the words function and k-function, which are concerned with specific pairs of initial and equilibrium states, because specific initial states provide key inputs to most biological circuits in development and physiology. The only equivalence between terms occurs when . Such a k-function is equivalent to the circuit's attractor landscape, because each of the initial states map onto themselves. In this case, the entire attractor landscape is embodied in the function.
We measure the robustness of circuits and of k-functions. The robustness of a circuit is calculated as the proportion of its mutational neighbors that have the same k-function, as follows. First, we remove the entries in the circuit's genotype G that correspond to inactive regulatory interactions. This results in a new vector that may differ from G. Second, we determine the fraction of single mutants of that produce the same multifunction. This is achieved by flipping each bit in , one at a time, and determining whether the resulting genotype has the same k-function. We refer to this measure of circuit robustness as , which is the measure that is used throughout the main body of the text. The robustness of a k-function is calculated as the average robustness of all circuits with that k-function.
Alternatively, the robustness of a circuit can be calculated as the connectivity of its genotype G in a genotype network of a k-function, divided by the maximum possible connectivity L. We refer to this measure of circuit robustness as . In Fig. S13, we show that these two calculations result in measures of k-function robustness that are highly correlated (Spearmans ). The fact that the data are always below the identity line indicates that is a more conservative measure of robustness than .
To detect whether a combination of k functions may exhibit historical contingency, we consider all permutations of those functions. We define a combination of k functions to be contingent if there exists at least one permutation that violates, and at least one other permutation that satisfies, the following condition: For the functions in the permutation, there exists a such that the number of genotype networks in the genotype set of function is greater than the number of genotype networks in the genotype set of function . For example, in Fig. S4, the permutation satisfies this condition because the genotype set of comprises two genotype networks while the genotype set of comprises only one genotype network. All other permutations violate this condition. Therefore this combination of k-functions exhibits historical contingency. Since all monofunctions comprise a single, connected genotype network, it is impossible for any bifunction to satisfy the condition above. Thus, in these model regulatory circuits, historical contingency can only occur for .
|
10.1371/journal.ppat.1002087 | HIV/SIV Infection Primes Monocytes and Dendritic Cells for Apoptosis | Subversion or exacerbation of antigen-presenting cells (APC) death modulates host/pathogen equilibrium. We demonstrated during in vitro differentiation of monocyte-derived macrophages and monocyte-derived dendritic cells (DCs) that HIV sensitizes the cells to undergo apoptosis in response to TRAIL and FasL, respectively. In addition, we found that HIV-1 increased the levels of pro-apoptotic Bax and Bak molecules and decreased the levels of anti-apoptotic Mcl-1 and FLIP proteins. To assess the relevance of these observations in the context of an experimental model of HIV infection, we investigated the death of APC during pathogenic SIV-infection in rhesus macaques (RMs). We demonstrated increased apoptosis, during the acute phase, of both peripheral blood DCs and monocytes (CD14+) from SIV+RMs, associated with a dysregulation in the balance of pro- and anti-apoptotic molecules. Caspase-inhibitor and death receptors antagonists prevented apoptosis of APCs from SIV+RMs. Furthermore, increased levels of FasL in the sera of pathogenic SIV+RMs were detected, compared to non-pathogenic SIV infection of African green monkey. We suggest that inappropriate apoptosis of antigen-presenting cells may contribute to dysregulation of cellular immunity early in the process of HIV/SIV infection.
| Antigen-presenting cells (APCs) are critical for both innate and adaptive immunity. They have a profound impact on the hosts' ability to combat microbes. Dysfunction and premature death by apoptosis of APCs may contribute to an abnormal immune response unable to clear pathogens. Circulating blood monocytes exhibit developmental plasticity, with the capability of differentiating into either macrophages or dendritic cells (DCs), and they represent important cellular targets for HIV-1. We report that HIV infection renders monocytes/macrophages and DCs in vitro more prone to undergo apoptosis and this heightened susceptibility is associated with changes in the expression of anti- and pro-apoptotic molecules. Our results show that during the acute phase of SIV-infection of rhesus macaques, monocytes and DCs are more prone to die by apoptosis. They express lower levels of Mcl-1 and FLIP proteins, two anti-apoptotic molecules, but higher expression of the active form of Bax and Bak, the gatekeepers of the mitochondria, major sensor of the apoptotic machinery. Because the early events are important in the pathogenesis of this disease, early death of APCs should play a major role leading to the defective immune response. Strategies aimed at preventing death of APCs could be beneficial in helping the immune response to fight HIV-1.
| Monocytes originating from the bone marrow are released into peripheral blood, where they circulate for several days before entering tissues, and replenish tissue macrophage populations in the steady state. Monocytes constitute a considerable systemic reservoir of myeloid precursors. Monocytes exhibit developmental plasticity, with the capability of differentiating into either macrophages or dendritic cells (DCs) in vitro depending on the cytokine milieu. They can enter in lymphoid tissues during inflammation and give rise to macrophages and inflammatory DCs [1], [2], [3]. Classical DCs represent a distinct lineage of myeloid cells that are also present in the blood and can migrate into the tissues [3]. Mononuclear phagocytes are critical for both innate and adaptive immunity. Recruited to inflammatory sites, cDCs, inflammatory DCs and macrophages play a critical role in the protection against pathogens [3], [4], [5], [6].
Mononuclear phagocytes and DCs which express CD4 receptor and chemokine co-receptors represent important cellular targets for human immunodeficiency virus type-1 (HIV-1). Circulating monocytes can be latently infected and productive infection can be initiated during differentiation into macrophages [7], [8]. Mononuclear phagocytes are rendered defective specifically by the envelope glycoprotein that impairs maturation and cytokine secretion [9], [10]. This contributes to the development of immune deficiency observed during HIV infection [11], [12], [13], [14].
The most striking feature of AIDS is the increased death and progressive depletion of CD4+ T lymphocytes which leads to immunodeficiency [15]. CD4+ T cells from HIV-infected individuals and SIV-infected rhesus macaques are more sensitive to undergo apoptosis due to the effects of death-receptors [16], [17], [18], [19], [20], [21], [22], [23], [24], [25]. Moreover, in the absence of viral replication, HIV or SIV primes CD4+ T cells for apoptosis in vitro [25], [26], [27]. In contrast, the impact of HIV on apoptosis of monocytes and DCs has not been extensively studied.
Monocytes, but not macrophages, are prone to undergo apoptosis after death-receptor ligation [16], [28], [29], [30], [31]. Death receptors include Fas/CD95, TRAIL-Receptor, and TNF-Receptor. The engagement of death-receptors by their counterparts, FasL, TRAIL and TNF, either in soluble form or at the membrane surface of the cells, induce death-signaling cascades. The molecular ba.sis of resistance to death-receptors-mediated apoptosis involves FLIP (cellular-FLICE-inhibitory protein expressed during differentiation of APCs [31], [32], [33]), an inhibitor of the DISC (death-inducing signaling complex) [34]. Moreover, apoptosis initiated by growth factor deprivation can be prevented by a decoy-receptor that blocked Fas and FasL interaction [31], [35], [36], and mice carrying functional mutations of Fas-FasL displayed elevated monocytic cell counts [37]. In addition, to the extrinsic pathway that involves death-receptors and their counterparts, apoptosis regulation in mononuclear phagocytes includes also the intrinsic pathway. Thus, among the anti-apoptotic members, Mcl-1 predominates in differentiated cells [38]. Mitochondrial outer membrane integrity is highly controlled, primarily through interactions between pro- and anti-apoptotic of the members of the Bcl-2 protein family. On activation, Bax and Bak proteins undergo extensive conformational changes leading to mitochondria permeabilization and cell death [39].
Subversion of monocyte apoptosis by intracellular bacteria or parasites is used by pathogens to favor their own replication and dissemination within the host when death is inhibited [40], [41], [42], [43], [44], [45], [46]. In contrast, massive cell death of infected macrophages induced by the Ebola virus contributes to pathogenesis by abolishing innate immunity [47]. Several viral infections are also associated with the death of DCs [45], [48], although DCs, unlike monocytes, are mostly resistant to FasL-induced cell death [33], [49], [50], [51], [52].
Differentiated macrophages infected by HIV in vitro are more resistant to TRAIL-mediated cell death triggered by the envelope protein [53] whereas another report suggests that HIV-infected macrophages are more prone to undergo apoptosis [54]. In the peripheral blood of chronically HIV-infected individuals and SIV-infected rhesus macaques (RMs), reduced numbers of DCs are found [55], [56], [57], [58], [59], [60], [61] consistent with increased death of those cells [62], [63], [64]. Furthermore, in chronically SIV-infected RMs, massive turnover of peripheral monocytes undergoing apoptosis have been reported [65]. In viremic HIV-infected individuals it has been shown that both spontaneous and IFN-ã-induced monocyte cell death are elevated compared to controls [66] although another report describes monocytes resistant to cell death, associated with antiapoptotic gene profiles [67]. However, little information exists on the precise molecular mechanisms involved and only few studies have assessed these processes early after infection.
Indeed, an increasing amount of evidence suggests that the acute phase dictates the rate of progression towards AIDS. Experimental infection of RMs of Chinese origin is an extremely valuable model to investigate these early events [22], [68], [69], [70], [71]. The aims of the present study were to determine whether HIV/SIV infection early after viral exposure sensitizes mononuclear phagocytes for apoptosis and to elucidate the molecular mechanisms behind the process. We assessed the relevance of apoptosis inducing processes during the acute phase of pathogenic lentiviral infection of RMs.
We demonstrated that in vitro and in vivo, monocytes and DCs exposed to HIV/SIV are sensitized to death-receptors ligation-mediated cell death. Among death-ligands, TRAIL and FasL were the most potent at promoting apoptosis of monocytes and DCs, respectively. Lower amounts of FLIP and Mcl-1 and an increase in the levels of the active form of Bax and Bak proteins were found. A broad caspase inhibitor prevented cell death and increased the number of TNF-α productive mononuclear cells. Thus, the inappropriate death of circulating mononuclear phagocytes during the acute phase could favor the development of a state of immunodeficiency.
Blood monocytes are non-cycling, non-proliferating cells incapable of supporting viral replication. Indeed, establishment of productive infection coincides with entry into G1/S phase of the cell cycle [72], and GM-CSF is one of the main cytokines that promotes and sustains productive infection [7], [8], [73], [74], [75]. We infected monocyte-derived macrophages (MØ)- and monocyte-derived DCs (immature DCs) during differentiation. One day after the process of differentiation was initiated, with either GM-CSF and IL-6 for MØ or GM-CSF and IL-4 for DCs, the R5 HIV-1 tropic strain, HIV-1BaL was added to simulate the presence of HIV-1 during the maturation process. This contrasts with the addition of virus at the end of the differentiation process utilized in most, if not all, published studies [53], [76], [77]. After 5 days, we assessed the percentage of infected MØ and DCs based on intracellular p24 staining by flow cytometry. As expected, we found that the percentage of MØ infected by the R5 tropic strain HIV-1BaL was higher than DCs (Figure 1A). The percentage of HIV-infected MØ varied (40%±7) among individual preparations, whereas DCs from the same individuals displayed less than 3%±1 of infected cells, consistent with previous reports [13], [78]. To confirm intracellular staining of p24 antigen, the cells were lysed and western blots performed to detect the profile of viral antigens, using sera from HIV-infected individuals. In MØ, we observed a typical profile displaying both envelope glycoprotein and gag protein, whereas none of these bands were clearly observed in DCs (Figure 1B). We then assessed the capacity of MØ and DCs to produce cytokines and express co-stimulatory molecules in response to LPS and IFN-γ. We found that activated-MØ as well as activated-DCs, incubated in the presence of HIV-1, secreted less pro-inflammatory cytokines such as IL-6, IL-8 and TNF-α as compared to uninfected cells. No difference was observed for IL-1β secretion (Figure 1C). Moreover, stimulation with LPS and IFN-γ induced lower expression of the co-stimulatory molecule CD86 (Figure 1D, E) and the maturation marker CD83 (data not shown), at the surface of HIV-infected cells as compared to uninfected cells (CD86 mean expression, MØ: 350±150 vs 1090±230; DCs: 400±220 vs 1570±230). Thus, HIV infection during the process of APC differentiation impacted cytokine secretion and cellular maturation.
We then examined whether MØ and DCs were more prone to die at day 5 post-infection with HIV-1BaL. After stimulation with LPS and IFNγ, we observed a significant increase in the percentage of apoptotic cells from HIV-infected culture as compared to non-infected cells (MØ, 55%±7 vs 31%±3.8, p<0.01; DCs, 36%±7 vs 16%±3.9, p<0.01). No major difference was observed after stimulation of uninfected cells (Figure 2A, B). The extrinsic apoptotic pathway involves members of the death-receptor family including CD95 (Fas), TNF-R and DR4/DR5 (Trail-R1/R2) [79]. Upon ligation of these death-receptors by their ligands, the association of the adaptor molecule FADD with the initiator caspases forms a death-inducing signaling complex (DISC) leading to apoptosis [80]. We assessed whether MØ and DCs in the presence of HIV-1BaL are sensitive to death-receptor ligands including TNF-α, TRAIL and FasL. Actinomycin D (Act D) was used as a positive control for cell death. First, both uninfected and HIV-infected MØ (51±6% and 62±7%, respectively) were more sensitive to undergo apoptosis in response to TNF-α in comparison with the medium alone (30±4% and 31±5%, respectively) whereas no similar effect was observed on DCs (Figure 3A, B). Second, MØ infected with HIV-1BaL were more prone to die in response to TRAIL as compared to non-infected cells (60±5% versus 36±6%) (Figure 3A) but no difference was observed for FasL (38±4% versus 41±6%). Finally, HIV-1 infection increased the sensitivity of DCs to die spontaneously (20±4% uninfected versus 31±6% in infected DCs) and after FasL-ligation (64±7% versus 31±6% in medium alone) but not to the binding of TRAIL (32±6%) (Figure 3B). Apoptosis was dependent on the amount of death ligands (Figure 3C).
In order to determine if viral replication was necessary for sensitization to apoptosis, we treated the cells with ddI (5 µM, a dose that blocks viral replication in MØ; <5% of p24+). Our studies showed that in the presence or absence of ddI, both MØ and DCs remain sensitive to TRAIL and FasL, respectively (Figure 3D). Furthermore, we assessed whether stimulation with LPS/IFN-γ-mediated apoptosis may be modulated by antagonists to death ligands using decoy receptors. We demonstrated that decoy receptors of TNF (TNF-R1), and TRAIL (TRAIL-R2/TRAIL-R1), reduced monocyte cell death-mediated by LPS/IFN-γ stimulation, whereas decoys receptors of Fas (Fas-Fc) and TNF (TNF-R1) reduced DCs cell death (Figure 3E). Thus, despite the fact that soluble TNF-α has no effect on DCs, TNF-R1 partially inhibits cell death. Altogether, these results indicated that HIV induces APC apoptosis after death-receptors ligation.
Since cells were more sensitive to undergo apoptosis, we next assessed whether this effect was related either to a modulation in the expression of death-receptors or in the regulation of the signaling pathway. Although, MØ and DCs exhibited a greater sensitivity to die in the presence of death ligands, we did not observe any modulation of death-receptor expression, including TRAIL-R1 and –R2 and Fas/CD95 on cells infected with HIV-1Ba-L compared to uninfected cells (data not shown). The molecular basis of resistance to death-receptors-mediated apoptosis involves the expression of FLIP (cellular-FLICE-inhibitory protein), which is an inhibitor of the DISC (death-inducing signaling complex) [34], and is expressed during differentiation of APCs [31], [32], [33]. Therefore, we analyzed the expression of FLIP in HIV-infected MØ and DCs. We observed that FLIP expression is detectable by western blot at day 5 in both uninfected MØ and DCs but decreased in HIV-infected cells (Figure 4A). Thus, the amount of FLIP decreased by 57%±4 in MØ and 46%±5 in DCs following HIV infection (Figure 4B). Altogether, our data suggest that HIV-1 infection increased the propensity of mononuclear phagocytes to undergo apoptosis in response to death-ligands, possibly due to a decrease in the amount of FLIP. This process occurred independently of any modulation of death receptor expression.
The molecular basis of macrophage resistance to apoptosis includes the expression of the anti-apoptotic Bcl-2 family members, among which Mcl-1 predominates in differentiated cells [38]. In the absence of growth receptor engagement, Mcl-1 is degraded by the ubiquitin-proteasome pathway [81], [82], [83] or cleaved by proteases [84], [85]. We found a 50% decrease in expression of Mcl-1 protein in infected-MØ, which was not observed in DCs (Figure 4A). In addition, SDS-PAGE analysis revealed that Mcl-1 migrated as a doublet suggesting the presence of phosphorylated Mcl-1, primed by GSK-3, on threonine 163. This phosphorylated form undergoes accelerated degradation [81], [83]. In HIV-infected MØ, this change in MCL isoforms was clearly observed compared to uninfected cells, whereas no difference was observed for DCs (Figure 4A). Additional bands of approximately 34 KDa on western blots probed with Mcl-1 antibody were also detected (Figure 4A). These product bands correspond to different translational products (Mcl-1S/ΔTM versus Mcl-1Exon-1). It is important to note that Mcl-1Exon-1 is pro-apoptotic [38]. The amount of Mcl-1Exon-1 protein was clearly enhanced in DCs cultured in the presence of HIV-1BaL (fold increase 2.1) as well as in MØ (fold increase 1.7) (Figure 4B).
Members of the Bcl-2 protein family, in particular Bax and Bak proteins play a critical role in controlling apoptosis [39]. To assess the early commitment of Bax and Bak activation, we subfractionated the cells to isolate a mitochondria-enriched fraction. At day 5 of culture, we observed higher amounts of Bax and Bak proteins within the enriched mitochondrial fraction derived from HIV-1BaL-infected MØ and DCs compared to uninfected cells (Figure 4C and D). Membrane insertion of Bax and Bak supported a dynamic model in which mitochondria is a central sensor. Taken together, these results suggest that HIV shifts the balance towards pro-apoptotic molecules rendering APCs more sensitive to death stimuli.
Early events during the acute phase of SIV infection are critical in determining the onset of AIDS, we therefore investigated APC death in RM during acute SIV infection. We analyzed the percentage of monocytes CD14+ that were infected compared to CD4+ T cells in peripheral blood (the purity was more than 98% for HLA-DR+CD14+ cells as well for CD4+ T cells after cell sorting). HIV-1 has been reported to be isolated from CD14+ monocytes of patients under HAART, indicating that monocytes are competent for HIV infection [86]. Moreover, because monocytes circulate in the blood for only a few days before differentiating into macrophages in tissues, they represent important cells in viral dissemination. The frequency of monocytes and CD4+ T cells harboring proviral DNA was quantified using a nested SIV PCR assay in limiting dilutions of purified cells [22]. The frequency of SIV-DNA positive monocytes increased and peaked at days 11–14 (day 11, mean: 3.8±1.3; day 14, 3.6±1.2 of monocytes were infected) which is equivalent to the frequency found in CD4+ T cells (day 11, mean: 4.8±2.1 and day 14, 1.4±0.4) (Figure 5). Thereafter, the frequencies of SIV-infected monocytes decreased (mean: 0.5±0.18). The dynamics of SIV-DNA is consistent with viral load (viral RNA) measured in the plasma (data not shown) [22]. Unlike monocytes and CD4+ T cells, the frequency of SIV-DNA in myeloid DCs (HLA-DR+CD11+) was extremely low during the acute phase of infection (day 11, mean: 0.015±0.004; day 14, 0.015±0.008 of DCs were infected) consistent with a previous report [87].
We quantified the percentages of dying HLA-DR+CD3−CD20− and CD4+ T cells before and after incubation with death-ligands by monitoring FITC-labeled annexin V. As previously shown [21], [25], [88], CD4+ T cells derived from SIV-infected RMs at the peak of viral replication (day 14) were prone to undergo apoptosis spontaneously and after FasL ligation as compared to Trail or TNF-α ligation (Figure 6A and D) and consistent with other studies [89]. Unlike CD4+ T cells, among death-ligands, TRAIL and FasL were the most potent ligands to promote apoptosis of HLA-DR+CD3−CD20− at day 14 (Figure 6B, C and D). We then demonstrated that early after infection both monocytes (HLA-DR+CD14+) and DCs (Lin−HLA-DR+CD11c+CD123−) are more prone to undergo apoptosis spontaneously (Figure 6E). Thereafter, the levels of apoptosis decreased to reach those observed before infection (Figure 6E). In non pathogenic SIV-infected African green monkeys (AGM), apoptosis of CD4 and of HLA-DR+CD3−CD20− cells at day 14 was similar to the level observed from healthy monkeys (Figure 6F) consistent with the absence of apoptosis reported in this non pathogenic primate model, despite a similar level of viral replication comparable to RMs [23], [70], [88]. The biologically active forms of death ligands include both a soluble and a membrane bound form. Therefore, we quantified the presence of death ligands in the sera of SIV-infected monkeys. We found, concomitant with the increase of cell death in RMs, higher levels of FasL two weeks post-infection (Figure 6G). In contrast, we did not observe any increase in the levels of FasL in SIV-infected AGM (Figure 6G). We have reported during the acute phase the absence of TNF-α detection in the sera of both SIV-infected species [90], [91]. Although, we were unable to detect soluble TRAIL in the sera of SIV-infected monkeys due to the unavailability of appropriate reagents for its detection in non-human primates (data not shown), it has been reported that there is increased expression of Trail mRNA in SIV-infected RMs [92].
To assess the impact of soluble and membrane forms of death ligands, we investigated whether apoptosis of monocytes and DCs from SIV-infected RM may be modulated by antagonists to death ligands using decoy receptors. We demonstrated that decoy receptors of TNF (TNF-R1) and TRAIL (TRAIL-R2 but not TRAIL-R1), and to a lesser extent decoy receptor of Fas (Fas-Fc), reduced monocyte cell death, whereas decoys receptors of Fas and TNF (TNF-R1) reduced DCs cell death (Figure 6H). Interestingly, despite the fact that soluble TNF-α has no effect, antagonist antibodies partially inhibited death suggesting that TNF-α at the cell surface may participate in the death of APCs [93]. These results suggest that apoptosis of mononuclear phagocytes involved death-receptors and their counterparts.
In order to analyze the apoptotic pathways in monocytes, positive selection of CD14+ cells was performed from healthy and SIV-infected RMs. Western blots probed with specific antibodies to FLIP revealed that monocytes from SIV-infected RMs displayed lower amounts of FLIP (Figure 7A), as compared to healthy RMs. Thus, the absence of FLIP is consistent with the increase sensitivity of these cells to undergo apoptosis after ligation of death receptors. Moreover, we found that monocytes from SIV-infected RMs had lower amounts of Mcl-1 (Figure 7A). In one SIV-infected RM, we also detected an increased amount of the proapoptotic form of Mcl-1Exon-1. To assess the expression of active form of Bax and Bak proteins in APCs from healthy and SIV+RMs, we used specific antibodies that detect conformational changes as previously described [25]. In comparison to CD4+ T cells, we found that 20%±6 and 30%±11 of monocytes from SIV-infected RMs at day 14 express the active form of the pro-apoptotic Bax and Bak molecules respectively as compared to monocytes from non-infected RMs (less than 11%±4). Similar data were observed in DCs although to a lesser extent (Figure 7B and 7C). Thus, our results indicate that monocyte and DCs are engaged in a process leading to mitochondria damage supporting our observation that these cells are more prone to undergo apoptosis during the acute phase. Furthermore, we used a broad caspase inhibitor and demonstrated that by blocking caspase activation, cell death of APCs was also prevented (Figure 7D). We also demonstrated that the addition of caspase inhibitor led to an increase in the number of cells expressing TNF-α after stimulation with LPS + IFN-γ stimulation (Figure 7E). Altogether, our data demonstrated a critical role of both the intrinsic and extrinsic apoptotic pathways in controlling APC death during the acute phase of SIV-infection.
We demonstrate that monocytes and DCs are more prone to undergo apoptosis in response to death-receptor ligation after in vitro infection with HIV or ex vivo from SIV-infected RMs. In addition, our data show that HIV/SIV infection is associated with an increase in the active forms of the pro-apoptotic molecules Bax and Bak and with a decrease in the anti-apoptotic Mcl-1 and FLIP proteins in both cell types. Thus, these results suggest that both the extrinsic and intrinsic pathways are involved in the death of APCs during HIV/SIV infection. Broad inhibition of caspase activation using a synthetic peptide prevented this death and increased the number of TNF-α productive mononuclear cells.
Circulating monocytes are essential not only to replenish the pool of tissue macrophage populations but also may differentiate into inflammatory DCs in the tissues following microbial infection. Because peripheral monocytes and DCs represent crucial populations for the control of pathogens, this enhanced susceptibility to die by apoptosis in the presence of death ligands could have a major impact on the establishment of the adaptative immune response early after infection. Interestingly, other persistent viral infections such as lymphocytic choriomeningitis virus (LCMV) and measles virus (MV), which are associated with a generalized immune suppression in their natural hosts, also induce death of accessory cells early after infection [94], [95]. Our results also demonstrated in vitro that incubation of monocyte-derived MØ and DCs with HIV during differentiation not only increased the susceptibility of these cells to undergo apoptosis but also impaired their maturation and their capacity to produce inflammatory cytokines after stimulation. Their down modulation could have an impact on the hosts' ability to mount an effective SIV-specific immune response.
Our results showed abnormal early death of APCs was associated with AIDS. The low level of infection of DCs suggests that apoptosis is not necessarily associated with productive infection. Moreover, during the acute phase, the percentage of monocytes prone to undergo apoptosis (and expressing active form of Bax and Bak) was higher than the frequency of SIV DNA+ cells. Our data revealed that also in vitro HIV primes both monocytes and DCs to undergo apoptosis in response to death ligands despite the presence of an inhibitor of viral replication, ddI. In a similar manner, the non pathogenic-primate model suggests that despite intense viral replication during the acute phase [70], APCs are not prone to undergo apoptosis. Altogether, these results point to the involvement of indirect mechanisms leading to cell death. This may suggest that triggering of TLRs or other pattern recognition receptors such as a mannose C-type lectin receptor, by HIV could lead to the observed changes in the sensitivity of these cells to undergo apoptosis without active replication [96]. We and others have previously reported the critical role of cytokines determining the sensitivity of monocytes to undergo apoptosis [28], [29], [97], [98]. Among them IL-10 has been shown to be a potent cytokine to induce monocyte death [98] but also increases membrane-bound TNF-α[99], and the expression of CCR5, a co-receptor for SIV [100]. Recently, it has been reported that IL-10 results in the rapid elimination of mature DCs by NK cells but is associated with the accumulation of DCs having an immature phenotype [101]. Blockade of IL-10, in addition to blockade of PD-1 signaling has been suggested as a means to restore anti-viral T cell responses in chronic LCMV infection [102] and to prevent apoptosis [16], [103], [104], [105], [106]. Therefore, whether a therapy based on neutralizing IL-10 antibody would be able to prevent monocyte cell death as well DCs in vivo remains an open question.
During inflammation, or in the presence of microbial antigens, monocytes become resistant to death associated with increased expression of anti-apoptotic molecules [28], [29], [31], [32], [107], [108], but resistance to death is counteracted by interferons (IFNs). Indeed, type I IFN production associated with bacterial pathogens such as Listeria monocytogenes [109] or in combination with LPS [110] induced apoptosis of APCs. We found that in vitro stimulation with LPS + IFN-γ induced the death of HIV-infected APCs as compared to uninfected cells. In HIV-infected patients, a recent study has reported increased levels of monocyte apoptosis after IFN-γ stimulation [66]. Interestingly, bacterial translocation associated with AIDS has been correlated with activation of innate immunity and especially with increased plasma levels of IFNα[111]. Moreover, in pathogenic primate models of SIV infection, increased levels of type I IFN related to the recruitment of pDCs in the lymph nodes (LN) have been reported [91] and associated with disease progression [112]. We found that mononuclear cells were abnormally sensitive to die through apoptosis concomitant with the peak of type I IFN production [91]. Thus, the presence of type I/II IFN may result in increased sensitivity of APCs to undergo apoptosis during the acute phase. After the peaks of viral replication and type I IFN, our data revealed a decrease in the susceptibility of these cells to undergo apoptosis. In chronically HIV-infected persons, a resistance to undergo apoptosis was reported [67]. Therefore, these results together reinforce the idea that death of APCs early after infection contributes to immune deficiency and further progression to AIDS.
Our data clearly demonstrated that in vitro and ex vivo, APCs were more sensitive to undergo apoptosis in response to specific death-ligands. Monocytes were more prone to die after TRAIL binding than FasL, whereas FasL was more efficacious to induce death in DCs. This increased propensity to undergo apoptosis after death-receptor ligation was not related to an increased expression of the death-receptors at the surface of APC's, but was associated with increased levels of FasL in the sera of SIV-infected RMs and not in non-pathogenic SIV-infected AGMs. Of note, FasL levels in plasma of HIV-positive individuals have been reported to be elevated, and correlated with HIV RNA burden [113], [114]. Elevated levels of TRAIL have also been reported in HIV-infected individuals early after infection [115], [116]. However, our attempts to measure TRAIL in the sera were unsuccessful due to the absence of available reagents for its detection in monkeys. Whereas the incubation of soluble TNF-α had no effect on the death of APCs from RMs, neutralization of TNF-α reduced death. Biologically active forms of death-ligands include both membrane bound and soluble forms, suggesting that TNF-α can be a co-factor for death-receptor sensitization at the cell surface [93]. Taken together, these results suggest that peripheral blood monocytes and DCs from pathogenic SIV-infected macaques are exposed to death-ligands during the acute phase.
Cell death via death receptors can be regulated at different levels, including altered expression of death-ligands or by inhibition of intracellular signaling events. In this context, our results showed that despite death-receptor expression and the presence of death-ligands in the culture, uninfected monocytes and macrophages were resistant to apoptosis indicating that death-ligands and receptors were not sufficient to induce apoptosis. These data are consistent with a model in which infection induced changes in the susceptibility of monocyte and DC populations to undergo apoptosis. The susceptibility of the cells to undergo apoptosis depends on the balance between pro- and anti-apoptotic molecules. We observed a downregulation of FLIP, an inhibitor of the DISC and caspase activation, in HIV-infected MØ- and DCs-derived monocytes. In purified CD14+ monocytes from monkeys, we also found a downregulation of FLIP in SIV-infected RMs as compared to healthy RMs. Moreover, the susceptibility of these cells to undergo apoptosis was also associated with increased expression of the active forms of the pro-apoptotic molecules Bax and Bak. Finally, we demonstrated a reduction in the expression of the anti-apoptotic Mcl-1 proteins but its pro-apoptotic form, Mcl-1Exon-1, was increased. Thus, our results demonstrated a dysregulation in the balance of pro- and anti-apoptotic molecules, which could contribute to mononuclear phagocyte death. Our results indicated therefore that both the extrinsic and the intrinsic pathways could be closely linked in determining mononuclear cell apoptosis outcome after HIV/SIV infection. In this sense a broad caspase inhibitor prevented cell death. Since monocytes are a heterogeneous population [117], it remains to be determined whether CD14dimCD16+ monocytes display similar susceptibility to die as CD14+ monocytes during HIV/SIV infection.
In conclusion, our findings demonstrate that HIV/SIV infection primes mononuclear cells to undergo apoptosis. Since circulating blood monocytes and DCs extravasate into tissues in response to pathogens, such sensitization to death-receptor mediated apoptosis may be a major factor leading to the defective immune response observed during the acute phase. Taken together, our results highlight the confounding role of apoptosis induction in the physiopathology of HIV/SIV infection associated with the death of mononuclear cells during the acute phase of SIV infection. Thus a strategy aimed at blocking their death could be beneficial in restoring an effective anti-viral response in HIV-infected persons.
Ten RMs (Macaca mulatta) seronegative for STLV-1 (Simian T Leukemia Virus type-1), SRV-1 (type D retrovirus), herpes-B viruses, and SIVmac were utilized. RMs were inoculated intravenously with ten 50% animal-infectious doses of the SIVmac251 strain (provided by AM. Aubertin, INSERM U74, Strasbourg, France). Four AGMs of sabaeus species were experimentally infected with 300 TCID50 of SIVagm.sab92018 strain [70].
All the animal experiments described in the present study were conducted at the Institut Pasteur according to the European Union guidelines for the handling of laboratory animals (http://ec.europa.eu/environment/chemicals/lab_animals/home_en.htm). The protocol was approved by the committee on the ethics of animal experiments of Ile de France (PARIS 1, #20080007). All surgery was performed under sodium pentobarbital anesthesia, and all efforts were made to minimize suffering.
The frequency of SIV-infected cells was measured by limiting dilution PCR. Cells were isolated from blood and stained with mAbs (CD3, CD4, HLA-DR and CD14 mAbs) (BD Biosciences, San Jose, CA). Cells were purified by cell sorting (FACS Vantage, BD Bioscences) on the basis of their size, granularity and phenotype (CD14+HLA-DR+CD3−CD20−; CD14−HLA-DR+CD3−CD20− versus CD3+CD4+); purity exceeded 98%. Purified cells were counted and serially diluted in a constant number of carrier CEMX174 cells as previously described [22]. The proportion of SIV DNA+ cells in purified cells was determined using Poisson law. The limiting-dilution PCR method detected one SIV DNA+ cell in 10,000 uninfected cells (CEMX174) validated with SIV-1C cells (provided by F. Villinger), that contain a single provirus of SIVmac251 per cell.
Fresh PBMC were isolated by density gradient centrifugation from blood of healthy donors. The blood samples were obtained from the Institut National de la Transfusion Sanguine. Monocytes were obtained by plastic adherence after extensive washing with media to remove non-adherent cells as described [98]. The adherent monocytes were carefully removed from the culture by incubating the plate 30 min at 4°C and use cold PBS and pipetting (not scraping). The purity of the monocytes exceed 90–95% as determined by flow cytometry after cell staining with mAbs CD14, CD3, CD20 and HLA-DR. The cells were then incubated in RPMI-1640 supplemented with 10% FCS, 1% glutamine, 1% pyruvate, and 1% antibiotics. Macrophages were derived from monocytes in the presence of GM-CSF (10 ng/ml) and IL-6 (5 ng/ml) (R&D system) while dendritic cells were derived in the presence of GM-CSF and IL-4 (10 ng/ml). At day one, the cells were incubated in the presence of the R5 HIV-1BaL strain (100 pg/ml of p24). At day 5, the cells were stimulated overnight with LPS (10 ng/ml) and IFN-γ (103 U/ml). Cells were also incubated in the presence or absence of TNF-α, TRAIL and FasL (200 ng/ml). Fresh PBMC from Non Human primates were isolated by density gradient centrifugation from blood and use to perform the different assays [21], [25], [88].
Infections of monocyte-derived MØ and DCs, respectively, were measured by flow cytometry based on the detection of intracellular p24 antigen (RD1-labeled mAb anti-p24 antigen, KC-57, Beckman coulter) after fixation and permeabilization of the cells (Intraprep permeabilization reagent, Coulter Coultronics). Productive HIV infection was also visualized by western blotting that allows detection of the presence of viral antigens in cell extracts. The immunoblots were incubated with sera obtained from a pool of HIV+ infected patients (kindly provided by F Mamano, Institut Pasteur). After treatment with horseradish peroxidase-linked goat anti-human secondary antibodies (Amersham Biosciences), immunoreactive proteins were detected using enhanced chemiluminescence (ECL+ from GE Healthcare) using a CCD camera (GBOX, SYNGENE).
The frequencies of SIV-infected CD4+ T cells, CD14+ and DCs were measured by limiting-dilution PCR [22] of purified cells by cell sorting (FACS Vantage; Becton Dickinson Biosciences, Le Pont de Claix, France) using the positive selection of cells stained with specific antibodies. Purified cells were counted and diluted in series in a constant number of carrier CEM X 174 cells. Cells were directly lysed with TPK buffer (10 mM Tris-HCl pH 8.3, 50 mM potassium chloride, 2.5 mM magnesium chloride, 0.5% Nonidet P-40, 0.5% Tween 20, 100 µg/ml of proteinase K). After 1 h at 56°C, proteinase K was inactivated at 95°C for 10 min. Twenty replicates of limiting dilutions were submitted to a nested PCR. SIV proviral DNA was amplified by nested PCR with SIV251-specific primers surrounding the nef region. After 35 cycles (95°C for 30 s, 60°C for 30 s, 72°C for 1 min.) with the first set of primers, Preco (59-CAG AGG CTC TCT GCG ACC CTA C) and K3 (59-GAC TGA ATA CAG AGC GAA ATG C), amplified a fragment of 961 base pairs, 10 µl of product was re-amplified (30 cycles 95°C for 30 s, 55°C for 30 s, 72°C for 1 min) with primers K1 (59-TGG AAG ATG GAT CCT CGC AAT CC) and A2 (59-GGA CTA ATT TCC ATA GCC AGC CA). Nested PCR products were electrophoresed through a 1.8% agarose gel. The proportion of infected cells was determined using Poisson law. The limiting-dilution PCR method was able to detect one infected cell in 10 000 uninfected cells (CEM X 174) demonstrated with SIV-1C cells (provided by F. Villinger), which contain a single provirus of SIVmac251 per cell.
Supernatants were collected after overnight stimulation and 6 days of culture. IL-1ß, IL-6, IL-8, and TNF-α were detected simultaneously by using the human inflammatory cytokine cytometric bead array (CBA) kit (BD Bioscience) [90]. The CBA working range was 20–5000 pg/ml for each cytokine. Cytokine levels were quantified by flow cytometry according to the manufacturer's directions. For intracellular TNF-α staining, the cells were incubated in the absence or presence of a broad caspase inhibitor Q-VD(OMe)-OPH (10 µM, MBL biomedical), and stimulated with LPS and IFN-γ. After 8 h, the cells were first stained with HLA-DR, CD3, and CD20 mAbs, washed and then permeabilized, before staining with PE-TNF-α mAbs (BD Biosciences). The number of HLA-DR+CD3−CD20− cells expressing TNF-α was measured by flow cytometry.
FasL in the serum was measured using a solid-phase immunoassay (MBL). The assay uses anti-FasL mAbs (clones, 4H9 and 4A5). The peroxidase substrate was used to quantify FasL and the optical density measured at 450 nm. The concentration was determined using a standard curve based on recombinant FasL. Three distinct ELISA specific for human TRAIL purchased from R&D system, Diaclone and Kamiya Biomedical Company, however, were unable to detect monkey Trail in the sera/plasma.
Monocyte-derived MØ and DCs, respectively, cultured in the absence or presence of the R5 HIV-1Bal strain, were then incubated in the absence (medium) or presence of LPS (10 ng/ml) plus IFN-γ (103 U/ml). Cells were then stained with FITC-CD14, APC-CD11c, PerCP-HLA-DR, and PE-conjugated antibodies to either CD83 (HB15e) or CD86 molecules (FUN-1) (BD Biosciences). Five hundred thousand events corresponding to mononuclear cells were acquired using a FACScalibur instrument (BD Biosciences).
Fresh PBMC from SIV-infected RMs at different days post-infection were isolated by density gradient centrifugation; apoptosis of monocytes and dendritic cells was determined at 24 h of culture by cell surface staining and with FITC-labeled annexin-V which is an early marker of dying cells detecting both caspase-dependent and -independent cell death programs [118]. The level of apoptosis was determined by flow cytometry as previously described [98]. We also used decoy receptors of Fas (Fas-Fc), TRAIL (TRAIL-R1-Fc and TRAIL-R2-Fc) and TNF-α (TNF-R1-Fc) at a dose of 10 µg/ml (Alexis corporation) as previously described [21].
Cells from SIV-infected RMs and healthy RMs were first labeled for cell surface markers (APC-HLA-DR, PE-CD14, Lineage PerCP-CD3/CD20 versus APC-CD11c, PE-HLA-DR and Lineage PerCP-CD3/CD20) and then fixed and permeabilized. Cells were then incubated with anti-Bax (BD Biosciences) or anti-Bak Abs (Calbiochem) as previously described for primates [25]. After washing, FITC-labeled goat anti-rabbit IgG Ab (Molecular probes) was added for 30 min at 4°C in the presence of mouse immunoglobulins. Cells were then washed and analyzed by flow cytometry.
Pellets of 3×106 monocyte-derived MØ and DCs, respectively, were lysed in Nonidet P-40 buffer (1% NP-40, 50 mM Tris-HCl (pH 7.4), 150 mM NaCl) containing protease and phosphatase inhibitors. Total lysates were resolved by SDS-PAGE (10–20% Tricine gels, Novex) and transferred to nitrocellulose membranes (Amersham Biosciences). After blocking nonspecific sites for 1 hour at room temperature with 5% nonfat milk and 0.2% Tween 20 in phosphate-buffered saline (pH 7.4), the membrane was incubated with rabbit anti-Bak (Calbiochem, clone 2–14), rabbit polyclonal anti-Bax (Santa-Cruz, N-20), rabbit anti-Mcl-1 (S19, Santa-Cruz), or rat anti-FLIP (DAVE-2, Alexis Corporation). To confirm equal protein loading and transfer, membranes were reprobed with anti-actin monoclonal antibodies (Sigma). After treatment with horseradish peroxidase-linked goat anti-mouse or anti-rabbit secondary antibodies (Amersham Biosciences), immunoreactive proteins were detected using enhanced chemiluminescence (ECL+ from GE Healthcare) using a CCD camera (GBOX, SYNGENE).
Data are reported as means ± SEM, and groups were compared using Mann-Whitney test (Prism software, GraphPad, San Diego CA). A p value <0.05 was considered significant.
|
10.1371/journal.pcbi.1000949 | A Computational Framework for Influenza Antigenic Cartography | Influenza viruses have been responsible for large losses of lives around the world and continue to present a great public health challenge. Antigenic characterization based on hemagglutination inhibition (HI) assay is one of the routine procedures for influenza vaccine strain selection. However, HI assay is only a crude experiment reflecting the antigenic correlations among testing antigens (viruses) and reference antisera (antibodies). Moreover, antigenic characterization is usually based on more than one HI dataset. The combination of multiple datasets results in an incomplete HI matrix with many unobserved entries. This paper proposes a new computational framework for constructing an influenza antigenic cartography from this incomplete matrix, which we refer to as Matrix Completion-Multidimensional Scaling (MC-MDS). In this approach, we first reconstruct the HI matrices with viruses and antibodies using low-rank matrix completion, and then generate the two-dimensional antigenic cartography using multidimensional scaling. Moreover, for influenza HI tables with herd immunity effect (such as those from Human influenza viruses), we propose a temporal model to reduce the inherent temporal bias of HI tables caused by herd immunity. By applying our method in HI datasets containing H3N2 influenza A viruses isolated from 1968 to 2003, we identified eleven clusters of antigenic variants, representing all major antigenic drift events in these 36 years. Our results showed that both the completed HI matrix and the antigenic cartography obtained via MC-MDS are useful in identifying influenza antigenic variants and thus can be used to facilitate influenza vaccine strain selection. The webserver is available at http://sysbio.cvm.msstate.edu/AntigenMap.
| Influenza antigenic cartography is an analogy of geographic cartography, and it projects influenza antigens into a two- or three-dimensional map through which we can visualize and measure the antigenic distances between influenza antigens as we visualize and measure geographic distances between the cities in a geographic cartography. Thus, influenza antigenic cartography can be utilized to identify influenza antigenic variants, and it is useful for influenza vaccine strain selection. Here we develop a new computational framework for constructing influenza antigenic cartography based on hemagglutination inhibition assay, a routine antigenic characterization method in influenza surveillance and vaccine strain selection. This method can be used for antigenic characterization in vaccine strain selection for both seasonal influenza and pandemic influenza.
| An influenza virus is a negative-stranded RNA virus that belongs to the Orthomyxoviridae family. There are three serotypes, A, B, and C, of which B and C are reported to infect mammals only. The influenza A viruses have genomic segments (segment ) with varying lengths from about to nucleotides which encode at least proteins: PB2 by segment , PB1 and PB1-F2 by , PA by , haemagglutinin (HA) by , nucleoprotein (NP) by , neuraminidase (NA) by , matrix protein M1 and M2 by , and nonstructural protein NS1 and NS2 by . Among these proteins, the surface proteins HA and NA are involved in virus attachment and cell fusion. Both HA and NA are the primary targets for host immune systems. The serotypes of influenza A viruses are based on HA and NA subtypes. To date, HA and NA subtypes have been reported in influenza A viruses [1]. For instance, H1N1 influenza A virus is named since it has HA and NA recognized by HA subtype and NA subtype antibodies, respectively. Influenza B viruses have segments while Influenza C has segments. There is not yet an HA-NA nomenclature system in Influenza B and C viruses.
The peak influenza season in the northern hemisphere is from January to April every year. More than hospitalizations and deaths are caused by influenza in the United States each year [2], [3]. The influenza A virus may cause a pandemic disaster that will impact multiple continents. In the 20th century, three influenza A pandemics occurred in 1918, 1957, and 1968, respectively [4], [5]. More than million people were killed in the 1918 influenza pandemic, which was caused by the H1N1 influenza A virus. This influenza pandemic shortened global life expectancy by more than years. During March and early April , a new H1N1 influenza A virus epidemic was detected in Mexico and the United States [6], and the virus spread rapidly through human-to-human transmission, resulting in WHO declaring a pandemic, which was the first influenza pandemic in the past years. This virus was estimated to cause about million infections and deaths solely in United States through Jan 14, 2010 (www.cdc.gov). If we consider all cases in five continents, the numbers will become significantly larger.
In the United States, vaccination is the primary option for reducing the effects of influenza. The seasonal influenza vaccines used in the past decades include three viral components: H1N1 influenza A virus, H3N2 influenza A virus, and influenza B virus. In an effective vaccination program, vaccine strain selection will be the most important step since the highest protection could be achieved only if there is an identical antigenic match of the vaccine and epidemic virus HA and NA antigens, especially HA, which is the primary target of human immune system. However, as an RNA virus, influenza A virus has rapid mutations in these two proteins, and such mutations can cause a change of antigenicity, thus making vaccines ineffective. Mutations in HA and NA are also referred as antigenic drift.
Immunological tests, such as hemagglutination inhibition (HI) assay, enzyme-linked immunosorbent assay (ELISA), and microneutralization assay, have been utilized to identify antigenic variants among the circulating influenza strains. Among these assays, HI, has been one of the routine procedures in influenza vaccine strain selection. HI assay is an experiment to measure how a testing influenza antigen (virus) and a reference antiserum (antibody) react. The antibody is usually diluted in fold first and then diluted in powers of . Thus, the titre from HI assay will be , . The larger the is, the more closely the testing antigens match the reference antigens, for which the reference antisera are generated. Usually a number smaller than is considered as a low reaction between antigen and antibody. In many cases, HI experiments are used to measure the antigenic distance between two testing antigens through their immunological reactions to the same reference antiserum. For instance, if one testing antigen is a high reactor for the reference antiserum (e.g. with a titre of ) while another testing antigen is a low reactor (e.g. with a titre of ). The antigenic distance could be approximately units, which is . In reality, the antigenic distances are usually measured by a set of reference antisera, thus the calculation is much more complicated. Such measurements from HI data are generally used to determine the antigenic distances between testing antigens.
In a typical influenza HI assay, generally less than reference antisera are used but the number of test antigens can be more than . However, interpretations of HI results are not straightforward due to the following two challenges: (1) HI assay only shows the indirect relationship between antigens and antisera since each value reflects a reaction from antigen, red blood cell (RBC), and antibody. Many variables from RBC and antibody will interfere the HI titres; (2) it is not be possible to perform HI for all pairs of antigen and antisera reactions. Thus, the resulting HI table is generally incomplete, and the percentage of missing data could be up to . By applying the metric multidimensional scaling method (MDS) to reduce the shape space into less than three dimensions, Lapedes and Farber [7] showed a linear correlation between logarithm values of HI titers and the space distances between influenza antigens. Based on this method, Smith et al. [8] constructed influenza cartography to visualize the distances among influenza antigens from HI tables by further developing the metric MDS method. Their method assumes that antigens and antibodies are mapped into the same low-dimensional space, and their interactions are the distances between the embedded points. However, in our implementation of their algorithm, the resulting influenza cartography depends on the initial values selected, and thus may not be stable. Moreover, this method results in cartographies in which global distances may contain relatively large errors. This is because the algorithm does not incorporate temporal modeling to reduce the inherent temporal bias in HI tables. The temporal bias is caused by the fact that HI table entries are not missing uniformly at random, and off diagonal entries are more likely to be missing or become low reactors (Figure 1). The underlying biological reason for this bias can be explained by the herd immunity effect, where influenza antigens evolve rapidly under the accumulating immune pressures of human population [9]. A more detailed illustration of this phenomenon will be given later.
The goal of this paper is to present a computational framework for influenza cartography construction which we call Matrix Completion-Multidimensional Scaling (MC-MDS). An important aspect of this framework is that temporal modeling can be easily incorporated, which as we shall show, is useful for dealing with HI tables with herd immunity induced temporal bias. Our framework includes two integrated steps: (1) a low rank matrix completion algorithm is first employed to fill in the entries of the HI matrix; (2) a MDS algorithm is utilized to map the antigens (or similarly, antibodies) into a two dimensional space for visualization. Our approach explicitly separates the visualization (cartography) step from the matrix completion step, making it easier to incorporate temporal models. Our experience shows that while temporal modeling is beneficial in both steps, it is less important in the first step, for which we may simply employ a sliding window approach; however it is more essential in the second step, for which we propose a more complex herd-immunity temporal regularization model as described in the Materials and Methods section. The reason for the difference is that the inherent temporal bias tends to give rise to incorrect global distances if not handled explicitly, and thus affect the 2D cartography process more significantly. The two step procedure in our approach is thus flexible in the first step, where we can simply use a standard low rank matrix completion algorithm. On the other hand, we have to pay special attention to temporal modeling in the second step, which is essential for accurate cartography construction. Both simulation and a practical application in H3N2 influenza A viruses demonstrate that this method is able to overcome some limitations in the original metric MDS method of [8] and it results in better influenza antigenic cartographies from HI data. Therefore the proposed framework can potentially facilitate more accurate interpretation of HI data in influenza surveillance as well as more accurate identification of influenza antigenic variants. Both are essential for influenza vaccine strain selection.
While greater details are given in the Materials and Methods section, we shall summarize the most important observations and intuitions in our computational framework before presenting the actual experimental results.
In this work we are specifically interested in HI datasets existing accumulating original, such as the immunological datasets of human origin. In a typical HI dataset, three types of data entries are present: Type I, a regular HI titre; Type II (low reactors), the value is defined as “less than a threshold”, e.g. and this threshold is caused by the lower bound experimental limit in HI assays indicating a weak (or low) immunological reaction between a testing antigen (virus) and an antiserum (antibody); Type III, missing values. A major characteristic of HI dataset is that the distributions of type I, type II, and type III data are not random. Specifically, if we arrange both antigens and antibodies in a HI matrix according to time, then there is a banded structure, where most Type I data appear very close to the diagonal of the matrix; Type II data tend to be slightly off diagonal, while Type III data are more likely to occur in matrix entries that are significantly off diagonal (Figure 1). This data characteristic introduces a “temporal bias” concerning the data distribution (in comparison to uniformly random distribution) that needs to be corrected. As we will show, if the problem is not handled appropriately, then inaccurate result will be produced. This is because classical methods assume uniformly random data distribution, which does not take the temporal bias effect into consideration. Our paper shows that temporal modeling, which reduces the data distribution bias in HI tables, is important in HI based influenza cartography.
The specific benchmark dataset used in our study includes entries, representing of all table entries (Figure 2). Among these entries, () are Type II values (that is, they are recorded as ) with . For algorithmic comparison purposes, we also include results on a simulation dataset with ground truth, which is generated according to characteristics of real HI datasets.
As pointed out above, most Type I data are located across the diagonal line of the HI matrix, which significantly deviates from the “missing uniformly at random” assumption in classical matrix completion. In order to reduce this bias, we adopt a sliding window approach where each low rank matrix completion will be performed in a HI sub-matrix, which has fewer amount of Type II and Type III data that more closely satisfy the “missing uniformly at random” assumption. The remaining entries that are not covered by the (sliding window) sub-matrices can be filled with a global matrix completion algorithm – those entries will be predicted with less accuracy due to the banded-structure of the HI data that violates the “missing uniformly at random” assumption.
The windows are based on the temporal spans of influenza A viruses. In order to complete the entire matrix, the algorithm will slide yearly along with both the dimensions of antigens and antisera to ensure the time difference between all antigens and antisera are within a certain window size. In order to obtain an optimal window size and best rank in matrix competition, we tested six different sizes, including , , , , , and , and ranks to . A -fold cross validation suggested that the time frames of -year and -year with rank are two best ones towards achieving the lowest RMSE (root mean squared error) value in matrix completion of H3N2 dataset (Table 1). The average RMSE from -year experiment is slightly better than that from -year experiment. Both the average RMSE for -year and -year experiment are better than that from the entire HI matrix. Thus, during matrix completion, a window of and a rank of will used. Similarly, our optimization method demonstrated that the window size of and the rank of are the best parameters for our simulation data.
After the matrix completion step, we need to project the influenza antigens onto a two-dimensional (2D) map. In order to obtain accurate global distances, we incorporate a temporal model in MDS based on the fact that the influenza antigens continue to evolve under the accumulating immune pressures of human population [9]. In order to evade the herd immunity, an influenza virus will most likely evolve into a strain with different antigenicity from recently circulating strains in human population. This intuition is mathematically incorporated in our temporal MDS model, where we assume that on the 2D cartography, influenza viruses tend to evolve along (approximate) straight-line segments during short time spans; that is, they tend to evolve in directions as far away from recently appeared viruses as possible. The detailed mathematical formula is presented in the Materials and Methods section.
In HI tables, a Type II value is resulted from experimental limitation of HI assay and reflects a weak (or low) immunological reaction between a testing antigen/antiserum pair. Although this value is not as informative as a Type I value, it is more useful than a Type III value (missing value). In particular, if a particular virus has type I values with a certain set of antibodies that show strong reactions, while another virus reacts weakly with the same set of antibodies (resulting in type II values), then the global distance between their 2D cartography embeddings should be relatively large. A set of constraints on global distances can be derived from this observation. The details can be found in the Materials and Methods section.
There are four parameters to be optimized in our temporal MDS model. We use -fold cross validations to select the optimal parameters that achieve the lowest RMSE while satisfying global distance constraints derived from Type II data. Our cross-validation results led to for the real data and for the simulation data.
To demonstrate the potential impacts of Type II data (low reactors) and Type III data (missing values) on the influenza cartography, we performed experiments using simulated HI matrices containing antigens versus antibodies in which we know the ground-truth. Three simulated HI matrices were generated, where one was based on the distributions of H3N2 1968–2003 HI data: (1) HI matrix ( data absence) with neither Type II nor Type III data; (2) HI matrix ( data absence, data structure: randomly distributed) with Type III data but without Type II data; (3) HI matrix ( data absence, data structure: with a temporal data missing bias similar to H3N2 data as shown in Figure 2) with both Type II data and Type III data. The first HI matrix serves as the benchmark data (ground truth). The second HI matrix is used to test the efficiency of standard matrix completion algorithms under the missing uniformly at random assumption. The third matrix is used to examine the efficacy of the temporal model in MDS. A more effective computational method would be expected to produce a cartography more similar to that of the benchmark matrix. Using these simulated HI matrices, we are able to compare the MC-MDS method proposed in this work to the original metric MDS method of [8] in terms of HI matrix completion and cartography construction accuracies.
To assess whether MC-MDS and metric MDS can accurately recover the HI values in the HI data, we calculated the local RMSEs for the Type I data using -fold cross validation (Table 2). The experimental data were partitioned into parts, and each time we use parts for training and part for testing. The RMSE values were calculated using the Type I values in the testing part. Here we only use Type I values for RMSE calculation in order to be consistent with our real-data experiment, where we do not know the ground-truth corresponding to Type II and Type III data. The local RMSE values were for MC-MDS and for metric MDS, where the notation of is used. Since a typical matrix value is about , these local RMSE values indicate that both methods were able to recover HI values effectively. The small difference between the two means of MC-MDS and metric MDS is significantly smaller than the standard deviations. Hence they are statistically insignificant. However, we note that metric MDS has a larger standard deviation, which is consistent with our observation that it is less stable.
The effectiveness of a cartography construction algorithm can be assessed using figures of merit that measure its robustness and correctness. The robustness of a method is determined by the correlation coefficient (CC value) that is calculated from the pairwise distances among antigens for every two independent runs. The correctness of cartography is measured by two values: the difference between the maximum distances (MD value) between any antigens in the benchmark cartography and that from the method being evaluated (either MC-MDS or metric MDS); the pairwise distance RMSEs (PD value), calculated by measuring the difference between the pairwise distances among all antigens in the benchmark cartography and those from the method being evaluated. We performed independent runs, and the mean and standard deviation for each figure of merit can be found in Table 2.
As specified in the Materials and Methods section, the matrix completion method employed in this paper was Alternating Gradient Descent (AGD). In Figure 3, the ground-truth cartography is given in Figure 3a. Figure 3b shows a typical result when matrix entries are missing uniformly at random (the second matrix generated in our simulation study), where the standard AGD method accurately reconstructed cartography since the resulting cartography is similar to that from the benchmark matrix. Figure 3c shows a typical result with temporally biased HI table (the third matrix generated in our simulation study), where the cartography was constructed from a combination of AGD for matrix completion and the conventional MDS (without temporal modeling) for cartography generation. It shows that this combination is unable to accurately recover the cartography of the benchmark data since the global distances are incorrect. In comparison, the combination of AGD with temporal MDS, shown in Figure 3d, does achieve significantly more accurate global cartography. This experiment demonstrates the need to explicitly incorporate temporal modeling into the MDS step. Moreover, our experiment shows that cartographies generated by AGD and temporal MDS are stable. The CC value and PD value for the independent runs are and , respectively. The MD value for the independent runs is , which is close to the ground-truth value of in the benchmark cartography (Figure 3a).
For comparison, we implemented the metric MDS method of [8] and applied to the third HI matrix which was generated with temporally biased data type distributions. Our results indicate that the cartographies from metric MDS are less stable, with two typical runs given in Figure 3e and 3f. In the independent runs of metric MDS, the CC value and PD value for the independent runs are and . The MD value is . These numbers are significantly worse than the corresponding numbers from the MC-MDS method proposed in this work. We shall especially note that the metric MDS method tends to over-estimate the global distances in this stimulation study. Moreover, the large standard deviations in the results also indicate that metric MDS is not very stable.
While these two methods achieve similar matrix completion accuracies, the reconstructed cartographies reveal a more significant difference. As we pointed out earlier, this is because the temporal bias (of data type distribution) in HI tables has stronger impact in the MDS step, especially when we compare global distances. Without temporal modeling, the accuracy of global distances between two points (representing two viruses) in the 2D cartography decays more rapidly when the two points become further apart in time. While this reduction of accuracy is an unavoidable limitation of the banded structure in HI tables (Figure 1) that makes it harder to reliably compare points far away in time, a good temporal model can alleviate its impact, and thus increase the accuracy of the resulting cartography.
Finally we summarize the main observations from this simulation study as follows. Both MC-MDS and metric MDS methods achieved similar accuracy in recovering HI values. This means that they achieve comparable performance in the matrix completion sub-task, which is less sensitive to the temporal bias problem in HI tables. However, without temporal modeling, the global distances among far away points in the reconstructed cartography become inaccurate. Therefore it is helpful to incorporate temporal modeling into the MDS step in order to reduce the temporal bias effect. The proposed MC-MDS framework (with herd-immunity temporal model) is effective in reducing the bias problem, and it leads to more accurate cartography. The metric MDS appears to be less stable and it generates less accurate cartographies because the method does not address the temporal bias problem.
In the second experiment, we use MC-MDS to construct influenza cartography for H3N2 influenza A viruses from 1968 to 2003 using the HI datasets from Smith et al. [8]. The antigenic map is shown in Figure 4. The scale of antigenic cartography is based on the antigenic distances from HI tables, e.g. each unit (grid) in the antigenic cartography represents of a 2-fold change in HI titres. These viruses are specifically labeled as eleven clusters (HK68, EN72, VI75, TX77, BK79, SI87, BE89, and BE92, WU95, SY97, and FU02). Our results indicate that the antigenic distance between HK68 and FU02 is approximately units.
The resulting cartography can be compared to the published antigenic map in Smith et al. [8]. The overall trend in our results is similar to the cartography from Smith et al. [8]. However, there are two major differences: (1) The global distances in our cartography are smaller than those of Smith et al. [8]. For example Smith et al. [8] shows a distance of units between HK68 and FU02. Although we have no ground truth for this data, we note that this discrepancy is consistent with our simulation study, where the metric MDS method also produces larger global distances. In that case, the metric MDS method over-estimated the global antigenic distance between A and J by units more than the true distance. (2) The local cartographies between some clusters are different. For instance, the distance between WU95 and BE89 from our method is larger than those shown in Smith et al. [8]. In order to examine which antigenic cartography is likely to be more accurate, we performed a small cartography for H3N2 HI data from 1987 to 1995. Since the number of Type II data on the HI data from 1987 to 1995 is quite small, the effects of Type II on the antigenic cartography is minimal. Therefore, the cartography for the viruses between 1987 to 1995 using data from the limited span will not suffer much from the temporal bias problem discussed in the paper, and thus should be close to the true cartography. Our result shows that the distance between WU95 and BE89 should indeed be larger than that between BE95 and BE92 (Figure 5), and this is consistent with the local cartographies from MC-MDS.
Similar to the simulated HI data experiments, we can assess the robustness of MC-MDS and metric MDS on the H3N2 data (Table 2). The best local RMSE was for MC-MDS and for metric MDS. Therefore there is no statistically significant difference in matrix completion quality. The CC values from the independent runs are and for MC-MDS and metric MDS, respectively. The MD value was for MC-MDS and for metric MDS. These numbers are consistent with the simulation study, showing again that MC-MDS is more stable for antigenic cartography construction.
From the runs of metric MDS, we were not able to generate the exact cartography in Smith et al. [8]. One reason might be that the initial values we randomly chose were not exactly the same as those from [8], which were not specified clearly from [8]. The source code of our implementation of the metric MDS method in [8] is available upon request. We shall point out that our implementation is strictly based on what was described in [8]. While we have spent great effort to ensure the correctness of our implementation, it is possible that there are undocumented improvements in the optimization algorithm used to solve the metric-MDS problem. In such case, their actual implementation might not suffer from the issues observed in our study. Nevertheless it is still useful for us to examine problems of the algorithm presented in their original paper, the underlying causes of these problems and their potential mathematical remedies. This is what this study tries to achieve.
Each year, about World Health Organization (WHO) collaborating laboratories and National Respiratory and Enteric Virus Surveillance System (NREVSS) that are located throughout the United States participate in virologic surveillance for influenza. By collaborating with over other National Influenza Centers in the WHO Global Influenza Surveillance Network, the vaccine strains for next influenza season are determined in the middle of February each year for northern hemisphere (these strains are used as vaccine strains in the United States) and September for southern hemisphere. The pandemic vaccine strains are also selected through collaborative efforts among different laboratories across the WHO Global Influenza Surveillance Network. Influenza vaccine strain selection is a very labor intensive procedure that depends on both antigenic characterization and genetic characterization. In general, whether an isolate will be sequenced or not is based on the result from antigenic characterization, and only highly potential antigenic variants are sequenced. Therefore, antigenic characterization is critical for vaccine strain selection. In order to identify a potential influenza vaccine strain, we have to integrate the HI tables from different experiments in the same laboratories or even from different laboratories. Each experiment only includes up to 15 reference antisera, which are updated at each influenza season or even each month within the same influenza season. In addition, it is common for individual laboratories to use different antisera. Therefore, the integrated HI table is typically an incomplete matrix. This incompleteness and the limitation of HI experiments (see the introduction section) present a challenge in interpreting HI results and thus antigenic variant identification. Another important challenge of HI data is the temporal bias effect, which means that entries in an HI matrix are not missing uniformly at random (Figure 1). These are the problems this paper addresses.
As an analog of geographic cartography, influenza cartography can be used to visualize and measure antigenic distances between influenza viruses. An essential criterion for a new influenza vaccine strain is significant antigenic divergence (e.g. fold change in HI test) from the current vaccine strain. Influenza antigenic cartography can help us identify whether a testing antigen (virus) is antigenically far away from a specific vaccine strain or a specific cluster of antigens (e.g. circulating strains at a specific time period).
In this study, we proposed a new computational framework for constructing an influenza antigenic cartography, and demonstrated its usefulness in antigenic characterization. This computational framework has two integrated steps: (1) through a matrix completion algorithm, influenza antigenic distance matrices are constructed; (2) through MDS (with herd-immunity temporal model), influenza antigens (viruses) are projected onto a two-dimensional cartography. We specifically pay attention to the major challenge that is caused by the temporal bias in HI datasets. That is, the banded structure of HI entries indicates that the matrix entries are not missing uniformly at random (Figure 1), which violates the standard assumption in conventional methods. Our experiment showed that standard approach will not handle this problem very well, and will produce cartographies with incorrect global distances. This paper addresses the problem through a biologically motivated temporal evolution model that is mathematically incorporated into the MDS algorithm. It is shown that more accurate antigenic distances can be obtained from this approach.
Although MC-MDS is presented as a 2D cartography construction method in this paper, it can be extended easily for 3D (or even higher dimensional) cartography by modifying the resulting cartography dimension in the MDS step of our computational framework.
The temporal regularization in MC-MDS is based on the fact that the influenza antigens continue to evolve under the accumulating immune pressures of human population [9]. Within a short time period, the antigenic distances among viruses tend to become larger in temporal order. Such a regularization is important since it can effectively minimize the biases of Type II data. However, such regularization does not necessarily imply that the antigen would always evolve forward. Theoretically, it is possible that the antigenicity (not genetic sequence) of influenza viruses could become similar to earlier circulating strains when the selective pressure from herd immunity disappears. This is supported indirectly by the report that 2009 pandemic H1N1 virus cross-reacted with the serum from the ages over , who were likely to infect the seasonal H1N1 virus circulating in human population before 1957 [10].
Besides the immunological datasets for the influenza viruses (such as those of human origin) with the accumulating immunity from their hosts, there are other immunological datasets for the influenza viruses from mutations (not necessarily accumulating immunity), such as those of swine or avian origin. For the latter case (e.g. the data of swine or avian origin), our limited experiments in H5 and H7 studies suggested that the users can use MC-MDS directly without temporal model (data not shown). However, there might be additional structures to explore in such data. This requires more extensive investigations in the future.
We introduced a new computational framework for influenza antigenic cartography construction from HI datasets. This approach, which we refer to as MC-MDS, integrates two mathematical procedures: matrix completion and MDS projection (with temporal modeling). Using the AGD matrix completion algorithm on HI datasets from 1968 to 2003, we successfully identified the eleven reported clusters of antigenic variants that represent major antigenic drift events during these years. Thus, this method is useful in both influenza antigenic variant identification and influenza vaccine strain selection. Our results also demonstrated that MC-MDS is more robust and effective than our implementation of the metric MDS method [8] in influenza antigenic cartography construction.
The goal of matrix completion is to fill the missing entries in an incomplete matrix based on appropriate mathematical models of the matrix. It is a traditional mathematical problem that has been studied for many decades. Early contributions on this problem include Schur [11], Farahat and Ledermann [12], Friedland [13], Hershkowitz [14], London and Minc [15], Mirsky [16] and Oliveira [17]–[19]. In the past decade, interest in the problem has grown substantially, especially after the launch of Netflix competition [20] in 2007. The Netflix problem is to predict each user's movie preference (in order for Netflix to make appropriate movie recommendations to each user) from approximately observed user ratings. This can be regarded as a matrix completion problem, where we predict missing user/movie ratings from incomplete observations. This is exactly like the problem of predicting antigen/antibody interactions which we consider in this paper. In general, matrix completion is ill-posed and computationally intractable [21], [22]. However, recently, Candes and Rect [22] and Recht et. al [23] proved that under appropriate conditions, the minimum rank matrix solution can be recovered from incomplete entries by solving a convex optimization problem. These theoretical developments generated further interest, and afterwards, a number of new methods have been proposed [24]–[30].
If we do not consider the temporal bias effect, then the antigenic cartography task can be formulated as a matrix completion problem. Simply, in an HI matrix, there are antigens corresponding to the rows, and antisera corresponding to the columns. Let denotes the HI value from the reaction between testing antigen and antiserum . The HI matrix can be represented asLet denote the subset of 's entries corresponding to Type I and Type II data. In practice . The goal of matrix completion is to estimate the HI values in the Type II and Type III entries as accurately as possible. In addition, matrix completion can re-estimate Type I entries and remove embedded noises, which were from the uncertainties in experimental measurements. This is a standard matrix completion problem. The standard approach to this problem is to assume that the matrix is low rank, with rank . In our application, this means that each antigen can be embedded into the -dimensional space as , and each antiserum can be embedded into the -dimensional space . In the low rank model, the interaction between antigen and antiserum is given by for some matrices . We can aggregate vectors into a matrix and aggregate vectors into a matrix . Mathematically, the low-rank model is to find matrices with dimensions , with dimensions and a diagonal matrix with dimensions such that(1)
Here we describe AGD matrix completion method, which is developed based on gradient decent method. AGD method assumes the low rank matrix completion model (1).
If type II data are not present, one can employ the following optimization formulation to estimate the missing values(2)where when and otherwise.
The function is a regularization condition for the matrix , which is introduced to stabilize the solution. The solution of the optimization problem (2), which does not contain any missing value, will replace (which has missing values) as the true (and denoised) HI table, which we can then use for other purposes, such as cartography construction.
In the AGD method, we take in (2), where when and otherwise. () denotes the th row of () and .
First, the algorithm uses SVD to obtain the factorization . Here is the trimmed matrix of where we randomly set some observed values to from the rows (columns) when a row (column) contains more than () observed values. The purpose of this trimming step is to guarantee that each row (column) has less than () non zero values. This is based on the observations of Keshavan et. al. [27] that when , the corresponding singular vectors are highly concentrated on high-weight column (or row) indices. It means that those vectors do not provide useful information. After SVD, we set the initial value to and to where and are the first columns of and respectively.
We then apply the following alternating optimization procedure until convergence or when certain number of iterations are reached.
The gradient of and are:(3)(4)whereand
|
10.1371/journal.pcbi.1004265 | Energy Efficient Sparse Connectivity from Imbalanced Synaptic Plasticity Rules | It is believed that energy efficiency is an important constraint in brain evolution. As synaptic transmission dominates energy consumption, energy can be saved by ensuring that only a few synapses are active. It is therefore likely that the formation of sparse codes and sparse connectivity are fundamental objectives of synaptic plasticity. In this work we study how sparse connectivity can result from a synaptic learning rule of excitatory synapses. Information is maximised when potentiation and depression are balanced according to the mean presynaptic activity level and the resulting fraction of zero-weight synapses is around 50%. However, an imbalance towards depression increases the fraction of zero-weight synapses without significantly affecting performance. We show that imbalanced plasticity corresponds to imposing a regularising constraint on the L1-norm of the synaptic weight vector, a procedure that is well-known to induce sparseness. Imbalanced plasticity is biophysically plausible and leads to more efficient synaptic configurations than a previously suggested approach that prunes synapses after learning. Our framework gives a novel interpretation to the high fraction of silent synapses found in brain regions like the cerebellum.
| Recent estimates point out that a large part of the energetic budget of the mammalian cortex is spent in transmitting signals between neurons across synapses. Despite this, studies of learning and memory do not usually take energy efficiency into account. In this work we address the canonical computational problem of storing memories with synaptic plasticity. However, instead of optimising solely for information capacity, we search for energy efficient solutions. This implies that the number of functional synapses needs to be small (sparse connectivity) while maintaining high information. We suggest imbalanced plasticity, a learning regime where net depression is stronger than potentiation, as a simple and plausible means to learn more efficient neural circuits. Our framework gives a novel interpretation to the high fraction of silent synapses found in brain regions like the cerebellum.
| The brain is not only a very powerful device, but it also has remarkable energy efficiency compared to computers [1]. It has been estimated that most of the energy used by the brain is associated to synaptic transmission [2]. Therefore to minimise energy consumption, the number of active synapses should be as low as possible while maintaining computational power [1, 3, 4]. The number of active synapses is the product of the activity and the number of synapses. Energy can thus be reduced in two ways: 1) by employing a sparse neural code, in which only few neurons are active at any time, 2) by removing synapses leading to sparse connectivity, leaving only few synapses out of many potential ones. This latter process is also called dilution of the connectivity. Remarkably, during human development brain metabolism neatly tracks synapse density, rapidly increasing after birth followed by a reduction into adolescence (e.g. compare the data in [5] to [6]).
Most computational algorithms of learning, however, optimise storage capacity without taking energy efficiency into account (but see [3]) and as a result only limited agreement between models and experimental data can be expected. The best studied artificial example of learning is the perceptron which learns to classify two sets of input patterns. Despite its simplicity, results of perceptron learning are crucial as they for instance guide the design of recurrent attractor networks [7–9]. Provided the task can be learned, the perceptron learning rule is guaranteed to find the correct synaptic weights. The traditional perceptron learning algorithm assumes that weights can have any value and can change sign. In that case a perceptron with N synapses can on average learn 2N random patterns. At the maximum load the corresponding weight distribution is Gaussian, i.e., the connectivity is dense and hence energy inefficient [10]. If one restricts the synapses to be excitatory, the capacity is halved [9, 11].
In this work we ask which learning algorithm maximises energy efficient storage, and thus maximises the number of silent synapses while still being able to perform a learning task [3]. However, finding the weight configuration with the fewest possible (non-zero) synapses is a combinatorial L0-norm minimisation task. This is in general a NP-hard problem [12, 13] and thus difficult to solve exactly. Using the replica method from statistical mechanics it is possible to calculate limits on the achievable memory performance with a fixed number of synapses [10], but such methods do not yield insight on how to accomplish this. An earlier approach prunes the smallest synapses after learning. If synapses are to be removed after learning, this procedure is optimal [14, 15]. Yet, as we will show it is far better to incorporate a sparse connectivity objective during the learning process.
Here we explore imbalanced plasticity as a simple and biologically plausible way to reduce the number of required synapses and thus improve information storage efficiency. In many memory models the amount of potentiation and depression are precisely matched to the statistics of the neural activity [16–19], but here we deliberately perturb the optimal plasticity rule by introducing a bias towards depression. This imbalanced plasticity finds weight configurations that require less functional synapses and that are thus more energy efficient.
We consider a recognition task from positive examples [20–22]. The perceptron should learn to give a response whenever a sample from a given category is presented. In contrast to the standard perceptron algorithm, which ‘unlearns patterns’ for which the neuron should not be active, the synapses are not modified for negative samples. It has been argued that this setup is relevant to biology in particular when the set of negative samples is very large and/or its statistics unknown [22]. For instance, one might want to train a neuron to recognise fruits, but not update the synapses for all other objects. This setup is also relevant when studying reinforcement learning, where learning is gated by reward feedback elicited by positive samples. Finally, it resembles the one-class support vector machine used in statistical learning, which detects whether a sample belongs to a class and which has applications in anomaly detection [23, 24].
The setup is illustrated in Fig 1. A single postsynaptic neuron calculates the weighted sum of its N excitatory inputs and compares it to a positive threshold θ N. Whenever h = ∑ i = 1 N w i x i − θ N is non-negative, the perceptron fires. The inputs xi are randomly chosen to be -1 or +1 with equal probability, and independently of the other inputs (see below for extensions). The N in the threshold is a mathematical convenience that ensures scaling of the system as the number of inputs is varied [11, 25].
During learning the neuron is provided with a set of K positive patterns, x1, …, xk, …, xK. As in the standard perceptron, we cycle through the set of patterns until the task is learned. The goal of the perceptron is to ‘fire’ for all these patterns. This should be contrasted to setups in which samples are presented only once (one-shot learning), which generally lead to a lower capacity [25]. We assume that initially all weights wi are zero (tabula rasa). The learning rule is as follows: whenever a positive pattern is presented and only if it does not lead to postsynaptic activity, the synapse is updated. For high inputs, i.e., xi = 1, potentiation occurs
Δ w i + = a [ 1 - Θ ( h ) ] , (1)
where Θ(⋅) is the Heaviside step function which is zero if its argument is negative and one otherwise, and a ≪ 1 is the potentiation rate. Similarly, when an input xi is low, the synapse depresses
Δ w i - = - b [ 1 - Θ ( h ) ] , (2)
where b is the amount of depression. Depression is followed by rectification so that all synapses remain excitatory, wi ≥ 0 [26–30]. If the pattern does already lead to firing of the perceptron, no synapse is altered. This stop-learning condition is also present in a standard perceptron; possible biophysical mechanisms are discussed in [31].
For the simple, random pattern statistics used here, the non-negativity constraint limits the maximal amount of patterns that can be learned to Kmax = N [9, 11], which is half of the number of patterns an unconstrained perceptron can learn. Below this limit the learning process finishes with high probability in a number of steps that is polynomial in N. We define the memory load α = K/N, which becomes αmax = 1 at the maximal load in the balanced case.
Unlike the traditional perceptron rule, we allow for distinct amounts of potentiation and depression. By introducing imbalance in favour of depression the learning dynamics is biased towards the hard bound of the weight at zero. We rewrite the plasticity rule using the learning rate ε ≡ (a+b)/2 and an imbalance parameter λ ≡ (b−a)/2ε. Provided the synapse does not hit the zero bound, the weight update is
Δ w i = ϵ [ 1 - Θ ( h ) ] ( x i - λ ) . (3)
The parameter λ is zero for balanced learning; depression is stronger than potentiation if 0 < λ ≤ 1. We find somewhat improved faster learning when we also depress even when the pattern has already been learned, i.e.
Δ w i = ϵ { [ 1 - Θ ( h ) ] ( x i - λ ) - Θ ( h ) λ } . (4)
For that case it can be shown that the learning dynamics minimises the energy function
E = ∑ k = 1 K [ θ N - ∑ i = 1 N w i x i k ] + + λ ∑ i = 1 N w i , (5)
where [⋅]+ denotes rectification. The first term of the energy sums over all patterns and promotes low false negative rates; it is zero if the perceptron fires, while it attributes a cost proportional to the distance to the firing threshold whenever a pattern is not yet learned. The second term acts as a linear regulariser; the depression-potentiation imbalance λ penalises synaptic weight configurations that have large linear norms ∣ w ∣ ≡ ∑ i = 1 N w i. The regularisation term has a simple interpretation, as it is proportional to the mean synaptic weight, ∣w∣ = N⟨w⟩. The plasticity rule, Eq 4, minimises this energy by performing a stochastic sub-gradient descent [32], projected onto the subspace {w: wi ≥ 0, i = 1, …, N}.
Rewriting the learning rule as the minimisation of the energy Eq (5) shows explicitly why introducing imbalance towards depression promotes weight sparseness. In linear regression and classification, optimising over regularised energy functions that penalise the L1-norm ∣ ∣ w ∣ ∣ 1 ≡ ∑ i = 1 N ∣ w i ∣ of the weights is well-known to induce sparseness [33–35]. Below the critical load αmax the weight configuration with minimal linear norm is known to be sparse [27]. Thus, the learning rule Eq (4) with imbalance λ > 0 will try to find solutions that satisfy the learning conditions but that are sparser than those obtained when λ = 0.
While the linear norm constraint promotes sparseness, it is not guaranteed to produce the sparsest possible solution. The true optimisation problem would be to minimise the L0-pseudo-norm ∣∣w∣∣0. The L0-pseudo-norm simply counts the number of non-zero synapses. However, this leads to a difficult NP-hard combinatorial optimisation task [12, 13]. Instead, optimising under the L1-norm constraint is a convex relaxation of the original problem for which efficient computer algorithms exist (e.g. [36]). Moreover, imbalancing plasticity has the advantage of being an online procedure that only requires tuning the potentiation and depression event sizes and is thus biologically plausible.
Ideally our perceptron learns all examples, and minimises the false positive rate. To characterise the performance we present the perceptron with learned samples and lures (other random patterns), both presented with equal probability. The mutual information between the class of the input pattern and the perceptron’s output on a given trial is
I = ∑ x ∈ { p , l } ∑ r = 0 , 1 P ( x ) P ( r ∣ x ) log 2 P ( r ∣ x ) P ( r ) , (6)
where P(x) = 1/2 is the probability that the test pattern is a positive pattern (p) or negative lure pattern (l), P(r) is the probability that the perceptron remains silent or fires, and P(r∣x) is the conditional probability that we observe a given response given the true pattern class.
The information can be expressed in terms of the false positive rate p01 and the false negative rate p10. Below the critical capacity (α ≤ αmax), the positive samples are recognised perfectly after learning, i.e. there are no false negatives (p10 = 0), so that the information is determined by the false positive rate only. As we have 2K trials, the total information normalised per synapse, C = 2 K N I, equals
C = 2 K N ( 1 - 1 2 [ ( 1 + p 01 ) log 2 ( 1 + p 01 ) - p 01 log 2 p 01 ] ) . (7)
Although this type of information calculation is common, we note that testing with equiprobable lures and learned patterns is somewhat sub-optimal in terms of information [37]. For the one-class perceptron, testing exhaustively with all 2N−K possible lures gives about 60.6% more information when p01 = 1/2 with a weak dependence on p01.
As the mutual information does not take energy efficiency into account, we consider a recently suggested capacity measure that includes the sparseness of the final weight configuration [3]. The memory efficiency S measures the information per non-zero synapse by normalising the information to the fraction of non-zero synapses F,
S = C F . (8)
Memory efficiency is thus measured in bits per functional synapse. Learning rules that achieve high information C using few resources will have high efficiency. If one assumes that a non-zero synapse has a certain energy cost (independent of synaptic weight) and a zero synapse has none, the memory efficiency S measures the energy cost of the stored memory.
A variant of the sign-constrained perceptron convergence theorem (see Methods) shows that the learning algorithm Eq 3 converges below a critical imbalance λmax(α) that depends on the memory load α. In computer simulations we focus on the two extreme cases, i.e., balanced (λ = 0) and maximally-imbalanced λ = λmax(α) plasticity. In principle it is possible to find the maximum imbalance by trying various values of λ and checking convergence of the learning process. However, it is much quicker to use that the problem is equivalent to learn the patterns while minimising the linear norm ∣w∣, see Eq 5. This was done with a linear programming solver (see Methods) which requires no manual search for the maximal imbalance.
For strongest depression (λ = λmax), the information C is only slightly below the information of balanced learning, Fig 2A (magenta vs. blue curve). However, imbalanced plasticity provides a large increase in memory efficiency S, Fig 2B. The reason is that the learning dynamics converges to synaptic configurations with a considerably larger number of silent synapses, Fig 2C. As the memory load α increases, the efficiency approaches that of the balanced solution. This is expected; by increasing the task difficulty we are imposing additional constraints on the synaptic weights. As a result the volume of the solution space shrinks and the constraint on the mean weight has to be relieved, therefore leading to smaller gains in memory efficiency. As α approaches its critical value, the space of solutions collapses to a single point, i.e., no additional constraints can be imposed at critical capacity and λmax = 0 [7].
We also considered alternative learning algorithms: first, a minimal-value pruning rule, where all weights below a certain threshold are set to zero after learning has converged. We set the deletion threshold of the offline pruning algorithm to produce the same number of zero-weight synapses as the imbalanced solution. This is optimal in the one-shot learning case [14, 15]. In this case we find a more pronounced loss of information and, interestingly, almost no efficiency increase (dark green curve). The superiority of imbalancing makes intuitive sense: imbalanced plasticity is an online protocol that accommodates for sparseness constraints by redistributing weights dynamically, while the pruning procedure is performed after learning and does not allow for further re-adjustments. Finally, we also tried random pruning after learning, which as expected, performs very poorly (light green curve).
For completeness, we compared these results to the solution that maximises information without requiring sparseness. The optimisation can be formulated as a quadratic programming (QP) problem (see Methods), and the best solution can be found with a high performance barrier method convex optimiser [38]. This algorithm clearly lacks biological plausibility, and does not provide a significant improvement in information over balanced (λ = 0) online learning, Fig 2A. In other words, perceptron learning works well for our problem, provided that the firing threshold θ is large enough (see Methods). Under QP the fraction of silent synapses slightly increases to around 50%, Fig 2C, which leads to a moderate improvement in memory efficiency, Fig 2B. Finally, one can resort to the min-over learning rule, which only applies a weight update for the pattern that evokes the minimal output h [39]. The synaptic weights are guaranteed to asymptotically converge (as θ → ∞) to the QP solution and unsurprisingly the information matches that which is obtained with the quadratic solver. This procedure is difficult to reconcile with biology as well, as each single learning iteration requires access to every pattern.
The learning algorithm and the threshold setting also determine the shape of the synaptic weight distribution. This distribution is of importance, as it can be compared to experimental data. For instance, the electro-physiologically determined synaptic weight distribution was used to link Purkinje cell learning to perceptron learning theory [28, 40]. We recorded the obtained synaptic weight histograms (see Methods), averaged over many trials (each with different pattern sets). While collecting results across trials is strictly only approximates the synaptic weight density, it is a good estimate of the actual observed distribution for a single realisation of the system, since the underlying weight density is strongly self-averaging [27, 28].
Balanced learning (λ = 0) leads to an approximately exponential distribution, Fig 3A. Interestingly, although the QP solution did not increase information compared to online balanced learning (Fig 2A), the shape of the distribution of synaptic weights changes considerably (cf. Fig 3A and 3B). At any memory load α ≤ αmax the fraction of zero-weight synapses always remains close to 50% while the remaining weights assume a truncated Gaussian distribution centred around w = 0. The problem that we are dealing with is thus not ‘intrinsically sparse’ in weight space. This should be contrasted with the non-negative perce ptron classifier with 0/1-coded inputs that was recently studied [28–30]. In that case, maximising information in the presence of postsynaptic noise automatically leads to sparse weight configurations (F < 0.5), provided that the memory load is below the critical point. Interestingly, at the critical load, the distribution becomes identical to the truncated Gaussian that we report here as the optimal one.
Imbalanced plasticity boosts the fraction of zero-weight synapses and stretches the weight distribution, Fig 3C. Although the mean weight is lower due to the increased sparseness of the weight configuration, the surviving synapses are stronger. This can be understood through theoretical arguments (see Methods). It can be shown that learning rules that lead to a large minimum postsynaptic sum, min k ∑ i = 1 N w i x i k (together with a normalisation condition that fixes the Euclidean norm ∣∣w∣∣2) give better recognition performance against lures. As some synapses are zeroed-out, specific strengthening keeps the postsynaptic sum large for learned patterns.
The non-zero weight distribution for maximal imbalance can be reasonably fitted to a compressed exponential P(w) ∼ exp(−cwβ), with an exponent β = 1.4. The two-class perceptron model yields β = 2 (a truncated Gaussian) at critical capacity [28]. The best fit of this type of distribution to the cerebellar data published [40] has an exponent β = 0.7±0.4, however it should be noted that the limited amount of data allows for a broad range of possible β.
Next we explore if our findings depend on the details of the coding. So far we assumed the inputs were -1 or +1, as in earlier studies of the non-negative perceptron [9, 26, 27]. This is hard to imagine biologically, unless an inhibitory partner neuron is introduced [19, 31, 41, 42]. An arguably more faithful biological model is obtained by representing low inputs as silent, xi = 0 [16, 19, 20, 28, 43]. Furthermore, we wish to generalise to a case where the probability for a high input is variable rather than fixed to 1/2.
The capacity of the above model can be fully recovered without drastically changing the neural circuit. In fact, two ingredients suffice: one has to rebalance the plasticity rules as a function of the activity level f, and, secondly, introduce a dynamic mechanism that adapts the firing threshold as a function of the linear norm ∣w∣. With these modifications, both the information C and the memory efficiency S are exactly identical to those reported in the previous section.
First, we generalise the model to deal with an arbitrary coding level f. When f = 1/2, the original model is recovered up to scale factors. To preserve the zero mean, we consider activity patterns that are coded as zi ∈ {−f,1−f}, with P(zi = 1−f) = f. Stochastic sub-gradient descent dynamics over the energy Eq (5) gives the adjusted potentiation rule for high inputs
Δ w i + = ϵ { ( 1 - f - λ ) [ 1 - Θ ( h ) ] - λ Θ ( h ) } , (9)
while depression at low inputs becomes
Δ w i - = ϵ { - ( f + λ ) [ 1 - Θ ( h ) ] - λ Θ ( h ) } , (10)
followed by rectification. Here h = ∑ i = 1 N w i z i − θ f N.
Next, a zero-mean input zi is related to 0/1 coding by the simple relation xi = zi+f, xi ∈ {0,1}. Therefore the net input of the neuron in response to a 0/1 pattern can be written through a change of variables as
h = ∑ i = 1 N w i x i - f ∑ i = 1 N w i - θ f N = ∑ i = 1 N w i x i - γ , (11)
where we defined a new threshold variable
γ = f ∑ i = 1 N w i + θ f N .
Note that this threshold grows during learning so as to compensate the increasing weights. This can be viewed as a kind of homeostatic adaptation process: as learning progresses, the neuron self-regulates so that it becomes harder to reach the firing threshold. While the incorporation of an auxiliary feed-forward inhibition circuit has been used in related models to increase capacity in the presence of non-negativity constraints [19, 31, 41, 42], the mechanism here does not directly depend on the precise pattern x of the presented input. It thereby obviates the need for coordinated plasticity with a partner interneuron as well as for precise temporal integration of inhibitory signals. Instead it could be implemented sub-cellularly without the aid of additional circuitry. Using the adaptive threshold, the information becomes independent of the input coding level f (Fig 4 solid line), while it decreases when the threshold is fixed (dashed curve). We note that, unlike for two-class learning, for one-class learning a high threshold suffices to implement a large-margin classifier.
An alternative route to recover capacity is to employ sparse coding, a finding that has been previously reported for the non-negative perceptron in a more general classification framework [43]. Here the asymptotic situation is rather simple, because as f → 0 and N → ∞ the original model is recovered and performance at low f approaches the ideal performance, Fig 4.
Activity correlations can severely limit the performance of learning rules, depending on the task and the nature of the correlations. For instance, in supervised memory tasks, Hebbian learning deteriorates under almost any type of correlation in the patterns [25, 44]. In contrast, more powerful plasticity rules equipped with a stop-learning condition, like the perceptron rule, are resistant to spatial input correlations [45], and can in some cases take advantage of input-output redundancies to store more patterns [29, 46].
To test the robustness of imbalanced plasticity to correlated activity we draw random patterns from a generative model that induces spatial presynaptic activity correlations (characterised by a parameter g, see Methods, [21, 45]). We first correlated the patterns such that the mean activity remained homogeneous across the inputs. Consistent with the standard two-class perceptron without synaptic sign-constraints [45], neither the imbalanced learning, nor the balanced rule are affected by input correlation, Fig 5A.
Next, we implemented a variation of the generative model that introduces heterogeneities in the input activity rates where some inputs tend to be active more often than others. Interestingly the imbalanced rule is robust to this type of correlation, Fig 5B. Whereas the efficiency of the other rules drops off, the efficiency of the imbalanced rule remains constant. The intuitive explanation is that the high activity synapses effectively experience balanced net potentiation and depression for non-zero imbalance λ. The imbalanced rule finds a high-information solution by silencing and ignoring the low activity inputs and subjecting the remaining synapses to the usual imbalanced learning protocol.
So far we have considered the recall of noise-free patterns, however, in the light of the many noise sources in the nervous system, it is important to confirm the noise robustness of the results.
First, we introduce transmission failures and spontaneous presynaptic activity, and test the learning with corrupted patterns, denoted x ˜. An active input is switched off with probability δ 10 = P ( x ˜ i = 0 ∣ x i = 1 ), while an otherwise silent presynaptic input fires with probability δ 01 = P ( x ˜ i = 1 ∣ x i = 0 ). The lures are generated with a matching mean activity, ⟨x⟩ = (1−f)δ01+f(1−δ10), to ensure that lure statistics match the patterns.
We examined the performance of the balanced and maximally-imbalanced rules, as well as thresholded synaptic pruning, under this stochastic synapse model, Fig 6A and 6B. The information of all three rules decreases smoothly as the input distortion increases. For dense patterns, f = 1/2, the efficiency of the maximally-imbalanced rule is initially the most affected by the introduction of noise, and becomes comparable to the thresholded deletion one for higher noise levels. For sparse patterns, Fig 6B, the efficiency is affected similarly by the noise for all three rules. The maximally-imbalanced and the thresholded solutions remain more efficient than balanced plasticity.
Next, we examined the role of postsynaptic current noise by adding a zero-mean Gaussian variable to the postsynaptic current h [28], the variance of which sets the noise intensity, Fig 6C. In contrast to the above, the magnitude of the random contributions is decoupled from the actual learned weights. For this noise model, the relative information reduction is comparable for both balanced and imbalanced plasticity.
In the above the imbalance parameter λ was optimised for automatically in an unbiological fashion. To examine suboptimal values we simulated learning while raising λ towards the critical imbalance λmax, above which the learning algorithm no longer converges. The memory task difficulty, set by the memory load α, limits the allowed imbalance (see Methods). Indeed, we find that λmax shrinks as α increases, Fig 7. Akin to the margin parameter which sets the noise robustness of the non-negative perceptron [28, 29], the actual λmax depends on the exact set of patterns the neuron should learn. However, for random patterns drawn from the same distribution, the system is self-averaging as N → ∞ [7]. In simulations we observe a similar behaviour across different runs, although some finite-size effects are still apparent in networks of moderate dimension, Fig 7 (rightmost curves). In other words, λmax can be reasonably estimated independent of the precise pattern set. Finally note that the figure implies that the parameter can be set conservatively, based on the maximum number of patterns to be expected. Of course, the efficiency gain is not maximised in this case, but still better than the balanced case.
The brain’s energy consumption is thought to be dominated by synaptic transmission [2, 47, 48]. We have considered how synaptic learning rules can lead to sparse connectivity and thus to energy efficient computation. We studied a one-class perceptron problem in which a neuron learns from positive examples only. One-class learning is relevant for learning paradigms such as recognition and reinforcement learning. One-class learning is also well-known in machine learning [24, 49, 50]. The two-class perceptron requires sampling the space of ‘negative’ patterns that is necessarily large under a sparse firing constraint [22] and secondly, it requires reversing plasticity (‘unlearning’) whenever appropriate. For instance, it is unclear how can a pattern be actively unlearned under spike-timing-dependent plasticity [51]. In contrast to two-class perceptrons, negative samples in the one-class perceptron do not cause plasticity which leads to further energy saving as plasticity itself is an energetically costly process [52].
We imbalance potentiation and depression to achieve sparse connectivity. In other memory tasks, the information loss can be substantial for imbalanced plasticity; for instance, postsynaptic-independent (i.e., without a stop-learning mechanism) online learning rules are severely affected when depression does not match potentiation [17–19]. However, here imbalance leads to a substantial energy reduction in storage as long as the task is below maximal capacity. Furthermore, it is robust against noise and correlated patterns. Imbalanced plasticity is not only a local and biophysically plausible mechanism, but it is also theoretically well-grounded, as it implements L1-norm regularisation, which is well-known to induce sparseness [27, 33, 34, 53]. Due to the biased drift towards zero in the learning rule, the probability of finding silent synapses is increased. Our learning rule reaches high information using a novel, biologically-plausible adaptive threshold without the need for an inhibitory partner neuron. The learning rule is unlike a previous approach to achieve sparse connectivity in which a pruning procedure removes the weakest synapses after learning [14, 15]. Such strategy can lead to as much weight sparseness as desired, but a significant drop in information and efficiency occurs.
Despite the large efficiency gain found, it should be noted that imbalanced plasticity probably does not maximise the efficiency fully. In the limit of many synapses the replica technique from statistical mechanics can provide an estimate on the minimal number of synapses required for a given performance. Extrapolation of such an analysis of the traditional perceptron without sign constraints [10], suggests that even more efficient solutions exist, although it is unclear how to obtain them via online learning. Unfortunately, the weight configuration that truly maximises memory efficiency requires resorting to an impractical and unbiological exhaustive search method, with a search time exponential in the number of synapses. A feasible alternative is to use greedy L0-norm minimisation methods [54], that are in general not guaranteed to achieve the theoretical limiting weight sparseness. Preliminary simulations suggest that the efficiency in this case is not substantially higher than when minimising the linear norm, as the increased number of zero-weight synapses is offset by a steep loss in information.
We note that sparse network connectivity can arise even when energy efficiency is not explicitly optimised for. Weight sparseness also emerges when maximising the information output of a sign-constrained classifier that is required to operate in the presence of postsynaptic noise [28, 30]. The reported weight distribution displays a large fraction of silent synapses [28]. In that learning setup, depression occurs for negative examples to drive the postsynaptic potential well below threshold and thus ensures that the activity of the neuron is suppressed even if noise is present.
In order to implement imbalanced learning various ingredients are needed. 1) As in the classical perceptron a stop-learning condition needs to be implemented. While in the cerebellum the complex spike might fulfil this role, neuromodulatory systems have also been suggested [31]. 2) The balance parameter needs to be precisely set to obtain the most efficient solution and its value depends on the task to be learned. A conservative imbalance setting will increase efficiency, but not as much. We note that the need for precisely tuned parameters is common in this type of studies, just like the standard perceptron requires a precise balance between potentiation and depression, which is also not trivially achieved biologically. 3) For one-class learning, plasticity only occurs when the neural output should be high but it is not (which contrasts the model in [28], where plasticity only occurs when the input is high). A separate supervisory input to the neuron could achieve this. Nevertheless, despite the details of this particular study the general imbalancing principle could well carry over to other systems. In particular including precise spike-timing perceptron learning [55, 56], or temporal STDP [57]. In the latter case, interestingly, energy constraints have also been used to define unsupervised learning rules.
Our study is agnostic about the precise mechanism of pruning. There is a number of biophysical ways a synapse can be inactivated [58, 59]: 1) The presynaptic neuron releases neurotransmitter, but no receptors are present (postsynaptically silent synapse). 2) Alternatively, presynaptic release is turned off (mute synapses). Finally, 3) the synapse is anatomically pruned and thus absent altogether (although it could be recruited again [60]). The first and second would presumably allow the system to rapidly re-recruit the synapse, while the third option not only saves energy, but also reduces anatomical wiring length and volume.
It is worthwhile to ask if our model is consistent with neuroscience data. Naively, one might think that imbalance would predict that LTD would be stronger than LTP, which would contradict typical experimental findings. However, for sparse patterns LTD has to be weakened to prevent saturation, so that the imbalance condition becomes f ⋅ LTP < (1−f) ⋅ LTD. It is unclear whether this condition is fulfilled in biology. Next, one could expect that the theory would predict a net decrease of synaptic strength during learning. However, this is not the case: after all, in the simulations all weights are zero initially, so that synaptic weights can only grow during learning. The reason for this apparent paradox is that learning is gated, unlike unsupervised learning, so the number of LTP and LTD events on a synapse does not necessarily match. While our findings also hold when we start from random weights, there is no obvious initial value for biological synaptic weights.
Finally, one can compare the resulting weight distributions and the number of silent synapses to the data. An advantage of the cerebellum is that also the fraction of zero-weight synapses is known, which is not the case for other brain regions. The weight distribution in the cerebellum matches theory very well when the capacity of a two-class perceptron is maximised in the presence of noise. The fraction of silent synapses exhibits a strong dependence on the required noise tolerance; it is significantly decreased in the low noise limit [28]. Our current model finds a similar distribution from a very different objective function, namely minimising the energy of a one-class perceptron. Which of these two is the appropriate objective for the cerebellum or other brain regions remains a question for future research.
Provided that the memory problem is realisable, perceptron learning ensures that each of the K patterns leads to postsynaptic firing activity (h ≥ 0), i.e., the FN error probability is zero, p10 = 0. In this case the information increases as the FP error probability p01 decreases (see main text, Eq 7). With the additional assumption that the lures are uncorrelated to the learned patterns, it can be shown that our perceptron learning rule leads to a decrease of the FP error.
To see why, we write p01 as a function of the learned synaptic weights. As the lure patterns are uncorrelated to the learned ones, each input xi will be uncorrelated to its corresponding weight wi. The total synaptic current is given by a sum of many terms. Assuming that there are sufficient non-zero weights, the probability distribution p(hl) of the net input hl in response to a lure is Gaussian, h l ∼ 𝓝 ( ⟨ h l ⟩ , ⟨ δ h l 2 ⟩ ). Under this approximation,
p 01 ≈ ∫ 0 ∞ d h l p ( h l ) = 1 2 erfc ( - ⟨ h l ⟩ 2 ⟨ δ h l 2 ⟩ ) , (12)
where erfc ( x ) = 2 π ∫ x ∞ e − t 2 d t is the complementary error function. The approximation improves as N → ∞, as the fraction of non-zero synapses F remains finite irrespective of the imbalance λ (for λ ≤ λmax) and as long as the memory load α does not vanish [10].
As the inputs are in zero-mean bipolar form, ⟨x⟩ = 0, ⟨x2⟩ = 1. The mean current elicited by lures is just ⟨ h l ⟩ = N ⟨ x ⟩ ⟨ w ⟩ − θ N = − θ N, independent of the weights. The variance in the current
⟨ δ h l 2 ⟩ = ⟨ ( δ ( h l + θ N ) ) 2 ⟩ = N ( ⟨ x 2 ⟩ ⟨ w 2 ⟩ - ( ⟨ x ⟩ ⟨ w ⟩ ) 2 ) = N ⟨ w 2 ⟩ (13)
is proportional to the second raw moment ⟨ w 2 ⟩ = ∫ 0 ∞ d w p ( w ) w 2 of the weight distribution. For a particular realisation of the system one has N ⟨ w 2 ⟩ = ∣ ∣ w ∣ ∣ 2 2, the squared Euclidean norm of the synaptic weight vector. This is illustrated in Fig 8. The information of the system is thus given by the Euclidean norm of the weight vector alone. This is true for the learned-vs-lure discrimination task as long as the Gaussianity of the lure current hl holds, irrespective of the particular learning rule that is employed. For instance, p01 takes the same form for postsynaptic-independent learning [19] or for rate-coded inputs that are learned via the offline pseudo-inverse rule [22].
Thus, the perceptron with the most information satisfies the firing condition h ≥ 0 for every learned pattern, but has a minimal Euclidean length weight vector. This coincides exactly with the traditional perceptron that is optimal with respect to the maximal-stability criterion [39], that prescribes the weight configuration with largest stability Δ ≡ θ N / ∣ ∣ w ∣ ∣ 2. This is a widely used principle that enlarges the basins of attraction in recurrent networks and improves the ability to generalise in classifiers [39, 61]. Notice that for a fixed threshold, increasing Δ can only increase information, as it is inversely proportional to the Euclidean weight vector length. Information maximisation thus reveals an interesting close link between recognition memory and the more usual two-class learning problems.
Furthermore, at least for random patterns, we can expect the perceptron learning rule to perform well. Below the critical load αmax the algorithm is known to converge to solutions with large Δ [62]. In other words, although the learning dynamics is not guaranteed to maximise information, it should achieve high C in the recognition task. As shown in the main text, Fig 2, the improvement indeed is minimal when the full quadratic program is actually solved.
The crucial condition that must be observed to achieve good performance is that the firing threshold θ should be large. Here θ plays the role of an indirect (unnormalised) stability parameter. It can be shown [39] that raising θ will indirectly lead to solutions with larger Δ. Lower bounds on how close the learning rule gets to maximal stability with a certain setting of θ and a, b can be extracted from the perceptron convergence proof [39].
Note that the above reasoning requires zero-mean inputs and balanced plasticity. For 0 or 1 inputs, the distribution of the unthresholded output hl that is obtained in response to lures is still well characterised by a Gaussian, as an uncorrelated input pattern gives a sum over on average fN randomly selected weights. The expressions for the mean ⟨hl⟩ and the variance ⟨ δ h l 2 ⟩ now include terms that depend on first- and second-order moments of the synaptic weight distribution. For a particular realisation of the random system the mean is ⟨ h l ⟩ = f N ⟨ w ⟩ − θ N = f ∣ w ∣ − θ N, and the variance 〈δhl2〉=N( f〈w2 〉)−f2 〈w〉2=f‖w‖22−f2N −1∣w∣2. Thus, when the inputs are in 0 or 1 form, the information per synapse C is no longer a simple function of the squared Euclidean norm as before. The output error probability p01, and therefore the information, is affected by the coding level f and the linear norm ∣w∣ as well.
To gain further insight on the effects of allowing a depression-potentiation imbalance, we prove the convergence of perceptron learning rule Eq 3 for non-zero λ, a variation of the detailed proof given by [29]. Besides the inclusion of the parameter λ, differences arise because our inputs are in bipolar form and because all patterns should elicit a high output.
We study a problem that can provably be solved in a finite number of learning steps by balanced postsynaptic-dependent learning (λ = 0). Therefore we can assume the existence of a weight configuration w* that solves the recognition task
∑ i = 1 N w i * x i k - ( θ + κ ) N ≥ 0 , k = 1 , … , K , (14)
while simultaneously satisfying the N non-negativity constraints w i * ≥ 0, i = 1, …, N. The variable κ ≥ 0 relates the threshold ( θ + κ ) N of the solution to the threshold θ N that is used in the learning algorithm.
We assume that initially all synapses are silent, i.e., we start from the tabula rasa condition wi = 0, i = 1, …, N. Learning proceeds by presenting patterns in random order. Since plasticity only occurs when the postsynaptic current h = ∑ i = 1 N w i x i − θ N is not large enough to activate the perceptron, we index time with m = 1, …, M, m being incremented only when h < 0. Whenever each synapse wi changes, it does so according to the update, Eq 3 Δ w i ( m ) = max { - w i ( m ) , ϵ η i ( m ) } , (15)
where ηi(m) = xi(m)−λ is the weight update before rectification and x(m) ∈ {x1, …, xK} is the pattern that led to the update at time m.
The analysis is carried out by tracking the quantity
a ( m ) = w * · w ( m ) | | w * | | 2 | | w ( m ) | | 2 (16)
over time. If we find that after a finite number of updates a(m) would become larger than one, then the learning process is convergent, as the Cauchy-Schwarz inequality implies that a(m) ≤ 1. To monitor the time evolution of a(m) we bound the scalar product w* ⋅ w(m) from below and the norm ∣∣w(m)∣∣2 from above.
After one update, the change Δ(w* ⋅ w(m)) ≡ w* ⋅ w(m+1)−w* ⋅ w(m) in the scalar product is
Δ ( w * · w ( m ) ) = w * · Δ w ( m ) = ϵ w * · η ( m ) + ∑ i ∈ B ( m ) w i * ( ϵ + ϵ λ - w i ( m ) ) = ϵ w * · x ( m ) - ϵ λ | w * | + ∑ i ∈ B ( m ) w i * ( ϵ + ϵ λ - w i ( m ) ) > ϵ θ N + ϵ κ N - ϵ λ | w * | + ∑ i ∈ B ( m ) w i * ( ϵ + ϵ λ - w i ( m ) ) , (17)
where B(m) = {i: wi(m) < ε+ελ ∧ xi(m) = −1, i = 1, …, N} is the set of all synapses that are set to zero due to the lower bound. Note that the lower bound can only be triggered by depression, which in turn can only occur for low inputs. The inequality is obtained by plugging in the definition Eq (14) of w*.
A bound on the scalar product w* ⋅ w(m) itself after m such updates can then be obtained by iteratively applying Eq (17):
w * · w ( m ) > ϵ m N ( θ + κ - λ N | w * | ) + ∑ l = 1 m ∑ i ∈ B ( l ) w i * ( ϵ + ϵ λ - w i ( l ) ) . (18)
Meanwhile, the change Δ ∣ ∣ w ( m ) ∣ ∣ 2 2 ≡ ∣ ∣ w ( m + 1 ) ∣ ∣ 2 2 − ∣ ∣ w ( m ) ∣ ∣ 2 2 in the squared norm of w(m) after one step can be obtained by expanding the square ∣ ∣ w ( m + 1 ) ∣ ∣ 2 2 = ∣ ∣ w ( m ) + Δ w ( m ) ∣ ∣ 2 2, so that
Δ | | w ( m ) | | 2 2 = 2 w ( m ) · Δ w ( m ) + | | Δ w ( m ) | | 2 2 . (19)
We have Δwi(m) ∈ {εηi(m), −wi(m)}, with wi(m) < ε+ελ, as Δwi(m) = −wi(m) only for i ∈ B(m). Thus, the squared norm of the update is dominated by the terms that come from low inputs at synapses that do not cross the lower bound. This gives the inequality
| | Δ w ( m ) | | 2 2 < ϵ 2 N ( 1 + 2 λ q + λ 2 q ) , (20)
where q ≡ max k 1 / N ∑ i = 1 N δ x i k , − 1 denotes the maximum fraction of low inputs observed across the K patterns.
The scalar product is expanded as before:
w ( m ) · Δ w ( m ) = ϵ w ( m ) · η m + ∑ i ∈ B ( m ) w i ( m ) ( ϵ + ϵ λ - w i ( m ) ) = ϵ w ( m ) · x ( m ) - ϵ λ | w ( m ) | + ∑ i ∈ B ( m ) w i ( m ) ( ϵ + ϵ λ - w i ( m ) ) < ϵ w ( m ) · x ( m ) + ∑ i ∈ B ( m ) w i ( m ) ( ϵ + ϵ λ - w i ( m ) ) . (21)
Note that the update condition h < 0 is always satisfied at time m, so that ε w ( m ) ⋅ x ( m ) < ε θ N. Together with the bound Eq (20), iterating over Eq (19) gives
| | w ( m ) | | 2 2 < ϵ m N ( 2 θ N + ϵ ( 1 + 2 λ q + λ 2 q ) ) + 2 ∑ l = 1 m ∑ i ∈ B ( l ) w i ( l ) ( ϵ + ϵ λ - w i ( l ) ) (22) < ϵ m N ( 2 θ N + ϵ ( 1 + 2 λ q + λ 2 q ) + 2 q ϵ ( 1 + λ ) 2 ) . (23)
The last inequality is obtained by noticing that wi(l) < ε+ελ inside the sum over l; the factor q arises from the iteration over the N synapses, conditioning on the low inputs. The bound Eq (23) implies that as learning proceeds ∣∣w(m)∣∣2 cannot grow faster than m.
From Eq (22) we collect
ϵ m θ N > - 1 2 ϵ 2 m N ( 1 + 2 λ q + λ 2 q ) - ∑ l = 1 m ∑ i ∈ B ( l ) w i ( l ) ( ϵ + ϵ λ - w i ( l ) ) . (24)
Turning back to Eq (18) and using the previous result Eq (24) yields
w * · w > ϵ m N ( κ N - 1 2 ϵ ( 1 + 2 λ q + λ 2 q ) - λ N | w * | ) + ∑ l = 1 m ∑ i ∈ B ( l ) ( w i * - w i ( l ) ) ( ϵ + ϵ λ - w i ( l ) ) > ϵ m N ( κ N - 1 2 ϵ ( 1 + 2 λ q + λ 2 q ) - λ N | w * | - q ϵ ( 1 + λ ) 2 ) . (25)
The last inequality stems from wi(l) < ε+ελ. The first bracketed factor is always larger than −(ε+ελ), while the second one is bounded from above by ε+ελ. Iterating over the constrained sum introduces the factor Nq as before.
We now have a bound for the cosine a(m). Substituting in Eqs (23) and (25) gives
a(m)>ϵmN[ κN−1/2−λN−1|w*|−12ϵ(1+2λq+λ2q)−qϵ(1+λ)2 ]||w*||22θN−1/2+ϵ(1+2λq+λ2q)+2qϵ(1+λ)2. (26)
Note that while the neural parameters {ε, θ, λ} can be set at will, for a certain task the solution margin κ and the norms are constrained by the existence of a vector w* that can satisfy the learning conditions. Thus, they cannot be varied arbitrarily. In fact, if one keeps ∣∣w*∣∣2 fixed, it will only be possible to increase κ up to a certain point, where we will have found the maximally-stable configuration. Similarly, the linear norm ∣w*∣ will have a minimum value. Furthermore, in general it is not possible to achieve simultaneously minimal ∣w*∣ and maximal κ with a single configuration.
From Eq (26) a number of conclusions can be drawn. The straightforward condition for convergence is to check whether that bound becomes larger than one. Another way to show that the learning algorithm stops is to check if a(m) is a monotonically increasing function of m. When λ = 0, the process is convergent, as long as ε ≤ 2 κ / [ N ( 1 + 2 q ) ]. For λ > 0, the crucial observation is that we can only show that learning converges if κ can be raised so as to compensate for the negative terms in the numerator.
Thus, as expected, we find that the imbalance λ is related to the linear norm of the solution vector (one can increase λ as ∣w*∣ can be made smaller), and to the occurrence of depression events (through q). But more importantly, λmax writes directly as a function of κ as well, which here sets the task difficulty, since the maximal value for κ shrinks as the memory load α increases. What is more, the minimum of ∣w*∣ depends itself on α. This theoretical prediction is confirmed by our numerical work. As α increases, the achievable imbalance λmax becomes closer to zero, and the fraction of silent synapses approaches that which is obtained with balanced (λ = 0) learning, cf. Fig 2C.
We generate correlated patterns following previous work in recognition memory [21]. In the first model we generate a template pattern x ^ with each input x ^ i being set high (+1) or low (-1) independently and with equal probability 1/2. To maintain balance we also use its negative, − x − x ^, as a template.
The K patterns the neuron should learn are generated conditioned on either template, such that P ( x i k = x ^ i ) = 1 + g 2. Lure patterns follow the statistics of the learned patterns and are produced from the same templates. The parameter g controls the level of input correlations; with the choice g = 0 the original statistics are recovered, while at g = 1 the recognition task is impossible, as all patterns are perfect copies or reversals of one another.
In the second model patterns generated according to the process described above, but only using a single template. This procedure introduces inter-pattern correlations at the same presynaptic site xi, as the arriving patterns become more similar to one another. It also leads to heterogeneous mean activity levels across neurons; although the mean number of active presynaptic neurons per pattern remains 1/2, increasing g leads to a bimodal presynaptic firing distribution. For g > 0, neurons that are active in the template fire more often and, conversely, the remaining neurons fire less frequently.
All our computer simulations were implemented on Matlab R2013a (MathWorks) and were performed on a standard desktop computer. We simulated a single postsynaptic neuron that was driven by N = 1000 presynaptic random inputs. We varied the memory load parameter within the range α ∈ [0.1,0.8] to avoid both the appearance of unsolvable problem instances and excessive simulation time. We chose a small learning rate ε = 1/N and a sufficiently large firing threshold at N (i.e., θ = 1) except when otherwise noted. The threshold was set so that typically no increase in information could be obtained by raising it further. In the figures we included second-degree polynomial fits to average values.
The online perceptron learning rule was iterated until all patterns were learned. To obtain the maximally-imbalanced solution (λ = λmax) we minimised the linear norm ∣w∣ using a linear programming algorithm [38], subject to the set of inequality constraints that ensured that every pattern would lead the neuron to fire. Specifically, using Matlab’s interior-point solver, available via the linprog command (Optimization Toolbox), we minimised ∣w∣ subject to N non-negativity constraints wi ≥ 0 and K linear pattern imprinting constraints specified in matrix form as X ⊤ w ≥ θ N 1, where X⊤ is the K × N design matrix whose rows are the positive examples.
For the balanced case, the maximum-information weight configurations were obtained using the Krauth-Mézard min-over algorithm [39], followed by rectification after each learning step in order to enforce the non-negativity synaptic constraints. This is a batch learning algorithm that employs the balanced rule (Eq 3, λ = 0). At each step the pattern x k min with lowest stability, k min = arg min k = 1 K ∑ i = 1 N w i x i k, is determined on the forehand. Then, only x k min is learned; plasticity is silenced for all other patterns. To confirm optimality and validate our mathematical results we also resorted to an interior-point convex optimiser [38] and solved the quadratic programming problem of finding the weight vector with minimal Euclidean norm ∣∣w||2. We resorted to Matlab’s quadprog command (Optimization Toolbox) to minimise ∣ ∣ w ∣ ∣ 2 2 subject to the same N non-negativity and the K pattern imprinting constraints imposed on the linear program. Up to numerical precision the obtained pattern stabilities Δ matched those given by the min-over algorithm.
To calculate the information Eq (7) we tested the neuron with a set of K lures generated with the same statistics as the K learned patterns and recorded the number of FP errors. To determine the fraction of silent synapses, one has to take care of numerical rounding errors as it might be unclear when a synapse can truly be considered zero. We removed the weakest synapses one by one while probing the neuron with a large number of lures, until a drop in information occurred. With this procedure we could distinguish the true zero-weight synapses from small ones while avoiding numerical precision issues and arbitrary threshold setting. The results did not qualitatively change if we simply counted the number of synapses below some small weight w zero ≪ max i = 1 N w i, held constant across trials.
Since we expected self-averaging of the synaptic weights distribution from the validity of the replica trick [7], the averaged synaptic weight histograms were collected from 1000 trials. To set a common weight scale across different learning rules and input statistics, we normalised the synaptic weights so that the threshold became unity, i.e., we re-scaled the weights by a factor w i / min k = 1 K ∑ i = 1 N x i k w i.
|
10.1371/journal.ppat.1001202 | Autoacetylation of the Ralstonia solanacearum Effector PopP2 Targets a Lysine Residue Essential for RRS1-R-Mediated Immunity in Arabidopsis | Type III effector proteins from bacterial pathogens manipulate components of host immunity to suppress defence responses and promote pathogen development. In plants, host proteins targeted by some effectors called avirulence proteins are surveyed by plant disease resistance proteins referred to as “guards”. The Ralstonia solanacearum effector protein PopP2 triggers immunity in Arabidopsis following its perception by the RRS1-R resistance protein. Here, we show that PopP2 interacts with RRS1-R in the nucleus of living plant cells. PopP2 belongs to the YopJ-like family of cysteine proteases, which share a conserved catalytic triad that includes a highly conserved cysteine residue. The catalytic cysteine mutant PopP2-C321A is impaired in its avirulence activity although it is still able to interact with RRS1-R. In addition, PopP2 prevents proteasomal degradation of RRS1-R, independent of the presence of an integral PopP2 catalytic core. A liquid chromatography/tandem mass spectrometry analysis showed that PopP2 displays acetyl-transferase activity leading to its autoacetylation on a particular lysine residue, which is well conserved among all members of the YopJ family. These data suggest that this lysine residue may correspond to a key binding site for acetyl-coenzyme A required for protein activity. Indeed, mutation of this lysine in PopP2 abolishes RRS1-R-mediated immunity. In agreement with the guard hypothesis, our results favour the idea that activation of the plant immune response by RRS1-R depends not only on the physical interaction between the two proteins but also on its perception of PopP2 enzymatic activity.
| Plant and animal bacterial pathogens have evolved to produce virulence factors, called type III effectors, which are injected into host cells to suppress host defences and provide an environment beneficial for pathogen growth. Type III effectors from pathogenic bacteria display enzymatic activities, often mimicking an endogenous eukaryotic activity, to target host signalling pathways. Elucidation of strategies used by pathogens to manipulate host protein activities is a subject of fundamental interest in pathology. PopP2 is a YopJ-like effector from the soil borne root pathogen Ralstonia solanacearum. Here, in addition to demonstrating PopP2 ability to stabilize the expression of its cognate Arabidopsis RRS1-R resistance protein and physically interact with it, we investigated the enzymatic activity of PopP2. Bacterial YopJ-like effectors are predicted to act as acetyl-transferases on host components. However, only two YopJ-like proteins from animal pathogens have been shown to be active acetyl-transferases. We show that PopP2 displays autoacetyl-transferase activity targeting a lysine residue well-conserved among YopJ-like family members. This lysine is a critical residue since its mutation prevents autoacetylation of PopP2 and abolishes its recognition by the host. This study provides new clues on the multiple properties displayed by bacterial type III effectors that may be used to target defense-related host components.
| To defend themselves against pathogen infection, plants have evolved a bipartite, inducible innate immune system that efficiently recognizes and wards off pathogens [1]–[3]. Perception of conserved molecules, essential to many pathogens and called pathogen-associated molecular patterns (PAMPs), triggers the so-called PTI (PAMP-Triggered Immunity), which represents the first layer of host defence [4]. The recognition of different PAMPs occurs by specific pattern recognition receptors (PRRs) acting at the plant cell surface to activate various defence responses in the host [1], [5], [6]. Plants can also perceive specific effectors [previously referred to as avirulence (Avr) proteins] through additional receptors –typically nucleotide binding site-leucine rich repeat (NB-LRR) resistance (R) proteins. This second layer of defence is called effector-triggered immunity (ETI). The simplest model to explain perception of a given Avr protein by the corresponding R protein is that they physically interact as a ligand-receptor couple, leading to activation of immunity. Nevertheless, direct interaction between Avr and R proteins has been only described in a few cases indicating that this type of recognition is an exception rather than the rule [7]–[10]. To explain the lack of evidence for direct interaction for most of known Avr-R pairs, the guard model was proposed [11], [12]. This model, validated in the case of several R-Avr interactions, postulates that R proteins guard effector targets or “guardees” from effector-triggered manipulation in host cells [13]–[16]. Several effector targets have been identified and shown to play crucial a role in the establishment of plant resistance [17], [18].
The elucidation of the function of type III effectors (T3Es), and more specifically of their enzymatic activities perceived by guard proteins, remains a major challenge in phytopathology. One of the bacterial strategies to suppress host innate immunity consists in the manipulation or inactivation of plant components playing a role in defence-related signalling pathways [19]. Bacterial effectors exhibit diverse activities that mimic eukaryotic functions involved in defence against infection. Known biochemical activities of T3Es include manipulation of host protein turnover, either through protease activity [14], [20], [21] or protein degradation via the 26S proteasome [22]–[24], modification of host transcription or RNA stability [25]–[28] and alteration of the phosphorylation state of plant proteins [29]–[34].
The RRS1-R resistance gene, present in Arabidopsis thaliana plants of the Nd-1 ecotype, confers broad-spectrum resistance to several strains of Ralstonia solanacearum. R. solanacearum is the causal agent of bacterial wilt in more than 200 plant species, including agronomically important crop plants of the Solanaceous family [35]. RRS1-R encodes an R protein with original structure since it belongs to the Toll/Interleukin1 receptor (TIR)-NBS-LRR subclass of R proteins and presents a C-terminal WRKY motif that is characteristic of the zinc-finger class of WKRY plant transcription factors [36]. Although genetically defined as recessive, RRS1-R behaves as a dominant gene in transgenic Arabidopsis plants. The dominant RRS1-S allele from the Col-0 susceptible accession encodes a highly similar TIR-NBS-LRR-WRKY protein lacking the last 96 amino acids of the RRS1-R protein [37], [38].
Among the more than 80 putative type three effectors (T3Es) encoded by R. solanacearum, PopP2 elicits RRS1-R-mediated specific disease resistance in Arabidopsis [9]. PopP2 belongs to the YopJ-like family of effectors. YopJ-like proteins are found in mammalian and plant pathogens, suggesting that they play important roles in the interaction with the host. Based on their structural characteristics, YopJ-like effectors have been assigned to the C55 peptidase family of the clan CE of cysteine proteases, which share a nucleophile cysteine and a predicted catalytic core composed of three conserved amino acid residues (Figure 1A) [39]–[41]. YopJ-like family members, such as AvrRxv and AvrXv4 from Xanthomonas campestris pv. vesicatoria, are SUMO (Small Ubiquitin-like MOdifier) isopeptidases [42]. It was proposed that removal of SUMO from host plant proteins would allow binding of ubiquitin leading to protein degradation. The YopJ protein from Yersinia spp displays de-ubiquitinating, de-SUMOylating and acetylating activities. YopJ-mediated acetylation of critical serine and threonine residues in target MAPKs, which are essential for the host cell inflammatory response, blocks their activation by phosphorylation [43]–[45].
Co-expression of PopP2 and either RRS1-R or RRS1-S in Arabidopsis protoplasts previously revealed that RRS1 proteins are specifically targeted to the plant cell nucleus. The interaction between PopP2 and RRS1-R/S was also demonstrated in yeast [9]. However, the enzymatic activity of PopP2 and its mode of action remain unknown. In this study, using the FRET-FLIM technique, the physical interaction between PopP2 and RRS1-R was shown to occur in the nucleus of N. benthamiana and Arabidopsis epidermal cells. In addition, we show that PopP2 stabilizes the expression of RRS1-R in planta. Similar results were obtained with the dominant RRS1-S gene product. Conservation of the YopJ catalytic core within the PopP2 sequence prompted us to test the hypothesis that PopP2 displays acetyl-transferase activity. We show that PopP2 autoacetylates on a conserved lysine residue essential for RRS1-R-mediated immunity in Arabidopsis.
A conserved catalytic triad is present in representatives of the clan CE of cysteine proteases, including the T3Es YopJ from Y. pestis and PopP2 from R. solanacearum (Figure 1A). If PopP2 avirulence function was dependent on its enzymatic activity, mutation of cysteine 321 in the predicted catalytic core to a non-reactive residue such as alanine (PopP2-C321A) should abolish PopP2 ability to trigger RRS1-R-mediated resistance in Arabidopsis. To test this hypothesis, a PopP2-C321A mutant version was engineered. PopP2 and PopP2-C321A were next tagged at their C-terminus with a triple hemagglutinin (3x-HA) epitope, and expressed under the control of the native popP2 promoter. These constructs were introduced in a previously described strain of R. solanacearum (ΔpopP2) from which the popP2 gene had been deleted, and therefore no longer recognized by Arabidopsis Nd-1 plants carrying the RRS1-R gene [9]. Nd-1 Arabidopsis plants were inoculated with the ΔpopP2 strain expressing either PopP2, PopP2-C321A, or a GUS control. As a consequence of the recognition of PopP2 by RRS1-R, Nd-1 plants remained symptomless 8 days after inoculation with ΔpopP2 carrying wild-type popP2, whereas complete wilting was observed after inoculation with the two ΔpopP2 strains expressing either mutant PopP2-C321A or the GUS control (Figure 1B,C). These data demonstrate that mutation of the conserved putative catalytic C321 in PopP2 abolishes its ability to trigger an RRS1-R-specific resistance response in Arabidopsis.
PopP2 is targeted to the plant cell nucleus where it is predicted to physically interact with RRS1-R, as indicated by previously published yeast two-hybrid data [9]. The inability of mutant PopP2-C321A to initiate the RRS1-R resistance response is consistent with the hypothesis that PopP2 functions as a cysteine protease and that its enzymatic activity is perceived by RRS1-R. Alternatively, we cannot exclude that the C321A mutation leads to a conformational change in PopP2 that may impair its nuclear targeting and/or its ability to interact with RRS1-R. Agrobacterium-mediated transient expression of PopP2 and PopP2-C321A fused to CFP (Cyan Fluorescent Protein), under the control of a 35S promoter, led to detection of a fluorescent signal in the nucleus of both Arabidopsis and N. benthamiana epidermal cells (Figure 2), demonstrating that mutation of the conserved cysteine residue in the catalytic triad of PopP2 does not affect its nuclear targeting.
A quantitative non-invasive fluorescence lifetime imaging (FLIM) approach was then used to monitor the physical interaction between RRS1-R and PopP2. Expression of RRS1-R fused to YFPv (Yellow Fluorescent Protein venus) was not detectable in Arabidopsis cells whereas a weak fluorescent signal was observed within the nuclei of N. benthamiana cells after transient expression using Agrobacterium (Figure 3A). Therefore, FRET-FLIM studies were conducted by co-expressing PopP2, or PopP2-C321A, fused to CFP, together with RRS1-R-YFPv in N. benthamiana. This transient expression system was previously used to demonstrate the physical interaction between PopP2 and RD19, a vacuolar Arabidopsis cysteine protease that is relocalized to the nucleus in the presence of PopP2 [46]. The average CFP lifetime in nuclei expressing PopP2-CFP was 2.386±0.030 ns (mean ± SEM). A significant reduction of the average CFP lifetime to 1.979±0.024 ns (p-value = 6.4×10−21) was measured in nuclei co-expressing the PopP2-CFP and RRS1-R-YFPv fusion proteins (Figure 3C,4A; Table 1), showing that RRS1-R is able to interact with PopP2 in the nucleus. Mutant PopP2-C321A also interacts with RRS1-R, as shown by a significant reduction of the average CFP lifetime in nuclei co-expressing PopP2-C321A-CFP and RRS1-R-YFPv, as compared to nuclei expressing PopP2-C321A-CFP alone (Figure 3D,4B; Table 1). This result indicates that the nuclear interaction between PopP2 and RRS1-R does not depend on the presence of the conserved cysteine residue in the catalytic triad of PopP2. To confirm that reduction of PopP2-CFP lifetime, in the presence of RRS1-R-YFPv, is not due to non-specific transfer of energy between the two fluorophores, PopP2-CFP was co-expressed with either untagged RRS1-R or YFPv. No physical interaction could be detected in either case, as shown by an average CFP lifetime which is not significantly different from that of PopP2-CFP alone (Figure 4A and Table 1). Importantly, the interaction between PopP2 and RRS1-R was confirmed in Arabidopsis cells after transient co-expression of both proteins (Table 1). Taken together, these data demonstrate a specific interaction between PopP2 and RRS1-R in planta, and strongly suggest that this protein association is independent of the enzymatic activity of PopP2.
Detection of the RRS1-R/PopP2 interaction prompted us to check whether the RRS1-S protein, present in susceptible Col-0 plants, was also able to associate with PopP2. RRS1-S was previously shown to colocalize with PopP2 in the nucleus of Arabidopsis protoplasts and the interaction between the two proteins was demonstrated in yeast [9]. Lack of in planta association between these two proteins might explain why RRS1-S is not able to trigger the resistance response [38]. To test this idea, RRS1-S-YFPv was co-expressed with PopP2-CFP in N. benthamiana epidermal cells. Reduction of the average CFP lifetime to 1.889±0.031 ns (p-value = 1.8×10−22) was detected in nuclei co-expressing PopP2-CFP and RRS1-S-YFPv fusion proteins (Figure 3E,4A and Table 1), demonstrating that, similarly to RRS1-R, RRS1-S associates with PopP2 in the nucleus. This result confirms the previously published RRS1-S/PopP2 interaction in yeast cells and indicates that the susceptibility of Col-0 plants to the Ralstonia GMI1000 strain cannot be explained by the lack of interaction between RRS1-S and PopP2. Taken together, our data strongly suggest that activation of the RRS1-R-mediated resistance response requires not only its physical interaction with PopP2, but also perception of PopP2 enzymatic activity.
Protein co-expression experiments in planta indicated that PopP2 may be able to increase the accumulation level of YFPv-tagged RRS1-R and RRS1-S (Figure 3A,C and 3B,E). This observation was further confirmed by transiently co-expressing RRS1-R or RRS1-S, tagged at their C-terminus with a triple Flag epitope (3x-Flag), and PopP2 tagged at its C-terminus with a 3x-HA epitope in N. benthamiana. Immunoblot analysis of total protein extracts showed higher accumulation of RRS1-R and RRS1-S in the presence of PopP2 (Figure 5A). Similar results were obtained with the PopP2 catalytic mutant C321A, suggesting that the ability of PopP2 to enhance in planta accumulation of RRS1 proteins is not dependent on its enzymatic activity.
To investigate the specificity of this observation, we next tested whether PopP2 may also promote stabilization of the previously described PopP2 interacting partner, RD19. No modification of the RD19 protein level was detected after co-expression with PopP2 (Figure S1), indicating that protein stabilization by PopP2 may be restricted to a subset of its interacting partners.
We next studied whether the turnover of RRS1 proteins is post-translationally regulated through proteasomal activity. In the presence of the proteasome inhibitor MG132, both RRS1-R and RRS1-S accumulated, whereas expression of a GUS control was not altered in the same conditions (Figure 5B).
Taken together, our data (i) show that PopP2 is able to specifically promote the accumulation of its cognate partners, RRS1-S and RRS1-R, regardless of the integrity of its catalytic triad, and (ii) strongly suggest that the interaction between PopP2 and RRS1-R/S may block a molecular mechanism that leads to RRS1-R/S proteasome-dependent degradation.
The observation that PopP2-C321A is affected in its avirulence function strongly suggests that perception of the enzymatic activity of PopP2 is required to trigger RRS1-R-mediated resistance. YopJ and VopA are active acetyl-transferases previously described to be able to autoacetylate in vitro [43], [47]. We thus investigated whether PopP2, being a member of the YopJ-like family of effectors, exhibits a similar enzymatic activity. Towards this goal, GST-tagged PopP2 and PopP2-C321A were purified from overexpressing bacteria and subjected to liquid chromatography/tandem mass spectrometry (LC-MS/MS) to assess their acetylated states. Analysis of relative tryptic digestions revealed the presence of four acetylated lysine residues (at positions 268, 285, 335 and 383) in GST-PopP2 (Figure 6A, Table S1 and S2). However, acetylation of lysines 268, 285 and 335 is probably the result of endogenous E. coli acetyl-transferase activity, since it was also detected in mutant GST-PopP2-C321A, whereas acetylation of K383 is dependent on the presence of an integral PopP2 catalytic core. Together, these data strongly suggest that PopP2 is an active acetyl-transferase that autoacetylates at its K383 residue.
The differential acetylation status of GST-PopP2 and GST-PopP2-C321A was further investigated by immunoblot analysis using an antibody directed against acetylated lysine residues (α-Ac-K). This antibody allowed the detection of a strong signal corresponding to acetylated GST-PopP2 (Figure 6B, lane 1). In contrast, a barely detectable signal was observed in the case of GST-PopP2-C321A (Figure 6B, lane 2), perhaps due to weak acetylation at residues K268, K285, and K335 residues. Lysine to arginine substitutions were next engineered to generate GST-tagged PopP2-K268R, PopP2-K285R, PopP2-K335R and PopP2-K383R mutant versions. Importantly, PopP2-K268R, PopP2-K285R and PopP2-K335R, but not PopP2-K383R, were detected by immunoblot using the α-Ac-K antibody (Figure 6B), strongly indicating that only the K383R mutation prevents PopP2 autoacetylation (Figure 6B, lane 6). Furthermore, as in the case of the inactive GST-PopP2-C321A mutant, LC-MS/MS analysis of purified GST-PopP2-K383R led to the detection of a set of three acetylated peptides containing K268, K285 and K335 (Table S3). Together, these data confirm that (i) acetylation of K268, K285 and K335 is most likely due to an endogenous acetyl-transferase activity from E. coli and that (ii) K383 is the amino acid residue specifically targeted by the auto-acetylation activity of PopP2.
In order to obtain further proof that acetylation of PopP2 at K383 is due to PopP2 autoacetyl-transferase activity, rather than acetylation by an endogenous E. coli acetyl-tranferase unable to target the PopP2-C321A or PopP2-K383 mutants, we next checked whether PopP2 is able to autoacetylate in trans. If this is the case, PopP2 should be able to acetylate the catalytically inactive PopP2-C321A mutant. To be able to distinguish between PopP2 and PopP2-C321A after SDS-PAGE analysis, we used a truncated form of PopP2 that lacks its first 80 amino acids (PopP281–488) but retains its acetyl-transferase activity (Figure 7A, lane 3). Recombinant GST-PopP2 and GST-PopP2-C321A proteins were respectively co-expressed with active GST-PopP281–488 or inactive GST-PopP2-C321A81–488. Purified proteins were subjected to immunoblot analysis using an α-Ac-K antibody. In the presence of active GST-PopP281–488 and GST-PopP2, inactive GST-PopP2-C321A and GST-PopP2-C321A81–488 are respectively acetylated, demonstrating that the C321A mutation does not prevent PopP2 autoacetylation in trans. Although this experimental set up does not allow us to exclude the possibility that PopP2 also autoacetylates in cis, our data clearly demonstrate the intermolecular autoacetylation displayed by PopP2.
We next investigated whether the K383 residue, which is very likely the main acetyl-CoA acceptor site in PopP2, is required for PopP2 trans-autoacetylation activity. First, when active GST-PopP2 was co-expressed with a truncated form of the K383R mutant (GST-PopP281–488-K383R) no GST-PopP281–488-K383R acetylated form could be detected, strongly indicating that K383R mutation prevents its trans-acetylation by active GST-PopP2 (Figure 7B, lane 3). Second, GST-PopP281–488-K383R behaves like inactive GST-PopP281–488-C321A, which is not able to acetylate GST-PopP2-C321A (Figure 7B, lanes 5 and 6, respectively). Thus, despite the integrity of its catalytic triad, GST-PopP281–488-K383R is impaired in its trans-acetylation activity. Together, our data show that K383 in PopP2 represents an acetyl-CoA binding site that is (i) targeted by autoacetylation and (ii) required for intermolecular acetylation of PopP2.
Finally, the protein sequences of members of the YopJ-like family of T3Es were inspected. Figure 8 shows a sequence alignment of several YopJ-like proteins in the neighbouring region of the conserved residue K383 that is targeted by autoacetylation in PopP2. This analysis showed that the K383 residue in PopP2 is perfectly conserved among all members of the family, strongly suggesting that this residue may represent a key acetyl-CoA acceptor site essential for the trans-acetylation activity of PopP2 and other members of the YopJ family.
The effect of the K383 mutation, which compromises PopP2 autoacetylation in E. coli, was further tested after transient expression in N. benthamiana. Expression of a PopP2-K383-CFP fusion was detected in the nucleus of N. benthamiana epidermal cells (Figure S2), confirming that this mutation does not affect nuclear targeting of the protein. In addition, co-expression of RRS1-R-YFPv and PopP2-K383-CFP led to detection of YFP fluorescence within the nucleus indicating that, as wild-type PopP2, PopP2-K383R is able to stabilize RRS1-R expression. Furthermore, significant reduction of the CFP lifetime in FRET-FLIM assays following transient co-expression of PopP2-K383-CFP and RRS1-R-YFPv, as compared to PopP2-K383-CFP expressed alone, demonstrates that PopP2-K383R also interacts with RRS1-R (Figure 4C and Table 1).
In order to investigate whether PopP2 avirulence activity is compromised by mutation of K383, the PopP2-K383R mutant was next used to transform the ΔpopP2 strain of R. solanacearum. First, in vitro secretion and stability of wild-type PopP2, PopP2-C321A and PopP2-K383R were tested by immunoblot analysis. Upon incubation of the different complemented ΔpopP2 strains in secretion medium, all PopP2 variants were detected in total extracts and in culture supernatants (Figure 9A). To ensure that the signal observed from supernatant was T3SS (Type 3 Secretion System)-dependent, popP2 constructs were also introduced into a ΔpopP2/ΔhrcV strain mutated both in popP2 and in hrcV, a gene coding for a conserved inner membrane component of the T3SS [48]. As expected, no signal corresponding to PopP2 was detected in culture supernatants of ΔpopP2/ΔhrcV bacteria cells expressing the various PopP2 variants, demonstrating the proper secretion of PopP2, PopP2-C321A and PopP2-K383R by the T3SS. In addition, to demonstrate that no bacterial lysis had occurred, all protein samples were probed with an antibody directed against the cytoplasmic chaperonin GroEL (Figure 9A).
Growth of ΔpopP2 expressing PopP2, PopP2-C321A or PopP2-K383R was then measured in root-inoculated Nd-1 plants. Between 4 and 7 days after inoculation, ΔpopP2 strains expressing HA-tagged PopP2-C321A or PopP2-K383R reached about ten-fold higher bacterial growth rates than the ΔpopP2 strain expressing HA-tagged wild-type PopP2 (Figure 9B), showing that, as previously shown for PopP2-C321A (Figure 1B,C), PopP2-K383R is not able to trigger the RRS1-R-dependent resistance response. High bacterial growth rates observed after inoculation with the ΔpopP2 strain expressing mutant PopP2-C321A or PopP2-K383R correlated with the complete wilting of Nd-1 plants 7 days after inoculation (Figure 1B,C, Figure S3). As controls, ΔpopP2 strains expressing PopP2-K268R, PopP2-K285R or PopP2-335R were also root-inoculated in Nd-1 plants. Importantly, these PopP2 mutants were able to trigger the RRS1-R resistance response and to multiply to the same extent as wild-type PopP2 expressing strain, demonstrating that these 3 lysine residues are not required for PopP2 avirulence activity.
Together, our data show that substitution of the K383 residue in PopP2 mimics the mutation in catalytic C321, leading to inactivation of the plant immune response. These findings strongly suggest that K383 is an acetyl-CoA-binding site in PopP2, essential for RRS1-R-mediated recognition of PopP2 activity in planta.
Although the characterization of many R-Avr gene pairs is consistent with the gene-for-gene hypothesis [49], the underlying perception mechanism has been subject of debate for many years. Initially, the ligand-receptor model, postulating that products of R genes act as receptors that directly interact with the products of Avr genes [50], was supported by the colocalization of some Avr and R proteins, most of which encode receptor-like proteins carrying Leu-rich repeats (LRRs). However, a direct interaction between R and Avr proteins has been reported only in a few cases [8]–[10], [51], [52]. Our study represents the first report of an Avr protein (PopP2) that interacts with its matching resistance protein (RRS1-R) in living plant cells. These data are in agreement with an earlier report showing that both proteins associate in yeast cells [9]. Here, we demonstrate that both dominant RRS1-S and recessive RRS1-R gene products (i) can associate with PopP2 within the plant nucleus and (ii) are stabilized by PopP2 in this subcellular compartment. Therefore, it is tempting to propose that recessiveness of RRS1-R-mediated resistance may be due to a difference of relative RRS1 binding affinities to PopP2. Comparison of FRET-FLIM data presented in this study is not suitable to address this question since FRET efficiencies reflect only the interaction between two proteins but do not provide quantitative data. In addition, previous data showing that RRS1-R behaves as a dominant resistance gene in Col-0 transgenic plants are not in favour of this model. Our data strongly suggest that functionality of RRS1-R, leading to the activation of the resistance response, requires not only its interaction with PopP2 but also the perception of its enzymatic function. This recognition would be restricted to RRS1-R, the RRS1-S protein being unable to perceive PopP2 acetyl-transferase activity. This hypothesis is strengthened by the fact that PopP2 mutated in the conserved cysteine residue of its catalytic triad is unable to trigger RRS1-R-mediated immunity in Arabidopsis, despite its nuclear localization and its ability to associate with RRS1-R. Although, consistent with the gene-for-gene hypothesis, the recognition of PopP2 by RRS1-R, and not RRS1-S, is also in agreement with the guard model, according to which R proteins detect changes in host proteins after modification by pathogen-derived effectors. Alternatively, rather than a guard, RRS1-R might act as an “enabler” of PopP2 activity. In this model, targeting of host components by PopP2 with the goal of altering plant physiology would be facilitated by the RRS1-R-containing complex, allowing the effector to bind to its substrate(s). However, the outcome of this scenario would be the activation of plant defence, following perception of PopP2 enzymatic activity.
In this study, we show that PopP2 not only interacts with RRS1-R/S but is also able to specifically promote their accumulation in the nucleus. Alteration of the levels of defence-related components is a common strategy used by bacterial pathogens to suppress innate immunity in the host. For instance, AvrPphB papain-like cysteine protease from Pseudomonas syringae targets the Arabidopsis protein kinase PBS1 [20] whose cleavage is perceived by the resistance protein RPS5 to initiate ETI [16]. An additional T3E from P. syringae, AvrRpt2, is a staphopain-like cysteine protease that cleaves various host proteins including RIN4, a negative regulator of PTI, whose modification is recognized by the R protein RPS2 [21], [53], [54]. In addition to direct protein degradation resulting from protease activity, mimicking the host ubiquitination machinery represents an alternative strategy that leads to alteration of protein levels in the host. For example, the AvrPtoB E3 ligase from P. syringae specifically ubiquitinates the tomato protein kinase Fen and promotes its degradation in a proteasome-dependent manner [23].
To the best of our knowledge, PopP2 represents the first example of a T3E from phytopathogenic bacteria that mediates stabilization of the expression of its interacting partner(s) in the host. Previous work showed that YopJ, a well-studied bacterial T3E from the human pathogen Y. pestis, requires its de-ubiquitinating and acetylating activities to stabilize host components [44], [55]. In resting mammalian cells, the dimeric transcription factor NF-κB, involved in the inflammatory response to Yersinia, forms a complex with its inhibitor IκB. Upon its phosphorylation by the IκB kinase (IKK), IκB is ubiquitinated and subsequently degraded. YopJ promotes indirect stabilization of IκB by inhibiting IKK activity through the acetylation of a threonine residue in the activation loop of the kinase [43], thereby blocking the cell inflammatory response. In contrast, perturbation of RRS1-R proteasome-mediated degradation is unlikely to be induced by the acetyl-transferase activity of PopP2, since a catalytically inactive PopP2-C321A mutant is also able to promote accumulation of RRS1-R. Examples of eukaryotic acetyl-transferases that induce protein stabilization independent of their enzymatic activity have been previously reported. For instance, the transcriptional coactivator p300 is a histone acetyl-transferase critical for regulating gene expression in mammalian cells. The activating transcription factor 4 (ATF4) plays a crucial role in multiple stress responses and is stabilized by p300 independent of its acetyl-transferase activity [56]. The molecular mechanism involved in ATF4 stabilization by p300 is still unknown but authors hypothesized that p300, upon its interaction with ATF4, might prevent ATF4 ubiquitination or its targeting to the nuclear proteasome. Similarly, we hypothesize that PopP2, through its physical association with RRS1-R/S, may prevent RRS1-R/S ubiquitination and thus protect them from proteasome-mediated degradation. Interestingly, PopP2-mediated stabilization of RRS1-R and RRS1-S expression appears to be specific, since accumulation of the PopP2-interacting protein RD19 is not affected by PopP2. However, RRS1-R/S stabilization by PopP2 is independent of their nuclear localization. Indeed, previously published data showed that co-expression of RRS1-R/S with a PopP2 derivative deleted from its nuclear localization signal (NLS) (PopP295–488), which is localized both in the nucleus and in the cytosol, leads to their detection in the same subcellular compartments [9]. Since RRS1-R has been shown to act as a negative regulator of plant resistance [57], PopP2-mediated RRS1-R stabilization may represent a bacterial strategy to promote pathogen virulence. Alternatively, RRS1-R stabilization and binding to PopP2 in the nucleus may lead to regulation of its transcriptional activity to trigger plant resistance. Anyhow, as observed in N. benthamiana, both C321A and K383R PopP2 mutant versions may promote RRS1-R accumulation in Arabidopsis, despite the loss of their avirulence activity. In this scenario, PopP2-mediated stabilization of RRS1-R expression would not be sufficient per se for activation of the immune response.
A number of T3ES exert their function by covalently modifying target proteins in the host cell. These covalent modifications are generally reversible and presumably aimed at modulating cellular functions by transiently altering the activity of pathogen cellular targets. In plants, T3Es have been previously found to modulate postranslational modifications (PTMs) in host proteins, including phosphorylation, SUMOylation and ubiquitination [19]. Here, we describe acetylation as a novel PTM displayed by a T3E from a plant pathogenic bacterium. Acetylation is a major PTM first identified on lysine residues from histones [58]. Addition of acetyl groups to the lysine residues of histone tails facilitates access of transcription factors to DNA by disrupting higher-order packaging of the chromatin [59] and also by neutralizing the positive charge of the histone proteins, which reduces the affinity of histones for DNA [60]. Acetylation also impairs the ability of the lysine side chain to form hydrogen bonds thereby enhancing specific or inhibiting non-specific DNA-binding activities of transcription factors [61]–[63]. In addition, acetylation forms docking sites for recruitment of transcriptional coactivators [64]. Several bacterial effectors have been found to affect host transcription. For example, members of the AvrBs3 family from the phytopathogenic bacterium Xanthomonas campestris, also referred to as TAL (Transcription Activator-Like) effectors, are targeted to the plant cell nucleus where they directly induce expression of plant genes [25], [26]. In addition, the SUMO-protease T3E XopD from X. campestris is also targeted to the nucleus of host cells and functions as a transcriptional repressor, resulting in suppression of host defence responses through an unknown molecular mechanism that may involve deSUMOylation of transcription factors or chromatin remodelling, for example [65]. Likewise, we propose that PopP2 autoacetylation and/or acetylation of its interacting partner(s), perhaps RRS1-R, may affect gene transcription in host cells.
Similar to phosphorylation, which is described to target kinases and phosphatases, acetyl-transferases can be acetylated via either intra (cis) or intermolecular (trans) mechanisms. Although it is possible that PopP2 autoacetylation additionally occurs in cis, our data clearly demonstrate the intermolecular autoacetylation of PopP2 on a lysine residue that is well conserved among all members of the YopJ-like family. It is thus tempting to speculate on the biological significance of the acetylation of this particular residue. Indeed, YopJ is predicted to use a two-substrate ping-pong mechanism whereby acetyl-coenzyme A (acetyl-CoA; the first substrate) interacts with the enzyme to form a high-energy acyl-enzyme intermediate, that is attacked by the second substrate resulting in a modified product [44], [45]. A previous study suggested that, in addition to the conserved catalytic core essential for acetyl-transferase activity, each member of the YopJ family, whether from a plant or animal pathogen, presents a conserved site, possibly located at the C terminus, that binds acetyl-CoA [45]. However, the identity of this acetyl-CoA-binding site and the reason(s) behind its conservation within the YopJ protein family are still unknown. The observations that (i) PopP2 autoacetylates in K383, a key residue that is required for its intermolecular acetylation activity, (ii) K383 in PopP2 is conserved among all members of the YopJ family of T3Es, and (iii) despite the presence of an integral catalytic triad, a PopP2-K383R mutant is no longer recognized by RRS1-R-expressing Arabidopsis plants, strongly support the idea that this lysine residue represents a good candidate to be the unknown acetyl-CoA binding site within the YopJ family. As a consequence of the mutation of K383 that prevents PopP2 to bind acetyl-CoA, PopP2 would be impaired not only in its autoacetyl-transferase activity in planta but also in its ability to acetylate its putative host substrate(s).
At present, it is difficult to distinguish between the two following hypotheses: is PopP2 autoacetylation and/or acetylation of its host substrate(s) required to trigger the RRS1-R resistance response? Indeed, both possibilities, which are not mutually exclusive, are consistent with the loss of avirulence activity of the PopP2-K383R mutant that behaves like the catalytic (acetyl-transferase defective) mutant PopP2-C321A, after inoculation of RRS1-R-expressing plants. According to the guard model [12], RRS1-R may survey the acetylation state of host components targeted by PopP2. The identified PopP2-interacting proteins RRS1-R and the vacuolar RD19 cysteine protease [46] are potential substrates that may be targeted by the acetyl-transferase activity of PopP2. However, despite several attempts, immunoprecipitation experiments, followed by Western blot analysis with α-Ac-K antibody, did not allow detection of any acetylated form of RRS1-S, RRS1-R or RD19 upon co-expression with PopP2. Indeed, additional plant or bacterial components may be required for PopP2 acetylation of its substrate(s). Alternatively, as previously described for YopJ, which is able to acetylate target proteins on serine and threonine residues [45], PopP2 acetyl-transferase activity may target additional amino acid residues other than lysine, which would be undetectable under our experimental conditions.
Future work will address these questions as well as the acetylated state of PopP2 in planta and whether its activity may be modulated by the presence of putative plant inhibiting and/or activating co-factors. Indeed, rather than acting as substrates, PopP2-interacting proteins, through their physical association with PopP2, might also modulate its enzymatic activity. According to this hypothesis, it is tempting to hypothesize that RRS1-R, unlike RRS1-S, might potentiate PopP2 enzymatic function, thereby leading to its recognition and activation of the resistance response. Identification of host proteins targeted by PopP2 activity and/or involved in its regulation will significantly contribute to the understanding of the molecular role(s) played by protein acetylation during plant innate immunity.
All experiments reported in this article were performed at least three times with similar results.
Plasmids used in this study were constructed by Gateway technology (GW; Invitrogen) following the instructions of the manufacturer. PopP2 mutants were generated from pENTR-PopP2 [9] by a two-step PCR-based site-directed mutagenesis using PrimeStar HS DNA polymerase from Takara Bio Inc. (Otsu, Japan) to introduce the following nucleotide substitutions: K268R: codon 268 AAG to CGG (K268R-F: AT ATT CGC CGG GAC GCC TCT GGT ACG AGC GTG ATC, K268R-R: GA GGC GTC CCG GCG AAT ATC TGC GGC TCT GGT); K285R: codon 285 AAA to AGA (K285R-F: C CTC CGA AAG GAA AGA GAT GAA AGC GCG TAC GTC GA, K285R-R: C ATC TCT TTC CTT TCG GAG GGG ATC GAC AAC G); K335R: codon 335 AAG to CGG (K335R-F: C AAG ATG CAT GAC CGG GAC GAC GCG TTT GC, K335R-R: C GTC CCG GTC ATG CAT CTT GAG TGC AAG TGA); C321A: codon 321 TGC to GCC (C321A-F: C TTC TTC GAT GCC CGG ATA CTC TCC CTG TCA CT, C321A-R: GAG TAT CCG GGC ATC GAA GAA GGA CTT CTG A); K383R: codon 383 AAA to CGG (K383R-F: GT ATG ATG CGG CAT GGT CAA GCC GCA, K383R-R: C TTG ACC ATG CCG CAT CAT ACG GGC GT). Second step PCR was performed using popP2-specific primers (AttB1-popP2-F: GGGG ACA AGT TTG TAC AAA AAA GCA GGC TTA ATG AAG GTC AGT AGC GCA; AttB2-popP2-R: GGGG ACC ACT TTG TAC AAG AAA GCT GGG TCG TTG GTA TCC AAT AGG GAA TCC). pENTR-PopP2 truncated clones (PopP281–488, PopP2-C321A81–488, PopP2-K383R81–488) were generated by PCR performed on corresponding pENTR-PopP2 full length clones (AttB1-PopP2-81F: GGGG ACA AGT TTG TAC AAA AAA GCA GGC TTA ATG CAT GTG CCA CTC CTA GAC A and AttB2-PopP2-R). Gateway PCR products flanked by attB sites were recombined into pDONR207 vector (Invitrogen) via a BP reaction to create corresponding entry clones with attL sites. Relative pENTR-PopP2 mutants were recombined into appropriate destination vectors via an LR reaction. pENTR-RRS1-S, pENTR-RRS1-R and pENTR-RD19 (pDONR207 vector backbone) used in this study have been previously described [9], [46]. In planta expression of the various protein fusions described in this study was performed using Gateway compatible pAM-PAT-P35S-GW binary vectors [46] to allow C-terminal protein tagging with the following epitopes: CFP, YFPv, 3x-Flag, 3x-HA and c-myc.
Plant root inoculations and bacterial internal growth curves were performed as previously described [37]. Plant phenotypic responses were scored daily, using a disease-index scale ranging from 0 to 4, according to the percentage of wilted leaves (0 = no wilt, 1 = 1 to 25%, 2 = 26 to 50%, 3 = 51 to 75%, 4 = >75%).
The CFP and YFP fluorescence in N. benthamiana leaves was analyzed with a Confocal Laser Scanning Microscope (TCS SP2-SE, Leica, Germany) using a 63× water immersion objective lens (numerical aperture 1.20, PL APO). CFP fluorescence was excited with the 458-nm ray line of the argon laser and recorded in one of the confocal channels in the 465- to 520-nm emission range. YFP fluorescence was excited with the 514-nm line ray of the argon laser and detected in the range between 520 and 575 nm. Images were acquired in the sequential mode using Leica LCS software (version 2.61).
Fluorescence lifetime of the donor was experimentally measured in the presence and absence of the acceptor. FRET efficiency (E) was calculated by comparing the lifetime of the donor in the presence (τDA) or abscence (τD) of the acceptor: E = 1−(τDA)/(τD). FRET-FLIM measurements were performed using a multiphoton FLIM system coupled to a streak camera (Krishnan et al., 2003). The light source was a mode-locked Ti:sapphire laser (Tsunami, model 3941, Spectra-Physics, USA), pumped by a 10 W diode laser (Millennia Pro, Spectra-Physics), delivering ultrafast femtosecond pulses with a fundamental frequency of 80 MHz. A pulsepicker (model 3980, Spectra-Physics) was used to reduce the repetition rate to 2 MHz. All the experiments reported in this work were carried out at λ = 820 nm, the optimal wavelength to excite CFP in multiphoton mode while minimizing the excitation of YFP (Chen and Periasamy, 2004). The power delivered at the entrance of the FLIM optics was 14 mW. All images were acquired with a 60× oil immersion lens (Plan Apo 1.4 numerical aperture, IR) mounted on an inverted microscope (Eclipse TE2000E, Nikon, Japan) coupled to the FLIM system. The fluorescence emission was directed back out into the detection unit through a short pass filter (λ<750 nm). The FLIM unit was composed of a streak camera (Streakscope C4334, Hamamatsu Photonics, Japan) coupled to a fast and high-sensitivity CCD camera (model C8800-53C, Hamamatsu). For each nucleus, average fluorescence decay profiles were plotted and lifetimes were estimated by fitting data with tri-exponential function using a non-linear least-squares estimation procedure with Origin 7.5 software (OriginLab, Northampton USA).
For Agrobacterium-mediated Nicotiana benthamiana leaf transformations, the relevant GV3101 strains were grown in Luria-Bertani liquid medium containing 100 µg mL-1 rifampicin, 25 µg mL-1 gentamicin and 25 µg mL-1 carbenicillin at 28°C for 24 h before use. Bacteria were harvested and resuspended in infiltration medium (10 mM MES pH 5.6, 10 mM MgCl2, 150 µM acetosyringone) to an OD600nm of 0.5 and incubated for 2 h at room temperature before leaf infiltration. The infiltrated plants were incubated for 36 h in growth chambers under controlled conditions [46]. Arabidopsis transient assays were performed as previously described [66]. Four day-old seedlings were infiltrated with appropriate A. tumefaciens strains and harvested 3 days after infiltration. MG132 treatment (100 µM) was performed 24 h after agroinfiltration. Leaf disks were harvested 12 h later and grounded in Laemmli buffer.
For bacterial overexpression, popP2 wild-type and popP2 mutant versions were recombined into a pGEX-GWY vector by LR recombination (Invitrogen). Relative GST-PopP2 constructs were expressed in Escherichia coli Rosetta DE3 cells (Novagen). The intermolecular acetylation assay was performed by co-expressing truncated PopP2 variants (PopP281–488, PopP2-C321A81–488 and PopP2-K383R81–488) with either PopP2 or PopP2-C321A in Rosetta DE3 cells as GST fusion proteins. A fragment containing the GST-PopP2 or GST-PopP2-C321A coding sequences under the control of a tac promoter was introduced into the pGEX-GWY vector to generate pGEX-GWY-GST-PopP2 and pGEX-GWY-GST-PopP2-C321A destination vectors, respectively. Relative pENTR-PopP2 truncated clones (PopP281–488, PopP2-C321A81–488, PopP2-K383R81–488) were recombined into pGEX-GWY-GST-PopP2 or pGEX-GWY-GST-PopP2-C321A vectors by LR recombination. Cells were grown at 37°C to an OD600nm of 0.6 in LB medium (50 µg mL-1 carbenicillin and 30 µg mL-1 chloramphenicol) and induced with 400 µM isopropyl-βthiogalactopyranoside (Roche Applied Science) for 3 h at 28°C. Before lysis, pelleted cells were concentrated 10 times in PBS [phosphate-buffered saline (pH 8.0)] supplemented with 1 mM phenylmethylsulfonyl fluoride (PMSF, Sigma) and 10 mM Sodium butyrate. GST purifications were performed using Glutathione Sepharose (GE Healthcare) according to the instructions of the manufacturer. For a given experiment, between 3 and 5 µg of purified proteins were used for Western blot analysis.
Equal amounts of protein were loaded onto a sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE), followed by electrophoresis and transfer to Protran BA85 nitrocellulose membranes (Whatman, Germany). Transferred proteins were visualized by Ponceau S red staining. Plant protein samples obtained from both N. benthamiana leaves (4 discs of 8 mm diameter harvested 24 to 36 hours post-infiltration) and Arabidopsis (20 seedlings), were homogenized in 250 µL of Laemmli loading buffer. Antibodies used for Western blotting were anti-Flag-HRP (M2, Sigma, dilution 1∶5000), anti-HA-HRP (3F10, Roche Applied Science, dilution 1∶5000), anti-myc-HRP (9E10, Roche Applied Science, dilution 1∶3000), monoclonal mouse anti-acetylated-Lysine antibody (Ac-K-103, Cell Signaling Technology; dilution 1∶2000), polyclonal rabbit anti-GroEL (Stressgen Biotechnologies Corporation, dilution 1∶20000), goat anti-mouse and anti-rabbit IgG-HRP antibodies (Santa Cruz, dilution 1∶10000).
GST affinity purified proteins were migrated on an SDS-PAGE. The gel band samples were washed several times by incubation in 25 mM NH4HCO3 for 15 min and then in 50% (v/v) acetonitrile containing 25 mM NH4HCO3 for 15 min. 0.15 µg of modified trypsin (Promega, sequencing grade) in 25 mM NH4HCO3 was added to the dehydrated gel spots for an overnight incubation at 37°C. Peptides were then extracted from gel pieces in three 15 min sequential extraction steps in 30 µL of 50% acetonitrile, 30 µL of 5% formic acid and finally 30 µL of 100% acetonitrile. The pooled supernatants were then dried under vacuum. For nano-LC-MS/MS analysis, the dried extracted peptides were resuspended in 30 µl of water containing 2.5% acetonitrile and 0.1 fluoroacetic acid. A nano-LC-MS/MS analysis was then performed (Ultimate 3000, Dionex and LTQ-Orbitrap XL, Thermo Fischer Scientific). The method consisted in a 60-minute gradient at a flow rate of 300 nL/min using a gradient from two solvents: A (5% acetonitrile and 0.1% formic acid in water) and B (80% acetonitrile and 0.08% formic acid in water). The LC system includes a 300 µm×5 mm PepMap C18 precolumn and a 75 µm×150 mm C18 column (PepMap C18 phase). MS and MS/MS data were acquired using Xcalibur (Thermo Fischer Scientific). The raw data were automatically processed through Mascot Daemon software (Matrix Science) and the MS/MS spectra were searched against the SwissProt/Trembl database using trypsin as the specific enzyme. Peptide modifications allowed during the searches were: N-acetyl (N-ter protein), oxidation (M), dioxidation (M), trioxidation (C) and K-acetylation.
popP2 allelic sequences were introduced by LR recombination in pLAFR6-P2GFH (popP2promoter::GW::3x-HA)vector that allows the expression of PopP2 tagged in the C-terminus end with 3x-HA epitope tag. To drive the expression of epitope tagged-PopP2 variants, a fragment that encompass 384 bp of PopP2 promoter sequence (from ATG, amplified from GMI1000 genomic DNA) was used. ΔhrcV mutation was introduced into ΔpopP2 strain by natural transformation of bacteria grown in MMG medium supplemented with glycerol (20 g L-1) with 3–5 µg of ΔhrcV genomic DNA. The derived pLAFR-P2GFH plasmids were introduced into R. solanacearum ΔpopP2 (GRS100, [9]) and ΔpopP2/Δhrcv (this study) strains by electroporation and selected on 10 µg mL-1 tetracyclin and 15 µg mL-1 gentamicin. Production of R. solanacearum concentrated supernatants from 2 mL of culture was performed as previously described [67]. Total extracts were obtained from the lysis of pelleted cells (2 mL) in Bugbuster protein extraction reagent (Novagen) following the instructions of the manufacturer.
Sequence data from this article can be found in the Arabidopsis Genome Initiative or GenBank/EMBL data libraries under the following accession numbers: AvrA (AAB83970), AvrBst (AAD39255), HopZ2 (ABK13722), PopP1 (CAF32331), PopP2 (CAD14570), RD19 (At4g39090), RRS1-S (At5G45260), RRS1-R (HQ170631), VopA (AAT08443), XopJ (YP_363887), YopJ (NP_395205).
|
10.1371/journal.pgen.1003784 | MEIOB Targets Single-Strand DNA and Is Necessary for Meiotic Recombination | Meiotic recombination is a mandatory process for sexual reproduction. We identified a protein specifically implicated in meiotic homologous recombination that we named: meiosis specific with OB domain (MEIOB). This protein is conserved among metazoan species and contains single-strand DNA binding sites similar to those of RPA1. Our studies in vitro revealed that both recombinant and endogenous MEIOB can be retained on single-strand DNA. Those in vivo demonstrated the specific expression of Meiob in early meiotic germ cells and the co-localization of MEIOB protein with RPA on chromosome axes. MEIOB localization in Dmc1−/− spermatocytes indicated that it accumulates on resected DNA. Homologous Meiob deletion in mice caused infertility in both sexes, due to a meiotic arrest at a zygotene/pachytene-like stage. DNA double strand break repair and homologous chromosome synapsis were impaired in Meiob−/− meiocytes. Interestingly MEIOB appeared to be dispensable for the initial loading of recombinases but was required to maintain a proper number of RAD51 and DMC1 foci beyond the zygotene stage. In light of these findings, we propose that RPA and this new single-strand DNA binding protein MEIOB, are essential to ensure the proper stabilization of recombinases which is required for successful homology search and meiotic recombination.
| Homologous recombination allows faithful repair of damaged DNA; in mitotic cells, it necessitates the formation of single strand DNA (ssDNA), which is first protected by RPA and then coated by the RAD51 recombinase to mediate homology search. Specific modifications are made to this mechanism during meiosis, a specialized division that allows halving the ploidy of the genome and the production of haploid gametes. Among others a specialized recombinase DMC1 is added to its somatic paralog RAD51 to perform homology search. We identified a new meiotic protein that we named meiosis specific with OB domains (MEIOB). Our findings indicate that MEIOB binds ssDNA, and we propose that MEIOB is a meiotic paralog of RPA, another OB-domain containing protein. Meiob mutant mice were infertile and unable to complete meiotic recombination, most likely due to destabilization of DMC1 and RAD51 in the absence of MEIOB. Meiosis appears thus to be a ‘game of two pairs’ using both the canonical players in homologous recombination (RPA and RAD51) and a second set of paralogs (MEIOB and DMC1). Identifying such new players should help clarify some genetic causes of infertility and shed new light on the interplay between the molecular actors involved in maintaining genome stability.
| Meiosis is a central process of sexual reproduction. This specialized cell division program allows halving the genome of diploid germ cells to produce haploid gametes. In order to ensure proper segregation of homologous chromosomes during the first meiotic division, these must become connected through chiasmata [1]. Crucially, formation of chiasmata depends on the occurrence of inter-homolog crossovers (CO) during the first meiotic prophase. COs originate from the recombination mediated-repair of programmed double strand breaks (DSBs) during meiotic prophase I. Meiotic recombination differs from mitotic recombination in that it uses a chromatid from the homolog instead of the sister chromatid as a template for repair [2]. It also favors CO formation and involves specific proteins [3]. In mice, about 250–300 DSBs are generated during the leptotene stage by the catalytic activity of the conserved topoisomerase-like transesterase SPO11 [4]–[6]. DNA ends at DSBs are resected to produce single stranded DNA for homology search [7]. Only a subset of DSBs form COs, the remaining DSBs being repaired without chromosome arm exchanges. The decision to form or not a CO is thought to be made before or during strand invasion of the homologous chromosome [8]. This being the case, the nature and the regulation of proteins loaded at broken ends is likely to be important for the outcome of the DSBs. Single-strand DNA (ssDNA) formed during DNA metabolism is coated by the trimeric replication protein A (RPA) complex (composed of RPA1 70 kDa, RPA2 32 kDa, RPA3 14 kDa) serving to protect from degradation and to prevent secondary structure formation. RPA binds ssDNA with high affinity through oligonucleotide binding (OB) domains [9], [10]. During homologous recombination, RPA has to be removed from the 3′ssDNA of broken ends to allow the formation of a presynaptic nucleofilament. This is essential for homology search and the formation of a physical connection between the invading ssDNA and a homologous duplex DNA template. Whereas mitotic recombination only needs RAD51 to search and invade a homologous sequence, meiotic recombination requires an additional recombinase, the meiosis-specific DMC1 protein [11]. In plants and mammals, the presence of BRCA2, which directly interacts with RAD51 and DMC1, is necessary for their localization, however, the exact mechanisms involved in RAD51 and DMC1 loading during meiosis remain to be determined [12]–[14]. Working models based on data mainly obtained in S. cerevisiae and A. thaliana proposed that Rad51 is essential for proper Dmc1 loading and that Rad51 recombinase activity has to be inhibited to favor the homolog search bias [15]–[17].
An event tightly associated with meiotic recombination is the pairing of homologous chromosomes through the formation of the synaptonemal complex (SC). SC is a tripartite structure comprising two lateral or axial elements (AE) and a central element. Early during the leptotene stage an AE is formed along each chromosome. However, it is only after successful homology search, that complete synapsis between homologs is observed connecting the AEs of homologous chromosomes with transverse filaments at the pachytene stage [18], [19]. Subsequently, proper SC formation is required to the integrity of CO formation [20], [21].
In order to better understand the complex events that occur during meiosis, it is crucial to identify the specific proteins required for meiotic recombination. Hereby, we characterized a conserved and meiosis-specific protein containing ssDNA binding domains homologous to those of RPA1. We named this protein meiosis-specific with OB domains (MEIOB) in vertebrates. In fly, mutation of the likely Meiob gene homolog hold'em (hdm) reduces meiotic crossover formation and sensitizes somatic cells to DNA-damaging agents [22]. Unlike MEIOB, hdm activity is not specific to meiosis. In the present work, we demonstrate that murine Meiob is specifically required after DSB formation during early steps of meiosis to ensure proper DSB repair by homologous recombination, a prerequisite for efficient crossover formation and male and female fertility.
To identify candidate genes possibly contributing to meiosis, we performed a transcriptome analysis of magnetic-activated cell sorted (SSEA1+) male and female embryonic germ cells at 13.5 days post-coïtum (dpc). At this stage, female but no male germ cells enter meiosis. We identified Meiob (referred to as RIKEN cDNA 4930528F23 gene at the time) among the most differentially expressed genes in female germ cells (complete data set will be published elsewhere). The Meiob murine gene is located on chromosome 17 and is composed of 14 exons (Figure S1A) coding for a predicted protein of 470 amino acids. Amplification of the full length Meiob transcript from 13.5 dpc mouse ovary gave a unique band (Figure S1B), the sequencing of which confirmed the predicted sequence (data not shown).
Search using tBlastn (http://blast.ncbi.nlm.nih.gov/Blast.cgi) with either the full length sequence or short conserved motifs identified Meiob homologs in the genomes of almost all metazoans (Figure 1A), all except Nematoda. Multiple alignments of full length amino acid sequences indicated a high degree of conservation of the various Meiob homologs in vertebrates (e.g. 91% of similarity and 85% of identity between mouse and human) while invertebrates displayed sequences with a much lower conservation (e.g. 23% of identity between mouse and fly). No ortholog could be retrieved in the genomes of fungi or plants, although a homolog was identified in the single-celled organism, Capsaspora owczarzaki. Interestingly, the closest Meiob paralog identified was the replication protein A large subunit (Rpa1). Consistently, InterProScan Search (http://www.ebi.ac.uk/Tools/pfa/iprscan/) identified three DNA binding domains (dbd), termed oligonucleotide/oligosaccharide binding (OB) folds, which were homologous to those of RPA1 (Figure 1B). However, the orthologs of the Meiob group formed a family distinct from the Rpa1 group (Figure 1A, phylogenetic tree of MEIOB and RPA1 using MULTALIN) [23]. These data suggest that Meiob may have evolved from an ancestral Rpa1 shortly before the appearance of multicellularity.
RT-qPCR analysis of Meiob transcription pattern in adult and fetal tissues detected Meiob transcripts exclusively in fetal ovary, postnatal testis and liver (Figures S2A and S2B). In the mouse ovary, Meiob expression started at 12.5 dpc, reached a maximum at 15.5 dpc and decreased to become undetectable in post natal life (Figure 1C). In testes, Meiob expression started at 10 days post partum (dpp), reached a maximum at 20 dpp and was maintained throughout adult life (Figure 1C). This expression profile was in accordance with previously published data screening genes expressed in spermatogenic cells [24]. Additionally, in human fetal gonads, MEIOB expression was only detected in the ovary starting at the 14th weeks post fertilization (Figure 1D). Thus patterns observed in both mouse and human gonads correlate with an expression during the meiotic prophase I suggesting a conserved role within this process.
We next investigated which cell type expressed Meiob in the gonads using fluorescence activated cell sorted (FACS) germ and somatic cells from 13.5 dpc Oct4-GFP ovaries. Meiob expression levels were 112 fold higher in germ cells compared to somatic cells (Figure 1E). This germ cell-specific expression was further confirmed using w/w mouse fetal ovaries devoid of germ cells (Figure S2C). To determine whether the corresponding MEIOB protein was produced in these tissues, we generated two rabbit polyclonal antibodies raised against different MEIOB peptides and confirmed their specificity using a human tagged-MEIOB produced in HEK-293 cells (Figures 2A and S5A). Western blot analysis of various populations of FACS-sorted spermatocytes (Figure S3) detected a 52 kDa band consistent with the predicted MEIOB molecular weight. MEIOB protein was detected solely during early meiosis prophase I (‘early 4N’ fraction containing leptotene, zygotene and few pachytene spermatocytes) and not during later stages of meiosis (Figure 2A). No obvious signal was detected in whole adult testis protein extracts, possibly reflecting the few number of cells expressing MEIOB in this tissue (see ‘early 4N’ in Figure S3) and/or low expression levels.
MEIOB immunostaining performed on chromosomal spreads from adult testes indicated that MEIOB form foci located on chromosome axes stained with SYCP3 in spermatocytes (Figure 2B). These MEIOB foci were visible as early as leptotene stage and persisted throughout the zygotene stage, with about 250 foci per spermatocyte, and decreased drastically during the pachytene stage (Figure 2C). We observed a similar staining in 15.5 dpc oocytes, corresponding to leptotene, zygotene and early pachytene stages (Figure S4). Hardly any signal was observed on chromosome axes in spreads from Meiob−/− mice confirming the specificity of our antibodies. Signal off the axes was considered as non specific as it was retrieved in Meiob−/− spermatocytes and only foci on the axes were repeatedly observed when using chromosome spreads prepared with various protocols (see materials and methods section and Figure S7). At the diplotene stage, about 25 weak MEIOB foci were detected. However we cannot formally exclude that some of these foci are unspecific due to the absence of the equivalent stage in the Meiob−/− mice. During the pachytene stage, γH2AX labels the sex body containing the X and Y chromosomes that can only pair on a limited portion called the pseudo-autosomal region (PAR). Immunostaining for MEIOB and γH2AX indicated that MEIOB was retrieved on chromosomes located in the sex body at this stage and its distribution on the X appeared similar to that observed on autosomes (Figure 2D). Of note, we always retrieved at least one bright focus in the PAR and foci in the PAR were observed until mid/late pachytene stage, when few foci persisted on autosomes.
The MEIOB being related to RPA1, a conserved ssDNA binding protein, we hypothesize that MEIOB too may bind ssDNA. To investigate whether MEIOB OB-domains confer any ssDNA binding activity, we performed ssDNA and double-strand DNA (dsDNA) oligonucleotide pull-down assays using recombinant tagged-MEIOB protein produced in a cell-free system. Pull-down assays performed with biotin 60-mer revealed a significantly greater retention of tagged-MEIOB protein on ssDNA than on dsDNA (Figure 3A). A similar experiment carried out with whole testis protein extract indicated that the endogenous protein also preferentially binds ssDNA (Figure 3B). Of note, MEIOB binding efficiency on ssDNA increased with oligonucleotide length as 30 mer ssDNA retained two fold less MEIOB than did 60 mer ssDNA. Extract from HEK-293 cells expressing tagged-MEIOB produced the same results (Figure S5B).
In order to evaluate the robustness of MEIOB binding to ssDNA, we performed ssDNA affinity chromatography. Protein extracts from HEK-293 cells expressing tagged-MEIOB were loaded on a column containing a ssDNA matrix, and proteins were eluted by progressively increasing salt concentration. Detection of eluted MEIOB by western blot revealed that MEIOB protein was eluted at a concentration of 0.75M NaCl (Figures 3C and S5C), indicating high affinity for ssDNA all be it lower than that of the trimeric RPA (most RPA being eluted at 1M and above, data not shown).
Based on our in vitro observations and protein domain predictions we hypothesized that MEIOB could target ssDNA in vivo. We therefore compared MEIOB localization to that of RPA, RAD51 and DMC1 in spermatocyte chromosome spreads. MEIOB and RPA foci overlapped considerably (Figures 4A and S6A), indicating that MEIOB is present on recombination initiation sites. However, some foci appeared solely stained for MEIOB. Respectively, 71, 76 and 86% of RPA2 foci were stained for MEIOB and 26, 49 and 61% of MEIOB foci were stained for RPA2 at leptotene, zygotene and early pachytene stages. MEIOB staining was retrieved in most RAD51 or DMC1 foci (Figure 4A), particularly in the early pachytene stage when over 80% of RAD51 or DMC1 foci were stained for MEIOB (Figures S6B and S6A). On the other hand, only about half of the MEIOB foci were stained for RAD51 or DMC1. To better characterize MEIOB recruitment we investigated its localization on chromosome spreads from Spo11−/− and Dmc1−/− mice. In the absence of SPO11, no DSBs are formed [4], [5] and we observed no MEIOB foci on the chromosome axes in spermatocytes (Figure 4B). In the absence of the DMC1 recombinase, DSBs cannot be repaired and are believed to accumulate with extensive resection [25]. As a consequence Dmc1−/− spermatocytes failed to complete synapsis and arrested at a stage termed ‘pachytene-like’ equivalent to late zygotene or early pachytene cells in wild type cells based on the stage of AE formation (i.e. with SYCP3 staining) [26], [27]. In Dmc1−/− spermatocytes, MEIOB accumulated on chromosome axes suggesting that MEIOB is recruited to the DSB sites before strand invasion (Figure 4C). Consistent with an expected localization on ssDNA and with presumed longer single stranded DNA ends at DSB sites in Dmc1−/− meiocytes, MEIOB staining was brighter in Dmc1−/− as compared to wild type spermatocytes (Figure 4C). Furthermore, in Dmc1−/− spermatocytes, MEIOB showed near perfect colocalization with almost all RPA2 foci (Figure S7A) and also colocalized with ATR foci (Figure S7B) a protein involved in ssDNA signaling [28].
To investigate the role of MEIOB in vivo, Meiob+/− mice were generated through the deletion of exons 2 to 8 (Figure S1A). Transmission of Meiob mutant allele exhibited the expected Mendelian distribution. RT-qPCR amplifying various parts of the Meiob transcript (including exons 9 and 10 still present in the mutant) confirmed the complete absence of Meiob mRNA in the adult testis of homozygous mutants and confirmed that our genetic model was indeed a null mutant (Figures 5A and S1C). Meiob−/− mice developed and grew normally. Histological analysis revealed no anatomical defect (including in the liver, data not shown). However, Meiob−/− male and female adult mice mated for four months with Meiob+/− or Meiob+/+ counterparts produced no offspring, though vaginal plugs were formed normally (Figure 5B). Meiob thus appears mandatory for both female and male fertility.
Next we analyzed gametogenesis in Meiob−/− mice to identify the defective steps. Meiob−/− adult testes and ovaries presented a strong reduction in size when compared to Meiob+/+ gonads (Figure 5C) (testes being 3.8 times smaller (Figure S8A) and mutant ovaries hard to distinguish). During fetal life, histological analysis revealed no morphological alteration in the Meiob−/− gonads (data not shown). During post-natal life however, in 1 dpp Meiob−/− ovary, despite there being no obvious diplotene stage, the marker of late pachytene/diplotene oocytes, P63-staining, indicated the presence of numerous oocytes similar to those forming primordial follicles (Figure 5D). Germ cell counting based on immuno-detection of the germ cell marker DDX4 revealed a subsequent drop in oocyte number starting at 3 dpp (Figure S9A) until 10 dpp when no more germ cells could be observed in the Meiob−/− ovaries (Figures 5E and S9A). Meiob−/− post-natal testes histology revealed no alteration until 10 dpp. In Meiob−/− adult testes, no stage beyond primary spermatocyte was observed. Some tubules contained only spermatogonia and others contained an accumulation of spermatocytes that were mostly at the leptotene, zygotene and pachytene-like stages based on chromatin compaction and shape (Figure 5F). Furthermore, no spermatozoon was observed in Meiob−/− epididymis (Figure S8B). Cleaved caspase-3 staining indicated a significant increase of apoptotic germ cells starting after meiotic initiation in both sexes: from 18.5 dpc in Meiob−/− ovaries and from 10 dpp in Meiob−/− testes (Figures 5G and 5H). Meiob+/− male or female mice did not show any reproductive alteration when compared to Meiob+/+ mice, including number of pups per litter (Figure 5B), size of adult ovaries and testes (Figures 5C and S8A), follicle population size in adult ovaries (Figure S9B), germ cell apoptosis (Figures 5G and 5H) or seminal vesicle weight (Figure S8C).
As defects in Meiob−/− occur during meiosis prophase I, we first investigated bivalent chromosome formation in wild type and Meiob null meiocytes. Homologous chromosomes pair and become connected along their lengths by synaptonemal complexes (SCs). AE formation is stained by SYCP3 initiated along the shared cores of sister chromatid pairs during leptotene. During zygotene a zipper like connection (stained by SYCP1) between the two AEs is initiated and completed by pachytene to fully connect homologous chromosomes. As illustrated on chromosome spreads in figure 6, the linear co-staining of SYCP3 and SYCP1 reflects fully synapsed bivalents in pachytene wild type spermatocytes. In Meiob−/− spermatocytes, as shown with SYCP3 staining, formation of AE appear to be unaffected as cells progress from leptotene to zygotene stages (Figure 2B). However, no cell with complete synapsis could be observed in adult testes (Figures 6A and 6B). The most advanced stage with regards synapsis presented an abnormal pairing of chromosomes resembling a stage between zygonema and pachynema (here termed ‘pachytene-like’ stage) (Figures 2B and 6). In such pachytene-like Meiob−/− spermatocytes, SYCP1 staining presented considerable cell-to-cell heterogeneity with regards the number of SYCP1 stretches per cell (Figure 6A). Most (∼60%) pachytene-like spermatocytes contained no SYCP1 complex and others presented from one up to nineteen (in very rare cells, less than 1%) SYCP1 stretches (Figure 6B). Similarly, 1 dpp Meiob−/− oocytes presented pachytene-like features with incomplete SYCP3/SYCP1 colocalization indicating a major defect in synapsis (Figure 6C). Both adult testes and newborn ovaries frequently displayed non-homologous pairing in Meiob deficient meiocytes with pairing between chromosomes of different sizes or combinations of more than two chromosomes (Figures 6A and 6C, white square magnifications).
DSB formation and repair is essential during prophase I of meiosis. We monitored these events using γH2AX staining, γH2AX being a marker of DNA DSB [29], [30]. In sections from Meiob−/− adult testes, all spermatocytes presented a robust staining for γH2AX while such a signal is usually observed in only few cells in wild type testes (Figure 7B). Analysis of chromosome spreads of spermatocytes from wild type and Meiob−/− mice confirmed the presence of γH2AX in leptotene and zygotene cells suggesting that DNA DSBs are normally formed. In all observed chromosome spreads of pachytene-like spermatocytes from Meiob−/− mice, the γH2AX signal was maintained whereas it had disappeared as expected in pachytene spermatocytes from wild type mice (Figure 7A). Additionally, in the Meiob−/− pachytene-like spermatocytes, no structure resembling the sex body was formed (Figure 7A). In 1 dpp ovaries, wild type oocytes are mostly in late pachynema and diplonema and none are stained for γH2AX whereas Meiob−/− oocytes retained a robust γH2AX-staining at this age (Figure S10). Thus, Meiob appears to be required for DNA DSB repair during meiosis. In several meiotic mutants the persistence of γH2AX staining is frequently associated with synapsis defects (i.e. Sycp1−/− [31]). Considering our finding of Meiob deficiency impairing synapsis, we attempted to correlate the extent of synapsis to DNA repair through SYCP1/γH2AX co-immunostaining in Meiob−/− spermatocytes. In these pachytene-like cells, even those with the highest rates of synapsis showed persistence of γH2AX staining (Figure 7C) indicating that the observed γH2AX signal is unlikely the consequence of the synapsis defect and strongly suggests persistence of unrepaired breaks.
Taking into account MEIOB spatial and temporal localization and its requirement for proper DSB repair, we investigated homologous recombination in Meiob−/− spermatocytes through immunolocalization of RPA2, RAD51, DMC1 and MLH1. In Meiob−/− leptotene and zygotene/pachytene-like cells, RPA foci were formed on chromosome axes and in normal abundance when compared to wild type equivalent stages (Figures 8A and 8B). This equally confirms that DNA DSBs were produced with no overt defect in the absence of MEIOB. Immunolocalizations of RAD51 and DMC1 were performed with antibodies recognizing specifically RAD51 or DMC1. In Meiob−/− leptotene stage spermatocytes, RAD51 and DMC1 foci were localized on chromosome axes with no overt differences in comparison to wild type counterparts (Figure 9). In wild type zygotene stage these foci were maintained and their numbers decreased over the course of the pachytene stage. However in Meiob−/−, we observed a massive decrease in both RAD51 and DMC1 stainings at zygotene and pachytene-like stages. By comparison with wild type, the number of RAD51 and DMC1 foci respectively decreased by 70% and 69% at mid-zygotene in Meiob−/− (Figure 9B). Measurement of RAD51 and DMC1 foci intensities indicated no significant change in the mean intensity during leptotene and zygotene stages (Figure S11A). Intense (high and medium) foci tended to decrease first in mid-zygotene and all foci intensely and mildly (low) stained considerably decreased in late zygotene/pachytene like stage in Meiob−/− spermatocytes (Figure S11B). These data indicate that while the formation of RAD51 and DMC1 foci is unaltered in Meiob−/− mutants, their stabilization is impaired. The absence of synapsis, the persistence of γH2AX and the reduction of RAD51 and DMC1 foci observed in the absence of MEIOB suggest a strong defect in recombination with consequential impairment of CO formation. To test this hypothesis we investigated CO formation using MLH1 immunostaining. MLH1 is believed to mark future CO sites in mid-to-late pachytene, and is essential at the late stages of recombination in the formation of CO [32]. As expected at least one MLH1 focus was observed per bivalent chromosome in wild type pachytene spermatocytes. In contrast, no MLH1 foc1 were observed in Meiob−/− spermatocytes (Figure 10A). This likely reflects a blockade prior to mid-pachytene preventing thus CO formation. Meiob−/− oocytes are eliminated at a developmental time-point where most wild type oocytes reached a stage beyond pachytene. This allowed us to address whether MEIOB is required or not for the formation of MLH1 foci in cells that likely correspond to pachytene stage based on the developmental time-course of wild type cells. Many oocytes are in pachytene stage, and have MLH1 foci in wild type 17.5 dpc embryos. Interestingly we were unable to observe bright MLH1 foci along the AEs in the oocytes of 17.5 dpc Meiob−/− embryos (Figure 10B). Altogether these data suggest that MEIOB may be involved in both early/homology search –related and possibly also in later/post-homology search steps of meiotic recombination and CO formation.
In this study, we have identified a new major player in mammalian meiosis and in particular have demonstrated for the first time an absolute requirement of MEIOB for meiosis prophase I completion in the mouse. The expression profile and deletion of Meiob indicated a specific role for this protein during meiosis prophase I. Furthermore, our data indicate that MEIOB is mandatory for DSB repair and crossover formation. MEIOB also appears to favor faithful and complete synapsis. As observed with other mutant mice in which these steps are impaired, these defects caused meiotic arrest, meiocyte death by apoptosis and infertility. The complete infertility of both male and female mice indicates that Meiob is one of the core genes required for meiosis, as the deletion of other meiotic genes in mice sometimes only induces male infertility [12], [34]–[35]. In summary, this mammalian meiotic mutant impairs meiotic recombination in both male and female and impairs RAD51 and DMC1 stabilization.
Our data support the idea that MEIOB binds ssDNA. First, protein domain prediction proposed that MEIOB contains three ssDNA binding domains, denoted OB domains, similar to those of RPA1 known to have a high affinity for ssDNA [36]. Second, in vitro DNA binding assays demonstrated that MEIOB is particularly retained by ssDNA. Moreover, MEIOB foci formed along AE and colocalized to a large extent with RPA which marks ssDNA filament [37], [38]. The number of foci corresponded to the expected number of DNA DSBs generated by SPO11 and no MEIOB foci were observed in the absence of SPO11-mediated DSBs. Lastly, when 3′ssDNA accumulated (i.e. in Dmc1−/−, in the absence of strand invasion) MEIOB colocalized almost perfectly with both RPA and ATR, known to sense ssDNA [28], [39]. Altogether these results lead us to propose that the MEIOB protein binds ssDNAs generated during the 5′ to 3′ resection of DNA ends at meiotic DSB sites. We thus propose that MEIOB may be an RPA paralog specifically required for meiotic recombination. Although the data in the present study do not formally prove that the ssDNA binding activity of MEIOB is required for meiosis, this is an appealing hypothesis that needs pursuing in future studies.
Meiob−/− meiocytes are arrested during prophase I of meiosis at a stage resembling late zygotene stage and characterized by persistent DSBs and defects in chromosomal pairing. This incomplete synapsis is likely the consequence of the defective DSB repair as is observed in most mutants with impaired homologous recombination [12], [26]. In Meiob−/− meiocytes the number of DSBs appears unaffected as we observed the expected number of RPA foci [38] and the RAD51 and DMC1 foci were formed but did not persist. These observations point towards the lack of MEIOB primarily causing a defect in the process of recombination. Unfortunately, due to their embryonic lethality, most mutations that impair homologous recombination have not been characterized during mammalian meiosis. Notably, the lethality of Rad51−/− disallows the comparison of Meiob−/− to these murine meiocytes. Of note, we would like to point out that Meiob−/− phenotype is similar to that of Dmc1−/− mice [26], [27]. However in the Dmc1−/− spermatocytes, RAD51 persists in the pachytene-like stage in contrast to that observed in Meiob−/− mice ([26], Figure S12). Interestingly, in Meiob−/− pachytene-like cells, despite the absence of RAD51 and DMC1 foci, we observed abundant and brighter RPA foci and strong γH2AX staining indicating the presence of unrepaired DSBs and of ssDNA. We thus conclude that the transitory loading of RAD51 and DMC1 observed in Meiob−/− is insufficient to allow the completion of homology search, homolog alignment and SC formation. As RAD51 recombinase activity is believed to be inactivated during meiotic recombination, DMC1 is proposed to catalyze homology search and strand exchange of most meiotic recombination events [40]. Thus, in the Meiob mutant, DMC1 would appear to have been prevented from performing its role due to its precocious destabilization. To our knowledge, no murine meiotic mutant has yet presented such a phenotype with a premature disappearance of RAD51 and DMC1 before pachynema.
The formation of RAD51-DNA presynaptic filament is promoted by BRCA2 in mammals to overcome the inhibitory effect on the heterotrimeric RPA [41]. BRCA2 is known to bind both RAD51 and DMC1; Brca2−/−;Tg have an impaired number of RAD51 and DMC1 foci in leptotene and zygotene mouse spermatocytes [12]. Thus the Meiob−/− defect does not seem to involve BRCA2 in a general manner as one would then have also expected an earlier defect in the loading of RAD51 and DMC1. Moreover, the growing percentage of RAD51 and DMC1 foci that are stained for MEIOB at late zygotene and early pachytene stages support a later role of MEIOB on the activity of recombinases. In this line, biochemical findings raise the possibility that the maintenance of RAD51 presynaptic filament in vivo might involve some RAD51 accessory factors such as RAD54, HOP2-MND1 and recently the SWI5-SFR1 complex [42]–[44]. The spermatocyte defects observed in Rad54−/− and Hop2−/− mice differ from those observed in the Meiob mutant. Indeed meiotic recombination is only slightly affected in Rad54−/− mice which in turn show no major fertility defect [45]. In Hop2−/− mice, despite a failure of meiotic recombination, RAD51 and DMC1 foci accumulate and persist through the pachytene-like stage [46]. The role of SWI5-SFR1 has to date only been investigated in mouse embryonic stem cells [47], however genetic studies in yeast have found evidence for the SWI5-SFR1 yeast orthologs Mei5 and Sae3, being involved in meiotic recombination [48], thereby suggesting that the mammalian SWI5-SFR1 may also play a role in meiotic recombination. Interestingly, this complex is devoid of DNA-binding activity in mice [49] whereas their respective orthologs Mei5 and Sae3 in S. cerevisiae possess DNA-binding activity [50]. Furthermore, the expression of Mei5 and Sae3 is restricted to meiosis and mediates Dmc1 activity by enhancing its ability to form nucleofilaments on ssDNA [51]. One may reasonably speculate therefore that MEIOB may influence the stability of the RAD51-DMC1 filament in cooperation with SWI5-SFR1 or a meiotic counterpart. Such a hypothesis will of course require further investigation. Of note, SWI5-SFR1 and Mei5-Sae3 complexes directly interact with RAD51 [49], [50]. We were unable to detect a direct interaction between MEIOB and RAD51 (data not shown). However our experiments were performed in a heterologous system and with an overexpressed protein, thus we cannot formally allow us to exclude a direct interaction between MEIOB and RAD51. Finally, we propose that the presence of two ssDNA binding proteins, namely RPA and MEIOB, confers special properties to resected DNA allowing the proper stabilization of two proteins on the meiotic presynaptic filament: DMC1 and RAD51. Along this line, one may consider MEIOB as a new meiotic-specific mediator for RAD51/DMC1.
While meiosis recombination is a conserved process in eukaryotes, there are clear differences among organisms [52]. Interestingly, yeast, plants and C. elegans do not have a Meiob homolog (Figure 1A). One may therefore consider that in these organisms, either meiosis is achieved via slightly different mechanisms or the function of MEIOB is performed by other proteins. For instance, in plants, there are multiple copies of RPA1 (RPA1-like proteins) [53], [54] some of which may assume the role of MEIOB. In plants, RPA1-like proteins have been proposed to form part of trimeric RPA complexes alongside RPA2 and 3 sub-units [55]; to our knowledge no such additional sub-units are conserved among vertebrates and the analysis of transcriptomic data did not allow us to pin point any meiosis-specific additional RPA2-like protein in mice. We therefore propose that either MEIOB acts with the canonical RPA2 and 3 (thus replacing RPA1) or by its own and possibly with different partners. The only Meiob homolog that has been described previously is hold'em (hdm) in Drosophila [22] that is implicated in crossover formation. However hdm−/− fly and Meiob−/− mice present drastic divergence during meiosis suggesting that their function may have diverged during evolution. Indeed in hdm−/− flies, some crossovers still occur though with lower frequencies compared to wild-type fly and the DSBs are repaired though with a delay. Furthermore, hdm is also involved in DNA repair in somatic cells [22] while Meiob is almost exclusively expressed in meiotic germ cells. Joyce et al. also provided evidence that hdm physically interacts with Ercc1, a member of the exchange class of gene products, and proposed that the Ercc1/Mei9/Mus81/hdm complex functions in the meiotic recombination pathway to resolve DSB-repair intermediates into crossovers [22]. This function appears to be required later by comparison to the here-described recruitment of MEIOB to DSB sites. Furthermore, such an interaction is unlikely to explain the entirety of Meiob−/− phenotype described here. Indeed, in mice, Ercc1 deletion does not prevent synapsis or sex body formation and males are able to produce few spermatozoa albeit with DNA damage [56]. We thus conclude that, in mice, the function of MEIOB is not through an interaction with ERCC1 in late recombination nodules and that additional partners need to be identified. On the other hand, one could also argue that this may be due to the meiotic process itself having slightly different requirements in-between mammals and drosophila. Therefore, we cannot exclude an additional role of MEIOB in CO resolution. Such a role could be responsible of the late staining observed for MEIOB (Figures 2B and 2C). In this line we could speculate that MEIOB might target additional ssDNA sites such as the D-loop intermediate, as it is thought for RPA [57], or the second end of the DSB (i.e. the one not involved in the initial homology search), that are formed during the process of CO. The second end is released to be captured on the recombination intermediate to form a double holiday junction (dHJ). This would fit with our reports of rare cells displaying some amount of pairing, thus having performed the initial homology search (first end), and with the complete lack of MLH1 foci observed in Meiob−/− cells. Such a speculative proposition might also be fueled by the persistent colocalizsation of MEIOB with recombinases and will require additional studies. Putative MEIOB partners are under investigation and their identification should help define the precise function of MEIOB during meiotic recombination and the mechanisms allowing recombinase stabilization in this process.
All animal studies were conducted in accordance with the guidelines for the care and use of laboratory animals of the French Ministry of Agriculture. Mice were raised and mated, and fetal gonads were isolated as previously described [58], [59]. Mice used in this study were NMRI mice (Naval Maritime Research Institute), Dmc1 and Spo11 mutant mice, w/w mice, Oct4-GFP mice (that have been previously described, respectively [5], [26], [60], [61]) and Meiob mutant mice (see below).
The Meiob null allele was established at the Mouse Clinical Institute/Institut Clinique de la Souris (MCI/ICS), Illkirch, France (http://www.ics-mci.fr/). The targeting vector was constructed as follows: a 4.5 kb fragment encompassing exon 2 and part of the first intron, Red Fluorescent Protein (Rfp) cDNA, followed by a neomycin resistance cassette surrounded by two loxP sites and a 4 kb fragment encompassing exons 9 and 10 (Figure S1A). Meiob endogenous ATG in exon 2 was conserved to ensure Rfp cDNA expression and there was a stop codon at the end of Rfp cDNA (Figure S1A). The linearized construct was electroporated in BD10 C57BL/6J mouse ES cells. After selection, targeted clones were identified by PCR using external primers and further confirmed by Southern blot with a Neo probe, 5′ external probe and 3′ external probe (data not shown). Two positive ES clones were injected into 129/Sv blastocysts and the derived male chimaeras produced germ-line transmission. Meiob+/Rfp-loxPNeoloxP mice were crossed with mice carrying ubiquitous Cre in order to remove the Neo resistance cassette and generate the final Meiob null allele (Figure S1A). Mice were genotyped by PCR of tailed DNA using REDExtract-N-Amp Tissue PCR Kit (Sigma) according to the manufacturer's instructions. Sequences of primers used are shown in Table S2.
Male and female Meiob+/+, +/− & −/− mice were mated over the course of four months. Mating partners were inverted every two months. Vaginal plugs were regularly checked to verify that coïtum occurred normally. Births and pups were counted and referenced every day and pups were sacrificed. Results are expressed in cumulative numbers of pups per couple per ten days.
Human fetal material was provided by the Department of Obstetrics and Gynecology at the Antoine Béclère Hospital, (Clamart, France) following legally induced abortions in the first trimester of pregnancy and therapeutical termination of pregnancy in the second trimester. Human fetal gonads were harvested as previously described [62]. Our study was approved by the Biomedicine Agency (reference number PFS12-002), and all women gave their informed consent.
Rabbit polyclonal anti-MEIOB antibodies were generated and affinity purified by Eurogentec (Angers, France) using double X protocol with ADP TEA SRN LAR QGH T and IRE NKE TNV ADE IDS polypeptides.
Testicular single-cell suspensions and cell sorting were processed as previously described [63]. Dissociation was performed with two wild type adult testes resulting in the sorting of two hundred million testicular cells. Seven hundred thousand cells from each sorting gate were lysed and submitted to western blot analysis.
RNA extractions and RT-qPCR were performed as previously described [64]. Sequences of primers are shown in Table S2. β-actin mRNA was used as the endogenous reporter. Data are expressed as a percentage of the maximum mRNA expression unless otherwise stated in which case an external reference (F9 cells) was used to normalize the expression of numerous points.
Human Epithelial Kidney cells (HEK-293, ATCC) were cultivated in DMEM High Glucose (Gibco) containing 15% FSB (Gibco). Human MEIOB complete cDNA was inserted into a pCMV6-Entry plasmid (Origene cat# RC228391) containing c-Myc and Flag tags in the C-terminal domain of the protein. This plasmid was transfected into HEK-293 cells using Lipofectamine 2000 (Invitrogen) according to the manufacturer's instructions. Tagged-MEIOB was observed in both nuclear and cytoplasmic compartments of the transfected HEK-293 cells (Figure S5A).
Protein extracts were produced from cell lines or tissues under native conditions. HEK-293 cells were harvested, centrifuged 5 min at 500 g and then lysed in Cell Lysis Buffer (Cell signaling, Danvers, USA) complemented with 1 mM 2-mercaptoethanol and Complete protease inhibitor (Roche, Mannheim, Germany). For adult testis, the albuginea was first removed and testis was lysed in the same buffer using ceramic spheres and FastPrep-24 Instrument (MP Biomedicals) with two pulses of 20 seconds. Extracts were then gently sonicated (two pulses of 30 seconds), centrifuged 10 min at 14 000 g and supernatant was immediately used for functional applications. Human recombinant tagged-MEIOB protein was produced in a cell-free system. The TnT T7 Quick Coupled Transcription/Translation System (Promega) was used with 1 µg of the tagged-MEIOB plasmid in a 50 µl reaction according to the manufacturer's instructions.
Prior to gel migration, protein samples were supplemented with Laemmli buffer and resolved by 10% SDS-polyacrylamide gel electrophoresis (SDS-PAGE). Gels were electrophoretically transferred to polyvinylidene difluoride membranes (PVDF) (Amersham Biosciences, Buckinghamshire, England) before hybridization with appropriate primary antibodies and fluorescent dye coupled secondary antibodies (See Table S1). Images were acquired using Typhoon 9400 scanner (Amersham Biosciences) and quantified with ImageJ software [65].
Histological sections, germ cell counting and immunohistochemical stainings using appropriate antibodies (Table S1) were performed as previously described [59].
Testes used for spermatocyte spread preparation were harvested from mice aged approximately 2 to 6 months. Ovaries used for the preparation of chromosome spreads were harvested in 15.5 and 17.5 dpc embryos and 1 dpp pups. Three types of meiocyte chromosome preparations were used. For γH2AX, SYCP1 and RPA2 stainings, chromosome spreads were prepared according to the drying-down technique [66] with minor, previously described modifications [67]. For RAD51, DMC1 and ATR stainings, chromosome spreads were prepared as previously described [68]. This later protocol was also used for MLH1 staining in oocytes and the former protocol was used for MLH1 staining in spermatocytes. For MEIOB staining, the two above protocols were used and an additional spermatocyte surface-spread technique was applied as follows. Spermatocytes were dispersed in an 80 mM NaCl drop, attached to glass slides, fixed with a 4% paraformaldehyde 0.03% sodium dodecyl sulfate solution and washed in 0.4% Photo-Flo 200 (Sigma). RAD51 and DMC1 antibodies were previously published [69]. The specificity of the DMC1 antibody was further validated using Dmc1−/− spermatocytes and no signal was retrieved (data not shown). Due to the lack of Rad51−/− spermatocytes (i.e. Rad51 deletion is lethal), no such control was performed for the RAD51 antibodies and one cannot formally exclude that these might crossreact.
Spermatocyte and oocyte preparations were washed with 0.4% Photoflo 200 (Sigma) and slides were air-dried 15 minutes before permeabilizing and blocking. They were then incubated overnight with the appropriate primary antibodies in blocking solution at room temperature, followed by 1 hour incubation with secondary antibodies at 37°C (see Table S1 for antibodies). Slides were mounted with Vectashield DAPI medium (Vector Laboratories). Imaging was performed using a AX70 epifluorescent microscope (Olympus) equipped with a charge-coupled camera (Roper Scientific) and IPlab software (Scanalytics) or with a DM5500 B epifluorescent microscope (Leica Microsystems) equipped with a CoolSNAP HQ2 camera (Photometrics) and Leica software (Metamorph). Images were processed and specific structures were quantified with ImageJ software (Cell Counter and Foci Picker3D programs). For quantification of foci colocalization, foci overlapping for the most part were considered as colocalized without restriction to foci that share the same centroid (e.g. immediately adjacent foci were not considered as colocalized).
HPLH purified biotinylated oligonucleotides were used for the DNA pull down assays: ss30-mer 2: 5′-GAT CTC AGC GAT TCA CAC GCG TCC TAA CTC G-3′-BiotinTEG, ss60-mer: 5′-GAT CTG CAC GAC GCA CAC CGG ACG TAT CTG CTA TCG CTC ATG TCA ACC GCT CAA GCT GC-3′-Biotin TEG (Eurogentec). Double strand hybridizations were performed in 50 mM NaCl, 25 mM Tris-HCl, pH 7.5 buffer with complementary sequences at molecular equivalence by a denaturing step (3 minutes 94°C) and a slow progressive return to room temperature. 0.2 pMol of DNA was immobilized onto 1 µg Dynabeads M-280 Streptavidin (Dynal) following the protocol supplied by the manufacturer. Protein extracts were pre-incubated on ice for 10 minutes in modified DBB (DBB with 25 mM Tris-HCl, 1 mM EDTA plus 5 mg/ml BSA for the cell-free protein assays or plus 10 µg/ml BSA for assays with total protein extracts) before addition of 500 µg Dynabeads with immobilized ss- or ds-DNA probes. DNA binding assays were conducted either with 3 µl of the tagged-MEIOB cell-free productions, with 600 µg of protein from HEK-293 cells transfected with hMEIOB (Figure S5A) or with 3 mg testis protein extracts. The DNA-protein mixture was incubated for 1 hour at 4°C under gentle agitation. After magnetic separation, the beads were washed three times in 500 µl binding buffer without BSA, before being washed once in 500 µl rinsing buffer (modified DBB with 150 mM NaCl). Elution of DNA binding proteins was performed by resuspending the beads in 20 µl Laemmli buffer and boiling the samples for 5 minutes before magnetic separation of the beads and western blotting of the samples.
HEK-293 cells were harvested 48 h after transfection of the tagged-MEIOB plasmid. Three mg of total protein extract were diluted 1/5 in DNA-Binding Buffer (DBB: 50 mM Tris-HCl, 100 mM NaCl, 10% (w/v) glycerol, Complete Protease Inhibitor, 1 mM 2-mercaptoethanol, pH 7.4), 1 ml of which was loaded in Poly-Prep columns (Biorad), previously loaded with 500 µl of ss-DNA-cellulose (Amersham Biosciences). Protein extract was incubated 90 min at 4°C under gentle agitation. Columns were washed with forty bed volumes of ss-DBB and the last 1 ml of wash was kept to ensure that complete elution occurred (0.1M LW). Columns were then washed with elution buffer containing increasing concentrations of NaCl: 0.25, 0.5, 0.75, 1, 2M. For each step columns were washed with ten bed volumes, the first 1 ml (Figure 3C) and last 1 ml were kept (Figure S5C). Eluted fractions were concentrated using Nanosep 3KDa (Pall) according to the manufacturer's instructions. Salt concentration was then equalized with ss-DBB. Concentrated protein fractions were recovered in Laemmli buffer and subjected to western blot with the appropriate antibodies (Table S1).
|
10.1371/journal.pcbi.1005029 | A Monte Carlo Study of Knots in Long Double-Stranded DNA Chains | We determine knotting probabilities and typical sizes of knots in double-stranded DNA for chains of up to half a million base pairs with computer simulations of a coarse-grained bead-stick model: Single trefoil knots and composite knots which include at least one trefoil as a prime factor are shown to be common in DNA chains exceeding 250,000 base pairs, assuming physiologically relevant salt conditions. The analysis is motivated by the emergence of DNA nanopore sequencing technology, as knots are a potential cause of erroneous nucleotide reads in nanopore sequencing devices and may severely limit read lengths in the foreseeable future. Even though our coarse-grained model is only based on experimental knotting probabilities of short DNA strands, it reproduces the correct persistence length of DNA. This indicates that knots are not only a fine gauge for structural properties, but a promising tool for the design of polymer models.
| We develop a coarse-grained model of double-stranded DNA which is solely based on experimentally determined knotting probabilities of short DNA strands. Our analysis is motivated by the emergence of DNA nanopore sequencing technology. The main advantage of nanopore sequencing in comparison to next-generation devices is its capability to sequence rather long DNA strands in a single run, currently up to ≈10,000 base pairs. Unfortunately, long DNA strands easily self-entangle into knotted conformations, and sequencing knotted DNA with nanopores may be subject to error. In our manuscript, the typical extent and likelihood of DNA knots is computed for DNA chains of up to half a million base pairs, and we estimate the abundance of complex and composite knots in relation to DNA length. Our analysis indicates that DNA knots may be a formidable roadblock for the development of devices which support substantially longer read lengths. We also show that structural properties of DNA, like its resistance to bending, are intimately linked to the molecule's tendency to form knots. We demonstrate how this connection can be utilized to introduce mathematical models of DNA which account for the molecule's overall statistical properties.
| Entanglements in molecular cords like polymers or semi-flexible biopolymers like double-stranded DNA (dsDNA) often lead to knotted chain conformations. The significance of DNA knots has been discussed in biological contexts [1, 2, 3], as well as in technological settings: Recent studies [4, 5] investigated how knots in DNA change the translocation of long DNA molecules through nanopores. DNA translocation dynamics are of practical importance in the context of DNA nanopore sequencing, where a single molecule of either single- or double-stranded DNA is electrophoretically driven through a nano-scale pore across an impermeable thin membrane. The DNA's translocation through the nanopore alters the amplitude of the electrochemical current by perturbing the charge transport along the pore [6, 7]. In a common approach, the chain's nucleotide sequence is read by directly analysing the time-dependence of the current signal [8].
Mathematically, knots are only well-defined in closed curves. Nevertheless, a physical definition is often applied to open strings [9, 10]: Ends are connected in a systematic manner before the knot type is analysed (see the Methods section). Knots are categorised by their minimal number of crossings in planar projection. The simplest (non-trivial) knot is the so-called trefoil, which has three crossings and can be obtained by closing an overhand knot. Intriguingly, some of our intuitive understanding of macroscopic knots carries over to the nano-scale. About 50 years ago, Frisch and Wasserman conjectured [11] that any molecular cord will eventually be knotted as the chain length increases. This conjecture was later proven for certain classes of lattice polygons [12], but does not state how polymer length and polymer properties are related to knotting probability. Polymers in globular states [13–15, 9] and DNA in viral capsids [16, 17, 3, 18] are known to be highly knotted, whereas unconstrained polymers [15, 19, 9] or DNA in good solvent conditions are less prone to self-entanglements. The probability of knotting in dsDNA was first measured in the early 1990s by gel electrophoresis [20, 21]. Strand sizes of up to ≈10,000 base pairs were considered and depending on salt conditions, the probability of observing knots was at most a few percent. Consequently, self-entanglements and knots have mostly been ignored in the context of nanopore sequencing. Reference [20] also describes for the first time a coarse-grained model for DNA which is partly based on knotting probabilities. Very recently, a ground breaking study (Plesa et al., Nature Nanotechnology in press) has pushed these boundaries even further and estimated knotting probabilities for significantly larger strands in high salt concentrations by analysing translocation events in solid-state nanopores.
The focus of this work is two-fold. First, we introduce a coarse-grained model of dsDNA which is solely based on experimentally determined knotting probabilities. This model is then used to analyse the statistics of DNA knotting and determine typical knot sizes, motivated by the emergence of DNA nanopore sequencing technology. This analysis naturally precedes further experimental or theoretical investigations designed to address the problem of how to avoid or control knots in DNA nanopore sequencing devices. Employing the aforementioned coarse-grained model, the abundance and size of knots in chains exceeding half a million base pairs is studied for physiologically relevant salt concentrations of 0.15M NaCl. Although it is well known that knots are likely to form in long polymer chains, no quantitative estimation of DNA knotting probabilities is available for chains beyond ≈50,000 base pairs [20, 21, 22]. To estimate knotting probabilities for DNA chains of up to half a million base pairs, model parameters are chosen so that predicted knotting probabilities of short DNA chains match knotting probabilities from electrophoresis experiments: DNA is modelled as a semi-flexible chain of impermeable spherical beads of diameter d≈4.5nm, which corresponds to ≈13 base pairs (bp). An intrinsic stiffness controls the bending of the chain and the effective DNA diameter d subsumes excluded volume effects as well as screened electrostatic interactions [20]. The simplicity of the model is the key to deriving optimal model parameters and obtaining statistical estimates for chain lengths which are relevant in the context of future applications of nanopore sequencing technology. Mathematical details of the model, as well as the rationale for choosing particular values of the model parameters are discussed in the Methods section.
Intriguingly, the DNA model predicts a persistence length of ≈50nm in excellent agreement with experimental findings [23]. We stress that experimentally measured knotting probabilities of short dsDNA chains are the only input to our model (see the Methods section). This observation is non-trivial, as it implies that metric properties of DNA can be predicted from purely (non-metric) topological properties, which has never been demonstrated before. Hence, basing a simple real chain model with stiffness on knotting probabilities allows for the description of physical properties of the chain. The model is then employed to predict knotting probabilities of long dsDNA chains as well as typical sizes of DNA knots. This is to be contrasted with the approach in [20], which in addition to knotting probabilities also requires the persistence length of DNA as an input parameter. Although the dsDNA model in [20] is coarser, and introduces a sequence of impermeable cylinders to model DNA, both approaches derive similar values for the effective diameter of DNA. In comparison, simple random walk models of DNA which lack excluded volume interactions [24] tend to overestimate the occurrence of knots. E.g., our real chain model predicts that a chain of 150,000 base pairs is knotted in roughly 40% of all cases, whereas random walks of 500 segments (assuming a Kuhn length of 300 base pairs) exhibit knots in ≈80% of all cases if the same closure is applied.
In Fig 1, a typical trefoil (light green) in a DNA chain of ≈13,000bp (represented by a coarse-grained model chain of N = 1,000 beads) is displayed in relation to characteristic nanopore sizes. In Fig 2 computed probabilities for observing knots under physiological salt conditions are shown for DNA strands of up to ≈525,000bp (N = 40,000, S1 Table, supporting information). At this length, ≈88% of all chains already contain at least one knot. Remarkably, more than ≈68% contain complex knots with more than three crossings or composite knots. The transition from a mostly unknotted to a mostly knotted ensemble of DNA chains is indicated by the base pair count B0 at which the probability of obtaining an unknotted conformation is 1/e≈0.37 [25]. B0≈250,000bp(N≈19,000 beads) also characterizes the regime where knots with higher crossing number (≥4) become more abundant than trefoil knots. Intriguingly, Fig 3 indicates that more complex entanglements are mainly made up of composite knots which include trefoil knots as prime factors. Beyond 300,000bp (N≈22,850) the triple trefoil knot and even the 31#41 composite knot occur more often than the figure-eight knot 41, and formation of prime knots with more than four crossings is very unlikely (S1 Fig). Hence, probabilities of composite knots are not mere product probabilities of the constituent prime factors, reflecting the non-local structure of emerging polymer entanglements. Note that even though the Alexander polynomials of the analysed composite knots (Fig 3) are identical to the Alexander polynomials of specific prime knots with eight crossings (e.g. 31#31 and 820 share the same polynomial), the influence of prime knots with eight crossings on observed knotting probabilities is expected to be negligible, since all prime knots with seven crossings already have vanishingly small probabilities.
In addition to estimating the mere abundance of knotted DNA chains, typical knot sizes can be determined as well: To identify the knotted region, a chain is trimmed from both ends until the remaining part becomes unknotted (see the Methods section). In Fig 4, the size of a trefoil knot refers to the contour length of the knotted region, and its distribution is shown for various DNA lengths. Intriguingly, the most likely size of a trefoil knot is around 3,000bp (N≈230), independent of strand size. This observation as well as computed distributions of trefoil contour lengths are in excellent agreement with recent simulation results [22]. In [22], typical knot sizes have been determined for a similar coarse-grained model, and for a range of model parameters which can be mapped onto dsDNA at various salt concentrations. Note that DNA models based on random walks predict smaller knots: The maximum of the size distribution is at around 7 segments, corresponding to 2100 base pairs [26]. As opposed to its most likely value, the expectation value of the knot size increases with DNA length (Fig 4). To estimate a trefoil's geometrical extent, the radius of gyration ⟨Rg2⟩ of the trefoil contour is computed, and its diameter is then taken to be 2⋅⟨Rg2⟩. The inset in Fig 4 displays the distribution of trefoil diameters for a DNA chain of ≈13,000bp (N = 1000), and the most likely value is ≈200nm. The trefoil displayed in Fig 1 (light green) consists of ≈4,000 base pairs (N≈300) and has a diameter ≈160nm. It may thus be regarded as a typical representative, even if longer DNA chains are considered.
As our DNA model reproduces the correct persistence length of DNA by adjusting model parameters to match experimentally measured knotting probabilities of short dsDNA chains, it can be inferred that knots may be employed as a tool in polymer physics: Given sufficient experimental data on knot statistics of a particular polymer species, the parameters of any suitable polymer model may be chosen so that theoretical knotting probabilities match the experimental ones. As knot statistics are a means to quantify global topological properties of polymer chains, the resulting model is expected to thoroughly reproduce polymer entanglements. Whether such a procedure leads to a good imitation of polymer behaviour and physically consistent results depends, among other things, on the selection of a proper set of adjustable model parameters.
The successful derivation of the persistence length of dsDNA indicates that at least for physiological salt concentrations, our DNA model is capable of representing basic physical properties of DNA strands. Other salt concentrations will be tested in future investigations.
Simulations of this model indicate that DNA molecules beyond 250,000 base pairs are likely knotted and contain composite knots. Our analysis of knotting in dsDNA for physiologically relevant salt concentrations of 0.15M NaCl estimates lower bounds for DNA knotting in nanopore sequencing devices which adapt double-stranded DNA for sequencing: Most nanopore setups operate in high salt [27], which increases the likelihood of knotted conformations [20]. Furthermore, while most nanopore sequencing techniques keep DNA in its single-stranded form [28], very successful alternatives [29, 30] approach the problem by ligating adaptors to the ends of dsDNA, which subsequently help to thread the dsDNA into the pore, and afterwards control the translocation of a single DNA strand.
More specifically, it was demonstrated [31, 29] that DNA polymerase can slow down and control the transport of dsDNA through a biological nanopore: A polymerase molecule is anchored at the entrance of the nanopore, which splits the double-stranded DNA and drives the translocation of a single DNA strand. With this measure, sequence information can be obtained more reliably, at least for DNA strands of up to ≈4,500 base pairs [30].
If translocation is driven by DNA polymerase, every step in sequential movement takes several milliseconds [31]. This timespan should be long enough to equilibrate the knotted region or at least a substantial part of it, whereas diffusion along the contour of the DNA can probably be neglected [32]. The momentary size of a threaded trefoil can thus be estimated from ensemble statistics to be 2⋅⟨Rg2⟩ (see the inset in Fig 4), and pulling a trefoil knot through the pore should not tighten it mechanically. However, knots with high crossing number or composite knots may behave differently, in which case 2⋅⟨Rg2⟩ may not be a proper measure of knot size. If a knot is located close to the nanopore, the ion flux may be perturbed, and the magnitude of the perturbation is probably related to its geometrical extent.
Therefore, knotty problems may even occur for chain lengths which are within reach of current technology. The presence of complex and composite knots in long DNA chains might lead to a blockage of the nanopore's entrance. Though a blockage can potentially be avoided for simple knots as discussed in [4, 5], knots of any crossing number might significantly alter the ion transport along the pore or even the translocation dynamics of the DNA. As soon as knots become abundant in long DNA chains, interpretation of the current signal and discrimination of individual nucleotides may be prone to errors, even if DNA molecules can still be threaded through nanopores at reasonable rates, since the fingerprint of the DNA's nucleotide sequence sensitively depends on the resulting ion flux and DNA translocation dynamics. It is a very hard problem to ascertain how the time-dependent electrochemical current changes in the presence of a DNA knot. A quantitative analysis clearly goes beyond the scope of this paper.
We hope that our work will stimulate further experimental and theoretical investigations of the aforementioned issues. The vision that one day, a nanopore sequencing device could read a significant portion of a chromosome from just a single DNA molecule will have to include an idea of how to avoid knots in long DNA chains [33, 34].
We employ a discrete worm-like chain (Kratky-Porod) model [35] with hard sphere interactions between beads and fixed bond lengths. The bending potential is given by
U/kBT=−g∑icos(θi)
with the θi, i = 1,…,N – 1, being the angles between adjacent bond vectors. The computational model is fully determined in terms of the number of beads N and stiffness parameter g ≥ 0. Knotting probabilities of Kratky-Porod chains with excluded volume interactions depend on g in a non-trivial manner [36], whereas the knotting of ideal Kratky-Porod chains monotonously decreases with stiffness. Screened electrostatic interactions are absorbed in an effective hard sphere diameter d, which depends on the salt concentration [20]. In dsDNA the distance between adjacent base pairs is 0.34nm. A DNA strand of B base pairs is thus modelled as a chain of N = B · 0.34nm/d beads.
In previous experimental studies [20, 21], DNA knotting probabilities have been obtained by gel electrophoresis for dsDNA molecules with a length of 5.6kbp, 8.6kbp [21] and 10kbp [20] for different salt concentrations. Even though DNA had been cyclized before the knot type was determined, knots formed when DNA was still in a linear state. Therefore, experimental knotting probabilities are more likely to reflect probabilities in linear DNA. Either way, probabilities for knots in loops and knots in open chains are very similar, as has been demonstrated for random walks in [10] and for self-avoiding chains in [15]. DNA lengths from these experiments can be converted to chain lengths of the computational model for a given d. To obtain an optimal set of parameters (g,d) to model dsDNA under physiologically relevant salt concentrations of 0.15M NaCl, knotting probabilities are computed for 16 × 16 = 256 points of an equispaced grid with boundary points N = 250, 1,000, g = 6.5, 14. Each chain is simulated with Markov chain Monte Carlo (MCMC) methods, applying generalized MOS (inversion, reflection and interchange) [37], crank-shaft and pivot moves [38]. Typical MCMC errors are two orders of magnitude smaller than corresponding experimental errors [20, 21] and neglected in subsequent analysis. With a non-parametric regression in R [39] (library method loess), a surface is fitted to the grid of simulated knotting probabilities. The comparison of the interpolated knotting probabilities with the experimental results for a salt concentration of 0.15M NaCl defines a smooth (least squares) error function E(g,d), which is minimized with the aid of the Levenberg-Marquardt algorithm (R [39] library method nls.lm): The minimum (g,d) of the error function E is interpreted as an optimal parameter set, yielding g≈11.673 and d≈4.465. Even though the physical diameter D of dsDNA is only about 2nm, the effective diameter d also accounts for the influence of screened Coulomb interactions in addition to excluded volume. Thus, d is in general larger than D and would approach D for high salt concentrations (for which electrostatic interactions of dsDNA are completely screened) [20]. Production runs (Figs 2–4 and S1) employ this parameter set to predict knotting probabilities for dsDNA strands of up to ≈525,000 base pairs (computed knotting probabilities and MCMC errors are documented in S1 Table, S2 Table and S3 Table). For the (ideal) Kratky-Porod chain, the functional dependence of the persistence length lp(g,d) is given by lp(g,d) = −d/ln (coth(g) − 1/g) [40], yielding ≈49.85nm.
As a topological knot is necessarily a closed space curve, the open DNA chain has to be closed prior to knot detection. Here, we join the ends of the chain by first extending them away from the centre of mass, and then connecting them by the legs of a triangle, so that the additionally constructed line segments do not interfere with the original chain volume [41]. For each closed curve, the Alexander polynomial is evaluated and used to identify the knot type [10]. Note that in principle, the implementation of the closure may create additional entanglements and knots. In practice, this effect only plays a minor role when calculating ensemble averages. Different closures result in almost identical knotting probabilities as was demonstrated for random walks in [10] and for self-avoiding chains in [15]. To determine the knotted region of a trefoil knot as shown in Fig 4, a chain is first trimmed bead by bead from one end and subsequently analysed until the remaining part becomes unknotted. The same procedure is then applied starting from the other terminus. The remaining beads define the contour length of the knot, and its radius of gyration ⟨Rg2⟩ is computed to describe the knot’s physical extent as 2⋅⟨Rg2⟩.
|
10.1371/journal.pgen.0030091 | A Transcriptional Program Mediating Entry into Cellular Quiescence | The balance of quiescence and cell division is critical for tissue homeostasis and organismal health. Serum stimulation of fibroblasts is well studied as a classic model of entry into the cell division cycle, but the induction of cellular quiescence, such as by serum deprivation (SD), is much less understood. Here we show that SS and SD activate distinct early transcriptional responses genome-wide that converge on a late symmetric transcriptional program. Several serum deprivation early response genes (SDERGs), including the putative tumor suppressor genes SALL2 and MXI1, are required for cessation of DNA synthesis in response to SD and induction of additional SD genes. SDERGs are coordinately repressed in many types of human cancers compared to their normal counterparts, and repression of SDERGs predicts increased risk of cancer progression and death in human breast cancers. These results identify a gene expression program uniquely responsive to loss of growth factor signaling; members of SDERGs may constitute novel growth inhibitors that prevent cancer.
| Cells constantly sense their environment to decide whether to divide. Many genes that control the entry into cell division are known, and their excessive activation may cause cancer. In contrast, the way that cells cease to divide was thought to be a passive process, where signals for cell division gradually decay. In this study, the authors show that the decision to cease cell division and enter a state termed quiescence is also an active process. By monitoring the changes in activity over all genes, the authors identify a set of genes that respond specifically to decrements of external stimuli and ensure cessation of cell division. These genes act as brakes to prevent excessive cell division, and their inactivity is characteristic in many human cancers, particular those that progress to life threatening disease.
| Quiescence, also termed G0, is defined as reversible cell cycle arrest where cells are poised to re-enter the cell cycle. Most eukaryotic cells spend the majority of their lifespan in the state of quiescence. In response to injury or specific extracellular stimuli, many types of somatic cells can quickly leave the quiescent state and enter the cell division cycle. For instance, in the skin, dermal fibroblasts and hair follicle stem cells are for the most part quiescent [1,2]. Injury to the skin stimulates fibroblasts and epidermal stem cells to rapidly proliferate; once tissue repair has been accomplished, the cells exit the cell cycle and reenter quiescence. Similarly, memory lymphocytes are quiescent as they circulate and survey the body, dividing only when stimulated by cognate antigenic stimuli to mount an immune response [3]. In addition to the absence of cell division, quiescent cells exhibit systematic differences in their metabolism and propensity for differentiation, which may help to ensure the reversibility of quiescence [4]. The ubiquity of quiescence as a central feature of cell life suggests that its regulation may be critical to normal development, degenerative diseases, and cancer [1,3,4].
Serum, the soluble fraction of clotted blood, is an important mitogenic signal in wound healing and tissue homeostasis. Many key genes involved in cell cycle entry were initially identified by their unique temporal patterns of expression in response to serum stimulation (SS) and are dysregulated in cancer [5]. In addition to cell cycle entry, serum induces a transcriptional program activating many aspects of wound healing [6]. This wound response program is recapitulated in many human cancers and is a strong predictor of tumor progression for these cancers [7,8]. While much is known about the signal transduction pathways, transcription factors, and immediate early genes that mediate the exit from quiescence and entry to the cell division cycle [5,6,9,10], comparatively much less is known about the mechanisms by which cells enter the quiescent state. Growth factor deprivation, contact inhibition, and loss of adhesion can each induce a shared set of quiescence genes [4], indicating the potential existence of multiple pathways to quiescence. Several tumor suppressor genes, such as Rb and PTEN, are required for quiescence maintenance in low serum conditions [11–13]. Yamamoto and colleagues have identified a set of antiproliferative genes whose repression requires ongoing activity of the mitogen-responsive kinase ERK during cell cycle progression [10]. The induction of these genes during quiescence is therefore simply a consequence of the absence of mitogen-induced signaling. If this mode of balanced regulation were generally applicable, then one might predict a symmetric network of gene regulation during quiescence entry and exit. Alternatively, an inducer of quiescence may engage a unique transcriptional program that is not regulated by cell cycle entry. Such quiescence entry-specific genes may represent novel growth inhibitors that link extracellular stimuli to the physiologic state of quiescence.
In this report, we characterize the genomic expression program of serum deprivation (SD) in fibroblasts and identify the predominance of asymmetric regulation in quiescence induction. We identify two putative tumor suppressor genes, SALL2 and MXI1, as key regulators of the serum deprivation early response genes (SDERGs) and demonstrate the roles of the SDERGs in cell cycle exit and human cancer progression.
To understand the genomic program of entry into quiescence, we characterized the temporal pattern of the genome-wide transcriptional profiles of fibroblasts in response to SD. We employed the same diploid fibroblast culture and experimental time points that we previously used to delineate a detailed transcriptional response to SS [7], thereby enabling a systematic comparison of entry and exit of quiescence. Human foreskin fibroblasts were grown in media containing 10% fetal bovine serum (FBS) for 48 hours, and following switch to low serum media containing 0.1% FBS, harvested at 15 time points ranging from 15 minutes to 48 hours after SD. Total RNA was extracted, amplified, and hybridized along with human universal reference RNA onto human cDNA microarrays containing ~43,000 elements, representing ~23,000 unique genes.
Comparison of the temporal expression profiles of SD and SS revealed two dynamic programs with marked asymmetry in the early responses (Figure 1A). For each gene, we quantified the similarity of its pattern of expression during SD and SS by a Pearson correlation. (A correlation of −1 indicates exact opposite pattern; a correlation of 0 indicates no relationship; a correlation of +1 indicates identical pattern.) Strikingly, genes with an induction or repression onset after eight hours of treatment showed symmetric regulation by SS and SD with a Pearson correlation of −0.80 (i.e., genes induced by SS are repressed by SD and vice versa). In contrast, genes regulated within the first three hours of SS or SD were asymmetrically governed by these two stimuli (Pearson correlation −0.2 to +0.2). Genes that were regulated within three to eight hours of the treatments had an intermediate level of symmetry. These results thus suggest an asymmetric regulation of quiescence entry and exit with distinct sets of early response genes.
To gain a higher order view of the transcriptional programs of SS and SD, we next identified the functional groups and upstream cis-regulatory sequences of genes that are regulated in expression by these two stimuli (Figure 1B and 1C). Using the module map method [14], we identified for each array the coordinate induction or repression of 1,735 gene sets, each defined as a group of genes encoding proteins possessing a shared biological function, biochemical process, or subcellular localization by gene ontology (p < 0.05; false discovery rate [FDR] < 0.05) (Figure 1B) [15]. This higher order view confirmed that many functions known to be regulated by serum [6], such as cell proliferation, RNA metabolism, and sterol synthesis, are symmetrically regulated by SS and SD in the late phase of each program (Figure 1B; Figure S1). In contrast, while coordinate induction of genes encoding transcription factors and signaling proteins characterized the early response of SS, the genes that were regulated in the early response of SD were enriched for functions in immune response, redox, and extracellular matrix metabolism (Figure 1B). The paucity of gene ontology terms describing functional groups characterizing early SD may reflect the current scarcity of knowledge about this cellular state.
Transcriptional regulation is mediated in large part by binding of trans-acting factors to cis-regulatory elements upstream of genes. To better understand the regulation of SS and SD, we therefore mapped the genome-wide occurrence of 175 phylogenetically conserved cis-regulatory motifs [16] in four kilobases surrounding transcriptional start sites genome-wide. For each array, we identified cis-regulatory motifs that are significantly enriched in the genes that are induced or repressed, yielding a map of the regulatory motifs active in quiescence entry and exit (Figure 1C). This unbiased approach highlighted many known transcription factors that play key roles in this process and again reveals the marked asymmetry in the early response of SS and SD (Figure 1C). For instance, the gene set defined by enrichment of motifs for E2F and DP transcription factors contained many cell cycle genes and was symmetrically regulated late in the response by SS and SD (Figure 1C; Figure S1). In contrast, gene sets defined by the motif of AP-1, MEF2, or NF-κB were induced early in response to SS but were not substantially regulated by SD. Once again, we noted a relative paucity of cis motifs that identify early response genes to SD. This result may reflect incomplete information of the relevant transcription factors and their cognate cis motifs or may reflect additional post-transcriptional mechanisms in regulating mRNA levels in SD.
To begin to understand the asymmetric regulation of quiescence entry and exit, we focused on genes that are induced early in response to either SS or SD. Approximately half of the well-known early response genes to SS, such as EGR1, CYR61, and GADD45B, were correspondingly repressed by SD, albeit with lower amplitude and delayed kinetics that peak at approximately three hours after SD (Figure 2A). In contrast, other key SS early response genes, such as FOS, JUNB, and MYC, are transiently induced by SS and showed little change in transcript level in SD (Figure 2A). Conversely, a group of 135 genes defined by early induction in response to SD, which we term SDERGs, showed immediate induction within the first three hours after SD, but demonstrated substantially less or delayed regulation by SS (Figure 2B and 2C; Table S1). Some SDERGs, including SALL2, MXI1, and TNKSBP1, are induced in a sustained manner by SD while other SDERGs, such as SPRY4 and SMAD7, are induced by SD in a transient manner. Intriguingly, SALL2 and MXI1 are both putative tumor suppressors that can suppress cell growth when overexpressed [17,18]. SALL2 encodes a zinc finger transcription factor and is a homolog of Drosophila homeotic gene Spalt. Human SALL2 resides in a chromosomal region frequently deleted in ovarian cancers, and SALL2 protein is a binding target of the oncogenic large T-antigen from polyoma virus [19]. MXI1 is a member of the MAD family of potent antagonists of MYC oncoproteins [20]. MXI1 resides on a locus in human Chromosome 10 that is deleted in several types of human cancers, including prostate cancer, and deletion of MXI1 in mouse leads to a cancer-prone phenotype [18]. We noted that several SDERGs are well-known interferon-induced genes, such as STAT1, ILIR1, BDKRB2, and PLSCR1; several additional SDERGs, such as JUND, IFIT2, and G1P2 are predicted by our cis-regulatory map to contain a motif for the interferon regulatory factor (IRF) family of transcription factors. A likely candidate is IRF1 because IRF1 can induce several of these genes, is itself induced by SD, and possesses antiproliferative properties [21,22]. The asymmetric regulation of early response genes suggests that the transition to a quiescent state may be enforced by employing signaling pathways unique to SD.
To address the functional role of the SDERGs, we used RNA interference to examine the requirement of specific SDERGs for cell cycle exit. We selected four candidate genes (SALL2, MXI1, IRF1, and TNKS1BP1) that encode transcription factors or signaling proteins that may regulate quiescence induction. TNKS1BP1, a putative telomere binding protein, was included because of the reported roles of telomerase in enhancing S-phase progression [23,24]. We transfected primary human foreskin fibroblasts grown in high serum (10% FBS) media with silent interfering RNA (siRNA) pools corresponding to each of four candidate genes or a control siRNA targeting GFP. The cells were switched to 0.1% serum media 72 hours after transfection to induce quiescence, and DNA synthesis was measured 16 hours later by 5-bromo-2′-deoxyuridine (BrdU) incorporation. Reverse transcription-PCR confirmed decreased expression of the target mRNAs (Figure S2). In control cells treated with siRNAs targeting GFP, SD lowered the percentage of BrdU+ cells from 65% to approximately 25%, indicating induction of quiescence and efficient cell cycle exit (Figure 3A and 3B). Strikingly, cells treated with siRNAs targeting SALL2 doubled the number of BrdU+ cells after SD (p < 0.0001), while siRNAs targeting MXI1 showed modest but consistent increase in BrdU+ cells (p < 0.02). siRNAs to IRF1 or TNKS1BP1 did not significantly affect quiescence induction under the conditions tested. To test the potential functional relationships between SALL2, MXI1, and IRF1, we treated cells with pair-wise combinations of siRNAs. Silencing of IRF1 strongly cooperated with silencing of MXI1, but silencing of neither IRF1 or MXI1 cooperated with silencing of SALL2 to prevent cell cycle exit. Fluorescence-activated cell sorting analysis of DNA content confirmed that depletion of SALL2 or MXI1 blocked the ability of cells to arrest in G0–G1 after growth factor deprivation and instead led to inappropriate progression through S and G2/M phases of the cell cycle (Figure 3C). These results suggest that several of the SDERGs identified by our microarray screen are required for entry to cellular quiescence and that SALL2 and MXI1 may trigger different pathways to enforce quiescence.
To further delineate the mechanisms of SALL2 and MXI1 action, we identified genes that required SALL2 and MXI1 for proper regulation during quiescence induction. RNA from cells transfected with silent interfering (si)SALL2 and siMXI1 was extracted, amplified, and compared with RNA from cells treated with siGFP on cDNA microarrays.
Genes whose expression levels were consistently changed by loss of SALL2 or MXI1 were identified, and their temporal regulation by SS and SD were systematically organized by hierarchical clustering. We observed three main patterns of gene regulation. First, both SALL2 and MXI1 are individually required for the induction of a cluster of SD middle response genes (Figure 4, cluster 1). After knockdown of either SALL2 or MXI1 in SD, these genes reverted to a pattern of expression more closely resembling their normal behavior in SS than in SD. Second, in contrast to this shared role in SD gene induction, SALL2 and MXI1 acted to repress mutually exclusive sets of SD-repressed genes (Figure 4, cluster 3). Third, SALL2 appears to have a unique role in limiting expression of a set of middle response genes to SD, as their expression became super-induced when SALL2 was silenced (Figure 4, cluster 2). These results confirm the distinct roles of SALL2 and MXI1 in quiescence induction and suggest multiple roles for SALL2 in gene regulation throughout quiescence entry and maintenance.
Having discovered the SDERGs as a set of 135 genes specifically induced when fibroblasts enter quiescence, we next tested whether the SDERGs might have broad roles in growth inhibition. We reasoned that if SDERGs were generally required to induce cell quiescence, then SDERGs might be coordinately repressed in conditions of excessive cell proliferation, such as in cancer. We therefore interrogated a compendium of 1,973 microarrays representing 22 human tumor types to search for enriched coregulation of the 135 SDERGs, using the gene module map method [14]. The SDERGs were indeed coordinately repressed in many conditions that represent pathologic proliferation, specifically the subset of fast doubling cell lines in the NCI60 collection of tumor cell lines (p < 10−10) and several human cancers relative to their normal tissue counterpart including cancers of prostate (p < 10−6), blood (p < 10−12), and lung (p < 10−4) (Figure 5A). These results suggest that SDERGs likely antagonize cell proliferation in many cell types.
Cancer consists of a broad range of clinical behaviors ranging from indolent tumors to aggressive metastatic disease. To further dissect the potential molecular variation underlying this clinical heterogeneity and to extend and validate our results, we tested the prognostic power of the SDERG gene set in independent datasets and different subtypes of human cancer. Analyzing DNA microarray data from a study of 103 prostate tissues and cancer [25], we found that coordinate repression of SDERGs could identify over 90% of prostate tumors relative to normal prostate, a finding very unlikely due to chance alone (p < 10−11) (Figure 5B). Furthermore, expression of SDERGs in a set of 295 breast cancers [26] naturally divided the breast tumors into two groups (Figure 5C). Patients with breast cancers that diminished expression of SDERGs had significantly worse survival (p < 10−5) and significantly increased probability of metastasis (p < 10−4). This group of tumors with repression of SDERGs also tended to be of the grade 3 tumors (p < 10−9), which are defined by higher cell proliferation and less differentiation. Interestingly, the mRNA levels of SALL2 and MXI1, either alone or in combination, were insufficient to predict overall survival or metastasis-free survival (p > 0.05, Cox-Mantel test); conversely, removal of SALL2 and MXI1 from SDERGs did not affect the prognostic power of the SDERG gene set (unpublished data). These results further suggest a role for SDERGs to prevent excessive and pathologic proliferation. By reflecting the propensity for quiescence, the expression level of the SDERG gene set as whole may aid in predicting tumor behavior in two of the most common human cancers.
To better understand the transcriptional regulation of quiescence initiation versus maintenance, we compared SDERGs (135 genes) with 116 genes previously found to be concordantly induced by prolonged entry into quiescence, four days after growth factor deprivation, contact inhibition, or loss of adhesion [4]. A total of nine genes were in common between these two gene sets, while only one overlap gene is expected by chance alone (p < 10−7, hypergeometric distribution). The overlap genes include MXI1 and STAT1, thus indicating an interesting but limited overlap between the SDERGs and genes expressed during quiescence maintenance (Figure 6A). We next examined the coordinate expression of these 116 quiescence maintenance genes in 1,973 microarray representing 22 human tumor types (Figure 6B). SDERGs and quiescence maintenance genes showed overlapping but distinct patterns of expression, with some tumors coordinately repressing both gene sets but many that repress only one of the two sets. In human prostate cancer, quiescence maintenance genes are typically repressed but in far more haphazard fashion compared to SDERGs (p value of the separation is five logs of magnitude worse) (Figure 6C). Similarly, coordinate repression of quiescence maintenance genes is modestly predictive of primary breast cancer survival but not predictive of metastasis-free survival (Figure 6D and unpublished data). These results suggest that genes mediating entry into quiescence are largely distinct from those associated with quiescence maintenance, and the two programs may be repressed in distinct fashions to facilitate the progression of specific types of human cancers.
The general association between SDERGs and cell cycle exit raises the possibility that SDERGs may be induced by additional stimuli. For instance, in response to variety of noxious stress, cells will exit the cell cycle as part of the stress response. To test the possibility that SDERGs may be induced by stress, we queried the expression pattern of SDERGs in the published gene expression data of fibroblasts exposed to multiple types of stress [27]. In contrast to the strong induction of SDERGs by SD, exposure of fibroblasts to the reducing agent dithiothreitol (causing protein unfolding and endoplasmic reticulum stress), heat shock, or menadione (inducing oxidative stress) did not induce the SDERGs (Figure 7A). These results reaffirm SDERGs as a program uniquely responsive to loss of growth factor signaling as represented by SD.
By determining the genomic transcriptional program in response to SS and SD, we observed the asymmetric regulation of quiescence entry and exit (Figure 1). While the late transcriptional responses to these two opposing stimuli are largely symmetric, this symmetric program results from two distinct early transcriptional responses. These data suggest that in addition to previously identified antiproliferative genes that require ongoing growth factor signaling for their suppression [10], a major mechanism of quiescence entry is the induction of a set of unique quiescence entry genes (Figure 2). By simple analogy, a speeding car may be slowed by releasing the gas pedal, but the car can also be brought to a screeching halt by releasing the gas pedal and stepping on the brakes. We suggest that the SDERGs may be a set of brakes for cell cycle progression and growth factor-induced gene expression. Just as a gain of growth factor signaling activates the classic immediate early genes to induce cell cycle entry, a loss of growth factor signaling uniquely activates the SDERGs to induce cell quiescence. The decision of cell proliferation or quiescence is thus determined by the balance of growth factor-induced genes and SDERGs (Figure 7B). SDERGs are also actively involved in the repression of growth factor-induced genes. For instance, SALL2 and MXI1 are required to repress distinct sets of serum-inducible genes (Figure 3C); in the case of MXI1 this may occur by direct competition with MYC for binding to promoters of serum-inducible genes [20]. Thus, there is likely cross regulation of the early transcriptional responses to growth factor stimulation and deprivation.
One potential reason for this dual transcriptional response to growth factor gain and loss is to achieve tight regulation. It has been well known that many early response genes to SS are induced as a precise pulse that then rapidly decays despite continued mitogen presence (Figure 2A), thereby providing a check against unlimited proliferation and the risk of cancer. For instance, the classic mitogen-induced proto-oncogene MYC is regulated by transcriptional autorepression [28], mRNA instability [29], and rapid protein turnover [30,31]; enforced MYC expression is sufficient to induce ectopic DNA replication and DNA damage within just one cell cycle [32]. The low level of MYC and other growth factor early response genes at steady state would render a system based on their further decrement an insensitive strategy to detect growth factor deprivation. Instead, a decrement of growth factor signaling triggers a robust transcriptional response of SDERGs, leading to additional quiescence gene induction and exit of cell cycle. Intriguingly, a large number of SDERGs (including SALL2 and MXI1, which are required for entry to quiescence) are induced and maintained in stable expression in response to SD (Figure 2B). The longevity of SDERG expression in contrast to the transient expression of growth factor early response genes may provide an explanation for cell quiescence as the default state of most eukaryotic cells.
Among the several the SDERGs we tested, silencing of SALL2 expression had the most substantial effect on cell cycle exit in response to growth factor deprivation. Previously, Benjamin and colleagues had identified SALL2 as an antiproliferative gene by virtue of a tumor host range selection procedure for the oncogenic polyoma virus [19]. Enforced SALL2 expression in ovarian cancer cells inhibits tumor xenograft growth in vivo and can induce the expression of cyclin-dependent kinase inhibitor p21, although it is unclear whether p21 is the sole mechanism by which SALL2 elicits cell cycle arrest [17]. The biological context in which SALL2 might exert its antiproliferative effect was also not known. Our results suggest that SALL2 is induced in response to loss of growth factor signaling (Figure 2B). Acute loss of SALL2 during SD blocked the ability to stop DNA synthesis and induce additional quiescence-associated genes, suggesting that SALL2 is required for quiescence induction in response to growth factor deprivation (Figure 3C). Among the mammalian SALL family of zinc finger transcription factors, mutation of SALL1 leads to developmental abnormalities [33], and SALL4 is required for maintenance of pluripotency in embryonic stem cells [34,35]. Surprisingly, SALL2 knockout mice are viable and have no obvious phenotype [36], raising the possibility that SALL2's function may be redundant or compensated by other SALL family members. Indeed, we found that knockdown of SALL4 also blocks cessation of DNA synthesis in response to SD even though SALL4 mRNA level does not change in response to SD (H. Liu and H. Y. Chang, unpublished data). Thus, it may be the total pool of SALL transcription factors in the cell that determines cell quiescence, and SALL4 may compensate for the chronic loss of SALL2 during development. Coller et al. have shown that quiescence in fibroblasts inhibits their trans-differentiation (such as into muscle cells in response to enforced MyoD expression) [4]; the role of SALL transcription factors in cell quiescence may therefore be intimately linked to their roles in stem cell pluripotency [34,35]. The mechanisms by which the SALL2 message accumulates during SD and the functional roles of newly discovered SALL-regulated genes in SD should be addressed in future studies.
Because cell quiescence has been postulated to be an important safeguard against cancer [1], we reasoned that a transcriptional program mediating entry into quiescence might be systematically repressed in human cancers. Our survey of nearly 2,000 microarrays representing diverse types of human cancers identified multiple tumor types in which SDERGs are coordinately repressed (Figure 4). In addition, we found that repression of SDERGs unambiguously distinguished prostate cancers from adjacent prostate tissues, and repression of SDERGs in human breast cancers further predicted aggressive clinical course of early stage tumors. These properties are specific to SDERGs and are present to a much lower extent in genes associated with quiescence maintenance (Figure 6). Combined with the evidence that several SDERGs are required for cell cycle exit, these results highlight a potentially important role for the ability of cells to sense decrements of growth factor signaling and respond by quiescence. SDERGs and other genes induced by stimuli that induce cell quiescence may represent previously unrecognized tumor suppressor pathways; better understanding of these transcriptional programs may lead to new avenues of cancer diagnosis and treatment.
Primary human foreskin fibroblasts (CRL 2091; American Type Culture Collection, http://www.atcc.org) were cultured in DMEM plus 10% FBS. Cells were plated at 10% confluence. Cells were switch to DMEM plus 0.1% FBS 48 h after the last passage and harvested at the indicated time points.
Construction of human cDNA microarrays containing approximately 43,000 elements, representing approximately 23,000 different genes, and array hybridizations were as previously described [8]. Total RNA was extracted using Trizol according to the manufacturer's instructions (Invitrogen, http://www.invitrogen.com) and amplified using the Ambion MessageAmpII aRNA kit (Ambion, http://www.ambion.com). For time course experiments, Human Universal Reference RNA (Stratagene, http://www.stratagene.com) was used as reference RNA to compare with RNA from individual time points. We took four independent samples at time zero, which functioned as the baseline for other sample time points. For siRNA samples, siGFP was used as the reference RNA to compare with RNA from cells transfected with siSALL2 or siMXI1; each comparison was performed in duplicate. Full microarray data are available for download at Stanford Microarray Database (http://genome-www5.stanford.edu) or Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo).
We selected genes for which the corresponding array elements had fluorescent hybridization signals at least 1.5-fold greater than the local background fluorescence in the reference channel and further selected genes for which technically adequate measurements were obtained from at least 80% of the samples in a given dataset. The zero time point for the SD experiment was performed in quadruplicate, and the four gene expression measurements were averaged and subtracted from those of the subsequent time points in order to visualize gene induction or repression over time. The gene expression profiles of the same cells to SS over 15 time points were similarly transformed by subtraction of expression value from each time point by that of the zero time point. The two datasets were then merged by matching Stanford clone IDs of the cDNA probes. We next focused on genes that exhibited substantial variation in expression and selected the subset of genes that were induced or repressed by at least four-fold in at least one array in either time course, yielding 444 cDNA probes (henceforth genes). These genes were organized and grouped based on the similarity in their patterns of expression by average linkage clustering using the Cluster software [37]. Clustering of genes revealed three main transition points of gene expression variation—immediately after SS or SD, three hours after the stimuli, or eight hours after the stimuli. We therefore defined a time for the induction or repression of each gene cluster as the time point at which the gene expression reaches half maximal induction or repression, and classified each cluster as being regulated early (<3 h), middle (3–8 h), and late (>8 h). For gene clusters that exhibit biphasic or more complex patterns of regulation, we defined the onset of activity based on the first peak of expression variation. To quantify the degree of divergence among early-, mid-, and late-response genes, we calculated the Pearson correlation between the expression pattern of the SS and SD time courses for each gene. After ordering genes by hierarchical clustering, we visualized the Pearson correlation of gene clusters by displaying the moving average of correlation values of the ten nearest genes along the y-axis of the heat map.
The gene module map of functional groups of genes was as described using the software package Genomica [14]. Briefly, for each microarray, we identified genes that were induced or repressed by at least 2-fold and tested for their enrichment in each of 1,735 gene sets defined by gene ontology terms [15] using the hypergeometric distribution. An FDR calculation was used to account for multiple hypothesis testing. Enrichments that had p < 0.05 and FDR < 0.05 were considered significant and are shown in Figure 1B, yielding a higher order view of each gene expression profile as sets of activated and repressed functions.
For the cis-regulatory motif map, we first defined motif modules from the data of Xie et al. [16], where each motif module is a group of genes that shared enrichment in an evolutionarily conserved cis-regulatory motif in their upstream regulatory sequences. The upstream regulatory region of each gene is defined as the 4,000 base pairs centered at the annotated transcription start site, as was done by Xie et al.; 175 motif modules were defined. Second, for each array, we identified genes that were induced or repressed by 2-fold and tested for their enrichment in each of the motif modules using the hypergeometric distribution as described above. Modules with significant enrichment (p ≤ 0.05) were identified and shown in Figure 1C, yielding a higher level view of each expression profile as a combination of activated and repressed cis-regulatory motifs.
To conduct microarray analysis of siRNA treated cells we employed a type I microarray design where mRNA of cells treated with siSALL2 or siMXI1 (labeled with Cy5) was compared to mRNA of cells treated with siGFP (labeled with Cy3) by competitive hybridization to cDNA microarrays. We selected for analysis genes for which the corresponding array elements had fluorescent hybridization signals at least 1.5-fold greater than the local background fluorescence in the reference channel, and we further restricted our analyses to genes for which technically adequate measurements were obtained from both duplicate arrays. We further selected genes that were induced or repressed by at least 2-fold in two arrays by siSALL2 or siMXI1, and the genes were organized by average linkage clustering.
Regarding the cancer compendium and clinical outcome data, a cancer compendium of 1,973 microarrays was as described [14]. Gene probes from different microarray platforms were mapped by LocusLink identification numbers ((http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=gene). To test for the coordinate regulation of SDERGs in human cancers, we defined the 135 SDERGs as one gene set and tested each expression profile in the compendium for coordinate induction or repression of the SDERGs using the module map method in Genomica [14]. Specifically, for each microarray, we identified genes that were induced or repressed by at least 2-fold, and tested for their enrichment in SDERGs over that expected by chance alone (p < 0.05, FDR < 0.05). Next, to identify enriched clinical annotations among samples that exhibit coordinate SDERG induction or repression, for each annotation we compared the frequency of SDERG induction or repression among the samples versus that expected by chance alone. We found many significant enriched annotations of SDERG repression but not of SDERG induction in cancer (p < 0.05, FDR < 0.05, hypergeometric distribution). Several of the top ten enriched clinical annotations are reported in Figure 5A. The exact same procedure was repeated for 116 quiescence maintenance genes defined by Coller et al. [4].
To examine the expression of SDERGS in an independent set of human prostate cancer, we used the published prostate cancer microarray data of Lapointe et al. [25]. Gene probes were matched by cDNA clone IDs as this dataset was also generated on Stanford cDNA microarrays. We computed the average expression value of the 135 SDERGs in each of the 103 samples and rank ordered the average SDERG expression values. We considered those samples with average SDERG expression value greater than the mean of all 103 samples to have induction of SDERGs and those samples with average SDERG expression value to be below the mean of all 103 samples to have repression of SDERGs. The significance of the observed grouping of over 90 percent of prostate cancers with repression of SDERGs compared to normal prostate was evaluated by a two-by-two chi-square test, yielding a p value of 10−11. p Values were also calculated using the same procedure for the 116 quiescence maintenance genes and the 107 genes unique to Coller et al. [4] and plotted in Figure 6C as the −log10(p value).
To examine the clinical significance of SDERG repression in human breast cancer, we used the published breast cancer microarray data of van de Vijver et al. [26]. Gene probes were matched by Unigene ID (http://www.sgn.cornell.edu/bulk/input.pl?mode=unigene), and the 295 breast cancer samples were organized by two-way hierarchical clustering based on the expression pattern of the SDERGs. The main bifurcation of the dendrogram separated the breast cancer samples into two groups, one group with coordinate induction of SDERGs (termed “SDERG up”) and one group with coordinate repression of SDERGs (termed “SDERG down”). We compared the differences in overall survival and metastasis-free survival of these two groups of patients as defined by SDERGs using the Cox–Mantel test in the program Winstat (R. Fitch Software, http://www.winstat.com). Data from the 295 breast cancer samples were obtained for the quiescence maintenance genes unique to Coller et al. [4] and the same procedures as described above were performed.
To examine the expression pattern of SDERGs in fibroblasts undergoing cell stress, we downloaded the published microarray data of Murray et al. [27]. Gene probes were matched by Stanford cDNA clone IDs. Expression of SDERGs during SD or exposure to DTT, heat, or menadione are shown as heat maps, and the average expression values are shown across each time course as a graph in Figure 5A.
Cells were transfected with 20 nM of siRNA pools corresponding to each of the target genes (SALL2, MXI1, IRF1, and TNKS1BP1) and a control (siGFP) using DharmaFECT3 according to the manufacturer's instructions (Dharmacon, http://www.dharmacon.com). Fibroblasts were transfected with 20 nM of siRNAs at a density of 2 × 105 cells/well (six-well plate) in high serum (10% FBS) media. After 24 h, the treated fibroblasts were replated at a density of 6 × 103 cells/well in four-well chamber slides and were allowed to recover in high serum for 48 h. The transfected cells were then transferred to low serum (0.1% FBS) media for 16 h. The transfection efficiency of each siRNA was verified by qRT-PCR (Figure S1). DNA synthesis was monitored by measuring the incorporation of the thymidine nucleotide analog BrdU (Sigma, http://www.sigmaaldrich.com) into DNA as previously described [11]. Briefly, cells were incubated with 10 μM BrdU in the media for 6 h; then washed with PBS, fixed, and stained with an anti-BrdU monoclonal antibody (Becton Dickinson, http://www.bd.com) and Alexa Fluor-conjugated goat anti-mouse antibody (Molecular Probes, http://www.invitrogen.com). The percentage of BrdU-positive cells among >200 DAPI-positive cells in four random fields was recorded. Propidium iodide staining of DNA content and FACS analysis were performed as described [38], with four replicate samples for each condition.
Gene expression levels for genes targeted by the siRNAs were quantitated using RNA extracted from the transfected cells by Taqman quantitative one-step RT-PCR (Applied Biosystems, http://www.appliedbiosystems.com). Taqman probes to SALL2, MXI1, and IRF1 were used. Assays were normalized to GAPDH levels, and relative abundance was calculated using a delta–delta threshold analysis as previously described [39]. The assay identification numbers for the Taqman probes are in the Accession Numbers list in the Supporting Information section of this paper.
The LocusLink (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=gene) and Unigene ID (http://www.sgn.cornell.edu/bulk/input.pl?mode=unigene) of genes discussed in this manuscript are listed in Table S1.
The National Center for Biotechnology (NCBI) Probe Database (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=probe) accession numbers for the Taqman probes discussed in this manuscript are GAPDH, Hs99999905_m1; IRF1, Hs00233698_m1; MXI1, Hs00365651_m1; and SALL2, Hs00826674_m1. |
10.1371/journal.pntd.0004124 | Trends of Mycobacterium bovis Isolation and First-Line Anti-tuberculosis Drug Susceptibility Profile: A Fifteen-Year Laboratory-Based Surveillance | Mycobacterium tuberculosis causes the majority of tuberculosis (TB) cases in humans; however, in developing countries, human TB caused by M. bovis may be frequent but undetected. Human TB caused by M. bovis is considered a zoonosis; transmission is mainly through consumption of unpasteurized dairy products, and it is less frequently attributed to animal-to-human or human-to-human contact. We describe the trends of M. bovis isolation from human samples and first-line drug susceptibility during a 15-year period in a referral laboratory located in a tertiary care hospital in Mexico City.
Data on mycobacterial isolates from human clinical samples were retrieved from the laboratory’s database for the 2000–2014 period. Susceptibility to first-line drugs: rifampin, isoniazid, streptomycin (STR) and ethambutol was determined. We identified 1,165 isolates, 73.7% were M. tuberculosis and 26.2%, M. bovis. Among pulmonary samples, 16.6% were M. bovis. The proportion of M. bovis isolates significantly increased from 7.8% in 2000 to 28.4% in 2014 (X2trend, p<0.001). Primary STR resistance was higher among M. bovis compared with M. tuberculosis isolates (10.9% vs.3.4%, p<0.001). Secondary multidrug resistance (MDR) rates were 38.5% and 34.4% for M. bovis and M. tuberculosis, respectively (p = 0.637). A rising trend of primary STR monoresistance was observed for both species (3.4% in 2000–2004 vs. 7.6% in 2010–2014; p = 0.02).
There is a high prevalence and a rising trend of M. bovis isolates in our region. The proportion of pulmonary M. bovis isolates is higher than in previous reports. Additionally, we report high rates of primary anti-tuberculosis resistance and secondary MDR in both M. tuberculosis and M. bovis. This is one of the largest reports on drug susceptibility of M. bovis from human samples and shows a significant proportion of first-line anti-tuberculosis drug resistance.
| Human tuberculosis caused by Mycobacterium bovis (HTBMb) is a lesser-known form of the disease. The main route of transmission of HTBMb is the consumption of unpasteurized dairy products, causing mostly extrapulmonary disease. M. bovis is naturally resistant to pyrazinamide, a drug that allows for a shorter treatment course. Therefore, if M. bovis is not properly identified or if there is resistance to other drugs, proper treatment may be hindered. Most laboratories in developing countries do not routinely perform mycobacterial cultures, and only a few laboratories can identify M. bovis. Therefore, HTBMb cases are believed to be underestimated. We report a large proportion of M. bovis isolates and an increasing isolation trend across time. We report a large proportion of M. bovis isolates from pulmonary samples, suggesting the possibility of human-to-human airborne transmission. Also, we showed that M. bovis isolates were more frequently resistant to streptomycin, perhaps as a result of antibiotic usage in cattle. This work underscores the need for identification to the species level, proper susceptibility testing, as well as a stricter control of bovine tuberculosis.
| Tuberculosis (TB) remains an important health problem in several regions of the world, especially in countries with a high prevalence of HIV infection. Mycobacterium tuberculosis complex (MTBC) includes closely related species among which M. tuberculosis, M. bovis, and M. africanum are the most frequently associated with human disease. [1] Unlike M. tuberculosis, M. bovis can infect a broad range of mammals, including cattle and, therefore, it is considered a zoonosis. The main mechanism of contagion in humans is the consumption of unpasteurized dairy products and, less frequently, animal-to-human and human-to-human contact. [2–4]
Historically, the burden of human TB caused by M. bovis (HTBMb) has been closely related to that of bovine TB (BTB) in the same region.[5] Unfortunately, data from most developing countries, where there is still inappropriate BTB control, is scarce.[6,7] There are several factors that explain the underreporting of HTBMb in these regions. First, control programs rely on acid-fast bacilli (AFB) smears as the primary diagnostic method in suspected TB cases; and mycobacterial culture is performed only when drug resistance is suspected, or treatment is failing. Second, most laboratories use culture medium containing glycerol, and this reduces the probability of M. bovis isolation. Lastly, only a few laboratories can identify MTBC at the species level.[8,9] In Mexico, one national reference laboratory (Instituto Nacional de Diagnóstico y Referencia Epidemiológicos) and 32 public health laboratories are capable of performing mycobacterial cultures. However, species-level identification is not routinely performed. Consequently, data regarding first-line anti-tuberculosis drug susceptibility for MBTC are scarce. Therefore, describing the burden of HTBMb and first-line drug susceptibility pattern has important implications for treatment, referral, and public health policies in developing countries. We describe the trends of M. bovis isolation and first-line drug susceptibility for a referral laboratory in a tertiary care hospital during a 15-year period.
This study was conducted at Instituto Nacional de Ciencias Médicas y Nutrición Salvador Zubirán, one of the National Institutes of Health in Mexico. This tertiary care referral center accepts adult patients from all over the country. The patient population involves complex medical and surgical cases including rheumatologic patients, bone marrow and solid organ transplant recipients, and HIV patients. The Laboratory of Clinical Microbiology receives local and referred clinical samples from other National Institutes of Health as well as other hospitals in Mexico City and nearby states for mycobacterial culture. Since 1992, as a regular practice, samples from patients suspected to have TB undergo mycobacterial culture, species-level identification and first-line drug susceptibility testing in all instances. We selected the 2000–2014 period because laboratory practices for mycobacterial isolation, identification, and susceptibility testing became more uniform then.
Institutional approval for this study was obtained from the Comité Institucional de Investigación Biomédica en Humanos. All data analyzed were anonymized.
A search was performed using the Laboratory of Clinical Microbiology database to investigate all MTBC isolates from January 2000 to December 2014. This database includes both local and referred samples. Only the first isolate was considered for the analysis when multiple samples within a 6-month period were available for a single patient. Whenever there were additional isolates from the same patient ≥ 6 months apart, they were considered as separate episodes and included for analysis.
Data on mycobacterial species, drug susceptibility, sample source, treatment, and referral status were obtained. HIV status was obtained for local samples only. Primary resistance (new cases) was defined as those samples from patients without previous anti-tuberculosis treatment. Secondary resistance (treated cases) was defined as those samples from patients who had previously received any anti-tuberculosis treatment. Resistance to both isoniazid (INH) and rifampin (RIF) was defined as multidrug resistance (MDR). Polydrug resistance was defined as resistance to two or more drugs but excluding those classified as MDR. Samples from sputum, bronchoalveolar lavage, endotracheal aspirate, gastric aspirate, lung and pleural biopsy, and pleural fluid were classified as pulmonary. Liver, spleen, and gastrointestinal biopsies, as well as ascites fluid, and fecal samples were classified as abdominal. Additionally, if samples from the same patient were obtained from a pulmonary and an extrapulmonary source, the case was classified as pulmonary.
As a laboratory standard procedure, all samples obtained from bronchoalveolar lavage, cerebrospinal fluid, and biopsies from any tissue or abscess were cultured in mycobacteria-specific culture medium, regardless of TB suspicion. Sputum and urine samples were cultured for mycobacteria at the request of the treating physician.
Samples were digested and decontaminated by the NALC-NaOH method as previously described. [10,11] After digestion, samples were inoculated in both Löwenstein-Jensen medium and MGIT tubes (Becton-Dickinson, Sparks, MA, USA) according to the manufacturer’s specifications. Biopsy samples were additionally inoculated in Stonebrink culture medium. Additionally, smears from all samples were prepared for Ziehl-Neelsen and Auramine-rhodamine stain. Isolates obtained from MGIT tubes were sub-cultured in Stonebrink and Löwenstein-Jensen medium. All positive cultures were further identified by DNA probe (Accuprobe, GEN-PROBE, San Diego, CA). Biochemical tests (niacin production, nitrate reduction, thiophen-2-carboxylic acid anhydride susceptibility, and pyrazinamidase deamidation) for the identification of M. bovis were performed in those positive cultures with dysgonic growth. [12] Spoligotyping was performed for local isolates as previously described, and data was entered into an international database (www.mbovis.org).[13] Susceptibility testing for anti-tuberculosis drugs was performed for INH, RIF, streptomycin (STR) and ethambutol (EMB). For this purpose, and according to the manufacturer’s specifications, the radiometric BACTEC 460 TB culture system (Becton-Dickinson, Sparks, MA, USA) was used from the years 2000 to 2010 with the following drug concentrations: INH (0.1 μg/mL), RIF (2.0 μg/mL), EMB (7.5 μg/mL), and STR (6.0 μg/mL); then, from 2010 on, the BACTEC MGIT 960 culture system (Becton-Dickinson, Sparks, MA, USA) was used with the following concentrations: INH (0.1 μg/mL), RIF (1.0 μg/mL), EMB (5.0 μg/mL), STR (2.0 μg/mL). This laboratory is subjected to regular quality control evaluations by the Centers for Disease Control and Prevention, and the College of American Pathologists for identification and susceptibility testing of mycobacteria species.
Statistical analysis was performed using STATA 11.0 software (StataCorp, College Station, TX, USA). Categorical data was summarized using frequency tables, and the X2 test was used for comparison between groups. The M. bovis case proportion by year was analyzed obtaining a X2 for trend by the Armitage test (regression). A p-value <0.05 was determined as statistically significant for all tests.
During the study period, 81,521 samples were processed for mycobacterial culture (Fig 1). Among these, 1,165 MTBC isolates were identified, 583 (50.0%) as local samples and 582 (49.9%) as referrals after eliminating duplicate cultures. Of these, 73.7% (859/1,165) were identified as M. tuberculosis, and 26.2% (306/1,165) as M. bovis. Sixty-two percent of the M. tuberculosis isolates and 35.2% of the M. bovis isolates were obtained from pulmonary samples (p<0.001; Table 1). Data on AFB stain were available for 954 isolates; the proportion of positive AFB stain was 75.9% (378/680) for M. tuberculosis, and 24.1% (120/274) for M. bovis (p = 0.001). Among the AFB-positive M. bovis isolates, 45.8% (55/120) were pulmonary, and 54.1% (65/120) were extrapulmonary (p = 0.018). Conversely, among the AFB-positive M. tuberculosis isolates, 79.3% (300/378) were pulmonary and 20.6% (78/378) were extrapulmonary (p<0.001).
One hundred and twelve (19.2%) local isolates were from HIV-infected patients; 63.3% (71/112) were M. tuberculosis, and 36.6% (41/112) were M. bovis (p = 0.054). Among the samples from HIV-infected patients, 52.6% (59/112) were pulmonary samples. Of these, 76.2% (45/59) were identified as M. tuberculosis, and 23.7% (14/59) were M. bovis.
The overall proportion of M. bovis isolation significantly increased from 7.8% in 2000 to 28.4% in 2014 (X2trend, p<0.001; Fig 2). Spoligotype pattern was available for 63.5% (108/170) of the local samples (S1 Table).
Data on first-line anti-tuberculosis drugs susceptibility were available for 1,139 (97.7%) isolates (Table 2). When considering monoresistance among all isolates, 10.9% of M. bovis and 3.2% of M. tuberculosis were resistant to STR (p<0.001). This association remained after stratifying by new and treated cases (p<0.001 and p = 0.032, respectively). Total MDR among all cases was 11.9% for M. tuberculosis and 7.6% for M. bovis (p = 0.038). This same association was observed among new cases (6.8% vs. 3%, p = 0.026). However, among treated cases no difference was observed.
An increasing trend of STR monoresistance among new cases was found when considering both species (3.4% in 2000–2004 vs. 7.6% in 2010–2014, p = 0.02) (Table 3).
This report demonstrates a high prevalence and a rising trend in the proportion of M. bovis isolation in our laboratory. This report is one of the largest on first-line anti-tuberculosis drug profile of M. bovis and shows a noteworthy proportion of first-line anti-tuberculosis drug resistance and secondary MDR isolates.
The proportion of M. bovis isolates in this study (26.2%) is much higher than that reported by other hospital-based studies in Latin America (0.4%) and by other hospitals in Mexico (<1%). [14,15] This may be explained by the larger study period of the present report and the high proportion of samples from immunosuppressed patients who are at a greater risk for M. bovis infection, as documented in previous studies.[16] In fact, isolates obtained from HIV-infected patients accounted for 19.2% of the local samples.[17]
We also identified a rising trend in the proportion of cases caused by M. bovis across time. HMBTb is considered a reflection of the BTB burden in the region. In fact, we recently identified a high burden of bovine and human TB in a dairy production facility in rural Mexico. [2] This also correlates with previous reports of M. bovis among artisanal dairy products, which have been linked to HMBTb cases in Mexico and along the south border cities of the United States.[18,19] Mexico is considered a country of “sporadic occurrence” of BTB by the World Organization for Animal Health. However, like in many countries in the region, the test-and-slaughter strategy for bovine tuberculosis control is not universally implemented. [6] The main obstacles for BTB control in Mexico are financial and cultural. Official government data reports an overall prevalence of BTB of 2.05% in 2015, but it reaches 16.5% among dairy farms.[20] The reason for this difference may be explained by the fact that meat producing regions require to be certified as BTB free for cattle export. On the contrary, in dairy production farms, BTB only mildly affects production and pasteurization eliminates M. bovis. [21] Unfortunately, about 30% of the milk production in Mexico is sold without pasteurization, mostly to small retailers and artisan cheese producers.[22]
The respiratory route of contagion is considered less efficient for M. bovis than for M. tuberculosis. However, recent data detailing outbreaks in the community and hospitals demonstrated that human-to-human contagion is not as unlikely as previously believed.[3,23,24] Interestingly, we observed an important proportion (16.6%) of M. bovis isolates recovered from pulmonary samples from a mainly urban population. Therefore, it may be hypothesized that airborne human-to-human transmission of M. bovis may occur in the community, but remains undetected in our region given that mycobacterial culture is not routinely performed. Unfortunately, as an important limitation of this study, the lack of clinical and epidemiologic data precludes us from reaching a definite conclusion.
We also report a high rate of first-line anti-tuberculosis resistance for MBTC. The proportion of INH (14%), RIF (7.6%), and STR (5.5%) primary resistance found among new cases is considerably higher for INH and RIF than in previous reports from our group in 1995 (INH 6%, RIF 2%, STR 6%).[25] These proportions are also higher than those from other reports in Mexico (1995 to 2006) among new cases (INH 9%, RIF 3%, and MDR 4.5%), and are also higher than recent data from the National Survey on TB Drug Resistance in Mexico (INH 3.5%, RIF 0.1%, STR 4% and MDR 2.3%).[26,27] This discrepancy may be explained by the different periods from which data was collected and the dissimilar patient population (ours being hospital-based and including more HIV-infected, and immunosuppressed patients).
When comparing the resistance profiles of M. bovis and M. tuberculosis, we observed a considerably higher primary MDR M. tuberculosis proportion, and a significantly higher STR monoresistance among M. bovis isolates. Data regarding drug susceptibility for M. bovis TB in humans and animals is limited. A study from San Diego reported 7% resistance for INH and 1% for RIF among 167 M. bovis TB cases.[28] Drug susceptibility has also been reported from outbreaks caused by MDR strains; however, most other case series of HTBMb report full susceptibility to all first-line anti-tuberculosis drugs.[14,15,29–32] A report from the National TB Genotyping Service of the United States informed of 17% of STR resistance among 165 M. bovis isolates; however, no explanation for this was proposed. [33] Drug susceptibility from M. bovis isolates collected from farm animals or wildlife is almost uniformly reported as fully susceptible to anti-tuberculosis drugs.[34–36]. We believe that a high proportion of primary resistance for STR among M. bovis isolates, may be explained by the use of aminoglycosides for treating other diseases in cattle.[37] Some have suggested that primary resistance to INH and RIF in M. bovis may indicate human-to-human transmission.[32] However, we surveyed BTB in a dairy farm, where samples from cows were obtained during necropsy. We recovered 150 M. bovis isolates; among these, we observed an even higher rate of STR resistance (15.6%) and a similar rate of INH (9.2%) and RIF (3.4%) resistance. (M. Bobadilla, Instituto Nacional de Ciencias Médicas y Nutrición Salvador Zubirán, Mexico City, personal communication).
It has been suggested that HTBMb cases may be at a higher risk for developing MDR strains if natural resistance to pyrazinamide is not considered and monoresistance to INH or RIF is present. [38,39] We identified a similar proportion of secondary MDR M. bovis and M. tuberculosis isolates. Unfortunately, only a few cases were analyzed and no data on previous treatments or outcomes were available, therefore we are unable to conclude if this assumption is true. Nevertheless, it should be recognized that TB cases caused by MDR M. bovis may result in disease that is harder to treat on a second- line drug regimen. This highlights the need for performing species-level identification and drug susceptibility testing whenever M. bovis is suspected.
In conclusion, we believe that data contained in this study is relevant in terms of public health and highlights the need for more stringent control of BTB in our country. It also underscores the importance of proper identification of M. bovis given that the considerable rate of primary resistance to INH and RIF along with the natural pyrazinamide resistance may result in treatment failures and select for MDR strains.
|
10.1371/journal.pntd.0003793 | International External Quality Assessment Study for Molecular Detection of Lassa Virus | Lassa virus (LASV) is a causative agent of hemorrhagic fever in West Africa. In recent years, it has been imported several times to Europe and North America. The method of choice for early detection of LASV in blood is RT-PCR. Therefore, the European Network for Diagnostics of ‘Imported’ Viral Diseases (ENIVD) performed an external quality assessment (EQA) study for molecular detection of LASV. A proficiency panel of 13 samples containing various concentrations of inactivated LASV strains Josiah, Lib-1580/121, CSF, or AV was prepared. Samples containing the LASV-related lymphocytic choriomeningitis virus (LCMV) and negative sera were included as specificity controls. Twenty-four laboratories from 17 countries (13 European, one African, one Asian, two American countries) participated in the study. Thirteen laboratories (54%) reported correct results, 4 (17%) laboratories reported 1 to 2 false-negative results, and 7 (29%) laboratories reported 3 to 5 false-negative results. This EQA study indicates that most participating laboratories have a good or acceptable performance in molecular detection of LASV. However, several laboratories need to review and improve their diagnostic procedures.
| A proficiency test panel for molecular diagnostic of Lassa virus provides objective evidence of testing quality of International diagnostic laboratories. Since there are no commercial assays available, it is very important to assess the quality of diagnostic test used as well as evaluate detection sensitivity and specificity performance. Participating laboratories have received samples containing different inactivated Lassa virus strains as well as two negative controls. Participants were asked to provide information on diagnostic test procedure and protocols used for analysis of samples of Lassa virus External Quality Assessment (EQA). Based on received information we were able to compare and evaluate the quality of diagnostic profile and facilitate further improvement. Participating laboratories may use results of Lassa virus EQA to become accredited for Lassa virus molecular diagnostic. Since different Lassa virus strains are not available for most of the laboratories, participants achieved very advanced training for diagnostic of rare and imported viruses.
| Lassa fever was first described in 1969 as the cause of a nosocomial outbreak of hemorrhagic fever in Nigeria [1]. Lassa fever is an acute viral infection associated with a wide spectrum of disease manifestations, which range from mild courses to multiorgan failure [1–3]. The etiologic agent of Lassa fever is Lassa virus (LASV, family Arenaviridae, genus Arenavirus) [4]. The natural host of LASV is the small rodent Mastomys natalensis, which lives close to human settlements [5]. The rodents can become chronically infected at birth and excrete infectious virus in urine and other body fluids, with subsequent transmission to humans [6]. There is evidence of human-to-human transmission in both hospital and community settings [7]. The fact that LASV may be transmitted from human to human gives rise to nosocomial or community-based outbreaks. LASV is endemic in the countries of Nigeria, Liberia, Sierra Leone, and Guinea [8, 9] and was detected in Mali [10, 11]. Seroepidemiological studies and imported cases of Lassa fever indicate that arenaviruses circulate somewhere in the region comprising Côte d’Ivoire and Burkina Faso [12]. The annual incidence is estimated at 300,000 cases, with 5,000 fatalities per year [13, 14]. Additionally, LASV has been introduced several times into Europe, Japan, and North America. Among the hemorrhagic fever viruses of risk group 4 (such as Crimean-Congo hemorrhagic fever, Ebola, and Marburg virus), LASV has been most frequently imported [15]. The virus usually is imported by returning travelers [16, 17]. Within Europe, LASV infections have been imported to Germany [18, 19], The Netherlands [20] and the United Kingdom [21].
Laboratory testing is required to establish a diagnosis, as Lassa fever can hardly be distinguished from other febrile diseases based on clinical symptoms [14, 22]. A suspected case must be rapidly ruled out or verified to facilitate appropriate case management, including treatment, the implementation of isolation measures, or the tracking of contact persons [18].
The method of choice for early detection of LASV in blood is reverse transcription (RT)-PCR [23–29]. However, the high degree of genetic variability of the virus poses a problem with the design of RT-PCR assays for the reliable detection of all virus strains [30]. The performance of the different techniques applied for molecular diagnosis of LASV may vary between laboratories. External quality assessment (EQA) studies for LASV molecular diagnostics have not been performed since 2004 [31]. An EQA study allows the participating laboratories to monitor the quality of their diagnostics and to identify problems with particular diagnostic assays. For these reasons, an EQA study for the molecular diagnosis of LASV was conducted by the European Network for Diagnostics of ‘Imported’ Viral Diseases (ENIVD) (http://www.enivd.org) in 2013. ENIVD is concerned with the development of laboratory diagnostic capacities for imported virus infections, quality control, standardization of laboratory procedures, and training of laboratory staff [32]. Based on the results of this study, the quality of LASV diagnostics may be improved.
Twenty-eight laboratories involved in diagnostics of viral hemorrhagic fevers were invited to participate in this study. Invitees were selected from the register of ENIVD network members and from the list of national and regional reference laboratories for rare, emerging, and dangerous viruses. The participation in the study was free of charge. Participants permitted publication of the results in a comparative and anonymous manner. This EQA was coordinated by ENIVD according to the established procedures of the network [33–35].
The proficiency test panel included 13 LASV preparations derived from culture supernatants of Vero E6 cells (ATCC—American Type Culture Collection) infected with 4 different LASV strains. Virus in cell culture supernatant was inactivated by heat (1 h at 60°C) followed by gamma irradiation (25 kilo gray). The test panel consisted of six samples of LASV strain Josiah from Sierra Leone, obtained by serial 10-fold dilution of cell culture supernatant (1:10 to 1:106), three samples of LASV strain Lib-1580/121 from Liberia (dilutions 1:103 to 1:105), LASV strain CSF from Nigeria (dilution 1:103), and LASV strain AV from Cote d’Ivoire or Burkina Faso (dilution 1:103). The samples were freeze-dried in 3% mannitol based formulation using an EPSILON 2-6D Pilot Freeze Dryer (Martin Christ GmbH, Osterode am Harz, Germany). In addition, we included one sample containing LASV strain Josiah at a dilution of 1:104 (sample #14) that was prepared with a new dry stabilizer method (Biomatrica, Inc., San Diego, CA, USA), and one sample containing LASV strain Lib-1580/121 at dilution of 1:104 (sample #6) that was prepared using a liquid stabilizer (Biomatrica). A sample containing lymphocytic choriomeningitis virus (LCMV), the prototype member of the family Arenaviridae, as well as two negative control sera were included in the test panel as specificity controls.
After lyophilized sample preparation, the samples were tested and quantified by an in-house real-time PCR assay for quality control purpose. The assay was performed by employing 12.5 pmol of forward primer LaV F2 (CCACCATYTTRTgCATRTgCCA), 13 pmol of reverse primer LaV R (gCACATgTNTCHTAYAgYATggAYCA) and 5 pmol of probe LaV TM (FAM-AARTggggYCCDATgATgTgYCCWTT-BBQ). The real-time PCR assay was carried out in one-step format on ABI 7500 real-time PCR system using the AgPath-ID One-Step RT-PCR Kit according to manufacturer´s instruction. Plasmid standards were used for the quantification of the genome copies of Lassa virus RNA.
The EQA was performed according to National Ethical regulations.
Before dispatching the panel, samples were sent to the reference laboratory for testing the quality and obtaining the reference results. Reference laboratory used RT-PCR protocol described by Ölschläger et al., 2010 [27]. Samples were resuspended in 100 μl of water and the RNA was extracted using the QIAamp viral RNA kit (Qiagen, Hilden, Germany). The presence of LASV or LCMV RNA in the samples was ascertained by RT-PCR and sequencing. The number of LASV genome copies present in these samples was determined by qRT-PCR. Samples were shipped by regular mail at ambient temperature. Participating laboratories were instructed to resuspend the samples in 100 μl of water and to analyze the material like serum samples potentially containing LASV using their routine nucleic acid detection assays. The EQA panel was accompanied by documentation including instructions and an evaluation form for results. Participants were asked to report the assay protocol, the result for each sample, the LASV strain identified, the number of genome copies as well as any problem encountered.
To guarantee anonymous data evaluation and reporting, each participating laboratory was coded with an identifier. The results were scored according to detection rate and specificity as in previous EQA studies of ENIVD [33–35]. We assigned one point for correct results; false-negative, false-positive, and indeterminate results did not count. Results were classified as “good”—when all results were correct; “acceptable”—when 1 to 2 results were incorrect; and “need for improvement”—when more than 2 results were incorrect. Results for the sample containing LCMV (sample #3) were not included in the score, as verification of the sequence was optional. In addition, we excluded from scoring the sample containing LASV strain Josiah at a dilution of 1:106 (sample #15) as this concentration is likely to be below the 95%-detection limit of most assays. Thus, obtaining a positive or negative result becomes a matter of chance. Each laboratory received the complete summary of the results in an anonymous way, by which only the own laboratory was recognizable.
Twenty-four (86%) of the 28 laboratories, which received the EQA material, reported results. The 24 participating laboratories, located in 17 countries—13 European, one African, one Asian, and two American countries (Table 1). The LASV detection rate varied among laboratories and scores ranged from 9 to the maximum of 14 (Table 2). Average score for all participating laboratories was 13 points. Good results were achieved by 13 (54%) laboratories, 4 (17%) laboratories achieved acceptable results, and 7 (29%) laboratories had need for improvement. Table 3 shows that 13 (54%) laboratories correctly detected LASV in all 12 LASV samples (100% detection rate). Three (12%) participants had one false negative result (92% detection rate). Eight participants had a detection rate between 58% and 83%. None of the laboratories reported false-positive results for the negative control samples.
All participating laboratories were able to detect LASV strain Josiah at 1:10, 1:100, and 1:104 dilution (Table 4). Two laboratories did not detect the Josiah strain at 1:103 dilution (sample #5) but detected the 1:104 to 1:106 dilutions of this strain (Table 2). A mix-up between sample#5 and sample #6 is a likely explanation. LASV strains AV from Cote d’Ivoire/Burkina Faso and CSF from Nigeria were detected correctly by all laboratories. Four laboratories did not detect at all the Liberian LASV strain. Further laboratories did not detect the higher dilutions of the Liberian strain (≥ 1:104).
Nineteen participants used published RT-PCR protocols, two laboratories used unpublished in-house RT-PCR assays, one laboratory used a combination of real-time and conventional RT-PCR, and two laboratories did not provide information about the protocol. Three laboratories confirmed their results by using a second published protocol. Table 5 shows that 11 (50%) laboratories used the protocol published by Ölschläger et al., 2010 [26]. These 11 laboratories reported nine false-negative results (detection rate of 93% for this protocol). Seven laboratories using the Ölschläger et al. protocol did not report any false-negative result. Six participants (27%) used the protocol of Vieth et al., 2007 [29] (Table 5). They reported 12 false-negative results (detection rate of 83% for this protocol). Two laboratories using this protocol did not report any false-negative results. Three other protocols (Demby et al., 1994 [23]; Coulibaly N’Golo et al, 2011 [28]; Drosten et al., 2002 [27]) were used by one or two laboratories (Table 5). In the supplementary file (S1 Fig) we are showing partial alignments of the S-segment sequences of LASV strains used in the EQA with position of some published primers: (A) forward primer sequence and position on S-segment alignment; (B) reverse primer sequence and position on S-segment alignment. In comparison to Demby et al., Lassa RT-PCR assay described by Ölschläger et al. had just modified reverse primer in order to detect all described Lassa virus strains.
Ten laboratories (42%) reported the presence of LCMV in sample #3. In addition, 9 (37%) laboratories reported this sample as negative, which is also considered a correct result because this EQA was conducted to test for the ability to detect LASV. Five laboratories (21%) reported the sample containing LCMV as positive for LASV. These laboratories used protocols for detection of Old World arenaviruses, including LCMV. This underlines the relevance of sequencing the diagnostic PCR products when using pan-virus family detection assays.
This EQA study indicates that most participating laboratories, located in various countries around the world, have a good or acceptable performance in molecular detection of LASV. However, several laboratories need to improve their performance, in particular with respect to detection of the Liberian strain. The data allow the participating laboratories to identify the weakness in their diagnostic procedures and to review and improve their protocols.
One published protocol has achieved 100% detection rate reported by single participant. However, the reference laboratory recommends Ölschläger et al., 2010 published protocol for LASV detection as most commonly used with good detection rate and ability to detect all described Lassa virus strains. The main aim of this EQA study was not to compare published protocols, rather to give chance to participating laboratories to evaluate their testing performance and provide practical exercise for molecular detection of LASV. There should be a follow-up EQA for molecular detection of LASV to evaluate a possible improvement.
|
10.1371/journal.pgen.1006554 | Post-Translational Dosage Compensation Buffers Genetic Perturbations to Stoichiometry of Protein Complexes | Understanding buffering mechanisms for various perturbations is essential for understanding robustness in cellular systems. Protein-level dosage compensation, which arises when changes in gene copy number do not translate linearly into protein level, is one mechanism for buffering against genetic perturbations. Here, we present an approach to identify genes with dosage compensation by increasing the copy number of individual genes using the genetic tug-of-war technique. Our screen of chromosome I suggests that dosage-compensated genes constitute approximately 10% of the genome and consist predominantly of subunits of multi-protein complexes. Importantly, because subunit levels are regulated in a stoichiometry-dependent manner, dosage compensation plays a crucial role in maintaining subunit stoichiometries. Indeed, we observed changes in the levels of a complex when its subunit stoichiometries were perturbed. We further analyzed compensation mechanisms using a proteasome-defective mutant as well as ribosome profiling, which provided strong evidence for compensation by ubiquitin-dependent degradation but not reduced translational efficiency. Thus, our study provides a systematic understanding of dosage compensation and highlights that this post-translational regulation is a critical aspect of robustness in cellular systems.
| Cells are exposed to environmental changes leading to fluctuations in biological processes. For example, changes in gene copy number are a source of such fluctuations. An increase in gene copy number generally leads to a linear increase in the amount of protein; however, a small number of genes do not show a proportional increase in protein level. We investigated how many of the genes exhibit this nonlinearity between gene copy number and protein level. Our screen of chromosome I suggests that genes with such nonlinear relationships constitute approximately 10% of the genome and consist predominantly of subunits of multi-protein complexes. Because previous studies showed that an imbalance of complex subunits is very toxic for cell growth, a function of the nonlinear relationship may be to correct the balance of complex subunits. We also investigated the underlying mechanisms of the nonlinearity by focusing on protein synthesis and degradation. Our data indicate that protein degradation, but not synthesis, is responsible for maintaining a balance of complex subunits. Thus, this study provides insight into the mechanisms for coping with the fluctuations in biological processes.
| Robustness in biological systems is a general trait of living cells and a fundamental feature involving the maintenance of stability during perturbation [1–4]. It is a universal challenge to cope with perturbations leading to fluctuations in biological processes because cells are exposed to changes in internal and external environments [5,6]. The robustness of cells to various perturbations can be understood as a consequence of fluctuations in gene expression and buffering of fluctuations [5–8]. Therefore, understanding buffering mechanisms is essential to understanding the optimization of gene expression and adaptation to changes in environmental conditions.
The decoding of genetic information is achieved through irreversible processes from DNA to RNA to protein as stated in the central dogma of molecular biology [9]. The gene expression level at each step is generally in a linear relationship with gene copy number, namely an increase in gene copy number leads to a proportional increase in messenger RNA (mRNA) and corresponding protein levels. However, in the face of perturbations, this linear relationship should become nonlinear for maintaining cellular homeostasis. This prediction highlights the importance of studying the quantitative aspects of the central dogma in the context of robustness. For example, previous studies have investigated the robustness of gene expression level under genetic perturbations caused by an increase in gene copy number [10–12]. These efforts have demonstrated that the copy number of a subset of genes in the genome correlates with mRNA levels but not directly with protein levels. This phenomenon is known as protein-level dosage compensation, reported in yeast and mammalian cells [13–15]. Although dosage compensation is expected to contribute to cell robustness, we lack a systematic understanding of the underlying mechanisms that confer robustness to biological systems.
Systematic investigations of the robustness in cellular systems have been performed by focusing on the effects of manipulating gene copy number on cell growth [12,16–18]. We previously measured cell robustness to gene overexpression using a genetic technique termed genetic tug-of-war (gTOW), by which fragility to protein overproduction is indirectly and quantitatively assessed as an upper limit of gene copy number in Saccharomyces cerevisiae [17,19,20]. The genome-wide gTOW analysis has revealed fragile points as a set of 115 dosage-sensitive genes that cause impaired growth when the gene copy number is slightly increased [17]. In other words, only 2% of the yeast genome (115 out of 5806 genes) is sensitive to gene dosage such that a copy number increase leads to breakdown of biological systems. Conversely, this result indicates that genetic perturbations to biological processes are generally buffered. However, the buffering mechanisms behind the robustness against gene overexpression remain to be investigated.
In this study, we developed a screening system for genes with dosage compensation based on the gTOW technique. Here, our findings suggest that the proportion of the dosage-compensated genes in the genome is approximately 10% and that these genes may encode subunits of protein complexes. We investigated the compensation mechanism by focusing not only on protein degradation but also on translational efficiency by using a ribosome profiling technique [21]. Our data suggest that the robustness of gene expression reflects transient degradation, dynamic changes in protein lifetime, produced in response to environmental changes.
To identify genes with dosage compensation, we developed a screening method as shown in Fig 1A. The key idea of this method is to determine the protein level expressed from a single copy of a target gene when its copy number is increased. We monitored the level of each target protein labeled with the tandem affinity purification (TAP) tag expressed from the genomic locus when the copy number of the same target gene without the TAP tag is increased by a multicopy plasmid (Fig 1A, middle and right panels). If the expression level of the TAP-tagged protein is reduced in this situation, we consider that the target gene is subjected to dosage compensation (Fig 1A, right panel), since the compensation mechanism should not distinguish the TAP-tagged endogenous protein from the non-tagged exogenous protein. Here, we call the condition where the target protein is expressed from the single genomic copy “Single” (Fig 1A, left panel) and the condition where the target protein is expressed from the genomic copy and the multicopy plasmid “Multi” (Fig 1A, middle and right panels). We used a series of strains in which the TAP tag is integrated into the 3´-region of each gene [22], and a plasmid collection in which each target gene with native regulatory regions, including promoter and 5´ and 3´ untranslated regions, is cloned into a multicopy plasmid, pTOWug2-836 [17].
We screened 54 genes on chromosome I whose TAP-tagged strains were available as representatives of the yeast genome (S1 Fig). By this screening, we identified five genes (RBG1, MTW1, POP5, SAW1, and ERP2) whose protein expression was reduced when their copy numbers were increased (Fig 1B). We did not detect off-target effects of an increase in gene copy number by the gTOW technique: the total cellular protein level measured in the total cell lysate did not differ in the Single and Multi conditions. An example of this observation is shown in S2A and S2C Fig. Quantification of fold change of the protein levels was carried out as shown in S2 Fig. The protein levels of the dosage-compensated genes were 0.2–0.6-fold (Fig 1C), when their copy numbers were 15–27 copies (S3 Fig). The dosage compensations are performed by post-transcriptional regulation because mRNA levels from the endogenous locus did not change even when the copy numbers were increased (Fig 1D). We thus identified five genes with dosage compensation via post-transcriptional mechanisms.
To verify the experimental setup for measuring only the endogenous protein levels, we measured the level of a target protein expressed from both the genome and plasmid. The experimental setup is the same with that used for the analysis of endogenous protein except that the plasmid encodes each of the TAP-tagged target proteins (S4A Fig). We measured the total TAP-tagged protein levels (S4B and S4C Fig) and the plasmid copy numbers (S4D Fig) and calculated the fold change of the protein levels per gene copy (S4E Fig). This analysis showed dosage compensation of all the five genes identified by the chromosome I screen when considering both endogenous and exogenous protein levels (S4F Fig). The fold change values were very similar with those calculated from the endogenous protein levels. Thus, we conclude that the experimental setup shown in Fig 1A, whereby we detect the TAP-tagged protein expressed from the genomic locus, can capture dosage compensation.
We further verified the experimental setup using green fluorescent protein (GFP) tag in order to assess the dependency of dosage compensation on the TAP tag. We used the yeast strains in which the GFP tag is integrated into the 3´-region of each target gene and measured the expression levels of GFP-tagged target proteins upon an increase in gene copy number. Western blot analysis for the dosage-compensated proteins Rbg1 and Mtw1 and the uncompensated protein Pop8 showed reduced levels of Rbg1 and Mtw1 but not Pop8 in the Multi condition (S5 Fig). Because the similar degree of the compensation was observed between the analyses using the TAP and GFP tags, dosage compensation is not a TAP-tag-mediated phenomenon.
Given that dosage compensation is performed by post-transcriptional mechanisms (Fig 1D), the deceleration of protein synthesis and/or the acceleration of protein degradation should be the mechanisms of dosage compensation (S6 Fig). We first examined the contribution of protein degradation by focusing on the ubiquitin–proteasome system, a major selective degradation pathway. We used cim5-1 strain as a proteasome-defective mutant [23] to test whether the compensation is not observed in this mutant. As shown in Fig 2A and 2B, the dosage compensations of Rbg1, Mtw1, and Erp2 were significantly weaker in cim5-1 than in wild-type cells (CIM5). The compensations of Pop5 and Saw1 also tended to be weaker in cim5-1 mutant, although the difference was not statistically significant (S7 Fig). The mRNA levels of these genes in cim5-1 and CIM5 cells did not differ (S8 Fig).
To further verify the participation of the ubiquitin–proteasome system in dosage compensation, we examined the ubiquitination of the compensated proteins. The TAP-tagged proteins were immunoprecipitated with IgG-coated beads and cleaved with tobacco etch virus (TEV) protease, and the cleaved proteins were analyzed by Western blotting using anti-ubiquitin antibody (Fig 2C). Because the expression levels of the dosage-compensated proteins and the pull-down efficiency were different among the samples (Fig 2D), we normalized the ubiquitination level by dividing it by loading amount of immunoprecipitated proteins as described in Fig 2E. We compared the amount of the TAP-tagged proteins captured on the beads before and after TEV cleavage, which reflects the amount of immunoprecipitates analyzed by Western blotting for ubiquitinated proteins. This analysis showed a tendency to accumulate the greater amount of ubiquitinated proteins in cim5-1 cells upon the Multi condition (Fig 2F). These results strongly suggest that protein degradation by the ubiquitin–proteasome system is the main mechanism of dosage compensation.
We also examined the contribution of translational control to dosage compensation. A high compensation level of Pop5 in cim5-1 cells (Fig 2B) prompted us to measure the translational efficiency change upon an increase in POP5 copy number. We performed ribosome profiling and RNA-seq and measured translation rate comparing between the Single and Multi conditions of POP5 gene copy number. While a high copy number of POP5 led to an increase in its mRNA expression (Fig 3A and 3C), the ribosome density per mRNA was not changed (Fig 3B and 3C). The RNA-seq analysis also indicates that an increase in POP5 copy number by the gTOW technique specifically increased its mRNA level and did not induce off-target effects on mRNA expression of the other genes. Therefore, we conclude that translational efficiency is not responsible for dosage compensation, at least in the case of Pop5. Residual proteasome activity in cim5-1 mutant or alternative systems may specifically degrade Pop5 protein upon an increase in its gene copy number.
We noted that all the five dosage-compensated genes identified by the chromosome I screen encode subunits of protein complexes, as listed in Table 1. To investigate the relationship between dosage compensation and complex subunits, we analyzed other subunits of the complexes. As shown in Fig 4, we found that six of seven subunits of the RNase MRP and nuclear RNase P complexes, NSL1 in the MIND complex, and EMP24 in the Erp2 complex were compensated at the protein level but not at the mRNA level. Quantification showed that the degree of compensation is very similar among the six subunits of the RNase MRP and nuclear RNase P complexes (Fig 4B). As listed in Table 1, we tested an additional 12 subunit genes and identified 7 dosage-compensated ones. This ratio is significantly higher than that identified in the initial screening (5 out of 54 genes) (p < 10−9, chi-square test), although not all subunit genes are compensated. Thus, we speculated that dosage compensation predominantly targets complex subunits.
As shown above, dosage compensation may be performed mainly through protein degradation and target predominantly complex subunits. We thus hypothesized that accelerated degradation of excess subunits that failed to construct a stable complex is the nature of dosage compensation. To examine this, we focused on the Rbg1–Tma46 complex as a model complex. Our working hypothesis is that when a subunit is overexpressed, there are two pools of subunit, the unstable pool that has not found a dimerization partner and the stable pool that is in a complex (Fig 5). The unstable pool is present but very small in the native condition where a large fraction of Rbg1 molecules are stable and a stoichiometric balance between Rbg1 and Tma46 is in the steady state. In contrast, when Rbg1 is overexpressed, the unstable pool of Rbg1 is predominant. In the unstable pool, accelerated degradation of excess subunits should be observed. We first assessed the degradation of Rbg1 upon its overexpression by measuring the amount of Rbg1 after treating cells with a translational inhibitor, cycloheximide (CHX). The CHX chase assay showed accelerated degradation of Rbg1 when its gene copy number was increased (Fig 6A and 6B), as we expected.
We next tested the effect of a loss and a high copy number of TMA46 on Rbg1 expression. In tma46Δ strain, the Rbg1 expression was reduced to less than 0.5-fold (Fig 6C and S9A Fig). On the other hand, the amount of Rbg1 was increased more than 1.3-fold when the TMA46 copy number was increased in wild-type cells (Fig 6D and S9B Fig). These compensations are performed post-transcriptionally because the RBG1 mRNA levels were not changed in these conditions (Fig 6C and 6D). We further examined whether dosage compensation directly contributes to a higher or lower levels of the resulting complexes. The levels of the Rbg1–Tma46 complex were assessed by Native-PAGE followed by immunoblotting. This analysis confirmed that the complex was almost not detected in tma46Δ strain (Fig 6E). In wild-type cells, the levels of the TAP-tagged version of the Rbg1–Tma46 complex decreased and increased upon an increase in RBG1 and TMA46 copy numbers, respectively (Fig 6F). These changes in the complex levels are consistent with the changes in the Rbg1 monomer levels in the same conditions. Therefore, we conclude that Rbg1 stability is modulated depending on the dosage balance against the partner molecule Tma46 and that dosage compensation affects not only subunit levels but also complex levels.
This study extends our understanding of the rescue mechanism for perturbations causing the breakdown of biological systems. Our results demonstrate that protein-level dosage compensation is responsible for robust expression of subunit genes under genetic perturbations. Correction of the subunit levels is performed at the final step in gene expression by protein degradation rather than earlier steps, mRNA transcription/degradation or translation. These results suggest that dosage compensation at the post-translational level is a critical step to mask the fragility caused by an increase in gene copy number. Furthermore, our findings in the context of systems biology provide a new foundation for the robustness of cellular systems.
The robustness in cellular systems to gene copy number changes has been investigated mainly using two approaches: generating aneuploidy of specific chromosomes [12,18] and introducing a plasmid carrying an individual target gene [17]. The generation of aneuploid cells containing one extra chromosome doubles the number of genes in the additional chromosome. Several recent studies using aneuploid yeast and mammalian cells have revealed fragility of cellular systems against gene copy number increase in a genome-wide manner [12,18]. The use of a multicopy plasmid carrying an individual target gene dramatically increases its copy number. A particular method for this approach is based on the gTOW technique [17]. The genome-wide gTOW analysis has revealed over 80% of the yeast genome with more than 100 copies of an upper limit of gene copy number.
The impact of an increase in gene copy number on cell fitness differs between doubled number of genes in an extra chromosome and many copies of a single gene. Previous studies have demonstrated that aneuploidy-induced proteotoxic stress causes cell fragility leading to growth impairment [10,13,29]. Because aneuploid yeast strains are very sensitive to perturbations at the RNA and protein levels, aneuploidy-induced proteotoxicity affects a wide range of biological processes. On the other hand, overexpression of most individual genes does not inhibit growth of wild-type yeast strain [17,30]. Thus, the gTOW technique allows us to study mechanisms for buffering against genetic perturbations by focusing on individual target genes in normal physiological condition. We expect that exploring the effects of an increase in individual gene copy number will identify novel mechanisms for maintaining cellular homeostasis. Indeed, a very recent study has shown that the fragility of aneuploid cells is caused by many genes on single additional chromosomes but not by duplicated dosage-sensitive genes that were identified by the gTOW analysis [31].
We first developed a screening method based on the gTOW technique to estimate how much of the genome is subjected to dosage compensation (Fig 1A). Our screen of chromosome I showed that 5 out of 54 genes are regulated by the compensation (Fig 1B), which estimates that dosage compensation confers robustness to 10% of the genome for buffering perturbed gene expression. Interestingly, all screened genes encode subunits of different complexes (Table 1) and, for 17 subunits included in these complexes, 70% (12 subunits) are subjected to dosage compensation (Fig 4 and Table 1). This result is in agreement with previous findings that protein levels of duplicated genes encoding complex subunits are reduced in aneuploid yeast strains [10]. However, Mtw1 and Rpp1, the dosage-compensated proteins identified in this study, are not compensated in aneuploid cells [12]. This difference may result from aneuploidy-specific physiological conditions associated with proteotoxicity [32].
Given that the biological function of dosage compensation is to maintain subunit stoichiometry, this result explains our previous observation that cellular systems are very fragile to subunit gene overexpression [17]. This is also consistent with previous observations that the stoichiometric imbalance caused by aneuploidy strongly correlates with impaired cell growth [18,33]. Similarly, our data support a classical hypothesis called the balance hypothesis that predicts deleterious effects due to imbalanced subunit stoichiometry [34].
Recent studies investigating the robust formation of protein complexes have elucidated the location where subunits are translated [35,36], the timing when subunits are assembled into complexes [37], and the mechanisms by which subunit stoichiometry is maintained [38,39]. Li et al. found a proportional synthesis strategy whereby protein synthesis rates of complex subunits correlate with subunit stoichiometry [39]. This strategy guarantees stoichiometry of some well-characterized complexes, with a small number of exceptions synthesized in excess. In agreement with previous studies [29,38,40–43], we also identify proteasomal degradation as a mechanism of dosage compensation. We further provide direct evidence for the ubiquitination of the individual dosage-compensated proteins (Fig 2C). Thus, this study enhances our understanding of dosage compensation as a general mechanism for the fine-tuning of subunit levels.
Protein-level dosage compensation might occur cotranslationally for the following reasons: (i) Subunits are assembled into complexes cotranslationally [37]. (ii) A large proportion of the proteome is cotranslationally ubiquitinated [44,45]. (iii) The degradation of subunits via an N-terminal degradation signal at the nascent chain level has been supported by experimental evidence [38]. In addition, autophagy might be included because higher expression of autophagy-related proteins has been detected in aneuploid mammalian cells [15,18].
We show no evidence for a contribution of translational efficiency to the compensation of Pop5 protein (Fig 3B and 3C). This result supports the robust translational efficiency of duplicated genes in aneuploid yeast strains [12,33]. However, it should be noted that an increase in a single gene to approximately 20 copies does not result in a decrease in ribosome occupancy for its mRNAs (S3 Fig, Fig 3B and 3C). We speculate that translational efficiency is not responsible for dosage compensation and that translation is quite robust against genetic perturbations caused by an increase in gene copy number.
Although our screen of chromosome I suggests that the dosage-compensated genes encoding complex subunits constitute approximately 10% of the genome, subunit genes constitute 33% of the yeast genome. This suggests that there are other rules to distinguish between the compensated subunits and the uncompensated ones. Pop8 might be helpful for further characterization of the dosage compensation mechanism since the compensation level of only Pop8 differed from those of all other tested subunits of RNase MRP and nuclear RNase P complexes (Fig 4). Pop8 has the smallest number of interacting partners in these complexes, although the other subunits have at least two or more potential partners [26,46]. Therefore, Pop8 is suggested to be located at the peripheral region of these complexes. It is also known that only depletion of the Pop8 does not result in deleterious effects on RNase MRP function [46–51]. A similar observation in a different protein complex, oligosaccharyl transferase (OST), was recently reported [42]. The OST complex consists of nine subunits, including the functionally redundant Ost3 or Ost6 components, which are potentially the last subunit assembled into the complex. Overexpression of Ost3 or Ost6 does not lead to reduction of its protein level, whereas many of the other subunits show accelerated degradation upon their overexpression. Moreover, deletion of the Ost3 or Ost6 gene does not affect the protein level of the other subunits and results in only a small decrease in enzyme activity of the OST complex [42,52,53]. As listed above, characteristic features with similarities between Pop8 and Ost3 or Ost6 include the order of assembly, number of interactions, and responsibility for the function of each complex. Consideration of these features seems to provide other rules to determine the complex subunits predominantly regulated by dosage compensation.
As shown in Fig 6, we show that the compensation of Rbg1 is performed in a stoichiometry-dependent manner between gene dosage of RBG1 and TMA46. This bidirectional regulation of Rbg1 level may reflect changes in its degradation rate (Fig 6A and 6B). These results are analogous to bidirectional changes of Cog1 level upon overexpression of itself or its partner subunits: Cog2, Cog3, and Cog4 [38]. Although dosage compensation has been postulated to contribute to the levels of subunits and also resulting complexes, there might be no direct evidence for changes in the complex levels. Our study provides direct experimental evidence that dosage compensation of Rbg1 affects the levels of the Rbg1–Tma46 complex under genetic perturbations (Fig 6E and 6F).
We conclude by noting that subunit stoichiometry potentially has a broad impact on robustness in cellular systems because of the fact that numerous biological processes are dependent on protein complexes. Furthermore, studies of mechanisms behind stoichiometry maintenance might be important for understanding diseases related to gene copy number alterations. For example, a recent study suggests that a set of specific genes on trisomic chromosome 21 have a causal effect on Down syndrome [54]. Again, our approach based on the gTOW technique for measuring robustness in cellular systems provides a fundamental framework for the quantitative assessment of cell robustness.
The yeast strain BY4741 (MATa his3Δ1 leu2Δ0 met15Δ0 ura3Δ0) [55] was used for the screening, ribosome profiling, and protein complex analysis. The W303-1B (MATα ade2-1 his3-11,15 leu2-3,112 trp1-1 ura3-1 can1-100) [56] and CMY765 (MATα cim5-1 ura3-52 leu2Δ1 his3Δ200) [23] strains were used for the analysis of the ubiquitin–proteasome system. The tma46Δ strain (MATa tma46Δ::KanMX his3Δ1 leu2Δ0 met15Δ0 ura3Δ0) was also used for the protein complex analysis. TAP-tagged or GFP-tagged strains (BY4741 background) and tma46Δ strain were obtained from Thermo Scientific. These strains were transformed with empty vector pTOWug2-836 or pTOW40836 or the same vector carrying the gene of interest. Transformation of the yeast strains was performed by the lithium acetate method [57]. The transformants were grown at 30°C in SC medium lacking the indicated amino acids.
The copy number of each gene was measured using the gTOW technique, as described previously [17]. Briefly, single colonies of yeast cells carrying pTOW plasmids were cultivated in a 96-well plate containing 200 μL of SC–Ura medium for 4 days at 30°C, and then, 5 μL of the culture was inoculated into 200 μL of fresh SC–Ura medium. After culturing for 50 h at 30°C, the cells were harvested by filtration followed by DNA extraction with zymolyase treatment. The extracts were subjected to real-time quantitative PCR with Lightcycler 480 (Roche) using SYBR Green I Master (Roche) to quantify the expression of LEU3 from the chromosome and leu2d gene from pTOW plasmids. The resulting copy number of the pTOW plasmid carrying each target gene was calculated according to the method described previously [19].
Yeast cells grown in the appropriate medium were harvested at log-phase and subsequently total RNA was extracted using the hot phenol method [58]. Contaminating genomic DNA was removed and reverse transcription was carried out with PrimeScript RT reagent Kit with gDNA Eraser (TaKaRa) according to the manufacturer’s instructions. The generated cDNA was amplified by real-time quantitative PCR with Lightcycler 480 using SYBR Green I Master. Quantification of TAP tag and ACT1 mRNA expression was performed with the following primers to amplify TAP-tag and ACT1 gene on the chromosome: TAP-tag-forward (5´-AATTTCATAGCCGTCTCAGCA-3´); TAP-tag-reverse (5´-CTCGCTAGCAGTAGTTGGAATATCA-3´); ACT1-forward (5´-TGCAAACCGCTGCTCAA-3´); and ACT1-reverse (5´-TCCTTACGGACATCGACATCA-3´). The fold change of mRNA levels was calculated as previously described [11].
Yeast cells were grown in 2 mL of the appropriate medium and subcultured in 3 mL of fresh medium. The optical density at 600 nm (OD600) was measured and 2 OD600 units were harvested at log-phase. The cells were treated with 1 mL of 0.2 N NaOH for 5 min at room temperature and then were suspended in 2× NuPAGE LDS Sample Buffer (Invitrogen) and heated at 70°C for 10 min. The supernatant corresponding to 0.5 OD600 units was labeled with EzLabel FluoroNeo (ATTO) and subjected to polyacrylamide gel electrophoresis with lithium dodecyl sulfate (SDS-PAGE), followed by Western blotting with PAP (Sigma-Aldrich) (1:2000) or an anti-GFP antibody (Roche) (1:1000) and peroxidase-conjugated secondary antibody (Nichirei Biosciences) (1:1000). We used NuPAGE 4%–12% Bis-Tris Gel (Invitrogen) for SDS-PAGE and iBlot Transfer Stack PVDF membrane (Invitrogen) for Western blotting. Chemiluminescence was induced by SuperSignal West Femto Maximum Sensitivity Substrate (Thermo Scientific) and detected using LAS-4000 image analyzer (Fujifilm) and ImageQuant LAS 4000 (GE Healthcare). The band intensity was quantified using ImageQuant TL (GE Healthcare), and the fold change of protein levels was calculated as shown in S2 Fig according to a previously described method [11].
TAP-tagged strains carrying pTOW plasmid were cultivated in 100 mL of SC–Ura medium. The whole cells were harvested at log-phase and lysed with glass beads in 750 μL of lysis buffer [20 mM HEPES, 2 mM EDTA, 100 mM NaCl, 20% glycerol, 0.05% IGEPAL CA-630 (Sigma-Aldrich), Protease Inhibitor Cocktail, EDTA-Free (Thermo Scientific)] with 20 mM N-ethylmaleimide. The supernatant was immunoprecipitated using Dynabeads coated with pan-mouse IgG (Life Technologies), as described previously [59]. In short, the supernatant was incubated with 40 μL of Dynabeads in a Thermomixer Comfort (Eppendorf) at 21°C for 2 h with shaking at 1300 rpm. The Dynabeads were washed one time with the lysis buffer and three times with the lysis buffer containing 150 mM NaCl and suspended in 16 μL of AcTEV buffer (Invitrogen) containing 1 mM DTT. Before TEV cleavage, for Western blot analysis of TAP-tagged protein, 2 μL of the suspension was removed and suspended in 10 μL of 2× NuPAGE LDS Sample Buffer and heated at 65°C for 20 min. The remaining Dynabeads were then treated with 1 μL (10 units) of AcTEV protease (Invitrogen) in a Thermomixer Comfort at 4°C for 16 h with shaking at 1300 rpm. The supernatant was subjected to Western blotting with polyclonal rabbit anti-ubiquitin antibody (DAKO) (1:500) as primary antibody and peroxidase-conjugated secondary antibody (Nichirei Biosciences). After TEV cleavage, the Dynabeads were suspended in 14 μL of 2× NuPAGE LDS Sample Buffer and heated at 65°C for 20 min, and 2 μL of the extracts were mixed with 8 μL of 2× NuPAGE LDS Sample Buffer and analyzed by Western blotting with PAP. Detection of chemiluminescence was performed as described above.
Yeast cells were grown to log-phase in SC–Ura, and 0.5 OD600 units were harvested for time point 0. Then, CHX was added to a final concentration of 200 μg/mL. Cells were harvested after 1, 2, 4, and 6 h of CHX treatment, followed by total protein extraction in 2× NuPAGE LDS Sample Buffer. The supernatant corresponding to 0.1 OD600 units was analyzed by Western blotting against the TAP tag as described above. The protein level at each time point was calculated as the intensity of Rbg1-TAP from Western blot divided by that of the 50-kDa protein, corresponding to enolase, from SDS-PAGE. The relative level was calculated by dividing the protein level at each time point by that at time point 0.
Yeast cells BY4741 expressing POP5-TAP from a single genomic locus and carrying pTOWug2-836 or pTOWug2-POP5 were grown in 150 mL of SC–Ura at 30°C with vigorous shaking. These cells were grown from an initial OD600 of approximately 0.2 to OD600 around 0.7, and the cells were then harvested by vacuum filtration. The cell pellet was immediately immersed in a 50 mL conical tube filled with liquid nitrogen and 2 mL of lysis buffer [10 mM Tris-HCl (pH 7.0), 10 mM Tris-HCl (pH 8.0), 150 mM NaCl, 5 mM MgCl2, 1 mM DTT, 1% Triton X-100, 200 μg/mL CHX, 25 U/mL Turbo DNase (Invitrogen)] was dripped into the tube.
Extracts were prepared as previously described [21], except that the frozen cells were pulverized with a mixer mill at 30 Hz. The total amount of RNA in the extracts was quantified using RiboGreen (Invitrogen), and then, 50 μg of total RNA was diluted to 300 μL with the lysis buffer. The sample was subjected to preparation of ribosome footprints according to a previously described method [60]. Briefly, total RNA was treated with RNase I (Epicentre), and then the ribosomal pellet was collected by sucrose cushion centrifugation. RNA was recovered from the pellet with TRIzol (Life Technologies) and purified with Direct-zol RNA MiniPrep (Zymo), followed by isopropanol precipitation. The resulting RNA was subjected to gel electrophoresis, and then, the 26–34-nucleotides regions were excised. The size-selected fragments were subjected to dephosphorylation with T4 PNK (New England Biolabs) and linker ligation with T4 Rnl2 (New England Biolabs). Ribosomal RNA was depleted from the sample using Ribo-Zero Magnetic Gold Kit for yeast (Epicentre). Reverse transcription was carried out with Protoscript II (New England Biolabs) on the rRNA-depleted sample. The reverse transcription product was then separated by gel electrophoresis, and the full-length product was excised.
The size-selected product was circularized with CircLigaseII (Epicentre). The circularized DNA was amplified by 6, 8, 10, 12, and 14 cycles of PCR with Phusion polymerase (New England Biolabs). The PCR products were loaded on gel, and the products of eight cycles were excised. The quality of the PCR product was assessed using Agilent 2200 TapeStation (Agilent Technologies). Deep sequencing (50 bp, single-end reads) was then performed on the Illumina HiSeq 4000 (Illumina). RNA-seq libraries were generated using TruSeq Standard Total RNA Library Prep Kit (Illumina) from total RNA prepared as described above, and then, deep sequencing was performed in the same run with ribosome footprint sequencing.
The profiling analysis was performed according to the method previously described [60,61] with modifications for the analysis of budding yeast profiling. In short, rRNA sequences were aligned to a set of budding yeast rRNA sequences, and then, non-rRNA reads were aligned to the budding yeast transcriptome. A-site offsets of ribosome footprints and mRNA fragments were estimated from 13 to 17 nucleotides for each read length of 26–30 nucleotides and 15 nucleotides for 22–51 nucleotides, respectively. The mapped reads excluding the first 15 codons and last 5 codons were counted based on the A-site offsets. DESeq was used to calculate fold change of RNA expression and translational efficiency [62]. Ribosome profiling and RNA-seq data analysis did not distinguish the reads from endogenous or exogenous POP5 copies.
Yeast cells were grown to log-phase in 6 mL of the appropriate medium and 5 OD600 units were harvested. The cells were washed with 1 mL of sterile water and lysed with glass beads in 250 μL of Digitonin buffer [1% Digitonin (Invitrogen), 1× NativePAGE Sample Buffer (Invitrogen), Protease Inhibitor Cocktail, EDTA-Free]. The supernatant corresponding to 0.2 OD600 units was mixed with NativePAGE 5% G-250 Sample Additive (Invitrogen) (final concentration 0.25%) and loaded on NativePAGE 4–16% Bis-Tris Gel (Invitrogen). The native gel electrophoresis was performed at room temperature with NativePAGE Running Buffer Kit (Invitrogen) according to the manufacturer’s instructions. After electrophoresis, the gel was treated with SDS buffer [1× NuPAGE MOPS SDS Running Buffer (Invitrogen), 1% SDS] for 15 min. The gel was washed five times with 1× NuPAGE MOPS SDS Running Buffer, and then, blotted onto PVDF membrane using the iBlot system. After blotting, the membrane was washed with methanol for 5 min for three times, rinsed with PBST [1× PBS, 0.1% Tween 20] for three times, and washed in PBST for 10 min. The membrane was blocked with 4% skim milk in PBST for 1 h at room temperature before incubation with PAP (1:4000) in the same condition. Chemiluminescence was induced and detected as described above. The membrane was stained with CBB-R250 after immunoblotting.
|
10.1371/journal.pcbi.1006762 | Convergent perturbation of the human domain-resolved interactome by viruses and mutations inducing similar disease phenotypes | An important goal of systems medicine is to study disease in the context of genetic and environmental perturbations to the human interactome network. For diseases with both genetic and infectious contributors, a key postulate is that similar perturbations of the human interactome by either disease mutations or pathogens can have similar disease consequences. This postulate has so far only been tested for a few viral species at the level of whole proteins. Here, we expand the scope of viral species examined, and test this postulate more rigorously at the higher resolution of protein domains. Focusing on diseases with both genetic and viral contributors, we found significant convergent perturbation of the human domain-resolved interactome by endogenous genetic mutations and exogenous viral proteins inducing similar disease phenotypes. Pan-cancer, pan-oncovirus analysis further revealed that domains of human oncoproteins either physically targeted or structurally mimicked by oncoviruses are enriched for cancer driver rather than passenger mutations, suggesting convergent targeting of cancer driver pathways by diverse oncoviruses. Our study provides a framework for high-resolution, network-based comparison of various disease factors, both genetic and environmental, in terms of their impacts on the human interactome.
| Cellular function and behaviour are driven by highly coordinated biomolecular interaction networks. A prime example is the protein-protein interaction network, often simply referred to as the “interactome”. Recent advances in systems biology have spawned the view of human disease as a manifestation of genetic and environmental perturbations to the human interactome, a key postulate being that similar perturbation patterns lead to similar disease phenotypes. Here, we took a structural systems biology approach to compare mutation-induced and virus-induced perturbations of the human interactome in diseases with both genetic and viral contributors. Specifically, we constructed a domain-resolved human-virus protein interactome and characterized the distribution of genetic disease mutations with respect to human domains either physically targeted or structurally mimicked by virus. Overall, we found significant convergent perturbation of the human domain-resolved interactome by viruses and mutations inducing similar disease phenotypes. Structure-guided, integrated analysis of host genetic variation and host-pathogen protein interaction data may help elucidate the molecular mechanisms of infection and reveal its connections to genetic diseases such as cancer, autoimmunity, and neurodegeneration. On a broader note, our finding implies that similar perturbations of the human interactome at the domain level can have similar phenotypic consequences, regardless of the source of perturbation.
| Cellular function and behaviour are driven by highly coordinated biomolecular interaction networks. A prime example is the protein-protein interaction (PPI) network, also known as the protein “interactome” or interactome for short. A central focus of disease systems biology is to use interactome networks to study genotype-phenotype relationships in complex diseases [1]. The idea of using interactome networks to infer gene function and gene-disease association comes from the well-validated principle of “guilt by association”, which states that physically interacting proteins tend to share similar functions and, by extension, tend to be involved in similar disease processes [1–4]. Recent advances in systems biology have spawned the view of human disease as a manifestation of genetic and environmental perturbations to the human interactome, a key postulate being that similar perturbation patterns lead to similar disease phenotypes [5–8]. A corollary is that, for diseases with both genetic and infectious contributors, similar perturbations of the human interactome by either disease mutations or pathogens can have similar disease consequences. This corollary has been tested for several viral species at the level of whole proteins [9, 10]. For example, Gulbahce et al. used yeast two-hybrid screens to map binary interactions between Epstein-Barr virus (EBV) and human papillomavirus (HPV) proteins and human proteins, and transcriptionally profiled human cell lines exogenously expressing HPV oncoproteins E6 and E7 [9]. They found that human genes associated with EBV- and HPV-implicated genetic diseases were often either directly targeted by the virus or transcriptionally regulated by viral targets. This finding led to the idea that oncoviral proteins may preferentially target host proto-oncogenes and tumour suppressors, which was experimentally validated in four families of DNA oncoviruses [10].
Despite insights from these studies on the etiology of virally-implicated genetic diseases, there has yet to be a systematic, structure-based comparison of mutation-induced and pathogen-induced perturbations of the human interactome. A high-resolution, structurally-resolved network biology approach is important for unravelling complex genotype-phenotype relationships, because mutations occurring in different PPI-mediating interfaces on the same protein often have distinct functional impacts and phenotypic consequences [5–8]. In this regard, structural systems biology has proved useful in uncovering evolutionary properties of single- and multi-interface PPI network hubs, systems-level principles governing human-virus interactions, and systems properties of disease variants [6, 11, 12]. For instance, by constructing atomic-resolution human-virus and within-human protein interactomes, Franzosa and Xia discovered that viral proteins tend to target existing endogenous PPI interfaces in the human interactome, rather than creating exogenous interfaces de novo, thereby efficiently perturbing multiple endogenous PPIs involved in cell regulation [12]. In a follow-up study, Garamszegi et al. expanded the coverage of the human-virus interactome using domain-resolved models of PPIs, and found that viral proteins tend to deploy short linear motifs to bind a variety of human protein domains [13]. The economical and pleiotropic nature of “host domain-viral motif” interactions reflects the efficiency with which viruses rewire the human interactome given limited genomic resources at their disposal. Meanwhile, Wang et al. constructed a domain-resolution within-human interactome where protein domains are annotated with disease variant information [6]. They found that mutations occurring in different PPI-mediating domains within the same protein tend to be associated with different disorders (“gene pleiotropy”). By contrast, mutations occurring in the domains of two different but interacting proteins, where the interaction is mediated by said domains, tend to be associated with the same disorder (“locus heterogeneity”). These studies attest to the utility of structural systems biology in the study of infectious and genetic diseases.
Here, we apply structural systems biology to the study of virally-implicated genetic diseases (VIDs), and rigorously test the postulate that endogenous genetic mutations and exogenous viral proteins give rise to similar disease phenotypes by inducing similar perturbations of the human interactome at the level of protein domains. Specifically, we constructed a domain-resolved human-virus protein interactome and characterized the distribution of genetic disease mutations with respect to human domains targeted by virus. Overall, we found that viral proteins and VID mutations induce similar perturbations of the human domain-resolved interactome, for individual viruses with clearly defined VIDs and sufficient numbers of host-virus PPIs (including EBV, HPV and HIV), for oncoviruses, as well as for all viruses combined. We first analyzed the disease associations of host proteins targeted by viral proteins and confirmed that virus-targeted proteins tend to be causally associated with VIDs rather than non-VIDs. We then analyzed the domain-level distribution of disease mutations in virus-targeted proteins and found that virus-targeted domains are significantly enriched for mutations causing VIDs rather than non-VIDs. Using a pooled analysis of all oncoviruses and all oncomutations, we found oncovirus-targeted domains to be significantly enriched for mutations causing cancer rather than other diseases. Furthermore, domains of oncoproteins either physically targeted or structurally mimicked by oncoviruses are significantly enriched for cancer driver mutations rather than passenger mutations, which implies convergent perturbation of cancer driver pathways by diverse oncoviruses. Finally, we also assessed the extent to which viral proteins and VID mutations perturb the same domain-domain interactions (DDIs) in the human interactome. We found that viruses preferentially target DDI partners of domains harbouring VID mutations, regardless of whether the DDI partners themselves are susceptible to known disease mutations. By correlating the equivalent pathogenicity of viral proteins and VID mutations with their convergent perturbation of the human domain-resolved interactome, we provide a framework for high-resolution, network-based comparison of the functional impacts of both genetic and environmental disease factors. On a broader note, our finding implies that similar perturbations of the human interactome at the domain level can have similar phenotypic consequences, regardless of the source of perturbation.
We first acquired human-endogenous and human-virus binary PPI data from IntAct, HPIDB 3.0, and the HIV-1 Human Interaction Database [14–18]. Only PPIs supported by at least one PubMed ID were included in the whole-protein resolution human-virus interactome, which consists of 173830 PPIs between 15995 human proteins, and 28531 PPIs between 7761 human proteins and 624 viral proteins. 7211 human proteins participate in both endogenous and exogenous PPIs. To build homology models of PPIs, we collected high-confidence domain-domain interaction (DDI) and domain-motif interaction (DMI) templates derived from 3D structures of protein complexes in the Protein Data Bank, and scanned protein sequences for the occurrence of Pfam domains and domain-binding linear motifs [19–23]. Structural models were assigned to each PPI by extracting all DDIs and DMIs possibly mediating the PPI. The resulting domain-resolved human-virus structural interaction network (hvSIN) consists of 61041 PPIs between 11596 human proteins, and 4654 PPIs between 1590 human proteins and 405 viral proteins. 1517 human proteins participate in both endogenous and exogenous portions of hvSIN.
We then obtained manually-curated disease variant data from UniProtKB and ClinVar [24, 25], selecting missense variants located inside Pfam domains for our analyses. Overall, 19047 mutations associated with 5383 diseases were mapped to 3585 domains of 2622 proteins. 14720 mutations associated with 4185 diseases were mapped to 2642 domains of 1864 human proteins in hvSIN. Table 1 lists the number of mutations by the type of domain in which they occur. Incidentally, 1272 domains of 957 human proteins in hvSIN are susceptible to disease mutations, but lack interacting domains or motifs. 850 of these 1272 domains harbour a total of 4154 mutations associated with 1381 diseases that are not accounted for by mutations occurring in PPI-mediating domains in hvSIN. Because the completeness of a domain’s PPI profile depends largely on the interactome search space and availability of 3D structures of protein complexes, and domains often have important biological functions besides mediating PPIs (e.g. enzymatic or nucleotide-binding activity), we included all domains of virus-targeted host proteins in a comprehensive analysis of the domain-level distribution of disease mutations.
To relate the equivalent pathogenicity of viral proteins and VID mutations to their equivalent perturbation of the host interactome, we first characterized the mutational landscape of human proteins targeted by EBV, HPV and HIV, three viruses with clearly defined VIDs and sufficient numbers of host-virus PPIs. Since most oncoviruses are causally implicated in only a few site-specific malignancies (e.g. HBV/HCV in hepatocellular carcinoma, KSHV in Kaposi’s sarcoma, and HTLV in adult T-cell lymphoma), and various types of cancer share common molecular hallmarks [26, 27], to increase the statistical power of our analysis and establish whether a general equivalence exists between endogenous and exogenous perturbagens of oncogenic pathways, we also performed a pooled analysis of host proteins targeted by diverse oncoviruses, by considering all types of cancer as interchangeable diseases, all oncomutations as interchangeable endogenous perturbagens, and all oncoviral proteins as interchangeable exogenous perturbagens. We found that for EBV, HIV, HPV and a broad spectrum of oncoviruses, virus-targeted host proteins tend to be causally associated with VIDs (Fig 1), and virus-targeted host domains tend to harbour mutations causally associated with VIDs (Fig 2). We discuss our findings for each type of virus below. A full list of VIDs and disease-associated proteins for EBV, HPV and HIV can be found in S1 Table.
A main challenge in cancer research is to distinguish mutations which confer clonal growth advantage (i.e. drivers), from mutations that do not cause clonal expansion (i.e. passengers) [75]. Large-scale cancer genome sequencing projects have enabled systematic identification of cancer driver proteins and mutations [76]. Rozenblatt-Rosen et al. previously constructed an oncovirus-human interactome and demonstrated, at the whole-protein level, comparability between oncoviral perturbation and conventional functional genomics approaches to cancer gene discovery [10]. However, by representing proteins and PPIs as generic nodes and edges, their approach is not sensitive enough to distinguish driver mutations from passenger mutations occurring in the same oncoprotein. As we demonstrated earlier in the case of pleiotropic oncoproteins, the oncogenicity or “driver-ness” of a mutation is often correlated with its occurrence in oncovirus-targeted domains (OVTDs).
To confirm that oncoviruses can help identify driver proteins, we first cross-classified human proteins in hvSIN by whether they are oncoviral targets, and whether they are curated by the Cancer Gene Census (CGC) as being causally implicated in cancer, i.e. driver proteins [76]. Out of 727 oncoviral targets, 93 (12.8%) are in CGC, whereas out of 10897 remaining human proteins in hvSIN, 514 (4.7%) are in CGC. In other words, there is a 3-fold enrichment of driver proteins among oncoviral targets (Fisher’s exact test, two-tailed P = 3 × 10−16) (Fig 5A). Next, to find out if oncoviruses can also help identify driver mutations, we cross-classified mutations in oncoproteins by whether they are drivers or passengers, and by whether they map to OVTDs. Oncogenic and resistance mutations with a ClinVar clinical significance value of “pathogenic” or “likely pathogenic” are considered drivers, while passengers include all other missense mutations in oncoproteins that are catalogued by ClinVar and COSMIC. Out of 194 oncoproteins with annotated driver mutations, we identified 30 oncoproteins as having at least one OVTD. Pooled analysis of all 30 oncoproteins mapped 340/398 (85.4%) driver mutations and 3673/7177 (51.2%) passenger mutations to OVTDs. In other words, the odds of finding a driver mutation in OVTDs is 5 times as high as that in non-OVTDs (Fisher’s exact test, two-tailed P < 2.2 × 10−16) (Fig 5B). Closer inspection identified 19 candidates for focused investigations into the common basis of viral and mutational oncogenesis (Table 2): (I) 7 oncoproteins where all domains are OVTDs, and the driver:passenger ratio is higher than the average ratio across all oncoproteins; (II) 8 oncoproteins where some domains are OVTDs, and driver mutations are exclusively found in OVTDs; and (III) 4 oncoproteins where some domains are OVTDs, and driver mutations are significantly enriched in OVTDs (Fisher’s exact test, two-tailed P < 0.05). An example of each type of candidate is given in Fig 6.
Viruses are known to encode structural homologues that mimic host domains in order to modulate the biological activities of host targets. Such viral homology domains (VHDs) play key roles in mediating immune response (e.g. PF00048 in CMV and KSHV), apoptosis (e.g. PF00452 in EBV and KSHV), cell differentiation (e.g. PF07684 in feline leukemia virus), and protein phosphorylation (e.g. PF06734 in CMV), among other cellular processes involved in virally-implicated diseases. VHDs often compete with cellular counterparts for interaction partners, thereby rewiring host signaling networks to the virus’s advantage. Table 3 lists instances of human proteins convergently targeted by human domains and oncoviral homology domains in hvSIN.
The preceding section established that oncovirus-targeted host domains are enriched for cancer driver mutations. Here, we test the hypothesis that oncovirus-mimicked host domains are also enriched for cancer driver mutations, independent of whether they are physically targeted by the virus. To this end, we identified 21 oncoproteins having at least one oncovirus-targeted domain (OVTD) and at least one viral homology domain (VHD). We further classified viral homology domains (VHDs) into those enriched in oncogenic viruses (oncoviral homology domains, or OVHDs), versus those enriched in non-oncogenic, i.e. “generic” viruses (generic viral homology domains, or GVHDs) (Methods, S2 Table). We found that domains structurally mimicked by oncoviruses (OVHDs) are more likely to harbour driver mutations, compared to domains structurally mimicked by generic viruses (GVHDs), independent of whether the domain is physically targeted by oncoviruses (OVTD) (CMH test, common odds ratio = 2.2, P = 5 × 10−5).
We then analyzed the mutational landscape of 44 oncoproteins having at least one oncoviral homology domain (OVHD) but not physically targeted by the virus, i.e. having no OVTDs. Pooled analysis of all 44 oncoproteins mapped 245/298 (82.2%) driver mutations and 5422/9554 (56.8%) passenger mutations to OVHDs. In other words, the odds of finding a driver mutation in OVHDs is 3 times as high as that in non-OVHDs (Fisher’s exact test, two-tailed P < 2.2 × 10−16) (Fig 5B). Closer inspection identified 23 candidates for focused investigations into the common basis of viral and mutational oncogenesis (Table 4): (I) 4 oncoproteins where all domains are OVHDs, and the driver:passenger ratio is higher than the average ratio across all oncoproteins; (II) 16 oncoproteins where some domains are OVHDs, and driver mutations are exclusively found in OVHDs; and (III) 3 oncoproteins where some domains are OVHDs, and driver mutations are significantly enriched in OVHDs (Fisher’s exact test, two-tailed P < 0.05). An example of each type of candidate is given in Fig 7. In summary, oncovirus-mimicked host domains are enriched for cancer driver mutations, regardless of whether these domains are physically targeted by the virus.
Gulbahce et al. previously hypothesized, and established at the whole-protein level, that viruses and VID mutations induce similar perturbations of the human interactome [9]. Here, we test the same hypothesis at the higher resolution of protein domains, by examining whether viruses and VID mutations perturb the same domain-domain interactions (DDIs) in the human interactome. In other words, do viruses tend to target DDI partners of domains harbouring VID mutations (viral disease domain-interacting domains, or VDDiDs), rather than DDI partners of domains harbouring non-VID mutations (non-viral disease domain-interacting domains, or nVDDiDs) (Fig 8A)? As some domains can interact with both VID domains and non-VID domains, we define VDDiDs as domains that interact with at least one VID domain, and nVDDiDs as domains that exclusively interact with non-VID domains. We found that EBV and HPV exhibit a slight preference for targeting VDDiDs, although the effect sizes are not statistically significant (42/62 VDDiDs vs. 58/104 nVDDiDs for EBV, and 20/29 VDDiDs vs. 41/69 nVDDiDs for HPV). HIV targets 218/309 (70.6%) VDDiDs and 193/346 (55.8%) nVDDiDs, representing a 1.9-fold enrichment of VDDiDs among HIV-targeted domains (Fisher’s exact test, two-tailed P = 1 × 10−4). Similarly, oncoviruses target 204/285 (71.6%) VDDiDs and 164/291 (56.4%) nVDDiDs, i.e. a 1.9-fold enrichment of VDDiDs among oncovirus-targeted domains (Fisher’s exact test, two-tailed P = 1 × 10−4). Finally, a meta-analysis on the common effect of all viral proteins and all mutations causing proliferative and immunological diseases found that viruses target 424/599 (70.8%) VDDiDs and 350/551 (63.5%) nVDDiDs, i.e. a 1.4-fold enrichment of VDDiDs among virus-targeted domains (Fisher’s exact test, two-tailed P = 0.01) (Fig 8B).
Virus’s preferential targeting of VDDiDs may be confounded by the tendency for viruses to target VID domains (Fig 2), and the tendency for VID domains to interact among themselves. We therefore excluded domains susceptible to known disease mutations and examined the extent to which virus targets “non-disease” domains that interact with VID domains. We found that HIV targets 179/250 (71.6%) VDDiDs and 164/285 (57.5%) nVDDiDs that do not harbour any known disease mutation (Fisher’s exact test odds ratio = 1.9, two-tailed P = 8 × 10−4). Similarly, oncoviruses target 165/230 (71.7%) VDDiDs and 137/237 (57.8%) nVDDiDs that do not harbour any known disease mutation (Fisher’s exact test odds ratio = 1.8, two-tailed P = 2 × 10−3). Pooled analysis of all viruses found that overall, viruses target 345/481 (71.7%) VDDiDs and 295/456 (64.7%) nVDDiDs that do not harbour any known disease mutation (Fisher’s exact test odds ratio = 1.4, two-tailed P = 0.02). Virus’s preferential targeting of VDDiDs supports our hypothesis that viruses and VID mutations inducing similar disease phenotypes convergently perturb the host domain interactome, possibly unveiling core disease modules underlying clinically heterogeneous virally-implicated diseases (Fig 9).
Structural interaction networks serve as a valuable tool for understanding the molecular mechanisms of genetic diseases, as well as the fundamental differences between endogenous and exogenous PPI networks. As experimental determination of protein structure remains an arduous task, homology modelling offers an efficient alternative for the structural annotation of protein complexes. This is based on the observation that PPIs are often mediated by evolutionarily conserved structural modules, such as domains and short linear motifs [77]. Here, we reassess the role of viral proteins as surrogates for human disease variants in relating interactome network perturbation to disease phenotypes, using a domain-resolved human-virus protein interactome where human domains are annotated with disease variant information. Compared to previous work demonstrating general proximity between viral targets and VID proteins in the human interactome, our results provide a structural explanation for the equivalent pathogenicity of viral proteins and VID mutations. Whereas previous studies merely recognized the existence of viral homologues of cellular domains, we delve deeper into the functional implications of oncoviral domain homology. Our approach can readily identify domains convergently targeted or mimicked by diverse oncoviruses for focused screening of driver mutations across various types of cancer. Further characterization of cellular domains and motifs interacting with domains targeted or mimicked by viruses may uncover immune evasion strategies exploited in common by cancer cells and pathogens, and shed light on pathways dysregulated in other virally-implicated disorders.
Although most of our findings are statistically significant, there are notable differences in the enrichment of VID mutations in virus-targeted domains, both among individual viruses (EBV, HPV and HIV), as well as between single-virus analysis and pooled analysis on multiple viruses. For single-virus analysis, enrichment effect size and significance are impacted by the number of virus-host protein-protein interactions and virus-specific diseases, which ultimately determine the statistical power. Pooled analysis on all oncoviruses detected trends in the same direction as analysis on single oncoviruses (EBV and HPV), but with higher statistical power. In addition to investigator bias resulting in some viruses having a higher number of mapped virus-host PPIs, it is also possible that certain viruses prefer to perturb host regulatory network, rather than host PPI network, which is beyond the scope of this work. Compared to direct targeting of VID domains (a “first-degree” effect), viral targeting of the interaction partners of VID domains is expected to have a weaker, “second-degree” effect on the VID domains. This partly explains why results of the “first-degree” analysis on EBV and HPV (Fig 2) are stronger than those of the “second-degree” analysis (Fig 8B).
Our pooled analysis of all oncoviral targets and all oncomutations is motivated by the assumption of convergent evolution and mimicry of endogenous oncogenic mechanisms by diverse oncoviruses. There is compelling evidence of different oncoviruses complementing each other’s replication and persistence strategies, thus eliciting multiple cellular responses associated with the hallmarks of cancer. One example is primary effusion lymphoma, a disease causally linked to KSHV but also having an EBV component. While expression of KSHV lytic genes such as vIL-6 and K1 promote VEGF secretion and angiogenesis, concomitant expression of EBV latent genes confers additional anti-apoptotic properties to infected cells in the initial phase of lymphomagenesis [78, 79]. Given the paucity of context-dependent (i.e. tissue- and disease-specific) host-endogenous and host-pathogen PPI data, here we focus on establishing viral proteins and genetic mutations that induce similar disease phenotypes as generally equivalent perturbagens of the human interactome. Future work will also consider the diversity of host range and tissue tropism among different viruses, and the potentially distinct functional impacts of the same mutation in different cell types and diseases.
One potential caveat of our interactome perturbation model is its incompleteness, due to the following reasons. Firstly, current mapping of the host-virus protein interactome is far from exhaustive. Secondly, some bona fide host-virus PPIs cannot be modelled by existing domain-based interaction templates. Thirdly, virus may not interact with a host protein via PPI, but rather regulate its expression via transcriptional or epigenetic mechanisms. Lastly, our study only considers missense mutations, because domain-based analysis of interactome perturbation requires precise positioning of mutations with respect to protein domains. Missense mutations can be unambiguously mapped to individual domains, whereas other types of mutations (e.g. nonsense or frameshift) may cause more drastic changes in the protein structure and are more difficult to map to individual domains. We are aware, however, of literature suggesting that nonsense and frameshift mutations tend to occur more frequently in tumour suppressor genes than in oncogenes [80]. Effects of these mutations on the integrity of the human interactome warrant further investigation. Still, despite the incompleteness of our model, we observed significant convergent perturbation of the human domain-resolved interactome by viruses and mutations inducing similar disease phenotypes.
The advent of high-throughput biotechnology has made it possible to comprehensively characterize genomic variations in and interspecies interactions between human and microbes, which play important roles in health and disease. As more data on pathogen-implicated diseases and host-pathogen interactions emerge, our approach may be extended to the study of bacterial diseases and co-infections involving multiple pathogenic species, such as the co-pathogenesis of HIV and Mycobacterium tuberculosis. By combining these data within the framework of structural systems biology, our work sets the stage for multi-scale, integrative investigations into endogenous and exogenous perturbagens of the human interactome, thus helping to elucidate the molecular mechanisms of infection and its possible connections to genetic diseases such as cancer, autoimmunity, and neurodegeneration.
Human-endogenous and human-virus binary PPI data were obtained from IntAct [14], HPIDB [15], and the HIV-1 Human Interaction Database [16–18]. Structural templates for domain-domain and domain-motif interactions were obtained from 3did [19], iPfam [21] and ELM [20]. Protein sequences were scanned for Pfam domains using InterProScan under default settings (version 5.30–69.0) [23, 81], and for the occurrence of domain-binding motifs as defined by 3did and ELM. Domain-based interaction models were assigned to each PPI by extracting all DDIs and DMIs possibly mediating the PPI. Disease association and clinical significance of variants were obtained from UniProtKB, ClinVar, and COSMIC [24, 25, 76]. Ensembl Variant Effect Predictor (VEP v93.0) was used for extracting variant genomic location, variation class, reference allele, HGVS notations, amino acid position, overlapping Pfam domains, among other features [82]. To facilitate counting of mutational events, variants are annotated with RefSNP IDs using VEP’s check_existing flag. Variants not co-located with any known variant are merged based on identical genomic location, variation class, and shared alleles, as per NCBI guidelines for merging submitted SNPs into RefSNP clusters (https://www.ncbi.nlm.nih.gov/books/NBK44417/). Only missense mutations located inside Pfam domains were retained for analyses. Assignment of each virally-implicated disease (VID) to EBV, HPV and HIV was based on at least two literature sources (S1 Text). To minimize redundancy in disease annotation, UMLS and OMIM IDs given to subtypes of the same disease were merged into the more general Disease Ontology [83], Orphanet [84] and MeSH IDs.
Oncoviruses are as classified by CDC, IARC, and MeSH (https://www.ncbi.nlm.nih.gov/mesh/68009858). Cancer is defined as any disease whose parent terms include “DOID:162”, “ORPHA:250908”, or MeSH IDs beginning with “C04.557|C04.588|C04.619|C04.626|C04.651|C04.666|C04.682|C04.692|C04.697|C04.700|C04.730|C04.834|C04.850”. Diseases without Disease Ontology, Orphanet or MeSH IDs are manually labelled as “cancer” if their names match the following regular expression: “blastoma|cancer|carcino*|glioma|leukemia|leukaemia|lymphoma|melanoma|neoplas*|sarcoma|tumour|tumor”. Proliferative diseases have parent terms “DOID:14566”, “ORPHA:250908”, or MeSH IDs beginning with “C04”. Immunological diseases have parent terms “DOID:2914”, “ORPHA:98004”, or MeSH IDs beginning with “C20”. All statistical analyses were conducted in R [85]. Plots of domain-level distribution of disease mutations were created with Protter [86].
Pfam domain annotation for all human and viral proteins in UniProt was retrieved from InterPro (Release 69.0) [87]. We define viral homology domains (VHDs) as Pfam domains conserved between human and viral proteins. For each VHD, the likelihood of it occurring in oncoviruses was calculated as the number of oncoviruses encoding the VHD, divided by the total number of unique oncoviral species in UniProt. Similarly, the likelihood of a VHD occurring in “generic” (i.e. non-oncogenic) viruses was calculated as the number of generic viruses encoding the VHD divided by the total number of unique generic viral species in UniProt. The observed likelihood ratio (LR) of an oncovirus vs. a generic virus encoding the VHD is then the ratio of the two likelihoods. We then permuted the label “oncovirus” and “generic virus” 10000 times among viruses encoding the VHD, thereby obtaining a null distribution for the LR. An empirical p-value for the enrichment or depletion of a VHD in oncoviral proteomes was calculated according to [88]. VHDs whose observed LR > 1 and Benjamini-Hochberg adjusted p-values (q-values) < 0.1 are considered enriched in oncoviral proteomes. These VHDs and other VHDs exclusively occurring in oncoviruses are called oncoviral homology domains (OVHDs). Likewise, VHDs whose observed LR < 1 and q-values < 0.1 are considered enriched in generic viral proteomes. These VHDs and other VHDs exclusively occurring in generic viruses are called generic viral homology domains (GVHDs).
|
10.1371/journal.pbio.1002578 | A Normalization Framework for Emotional Attention | The normalization model of attention proposes that attention can affect performance by response- or contrast-gain changes, depending on the size of the stimulus and attention field. Here, we manipulated the attention field by emotional valence, negative faces versus positive faces, while holding stimulus size constant in a spatial cueing task. We observed changes in the cueing effect consonant with changes in response gain for negative faces and contrast gain for positive faces. Neuroimaging experiments confirmed that subjects’ attention fields were narrowed for negative faces and broadened for positive faces. Importantly, across subjects, the self-reported emotional strength of negative faces and positive faces correlated, respectively, both with response- and contrast-gain changes and with primary visual cortex (V1) narrowed and broadened attention fields. Effective connectivity analysis showed that the emotional valence-dependent attention field was closely associated with feedback from the dorsolateral prefrontal cortex (DLPFC) to V1. These findings indicate a crucial involvement of DLPFC in the normalization processes of emotional attention.
| Attentional selection is the mechanism by which the subset of incoming information is preferentially processed at the expense of distractors. The normalization model of attention suggests that attention-triggered modulatory effects on sensory responses in the visual cortex depend on two factors: the stimulus size and the attention field size. However, little is known regarding whether emotional attention shapes perception by means of the normalization framework. To test this hypothesis, we manipulated the attention field by emotional valence—negative faces versus positive faces—while holding the stimulus size constant in a spatial cueing task. We observed that attention increased response gain for negative faces, with the largest cueing effects occurring at high contrasts and little to no effect at low and mid-contrasts; however, attention increased contrast gain for positive faces, with the largest cueing effects occurring at mid-contrasts and little to no effect at low and high contrasts. A complementary neuroimaging experiment confirmed that subjects' attention fields were narrowed for negative faces and broadened for positive faces. Across subjects, the self-reported emotional strength of negative faces and positive faces correlated, respectively, both with response-gain and contrast-gain changes and with narrowed and broadened attention fields in the primary visual cortex. Mechanistically, we found that the emotional valence-dependent attention field was closely associated with feedback from the dorsolateral prefrontal cortex to the primary visual cortex. Our findings provide evidence for a normalization framework for emotional attention and for the critical role of feedback from the prefrontal cortex to the early visual cortex in this normalization.
| Attentional selection is the mechanism by which the subset of incoming information is preferentially processed at the expense of distractors. Numerous studies have suggested that attentional selection modulates both visual performance and neuronal activity in striate and extrastriate visual cortices [1]. However, studies have found disparate attentional selection effects on stimulus-evoked neural responses, such as the contrast-response function (CRF) [2,3]. Some have reported that attentional selection primarily enhances neural responses to high-contrast stimuli (response gain) [4–9], whereas others have reported that attentional selection primarily enhances neural responses to medium-contrast stimuli (contrast gain) [2,3,10,11]. Still others have reported that attentional selection either enhances the entire contrast range or produces a combination of both response-gain and contrast-gain changes [12–16].
The normalization model of attention suggests that these seemingly conflicting modulatory effects of attention on sensory responses in the visual cortex may depend on two factors: the stimulus size and the attention field size [6,17–19]. Changes in the relative size of these two factors can tip the balance between neuronal excitatory and inhibitory processes, thereby resulting in response-gain changes, contrast-gain changes, or various combinations of the two [19]. Specifically, this model predicts that attention increases response gain when the stimulus is large and the attention field is small and increases contrast gain when the stimulus is small and the attention field is large. Previous psychophysical [17] and electroencephalography [20] studies have reported that the pattern of both behavioral performance and steady-state visual evoked potentials is consistent with the normalization model of attention. However, little is known about whether emotional attention also shapes perception by means of the normalization framework.
Emotional stimuli, both negative and positive emotion, tend to attract attention in humans as well as other primates [21–26]. However, there is a critical distinction between the perceptual correlates of negative and positive emotions, with negative emotion narrowing and positive emotion broadening the scope of attention or perception [27–30]. For example, negative emotion shows lower and positive emotion shows higher sensory responses to unattended extrafoveal stimuli than neutral emotion [31]. The narrowing of the attention field by negative emotion is sometimes referred to as “weapon focus,” in which peripheral details of stimuli are more poorly encoded, as measured in later memory [32] and repeated adaptation [31]. Similarly, negative emotion is associated with a greater tendency to perceive local components of visuospatial stimuli [33], whereas positive emotion is associated with a greater tendency to perceive their global components [34]. Therefore, emotional stimuli, negative versus positive, offer a unique opportunity to change the size of the attention field relative to the stimulus, differentially modulating the gain of attentional selection.
Here, the size of the attention field was manipulated by emotional valence—negative faces versus positive faces—while the stimulus size was held constant, and the stimulus contrast was varied in a spatial cueing task [19,35]. We measured the gain pattern of CRFs on the spatial cueing effect derived by the emotional faces and empirically revealed an interaction between emotion and attention: gain modulation depended on emotional valence, with a change in the spatial cueing effect consonant with a change in response gain for negative faces and a change in contrast gain for positive faces. A functional magnetic resonance imaging (fMRI) experiment confirmed that subjects’ attention fields were narrowed and broadened by negative faces and positive faces, respectively, as indexed by the decreased and increased primary visual cortex (V1) responses to flanking gratings. Furthermore, the self-reported emotional strength of the emotional faces significantly correlated with the psychophysical gain modulations, and with the V1 blood oxygenation-level-dependent (BOLD) signal changes, across individual subjects. Finally, effective connectivity analysis showed that emotional valence controlled the attention field through the modulation of feedback from the dorsolateral prefrontal cortex (DLPFC) to V1. These findings indicate that emotional attention interacts with the normalization processes depending on emotional valence, which is best explained by feedback modulation to the visual cortex from DLPFC.
In the psychophysical experiment, subjects performed an orientation discrimination task on one of two target grating patches; each was presented at five different contrasts (the contrasts of both gratings were identical on any given trial and covaried across trials in random order). Covert attention (without eye movements, S1 Fig) was captured by the emotional face (negative or positive), which also modulated the attention field: negative faces narrowed and positive faces broadened the attention field (Fig 1B and 1C). A response cue at the stimulus offset indicated the target location, yielding congruent cue (the emotional face matched the response cue) and incongruent cue (mismatched) conditions (Fig 1A). Comparing performance accuracy (d′) for congruent and incongruent trials revealed the spatial cueing effect for each target contrast.
The mean d′ plotted as psychometric functions of stimulus contrast and emotional valence are shown in Fig 2A: the negative emotion yielded a pattern that qualitatively resembled response gain (left), and the positive emotion yielded a pattern that qualitatively resembled contrast gain (right). The measured psychometric function for each emotional valence (negative and positive) and each trial condition (congruent and incongruent) was fit with the standard Naka–Rushton equation [37]. The two parameters d' max (asymptotic performance at high-contrast levels) and c50 (the contrast yielding half-maximum performance) determined response gain and contrast gain, respectively. The exponent n (slope) was fixed at 2 in the current analysis [17,38]. The d' max for emotional valence (negative and positive) and trial conditions (congruent and incongruent) are shown in Fig 2B and were submitted to a repeated-measures ANOVA with emotional valence and trial condition as within-subjects factors. The main effect of emotional valence (F1, 22 = 1.734, p = 0.201) was not significant, but the main effect of the trial condition (F1, 22 = 34.971, p < 0.001) and the interaction between these two factors (F1, 22 = 13.742, p = 0.001) were both significant. Further t tests showed that the d' max of congruent trials was higher than that of incongruent trials (t22 = 14.422, p < 0.001) for negative emotion, but not for positive emotion (t22 = 0.789, p = 0.438); the d' max for negative emotion was higher than that for positive emotion in the congruent trials (t22 = 2.181, p = 0.040), but not in the incongruent trials (t22 = 0.083, p = 0.934). Similarly, for the c50 (Fig 2C), the main effect of emotional valence was not significant (F1, 22 = 1.072, p = 0.312), but the main effect of the trial condition (F1, 22 = 40.884, p < 0.001) and the interaction between these two factors (F1, 22 = 30.950, p < 0.001) were both significant. Further t tests showed that the c50 of congruent trials was lower than that of incongruent trials for positive emotion (t22 = −7.676, p < 0.001), but not for negative emotion (t22 = −1.377, p = 0.182); the c50 for negative emotion was lower than that for positive emotion in the incongruent trials (t22 = −2.172, p = 0.041), but not in the congruent trials (t22 = 0.464, p = 0.647). These results thus suggest that gain modulation of attentional selection depends on emotional valence.
To evaluate further the role of emotional valence in the gain modulation of attention, we calculated the correlation coefficients between the self-reported emotional strength of the faces and psychophysical measures (d' max and c50) across individual subjects. The self-reported emotional strength of negative faces significantly correlated with the d' max difference between congruent and incongruent trials (r = 0.536, p = 0.008, Fig 2D, left), but not with the c50 difference between congruent and incongruent trials (r = 0.014, p = 0.948, Fig 2E, left). Conversely, the self-reported emotional strength of positive faces significantly correlated with the c50 difference between congruent and incongruent trials (r = −0.536, p = 0.008, Fig 2E, right), but not with the d' max difference between congruent and incongruent trials (r = 0.205, p = 0.348, Fig 2D, right). These results thus demonstrate a close relationship between emotional valence and gain modulation of attentional selection (response-gain and contrast-gain changes in psychophysical performance). Furthermore, given that subjects performed the negative and positive sessions on two different days (the order of the two sessions was counterbalanced across subjects), we performed an additional analysis to confirm that the order of these two sessions did not influence our psychophysical results (S2 Fig).
To directly investigate whether negative emotion narrowed and positive emotion broadened subjects’ attention fields, a block-design fMRI experiment was designed to measure the V1 responses to task-irrelevant gratings (Fig 3A). Each run consisted of 12 stimulus blocks of 16 s, interleaved with 12 blank intervals of 16 s. There were 6 kinds of stimulus blocks: 2 (visual field: left/right) × 3 (emotional valence: negative/neutral/positive), and each stimulus block was randomly repeated two times in each run. For each type of emotional valence, data from the left and right visual fields were pooled together for analysis. Each stimulus block consisted of 8 trials; on each trial, a target face was centered at 4.65° eccentricity in the left or right hemifield and flanked by four gratings. The center-to-center distance between the target face and nearby gratings and between the target face and far gratings was 2.54° and 4.52°, respectively (Fig 3A and 3B). The target face and flanking gratings were presented for 0.3 s, followed by a 1.7-s fixation interval, and subjects were asked to discriminate the gender of the target face (male or female) while maintaining central fixation throughout the trial (Fig 3C). The accuracy rates (mean percent correct ± standard error of the mean [SEM]) were 91.35% ± 1.09%, 91.95% ± 0.86%, and 92.44% ± 1.15%, while the reaction times (mean reaction time ± SEM) were 813.06 ± 19.95 ms, 821.96 ± 19.09 ms, and 827.94 ± 20.16 ms for negative, neutral, and positive conditions, respectively. For these measurements, there was no significant difference (all p > 0.05) in subject performance among the three types of emotional valence of the target faces.
Regions of interest (ROIs) in V1 were defined as the cortical regions responding significantly to the target face, nearby gratings, and far gratings (Fig 3B). We focused our analysis on V1 because activated areas in extrastriate cortex that corresponded to these three different stimuli showed a great deal of overlap. BOLD signals were extracted from these ROIs and then averaged according to emotional valence. For each stimulus block, the 2 s preceding the block served as a baseline, and the mean BOLD signal from 5 s to 16 s after stimulus onset was used as a measure of the response amplitude. The BOLD amplitudes in V1 evoked by the target face and flanking gratings (nearby + far) are shown in Fig 4B and 4C, respectively, and were submitted to a repeated-measures ANOVA with emotional valence as a within-subjects factor. For the target face, the main effect of emotional valence was not significant (F2, 28 = 2.416, p = 0.112). For the flanking gratings, however, the main effect of emotional valence was significant (F2, 28 = 16.582, p = 0.001); post hoc paired t tests revealed that V1 response during the neutral condition was significantly lower than that during the positive condition (t14 = −4.165, p = 0.003) but significantly higher than that during the negative condition (t14 = 3.806, p = 0.006). We further evaluated the role of emotional valence in the modulation of V1 responses to flanking gratings and calculated the correlation coefficients between the self-reported emotional strength of the faces and fMRI measures across individual subjects. Compared to the neutral condition, the decreased BOLD signal in the negative condition and the increased BOLD signal in the positive condition correlated significantly with the self-reported emotional strength of negative faces (r = −0.746, p = 0.001, Fig 4D, left) and positive faces (r = 0.633, p = 0.011, Fig 4D, right), respectively. Moreover, these decreased and increased BOLD signals also correlated significantly with the response-gain (Fig 4F, left) and contrast-gain (Fig 4G, right) changes, respectively, in the psychophysical experiment.
Our results thus indicated that negative emotion decreased and positive emotion increased the encoding of flanking gratings, as indexed by the BOLD signal changes in V1 evoked by four gratings (nearby + far). However, at least three potential mechanisms could explain the same result: (1) emotional valence modulates the scope of perceptual encoding, with negative emotion narrowing and positive emotion broadening the attention field (S4A Fig); (2) emotional valence modulates the brain state (e.g., arousal), with negative emotion decreasing and positive emotion increasing the V1 signal (S4B Fig); or (3) a combination of hypotheses 1 and 2, with negative emotion narrowing the attention field and decreasing the V1 signal and positive emotion broadening the attention field and increasing the V1 signal (S4C Fig). Accordingly, for each emotional condition, we analyzed the BOLD amplitudes in V1 evoked by nearby gratings and far gratings separately. We hypothesized that these different mechanisms would show different patterns in V1 responses to nearby gratings and far gratings (S4 Fig). The BOLD amplitudes in V1 evoked by nearby gratings and far gratings are shown in Fig 4E (left and right, respectively) and were submitted to a repeated-measures ANOVA with emotional valence (negative, neutral, and positive) and grating distance (nearby and far) as within-subjects factors. The main effect of emotional valence (F2, 28 = 17.227, p = 0.001), the main effect of the grating distance (F1, 14 = 8.140, p = 0.013), and the interaction between these two factors (F2, 28 = 8.887, p = 0.003) were all significant. Thus, these data were submitted to a further simple effect analysis. For the nearby gratings, the main effect of emotional valence was significant (F2, 28 = 13.487, p = 0.002); post hoc paired t tests revealed that there was no significant difference between neutral and positive conditions (t14 = −1.866, p = 0.250), and both were significantly higher than the negative condition (neutral versus negative: t14 = 5.211, p < 0.001; positive versus negative: t14 = 3.672, p = 0.008). For the far gratings, the main effect of emotional valence was also significant (F2, 28 = 18.989, p < 0.001); post hoc paired t tests revealed that there was no significant difference between negative and neutral conditions (t14 = −1.900, p = 0.235), and both were significantly lower than the positive condition (negative versus positive: t14 = −4.322, p = 0.002; neutral versus positive: t14 = −6.426, p < 0.001). For both the negative and neutral conditions, the nearby gratings were significantly higher than the far gratings (negative: t14 = 2.849, p = 0.013; neutral: t14 = 3.366, p = 0.005), but significant for the positive condition (t14 = 2.160, p = 0.049). These findings are consistent with the first hypothesis that emotional valence modulates the scope of perceptual encoding in V1 by narrowing and broadening the attention field.
To examine potential cortical or subcortical area(s) that showed a consistent pattern of activation with that in V1, where negative and positive emotions modulated its responses to flanking gratings in opposite ways (Fig 4C), we performed a group analysis and did a whole-brain search for cortical and subcortical area(s) that showed opposite modulations of flanking gratings for negative and positive emotions, relative to the neutral condition. The results showed that only early visual cortical areas, the pulvinar thalamic nucleus, and DLPFC demonstrated this effect. The BOLD amplitudes in the pulvinar and DLPFC for the three types of emotional valence are shown in Fig 4H and 4J, respectively, and were submitted to a repeated-measures ANOVA with emotional valence as a within-subjects factor. The main effect in both the pulvinar (F2, 28 = 9.092, p = 0.001) and DLPFC (F2, 28 = 23.081, p < 0.001) was significant; post hoc paired t tests revealed that, for the pulvinar, the negative condition was significantly higher than that during the positive condition (t14 = 3.801, p = 0.006), but no significant difference was found between the neutral and negative conditions or between the neutral and positive conditions (all p > 0.05). For the DLPFC, however, the neutral condition was significantly lower than that during the negative condition (t14 = −5.336, p < 0.001) but significantly higher than that during the positive condition (t14 = 3.779, p = 0.006). Furthermore, we found that V1 responses to flanking gratings were significantly correlated with DLPFC responses (Fig 4K), but not with pulvinar responses (Fig 4I). Compared to the neutral condition, V1’s decreased BOLD signal in the negative condition and increased BOLD signal in the positive condition correlated significantly with DLPFC`s increased BOLD signal in the negative condition (r = −0.631, p = 0.012) and decreased BOLD signal in the positive condition (r = −0.704, p = 0.003), respectively. Taken together, these findings suggest that the modulation of the attention field size in V1 by emotional valence may be derived by feedback from DLPFC.
Additionally, to further exclude the possibility that emotional valence modulation of the attention field size in V1 could be derived from feedback from other attention-specific (i.e., the frontal eye field [FEF] and the posterior parietal cortex [PPC]) or emotion-specific (i.e., the amygdala and medial orbitofrontal cortex [mOFC]) cortical areas, we performed a supplemental analysis and found that the BOLD responses in both the amygdala and mOFC, but not in either FEF or PPC, were significantly modulated by emotional valence. For the amygdala, as well as mOFC, both the negative and positive conditions were significantly higher than the neutral condition; however, no significant difference was found between these two conditions (S5 Fig), showing an inconsistent pattern of activation with that in V1, where the negative condition was significantly lower than the positive condition.
To directly confirm whether emotional valence modulated the attention field size in V1 through the modulation of feedback from DLPFC, we used dynamic causal modeling (DCM) to examine functional changes in directional connectivity among the amygdala, DLPFC, the pulvinar, and V1 related to negative and positive emotions. The pulvinar was selected in the models since it showed a consistent pattern of activation with DLPFC (Fig 4H), while the amygdala was selected in the models since it is well known as a critical brain area for emotion processing [24], showing significantly greater responses to emotional faces than neutral faces (S5 Fig). Given the extrinsic visual input into V1, we defined seven different models with modulatory inputs (either the negative emotion or positive emotion, Fig 5A). The modulatory inputs could modulate the feedback from the amygdala (Model 1), from the pulvinar (Model 2), from both the amygdala and pulvinar (Model 3), from DLPFC (Model 4), from both the amygdala and DLPFC (Model 5), from both DLPFC and the pulvinar (Model 6), and from all three areas (Model 7) to V1. We examined these seven models for modeling the modulatory effect by negative and positive emotions and fit each of these seven models for each subject.
For negative emotion, we computed the exceedance probability of each model [39]. The result showed that Models 1 through 7 had exceedance probabilities of 2.55%, 5.14%, 2.88%, 30.45%, 14.35%, 23.57%, and 21.05%, respectively, suggesting that Model 4 was the best one to explain the modulatory effect by negative emotion (Fig 5B, up). The negative emotion significantly increased the feedback connectivity from the amygdala to both DLPFC (t14 = 2.906, p = 0.011) and the pulvinar (t14 = 2.213, p = 0.044) but decreased the feedback connectivity from DLPFC to V1 (t14 = −3.792, p = 0.002) (Fig 5C, up). For positive emotion, the exceedance probabilities of Model 1 to Model 7 were 2.89%, 3.98%, 5.37%, 32.24%, 15.60%, 21.02%, and 18.91%, respectively, suggesting that the modulatory effect by positive emotion was also best explained by Model 4 (Fig 5B, down). However, the positive emotion significantly decreased the feedback connectivity from the amygdala to both DLPFC (t14 = −2.743, p = 0.016) and the pulvinar (t14 = −2.573, p = 0.022) but increased the feedback connectivity from DLPFC to V1 (t14 = 3.923, p = 0.002) (Fig 5C, down). Furthermore, we calculated the correlation coefficients between V1 responses and the effective connection strengths (the sum of the intrinsic and modulatory connectivities) from DLPFC to V1 across individual subjects. Compared to the neutral condition, the decreased V1 BOLD signal in the negative condition (r = 0.620, p = 0.014) and the increased V1 BOLD signal in the positive condition (r = 0.587, p = 0.021) correlated significantly with feedback connectivity from DLPFC to V1 (Fig 5D). Additionally, a supplemental DCM analysis with mOFC, instead of the amygdala, showed significantly greater responses to emotional faces than neutral faces, confirming these results (S6 Fig). Together, these results further support the idea that emotional valence-dependent modulations of the attention field size in V1 may be derived by feedback from DLPFC.
This study examined whether emotional attention shapes perception via a normalization framework. The normalization model of attention proposes that attention can affect performance by response- or contrast-gain changes, depending on the stimulus size and the attention field size [19]. Previous studies have suggested that negative emotion could narrow and positive emotion could broaden the scope of perceptual encoding [29,31,32], which offers a unique opportunity to change the size of the attention field relative to the stimulus size. Here, we measured the gain pattern of CRFs on the spatial cueing effect derived from negative and positive faces. We found a change in the spatial cueing effect consistent with a change in response gain for negative faces and in contrast gain for positive faces. The fMRI experiment confirmed that emotional valence modulated the attention field in V1; negative faces decreased and positive faces increased V1 responses to flanking gratings. Importantly, across subjects, the self-reported emotional strength of negative and positive faces correlated, respectively, both with response- and contrast-gain changes and with V1 decreased and increased responses to flanking gratings. Furthermore, effective connectivity analysis showed that the V1 attention field size controlled by emotional valence was best explained by increased and decreased feedback from DLPFC to V1.
Our data provide, to our knowledge, the first neural evidence that emotional attention interacts with normalization processes depending on emotional valence. Our behavioral data can be interpreted by a hypothesis that behavioral performance is limited by the neuronal activity with an additive, independent, and identically distributed noise, and the decision-making process with a maximum-likelihood decision rule [40]. Performance accuracy d', used in both Herrmann et al. [17] and our studies, is proportional to the signal-to-noise ratio of the underlying neuronal responses. Thus, it can reflect in parallel any change in neuronal CRFs in our study. Indeed, we found that a change in the cueing effect (Fig 2A) was consonant with a change in response gain of CRF (Fig 1B) for negative faces and a change in contrast gain of CRF (Fig 1C) for positive faces. These emotional valence-dependent gain modulations of attentional selection not only are consistent with existing psychophysical [30,32,41] and brain imaging [31,42] studies suggesting that negative emotion narrows and positive emotion broadens the scope of perceptual encoding, but also support and extend the normalization model of attention [19]. This model proposes that, in the absence of attention (e.g., in the incongruent cue condition), two factors determine the firing rate of a visually responsive neuron. One is the stimulus drive (excitatory component) determined by the contrast of the stimulus placed in the receptive field of a neuron. The other is the suppressive drive (inhibitory component) determined by the summed activity of other neighboring neurons, which serves to normalize the overall spike rate of the given neuron via mutual inhibition [43,44]. Attention (e.g., in the congruent cue condition) modulates the pattern of neural activity by altering the balance between these excitatory and inhibitory components, depending on the relative sizes of the attention field to the stimulus size, and thereby exhibiting response-gain changes, contrast-gain changes, and various combinations of the two. In our study, given the fixed size of the target stimuli in the spatial cueing task, the narrowed attention field by negative emotion led to response-gain changes because attentional gain enhanced the entire stimulus drive but enhanced only the center of the suppressive drive. Conversely, the broadened attention field by positive emotion led to contrast-gain changes because attentional gain was applied equally to the stimulus and suppressive drives.
The fMRI data confirmed these emotional valence-dependent changes of the attention field; negative emotion narrowed and positive emotion broadened the attention field in V1. Importantly, this result cannot be explained by brain state changes or by the combination of brain state and attention field changes (S4 Fig). Moreover, the result cannot be explained by a number of other factors, such as low-level features, task difficulty, target face processing, or eye movement. First, the size and contrast of the flanking gratings were identical on any given trial, and the phase and orientation were random across trials, suggesting no physical difference of the gratings among the different emotional conditions. Second, during scanning, the flanking gratings were never task relevant for the subjects, who performed a gender discrimination task on the faces. There was no significant difference in subject performance among the three types of emotional valence of the faces, suggesting no difference in task difficulty. Third, the finding of no significant activation difference in V1 for the emotional faces excluded the possibility of a trade-off between attention to the target face and flanking gratings (Fig 4B). Finally, the eye movement data showed that the subjects’ eye movements were small and their eye position distributions were statistically indistinguishable for the three types of emotional valence (S3 Fig). Although our eye movement data were recorded in a psychophysics lab (outside the scanner), it should be noted that the recordings were made when subjects performed the same task as the one in the fMRI experiments. Differences in eye movements for the three types of emotional valence may be a potential confound, but it is highly unlikely since our recordings outside the scanner did not detect any such differences.
One should note that emotional valence-dependent modulations of attention fields in our study were indexed by the decreased and increased V1 responses to flanking gratings, which were irrelevant and presumably ignored while subjects attended to the target face. Thus, how does emotional valence differentially modulate V1 responses to these distractors? Previous neurophysiological and brain neuroimaging studies have implicated prefrontal areas in the filtering of distractors [45–50], and our findings are consistent with such an influence. Our findings suggest that distractor suppression by emotional valence in V1 could be associated with feedback from DLPFC. First, DLPFC responses were significantly modulated by emotional valence and showed a pattern of activation consistent with that in V1, where negative and positive emotions modulated its responses to task-irrelevant distractors in opposite ways. This consistent pattern of activation between V1 and DLPFC was also confirmed by a group analysis and a whole-brain search for cortical and subcortical area(s) that showed opposite responses for negative and positive emotions (Fig 4J). Second, V1 responses to flanking distractors were significantly predicted by DLPFC responses (Fig 4K). Finally, the DCM analysis indicated that negative emotion increased and positive emotion decreased suppression from DLPFC to V1, and these suppression effects significantly predicted the V1 responses to flanking distractors (Fig 5C and 5D).
Our study succeeded in linking emotional valence-dependent feedback from the DLPFC to V1 directly with distractor suppression. Based on our fMRI findings, in conjunction with existing neurophysiological [51], behavioral [30,32], and neuroimaging [42,52] data, we speculate that emotional valence-dependent distractor suppression is derived from DLPFC influences on the scope of inhibitory control. Inhibitory control is thought to limit the amount of information entering the focus of attention [53], which, in turn, affects the scope of attentional selection, and DLPFC is thought to play a very important role in this function [48]. We speculate that, in our study, when the emotional faces were negative, both feedback from the amygdala to DLPFC and the BOLD signal in DLPFC increased and thus decreased effective connectivity (i.e., increased suppression) from DLPFC to V1, as revealed by the DCM analysis, which then would increase the inhibitory control, in other words, increase the inhibition of ignored distractors, resulting in a narrowed focus of attention and reduced processing of the flanking gratings. Conversely, when the emotional faces were positive, both feedback from the amygdala to DLPFC and the BOLD signal in DLPFC decreased and thus increased effective connectivity (i.e., decreased suppression) from DLPFC to V1, which would decrease the inhibition of ignored distractors, resulting in a broadened scope of attention and flanking gratings that were more fully processed. It should be noted that our speculation only provides a possible mechanism for emotional valence-dependent attention field in V1, which should be tested with neurophysiological techniques in the future.
Our results also indicate that the pulvinar may be involved in emotional valence-dependent modulations of distractor suppression in V1, consistent with previous lesion [54], neurophysiological [55], and brain neuroimaging [56] studies, implicating the pulvinar’s involvement in the filtering of unwanted information. However, it is important to note that the ROIs in the pulvinar defined in our study were across dorsal and ventral parts. The dorsal pulvinar predominantly projects to areas within the frontoparietal network and superior anterior temporal cortex [57]; the ventral pulvinar, conversely, exhibits reciprocal connections with successive occipitotemporal cortical areas along the ventral processing stream [58,59]. Thus, further work is needed to use high-spatial resolution fMRI or neurophysiological techniques to parse the relative contributions of the dorsal and ventral pulvinar to emotional valence-dependent modulations of distractor suppression.
One should note that our results cannot be explained by a number of other factors, including poststimulus modulation by the response cue, greater attention directed to negative faces, or an effect of emotional faces on decisional rather than perceptual processing of the target. First, although previous studies have suggested that the poststimulus cue (for example, the response cue in our study) can influence not only subjects’ nonperceptual decision [60] but also the perception of stimuli presented before it [15,61,62], the response cue in our study was totally randomized and uninformative about the target; we thus believe that our psychophysical results cannot be explained by the response cue. Second, although previous studies have found that negative faces tend to attract more attention and show a greater response than positive faces [25,26], this effect was not obtained in our study; no significant difference in response to negative and positive faces was found in V1 (Fig 4B), the amygdala, or mOFC (S5 Fig), thus eliminating the specific impact of negative faces as a factor affecting our fMRI results. Finally, if the emotional faces (negative versus positive) affected subjects’ decisional rather than their perceptual processing of the target stimuli, then these two conditions should have produced different responses in the orbitofrontal cortex (OFC), an area critically involved in decision making [63,64]; however, no significant difference between these two conditions was found in OFC (S5 Fig), indicating that the observed difference between the negative and positive conditions was most likely caused by perceptual rather than decision-making processes. In addition, our study used the normalization model to predict and explain psychophysical performance only. Our fMRI experiment did not measure the BOLD response for different contrast levels but instead examined whether negative emotion narrowed and positive emotion broadened subjects’ attention fields. The design of our fMRI study took into account the results of several papers reporting that the attentional effect on the BOLD response is constant across different contrast levels, showing a baseline increase/additive shift rather than either a response gain or contrast gain [12,14,15]. In those studies, the attention field was manipulated by focused (narrowed attention field) and distributed (broadened attention field) cues. These two cues, however, either enhanced the entire contrast range or produced a combination of both response-gain and contrast-gain changes, indicating inconsistent predictions of the normalization model [19]. Previous studies have suggested that their results may be because BOLD signals integrate the activity across neurons showing different attention modulatory effects, which would result in various combinations of both response and contrast gains [15,16]. Moreover, attention-triggered BOLD signals can be driven by both bottom-up stimuli and top-down goals [14,65]; hence, the increased BOLD signals to the attended low and mid-contrast stimuli may be mainly driven by the top-down modulation rather than the bottom-up stimuli.
In sum, our study provides strong evidence that gain modulation of emotional attention depends on emotional valence. Negative emotion and positive emotion modulate the attention field in V1 in opposite ways, maybe depending on the increased or decreased feedback from DLPFC, thereby changing the suppression of distractors [53]. The prominent role of the prefrontal cortex in distractor suppression evident here is consistent with recent neurophysiological findings that have begun to address how prefrontal areas directly influence sensory representations to filter out distractors [49,51,66]. Identifying DLPFC as a potential neural substrate of emotional valence-dependent normalization processing of attention gives insight into how the interaction between emotion and attention shapes our experience of the world.
A total of 23 human subjects (8 males, 21–41 y old) participated in the study. All 23 participated in the psychophysical experiment, and 15 (8 males, 21–41 y old) of them participated in the fMRI experiment. All subjects were naїve to the purpose of the study. They reported normal or corrected-to-normal vision and had no known neurological, psychiatric, or visual disorders. They gave written informed consent in accordance with protocols approved by the National Institute of Mental Health (NIMH) Institutional Review Board (93-M-0170).
Forty angry, forty happy, and forty neutral faces were chosen from the NimStim Set of Facial Expressions (http://www.macbrain.org/resources.htm) [36]. All faces were masked to exclude ears, neck, hair, and hairline and were scaled to the same size (diameter: 2.2°) (Fig 1A). In the psychophysical experiment, a pair of faces were centered in the left and right hemifields at 4.65° eccentricity, one of which was an emotional face. Target gratings (spatial frequency: 4.0 cycles/°; diameter: 2.2°; phase: random) were presented at five possible contrasts: 0.03, 0.08, 0.20, 0.45, and 0.75. In the fMRI experiment, a single face (diameter: 2.2°) was centered in either the left or right hemifield at 4.65° eccentricity and was flanked by four gratings (diameter: 1.4°; spatial frequency: 4.0 cycles/°; contrast: 0.20; phase: random; orientation: randomly chosen from 0° to 180°). The center-to-center distance between the face and nearby gratings and between the face and far gratings was 2.54° and 4.52°, respectively (Fig 3A and 3B).
Visual stimuli were displayed on a BENQ LCD monitor (model: XL2420Z; refresh rate: 60 Hz; resolution: 1,920 × 1,080; size: 24 in) at a viewing distance of 57 cm. The subjects’ head position was stabilized using a chin rest. A white fixation (diameter: 0.1°) cross was always present at the center of the monitor.
Each trial began with central fixation. A pair of faces (one emotional) were presented in the left and right hemifields for 150 ms, followed by a 50 ms fixation interval. The emotional face served as a cue to attract covert spatial attention. Then, a pair of gratings (with identical contrasts) were presented for 33 ms in the left and right hemifields at 4.65° eccentricity, one of which was the target. Subjects were asked to press one of two buttons to indicate the orientation of the grating (leftward or rightward tilted) and received auditory feedback if their response was incorrect. Target location was indicated by a peripheral 100 ms response cue (0.5° white line) above one of the grating locations, but not at the grating location to avoid masking. A congruent cue was defined as a match between the emotional face location and response cue location (half the trials); an incongruent cue was defined as a mismatch (half the trials). Participants were explicitly told that the emotional faces were randomized and uninformative about the target location (Fig 1A). The experiment consisted of two sessions (negative emotion and positive emotion of the faces), with the two sessions occurring on different days; the order of the two sessions was counterbalanced across subjects. Each session consisted of 30 blocks; each block had 80 trials, from randomly interleaving 16 trials from each of the five contrasts. Contrast varied from trial to trial in randomly shuffled order, and stimuli were presented briefly (i.e., 33 ms) to avoid any possible dependence of attentional state on stimulus contrast. The attentional effect for each grating contrast was quantified as the difference between the performance accuracy (d') in the congruent and incongruent cue conditions. After each session, subjects were asked to rate (on a seven-point Likert scale) their self-perception of the emotional strength of each emotional face. For each subject, the self-reported emotional strength of positive and negative emotion was the mean rating for 40 positive and 40 negative faces, respectively.
To quantitatively examine the pattern of gain (either response or contrast gain) separately for positive emotion and negative emotion, for each subject, performance—i.e., d' = z (hit rate)–z (false alarm rate)—was assessed across experimental blocks for each contrast and each trial condition (congruent and incongruent). A rightward response to a rightward stimulus tilt was (arbitrarily) considered to be a hit, and a rightward response to a leftward stimulus was considered to be a false alarm. For each subject, the mean d' contrast response functions (CRFs) obtained for congruent and incongruent trials were fit with the standard Naka–Rushton equation [37]:
d′(c)=d′max(cn/[cn+c50n]),
where d' is performance as a function of contrast (c), d' max determines the asymptotic performance at high contrasts, c50 is the contrast corresponding to half the asymptotic performance, and n is an exponent that determines the slope of the CRFs. In this analysis, n was fixed at 2 [17,38].
Using a block design, the experiment consisted of six functional runs. Each run consisted of 12 stimulus blocks of 16 s, interleaved with 12 blank intervals of 16 s. There were 6 different stimulus blocks: 2 (visual field: left/right) × 3 (emotional valence: negative/neutral/positive). Each stimulus block was randomly repeated two times in each run, and consisted of 8 trials; on each trial, a face flanked by four gratings was presented for 0.3 s, followed by a 1.7-s fixation interval, and subjects were asked to discriminate the gender of the face (male or female) while maintaining central fixation throughout the trial (Fig 3C).
The V1 boundary was defined by a standard phase-encoded method developed by Sereno et al. [67] and Engel et al. [68], in which subjects viewed rotating wedge and expanding ring stimuli that created traveling waves of neural activity in visual cortex. A block-design scan was used to localize the ROIs in V1 corresponding to the target face, nearby gratings, and far gratings (Fig 3B). The localizer scan consisted of 12 stimulus blocks of 12 s, interleaved with 12 blank intervals of 12 s. In a stimulus block, subjects passively viewed 8-Hz flickering patches. Each block type was repeated four times in the run, which lasted 288 s.
MRI data were collected using a 3T Siemens Trio scanner with a 32-channel phase-array coil. In the scanner, the stimuli were back-projected via a video projector (refresh rate: 60 Hz; spatial resolution: 1,280 × 800) onto a translucent screen placed inside the scanner bore. Subjects viewed the stimuli through a mirror located above their eyes. The viewing distance was 115 cm. BOLD signals were measured with an echo-planar imaging sequence (TR: 2,000 ms; TE: 30 ms; FOV: 192 × 192 mm2; matrix: 64 × 64; flip angle: 90°; slice thickness: 3 mm; gap: 0 mm; number of slices: 34; slice orientation: axial). The bottom slice was positioned at the bottom of the temporal lobes. A 3D MPRAGE structural dataset (resolution: 1 ×1 × 1 mm3; TR: 2,600 ms; TE: 30 ms; FOV: 256 × 224 mm2; flip angle: 8°; number of slices: 176; slice orientation: sagittal) was collected in the same session before the functional scans. Subjects underwent two sessions, one for retinotopic mapping and the other for the main experiment.
The anatomical volume for each subject in the retinotopic mapping session was transformed into a brain space that was common for all subjects [69] and then inflated using BrainVoyager QX. Functional volumes in both sessions for each subject were preprocessed, including 3-D motion correction, linear trend removal, and high-pass (0.015 Hz) filtering using BrainVoyager QX [70]. Head motion within any fMRI session was <2 mm for all subjects. The images were then aligned to the anatomical volume from the retinotopic mapping session and transformed into Talairach space [69]. The first 8 s of BOLD signals were discarded to minimize transient magnetic saturation effects. A general linear model (GLM) procedure was used to determine the V1’s boundary and ROI analysis. V1 boundaries were delineated by a standard retinotopic mapping method [67,68]. The ROIs within V1 were defined as regions that responded more strongly to the flickering patches than to the blank screen (p < 10−3, uncorrected).
To directly confirm whether emotional valence modulated the attention field size in V1 through the modulation of feedback from DLPFC, we applied DCM analysis in SPM10 to our fMRI data [39]. For each subject and each hemisphere, using BrainVoyager QX, the amygdala and V1 (including dorsal and ventral parts) voxels were identified as those activated by the emotional block and the flanking gratings at a significance level of p < 0.005, respectively; both the pulvinar and DLPFC voxels were identified as those activated by the stimulus block at a significance level of p < 0.005. The mean Talairach coordinates of these voxels and the standard errors across subjects in the amygdala, dorsal V1, ventral V1, the pulvinar, and DLPFC were [−22 ± 1.4, −7 ± 1.0, −13 ± 1.1], [−7 ± 0.8, −93 ± 1.0, −12 ± 1.3], [−3 ± 1.1, −84 ± 1.1, −16 ± 1.2], [−18 ± 1.9, −27 ± 1.2, 7 ± 0.9], and [−44 ± 1.6, 25 ± 1.7, 29 ± 2.7] for the left hemisphere and [25 ± 1.6, −9 ± 1.0, −15 ± 1.0], [7 ± 1.1, −94 ± 0.6, −8 ± 2.2], [3 ± 0.8, −83 ± 1.1, −14 ± 1.7], [17 ± 1.7, −29 ± 1.0, 7 ± 0.9], and [45 ± 1.7, 21 ± 3.1, 31 ± 2.3] for the right hemisphere, respectively. For each subject and each hemisphere, these Talairach coordinates were converted to Montreal Neurological Institute (MNI) coordinates using the tal2mni conversion utility (http://imaging.mrc-cbu.cam.ac.uk/downloads/MNI2tal/tal2mni.m). In SPM, for each of these areas, we extracted voxels within a 4-mm sphere centered on the most significant voxel and used their time series for the DCM analysis. The estimated DCM parameters were later averaged across dorsal and ventral V1 and the two hemispheres using the Bayesian model averaging method [39].
DCMs have three sets of parameters: (1) extrinsic input into one or more regions; (2) intrinsic connectivities among the modeled regions; and (3) bilinear parameters encoding the modulations of the specified intrinsic connections by experimental manipulations [39,71,72]. The third set of parameters is used to quantify modulatory effects, which reflect increases or decreases in connectivity between two regions given some experimental manipulation, compared with the intrinsic connections between the same regions that capture connectivity in the absence of experimental manipulation. FMRI data were modeled using GLM, with regressors for negative, neutral, and positive emotions, as well as a fourth condition comprising all visual input. The fourth condition was added specifically for the DCM analysis to be used as a direct visual input. Given the extrinsic visual input into V1, we defined seven different models with modulatory inputs (either the negative emotion or positive emotion, Fig 5A). The modulatory inputs could modulate feedback from the amygdala (Model 1), from the pulvinar (Model 2), from both the amygdala and pulvinar (Model 3), from DLPFC (Model 4), from both the amygdala and DLPFC (Model 5), from both DLPFC and the pulvinar (Model 6), and from all three areas (Model 7) to V1. We examined these seven models for modeling the modulatory effect by negative and positive emotions. We fit each of these seven models for each subject. Using a hierarchical Bayesian approach [73], we compared the seven models by computing the exceedance probability of each model, i.e., the probability to which a given model is more likely than any other included model to have generated data from a randomly selected subject [39,71,72]. In the best model (Model 4), we examined the modulatory effects by negative and positive emotions.
Eye movements were recorded with an ASL EyeTrac 6000 (Applied Science Laboratories, Bedford, Massachusetts) in a psychophysics lab (outside the scanner). Its temporal resolution was 60 Hz, and its spatial resolution was 0.25°. Recording was performed when subjects performed the same task as the psychophysical and fMRI experiments. S1 Fig and S3 Fig show that subjects’ eye movements were small and statistically indistinguishable across all conditions.
|
10.1371/journal.pgen.1008316 | The PSMA8 subunit of the spermatoproteasome is essential for proper meiotic exit and mouse fertility | The ubiquitin proteasome system regulates meiotic recombination in yeast through its association with the synaptonemal complex, a ‘zipper’-like structure that holds homologous chromosome pairs in synapsis during meiotic prophase I. In mammals, the proteasome activator subunit PA200 targets acetylated histones for degradation during somatic DNA double strand break repair and during histone replacement during spermiogenesis. We investigated the role of the testis-specific proteasomal subunit α4s (PSMA8) during spermatogenesis, and found that PSMA8 was localized to and dependent on the central region of the synaptonemal complex. Accordingly, synapsis-deficient mice show delocalization of PSMA8. Moreover, though Psma8-deficient mice are proficient in meiotic homologous recombination, there are alterations in the proteostasis of several key meiotic players that, in addition to the known substrate acetylated histones, have been shown by a proteomic approach to interact with PSMA8, such as SYCP3, SYCP1, CDK1 and TRIP13. These alterations lead to an accumulation of spermatocytes in metaphase I and II which either enter massively into apoptosis or give rise to a low number of aberrant round spermatids that apoptose before histone replacement takes place.
| Proteins within the cells that are unnecessary or damaged are degraded by a large protein complex named the proteasome. The proteins to be degraded are marked by a small protein called ubiquitin. The addition of a small modification (acetyl group) to some proteins also promotes their degradation by the proteasome. Proteasomal degradation of proteins is an essential mechanism for many developmental programs including gametogenesis, a process whereby a diploid cell produces a haploid cell or gamete (sperm or egg). The mechanism by which this genome reduction occurs is called meiosis. Here, we report the study of a protein, named PSMA8 that is specific for the testis proteasome in vertebrates. Using the mouse as a model, we show that loss of PSMA8 leads to infertility in males. By co-immunoprecipitation-coupled mass spectroscopy we identified a large list of novel PSMA8 interacting proteins. We focused our functional analysis on several key meiotic proteins which were accumulated such as SYCP3, SYCP1, CDK1 and TRIP13 in addition to the known substrate of the spermatoproteasome, the acetylated histones. We suggest that the altered accumulation of these important proteins causes a disequilibrium of the meiotic division that produces apoptotic spermatocytes in metaphase I and II and also early spermatids that die soon after reaching this stage.
| Intracellular protein content is controlled through the balance between the rates of their synthesis and degradation. In eukaryotic cells, the bulk of the degradation is carried out by the ubiquitin-proteasome system (UPS). The proteasome is a multi-subunit complex that eliminates proteins, typically labeled with ubiquitin, by ATP-driven proteolysis [1]. Proteasome complexes comprise a cylindrical catalytic core particle (CP, 20S) and different regulatory particles (RPs, 19S) that regulate the access to the CP by capping it at either end [2]. The CP is composed of seven α-type subunits and seven β-type subunits arranged as a cylinder of four rings (α1–7, β1–7, β1–7, α1–7) [1, 3]. RPs are composed of 20 subunits and their association with the CP is ATP-dependent. There are four additional activators, the 11S regulator PA28α/β/γ and the ubiquitous PA200 (Psme4) regulator that stimulates protein degradation independently of ubiquitin [4] and plays a main role in acetylation-dependent degradation of somatic core histones during DNA repair and spermiogenesis [5, 6]. Hybrid proteasomes enclosing a RP at one end and an activator at the other end are also possible [7]. In addition, there are paralogs for three β-genes that are expressed only in the immunological system, which constitutes the immunoproteasome [8], and one β5t gene expressed exclusively in the thymus, which constitutes the thymoproteasome [9]. Finally, there is a meiotic paralog of the α4 subunit (Psma7), named α4s (Psma8) [10], which might provide substrate specificity and heterogeneity to the α4s-cotaning proteasome.
The proteolytic activity of the proteasome is regulated by the rate of protein ubiquitylation, but also by its association with E3 ubiquitin ligases and deubiquitinating enzymes that edit their potential substrates [11, 12]. The classical targets of the UPS are misfolded or damaged proteins and/or short-lived regulatory proteins, whose concentration is regulated by fine-tuning of their synthesis and degradation kinetics [13, 14]. Typical examples of the latter proteins are cyclins [15, 16]. More recently, it has been hypothesized but not proven that the ZMM complex (also known as the synapsis initiation complex) involved in meiotic homologous recombination is similarly regulated in the mouse [17, 18].
Meiosis is a fundamental process in sexually reproducing species that ensures the production of genetic diversity and the generation of haploid gametes from diploid progenitors [19]. This reduction in genome content is achieved by the physical connections between homologs by chiasmata [20], which are mediated by the repair of self-induced DNA double-strand breaks (DSBs) as crossing-overs (COs). Meiotic recombination takes place on proteinaceous core structures or axial elements (AEs) that scaffold the chromosomal DNA content and physically connect (synapse) homologs through the assembly of the synaptonemal complex (SC) during prophase I [21].
The UPS regulates meiotic recombination in yeast and mouse via its physical association to AEs [17, 22]. Given the unknown function that the α4s-containing proteasome plays during spermatogenesis, we explored its function in the mouse. In this study, we show that PSMA8 is localized to and dependent on the central element of the SC, and promotes the assembly of the proteasome activator PA200. Accordingly, synapsis-deficient mice show delocalization of PSMA8. Also, Psma8-deficient mice are proficient in meiotic homologous recombination, but show alterations in the proteostasis of several key meiotic players including acetylated histones, SYCP3, SYCP1, CDK1 and TRIP13, which in turn leads to an aberrant meiotic exit, accumulation of apoptotic spermatocytes in metaphase I and II, and finally early spermatid arrest long before histone replacement takes place.
Psma8 mRNA expression in mouse tissues is almost exclusively restricted to the testis (GTEx database [23] and previous studies [10]). To elucidate the cell type in which PSMA8 is expressed, we examined by western blotting testis extracts at various postnatal ages during the first wave of spermatogenesis, which progresses more synchronously than in adult mice. PSMA8 expression (using a specific antibody against the PSMA8 C-terminus [10], see Fig 1A) was first detected at P12 and increased from P14 to P20. We also used a PSMA8-R2 antibody raised against the entire recombinant PSMA8 protein, which detected the expression of both PSMA7 (already apparent at P8, before meiosis has started) and PSMA8 (Fig 1A and S1 Fig). Analysis of testis cell lines (including spermatogonium GC1-spg, Leydig cell TM3, and Sertoli cell TM4 lines), revealed the expression of PSMA7 but not PSMA8 (Fig 1A). These results indicate that its expression is restricted to cells undergoing meiosis.
To explore the subcellular localization of PSMA8, we employed the R2 antibody (PSMA7/8) since the PSMA8 C-terminus antibody did not produce any specific labeling. Double immunolabeling of PSMA8 with the AE protein SYCP3 or with SYCP1, the transverse filament protein essential for synapsis (Fig 1B and S2 Fig), revealed PSMA7/8 presence at the central region of the synaptonemal complex (super resolution imaging, Fig 1B). We validated this localization by in vivo electroporating [24] an expression plasmid encoding GFP-PSMA8 in the testis (Fig 1C). These results agree with the recent localization of the proteasome to the chromosome axes [17].
To investigate the possible dependence of PSMA8 localization on synapsis, we analyzed synaptic mutants with mild (Rec8-/- [25]) and severe (Six6os1-/- [24]) phenotypes. Mutants for the meiotic cohesin REC8 show pseudo-synapsis between sister chromatids [25], and PSMA8 was detected at these atypical synapsed-like regions (Fig 1D). In mice lacking the novel central element protein SIX6OS1, in which AEs are physically separated and unsynapsed at pachynema [24], PSMA8 signal was not restricted to their AEs and showed a broader and more disperse labeling (Fig 1D). These results indicate that PSMA8 localization to the SC central region is consequently dependent on the assembly of the SC.
To study the role of PSMA8, we generated a targeted mutation in exon 1-intron 1 of the murine Psma8 gene by CRISPR/Cas9 genome editing (S3A and S3B Fig). Homozygous mutant testes showed no PSMA8 protein expression by western blotting when analyzed using two independent polyclonal antibodies (S3C Fig). Immunofluorescence analysis of PSMA8 expression (R2 antibody, S3D Fig) revealed a weaker signal in the SC of the mutant spermatocytes than in WT spermatocytes (51% less; 4.22±1.9 WT vs 2.05±1.7 KO), likely representing PSMA7 detected by the R2 antibody (also observed in the western blot; S3C Fig). These results indicate that the generated mutation is a null allele of the Psma8 gene (herein termed Psma8-/-).
Mice lacking PSMA8 did not display any somatic abnormalities; however, male but not female mice were sterile (S1 Table). Indeed, Psma8 mutation resulted in a reduction of the testis weight (63.09% decrease; N = 6) and the absence of spermatozoa in the epididymides (Fig 2A and 2B). Histological analysis of adult Psma8-/- testes revealed the presence of apparently normal numbers of spermatogonia, spermatocytes, Sertoli cells and Leydig cells (Fig 2B). Mouse seminiferous tubules can be classified from epithelial stage I to XII by determining the groups of associated germ cell types in histological sections. Following these criteria, we found that spermatogenesis in the mutant testes proceeded normally up to diplotene in epithelial stage XI. However, the proportion of tubules at stage XII was more than 2-fold increased in the mutant sections (12.5% in mutants versus 5.4% in WT, S2A Table). Given that spermatocytes in meiotic divisions were seen to occur at epithelial stage XII, we used p-ser10-H3 (pH3) staining to analyze the number of metaphase I and II cells present in these tubules, finding an increase in the mutant (Fig 2C and S2B Table). Quantitative analysis of seminiferous tubules in squashed preparations confirmed the increase in the number of metaphase I and metaphase II cells as compared with WT testes (77% and 89% respectively, Fig 2D and S2C Table). Moreover, a large proportion of these metaphases were positive for Caspase-3 and TUNEL indicating apoptosis (Figs 2D, 3A and 3B and S2C Table).
Together with the accumulation of apoptotic meiotic divisions, other apoptotic cells could be also observed that, from their size and molecular markers of the acrosome and chromatoid body, were round spermatids (Fig 3C and S4 Fig). Indeed, seminiferous tubules in PSMA8-deficient testes sometimes contained a few surviving round spermatids. However, these round spermatids were unable to form a proper acrosome but did accumulate some PAS positive material. Apoptotic round spermatids were also seen and no elongating spermatids were observed (Fig 2B). We corroborated that round spermatids were arrested at early stages by immunolabeling for H2AL2. H2AL2 is a transition histone essential for the first replacement of histones by TNP1 and TNP2 before protamine incorporation [26]. H2AL2 was absent from mutant spermatids (S5A Fig). We also used FACs analysis of whole cells from seminiferous tubules to verify this analysis. The results obtained confirmed the presence of a small haploid compartment in Psma8-/- testes (Fig 3D and S5B Fig). We conclude from these results that PSMA8 deficiency causes the accumulation of spermatocytes in metaphase I and II which either enter massively into apoptosis or give rise to a low number of aberrant round spermatids that finally apoptose long before histone replacement takes place.
Metaphase I accumulation can occur either because of a failure to enter anaphase or because of some event taking place during prophase (SC formation, DBSs repair or chromosome recombination) that aberrantly triggers a checkpoint-mediated delay.
To test this, we first analyzed the assembly/disassembly of the SC by monitoring the distribution of SYCP1, as co-labeling of SYCP3 and SYCP1 highlights regions of synapsis in spermatocytes. We did not observe any differences in this process from zygonema to diakinesis (S6 Fig).
We next studied the kinetics of DSB repair during meiosis. Meiotic DSBs are generated by the nuclease SPO11 and are then resected to form ssDNA ends that invade into the homologous chromosome. DSBs are marked by the presence of phosphorylated H2AX (γ-H2AX) [27]. The distribution of γ-H2AX in mutant spermatocytes was similar to that found in WT cells at prophase I (S7A Fig and S3 Table). We also did not observe any differences in the distribution of RAD51, a recombinase that promotes homologous strand invasion [28], (S7B Fig and S3 Table). Because defective DNA repair ultimately abrogates CO formation [29] and because of the involvement of ubiquitylation / sumoylation in CO designation [30], we analyzed the distribution of MLH1 foci [31], a mismatch repair protein (marker of crossover sites) that functions in the resolution of joint molecules at the end of crossover formation [32]. We found a similar value between the KO (24.9±0.9 foci) and the WT (24.3±1.1 foci; S7C Fig and S3 Table). These results indicate that the repair of meiotic DSBs and synapsis/desynapsis proceed normally during prophase I in the absence of PSMA8, and is not responsible for the observed metaphase I accumulation.
We also analyzed the morphology of the metaphase I / II cells by staining for tubulin (spindle) and SYCP3. The results showed an aberrant morphology, the presence of multipolar spindles (Fig 3E), and also a striking aberrant labeling of SYCP3 at the centromeres of metaphase II chromosomes (SYCP3 labeling is barely visible in metaphase II sister kinetochores in WT cells, Fig 3F). Finally, the arrested round spermatids showed the presence of multiple patches of heterochromatin after DAPI staining (Fig 3C and S4 Fig, chromocenter fragmentation), suggesting abnormal chromosome segregation or cytokinesis.
During spermiogenesis, most of the histones are replaced by basic transition proteins, and ultimately by protamines, facilitating chromatin compaction. Hyperacetylation of core histones during this process, and especially the acetylation of H4K16, is assumed to play a pivotal role in the initiation of histone displacement and chromatin ultracondensation [33, 34]. The proteasome activator subunit PA200 targets acetylated histones for degradation during histone replacement [5].
The core subunit PSMA8 co-immunoprecipitated PA200 (S4 Table). Given the stoichiometric relationship between the CP and RP, we analyzed the expression of PA200 by immunofluorescence in the absence of PSMA8. Whilst PA200 decorated the AEs of WT spermatocytes, we failed to observe any signal in the AEs of mutants (Fig 3G and S8 Fig). In addition, we were not able to detect PA200 by mass spectrometry analysis of PSMA7/8 immunoprecipitation of Psma8-deficient testis extracts (see section Purification of PSMA8-interacting proteins, S4 Table). These results indicate that PSMA8 is necessary or promotes the assembly of PA200 to the CP. Thus, within the limits of detection, the deficiency of Psma8 leads to a drastic decrease of PA200.
To understand the acetylated-dependent degradation of histones by the proteasome [5], we measured the acetylation status of three core histones, H2AK5ac, H3ac and H4ac (pan-H4ac and H4K16ac) in chromosome spreads by double immunolabeling for SYCP3 and the corresponding acetylated histone (Fig 4A–4D and S9–S12 Figs). This procedure enables a more precise staging of the spermatocytes and is a more efficient mean to quantitate signals than peroxidase immunostaining of testis sections [5]. The loss of PSMA8 led to the accumulation of H2AK5ac, H3ac, H4ac and H4K16ac, albeit to different degrees. Results showed that the levels of H2AK5ac, H3ac, H4ac and H4K16ac were moderately higher in Psma8-/- cells, with a relative increase at late prophase I (Fig 4A–4D and S9–S12 Figs). We failed to detect staining for H2AK5ac and H3ac in spermatocytes in late diakinesis and round/arrested spermatids. In contrast, pan-H4ac and H4K16ac also labeled metaphase I chromosomes, interkinesis nuclei and round/arrested spermatids, with greater intensity in mutant than in WT cells (Fig 4C and 4D and S11 and S12 Figs). The accumulation of acetylated histones during prophase I and particularly of H4ac and H4K16ac in the arrested round spermatids suggests that the PSMA8-containing proteasomes are involved in the acetylation-dependent degradation of histones.
We next investigated the biochemical activity of testis extracts lacking PSMA8-containing proteasomes by measuring chymotrypsin-like activity (corresponding to the catalytic subunit β1), caspase-like activity (corresponding to β5) and trypsin-like activity (β3) by a standard fluorogenic assay [35] in the presence and absence of SDS (activated proteasome). Results showed that proteasomal activity in Psma8-deficient testis extracts was not noticeably different from that in WT extracts. Indeed, the trypsin-like activity was the only proteolytic function with a modest reduction in the KO (Fig 4E). Overall, these results show that the general proteasome activity of the Psma8-deficient testis is not radically changed, which is likely due to the presence of PSMA7-dependent CPs (see dataset 1 in [36]).
To ascertain the degree of activity in vivo, we first investigated the steady-state levels of protein ubiquitylation in testis during mouse meiosis. Using immunofluorescence, we analyzed spermatocytes obtained from spreads and squashed preparations with ubiquitin antibodies (Fig 4F and S13 Fig). The results showed a slight decrease of chromatin bound ubiquitylated proteins but an increase in the soluble fraction of ubiquitylated proteins during prophase I (Fig 4F and S13 Fig). These results are partially in agreement with the observed increase in the ubiquitylation state of cultured spermatocytes treated with the proteasome inhibitor MG132 (18), and suggest a specific function of the PSMA8-containing proteasomes in the controlled degradation of ubiquitylated proteins during spermatogenesis.
The composition of the CP and its RPs has previously been established by mass-spectrometric analysis of crude preparation of proteasomes from whole testes [37]. To better understand the molecular mechanism underlying the mutant phenotype, we purified PSMA7/8-interacting proteins by single-step affinity chromatography (see Material and methods for a detailed description). Most of the canonical subunits of the CP and RP were present within the more than 596 proteins of the PSMA8 proteome (S5 Table, using a conservative cut-off, see methods). In agreement with previous results, among the two activators of the testis-specific proteasome detected (PA200 and Pa28γ) [5], PA200 was the most abundant. In contrast to previous observations, we were unable to detect Pa28α and Pa28β or the inducible catalytic subunits of the immunoproteasome (β1i, β2i and β5i) [5], suggesting a very low abundance or absence. We could not detect PA200 as an interacting protein of PSMA7/8 in testis extracts from Psma8-deficient testes (S4 Table).
Among the novel proteasome-interacting proteins (PIPs) detected were chaperones including CCT6b and CCT2, ubiquitin ligases (TRIP12, NEDD4, TRIM36 and RAD18), and novel ubiquitin specific proteases (USPs) such as USP9X, USP34, USP5 and USP47 (S6 Table). We studied the proteins enriched in the immunoprecipitation through functional (gene ontology, GO) and pathway analysis (KEGG). The top GO and KEGG results were related to the proteasome and to ribonucleoproteins. Pathway analysis showed links to spermatogenesis, cell cycle, and meiosis (see S1 Text), in accordance with the observed mutant phenotype.
Interestingly, we identified meiotic proteins a priori unrelated to the UPS such as DAZL (deleted in azoospermia), SPAG1 (Sperm-associated antigen 1), SPATA5/20 (Spermatogenesis-associated protein 5/20), the tudor domain proteins TDRD1/6/9, MAEL (repressor of transposable elements), and RNF17. These PIPs could represent proteins captured during ubiquitin-dependent targeted degradation [38] and/or proteins interacting via ubiquitin-independent proteasomal degradation, as has been shown for the related subunit α4/PSMA7 [39]. Altogether, the list of novel PIPs included novel potential readers, erasers and writers of the ubiquitin code [40] of the testis-specific proteasome, reflecting its complexity. Among these PIPs, we focused our attention on the following candidates for their role in chromosome segregation and synapsis: SYCP1, TRIP13, TEX30, PIWIL1, PIWIL2 and CDK1 (S6 Table).
Among the possible interactors, we first evaluated the transverse filament protein SYCP1. Because Sycp1 mutant mice are infertile but otherwise healthy [41], we analyzed the interaction of SYCP1 with PSMA8 and its localization in mutant meiosis. We co-transfected Sycp1 with Psma8 in HEK293T cells and we detected co-immunoprecipitation between SYCP1 and PSMA8 (Fig 5A). Despite the observation that SYCP1 is properly loaded to the SC and removed from desynapsed regions (S6 Fig), we observed an abnormal accumulation of SYCP1 in Psma8-deficient metaphase I cells, (Fig 5B). These results suggest defective degradation of SYCP1 with very likely detrimental functional consequences in the exit of meiosis.
We next extended the validation analysis of the remaining candidate interactors by co-immunoprecipitation with PSMA8, making use of the same heterologous system of HEK293T cells. These included TEX30, PIWIL1, PIWIL2, CDK1 and TRIP13. All protein-protein interaction assays carried out were negative (S14A Fig) with the exceptions of the cyclin dependent kinase CDK1 and the AAA-ATPase TRIP13 (AAA-ATPases associated with diverse cellular activities; see Figs 6A and 7A). Because of the relevance of CDK1 in metaphase transition, we first determined the expression levels of CDK1 by immunofluorescence. The results showed that more CDK1 but not the related kinase CDK2 [42] could be detected in the centromeres of metaphase I chromosome from mutant cells (Fig 6B and S15A Fig; KO 0.31±0.2 vs 0.19±0.1 WT; an increase of ~ 40%). To determine whether the increased level of CDK1 corresponded to its active or inactive phosphorylated form, we used an antibody against CDK1-Tyr15-p (inactive form, Fig 6C). The results showed no differences in the labeling at the centromeres of the metaphase I chromosomes, and therefore a decrease in phospho-CDK1/total CDK1 ratio in mutant cells. Given that CDK1 must be complexed with cyclin B1 to be active, we reasoned that if higher levels of active CDK1 are present, cyclin B1 would be similarly increased. Results showed an increase of cyclin B1 at the centromeres of metaphase I chromosomes (Fig 6D). This result was congruent with the increased amount of CDK1 and CyclinB1 observed by western blot and in squashed seminiferous tubules (Fig 6E and S15B and S15C Fig). Overall, these findings suggest that loss of PSMA8 causes an increase of CDK1 / CyclinB1 which would cooperate in the accumulation of metaphase I / metaphase II that ultimately results in apoptotic metaphase plates.
We also analyzed the distribution of TRIP13, a pleiotropic ATPase that participates in meiotic DNA repair and chromosome synapsis through HORMAD interaction and somatic spindle assembly checkpoint (SAC) proficiency through MAD2 interaction [43–46]. We first performed immunofluorescence analysis of TRIP13 in Psma8-deficient and WT spermatocytes. Results using two independent antibodies showed robust labeling of the telomeres from zygonema (two dots) to pachynema (fused to a single dot) in WT cells, which declined from diplonema to diakinesis. The staining pattern was similar but enhanced in mutant spermatocytes (Fig 7B). However, the staining pattern of TRIP13 at metaphase I differed between WT and mutant cells. Specifically, it was detected at the kinetochores of Psma8-/- spermatocytes but was absent in WT cells (Fig 7B). This labeling pattern at the metaphase I kinetochores resembles TRIP13 staining in somatic cells [47]. These results thus suggest that TRIP13 accumulates in the absence of a functional PSMA8-containing proteasome.
We next analyzed several downstream effectors of TRIP13, HORMAD1, HORMAD2, and the mitotic checkpoint protein MAD2 [48–50]. No differences were observed in the HORMAD1/2 labeling pattern between WT and mutant cells (S16 Fig). It has been shown in C. elegans that in the absence of TRIP13, MAD2 recruitment to kinetochores is delayed and that in addition to its role in checkpoint silencing, TRIP13 also contributes to spindle checkpoint activation [50]. It could thus be argued that an excess of TRIP13 would increase MAD2 loading to kinetochores thereby delaying mitotic exit. We confirmed this prediction and found that MAD2 expression at the kinetochores was enhanced in Psma8-/- spermatocytes (Fig 7C), further validating a functional consequence of TRIP13 accumulation at the kinetochores.
In order to validate the substrate specificity of the PSMA8-containing proteasome in protein degradation, we analyzed the expression levels of the separase inhibitor securin (PTTG1), a well-known substrate of the somatic proteasome. Immunofluorescence analysis showed similar levels of PTTG1 in Psma8-/- and WT spermatocytes (S17 Fig). This result suggests that PSMA8-containing proteasomes are not involved in the degradation of classical ubiquitylated substrates degraded by the somatic proteasome.
To investigate the molecular basis of PSMA8 localization in the SC, and considering the alteration of SYCP3 and SYCP1 in Psma8-/- spermatocytes (Fig 3F and Fig 5B), we used a candidate gene approach to identify additional putative interactors of PSMA8. We co-transfected Psma8 with cDNAs encoding each of the known central element proteins (SIX6OS1, SYCE1, SYCE2, SYCE3, and TEX12), and the AE protein SYCP3. As positive controls, we exploited the well-known interaction between SYCE2 and TEX12 [51] (S14C Fig). Surprisingly, we detected specific co-immunoprecipitation of PSMA8 with SIX6OS1 and SYCE3 (Fig 8A and S14B Fig). We were unable to immunoprecipitate transfected SYCP3 (using several tags or antibodies against SYCP3), likely due to the highly complex structures of transfected SYCP3, which prevented to perform co-immunoprecipitation experiments. Because SYCP3 forms filamentous structures in the cytoplasm of transfected cells, termed polycomplexes [52], co-expression of an interacting protein with SYCP3 may lead to its recruitment to polycomplexes [24], an indication of protein interaction. Indeed, we obtained self assembled higher structures when Psma8 was co-transfected with Sycp3 (Fig 8B). This SYCP3-dependent cytological interaction was not observed when Psma7 was co-transfected (Fig 8B), further validating the specificity of the interaction given the extensive protein similarity between both PSMA8 and PSMA7 (92%). To validate this interaction in vivo, we performed a detailed analysis of SYCP3 in mouse mutant squashed spermatocytes, a procedure in which no solubilization or protein extraction is performed. We observed SYCP3 aggregates/polycomplexes in the Psma8-deficient spermatocytes during prophase I and metaphase I / II (Fig 8C and 8D and S7 Table). SYCP3 accumulated in metaphase II chromosomes as abnormal SYCP3 labeling at the centromeres between sister kinetochores and as aggregates in the cytosol (Fig 3F and Fig 8D). Global accumulation of SYCP3 was also observed by western blot of whole testis under high denaturing conditions (Fig 8E) [53]. Interestingly, it has been previously shown that cultured spermatocytes chemically treated with the proteasome inhibitor MG132 form SYCP3 aggregates [17]. Overall, our results suggest that SYCP3 is targeted for degradation by the PSMA8-containing proteasome and that in the absence of PSMA8 its accumulation could mediate, at least in part, the arrest and apoptosis of spermatocytes.
The testis-specific proteasome is one of the three tissue-specific proteasomes identified in mammals (together with the immunoproteasome and the thymoproteasome); however, little is known about its biochemical and physiological function. The groundbreaking work of Xiao-Bo Qiu and colleagues showing the acetyl-histone preference of the PA200 subunit of the proteasome [5] has provided novel insights into the proteasome-dependent degradation of non-ubiquitylated proteins and led to the designation of spermatoproteasome to the PA200-containing proteasome. However, following the criteria employed for the designation of the thymoproteasome, which were devised based on the restricted expression of its β5t subunit in the thymus [9] (GTEx portal), we suggest that the term spermatoproteasome be restricted exclusively to the PSMA8-containing proteasome instead of the widely expressed PA200 subunit [5].
We have shown that genetic depletion of Psma8 causes the delocalization and the drastic decrease (loss of detection) of the proteasome activator PA200 in spermatocytes. Accordingly, Psma8-deficient spermatocytes accumulate acetylated histones. PSMA8 deficiency is comparatively more severe than that of the PA200 single mutant (subfertile) and of the PA200 and PA28γ double mutant, which do not show an arrest in spermatogenesis despite being infertile in vivo but not in vitro (spermatozoa are not motile but can fertilize in vitro [54]). From a genetic analysis perspective, this result would suggest that PSMA8 has additional functions that are independent of the activators PA200 and PA28γ. Our proteomic analysis, together with other data [10], supports this notion and indicates that PSMA8-containing proteasomes can be associated with other regulators such as the 19S subunit, expanding its targets.
Beyond its role in initiation of histone replacement [34], H4K16ac is involved in the three waves of H2AX phosphorylation during prophase I [55]. We have shown that Psma8 deficiency causes the accumulation of H4ac and H4K16ac during prophase I. However, we did not observe defects in this process in the form of a different staining pattern for γ-H2AX (leptonema and zygonema), including the expansion of γ-H2AX staining to the chromatin of the sex body (in pachynema). However, the observed premature accumulation of H4K16ac at early round spermatid might cause a defect in histone removal later on in spermiogenesis if the Psma8-/- mutants spermatids would not have entered apoptosis before this event.
We have shown that spermatoproteasome deficiency causes severe defects in protein turnover of key meiotic players that affect metaphase I/II exit, but not the complex process of meiotic recombination that occurs during prophase I (CO). By using a candidate approach of PIPs, we have identified CDK1 and TRIP13 as likely crucial proteins that have an abnormal expression pattern during meiotic metaphase in mutant mice. Given the key roles of these proteins in all aspects of mitotic/meiotic division (including SAC activation), the accumulation of aberrant metaphase I/II spermatocytes in Psma8-deficient mice is to be expected.
The role of CDK1 in the metaphase-anaphase transition is complex and is multifaceted. CDK1 inhibits and activates APC/C by promoting the SAC and also by a SAC-independent mechanism [56]. The balance between these opposing functions determines cyclin B1 destruction and separase activation, giving rise to cohesin cleavage and anaphase onset [57]. Based on the normal expression levels of PTTG1 in Psma8-/- metaphase I cells, it can be argued that there is no precocious APC activation in Psma8-deficient cells (S17 Fig). Given that CDK1 activation of the SAC is dominant over the activation of APCCdc20 [58] in oocytes, we suggest that the former effect is acting on Psma8-deficient spermatocytes. The question how CDK1 promotes the SAC is still unresolved in oocytes and even less is known about this in spermatocytes
Another group of proteins found to be deregulated in spermatoproteasome-deficient mice are the SC structural proteins SYCP1 and SYCP3. The precise effect of the accumulated SYCP1 in the cytoplasm of Psma8-/- spermatocytes cannot be experimentally analyzed. However, the coiled-coil structure and self-assemblance abilities of SYCP1 strongly suggest a functionally detrimental consequence. Similarly, the presence of SYCP3 aggregates during pachynema and metaphase I mutant spermatocytes and its persistence at metaphase II centromeres, where SYCP3 is barely visible in WT cells, also suggest a detrimental effect on these cells causing their entrance into apoptosis.
We have also shown that PSMA8 is delocalized in the severe synapsis Six6os1 mutant, which is consistent with the observed co-immunoprecipitation of PSMA8 with SYCP1, SIX6OS1 and SYCE3. All the synapsis-less mutants of CE proteins failed to load properly or lacked SYCP1 and the remaining CE proteins [24, 59–61]. Thus, we would predict delocalization of the spermatoproteasome from the SC in the remaining mouse mutants of the CE proteins. Overall, our results support the idea of a physical anchorage or recruitment of the spermatoproteasome to the SC especially through SYCP3, possibly facilitated or mediated by SYCP1, SIX6OS1 and SYCE3 as their most relevant structural partners. Supporting this notion, the Zip1 transverse filament protein of the yeast SC participates in the recruitment of the proteasome to the SC [22], suggesting an evolutionary conservation of the mechanism.
Yeast mutated for a nonessential subunit of the proteasome (pre9) showed abnormal meiotic recombination, pairing and synapsis [22]. Similar but milder defects were also observed in spermatocytes cultured with a proteasome inhibitor [17]. It has been proposed that the UPS regulates the proteostatic turnover of the ZMM which is required for efficient synapsis and CO [17], through the RNF212 (E3 sumo ligase)-Hei10 (E3 ubiquitin ligase) pathway [31]. Given this, the lack of a meiotic recombination phenotype (DSBs are generated and repaired and COs are generated normally) in our Psma8-deficient mouse is surprising. It can be argued that PSMA7-containing proteasomes are still present and at the early stages of meiosis are compensating for the loss of function of Psma8. Another possible but not mutually exclusive explanation is that the main targets of the PSMA8-containing proteasome are proteins from mid-prophase I onwards.
The spermatoproteasome through its complex interactome would serve as a hub for the fine tuning of several fundamental key molecules of the spermatogenic process such as those analyzed during the present work (SYCP1, SYCP3, TRIP13, CDK1 and acetyl-histones). Our data suggest that deregulation of proteostasis of key meiotic proteins promoting cell division leads to the presence of multipolar spindles and aberrant meiotic exit. Thus, we favor an explanation in which the joint contribution of several pathways is responsible for the observed infertility.
In relation to human disease, protein degradation was one of the top cellular functions found in an unbiased differential proteomic profiling of spermatozoa proteins from infertile men with a varicocele [62]. More specifically, PSMA8 is among the top 7 in this list of proteins that are differentially expressed, suggesting a causal role in the severity of the disease. From an organismal perspective, Psma8 transcription is mainly restricted to the human testis and to some tumors like Burkit lymphoma and melanoma (TCGC database). Altogether, and considering the PSMA8 dependency of the mouse male germline, we suggest that the spermatoproteasome may be an effective target for male contraception and for the treatment of some human malignancies.
Testes were freed from the abdominal cavity and 10 μl of DNA solution (50 μg) mixed with 1μl of 10×FastGreen (Sigma Aldrich F7258) was injected into the rete testis with a DNA embryo microinjection tip. After a period of 1 h following the injection, testes were held between electrodes and four electric pulses were applied (35 V for 50 ms each pulse) using a CUY21 BEX electroporator.
Psma8-sgRNAs G71 5’- GGGCATACT CCACTTGGAAA -3’ G84 5’-ACCGCGGTAAGCTGCTCCCC-3’ targeting exon 1 and intron 1 were predicted at crispr.mit.edu. Psma8-sgRNAs were produced by cloning annealed complementary oligos at the BbsI site of pX330 (#42230, Addgene), generating PCR products containing a T7 promoter sequence that were purified (NZYtech), and then in vitro transcribed with the MEGAshortscrip T7 Transcription Kit (Life Technologies). The plasmid pST1374-NLS-flag-linker-Cas9 (#44758; Addgene) was used for generating Cas9 mRNA. After linearization with AgeI, it was transcribed and capped with the mMESSAGE mMACHINE T7 Transcription Kit (AM1345; Life Technologies). RNAs were purified using the RNeasy Mini Kit (Qiagen). RNAs (100 ng/μl Cas9 and 50ng/μl each guide RNA) were microinjected into B6/CBA F2 zygotes (hybrids between strains C57BL/6J and CBA/J) [63] at the Transgenic Facility of the University of Salamanca. Edited founders were identified by PCR amplification (Taq polymerase, NZYtech) with primers flanking exons 1 and intron 1 (Primer F 5`-CTTCTCGGTATGACAGGGCAATC-3’ and R 5’- ACTCTACCTCCACTGCCAAC CTG-3’) and either direct sequenced or subcloned into pBlueScript (Stratagene) followed by Sanger sequencing. The predicted best null mutation was selected by PCR sequencing of the targeted region of Psma8 (S3B Fig). The selected mutant allele was 166 bp long versus 222bp of the wild-type. The founder was crossed with wild-type C57BL/6J to eliminate possible unwanted off-targets. Psma8+/- heterozygous mice were re-sequenced and crossed to give rise to Psma8-/- homozygous. Genotyping was performed by analysis of the PCR products of genomic DNA with primers F and R. Mouse mutants for Rec8 and Six6os1 have been previously developed [24, 25].
For histological analysis of adult testes, mice were perfused and their testes were processed into serial paraffin sections and stained with hematoxylin-eosin or were fixed in Bouin´s fixative and stained with Periodic acid–Schiff (PAS) and hematoxylin.
Slides were visualized at room temperature using a microscope (Axioplan 2; Carl Zeiss, Inc.) with 63 × objectives with an aperture of 1.4 (Carl Zeiss, Inc.). Images were taken with a digital camera (ORCA-ER; Hamamatsu) and processed with OPENLAB 4.0.3 and Photoshop (Adobe). Quantification of fluorescence signals was performed using Image J software. Squashed preparations were visualized with a Delta vision microscopy station. Stimulated emission depletion (STED) microscopy (SP8, Leica) was used to generate the super-resolution images. Secondary antibodies for STED imaging were conjugated to Alexa 555 and 488 (Invitrogen). Slides were mounted in Prolong Antifade Gold without DAPI.
Testes were detunicated and processed for spreading using a conventional "dry-down" technique or squashing [64]. Antibody against the C-term of PSMA8 was a gift from Dr. Murata (Univ of Tokyo, Japan) and has been previously described [10]. Rabbit polyclonal antibodies against PSMA8 were developed by Proteintech (R1 and R2) against a fusion protein of poly-His with full length PSMA8 (pET vector) of mouse origin (see S1 Fig for validation) and was used to validate the immunofluorescence and western results. The primary antibodies used for immunofluorescence were rabbit αSYCP1 IgG ab15090 (1:200) (Abcam), rabbit anti-γH2AX (ser139) IgG #07–164 (1:200) (Millipore), ACA or purified human α-centromere proteins IgG 15–235 (1:5, Antibodies Incorporated), mouse αMLH1 51-1327GR (1:5, BD Biosciences), mouse αSYCP3 IgG sc-74569 (1:100), rabbit αRAD51 PC130 (1:50, Calbiochem), Mouse αCDK1 sc-54 (1:20 IF; 1:1000 wb, Santa Cruz), rabbit αCDK1 Tyr15p #4539 (1:10, Cell Signaling), rabbit αCDK2 sc-6248 (1:20, Santa Cruz), rabbit αPTTG1 serum K783 (1:20 IF, 1:1000 wb), rabbit αTRIP13 19602-1-AP (1:20, Proteintech), rabbit αH2AL2 (1:100, from Dr. Saadi Khochbin), rabbit αPA200 (1:20, Bethyl A303-880A), rabbit α-Caspase3 #9661 (1:30, Cell Signaling), rabbit αH2AK5ac ab45152 (1:20, Abcam), Rabbit αH4K16ac #07–329 (1:50 Millipore), Rabbit αH3ac (K9 and K14) #06–599 (1:20, Millipore), Rabbit αH4ac (K5, K8, K12 and K16) #06–598 (1:20, Millipore), Mouse αUbiquitin 11023 (1:20 IF, 1:1000 wb, QED Bioscience), Rabbit αHORMAD1 and αHORMAD2 and chicken anti SYCP1 (1:50, from Dr. Attila Toth; [65]), Rabbit anti p-ser10-H3 06–570 (1:100, Millipore), Mouse anti α-tubulin T9026 (1:100, Sigma), Rabbit αCyclin B1 ab72 (1:20, Abcam), Rabbit αMAD2 (1:30 provided by Dr. Stemmann), Peanut agglutinin lectin L7381 (15μg/ml, Sigma), SMC6 ab18039 (1:50, Abcam), Human αVASA 560189 (1:100, BD), Rabbit αINCENP 1186 (1:50, provided by Dr. Earnshaw). TUNEL staining of chromosome spreads was performed with the in situ cell death detection kit (Roche).
Psma8+/+ and Psma8−/− testicular cells preparation and measurement of their DNA content were performed by a standard procedure [66]. Briefly, the testes were detunicated and the seminiferous tubules were kept in 5 ml of ice-cold separation medium (DMEM supplemented with 10% FCS, 0.1 mM NEAA, 1.5 mM sodium pyruvate, 4 mM L-glutamine and 75 μg/ml ampicillin). They were treated with 0.1 mg/ml collagenase at 37°C for 10 min under mild shaking. The sedimented seminiferous tubules were washed twice with separation medium and treated for 2 min at 37°C with 2.5 μg/ml trypsin and 1 U/ml DNAse I in separation medium and transferred to ice. Afterwards, single cells were extracted from the seminiferous cords with a Pasteur pipette and filtered through a 40 μm nylon mesh. The cell suspension (2 × 106 cells/ml) was diluted 1:1 with a solution containing 0.05 mg/ml propidium iodide and 0.1 mg/ml RNAse for 15 min. Finally, the cells were analyzed through flow cytometry in a cytometer FACSCalibur and the BD Cell-Quest software. The cell cycle distribution was analyzed with the Kaluza Analysis software (Beckman Coulter).
The 26S proteasome assay was carried out in a total volume of 250 μl in 96 well plates with 2 mM ATP in 26S buffer using 100 μg of protein supernatants from whole extracts of mouse testis. Fluorescently labeled substrates employed were: succinyl-Leu-Leu-Val-Tyr-7-amino-4-methylcoumarin (Suc-LLVY-AMC), Z-Ala-Arg-Arg-AMC (Z-ARR-AMC, Bachem), and Z-Leu-Leu-Glu-AMC (Z-LLE-AMC) for the detection of the chymotrypsin- (β5 catalytic subunit), trypsin- (β2 catalytic subunit) and caspase- (β1 catalytic) like activity measurements respectively. The final substrate concentration in each assay was 100 μM.
The HEK293T, GC1-spg, Leydig TM3, and Sertoli TM4 cell lines were directly purchased at the ATCC and cultured in standard cell media. HEK293T cell line was transfected with Lipofectamine (Invitrogen) or Jetpei (PolyPlus). Cell lines were tested for mycoplasma contamination (Mycoplasma PCR ELISA, Sigma).
Full-length cDNAs encoding PSMA8, PSMA7, CDK1, SYCP1 and SIX6OS1, SYCP3, SYCE2, TEX12, TEX30, PIWIL1 and PIWIL2 were RT-PCR amplified from murine testis RNA. Full-length cDNAs were cloned into the EcoRV pcDNA3-2XFlag or SmaI pEGFP-C1 expression vectors under the CMV promoter. In frame cloning was verified by Sanger sequencing.
200 μg of antibody R1 and R2 were bound to 100 μl of sepharose beads slurry (GE Healthcare). Testis extracts were prepared in 50mM Tris HCl (pH8), 500mM NaCl, 1mM EDTA 1% tritonX-100. 20 mg of proteins extracts were incubated o/n with the Sepharose beads. Protein-bound beads were packed into columns and washed in extracting buffer for three times. Protein were eluted in 100 mM glycine pH3. The whole immunoprecipitation of PSMA8 was performed in a buffer lacking ATP and glycerol to increase the stringency of the interactors and regulators/activators subunits. HEK293T cells were transiently transfected and whole cell extracts were prepared and cleared with protein G Sepharose beads (GE Healthcare) for 1 h. The antibody was added for 2 h and immunocomplexes were isolated by adsorption to protein G-Sepharose beads o/n. After washing, the proteins were eluted from the beads with 2xSDS gel-loading buffer 100mM Tris-Hcl (pH 7), 4% SDS, 0.2% bromophenol blue, 200mM β-mercaptoethanol and 20% glycerol, and loaded onto reducing polyacrylamide SDS gels. The proteins were detected by western blotting with the indicated antibodies. Immunoprecipitations were performed using mouse αFlag IgG (5μg; F1804, Sigma-Aldrich), mouse αGFP IgG (4 μg; CSB-MA000051M0m, Cusabio), rabbit αMyc Tag IgG (4μg; #06–549, Millipore), mouse αHA.11 IgG MMS- (5μL, aprox. 10μg/1mg prot; 101R, Covance), ChromPure mouse IgG (5μg/1mg prot; 015-000-003), ChomPure rabbit IgG (5μg/1mg prot.; 011-000-003, Jackson ImmunoResearch), ChomPure goat IgG (5μg/1mg prot.; 005-000-003, Jackson ImmunoResearch). Primary antibodies used for western blotting were rabbit αFlag IgG (1:2000; F7425 Sigma-Aldrich), goat αGFP IgG (sc-5385, Santa Cruz) (1:3000), rabbit αHA IgG (H6908, Sigma-Aldrich) (1:1.000), mouse αMyc obtained from hybridoma cell myc-1-9E10.2 ATCC (1:5). Secondary horseradish peroxidase-conjugated α-mouse (715-035-150, Jackson ImmunoResearch), α-rabbit (711-035-152, Jackson ImmunoResearch), or α-goat (705-035-147, Jackson ImmunoResearch) antibodies were used at 1:5000 dilution. Antibodies were detected by using Immobilon Western Chemiluminescent HRP Substrate from Millipore. Protein extracts for the analysis of SYCP3, CDK1 and CyclinB1 were extracted in Tris-HCl 250mM, SDS10%, Glycerol 50% (denaturing buffer).
Raw MS data were analized using MaxQuant (v. 1.5.7.4) and Perseus (v. 1.5.6.0) programmes 71. Searches were generated versus the Mus musculus proteome (UP000000589, May 2017 release) and Maxquant contaminants. All FDRs were of 1%. Variable modifications taken into account were oxidation of M, acetylation of the N-term and ubiquitylation remnants di-Gly and LRGG, while fixed modifications included considered only carbamidomethylation of C. The maximum number of modifications allowed per peptide was 5. For the case of the protein group of CDK1 to 3, experimental results showed that the protein detected was CDK1. For the PSMA8 antibodies R1 and R2, ratios of their respective iBAQ intensity versus the correspondent iBAQ intensity in the control sample were calculated. Proteins with ratio higher or equal to 5 and two or more unique peptides for at least one RP antibody were selected for ulterior analysis. Additionally, in order to avoid filtering rare proteins, those with at least one unique peptide and one peptide for both Rabbit antibodies (R1 and R2) and none for anti-IgG were also selected for further analysis.
GO and KEGG over-representation tests were performed using the R package clusterProfiler [67] using standard parameters except for a FDR cutoff of 0.01. KEGG pathways where some key genes (TRIP13, CDK1, SYCP1, DDX4, SYCP3, SYCE3, SIX6OS1) operate and the role of the co-immunoprecipitated proteins were studied using the R package pathview [68].
In order to compare counts between genotypes at different stages, we used the Welch´s t-test (unequal variances t-test), which was appropriate as the count data were not highly skewed (i.e., were reasonably approximated by a normal distribution) and in most cases showed unequal variance. We applied a two-sided test in all the cases. Asterisks denote statistical significance: *p-value <0.01, **p-value <0.001 and ***p-value<0.0001.
Mice were housed in a temperature-controlled facility (specific pathogen free, spf) using individually ventilated cages, standard diet and a 12 h light/dark cycle, according to EU laws at the “Servicio de Experimentación Animal, SEA”. Mouse protocols were approved by the Ethics Committee for Animal Experimentation of the University of Salamanca (USAL). We made every effort to minimize suffering and to improve animal welfare. Blinded experiments were not possible since the phenotype was obvious between wild type and Psma8-deficient mouse for all of the experimental procedures used. No randomization methods were applied since the animals were not divided in groups/treatments. The minimum size used for each analysis was two animals/genotype.
|
10.1371/journal.pntd.0001701 | Effects of Cu/Zn Superoxide Dismutase (sod1) Genotype and Genetic Background on Growth, Reproduction and Defense in Biomphalaria glabrata | Resistance of the snail Biomphalaria glabrata to the trematode Schistosoma mansoni is correlated with allelic variation at copper-zinc superoxide dismutase (sod1). We tested whether there is a fitness cost associated with carrying the most resistant allele in three outbred laboratory populations of snails. These three populations were derived from the same base population, but differed in average resistance. Under controlled laboratory conditions we found no cost of carrying the most resistant allele in terms of fecundity, and a possible advantage in terms of growth and mortality. These results suggest that it might be possible to drive resistant alleles of sod1 into natural populations of the snail vector for the purpose of controlling transmission of S. mansoni. However, we did observe a strong effect of genetic background on the association between sod1 genotype and resistance. sod1 genotype explained substantial variance in resistance among individuals in the most resistant genetic background, but had little effect in the least resistant genetic background. Thus, epistatic interactions with other loci may be as important a consideration as costs of resistance in the use of sod1 for vector manipulation.
| Driving resistance genes into vector populations remains a promising but underused method for reducing transmission of vector-borne diseases. Understanding the genetic mechanisms governing resistance and how resistance is maintained in vector populations is essential for the development of resistant vectors as a means of eradicating vector-borne diseases. We investigated the utility of one gene (cytosolic copper-zinc superoxide dismutase - sod1) for driving resistance associated alleles into populations of the snail Biomphalaria glabrata, a vector of the trematode parasite of humans, Schistosoma mansoni. Under controlled laboratory conditions we found no evidence for costs of resistance associated with carrying the most resistant allele at sod1 (in terms of growth, fecundity, or mortality). However, we did find a strong effect of genetic background on how strongly sod1 genotype influences resistance. Thus, epistatic interactions with other loci may be as important a consideration as costs of resistance in the use of sod1 for vector manipulation in the field.
| Although vector-borne diseases account for approximately one-sixth of the global human disease burden [1], [2], we still lack effective drugs and vaccines for many of these diseases. Even when effective drugs are available, high-risk populations often cannot be adequately treated due to a lack of funding and infrastructure in the heavily impacted countries [1], [3]. Therefore, in the absence of vaccines, eradication efforts that include both drug therapy and vector control can be the most effective approach [4]. Vector control methods most often utilize chemicals for eradication [1], [4]. This approach has obvious drawbacks because it results in habitat degradation and risk of human exposure to pesticides. Also, recurrent pesticide application is often necessary because it is nearly impossible, with a single treatment, to completely remove all possible vector individuals from an epidemiologically relevant site [5].
Recent advances in understanding the genetics of host-parasite interactions have led to increased interest in driving resistance genes into susceptible vector populations [6]–[11]. In this context, the term “resistance” describes a continuously varying trait we define as the probability of becoming infected after being challenged by a parasite, rather than to mean the absolute inability to become infected (i.e. a population or genotype can have high or low average resistance). Making vector populations more resistant to infection could be a better long-term solution and an ecologically safer way of breaking transmission cycles. Unfortunately, this approach faces major population-genetic hurdles. A non-exhaustive list includes: (1) genotype-by-environment (GxE), where the performance of a gene or gene(s) of interest depends on environmental conditions such that interactions can affect how a resistance gene performs in the field versus in the lab [12]–[16], (2) parasites and hosts are genetically more variable in the field, and there can be interactions between host genotypes and parasite genotypes (genotype-by-genotype (GxG) interactions; [16]–[19]), (3) genetic background can influence how a resistance gene performs in a natural versus a lab population. In other words, the gene of interest may perform differently depending on the genomic context in which it is interacting (epistasis), and (4), there may be a cost of resistance such that natural selection in the absence of parasites favors the “wild-type” alleles that we wish to replace.
Cost of resistance may be a particularly vexing problem for resistance-gene introduction programs. Such costs have been demonstrated in many host-parasite systems (reviewed in [20]–[26]). Nevertheless, costs of resistance are not universal [8], [27]–[31], and they may be context dependent (e.g. revealed only in stressful environments; [12], [32]–[36]). Costs of resistance presumably involve a reallocation of metabolic resources between one or more of the following life-history components: reproduction, growth, and somatic maintenance/immune function [24], [26], [37], [38]. Also, the severity of the cost should depend on the particular mechanism of resistance [29], [39]. For example, it was predicted that mechanisms involving over-expression of particular genes might be among the most costly [39].
This study was designed to measure costs of resistance and epistatic effects of genetic background associated with a single locus in Biomphalaria glabrata, a snail vector of the human pathogen Schistosoma mansoni. Schistosomiasis is responsible for approximately 200,000 deaths yearly, with 200 million people infected worldwide [40]–[42]. B. glabrata is a facultative, hermaphroditic freshwater pulmonate snail that occurs throughout much of the New World tropics [43]–[45]. The B. glabrata/S. mansoni system is a well-established model for investigating host-parasite interactions in a controlled laboratory setting [46].
Resistance to S. mansoni infection in B. glabrata is highly heritable in many lab and field populations, and is almost certainly controlled by multiple loci [47]–[52]. The expression patterns of known immune-related genes have been found to differ between individuals from more resistant and less resistant strains when each is challenged with the same strain of parasite [53]–[59]. However, to date only a single locus has been identified at which allelic variation clearly associates with resistance to the parasite: copper-zinc superoxide dismutase (sod1) [60], [61]. SOD1 is a ubiquitous protein involved in several cellular functions including signaling and immune response [62]–[65]. Among the various functions of SOD1, it catalyzes the reduction of highly reactive superoxide (O2−) to hydrogen peroxide (H2O2). Hydrogen peroxide is a known cytotoxic component of the oxidative burst, which is the primary defense mechanism for parasite clearance in molluscs [46], [66], [67]. When a schistosome invades a snail, hemocytes surround the invading parasite and are thought to generate H2O2 as part of the killing mechanism [46], [66], [68]. Consistent with this hypothesis, increased H2O2 production was correlated with the difference in resistance between snails from the M-line strain and the more resistant 13–16-R1 strain [46], [68]. An sod1 allele present in the 13–16-R1 strain was over-expressed relative to the other alleles, and correlated with a more effective defense against parasite infection [46], [61], [69]. More recently, Moné et al. [70] demonstrated a correlation between the ability of certain strains of B. glabrata to produce reactive oxygen species and the anti-oxidant defenses of their respective compatible S. mansoni strains. Thus, loci involved in the oxidative burst, such as sod1, may be very important in the evolution of schistosome-snail interactions. Therefore, sod1 is a promising candidate locus for driving resistance alleles into susceptible natural populations of snails.
Although sod1 seems a favorable candidate for genetic manipulation of snail populations, there are two reasons why one might expect a cost of resistance associated with the allelic polymorphism at sod1. First, increased expression of any gene is likely to be costly [39]. Second, increased expression of sod1 might incur a cost due to increased oxidative stress on the host [71], [72]. Therefore, investigating the fitness costs associated with allelic variation at sod1 is an important first step in evaluating the potential use of sod1 for creating highly resistant vector populations in the field.
We used a population of the 13–16-R1 strain of B. glabrata that has been maintained as a large population (hundreds) in C.J. Bayne's lab at Oregon State University since the mid-1970s. The 13–16-R1 strain was reportedly created by crossing highly resistant strains of snails isolated from Brazil and Puerto Rico [47] but it has been in culture for so long in so many laboratories that its history is not entirely clear. Our population has been maintained in the absence of parasite exposure, and therefore under relaxed selective pressure in regards to resistance to S. mansoni. B. glabrata is a facultative self-fertilizing hermaphrodite such that snails will preferentially outcross when given access to a mate, but when isolated will usually reproduce through self-fertilization (e.g. [73]–[75]; our laboratory population is in Hardy-Weinberg Equilibrium for sod1 and microsatellite loci: [61], [69]; unpub. data). We recently created 52 inbred lines: we started with haphazardly picked juvenile snails and completed three generations of selfing using a single offspring from each self-fertilization event to begin the next generation. The inbred lines are mostly fixed for one of three alleles of sod1 A, B and C, as described in [61]. These lines also vary substantially for resistance within each sod1 genotypic class (AA, BB, and CC). That there are highly resistant and highly susceptible lines within each sod1 class suggests that other loci besides sod1 have a large effect in determining resistance. These inbred lines can be used to compare directly the fitness effects of carrying a specific genotype at sod1 and the effects of genetic background on the association between resistance and sod1 genotype.
Several inbred lines were used to create three outbred F2 populations, each of which was segregating for the B and C allele (Figure 1). We hereafter refer to these three F2 populations as “genetic backgrounds” because we wanted to know if the phenotypic effects of variation at sod1 depend on the genomic context in which those alleles are expressed. These F2 individuals were then used to evaluate the effects of sod1 allele on life history traits and resistance. Inbred lines were chosen so that the three populations differed in average resistance. BB and CC fixed lines were chosen because the B allele confers the highest resistance and the C allele the lowest [61]. Additionally, in hemocytes (the defense cells) the B allele is constitutively over-expressed relative to the other two alleles [69]. To create the three F2 populations, we paired an individual from an inbred line fixed for the B allele with an individual from an inbred line fixed for the C allele (BB×CC), which resulted in offspring that were heterozygous at sod1 (BC). Three unique BB and CC inbred lines were used, and each cross was completed in triplicate with unique individuals (n = 9 crosses). To compare directly the effects of carrying the BB and CC genotypes within a family and among different backgrounds, we paired heterozygous offspring from each initial cross with a heterozygous individual from a different initial cross using a factorial design. This resulted in three different F2 populations of outbred individuals that had the same sod1 genotypes, but in different genetic backgrounds (Figure 1). The F2 individuals in each of the three populations carried the BB, BC and CC genotypes in the expected (1∶2∶1) Mendelian ratios (sod1 genotypes were verified by sequencing). We used these F2 individuals to test the effects of sod1 genotype on fecundity, growth and resistance in each of the three genetic backgrounds. Our three populations (genetic backgrounds) differed in overall resistance (77.8%, 63.8%, and 38.9%), which strongly correlated with the resistance of their grandparents (the original inbred lines) (Figure 2).
For each F2 population (genetic background), a total of 72 individuals were haphazardly chosen from a pool of offspring from the final set of crosses. We exposed single juvenile snails (4–5 mm diameter) to five S. mansoni strain PR-1 miracidia in 3 mL of artificial spring water (ASW; [76]) for two hours at 26°C, in 12-well culture plates. The PR-1 strain has been maintained in Syrian hamsters and the M-line (Oregon) strain of B. glabrata snails by the Bayne lab for 36 years. Challenged individuals were then reared in moderately dark tubs in groups of 24, with three replicate tubs for each background (n = 72). We examined the snails for infection at six, nine, and eleven weeks (we rarely see shedding after 11 weeks). Each examination week we induced cercarial shedding (parasite emergence) by exposing snails individually in 3 mL of ASW to direct fluorescent light for two hours at 26°C in 12-well culture plates. The presence of cercarial shedding indicated a positive infection. Infected snails were preserved in 95% ethanol (EtOH), and non-infected snails were returned to rearing tubs after each assay. After the final cercarial shedding attempt (eleven weeks) we preserved the remaining snails, and all tissue samples were processed for sod1 genotyping (described below in ‘Molecular Methods’ section). Resistance to parasite infection was scored in each tub group as the percentage of snails that did not shed cercariae by eleven weeks post-challenge. Snails that died prior to shedding assays were excluded from the experiment. Average mortality observed from the parasite challenge ranged from 8–12% among tubs, and did not differ among genetic backgrounds (One-way ANOVA, p = 0.442).
We collected single egg masses (n = 58) from Styrofoam substrate within 48-hours of egg mass deposition from individual pairs of the final set of crosses (i.e. embryos in the eggs are F2s). The single egg masses were reared individually and allowed to hatch. We measured offspring size (diameter of the shell) twelve weeks after egg mass deposition. All snails were then preserved in 95% ethanol for subsequent sod1 genotyping. Clutch sizes (the numbers of eggs/embryos in single egg masses) ranged from 2 to 34 (n = 58). Initial analysis revealed that average offspring size was correlated with clutch size, (adjusted R2 = 0.363, P<0.001) suggesting a strong density-dependent effect of number of snails per bowl on growth (same effect across all genetic backgrounds). Therefore, we restricted our analysis of effects of sod1 genotype to the offspring of clutch sizes between 13–17 eggs/embryos (there was no association between clutch size and snail size within that limited range of clutch sizes; adjusted R2 0.001, P = 0.28). We compared snail growth from 3–4 clutches in each genetic background (background 1: n = 45, background 2: n = 57, background 3: n = 58).
We also measured growth (shell diameter) in snails that were raised individually for 32 weeks as part of the egg production and hatch success experiments described below (hereafter referred to as “late growth” compared to the “early growth” measures described in the above experiment).
As in the growth study, we collected egg masses from individual pairs of the final crosses (i.e. the F2 offspring). From each population, we haphazardly chose 50 sexually immature offspring (4–5 mm shell diameter). Each snail was reared singly and a portion of a tentacle was excised to determine its sod1 genotype. We then randomly chose ten juveniles of each genotype (BB, BC, and CC) from each set of 50 genotyped snails, and reared them individually for subsequent fecundity comparisons (i.e. n = 30 per genetic background). Because B. glabrata is a facultative self-fertilizing hermaphrodite, we provided a mate to each snail prior to measuring egg production and hatch success to ensure offspring were not the result of selfing (because inbreeding depression is expected to affect egg survival). We chose to mate the genotyped individuals with snails from an isogenic inbred population to keep consistent the relative contribution of the “male-acting” snail to egg production. The isogenic inbred individuals were from a population of inbred M-line strain of B. glabrata established at the University of New Mexico through 32 generations of selfing (Si-Ming Zhang pers comm.). Because the M-line and F2 offspring look morphologically similar, we marked the M-line snails with a white dot using nail polish 24 hours prior to mating.
All snails were individually reared until reproductively active, as determined by the presence of well-formed egg masses containing developing embryos. B. glabrata preferentially use allosperm for fertilization and store sperm for up to 10 days [74]. Consequently, each snail was paired with a size-matched, painted, inbred M-line individual for one week, then separated and allowed to lay eggs for one week in a new cup. These eggs were thus presumably fertilized by allosperm, even though layed in the absence of a partner [73]–[75]. Egg numbers were counted at the end of each 1-week laying period, after which snails were re-paired with a different mate. We continued the mating/laying schedule for ten weeks, resulting in five one-week accumulated egg production measurements from each snail. We present the sum of the five one-week egg accumulation measures as the total egg production for each snail over five weeks.
We examined egg hatch success in the same set of genotyped individuals in which we surveyed egg production. Each snail was paired with a size-matched painted inbred M-line individual for 48 hours, and then isolated in a new cup. Two egg masses from each snail were carefully collected 72 hours post-transfer and reared individually (n = 180). Egg masses were surveyed for total egg count upon collection, and final hatch counts were conducted six weeks later. Hatch success (percent of eggs hatched at six weeks) from the two egg masses was averaged for each snail.
In addition to measuring egg production and egg hatch, we also monitored mortality at eight and twelve months in the same set of F2 snails used for the egg production and hatch success experiments. Mortality was measured as percent of individuals from each sod1 genotype alive at the time of census for each genetic background.
All snails were reared in an environmentally controlled room kept at 26°C and on a 12 hr day/12 hr night light cycle with full spectrum light. Snails were fed green leaf lettuce ad libitum throughout all experiments. In experiments other than those in which we measured resistance, egg masses and snails were reared, mated, and maintained in 500 mL cups with 300 mL of ASW. Complete water changes were carried out weekly. When generating the three different populations (i.e. the three different genetic backgrounds) for the fecundity experiments, the egg masses (and offspring) were reared in 2 L of ASW in aerated, lidded 1-gallon, clear plastic boxes (IRIS, USA). The egg masses monitored in the hatch success experiment were reared in petri-dishes (100×15 mm) with 5 mL of ASW. Finally, in the resistance assay we reared exposed snails in moderately dark, lidded 3-gallon plastic tubs (Dark Indigo Rubbermaid Roughneck boxes). Each contained 7.5 L of aerated dechlorinated water supplemented with 10 mL of calcium carbonate shell hardening solution (30 mg Ca++/L). Half of the water was changed with dechlorinated water between each infection assay.
We extracted genomic DNA from snail head foot tissue following the CTAB protocol [77], and used chelex extraction methods for tentacle tissue. sod1 genotype was determined using fragment analysis on an ABI 3730 capillary sequencer following amplification with AmpliTaq (Applied Biosystems, Inc.) (F-(VIC) - TCA TTG GTC GCA GCT TAG TG, R - GTC CTG TCA TGT AGC CAC CA). The B and C alleles are differentiated by a two base-pair (bp) insertion/deletion in the fourth intron that is fully resolved by the capillary system (the full sequences for the fourth intron are available for the B and C allele on NCBI GenBank from [61]). Sequence analysis of a subset of samples corroborated fragment analysis methods. Fragment analysis peaks were visualized using GENOTYPER (Applied Biosystems, Inc.), and sequence data were analyzed using SEQUENCHER (GeneCodes, Inc.).
Data were assessed for normality (Shapiro-Wilk) and equal variance. To examine the effects of genetic background on the association between carrying the B allele and resistance to parasite infection we used generalized linear models (logit function) to compare resistance (coded as a binomial response for each snail, infected = 1, not infected = 0) among genetic backgrounds and sod1 genotypes. We used regression coefficients from individual logistic regressions to quantify the relative effect sizes of substituting one allele for another in each of the genetic backgrounds. We compared fitness measures (growth rate, egg production, and hatch success) among genetic backgrounds and genotypes using two-way ANOVAs and Tukey post-hoc tests. For mortality we used generalized linear models (logit function, surviving snail at time of census = 1, dead snail = 0). No transformations were needed to normalize any of these data. We defined significance at the level of alpha = 0.05. For data analyses, we used the statistical packages SPlus version 8.1 for Windows (TIBCO Software, Inc) and SigmaPlot for Windows version 11.0 (Systat Software, Inc).
We found main effects of genotype and genetic background, and a background-by-genotype interaction (logit GLM; background: P = 0.09, genotype: P = 0.003, background×genotype: P = 0.022). As expected, the B allele was most protective. However, the strength of the association between sod1 genotype and resistance to infection depended on genetic background. The association was strongest in genetic background 1 and there was a similar but non-significant trend in background 2. In contrast, allelic variation at sod1 explained little of the variance in resistance in background 3 (Figure 3). Substituting a B allele for a C allele decreased the odds of infection by 6.2 in genetic background 1, and by 2.5 in genetic background 2 (logit GLM; P = 0.0027 and 0.0477, respectively). In genetic background 3 there was no significant additive effect. Thus, the effect of allelic variation at sod1 on resistance to infection was most important in predicting infection in the genetic background having high average resistance, and was largely irrelevant in the low-resistance genetic background.
With regard to early growth (size at 12 weeks), we found significant main effects of genetic background and sod1 genotype, but no interaction effect. Surprisingly, individuals with the CC genotype were smaller, on average, than those with BB and BC genotypes (two-way ANOVA; background: F2,151 = 11.07,P<0.001; genotype: F2,151 = 8.11,P<0.001; background×genotype: F4,151 = 0.68, P = 0.991) (Figure 4A). Thus the B allele was associated with faster growth and appeared almost completely dominant to the C allele for this trait (Figure 4A).
For late growth (size at 32 weeks), we again found significant main effects of genetic background and genotype, and no interaction (two-way ANOVA; background: F2,75 = 39.8, P<0.001; genotype: F2,75 = 3.68, P = 0.030; background×genotype: F4,75 = 1.54, P = 0.20). The CC individuals were still smaller than the BC and BB individuals, and the B allele appeared to act dominantly (Figure 4B).
In regard to egg production, we found a main effect of genetic background, but no main effect of sod1 genotype and no significant interaction (two-way ANOVA; background: F2,73 = 6.11, P = 0.0035; genotype: F2,73 = 0.533, P = 0.59; background×genotype: F4,73 = 0.472, P = 0.756). The BB genotype had the lowest estimated fecundity in genetic backgrounds 1 and 2, but the CC genotype had the lowest in background 3 (Figure 4C). However, we examined only 10 individuals per genotype within each genetic background, and thus had low power to detect all but strong main or interaction effects, as evidenced from a post-hoc power analysis. Our calculated effect size for the main effect of genetic background was 0.432, while effect sizes for the main effect of genotype and interaction were only 0.15 and 0.17, respectively. Additionally, our calculated power was 0.95 for the main effect of genetic background but only 0.22 and 0.27 for the main effect of genotype and for the interaction, respectively. Thus, an effect of sod1 genotype on fecundity would have had to be much stronger than observed to be detected with our sample sizes.
Average hatch success across all genetic backgrounds was 49%, and varied from 35% to 62% among genotypes (Figure S1). We did not find a significant main effect of genetic background or genotype on hatch success (two-way ANOVA; background: F2,60 = 0.47, P = 0.62; genotype: F2,60 = 1.52, P = 0.23; background×genotype: F4,60 = 0.99, P = 0.42). Thus, the B allele does not incur an obvious fitness cost associated with egg production (Figure 4C) or offspring hatch success. We note that although our average hatch rate of 49% is on the low side of rates reported in the literature, it is not unusually low (e.g. [78]).
At the 8-month census we found significant main effects of both genetic background and genotype on mortality (logit GLM, background: P = 0.002, genotype: P = 0.04), but no interaction (drop-in-deviance test, P = 0.19). CC individuals exhibited greater mortality, averaging 37% across genetic backgrounds, whereas BB and BC average 17% and 13% respectively (Figure 4D).
At 12 months we again found a significant main effect of genetic background, but the genotype effect was no longer significant (logit GLM, background: P = 0.02, genotype: P = 0.18), and there was no interaction (drop-in-deviance test, P = 0.39). These results suggest there is no cost to having the B allele in terms of increased mortality, and a possible advantage in early survival (Figure 4E).
In this study we considered the utility of a resistance-associated locus, cytosolic copper-zinc superoxide dismutase (sod1) in Biomphalaria glabrata, for vector-mediated control of Schistosoma mansoni. We looked for evidence of fitness costs in growth rate and reproduction. We also tested for epistatic effects of genetic background by assessing influence of the B and C alleles on resistance and on life history traits.
The association between allelic variation at sod1 and resistance to infection varied substantially among genetic backgrounds. The three genetic backgrounds differed in average resistance (78%, 64%, and 39%; Figure 2). sod1 genotype was most predictive in the genetic background having the highest average resistance, and had a negligible effect in the genetic background having the lowest average resistance (Figure 3). Thus, sod1 appears to interact epistatically with other genes that influence resistance, a result that might help us identify those other loci. That there are other resistance loci segregating in the 13–16-R1 population is evident because inbred lines having identical sod1 genotypes vary substantially in resistance (Bender and Larson, unpublished observations). Through gene expression studies, several other loci have been identified in B. glabrata as being potentially immune relevant [53]–[59], and various physiological differences have been noted between snail strains having high or low resistance to trematode parasites (reviewed in [67]). However, candidates that seem particularly likely to interact with sod1 as observed here include loci encoding proteins involved in non-self recognition and loci that control other steps in the oxidative burst pathways. Recognition loci are suggested because, as part of the effector mechanism used by the host to attack the parasite, sod1 would come into play only after the parasite has been recognized. Thus, sod1 genotype would be irrelevant in a low-recognition background, but very important in a high-recognition background. Possible recognition loci include lectin-like molecules such as FREPs [79]. Loci affecting numbers or some other property of hemocytes might also behave epistatically with sod1 in a similar manner such that if hemocytes were incompetent (or insufficient in number) to encapsulate the parasite, their ability to produce H2O2 would be irrelevant.
Costs of resistance have been demonstrated in many systems [21]–[26]. Even in B. glabrata, there is some evidence that strains with higher resistance to schistosomes differ from strains with lower resistance in components of fitness [49], [50], [80]–[85]. Furthermore, relative to the A and C alleles, the B allele of sod1 is over-expressed. The SOD1 protein produces H2O2, a highly reactive species with the potential to damage host tissue as well as the parasite [69]. Thus, it would be no surprise to see a cost of resistance associated with the B allele at sod1. Nevertheless, here we failed to detect any disadvantage due to the B allele in terms of reproduction, and observed an advantage over the C allele in terms of growth rate and survival to 8 months post-hatch (Figure 4). Furthermore, there were no significant interactions between sod1 genotype and genetic background with regard to life history traits. It is also interesting that the B allele acted dominantly to the C allele for growth rate (Figure 3), a result that might be expected if the difference really results from over-expression of the B allele.
Given our data suggest that the B allele may confer a slight advantage in terms of growth and early survival, one might wonder why our population has not become fixed for the B allele. Possible explanations include: (1) this laboratory maintained population is not in equilibrium and the selection pressure is not strong enough to have driven the allele to higher frequency yet (we have no data on allele frequencies of sod1 at the founding of this laboratory population); (2) there may be costs to having the B allele in other components of fitness that we did not measure; (3) perhaps there are complex interactions among the three major alleles in the population (A, B, and C) that prevent the B allele from increasing in frequency (e.g. see p 223–225 in [86]).
We showed the promising result of no obvious cost, and perhaps a life history trait advantage for the more-resistant allele at sod1. Obvious caveats include that our experiments were conducted in a (presumably benign) laboratory setting, and would need to be replicated under field conditions. Other studies have found that costs of resistance are more likely to manifest under specific environmental conditions, such as low food and temperature stress [12], [32], [35], [36]. Of perhaps greater concern is the strong epistatic effect on resistance between sod1 and other loci in the genome. Defeating an attempted infection is a complex process that involves many steps including recognition, signaling and implementing the effector (killing) mechanisms. SOD1 can participate in both signaling and effector mechanisms, and the products of many loci may need to interact properly to sufficiently clear an infection. Thus, it will be essential to assess the performance of sod1 in the field and in a variety of other genetic backgrounds.
There are also a number of basic questions, unrelated to those addressed here, about sod1 and resistance to S. mansoni that need to be answered before one could seriously consider using sod1 for vector manipulation in the field. We still need to prove that the association between resistance and sod1 alleles is actually causal, and if so, if the protective effect of allele B is really owing to its overexpression. It is theoretically possible that sod1 is not the actual causal locus, but is just in strong linkage disequilibrium with a closely-linked locus that actually controls resistance. This seems unlikely given the association between sod1 genotype and resistance was discovered using a functional approach (e.g. knocking down H2O2 production in B. glabrata hemocytes increases their susceptibility to infection [66]), but the functional basis of the association still needs to be proven. Additional work to test the causality of the association is underway. In the unlikely event it turns out that another locus is actually causal, then the results of this study are still quite relevant, but for the new locus of interest.
We also do not know yet if the effect of sod1 we observed is generalizable to other populations/strains of S. mansoni. We have only studied the PR-1 strain of S. mansoni in interaction with the 13–16-R1 population of B. glabrata. It is possible that the protective effect of sod1 alleles depends on the strain of parasite in addition to the strain of snail. In a similar vein, we also have no data on if, or how, sod1 genotype affects resistance to other pathogens. A field population of snails interacts with many pathogens in addition to S. mansoni, and there could be fitness tradeoffs associated with other pathogens that render the use of sod1 for vector manipulation ineffective in some environments.
In summary, we have here shown that, in a laboratory setting, there was no obvious cost to having the most protective allele at sod1, and perhaps a slight advantage. The generality of this result will need to be verified in other environments, and for other components of fitness. We also demonstrated an effect of genetic background on the association between sod1 genotype and resistance, a result that points to strong epistatic interactions with other loci in the genome. Clearly sod1 is not the only locus in the genome that influences resistance. So perhaps vector manipulation will require changes at several interacting loci to insure success. Further work of this sort on sod1 and other resistance-associated loci will be essential for evaluating the prospects for vector manipulation as a way to control transmission of S. mansoni.
|
10.1371/journal.pcbi.1005033 | Bipartite Community Structure of eQTLs | Genome Wide Association Studies (GWAS) and expression quantitative trait locus (eQTL) analyses have identified genetic associations with a wide range of human phenotypes. However, many of these variants have weak effects and understanding their combined effect remains a challenge. One hypothesis is that multiple SNPs interact in complex networks to influence functional processes that ultimately lead to complex phenotypes, including disease states. Here we present CONDOR, a method that represents both cis- and trans-acting SNPs and the genes with which they are associated as a bipartite graph and then uses the modular structure of that graph to place SNPs into a functional context. In applying CONDOR to eQTLs in chronic obstructive pulmonary disease (COPD), we found the global network “hub” SNPs were devoid of disease associations through GWAS. However, the network was organized into 52 communities of SNPs and genes, many of which were enriched for genes in specific functional classes. We identified local hubs within each community (“core SNPs”) and these were enriched for GWAS SNPs for COPD and many other diseases. These results speak to our intuition: rather than single SNPs influencing single genes, we see groups of SNPs associated with the expression of families of functionally related genes and that disease SNPs are associated with the perturbation of those functions. These methods are not limited in their application to COPD and can be used in the analysis of a wide variety of disease processes and other phenotypic traits.
| Large-scale studies have identified thousands of genetic variants associated with different phenotypes without explaining their function. Expression quantitative trait locus analysis associates the compendium of genetic variants with expression levels of individual genes, providing the opportunity to link those variants to functions. But the complexity of those associations has caused most analyses to focus solely on genetic variants immediately adjacent to the genes they may influence. We describe a method that embraces the complexity, representing all variant-gene associations as a bipartite graph. The graph contains highly modular, functional communities in which disease-associated variants emerge as those likely to perturb the structure of the network and the function of the genes in these communities.
| Genome Wide Association Studies (GWAS) have created new opportunities to understand the genetic factors that influence complex traits. Excepting highly-penetrant Mendelian disorders, the majority of genetic associations seem to be driven by many factors, each of which has a relatively small effect. In a recent study [1], 697 SNPs were associated with height in humans at genome-wide significance, yet these SNPs were able to explain only ∼20% of height variability; ∼9,500 SNPs were needed to raise that to ∼29%. In addition, ∼95% of GWAS variants map to non-coding regions [2], complicating biological interpretation of their functional impact.
To bridge the functional gap between genetic variant and complex trait, expression Quantitative Trait Locus (eQTL) analysis associates SNP genotype with gene expression levels. The first empirical, genome-wide linkage study with gene expression in yeast was published in 2002, linking expression levels of 570 genes to genetic loci [3]. In humans, loci have been associated with the expression of thousands of genes [2, 4], and eQTLs are enriched for phenotype associations and vice versa [5–7].
Most eQTL analyses have focused on cis-SNPs—those near the Transcriptional Start Site (TSS) of the gene in the association test. Recent computational developments [8] and work demonstrating the impact and replicability of trans-eQTLs [9, 10] have increased interest in identifying and understanding the role played by trans-acting SNPs.
However, new methods are needed to elucidate the potential functional impact of the thousands of GWAS SNPs and tens to hundreds of thousands of eQTL SNPs that can be detected in a single study. Here we present CONDOR, COmplex Network Description Of Regulators, (Fig 1) a method that incorporates both cis- and trans- associations to identify groups of SNPs that are linked to groups of genes and systematically interrogate their biological functions. The method has been implemented as an R package and is publicly available at https://github.com/jplatig/condor. We then validate this approach using genotyping and gene expression data from 163 lung tissue samples in a study of Chronic Obstructive Pulmonary Disease (COPD) by the Lung Genomics Research Consortium (LGRC).
We used the MatrixEQTL package in R to calculate cis- and trans-eQTLs, considering only autosomal SNPs, using age, sex, and pack-years as covariates (see Methods). The cis- and trans- associations were run separately, with an FDR threshold of 10%. This analysis identified 40,183 cis-eQTLs and 32,813 trans-eQTLs. Quantile-quantile plots for both cis- and trans- are shown in Fig 2. In total, 72,996 statistically significant associations were detected between 57,062 SNPs and 7,051 genes.
We represented these associations as a bipartite network consisting of two classes of nodes—SNPs and genes—with edges from SNPs to the genes with which they are significantly associated based on the eQTL FDR cut-off. The network had a Giant Connected Component (GCC) with 41,813 links, 28,593 SNPs, and 3,091 genes. As a network diagnostic, we estimated whether or not we could reject the hypothesis that the SNP and gene degree distributions were power-law distributed. To test this, we fit each degree distribution to a power law, and determined the goodness of fit using the method described in [11] (see Methods). If the edges from all connected components are considered, the p-value for the SNP degree is very low, Ppl ≈ 0, suggesting that we can rule out a power law distribution. However, if very small connected components (fewer than 5 SNPs and 5 genes) are excluded, the SNP degree may follow a power-law (Ppl < 0.8) as shown in Fig 3a. The gene degree distribution (Fig 3b) may be power-law distributed when considering all connected components or only those with more that 5 SNPs and 5 genes (Ppl < 0.4 in both cases) and there are multiple network hubs, shown in the tail of the distribution in Fig 3b. For our further analysis we considered all connected components with more than 5 SNPs and 5 genes.
It is often cited in complex networks literature that the hubs, those nodes in the network that are most highly connected, represent critical elements whose removal can disrupt the entire network [12, 13]. As a result, one widely-held belief about biological networks is that disease-related elements should be over-represented among the network hubs [14]. To test the hypothesis that disease-associated SNPs are concentrated in the hubs, we projected GWAS-identified SNPs associated with a wide range of diseases and phenotypes onto the SNP degree distribution (Fig 4). We used the gwascat package [15] in R to download GWAS SNPs annotated in the NHGRI GWAS catalog; 274 of those SNPs mapped to the eQTL network (S1 Table). To our surprise, the network hubs—the right tail of Fig 4—were devoid of disease-associated SNPs which were instead scattered through the upper left half of the degree distribution. The difference in degree distributions did not appear to be driven by linkage disequilibrium or distance to nearest gene (see Methods and S1, S2, S3 and S4 Figs). While the SNPs associated with a single gene are easier to interpret, the concentration of disease-associated SNPs in the middle of the distribution prompted us to look at other features of the network and its structure.
Given the low phenotypic variance explained by any single GWAS SNP, we expected groups of SNPs to cluster with groups of functionally-related genes in our eQTL network. Unlike previous work [16–18] which imposes “known” pathway annotations and other data to posit the function of GWAS SNPs or identifies modules with only a handful of SNPs [19], we used the structure of the eQTL network to identify densely connected groups of SNPs and genes and then tested those groups for biological enrichment.
Our goal is the identification of those densely connected communities in the bipartite network. Methods for finding bicliques (subgraphs with all-to-all connections within the larger bipartite network) have been described for bipartite networks with a small number (∼102) of nodes in each connected component [20]. However, these methods do not scale to networks with connected components containing thousands of nodes [20, 21]. Further, we do not expect biologically meaningful eQTL clusters to contain only all-to-all connections.
To cluster our eQTL network, we adapted a well-established strategy [22], community structure detection, which has been shown to scale well to large networks [23]. Many real-world networks have a complex structure consisting of “communities” of nodes [24]. These communities are often defined as a group of network nodes that are more likely to be connected to other nodes within their community than they are to those outside of the community. A widely used measure of community structure is the modularity, which can be interpreted as an enrichment for links within communities minus an expected enrichment given the network degree distribution [22].
To partition the nodes from the eQTL network into communities—which contain both SNPs and genes—we maximized the bipartite modularity [25]. As recursive cluster identification and optimization can be computationally slow, we calculated an initial community structure assignment on the weighted, gene-space projection, using a fast uni-partite modularity maximization algorithm [23] available in the R igraph package [26], then iteratively converged (ΔQ < 10−4) on a community structure corresponding to a maximum bipartite modularity.
The bipartite modularity is defined in Eq (1), where m is the number of links in the network, A ˜ i j is the upper right block of the network adjacency matrix (a binary matrix where a 1 represents a connection between a SNP and a gene and 0 otherwise), ki is the degree of SNP i, dj is the degree of gene j, and Ci, Cj the community indices of SNP i and gene j, respectively (see [25] for further details).
Q = 1 m ∑ i , j A ˜ i j - k i d j m δ ( C i , C j ) (1)
This analysis identified 52 communities across 10 connected components in the LGRC data, with 34 of those communities mapping to the GCC (Qgcc = 0.79; Fig 5). The density of these communities can be seen in Fig 5. In Fig 5b, there is visible enrichment for links within each community (colored links) compared to links between different communities (black links). These communities represent groups of SNPs and genes that are highly connected to each other and span multiple chromosomes (see Fig 6), suggesting that groups of genes may be jointly moderated by groups of SNPs that together represent specific biological processes.
To investigate this hypothesis, we tested each community for GO term enrichment using Fisher’s Exact Test (available in the R package GOstats [27]) and found 11 of the 52 communities contained genes enriched for specific Gene Ontology terms (see S2 Table) (P < 5e − 4; overlap >4), encompassing a broad collection of cellular functions that are not generally associated with COPD. Indeed, this is what one might expect as the genetic background of an individual should have an effect not only on disease-specific processes, but more globally on the physiology of his or her individual cells. A number of communities do, however, show enrichment for biological processes that are known to be involved in COPD, including genes previously associated with the disease.
For example, Community 29 (see Fig 5 and S2 Table) was enriched for chromatin and nucleosome assembly/organization and includes members of the HIST1H gene superfamily. Community 33 (see Fig 5 and S2 Table) included GO term enrichment for functions related to the HLA gene family, including T cell function and immune response; autoimmunity has been suggested as a potential contributor to COPD pathogenesis [28]. This community also contains PSORS1C1, which has been previously implicated in COPD [29].
Another of the genes in Community 33, AGER, has been implicated in COPD [30] and encodes sRAGE, a biomarker for emphysema. Its expression is negatively associated via eQTL analysis (β = −0.3) with rs6924102. This SNP has been observed to be an eQTL in a large blood eQTL dataset for a number of neighboring genes [9], but it has not previously been described as an eQTL for AGER. This SNP lies in a region containing a DNase peak in cell lines analyzed by ENCODE [31] (indicating it sits in a region of open chromatin) and there is evidence of POLR2A binding from ChIP-Seq data in the GM12878 cell line as reported by ENCODE (http://regulomedb.org/snp/chr6/32811382). This suggests that rs6924102 may inhibit the expression of AGER through disruption of RNA Polymerase II binding and subsequent mRNA synthesis. This SNP is located ∼700KB from the well-studied non-synonymous AGER SNP, rs2070600.
Examining Fig 5a, it is evident that within each community there are local hubs that are highly connected to the genes within that community. While a wide array of network node metrics exist (for example, [32, 33] and references in [33]), most of these metrics are global measures that do not consider a node’s role in its local cluster/community and so may miss SNPs that are central to their communities and therefore likely to alter gene expression of functionally associated genes. Such within-community hubs have been observed in protein-protein interaction networks [34] and metabolic networks [35].
We defined a core score that estimates importance of a SNP in the structure of its community. For SNP i in community h, its core score, Qih, Eq (2), is the fraction of the modularity of community h, Qh, Eq (3), contributed by SNP i. This allows for comparison of SNPs from different communities, as each community does not have the same modularity, Qh.
Q i h = 1 m ∑ j ( A ˜ i j − k i d j m ) δ ( C i , h ) δ ( C j , h ) Q h (2) Q h = 1 m ∑ i , j ( A ˜ i j − k i d j m ) δ ( C i , h ) δ ( C j , h ) (3)
If one views disease as the disruption of a process leading to cellular or organismal dysfunction, one natural hypothesis is that SNPs with the greatest potential to disrupt cellular processes might be enriched for disease association. To test this we used both the Wilcoxon rank-sum and Kolmogorov-Smirnov (KS) tests to assay whether the 274 NHGRI GWAS-annotated SNPs in the network were more likely to have high Qih scores. For both tests, the distribution of Qih scores for GWAS-associated SNPs were compared to the distribution of non-GWAS SNP scores.
To obtain an empirical p-value for these tests, we permuted the GWAS/non-GWAS labels and recalculated the KS and Wilcoxon tests 105 times. Histograms of the test statistics are shown in Figs 7 and 8. The red dot in the histogram represents the test score with the true labeling. Both tests had highly significant permutation p-values, with P < 10−5 for the KS and Wilcoxon tests, indicating that GWAS SNPs were over-represented among SNPs with high core scores. Furthermore, the median core score for the GWAS SNPs was 1.74 times higher than the median core score for the non-GWAS SNPs. To test this result for dependence on Linkage Disequilibrium (LD) and gene distance, we reran the KS and Wilcoxon permutation tests with a subset of SNPs matching the LD structure and distance to nearest gene of the 274 GWAS SNPs (see Methods for details). Neither the LD structure (P < 0.001 for KS and Wilcoxon tests, S5 and S6 Figs) nor distance from the nearest gene (P < 0.001 for KS and Wilcoxon tests, S7 and S8 Figs) of the GWAS SNPs was signficantly associated with the core score. Thus, while global hubs are devoid of GWAS associations with disease, local hubs within communities are significantly enriched for disease associations.
As a way of further assessing the link between GWAS significance and functional perturbation in COPD, we calculated a GWAS-FDR for all SNPs clustered in our network that had a reported p-value from a recent GWAS and meta-analysis of COPD [36] (see Methods). There were 30 SNPs with an FDR < 0.05, and 28 of the 30 had evidence of functional impact according to RegulomeDB [37], with 15 SNPs identified as likely to affect transcription factor binding and linked to expression (See Fig 9 and S3 Table). These 30 SNPs mapped to 3 different communities (see S3 Table) including Community 33, which contains other COPD-associated SNPs and genes, and is enriched for GO terms describing T cell function and immune response. One of the SNPs in this community likely to affect binding (S3 Table) is rs9268528, which is linked by our network to HLA-DRA, HLA-DRB4, and HLA-DRB5; the cis-eQTL associations between rs9268528 and both HLA-DRA and HLA-DRB5 have been previously observed in lymphoblastoid cells [38]. All three HLA genes lie in Community 33 and contribute to the community’s enrichment for T cell receptor signaling pathway (GO:0050852) [39].
To determine the network influence of these 30 SNPs, we compared their core score, Qih, to the core scores of SNPs with a GWAS-FDR ≥ 0.05 (See Fig 10). The median Qih value for the 30 GWAS-FDR significant SNPs was 20.3 times higher than the median for SNPs with an FDR ≥ 0.05. Using the KS and Wilcoxon tests described in the Methods, these core scores were not significantly associated with LD structure (P < 0.001, S9 and S10 Figs) or distance to nearest GSS (P < 0.001, S11 and S12 Figs).
Genome-wide association studies have searched for genomic variants that influence complex traits, including the development and progression of disease. However, the number of highly-penetrant Mendelian variants that have been found is surprisingly small, with most disease-associated SNPs having a weak phenotypic effect. GWAS studies have also identified many SNPs that do not alter protein coding and have found significant loci that are shared in common across multiple diseases. This body of evidence suggests that in most instances it is not a single genetic variant that leads to disease, but many variants of smaller effect that together can disrupt cellular processes that lead to disease phenotypes. The challenge has been to find these variants of small effect and to place them into a coherent biological context.
We chose to address this problem by analyzing the link between genetic variants and the most immediate phenotypic measure, gene expression. In doing so, we chose not to focus solely on cis-acting SNPs, but also to consider trans-acting variants. Our motivation was, in part, to try to understand SNPs found through GWAS studies to be associated with phenotypes, but that could not be immediately placed into a functional context. After performing a genome-wide cis- and trans-eQTL analysis, we identified a large number of many-to-many associations: single SNPs associated with many genes as well as single genes that were significantly associated with many SNPs. To represent those associations, we constructed a bipartite network, one that contains two types of nodes—SNPs and genes—with edges connecting SNPs to the genes with which they were significantly associated. Our analysis of that network led to a number of observations that independently speak to our intuition about disease and the genetic factors that control it.
First is the observation that the highly connected SNPs, the global hubs in the network, are devoid of variants that have been identified as being disease-associated in the hundreds of studies collected in the NHGRI GWAS catalog. While initially surprising, further consideration suggests that this may be the result of negative selection. Since a true hub SNP influences genes across the genome that are involved in many biological processes, highly disruptive variants that are hubs are likely to significantly affect cellular function. In fact, this is the expected impact of a hub—its disruption should lead to the catastrophic collapse of the network. And so, disruptive SNPs that would be network hubs are likely to be lethal or highly debilitating and therefore strongly selected against and quickly swept from the genome.
Second, we found that SNPs and their target genes form highly connected communities that are enriched for specific biological functions. This too speaks to our inituition and to the evidence about polygenic traits that has accumulated over time. They are not the result of a single SNP that regulates a single gene, but a family of SNPs that together help mediate a group of functionally-related genes.
Third, the enrichment for GWAS disease associations among the high core score SNPs has a very simple and intuitive interpretation. The SNPs that are most significantly connected within a particular functionally-related group are those most likely to disrupt that process and therefore be discovered in GWAS analysis. After all, diseases do not develop because the cell’s entire functionality collapses, but because specific processes within the cell are disrupted.
What our analysis provides is a new way of exploring cis- and trans-eQTL analysis and GWAS. What one must do is to consider not only the local effects of genetic variants, but also the complex network of genetic interactions that help regulate phenotypes, including gene expression.
This method also suggests a new way of filtering genes for inclusion in GWAS analysis. Since many disease-associated SNPs appear to be either cis-acting or those which are central to functionally-defined communities, one could focus on those SNPs most likely perturb specific biological processes rather than considering the entirety of SNPs in the genome.
One might note that this analysis was carried out using data on genetic variation and gene expression from the LGRC representing COPD and control lung tissue and question both the generalizability of the results and the use of GWAS-associated disease SNPs from many diseases in the analysis. While these are potentially legitimate concerns, many of the community-based processes we find are not specific to COPD or to the lung but instead are active in nearly all human cell types.
Although one might expect some processes to change in different disease states, the impact of common variants and the structure of the network is likely to be highly similar. Consequently, although there may be some SNPs whose impact is disease- and tissue-specific, many are likely to be independent of disease state. This suggests that it may be useful to develop eQTL networks across disease states and tissue types and to explore changes in the overall network and community structure across and between phenotypes due to rare variants and tissue-specific expression.
Validating individual associations in the eQTL network is a difficult challenge. Most eQTL studies limit their validation efforts to downstream effects of high-confidence cis-acting eQTLs. The bipartite network presented here captures not only these strong cis-eQTLs but also the weak effects of many more cis- and trans-acting SNPs. So the likelihood that any individual association can be easily validated may not be that great, as it is likely to be of small phenotypic effect and important in only a subset of individuals. However, this is not the point. What is important for the phenotype is not any single SNP-gene association, but the “mesoscale” organization of genes and SNPs represented by the communities in the network. We believe this intermediate structure better reflects the aggregation of weak genetic effects that contribute to late-onset complex diseases. What we hope to have demonstrated in this manuscript is that the higher order structure, which was not an input to the network model, provides insight into a number of aspects of the genetics and manifestation of polygenic traits.
We began by downloading gene expression data from the LGRC web portal (https://www.lung-genomics.org/download/) representing data from COPD-case and control samples generated by the Lung Genomics Research Consortium (LGRC). This included GCRMA-normalized gene expression data obtained using Agilent-014850 Whole Human Genome 4x44K and Agilent-028004 SurePrint G3 Human GE 8x60K Microarrays. We then obtained matching genotyping data (dbGAP accession phs000624.v1.p1) collected using the Illumina Infinium HD Assays with Human Omni 1 Quad and Human Omni 2.5 Quad arrays. All subjects were reported to be of Caucasian descent and were selected based on a variety of parameters including clinical measures associated with diagnosis. Samples that did not meet standards for lack of relatedness as measured using Identity by Descent (IBD) and inbreeding coefficient, F, were excluded. Those samples with discordance between reported and genetic sex were not included. Samples missing more than 10% percent of genotyped SNPs were also removed. SNPs with minor allele frequency (MAF) < 0.05 or Hardy Weinberg Equilibrium P-value < 0.001 were removed. After all quality controls, 163 samples remained. All SNPs were mapped to human genome 19, and the Ensembl IDs provided by the LGRC web portal were mapped to the GRCh37 build of human genome 19 using the biomaRt library [40] in R. The cis-window was defined as +/- 1 MB of the Ensembl-defined GSS. The COPD GWAS data from a meta-analysis of COPDGene non-Hispanic whites and African-Americans, ECLIPSE, GenKOLS, and NETT/NAS studies was obtained from the authors of [36]. The bipartite clustering via modularity maximization took 95 seconds on a 64-bit Linux server with 189 GB of RAM running R 3.1.3.
For each empirical degree distribution, we fit the two parameters for a power-law: the minimum degree at which the power-law behavior starts, dmin, and the exponent, α. A Kolmogorov-Smirnov test was then used to estimate the goodness of fit between 5,000 randomly generated power-law distributed synthetic data sets given dmin and α and their corresponding power-law fit. The p-value, Ppl, used to reject the power-law hypothesis is then the fraction of times a synthetic data set has a KS statistic larger than that of the true test. For both the SNP and gene degree distributions, Ppl was calculated using the 5,000 goodness of fit values (code for the parameter estimation, goodness of fit and probability estimation was obtained from the website associated with [11]).
To test the effect of LD and distance from Gene Start Site (GSS) on the degree distribution and core score (Qih) distribution of a set of GWAS SNPs, we created equivalently sized sets of SNPs that matched on a given characteristic of interest (LD or GSS) and compared that subset to all other SNPs. We repeated this process for each GWAS SNP set 1000 times. For the LD testing, we calculated LD blocks using the PLINK [41] “blocks” flag, estimating blocks using all SNPs that passed quality control. To achieve adequate sample sizes in the resampling, we binned LD blocks in 5kb windows, grouped all blocks >100kb into one bin and grouped all SNPs not in a block into one bin. For each resampling, the random set matched the GWAS set for both the LD bin and the number of SNPs in LD together within a block.
As a proxy for the gene density of a region, we used each SNP’s distance from the nearest GSS. Distances were grouped into 1kb bins, with all distances >400kb grouped into one bin. The resampled sets were then matched on the GWAS SNP sets such that the number of SNPs in each bin was the same.
|
10.1371/journal.pcbi.1003664 | CCAST: A Model-Based Gating Strategy to Isolate Homogeneous Subpopulations in a Heterogeneous Population of Single Cells | A model-based gating strategy is developed for sorting cells and analyzing populations of single cells. The strategy, named CCAST, for Clustering, Classification and Sorting Tree, identifies a gating strategy for isolating homogeneous subpopulations from a heterogeneous population of single cells using a data-derived decision tree representation that can be applied to cell sorting. Because CCAST does not rely on expert knowledge, it removes human bias and variability when determining the gating strategy. It combines any clustering algorithm with silhouette measures to identify underlying homogeneous subpopulations, then applies recursive partitioning techniques to generate a decision tree that defines the gating strategy. CCAST produces an optimal strategy for cell sorting by automating the selection of gating markers, the corresponding gating thresholds and gating sequence; all of these parameters are typically manually defined. Even though CCAST is optimized for cell sorting, it can be applied for the identification and analysis of homogeneous subpopulations among heterogeneous single cell data. We apply CCAST on single cell data from both breast cancer cell lines and normal human bone marrow. On the SUM159 breast cancer cell line data, CCAST indicates at least five distinct cell states based on two surface markers (CD24 and EPCAM) and provides a gating sorting strategy that produces more homogeneous subpopulations than previously reported. When applied to normal bone marrow data, CCAST reveals an efficient strategy for gating T-cells without prior knowledge of the major T-cell subtypes and the markers that best define them. On the normal bone marrow data, CCAST also reveals two major mature B-cell subtypes, namely CD123+ and CD123- cells, which were not revealed by manual gating but show distinct intracellular signaling responses. More generally, the CCAST framework could be used on other biological and non-biological high dimensional data types that are mixtures of unknown homogeneous subpopulations.
| Sorting out homogenous subpopulations in a heterogeneous population of single cells enables downstream characterization of specific cell types, such as cell-type specific genomic profiling. This study proposes a data-driven gating strategy, CCAST, for sorting out homogeneous subpopulations from a heterogeneous population of single cells without relying on expert knowledge thereby removing human bias and variability. In a fully automated manner, CCAST identifies the relevant gating markers, gating hierarchy and partitions that isolate homogeneous cell subpopulations. CCAST is optimized for cell sorting but can be applied to the identification and analysis of homogeneous subpopulations. CCAST is shown to identify more homogeneous breast cancer subpopulations in SUM159 compared to prior sorting strategies. When applied to normal bone marrow single cell data, CCAST proposes an efficient strategy for gating out T-cells without relying on expert knowledge; on B-cells, it reveals a new characterization of mature B-cell subtypes not revealed by manual gating.
| Understanding cancer heterogeneity is increasingly being regarded as critical in understanding cancer progression and overcoming therapeutic resistance [1]–[4]. Different types of heterogeneity are commonly observed among the cells composing a single tumor, including genetic [5], [6], epigenetic [7], and phenotypic heterogeneity [3], [4]. Although technological challenges have limited our ability to fully characterize intra-tumor heterogeneity, in recent years characterizing heterogeneous populations of cells at the single-cell level using multidimensional fluorescence and mass flow cytometric data, combined with novel computational tools, has greatly improved our understanding of the extent of cellular heterogeneity [8], [9]. Moreover, by sorting out homogeneous subpopulations, researchers can measure and compare genomic and other functional properties of different subpopulations. However, in spite the high-throughput nature of these single cell measurements, current methods for sorting specific cell subpopulations rely on a low dimensional, often user-defined, process known as gating. Gating on a fluorescence-activated cell sorting (FACS) machine commonly refers to a manual process, performed by sequentially selecting regions from bivariate graphs that depict the expression of two markers at a time across all the cells. The gating strategy often relies on an expert's assessment of the choice of gating markers, the order of gating and cut points to identify each gated region; this assessment is often based on a subjective analysis using packages such as flowJo and FlowCore [10]. It is well documented that minor differences in gating strategy can lead to significantly different quantitative conclusions [11], [12]. We present a gating strategy that is optimized for cell sorting. Because our gating strategy is data derived, we argue that is optimal compared to manually-derived gating strategy which can be biased and highly variable. In our work, we make a distinction between gating algorithms that are optimized for sorting single cells versus analyzing a heterogeneous population of single cell data. Even though our gating strategy is optimized for cell sorting, it also has value when used in analysis of population data at the single cell level.
When analyzing a population of single cells, several gating algorithms have been developed to reduce the technical, biological and human sources of variation involved in identifying and analyzing clusters of similar cell subpopulations [8], [9], [13]–[17]. Bashashati and Brinkman provide a comprehensive overview of analysis tools for flow cytometry (FCM) data [18]. More recently, the FlowCAP-II project [12] compared the accuracy and reproducibility across several gating algorithms in terms of identifying cell clusters. All gating algorithms, including ours, have some form of a clustering algorithm, which is used to identify homogeneous subpopulations, as a major component. Many unsupervised clustering algorithms take into account the uncertainty in cluster assignments by modeling the data as mixtures of parametric distributions [18]. Although parametric mixture models have been developed to analyze FCM data [14], computational, as well as estimation errors, in clustering could still arise from outliers and skewness in the data which may not reflect the underlying assumptions of the parametric model. As an alternative, we propose a modified version of the non-parametric multivariate mixture modeling approach by Benaglia et al. [19] for clustering FCM data, where our modification includes the use of silhouette measures. This clustering algorithm handles uncertainty regarding to which cluster an event should be assigned as well as the uncertainty in the number of underlying cell states in the heterogeneous parent population and makes little or no prior assumptions on the underlying model structure. In addition, we implement an alternative clustering algorithm, namely hierarchical clustering [20], to show that the results from our gating strategy are independent of the particular clustering method used. The goal of our study is not to provide an optimal clustering strategy, but instead to provide an optimal gating strategy for sorting homogeneous cell subpopulations given any reasonable clustering algorithm.
A commonly neglected area in studying populations of single cells is identifying an optimal gating strategy for cell sorting. Sorting cells for downstream analysis relies not only on the identifying the clusters but also on the gating strategy, which is defined by the gating markers, thresholds and sequence. For manual gating at the FACS machine, typical gating strategies are organized like a family tree. For example, from mature bone marrow cells, lymphocytes are gated from the parent cells and from that gate, T-cells or B-cells are gated, and from those gates, specific T-cell and B-cell types are gated [9]. In particular, sorting out T-cells is equivalent to isolating a CD4+/CD8+ population; the user would first isolate the lymphocytes, then derive the CD3+ cells and from there, would draw a gate around the CD4 positive and CD8 positive subpopulations. This approach assumes prior knowledge of the underlying set of markers that define cell types, the gating hierarchy and relative boundaries for isolating pure cell subpopulations of interest. Selecting these parameters based solely on literature and human perspective introduces bias and variability and could result in contamination among the cell subpopulations. We make this process data-driven and fully automated by applying a recursive partitioning technique that generates a decision tree representing a reproducible gating strategy for all subpopulations of interest.
Recognizing the current reliance on human perspective and intuition in manual gating, Ray and Pyne [17] recently developed a gating framework which emulates the human perspective in FCM data analysis based on a mathematical map of the high dimensional data landscape. They propose flexible, sample-specific templates for extracting features of interest, which may have unusual shapes and distributions. An alternative approach by Lee et al [21] uses transfer learning technique combined with the low-density separation principle; this approach transfers expert knowledge on training FCM data sets to a new data. A more recent study by Aghaeepour et al. [22] developed a supervised learning computational framework that automatically reveals cell subsets that correlate strongly with clinical outcome and identifies their relevant set of markers for gating. In a follow-up study, Aghaeepour et al. [22] developed a computational tool, RchOptimyx [23], that uses dynamic programming and optimization techniques from graph theory to construct a cellular hierarchy, providing a gating strategy to identify target populations to a desired level of purity. One might argue that our work is most similar to RchOptimyx. However, as will be shown later, RchyOptimyx provides multiple approaches for gating a specific subpopulation, whereas our approach aims to find a single, optimal gating strategy in a fully automated manner without relying on qualitative judgement.
We present an algorithm, named CCAST for Clustering, Classification and Sorting Tree, to identify and sort homogeneous subpopulations from a heterogeneous parent population using a decision tree representation for a gating strategy that can be used to sort for homogeneous cell subpopulations. The gating strategy derived from CCAST is data-driven and fully automated and it does not rely on expert knowledge. While CCAST is optimized for cell sorting, CCAST also has value when applied to data analysis by filtering and retraining the decision tree to produce more homogeneous subpopulations. In addition, when used for data analysis, CCAST may identify new subpopulations among the initial clusters. We apply CCAST on populations of single cell measurements made on breast cancer and normal human bone marrow. On the breast cancer SUM159 cell line, CCAST reveals at least 5 distinct cell states based on two surface markers (CD24 and EPCAM). When applied to normal bone marrow data, CCAST reveals an efficient strategy for gating T-cells. In addition, CCAST reveals two new mature B-cell subtypes, which were not found by manual gating but show distinct intracellular signaling behaviors.
We demonstrate the performance of CCAST on simulated and actual populations of single cell data. The details of the CCAST algorithm are provided under Materials and Methods. In Figure 1, CCAST is summarized in a flowchart alongside its application to simulated data. Briefly stated, starting with a population of single cell data (Figure 1A), CCAST performs a cell clustering algorithm to identify groups of similar cells (Figure 1B). The clustering can be performed in a variety of ways. We implement a nonparametric mixture model denoted as “npEM” (see Materials and Methods), but show that other clustering algorithms, such as hierarchical clustering (HCLUST), produces similar gating strategies within the CCAST framework. Once the cell clusters (aka “cell types”) are established, CCAST derives a gating strategy that is represented by a decision tree (Figure 1C), where the nodes specify the gating markers and their thresholds (aka “split points”) as edges. The terminal leaves of the decision tree represent the final gated subpopulations. Often the final number of gated populations is greater than the number of cell clusters. When this happens some of the subpopulations capture cells from only one cluster, but others capture cells from multiple clusters. For the subpopulations which contain cells from multiple clusters, all but the cells from the dominant cluster are removed and CCAST is retrained on the remaining population, producing a more robust gating strategy because it is less influenced by “contaminating” cells (Figure 1D). The final decision tree can be used for cell sorting (Figure 1E) or data analysis (Figure 1F). Although not shown in Figure 1, it is possible that a single cluster may be distributed across multiple subpopulations, where each subpopulation only contains cells from that cluster; in those cases, the cluster may have more subpopulations than derived by the clustering algorithm. This feature and all other mentioned features of CCAST are demonstrated below.
We implemented CCAST as an R package and it has been made available as a zip file in the Supplement.
To illustrate the basic properties of CCAST, we applied it to a simulated dataset of 850 single cells comprised of a mixture of 5 cell types, as illustrated in Figure 1A. On each single cell, 3 markers are measured; the distributions of marker values for each cell type are summarized in Supplement Table S1. We sampled 100, 300, 150, 100, 200 cell vector expression values for each cell type respectively. Figure 1B shows the 3D scatter plot of the cell measurements with the 5 cell types color coded; from this figure, it is not automatically apparent how to optimally sort out these 5 clusters. Figure 1C shows the first CCAST-derived decision tree based on the entire dataset; this tree partitioned the data into 5 clusters as evidenced from the leaf nodes (5,6,9,10 and 11) of the tree. Nodes 9, 10 and 11 represents pure subpopulations of clusters 4, 3 and 5, respectively; node 8 shows a mixture of clusters 1 and 4; nodes 5 and 6 are dominated by cells from clusters 2 and 1 respectively. After CCAST removed the contaminating cells from the subpopulations that have more than one cluster and re-ran the decision tree algorithm, it generated the final decision tree in Figure 1D. Note also that these subpopulations were gated using only two markers even though 3 markers were measured. Figure 1E shows the application of the final decision tree (Figure 1D) on the entire dataset. When this gating strategy was applied to the filtered dataset for downstream analysis, the resulting subpopulations are shown in Figure 1F, represented with bar plots of the markers' expression and labeled by their corresponding cell cluster. Figure 1G shows the application of the gating strategy using the estimated cut-offs on the entire data using hierarchical clustering instead of “npEM” clustering. The similar partitions on the 2D data imply that using different a clustering algorithm results in similar homogeneous subpopulations.
We next demonstrate the applicability of CCAST on actual hematopoietic dataset obtained in the study by Bendall et al. [9]. This study analyzed normal bone marrow at the level of single cells using mass cytometry (MCM), which is a recently developed high throughput technology for labeling single cells with metal-chelated antibodies that reduce auto fluorescence effect. An appeal of this particular study is that hematopoietic cells have a well-established set of lineage markers defining their differentiation stages. In this study, unstimulated and stimulated human peripheral blood mononuclear cells (PBMCs) from a healthy donor were analyzed using thirteen surface parameters, namely: CD45, CD45RA, CD19, CD11b, CD4, CD8, CD34, CD20, CD33, CD123, CD38, CD90, and CD3. In addition, 18 intracellular signaling molecules were measured. The manual gating process and the characterization of the major cell populations are shown in Supplement Figure S5 of [9]. One part of this study focused on a T-cell subset that included naive CD4+ and CD8+ T-cells and mature CD4+ and CD8+ T-cells. The analysis of the induced intracellular signaling responses in these subpopulations, as compared with those of an unstimulated control, relied on a manually-defined gating process.
To demonstrate the robustness of CCAST, we consider a subset of the data from the study by Bendall et al. [9] in order to assess both the error and reproducibility of our results in a transparent manner. We focus on a 20,000 cell T-cell subpopulation which had been manually gated into 4 subtypes (see Figure 1 in Bendall et al. [9] for the manual gating scheme). Here we pool this manually gated T-cell data, then blind the data by removing all prior knowledge of cell types or marker labels. We then randomly separate this data into a training and test set of 10,000 cells each. Pairwise scatter plots across all 13 markers, unlabeled, are shown in Figure 2. We apply CCAST on the training data to obtain the final decision tree shown in Figure 3A. These results indicate that the 4 distinct homogeneous cell states can easily be isolated using only 2 of the 13 measured markers, namely Marker 5 and Marker 2. We next carried out a sensitivity analysis on the decision tree parameters, namely the optimal tree height, denoted as L, and the split points (see Materials and Methods). First we ask the question: what happens to the purity of the homogeneous subgroups if we increase the level of pruning the decision tree, L? Figure S1A in the supplement document shows exactly the same decision tree as in Figure 3A after increasing L to 3 or more levels. In fact, an L-sensitivity analysis with the simulated 3D data (described above) showed that increasing L above 4 produces the expected 5 homogeneous groups but decreases the expected number of cells per group (results not shown). CCAST automatically determines L based on the homogeneity of the subpopulations (Materials and Methods). Next, we performed a bootstrap analysis to assess the range of values for the split points in the optimal decision tree. More specifically, we performed a strata-sampling method with replacement to generate 200 bootstrap datasets of the same sample size as the training data. We ran CCAST on these samples to generate 200 decision trees with different split points. The hierarchy and selected markers for these bootstrap samples were exactly the same as shown in Figure 3A. We show the confidence intervals of the split points by minimum and maximum boundary estimates from our bootstrap analysis (see range located beside split point estimates in Figure 3A). Note that we could not calculate the normal confidence intervals for these split point estimates due to the multi-modal nature of the split point distributions (Figure 3B). To test the performance of CCAST, we applied CCAST on the test data using decision tree derived from the training set (Figure 3A). After data filtering, the final decision tree on the test dataset is shown in Supplemental Figure S1B. Note that all split point estimates lie within the previously estimated confidence intervals shown in Figure 3A. In addition the hierarchy of the tree remains the same. This result demonstrates that CCAST yields robust split point estimates and can produce reproducible results. Finally, we compare the CCAST result before (Figure 3C) and after data filtering (Figure 3D). Figure 3C show a 2D scatter plot of the 2 markers that partition the training data into clearly 4 clusters. Although there is a strong evidence of 4 clusters, it is apparent that sorting out the population in the yellow cluster without contaminating green cells would be challenging. Figure 3D show the results after applying CCAST on the training data for data analysis. Notice the pure subpopulations after applying the data-filtering step of CCAST. Hence, in addition to providing a gating strategy, CCAST can also produce a more homogenous representation of the original data for data analysis.
Using the T-cell dataset described above, we show that our CCAST-derived gating strategy reproduces the manual gating results in Bendall et al. [9] without relying on expert knowledge. Figure 3E shows that CCAST isolates the 4 distinct T-cell states using only 2 of the 13 measured surfaces markers. These two markers turn out to be CD4 and CD45RA. Figure 3F shows the distribution of the 4 labeled T-cell subtypes based on CD4 and CD45A expression. This result demonstrates that CCAST can identify the 2 of 13 markers that are known to be most relevant to identifying the subtypes of interest without relying on prior knowledge of the subtypes or the markers that are best known to define them. Moreover, for data analysis, CCAST provides more homogeneous subpopulations by filtering out the contaminating cells; an analogous step was not performed in the manually gated analysis [9].
We next applied CCAST only on the manually gated B-cell subpopulations of the Bendall et al. study [9]. In this study, the manually gated B-cell subtypes were: early Pre-B I cells, late Pre-B II cells, immature B-cells, naive mature CD38mid B-cells and mature CD38low B-cells (see Figure 1 in Bendall et al. [9]). To verify the existence of these 5 major B-cell subpopulations, we performed hierarchical clustering, with a cutoff of 5 clusters, on the pooled manually-gated B-cell data, which consisted of about 17,000 cells. The silhouette plot in Figure 4A shows strong evidence of 5 clusters. Figure 4B shows the CCAST-derived gating strategy as a decision tree whereby the 5 distinct cell types can be isolated using only 4, of the 13, surface markers (namely CD45, CD34, CD38 and CD123) with only 3 levels of branching. A cross classification analysis between the CCAST-derived versus the manually gated subtypes is summarized as a heatmap in Figure 4C. Based on Figure 4C, we predict that subpopulations comprising CCAST-derived Cell-types 1, 4, 3, and 5 are predominately immature B, mature CD38low B, Pre B II, and Pre B I cells, respectively. However, there is not a clear one-to-one mapping between the CCAST-derived and manually gated subtypes. In particular, Figure 4C shows strong evidence of a mixture of the mature B-cell subtypes in CCAST Cell-types 2 and 4. The heatmaps in Figure 4D show evidence of two CCAST-derived distinct cell types corresponding to Cell-types 2 and 4 which were considered as one major population, namely mature CD38low B-cells, by manual gating. Based on surface marker expression, the most striking difference between Cell-types 2 and 4 is the expression of CD123, a signaling molecule which promotes proliferation and differentiation within the hematopoietic cell lines and is associated with hairy cell leukemia [24]. Figure 5A provides the heatmaps of BCR, IFNa, FTL3, IL3, IL7, and SCF induced intracellular signaling responses in the 5 CCAST-derived B-cell subtypes compared with an unstimulated control. For the purpose of comparing with the results of Bendall et al. [9] signaling induction was calculated using the difference of the mean scaled arcsinh value of unstimulated condition and the mean scaled arcsinh value of a stimulated condition; moreover, only the 13 surface markers were used to predict the cell types in the stimulated conditions using the decision tree from the unstimulated controls. The difference is calculated as a difference of absolute fold changes. BCR, IFNa, IL7 and SCF stimulations induce strong intracellular signaling across the B-cells across the different development stages. The heatmap in Figure 5B provides heatmaps of BCR, IFNa, FTL3, IL3, IL7, and SCF induced intracellular signaling responses for various B-cell subtypes derived from the manual gating in [9]. In the manually gated cells, the strongest signaling differences are limited to mature B-cells particularly associated with P38 and Ki67. In the CCAST-gated cells, BCR stimulation induces strong differences in PLC-gamma2 signaling, STAT3, H3, S6, CREB; IL7 stimulation alters ERK1/2 and P38 signaling, INFalpha alters STAT3 signaling; and SCF induces changes in P38 signaling. Overall, compared to the manually gated cell types, the CCAST-derived cell types exhibit more differences in stimulated induced signaling, presumably because the CCAST-gated subpopulations are more homogeneous. Finally, as an aside, we note that CCAST produces 7 homogeneously gated subpopulations, 3 of which belong to Cell-type 3, suggesting that this cell type may be more heterogeneous than suggested by the clustering algorithm.
We applied CCAST on about 1 million cells of a SUM159 (triple negative) breast cancer cell line. We generated primary FACS analysis on SUM159 cell line for this study based on expression of EPCAM, CD24 and CD44 (see Materials and Methods). To assess the likely number of cell clusters in SUM159, we ran the “npEM” cluster algorithm, assuming 10 clusters, on a random subsample of about 3,000 cells and obtained 5 clusters. Using hierarchical clustering with a cut off of 5 clusters, on the entire SUM159 dataset, CCAST-derived the gating strategy that is shown in Figure 6. CCAST identified 9 homogenous subpopulations denoted as P1–P9 at the terminal nodes of the tree in Figure 6. A similar implementation on flowJo showing 9 homogeneous clusters is shown in Supplemental Figure S2. Figure 7A summarizes the results for the estimation process for all the split point statistics on all the inner nodes of the CCAST decision tree. The root node corresponding to EPCAM shows one global maximum indicating a strong split point. Nodes 3, 4, 8, 9, 13 and 14 have clear natural maxima indicating optimal splits for the data into clearly 9 subpopulations, each corresponding to 9 single mode histograms in the leaf nodes of the tree. Corresponding barplots for all 9 subpopulations with standard deviation intervals for each marker are shown in Figure 7B. A multivariate Hotelling's T square test showed significant differences between group pairs (p-value: 0), indicating that these 9 nine subpopulations are statistically different from each other. Interestingly, CCAST splits cluster 1 into the subpopulations P5, P6 and P8; it also splits cluster 3 into the subpopulations P3, P4 and P7.
Next we compare the results of the CCAST-derived gating strategy on SUM159 to the manually-defined gating strategy by Gupta et al. [3] on the same cell line. Gupta et al. identified three cell states (stem-like, basal-like and luminal-like cells) in SUM159 based on the three markers (EPCAM, CD24 and CD44). Based on prior knowledge, the stem-like cells were defined as CD44-high, CD24-neg, and EPCAM-low; basal-like cells were defined as CD44-high, CD24-neg and EPCAM-neg; and luminal-like cells were defined as CD44-low, CD24-high, and EPCAM-high. Figure 7C reproduces the Gupta et al. [3] gating strategy on FCM file analyzed in Figure 6. Gupta et al. strategy first gates the cells based on EPCAM high and low then gated the stem, luminal and basal like subpopulations based solely on CD24 low and high, as shown in Figure 7C. A cross classification table of our 9 subpopulations and the 3 Gutpa et al. cell states (labeled as stem, luminal and basal like subpopulations) is shown in Figure 7D. This analysis indicates that the basal-like subpopulation identified by the Gupta et al. gating is a combination of all the CCAST-derived cell states. Furthermore the analysis suggests that a mixture of basal-, stem- and luminal-cell like populations from the Gupta et al. sorting actually correspond to a single CCAST subgroup P9. This results implies that the cell-type specific analysis provided by Gupta et al. may have reflected the behavior of a single cell type. The Gupta et al. analysis may have been more informative if it were to investigate the distinct subpopulations, such as P1, P2, P5 and P7.
Finally, for experimental validation, we applied our CCAST-derived gating strategy on a SUM159 cell line in real time at a FACS machine. Supplemental Figure S3 shows the sorting result from this independent replicate; we are able to recover 5 distinct CCAST-derived subpopulations in real-time.
We compare the application of CCAST and RchyOptimyx algorithm on the FCM data of the SUM159 breast cancer cell line. As briefly described in the Introduction, RchyOptimyx provides a gating strategy to identify target populations at various levels of purity [23]. On SUM159, RchyOptimyx initially generates 27 subpopulations for analysis. Because there is no clinical outcome variable to filter through these 27 predicted phenotypes using the RchyOptimyx algorithm, we selected only the phenotypes that correspond to a combination of CD24 and EPCAM for comparison to CCAST. Recall that CCAST resulted in 9 homogenous subpopulations that can be characterized in terms of these 2 markers alone. Based on use of EPCAM and CD24 alone RchyOptimyx yielded 12 subpopulations that can be targeted by a variety of gating strategies as shown in Supplement Figure S4. In other words, RchyOptimyx provides several possible paths to a particular subpopulation; in comparison, CCAST offers only a single path to target homogenous subpopulations thereby circumventing any additional interpretation of the output from RchyOptimyx for choosing the gating strategy. The underlying formalism of RchyOptimyx and CCAST are different but a full description of those differences is beyond the scope of this analysis.
We presented a model-based gating strategy, CCAST, for sorting a homogeneous subpopulation from a heterogeneous population of single cells without relying on expert knowledge. To identify a hierarchical 2D gating scheme to sort out homogeneous cells, we propose CCAST as a new approach that addresses three key and often-neglected questions: (1) How do we select the optimal markers for gating? (2) What is the optimal ordering of markers for sorting? (3) How do we estimate the marker cut offs for drawing the gates? The answers to these questions are usually decided in a subjective and bias manner making it very difficult to draw precise conclusions from the resulting sorted data. CCAST is an automated and unbiased strategy, requiring minimal human expertise, for optimizing gating of single cell data. While CCAST is optimized for cell sorting it can be applied for analysis of purified subpopulations among heterogeneous single cell data.
In all applications of CCAST in the study, we show that it is possible to characterize and isolate cell types based on a subset of the measured markers. When applied to normal bone marrow data, CCAST reveals an efficient strategy for gating T-cells. CCAST also produced an alternative gating framework for B-cells that produced a new characterization of mature B-cells into CD123+ and CD123- cells. The ability to isolate important cell subpopulations based on limited markers is particularly important since high-throughput cytometry technologies are increasing the number of markers they can measure and one will need new approaches to optimally select important set markers for gating. Hence CCAST not only provides the relevant marker set, optimized gating scheme, and reduces the need for human expertise, it can also reduce the number of antibodies needed for cell sorting.
We further motivated the need for CCAST as an automatically-generated gating scheme that does not rely on prior knowledge of cell states or marker relevance on the SUM159 breast cancer cell line. On this cell line Gutpa et al. tested the hypothesis that cancer cells can transition in any of the several possible cell states which exhibit important functional properties [3]. This study aimed to demonstrate the evidence of phenotypic switching between stem, basal and luminal breast cell states, which were defined by CD24 and EPCAM. Establishing strong evidence of cell state transitions would require pure cell states at onset, however, pure sorting is not evident by the manual gating scheme used in the study. In an independent study on the issue of phenotypic switching of cancer cell states, Zapperi and Porta [4] gave an alternative interpretation of the Gupta et al. based on an imperfect marker scenario. The CCAST analysis also infers nonhomogeneous subpopulations under the Gutpa et al. gating strategy and provided an alternative, more homogeneous cell states using an alternative gating strategy based on the same markers, namely CD24 and EPCAM. CCAST identifies at least 5 distinct breast cancer cell states in SUM159 and sorted out these pure cell states automatically (Figure 6) using only two surface markers, namely EPCAM and CD24. These subpopulations warrant further investigation to validate the notion of phenotypic switching in breast cancer cells as proposed by the Gupta et al. study.
CCAST enables the possibility to sort out unique underlying unknown cell states from a heterogeneous parent population in an optimal and unbiased manner using a gating scheme based on a decision tree representation. CCAST identifies homogeneous cell subpopulations using a non-parametric mixture distribution. Although several other clustering algorithms can also be used, CCAST can handle the unknown number of true clusters without the mathematical optimization of a distribution function. Silhouette coefficients are used to optimize the cell subpopulations and a recursive partitioning technique on the complete data given the cell states is used to generate the optimal decision tree for isolating the various subpopulations of interest. The partitioning comes after a marker selection step, which depends on a non-parametric test statistic making it completely data driven. CCAST also provides a confidence interval for marker cut-offs taking into account possible variability in marker distributions. For future methodological improvement on CCAST to both the computational cost and the pruning level L, one might consider multi way splits at each node, instead of using binary splits. Another methodological direction could be to use the confidence intervals to further enhance the decision trees; in particular, methods proposed by Katz et al. [25] can be adapted for CCAST.
In summary, CCAST is a fully automated model framework to identify a gating strategy to isolate subpopulations from single cell data with greater homogeneity compared to manual gating procedures. More generally, the CCAST framework could be used on other biological and non-biological high dimensional data types involving a mixture of unknown homogeneous subpopulations.
CCAST formalizes the gating process of single cells as a statistical model and provides a simple unbiased hierarchical 2D gating scheme with the relevant set of marker cut-offs for gating a homogenous cell subpopulation given FCM data. Following, we describe the various steps in the non-parametric model framework of CCAST when applied to single cell data. A typical FCM dataset comprises simultaneous quantitative signal measurements of multiple biomarkers of single cells. These measurements can be fluorescence or atomic mass based. The data are stored in flow cytometry standard (FCS) files as a data frame with rows representing the cells or events and the columns corresponding to the markers of interest. Currently, we assume that the data have already been compensated to correct for spectral overlap during data generation and preprocessed using standard preprocessing steps in analysis of FCM data to remove spurious events. The data is then transformed using the recommended Arcsinh function [9] which can handle both positive and negative expression values.
Sum 159 cells were cultured in Ham F12 medium supplemented with 5% calf serum, insulin (5 ug/ml), hydrocortisone, Pen/Strep/L-Glutamine. Cells were grown at 37° C in a 5% CO2 incubator. Stock aliquots of cells were frozen in 10% DMSO and 90% FBS and stored in −80° C liquid nitrogen. The cells were thawed initially into T25 flasks and allowed to expand in culture for two weeks prior to sorting (expanded into T75 flasks). The day of sort, cells were trypsinized, washed with PBS and stained with antibodies specific for the following human cell surface markers: EPCAM (ESA)-FITC (AbD Serotec, MCA1870F), CD24-PE (BD Biosciences), CD44-APC (BD Biosciences), CD49f-PerCP/Cy5.5 (Biolegend). Roughly 1×107 cells were incubated with antibody (20uL antibody per million cells) for 15 min at room temperature in PBS with 1% BSA. Unbound antibody was washed off and cells were analyzed on a custom Stanford and Cytek upgraded FACScan (Beckman Center, Stanford) no more than one hour after staining. Cell sorting was performed on BD Aria II (Beckman Center, Stanford). The raw data is available in supplement Dataset S1 as an FCS file.
|
10.1371/journal.pcbi.1005248 | A human judgment approach to epidemiological forecasting | Infectious diseases impose considerable burden on society, despite significant advances in technology and medicine over the past century. Advanced warning can be helpful in mitigating and preparing for an impending or ongoing epidemic. Historically, such a capability has lagged for many reasons, including in particular the uncertainty in the current state of the system and in the understanding of the processes that drive epidemic trajectories. Presently we have access to data, models, and computational resources that enable the development of epidemiological forecasting systems. Indeed, several recent challenges hosted by the U.S. government have fostered an open and collaborative environment for the development of these technologies. The primary focus of these challenges has been to develop statistical and computational methods for epidemiological forecasting, but here we consider a serious alternative based on collective human judgment. We created the web-based “Epicast” forecasting system which collects and aggregates epidemic predictions made in real-time by human participants, and with these forecasts we ask two questions: how accurate is human judgment, and how do these forecasts compare to their more computational, data-driven alternatives? To address the former, we assess by a variety of metrics how accurately humans are able to predict influenza and chikungunya trajectories. As for the latter, we show that real-time, combined human predictions of the 2014–2015 and 2015–2016 U.S. flu seasons are often more accurate than the same predictions made by several statistical systems, especially for short-term targets. We conclude that there is valuable predictive power in collective human judgment, and we discuss the benefits and drawbacks of this approach.
| Despite advanced and widely accessible health care, a large number of annual deaths in the United States are attributable to infectious diseases like influenza. Many of these cases could be easily prevented if sufficiently advanced warning was available. This is the main goal of epidemiological forecasting, a relatively new field that attempts to predict when and where disease outbreaks will occur. In response to growing interest in this endeavor, many forecasting frameworks have been developed for a variety of diseases. We ask whether an approach based on collective human judgment can be used to produce reasonable forecasts and how such forecasts compare with forecasts produced by purely data-driven systems. To answer this, we collected simple predictions in real-time from a set of expert and non-expert volunteers during the 2014–2015 and 2015–2016 U.S. flu seasons and during the 2014–2015 chikungunya invasion of Central America, and we report several measures of accuracy based on these predictions. By comparing these predictions with published forecasts of data-driven methods, we build an intuition for the difficulty of the task and learn that there is real value in collective human judgment.
| It is perhaps unsurprising that societal advances in technology, education, and medicine are concomitant with reduced morbidity and mortality associated with infectious diseases [1]. This is exemplified by the staggering drop in infectious disease mortality in the United States during the 20th century [2]. Yet despite continued technological advances, the U.S. death rate due to infectious disease has not improved significantly since around the 1960s, and in fact has been rising since the 1980s [3]. It is evident that incremental advances in medical treatment are failing to reduce overall infectious disease mortality. One way in which the situation can be improved is through expanding our capacity for preparedness and prevention—we need forewarning [4]. This is the defining problem out of which the nascent field of epidemiological forecasting has risen.
There is widespread interest in predicting disease outbreaks to minimize losses which would otherwise have been preventable given prior warning. In recent years, the U.S. Centers for Disease Control and Prevention (CDC) has sponsored three challenges to predict Influenza epidemics [5–7], Defense Advanced Research Projects Agency (DARPA) has sponsored a challenge to predict the invasion of chikungunya [8, 9], several agencies working together under the Pandemic Prediction and Forecasting Science and Technology (PPFST) Working Group have sponsored a challenge to predict Dengue outbreaks [10], and the Research and Policy for Infectious Disease Dynamics (RAPIDD) group of the National Institutes of Health hosted a workshop for forecasting Ebola outbreaks [11]. Given the magnitude of time, energy, and resources collectively invested in these challenges by both participants and organizers, it is critical that qualitative and quantitative assessments be made to help understand where epidemiological forecasting excels and where it lags.
As exemplified by the fields of meteorology and econometrics, statistical and computational models are frequently used to understand, describe, and forecast the evolution of complex dynamical systems [12, 13]. The situation in epidemiological forecasting is no different; data-driven forecasting frameworks have been developed for a variety of tasks [14–16]. To assess accuracy, forecasts are typically compared to pre-defined baselines and to other, often competing, forecasts. The focus has traditionally been on comparisons between data-driven methods, but there has been less work toward understanding the utility of alternative approaches, including those based on human judgment. In addition to developing and applying one such approach, we also provide an intuitive point of reference by contrasting the performance of data-driven and human judgment methods for epidemiological forecasting.
Methods based on collective judgment take advantage of the interesting observation that group judgment is generally superior to individual judgment—a phenomena commonly known as “The Wisdom of Crowds”. This was illustrated over a century ago when Francis Galton showed that a group of common people was collectively able to estimate the weight of an ox to within one percent of its actual weight [17]. Since then, collective judgment has been used to predict outcomes in a number of diverse settings, including for example finance, economics, politics, sports, and meteorology [18–20]. A more specific type of collective judgment arises when the participants (whether human or otherwise) are experts—a committee of experts. This approach is common in a variety of settings, for example in artificial intelligence and machine learning in the form of committee machines [21] and ensemble classifiers [22]. More relevant examples of incorporating human judgment in influenza research include prediction markets [23, 24] and other crowd-sourcing methods like Flu Near You [25, 26].
Here we assess the performance of a forecasting framework based on collective human judgment—“Epicast”. In particular, we assess its performance as a competitor in the aforementioned influenza and chikungunya forecasting challenges. Each of these challenges dealt with a different set of data and objectives, and we analyze them separately. For both influenza challenges, the data of interest was population-weighted percent influenza-like illness (wILI) in 10 regions of the U.S. and the U.S. as a whole. wILI is syndromic surveillance defined as the percent of patients having flu-like symptoms (fever over 100°F and either cough or sore throat) without a known cause other than influenza. wILI is reported voluntarily by health care providers through ILINet and distributed by CDC [27, 28]. The DARPA chikungunya challenge focused instead on predicting case counts within 55 countries and territories in the Americas. The data of interest was the number of weekly cases (including suspected, confirmed, and imported) within each location, cumulatively since the beginning of 2014. These case counts are published by the Pan American Health Organization (PAHO) [29].
In regards to terminology in subsequent discussion, there is an important distinction to be made between a prediction and a forecast. These words are often used interchangeably elsewhere, but here we use them to refer to subtly different concepts. A prediction makes an absolute statement about the future and says nothing about other potential outcomes. In contrast, a forecast is a generalization of a prediction in which a probability is assigned to all possible outcomes. In the case of Epicast, we collect a prediction from each participant—a single possibility. Epicast aggregates many such predictions to produce a forecast—a probability distribution over all possibilities. Because predictions and forecasts make different claims, they are evaluated by different metrics; we use mean absolute error (MAE) to assess predictions and mean negative log likelihood (based on “logarithmic score” [30], also known as “surprisal” in other contexts [31]) to produce a figure of merit for forecasts.
As part of our evaluation of the Epicast system in forecasting flu, we compare with several competing statistical and/or data-driven systems. For the 2014–2015 flu contest, we compare Epicast with “Empirical Bayes” and “Pinned Spline”; for the 2015–2016 flu contest, we compare Epicast with “Stat” and “ArcheFilter”. All of these systems were serious and successful competitors in their respective contest years. Consequently, the performance of these systems provides a measure of the state of the art in flu forecasting. None of these competing systems were used in both flu contests, hence we compare against a different set of systems in different years.
The Empirical Bayes system [32] is based on the notion that future epidemics will generally resemble past epidemics, up to basic transformations and noise. This system defines a prior distribution over wILI trajectories, draws samples from that distribution, assigns a likelihood-based weight to each sample, and finally produces a posterior distribution over trajectories. The reported forecast for each target is derived from the posterior distribution of wILI trajectories.
The Pinned Spline system [33] attempts to smoothly interpolate current and past wILI. To do this, two partial wILI trajectories are defined. The first partial trajectory spans the start of the season through the current week and is defined to be wILI as published by CDC for the current season. The second partial trajectory spans the next week through the end of the season and is defined to be the week-wise mean of the wILI trajectories of past epidemics over the same span. Finally, the two partial trajectories are connected with smoothing splines. In subsequent analysis we only evaluate point predictions—not forecasts—made by the Pinned Spline system.
The Stat system is a weighted ensemble of statistical methods, including both Empirical Bayes and Pinned Spline. It additionally contains baseline components (including a uniform distribution and an empirical distribution) and other non-mechanistic methods (including delta density and kernel density methods). The cross-validation weight assigned to each constituent method is recomputed for each forecast. Stat forecasts are a weighted combination of the forecasts of each method.
The ArcheFilter system [33] assumes that there is a latent archetype wILI trajectory, and that the observed wILI trajectory of each flu epidemic is a transformed, noisy version of the archetype. The ArcheFilter defines this archetype roughly as the peak-aligned, week-wise mean wILI trajectory of all past epidemic seasons. As an epidemic progresses, a Kalman filter is used to estimate the time-shift and wILI-scale parameters that, when applied to the archetype, most parsimoniously explain the observed wILI trajectory of the current epidemic. Uncertainty in the state of the filter—the shift and scale parameter values—gives rise to a distribution over wILI trajectories from which the forecast for each target is derived.
This study was granted Carnegie Mellon University IRB exemption with ID STUDY2015_00000142.
We developed a website [34] for collecting predictions of epidemiological time series (Fig 1). Participants were shown a partial trajectory and were asked to hand-draw a continuation of the trajectory as a prediction. At regular intervals, user-submitted trajectories were collected and an aggregate forecast was generated. Participants were not shown the predictions made by other participants. We produced, in real-time, forecasts for the 2014–2015 and 2015–2016 U.S. flu seasons and predictions for the 2014–2015 chikungunya invasion of the Americas. For influenza, we collected predictions from the general public; for chikungunya, we only collected predictions from a set of selected experts in related fields. All three forecasting challenges were carried out as the event (epidemic or invasion) progressed; we did not produce retrospective forecasts.
Each week during the 2014–2015 and 2015–2016 flu seasons, we asked participants to predict wILI for each remaining week of the season. Each individually submitted prediction was a trajectory of varying length (depending on the week of submission) of wILI values, and we asked users to provide such predictions for each of the HHS and U.S. regions. Similarly, once each month, from August, 2014 through January, 2015, we asked participants to predict the cumulative weekly chikungunya case count in each of the 55 Pan American Health Organization (PAHO) locations through the end of February 2015.
Each predicted trajectory was extended to cover the entire time series of interest by concatenating the observed time series at the time of submission with the predicted time series. By doing this, all time series were made to have the same length (for example, 32 weekly wILI values spanning the 2014–2015 flu season).
The objectives of the influenza challenges were to forecast a number of features, or “targets”, of the weekly wILI time series. These included: “Peak Height”, the maximum value of wILI reported throughout the flu season; “Peak Week”, the Morbidity and Mortality Weekly Report (MMWR) week number [35] on which wILI reaches its maximum value; and the next four values of wILI, called “1–4 Week Lookaheads”. Each of these values was forecasted separately for each of the ten Health and Human Services (HHS) regions of the U.S. and also for the U.S. as a whole—a super-region which we include in our analysis as one of eleven total regions. The forecast for each target in each region consisted of a set of probability bins over a pre-defined range of outcomes and a prediction of the single most likely outcome. The chikungunya challenge objective was simply to predict the trajectory of cumulative case counts in each country or territory.
Of the six flu targets, two are season-wide: Peak Week and Peak Height. The rest—the 1–4 Week Lookaheads—are short-term targets. Peak Week is an integer representing the number of weeks elapsed between the start of the season and the week during which wILI peaks. The start of the season is defined as MMWR week 40 of the first year of the season, written as “2014w40” or “2015w40”, depending on the season. The remaining five targets are measured in wILI on a continuous scale from 0% to 100%. An additional season-wide target, the MMWR week of epidemic onset, is discussed in S1 Text.
The Epicast point prediction for any target was defined as the median of the target values measured on user predictions. The Epicast forecast for any target was a Student’s t distribution with location equal to the median value (the point prediction), scale equal to the sample standard deviation of values, and degrees of freedom equal to the number of participants. The same general methodology was used for both forecasting challenges, with the exception that we produced both a prediction and a forecast for influenza and only a prediction for chikungunya.
We primarily assess the quality of predictions in terms of mean absolute error (MAE). Given a set of N true outcomes, y, and corresponding predictions, y ^, MAE can be written as:
MAE = 1 N ∑ i = 1 N ∣ y i - y ^ i ∣ .
We further assess wILI and case-count predictions (for example, Peak Height) by measuring how often predictions fall within some range (± 10%, 20%, 30%, 40%, or 50%) of ground truth. We similarly assess predictions of weeks (for example, Peak Week) by measuring how often predictions fall within some range (± 1, 2, 3, 4, or 5 weeks) of ground truth. We report the fraction of the time that predicted values fall within each range, aggregated over regions and potentially also weeks. We refer to these analyses as “fraction of predictions accurate within a target range”.
We assess the quality of flu forecasts in terms of a likelihood-based score. We define the “log score” as the negative average of the logarithm of the probability assigned to a range of values surrounding the true outcome. In the case of Peak Week, we consider the log score of the range of the actual Peak Week plus or minus one week (for example, if the Peak Week was 5, we compute the log likelihood of the probability assigned to a peak being on week 4, 5, or 6). Suppose that PkWk r obs denotes the observed value of Peak Week in region r and that P(⋯) represents the probability assigned by the forecaster to a given outcome. Then the score across all regions can be written as:
score = - 1 11 ∑ r = 1 11 log P ( PkWk r ∈ [ PkWk r obs - 1 , PkWk r obs + 1 ] ) .
For the five wILI targets, we only have available a set of probability bins each of width 1 wILI, as this is what was required by CDC’s 2014–2015 flu contest. To determine which bins to include in the likelihood calculation, we select (1) the wILI bin containing the actual value and (2) the adjacent wILI bin nearest to the actual value. For example, the actual Peak Height in the U.S. National region was 6.002, and we select the two bins which together give the probability assigned to the event that actual Peak Height falls between 5 and 7. Suppose a forecast was made that P(5 ≤ wILI < 6) = 0.215 and P(6 ≤ wILI < 7) = 0.412; the log score assigned to this forecast is − log(0.215 + 0.412) = 0.467. For Peak Height (and similarly for the Lookahead targets) across all regions:
score = - 1 11 ∑ r = 1 11 log P ( PkHt r ∈ [ round ( PkHt r obs ) - 1 , round ( PkHt r obs ) + 1 ] ) .
To compensate for the varying difficulty over time of predicting and forecasting, we often treat accuracy as a function of “lead time”, the number of weeks preceding the region-specific Peak Week. Positive, zero, and negative lead times indicate predictions and forecasts made before, on, and after the epidemic peak, respectively. We consider lead times for the 2014–2015 flu season that range from +10 to −10 weeks. However, due to the unusually late Peak Week within most regions in the 2015–2016 flu season, we constrain lead times in that season to the range of +10 to −5 weeks.
To contextualize the accuracy of both predictions and forecasts, we compare Epicast with individual participants and/or other forecasting methods. To further contextualize log scores, we show also the log score of a hypothetical “Uniform” system in which uniform probability is assigned to all plausible outcomes. For Peak Week, we define this as a uniform distribution over weeks 2014w46 through 2015w12 and 2015w45 through 2016w12 (p = 1 20 per week, per season), and for the wILI targets we define this as a uniform distribution over 0% to 12% wILI (p = 1 12 per bin). The Uniform system is intended to provide a lower bound on the performance of a reasonable forecaster.
Our main challenge in presenting results is that the space in which comparisons can be made consists of several orthogonal dimensions: regions (U.S. and 10 HHS regions), targets (Peak Week, Peak Height, and 1–4 Week Lookaheads), season weeks (depending on season and target, up to 32), and error metrics (MAE and log score). To concisely compare system performance, we are given the non-trivial task of reducing this dimensionality, otherwise we would come to thousands of separate figures of merit. Several confounding issues impede aggregation along any one axis; forecasting difficulty varies over time as the season progresses, the various regions may peak at different times in the season, long-term targets are often more difficult to predict than short-term targets, and targets are measured in different units. To work around these complications in the case of point predictions, we rank systems and participants in terms of absolute error and perform our analysis on the relative ranking assigned to each predictor. More specifically, we consider the pairwise ranking in absolute error of Epicast versus individual participants and statistical frameworks. For each lead time, region, and target, we ask whether Epicast or the competitor had a smaller absolute error, and we measure the fraction of instances where Epicast had the smaller error—a “Win Rate”. To assess the statistical significance of each result, we use a Sign test with the null hypothesis that the pair of forecasters is equally accurate. It should be noted that this test assumes that all observations are independent, but results across adjacent weeks, for example, are likely to be correlated to some extent.
We define ground truth to be the version of wILI published by CDC 15 weeks after the end of each flu season—MMWR week 35. Specifically, we use values published on 2015w35 and 2016w35 for evaluating the results of the 2014–2015 and 2015–2016 flu contests, respectively.
For the 2014–2015 flu season we gathered a total of 5,487 trajectories from a set of 48 volunteer participants during the 32 week period spanning 2014w41 through 2015w19. For the 2015–2016 flu season we gathered a total of 3,833 trajectories from a set of 23 volunteer participants during the 30 week period spanning 2015w42 through 2016w19. Participants varied in self-identified skill, from experts in public health, epidemiology, and/or statistics, to laypersons. Participation varied over time with an average of 16 participants per week during the 2014–2015 season and 12 participants per week during the 2015–2016 season (Fig 2). In the following analysis we did not handle expert and non-expert predictions differently, but we compare the performance of the two groups in S1 Text—the experts on average made slightly more accurate predictions. In what follows, we group errors across regions for brevity, but a breakdown of performance within each region is also given in S1 Text.
We first consider the fraction of predictions accurate within a target range, aggregated over weeks of the season (Fig 3). For the four short-term Lookahead targets, the Epicast prediction is within 10% of the actual value just under half the time when predicting one week into the future; this falls to roughly one quarter of the time when predicting 4 weeks into the future. The trend is similar, though perhaps less abrupt, at other accuracy thresholds. Accuracy within 50% is achieved near or above 85% of the time, even predicting up to 4 weeks ahead.
We next consider the number of regional predictions accurate within a target range, as a function of lead time (Fig 4). For 2, 3, and 4 weeks ahead, the lead time with lowest accuracy is roughly 2, 3, and 4 weeks ahead of the Peak Week, respectively, which suggests that there is a distinct challenge in forecasting the Peak Height. This is to be expected because there is significantly more volatility around the peak of the epidemic. In general, accuracy in season-wide targets rises sharply 2–5 weeks before the epidemic peak and remains high through the remainder of the season.
For the 2015–2016 season, short-term accuracy (relative, not absolute) is surprisingly low in several regions around 10 weeks before the peak. This is likely due to the fact that the Peak Week in most regions was exceptionally late during this season. As a result, wILI was still near to baseline values for several weeks after predictions were made. Predicting a premature rise in wILI when ground truth was small in magnitude resulted in errors exceeding 50% of the actual value in several regions. Once it became clear that this would likely turn out to be a mild and/or late-peaking season, accuracy rose to nominal pre-peak levels.
The remainder of our analysis is focused on comparing the accuracy of Epicast with individual participants and competing methods; we begin with Win Rate (Fig 5). Overall, considering all targets, Epicast has Win Rates above 0.5 (lower absolute error than a competitor on a majority of predictions) when compared with all but one individual participant and all four statistical frameworks. In season-wide targets, Epicast performs reasonably well; however, six participants and the ArcheFilter method bring Epicast’s Win Rate below 0.5. In short-term targets, Epicast has Win Rates uniformly above 0.5. Epicast has a Win Rate significantly higher than the Spline method in all categories and significantly higher than the Empirical Bayes and ArcheFilter methods both overall and in short-term targets. Epicast never has a significantly lower Win Rate than any of the competing statistical systems.
Next, we compare predictions in terms of MAE. We calculate, separately for each target, MAE across regions as a function of lead time (Fig 6). In agreement with previous results, Epicast MAE in season-wide targets generally decreases with lead time and is highest in short-term targets when predicting the peak value of wILI. MAE is occasionally elevated in short-term targets on the Peak Week (lead time = 0), suggesting a relative increase in uncertainty immediately after the true peak (which is not known at the time to be so). Compared with the statistical methods, Epicast particularly excels when predicting short-term targets.
Finally, we compare forecasts in terms of log score. Our analysis in this context does not include the 2014–2015 Spline method, but the hypothetical Uniform method is included for both seasons. We compute the average log score for Epicast and competing methods separately for each target and lead time (Fig 7). In the 2014–2015 season, Empirical Bayes scored within the bounds of the Uniform system more consistently than Epicast. However, in the 2015–2016 season, all systems consistently scored within the Uniform bounds for wILI targets, likely due at least partly to a relatively low peak wILI in this season. Across both seasons, Epicast has average log score in short-term targets as good as, or better than, that of the statistical systems. However, the statistical systems almost uniformly outperform Epicast in season-wide targets.
In total, we gathered 2,530 trajectories from a set of 12 volunteers with expertise in vector-borne diseases, public health, and/or epidemiology (Fig 8). Predicting chikungunya fundamentally differed in two ways from predicting flu. First, the chikungunya invasion of the Americas was a rare event for which little historical precedent was available, whereas flu epidemics are a regular occurrence for which we have significant historical data. Second, errors in (cumulative) chikungunya predictions accumulated over weeks, whereas errors in (non-cumulative) flu predictions were separated out across weeks. While it would have been trivial to convert a cumulative trajectory into a non-cumulative trajectory, the published counts which were defined to be ground truth are only available sporadically over time, preventing us from converting the true cumulative trajectory into a non-cumulative trajectory. The increased difficulty of the task is reflected by a reduction in accuracy. At best (1 week ahead), less than one in three predictions were within 10% of the actual value; and at worst (10 weeks ahead), over half of the predictions were off target by more than 50%. Even in such conditions, when comparing pair-wise absolute error between Epicast and each user, Epicast more frequently predicts closer to the true value than any individual user.
Epicast was one of two winning methods in the 2014–2015 flu contest [7]. Epicast was one of three winning methods in the 2015–2016 flu contest [36] (the other winners were Stat and ArcheFilter). We expect a future CDC publication to provide additional details and analysis. Epicast was not selected as one of the six chikungunya challenge winners [37], but we are told that it ranked in the top quartile of submissions.
There are two caveats to point out regarding our incarnation of the crowd prediction method for forecasting flu. First, the ILINet data we showed participants, and also asked them to predict, was subject to weekly revision—in some cases significantly so (for example, wILI in HHS region 2 on 2015w03 was first reported as 6.2%, and then the next week as 5.6%; the final, stable value was reported as 5.0%). The changing values are due to a backfill process whereby data from late-reporting providers is used to retrospectively update prior values of wILI. This is one reason, for example, that MAE after the Peak Week is non-zero; even once the peak has been observed with high confidence, there is still some non-negligible chance that a subsequent update due to backfill will result in a revision of the peak timing. A further discussion of the effects of backfill can be found in S1 Text. Second, the data used in our present analysis is the same data collected for the various contests, but our methodology in the 2014–2015 flu contest differed slightly from what we present here. Namely, our 2014–2015 contest submissions assumed a normal distribution over user inputs, whereas here, and in our 2015–2016 contest submissions, we assumed a Student’s t distribution.
There are several important limitations of the human judgment approach relative to purely data-driven methods that should be made clear. First, these results are only representative of two flu seasons and a single chikungunya outbreak. This highlights one of the biggest shortcomings of this approach—collecting predictions is a tedious and time-consuming process. Unlike statistical methods which can be applied retrospectively to any outbreak, the approach here requires a significant amount of work from a large number of participants. For example, because of this we are unable to perform cross validation across seasons. Second, these results do not necessarily provide us with an improved understanding of epidemiological dynamics. In contrast, statistical methods aim to learn from past data in order to better describe and model the epidemic process. On the other hand, the human judgment approach does have unique advantages over purely data-driven systems. Humans have the innate and powerful ability to assimilate, with little to no effort, diverse data sources and considerations. An example of this is using news headlines, which we display within the Epicast interface, to inform predictions. Another advantage of human judgment is the ability to make reasonable predictions for events with little historical precedent, like the outbreak of a new disease or a disease invasion in a new location.
The task of predicting trajectories is not trivial, and we asked each of the participants to provide us with many such trajectories over quite a long period of time. This resulted in some tedium, which we suspect is the reason for the relatively high attrition rate in the flu Epicast. There are many guidelines describing ways to make crowd work streamlined and sustainable, and we made every effort to implement these ideas. To minimize the overall amount of effort required and to streamline the process as much as possible we: allowed participants to use their previously entered forecasts as a starting point; accepted any number of regional predictions (not requiring all eleven to be completed); reduced the entire process to one drag and one click per region; and sent URLs tailored with a unique identifier via email each week to bypass having to manually login. Additionally, we tried to increase interest and participation by including a leader board of both weekly and overall high scores. We also had the competing objective of collecting the most informed forecasts from the participants. To this end we included a section of links to educational resources, and for the flu Epicast we embedded within each participant’s home page a Google news feed on the topic of “flu”.
Epicast is not well suited for all forecasting situations. “Wisdom of crowds” methods are robust to high variance among individual predictions, but require that the overall distribution of predictions is unbiased. By showing wILI trajectories of past flu epidemics on Epicast’s forecasting interface, we undoubtedly bias predictions toward typical flu seasons. While this may be beneficial when forecasting a typical flu season, it is almost certainly harmful when predicting highly atypical flu epidemics, and especially pandemics. Epicast is not robust to “long-tail” events such as these. In general, Epicast is best suited for situations where: (1) the event is regularly occurring, (2) historical surveillance data is available for many examples of the event, and (3) ongoing surveillance data is available with relatively short lag in comparison to the length of the event. This may explain the increased difficulty of Epicast in predicting the chikungunya invasion: it was a one-time event with no historical data and relatively lagged and intermittently available ongoing data.
Alternative prediction methods which solicit and aggregate human judgment exist; the Delphi method [38] is one such example. Both the Delphi method and the Epicast method herein collect predictions from human participants and produce a single output prediction. However, the Delphi method is iterative, requiring more time, effort, and coordination. The Delphi method requires from each participant not only a prediction, but also reasoning or justification for that prediction. Then, all participants are shown the predictions, and reasons for them, of all other participants. Participants are then given the opportunity to revise their predictions, and the process continues iteratively either until convergence or some other stopping criteria are met. One of the design goals of Epicast was to minimize human time and effort required, and so we did not pursue the Delphi method. However, it would be of value to compare the Epicast and Delphi methods to learn the relative advantages and disadvantages of each method in terms of both human effort and prediction accuracy.
It was our hope that the number of participants would grow organically, for example through word of mouth and social media. Instead, we found it difficult to recruit new participants and to maintain participation throughout the flu season. The failure to achieve a true “crowd” is most likely due to the tedium of the task, and we are working on ways to both reduce this tedium and to make the task more gratifying for participants. While we strove to design the user interface in a way that minimizes the level of effort required to input predictions, there is always room for further improvement. One option we considered, but did not implement because of the small number of participants, is to reduce workload by asking participants to provide a prediction only for a randomized subset of regions. Another option we considered, but did not completely implement due to time constraints, was gamification. This was partially implemented in the form of leader boards, but it would be difficult to provide a more immediate reward because of the inherent delay between prediction and revelation of true outcomes.
There are several additional ways in which the Epicast method could be improved. First, there is an important relationship between a prediction, and the level of confidence in that prediction, that we were unable to capture. We asked participants to give us their best point predictions, but there was no way for them to communicate with us their level of confidence in those and other predictions—a forecast. We made the implicit assumption that disagreement among user predictions implies lack of confidence, which is probably true to some extent. The inverse however—that uniformity in predictions implies high confidence—is clearly untrue. Consider as an example the case where everyone believes that next week’s wILI has a 60% chance of staying the same as this week’s wILI, resulting in all point predictions strongly concentrated on the same wILI, and the distributional spread being very narrow, in contrast with the participants’ beliefs. It would be ideal to collect from each user a more informative measure of their confidence, but this would undoubtedly complicate the user interface and degrade the overall experience (which we were averse to).
Another improvement to consider is a weighted combination of predictions whereby participants who have historically had more accurate predictions are given more weight in the aggregation process. This is similar in spirit to weighting user recommendations and rankings, which has been shown to increase accuracy in those settings [39, 40]. In the case of Epicast, there is limited evidence suggesting that some participants are overall more (or less) accurate than other participants. One example of this is in Fig 5B where one participant has significantly higher Win Rate than Epicast and several other users. On the other hand, it is not clear whether the variance of prediction error is sufficiently small to learn which users are the most accurate in a reasonable amount of time—before the epidemic peak, for example. In other words, differences in accuracy may be exploitable, but only if precision is sufficiently high. If this is the case, then an adaptive weighting scheme could benefit the overall forecast. However, there is a critical obstacle that hinders the practical implementation of such a scheme: because of backfill, the final measure of accuracy is not known for many months. Despite this, we propose, implement, and analyze one such scheme in S1 Text—the result is a small and statistically insignificant increase in accuracy.
A natural evolution of systems such as those for epidemiological forecasting is the combination of human and statistical (machine) methods [41, 42]. The first question in such a project is whether human predictions should be given as input to statistical methods or whether the output of the statistical methods should be shown to humans for more informed predictions. In theory both directions are viable, and there are intuitive reasons for each. In support of the latter, people are naturally inclined to trust forecasts made by humans (or to distrust forecasts made by machines), a phenomenon known as algorithm aversion [43]. Supporting the former, on the other hand, is the observation that in many settings and in a variety of tasks, objective machine prediction is often superior to subjective human prediction [44, 45]. We have begun to explore both directions; currently, we show a subset of Epicast participants a confidence band derived from a separate statistical method, and we are developing a retrospective analysis wherein we compare performance of various statistical methods with and without a supplemental input of human prediction as an independent data source. In the meantime, we plan to continue to host Epicast and collect predictions for the current flu season, and we end this section with an open invitation to participate [34].
For years, both humans and machines have been employed to tackle difficult prediction problems, and the biases involved and the relative advantage of data-driven approaches are at least well documented [43, 44], if not well understood. We do not make the claim that human judgment is intrinsically more valuable or more capable than machines when making epidemiological forecasts, but we do posit that there is value in understanding the strengths in each approach and suspect that both can be combined to create a forecasting framework superior to either approach alone.
|
10.1371/journal.pcbi.1004164 | Theory of Choice in Bandit, Information Sampling and Foraging Tasks | Decision making has been studied with a wide array of tasks. Here we examine the theoretical structure of bandit, information sampling and foraging tasks. These tasks move beyond tasks where the choice in the current trial does not affect future expected rewards. We have modeled these tasks using Markov decision processes (MDPs). MDPs provide a general framework for modeling tasks in which decisions affect the information on which future choices will be made. Under the assumption that agents are maximizing expected rewards, MDPs provide normative solutions. We find that all three classes of tasks pose choices among actions which trade-off immediate and future expected rewards. The tasks drive these trade-offs in unique ways, however. For bandit and information sampling tasks, increasing uncertainty or the time horizon shifts value to actions that pay-off in the future. Correspondingly, decreasing uncertainty increases the relative value of actions that pay-off immediately. For foraging tasks the time-horizon plays the dominant role, as choices do not affect future uncertainty in these tasks.
| Numerous choice tasks have been used to study decision processes. Some of these choice tasks, specifically n-armed bandit, information sampling and foraging tasks, pose choices that trade-off immediate and future reward. Specifically, the best choice may not be the choice that pays off the highest reward immediately, and exploration of unknown options vs. exploiting known options can be a normatively useful strategy. We characterized the optimal choice strategies across these tasks using Markov Decision Processes (MDPs). The MDP framework can characterize optimal choice strategies when choices are affected by the value of future rewards. We found that uncertainty and time horizon have important effects on the choice strategies in these tasks. Specifically, in bandit and information sampling tasks, increasing uncertainty increases the value of exploring choice options that tend to pay off in the future, while decreasing uncertainty increases the value of choice options that pay off immediately. These effects are increased when time horizons are longer. Foraging tasks differ in that uncertainty plays a minimal role. However, time horizon is still important in foraging. Specifically, for long time horizons, travel delays to rewards become less relevant.
| Decision making has been studied with a wide array of tasks. Choices in many of these tasks either do not affect future choices or are modeled as if they do not affect future choices. For example, when asked to choose between gambles (e.g. 50% chance of $20 or 100% chance of $11), the choice in the current trial does not affect the gambles presented in the next trial, or the information on which one decides in the next trial. Correspondingly, even reinforcement learning tasks, where choices do affect the information that will be available for future choices, are often modeled using delta rule reinforcement learning (DRRL) or logistic regression, neither of which provides a normative description of the task. These modeling approaches assume that current choices should be driven entirely by past outcomes without considering how they will affect the future.
Many interesting decision making problems, however, require consideration of how current choices will affect the future [1–7]. For example, there has been interest in the explore-exploit tradeoff [8–17], information sampling [4,6], and foraging [18,19]. Explore-exploit trade-offs exist in any real-world decision making context where one has to choose between continuing to exploit a known option, for example a familiar restaurant, vs. exploring an unknown or novel restaurant. Similarly, information sampling underlies many deliberative choice processes where one collects information before committing to a decision. For example, one might study product reviews or ratings before making a large purchase. These tasks require more sophisticated choice strategies because choices can be driven by future expected values. In other words, the best choice may not be the one that delivers the largest immediate reward. The best choice may lead to larger rewards in the future at the expense of smaller immediate rewards. Choices in these tasks can be modeled with markov decision processes (MDPs). MDPs provide a general modeling framework, useful in tasks where the future depends upon what one chooses in the present. If one assumes that an agent is maximizing the expected (discounted or undiscounted) total reward, MDPs can be used to provide normative, or at least approximately normative, solutions to most current decision problems.
While the choice behavior of subjects often deviates from normative behavior [4], particularly in patient groups [5,20,21], normative models are still important. Specifically, normative models identify the information on which decisions should be based, and the computations that must be carried out on that information. These two points can be conceptualized as the strategy optimal for the task. Further, normative models can be parameterized to fit the behavior of individual subjects [4,5,22]. This approach can provide insight into how subjects are deviating from the normative model and therefore it can suggest specific deficits or biases, as opposed to an overall change in task performance.
Here, we used MDPs to model n-armed bandit, information sampling, and foraging tasks. Normative solutions to some of these tasks, to our knowledge, do not currently exist in the literature. There is, however, a long theoretical literature on binary bandit tasks [23,24], and some foraging tasks have been modeled using the marginal value theorem [25]. For MDPs, the development of approximation techniques using basis functions has opened up the solution of a much larger class of problems than was tractable previously [26]. The normative solutions provide insight into the optimal strategies. We also used the models to examine several specific questions. For example, when is it useful to explore in a bandit task, and which features of the task can increase the value of exploring? How can non-stationarity drive exploration? Furthermore, once the tasks were mapped into the MDP framework we could examine their similarities and differences. This showed that decisions in all of these tasks pose a trade-off between immediate and future expected rewards. Further, we identified two factors that are important to this trade-off in these tasks. The first is uncertainty and the second is the time horizon. In bandit and information sampling tasks future expected values are relatively higher for options about which there is more uncertainty. When there is less uncertainty, action values are driven more by immediate expected reward. Further, uncertainty, and the value of exploring uncertain options is more valuable when the time horizon is longer. We also show that reward rate maximization in foraging tasks with an undiscounted, infinite time horizon is insensitive to travel delays to patches. In general with MDPs, infinite horizon undiscounted models are insensitive to finite delays to rewards.
We used markov decision processes, either partially observed (POMDPs) or fully observed (MDPs) to model several choice tasks. For these models we are interested in the utility, ut. For MDPs the utility depends on the state, and for POMDPs it depends on the information state. We indicate both states and information states by st. The utility is then given by the action, a, that maximizes action value, Q(st,a):
utst = maxa∈AstQ(st,a)
The action value:
Qst,a = rst,a+C(st,a) +γ∑j∈Sp(j|st,a)ut+1(j)
can be broken down into the immediate: r(st,a)+C(st,a) and future: γΣj∊sp(j|st,a)ut+1(j) expected values, which we will call IEV and FEV respectively. The IEV is the expected reward, r(st,a), for taking action a in state st plus a possible cost to sample, C(st,a). These occur immediately. The FEV is the expected value of the utility of the next state, where the state transition function or the probability of transitioning to state j when taking action a in state st, is p(j|st,a). The FEV is an estimate of the (possibly) discounted future rewards that will be obtained, given the current action. The state is the information on which decisions are based. For most of the tasks, except the foraging tasks, the state is a hidden variable, and this hidden variable gives rise to observations through an observation model. In this case, one is dealing with a partially observable Markov Decision Process (POMDP). The observations define the information state, and one can infer the value of the hidden state using the observations. We explicitly refer to the state as the information state for POMDPs.
When the environment is unknown, and model-free reinforcement learning (RL) is used to learn the environment [27], exploration can be used to drive the RL algorithm to sample from the complete space of possible options. Here we deal with tasks where the environment is specified and MDPs (or POMDPs) can be used to calculate expected values for each state. Therefore heuristic exploration does not have to be used to make choices. Exploration, if it is defined as selecting options which have a smaller IEV but a larger action value, however, can still be optimal [27]. If an agent is maximizing total expected reward, an option with a smaller IEV can be selected if its FEV is relatively larger. Thus, immediate rewards can be foregone to obtain more total rewards over the relevant time horizon. We began by examining the explore-exploit trade-off in a stationary 2-armed bandit task, in which both bandits paid-off with the same fixed reward. The bandits varied, however, in the fraction of times they delivered a reward if chosen. In this case the explore-exploit trade-off affects the first few choices, before both targets have been sampled a few times. We modeled this as a finite state, finite horizon, undiscounted POMDP, where the information states were the number of times each bandit was chosen, Ci and the number of times each bandit was rewarded, Ri. This information state space is formed by the sufficient statistics for the two bandit processes. Transitions through the information state space occur after each choice and its associated outcome and they correspond to belief updates for the process.
To examine the information state space for the bandit task more quantitatively, we can examine the distributions over expected future reward values generated in the task. Each of the bandit options is represented by a tree of possible outcomes (Fig. 1A). Each node in the tree defines the information state (i.e. Ri,Ci) for that option. The information state can be used to estimate the underlying reward probability, q for each bandit option, where q is the hidden state of the system. As one of the options is sampled, the tree is traversed. With a binomial likelihood function and a beta(α,β) prior, the posterior over reward probability is given by
pqri,ci∝pri,ciqpq
p(q|ri,ci)∝qri(1−q)ci−riqα−1(1−q)β−1
p(q|ri,ci)∝qri+α-11-qci-ri+β-1
The beta prior is the natural conjugate prior for the binomial likelihood function. Therefore, the prior can be interpreted as data. The posterior expected value is q|Ri,Ci = RiCi = α+riα+β+ci, where we have defined the actual choices and rewards as ri,ci, and the posterior choices and rewards as the data plus the prior Ri = ri + α,Ci = ci +α + β. If we start with a beta(α = 1,β = 1) prior we have posterior values of Ri = 1,Ci = 2 for each bandit arm before any options have been sampled (Fig. 1A). The possible posterior expected values are given by the nodes of the tree (Fig. 1A). These nodes are also the immediate expected value for a choice, i.e. <rst,a> = RiCi, and these values also define the transition probabilities. Thus, if one is in state, Ri,Ci one transitions to Ri+1,Ci+1 with probability pj = Ri+1,Ci+1st = Ri, Ci,a = RiCi and one transitions to Ri,Ci+1 with probability pj = Ri,Ci+1st = Ri, Ci,a = 1-RiCi. This defines two of the terms on the r.h.s. of equation 2 (ignoring the cost to sample). The other term on the r.h.s. of equation 2 is the utility of the next state, ut+1. These utilities are recursively related to future utilities, ut+2, etc. However, in the final trial, assuming a task where there are a finite number of trials and the number is known a-priori, there is no FEV because there will be no choices in the following trial. Therefore, utilities in the final trial, t = N, are given by the IEV, <rt(st,a)>. The IEV for each state that can exist in the final trial can be directly calculated from these information states. Once these are calculated, one can calculate the utilities for t-1, and continue backwards until the utilities for the current trial can be calculated. This is the backwards induction algorithm ([28]; see methods).
When an option has not been sampled, any point in the tree can potentially be reached, although not under the optimal policy, and the distribution over reward probabilities is broad. This tree, therefore, represents the possible outcomes if one of the options is chosen repeatedly (Fig. 1A). The state space for the task is, however, the product space over the nodes of two of these trees (Fig. 1B), as it is constructed of all combinations of possible outcomes from each individual tree. When the FEV is calculated for one of the options, it is only calculated across the nodes in the full tree that are visited by the optimal policy. This is because the max operator in equation 1 is an expectation over the policy that optimizes choices in each state. Thus, when the FEV is calculated the expectation is taken over the portion of the product space (Fig. 1B) where the expected action value of an option is greater than the other option (thick lines in Fig. 1B). The expectation is not computed over the dotted lines (Fig. 1B) because an optimal policy does not choose these actions. If we examine the distribution of reward probabilities over a representative finite horizon (Fig. 1C) we see that options which have been sampled less have higher expected values, when IEVs are each 0.5. In this example it is less likely that one will encounter a reward probability (q) of 0.5 for options that have been sampled less, and more likely that one will encounter options that have a reward probability greater than 0.8. This increased mass over higher reward probability nodes in the tree drives exploration in bandit tasks.
As an example, we examined a scenario in which bandit option 1 was sampled 6 times, and rewarded three times (Fig. 2A). (Note that in this example the agent is not following the optimal policy. Rather we have defined choices and outcomes to illustrate action values under particular scenarios.) The action value for option 1 exceeds the action value for option 2 during the first three trials while it is being rewarded. The FEV, however, of option 2 is larger than the FEV of option 1, even in the first 3 trials, during which option 1 is being rewarded. After option 1 is not rewarded once, it becomes more valuable to sample option 2 (i.e. Q(s,2) > Q(s,1) in trial 5). After option 1 had been sampled 6 times and rewarded three times, its IEV is the same as option 2, which had an expected value of 0.5 because of its prior. However, the action value (IEV + FEV) favors option 2 at this point (i.e. trial 7). If option 2 is then sampled 6 times and rewarded 3 times, the action values of the two options are again the same (i.e. trial 13). The exploration bonus (here taken as the difference in FEV between the two options on trial 4) is also larger when the time horizon is longer (Fig. 2B). This is because option 2 can be exploited for a longer time horizon if it is sampled and found to be better. When the first option chosen is rewarded, and it continues to be chosen and rewarded, the action value of the second option will not exceed the value of the first option (Fig. 2C), given these finite time horizons.
The exploration bonus is driven by three factors. Continuing on the example above, assume option 1 has been sampled and option 2 has not been sampled. First, there is uncertainty about option 2 (i.e. the prior distribution over possible reward probabilities for unsampled options is broad, assuming a vague prior). Therefore, option 2 might be better than option 1. If option 1 cannot be better than option 2, because of the structure of prior knowledge, there is no exploration bonus. The second factor, as shown above (Fig. 2B) is the time horizon [17]. If the time horizon is too short one cannot obtain enough additional rewards when option 2 is found to be better than option 1, to make up for the scenarios (i.e. other episodes of the task) when option 2 is found to be not as good as option 1. This factor relies on the assumption that option 2 might be better than option 1. Third, if option 2 is sampled and it is not as good as option 1, then one can switch back to option 1. On the other hand, if option 2 is better than option 1, then one can stick with option 2. This preference for the option which will be found to be better in the future, drives choices in the present via the max operator over action values in the utility equation (equation 1), which operates on the distribution of future outcomes via the embedded recursion.
We next examined a novelty task [5,8,29]. This is a 3-armed bandit task similar in several ways to the 2-armed bandit task described above. The size of the reward is the same for each bandit option, but the probability of receiving a reward when each option is selected differs. In addition to this, however, choice options are replaced by novel choice options at stochastic intervals. Thus, after subjects accumulate experience with the current set of 3 bandit options for a period of time, one of the options is replaced by a novel option. These replacements are stochastic and not known in advance, but they are indicated to the subject. We modeled this task with an infinite horizon, finite state, discounted POMDP. Consistent with the 2-armed bandit, the information state is defined by Ri,Ci for each option. The full information state is now a product space across 3 trees (Fig. 1A), so it is larger.
To examine this task we considered a scenario similar to the one examined for the 2-armed stationary bandit. The action value of the chosen option (option 1) increased while it was being rewarded in trials 1–3 (Fig. 3A, for the choices and rewards see Fig. 3D; note that these actions are not chosen by the optimal policy. Rather they were chosen to illustrate the effect of experience with an option). The FEV also increased for all 3 options because of the overall increase in the expected reward in the environment (Fig. 3C). However, similar to what was seen in the 2-armed bandit (Fig. 2A), the FEV was larger for unexplored options (Fig. 3B). Further, when option 1 was replaced, after each of the options had been chosen a few times, its FEV increased relative to the other two options (Fig. 3B, trial 15). Similarly, when option two was replaced on trial 20, its FEV increased (Fig. 3B). As with the 2-armed bandit, when the discount parameter was increased towards 1 (Fig. 3E), the exploration bonus increased (Fig. 3F). Thus, when a long time-horizon is available to exploit a novel option if it is found to be more valuable, the FEV for exploring that option increases. Every time a novel option is introduced, it is equivalent to resetting that option to the beta(1,1) prior, resetting it to the start of the tree (Fig. 1A). Thus, uncertainty drives an exploration bonus as long as a sufficient time horizon is available to exploit the novel option if it turns out to be better than the alternative options available. Correspondingly, the substitution rate of novel options also affects the novelty bonus, by effectively limiting the time horizon (Fig. 3F). If the substitution rate is high, one likely will have less time to exploit novel options that turn out to be good, before they are again replaced.
To examine exploration in related bandit tasks, we used an infinite horizon, discounted, continuous state, POMDP to model a non-stationary two-armed bandit task [9]. The information state in this model is given by the mean and variance of the bandits, which are the sufficient statistics for the two processes. The bandits in this task returned continuous valued rewards (e.g. 0–100). The means of the returned values for each bandit were non-stationary in time, following independent, random walks that decayed to 50. The actual reward earned on an individual trial was given by a sample from a Gaussian distribution with the current mean, and a standard deviation of 4. The IEV is given by the estimated mean of each bandit. The utility depends on the estimated means of the two options (Fig. 4A) as well as the estimated variance of the options (Fig. 4B). The effect of variance on utility also depends on the time-horizon (Fig. 4B). The variance has a larger effect when the time horizon is longer. The effect of the variance of the utility can be understood in the framework developed above for the stationary bandit (Fig. 1). Specifically, when an option is not sampled its variance grows because of the nonstationarity of the underlying generative model, effectively driving it backwards in the tree (Fig. 1A). On the other hand, when an option is sampled its variance decreases, effectively driving it forwards in the tree (Fig. 1A). Thus, an option which has not been sampled for several trials becomes similar to a novel option, and it should be explored.
We examined the choice sequence of the algorithm for some examples. If we consider an artificial case where the means are locked at 45 and 55 (but the algorithm still assumes the means are non-stationary), and compare the sampling under two different discount parameters (effective time horizons) we see that the algorithm periodically samples the option with a smaller estimated mean, as its variance grows (Fig. 4C). In addition, when the discount parameter is larger (γ = 0.90 vs γ = 0.99) the algorithm samples more often, consistent with the larger difference in utility for a given standard deviation for larger discount parameters (Fig. 4B). This can also be seen clearly in the action values (Fig. 4E and F—Note that the algorithm stochastically sampled option 1 first in panel E and option 2 first in panel F, which gives rise to the initial downward vs. upward fluctuation). With the means fixed, the action values depend only on the variance of the two processes, if we ignore the decay of the process to 50. When an option is sampled its variance decreases and its utility decreases, and when an option is not sampled its variance increases and its utility increases. The combination of these eventually drives the action value of the recently unsampled option to exceed the action value of the option currently being sampled (Fig. 4E and F), and the option which has not been recently sampled is then sampled. This can be seen in example sequences drawn from the actual generative process as well (Fig. 4G-H, the actual process values are identical for these two examples). In this case when the algorithm is modeled with a longer time horizon it samples more (Fig. 4H).
We next examined an information sampling task, often referred to as the beads or urn task [4,5,20,21,30]. In this task subjects are shown a sequence of beads drawn from one of two possible urns (Fig. 5A). One of the urns has q orange beads and 1-q blue beads and the other has q blue beads and 1-q orange beads. After each bead is drawn subjects have three choices. They can either draw another bead from the urn, guess that beads are being drawn from the predominantly blue urn, or guess that beads are being drawn from the predominantly orange urn. Sampling another bead usually involves an explicit cost-to-sample. In other words, subjects are charged for collecting more information. In this task, the value of choosing an urn is given by the IEV, because no more samples are allowed after an urn is chosen so FEV is zero, whereas the value of sampling another bead is given by the FEV (minus the cost-to-sample), because there is no reward if one does not try to infer the urn. Thus, this task explicitly sets up a trade-off between immediate and future expected rewards, and in this sense it is similar to the explore-exploit trade-off in bandit tasks.
In most cases subjects are told that they can draw only up to a maximum number of beads and after the last bead is drawn they have to guess an urn. As such, the task can be modeled as a finite horizon, finite state, undiscounted, POMDP. The information state space is simpler than the state space in the bandit task, as it is given by a single tree (Fig. 5B), where instead of rewards and no rewards, the state is given by the number of blue (or orange) beads that have been drawn, and the total number of beads drawn. These form the sufficient statistics for the process. As one draws beads, one works through the state space, similar to the situation with the bandit tasks. For example, the first 3 bead draws for the example sequence shown in Fig. 5C would go through the set of states shown (Fig. 5B). Unlike the bandit task, this task was modeled with an uninformative prior on bead draws, because it is normally implemented by showing subjects one bead before asking them to decide [21]. The action values for guessing either of the urns or sampling again show that the value of guessing an urn increases as evidence for that urn increases (i.e. more beads drawn of the corresponding color), and decreases as evidence for the urn decreases, in a cumulative fashion (Fig. 5C-F). The value of sampling again is initially above the value of guessing an urn, but at some point it drops slightly below. Note that without a cost-to-sample (C(st,a) = -0.005 in panels C-E and C(st,a) = -0.025 in panel F) it is always best to sample all of the available beads. To examine the effect of the cost-to-sample, we calculated values for two costs, on identical sequences of bead draws (Fig. 5E-F). When the cost was lower (C(st,a) = -0.005; Fig. 5E), it was optimal to delay the decision until after the 11th bead was drawn, whereas when the cost was higher (C(st,a) = -0.025; Fig. 5F) it was optimal to decide after 2 beads. This task can be considered a pure exploration task: how long does one explore before committing to (exploiting) one of the choices? This is similar to exploring a novel option for several trials, while always considering whether to switch back to the known option, or sticking with the novel option. As the certainty about which urn is being drawn from increases, picking an urn (which will deliver an IEV), as opposed to drawing again (which is valuable because of the FEV), becomes more valuable.
The final tasks we considered were foraging tasks. Much like the tasks examined above, these tasks trade-off immediate and future expected values. Should one stay in the current patch whose resources are being depleted (i.e. choose IEV) or travel to a new patch (i.e. choose FEV) [19]? Or, should one sample again (i.e. choose FEV) or commit to the current gamble on offer (i.e. choose IEV) [18]? The state spaces for these tasks differ in a fundamental way from the state spaces in the bandit and information sampling tasks (Fig. 6A and 7A). The state spaces for the foraging tasks are recursive. Stated another way, the state spaces for the foraging tasks do not represent learning or information accumulation. Learning or information accumulation are not recursive because you do not return to the same state (technically, this is not completely accurate, as one can with some probability, return to a previous state in either the non-stationary bandit or the novelty bandit). Rather, in the foraging tasks the current state is provided to the animal and the animal does not have to estimate beliefs or distributions over states. Therefore these tasks are MDPs, as opposed to POMDPs where the state is hidden. In the foraging tasks one observes the state directly.
In the patch leaving time task the subjects chose between staying in the current patch or traveling to a new patch [19] in each trial. The state relevant to choices is defined by the current amount of juice and the travel delay. If they stay in the current patch, they receive a (slightly delayed) reward, and the amount of reward that they will receive in the next trial if they again choose to stay in the current patch is decreased. If they choose to leave the current patch they have to wait for a known travel delay and they receive no immediate reward. The amount of reward that will be received in the new patch is reset to a fixed level and the travel time to the next new patch is sampled from the distribution of possible travel times (Fig. 6A). The patch leaving time task was modeled as an infinite horizon, discounted MDP. The relevant state variables when a decision is made are given by the current travel delay and the current reward estimate (Fig. 6A). From the model one can calculate the difference in action value for staying in the patch vs. leaving for another patch (Fig. 6B and 6C). It can be seen that the longer the travel time, the longer one stays in the patch (Fig. 6D), consistent with what was shown previously with heuristic models [19]. However, this effect only occurs for discount parameters less than 1. The undiscounted model is insensitive to finite travel times (Fig. 6E and 6F). This is because undiscounted infinite horizon MDPs are insensitive to finite time delays. Stated another way, if K is the mean first passage time to a state st = j and from state j one follows the optimal policy, then with an infinite horizon the value function can be written [26]:
vNπs = limN = ∞1N∑t = 1K-1rst,asπ+limN = ∞1N∑t = KN-1rst,asπ
From this it can be seen that actions taken prior to entering state j, at time K, do not matter. This is because the first sum is finite if the rewards are finite, and so it goes to zero in the limit. In the foraging task, if K is the time to get to the inter-trial interval (ITI) after choosing to travel, it doesn’t matter how long K is for finite K.
The final task is a variant on standard foraging tasks. The state for this task is given by the current gamble pair on offer and the state space includes all the possible gamble pairs. In most foraging tasks a decreasing marginal reward in the current patch eventually drives the action value to leave the patch above the action value to stay in the patch, because leaving has a fixed expected value. This task, however, used a paradigm in which one samples, in each trial, two gambles from a set of six possible individual gambles (Fig. 7A). The six individual gambles from which the pairs were drawn were shown for the current foraging bout and their reward values were known (e.g. gamble 1 may have had a value of 12 points). In each round, a pair of gambles from the set of individual gambles was sampled (15 possible pairs assuming sampling without replacement from the 6, and symmetry of gambles). For example, if the gambles for a given session were g1 … g6, a subject might be shown in a single trial g3 and g5. They then have to decide whether to engage with that offer pair, or sample again. If they sampled again, a new pair was drawn from the current set of six possible gambles (perhaps g2 and g3). Every time the subjects sampled again they also incurred a cost-to-sample. (Note that a cost-to-sample is paid at the time of sampling, and it does not decrease the value of future gambles, in an MDP.) If they decided to accept the offer, they moved to a decision stage. In this stage the probability that the reward associated with each gamble would be delivered was revealed, and this probability was randomly assigned to each gamble every time the decision stage was entered. The subjects had to choose one of the two gambles in the decision stage based on its magnitude and the associated probability. For example they might be choosing between p1g1 and p4g4 where pi is the probability that the subjects will receive reward gi if they choose that gamble in the decision stage. The agent then selects the gamble that has the highest expected value.
The value of sampling again is given by the FEV. The FEV is not equal to the average values of the individual gambles. The FEV is the expected value of future draws (see methods), plus the cost-to-sample. The time horizon is long, and many future samples could be drawn. However, the cost-to-sample decreases the value of future samples linearly with time, when viewed from the present decision. This can be compared to exponential discounting which exponentially decreases the value of future samples. With a sufficient time horizon the FEV is fixed. The task provided no explicit time horizon so we modeled it as a finite (although long) time horizon MDP. Therefore one simply samples until the IEV of the offered pair exceeds the (constant) FEV (Fig. 7B). It is important to point out that sampling more in this foraging task, unlike the beads task, does not improve the IEV. In other words, the IEV does not necessarily increase with samples, although one can sample a pair with a better IEV. This is related to the state space of the problem. Additionally, without a cost-to-sample, the optimum strategy would be to sample until the pair with the highest value is drawn. The cost-to-sample creates a situation where choice of a gamble pair that is not the largest is optimal, because it may cost too much to obtain a better pair.
We have applied markov decision process models (MDPs/POMDPs) to choice tasks that have been used to study the explore-exploit tradeoff, information sampling and foraging. The models allowed us to determine the normative choice mechanisms in these tasks, and therefore they provide insight into their similarities and differences. All of the tasks manipulate trade-offs between immediate and future expected values (IEVs vs. FEVs). Temporal-discounting tasks explicitly manipulate this trade-off. We have not considered them in the present study, but we have modeled them previously using MDPs [5]. Normative exploration in the bandit tasks can be defined as the choice of an option whose IEV is smaller but whose FEV is larger than an option with which one has more experience. This also drives exploration when there is non-stationarity, or when novel options are presented, because both of these increase uncertainty. In information sampling tasks an explicit trade-off is setup between sampling again, the action value for which is driven by the FEV (plus a small cost-to-sample), vs. inferring an option, the action value for which is driven by the IEV. Finally, foraging tasks also present the option of staying in the current patch, which has a larger IEV but a decreasing FEV, or traveling to a new patch, an action for which the IEV is zero but the FEV is larger. Across these tasks the IEVs and FEVs are calculated in different ways. In other words, the mechanisms that underlie value estimation differ across tasks. When trying to understand the neural circuits and underlying neural processes that carry out decisions in these tasks, it will be important to understand how these computations are carried out in the brain.
We began by examining the explore exploit trade-off in a two-armed bandit task, in which the reward amount for both options was the same, but they differed on the fraction of times that they were rewarded. Bandit tasks have been used to study learning in healthy and clinical populations [31,32]. In the first few trials there is value to sampling both options, and unsampled options have a larger FEV. This future expected value depends on three factors. First, the distribution over possible reward probabilities for the unsampled option is broad, given by the prior. Thus, the unsampled option may be more rewarding than the options which have been sampled. Second, if the unsampled option is sampled, and it is not as good as the other options, the subject can switch back to the other options. However, if the (previously) unsampled option is better than the other options, the subject can stick with it. Finally, the time horizon must be long enough to reap the rewards of investing samples in the novel option.
Heuristically, one could consider the following approximate example. Assume that one has sampled one of two available options (call it option 1) and found that it is being rewarded 70% of the time, and that one now has 100 more trials. One could then sample the alternative option (option 2) 10 times. If option 2 is rewarded 80% of the time one could then stick with that option, gaining on average 80 rewards over the 100 trial horizon. If it is found that option 2 is only rewarded 20% of the time, then one could switch back to option 1, gaining 0.2*10 + 0.7*90 = 65 rewards on average. If option 2’s (i.e. all option 2’s that one encounters, in repeated plays of the task) are either rewarded 80% of the time, or 20% of the time, the average reward with this simplified strategy will be 72.5 over the 100 trials, whereas it would only be 70 if one always stuck with option 1. The 2.5 additional rewards on average is the exploration bonus. It depends on the possibility that the novel option is better than the current option, the fact that one will switch back to the alternative if it is better than the novel option, and having a sufficient time horizon.
We also examined two other tasks which are extensions of the bandit task. Specifically, a non-stationary bandit [9], and a novelty task [5,8,29,33]. In the non-stationary bandit the mean reward magnitudes of the two options follow independent random walks. When an option is sampled several times, a relatively accurate (i.e. low variance) estimate of its mean can be derived. However, when an option is not sampled, the distribution of its mean becomes broad. When a random walk is not observed for a period of time, the variance of its estimate grows linearly with time. One way to conceptualize this, relative to the stationary bandit, is to say that when an option has not been sampled for some time, it becomes like a new option, and there is value in exploring it. This is true of any tasks that have an underlying non-stationarity in the reward [34]. It is, however, variance in the estimate of the mean that drives the exploration bonus. When the variance gets large, the option might be better than the current options, and exploration is advantageous. Similarly in the novelty task, when a novel option is substituted for one of the options that has been sampled, the reward probability for the novel option is unknown, and therefore it is valuable to explore it.
Next we examined the beads or urn task [4,5,20–22,30,35]. This is an information sampling task, similar in structure to other sampling tasks [36]. The POMDP model for this task only optimized choices in single trials with an explicit cost-to-sample. It did not optimize reward rates over multiple trials. Subjects are given the option to sample as much information as they would like, before guessing an urn. The choice to sample rests on the belief that the FEV of sampling is greater than the IEV of guessing an urn. In this respect, sampling is similar to exploring, as it is a choice in favor of the FEV, relative to the IEV. It differs from exploration, however, in that exploration in bandit tasks usually has some IEV. That is, choice of the unknown option in bandit tasks usually leads to some reward. This does not need to be true in general. In information sampling tasks, however, choosing to sample usually leads to zero IEV (or a slightly negative IEV, given by the cost to sample). In this way, sampling is more similar to foraging. It is also worth pointing out that reaction time versions of perceptual inference tasks can be modeled within a framework that is equivalent to the approach used here to model information sampling [37,38]. Perceptual inference tasks, as well as many other choice tasks, are often modeled using a drift-diffusion framework, and it is assumed that when an evidence bearing particle crosses a threshold a decision is made. The “threshold” crossing is a choice to stop sampling. It is often inferred for drift diffusion models in perceptual inference tasks, on the basis of behavioral reaction times. But with an MDP the threshold can be calculated dynamically, on the basis of current levels of belief, costs-to-sample, and transition probabilities [38]. Thus, an optimal threshold can be inferred for any tractable task.
The final tasks we considered were foraging tasks. These tasks also trade-off immediate vs. future expected values. The choice to forage leads to a zero IEV. The action value of choosing to forage is entirely an FEV. Foraging tasks differ from the tasks considered above, because their state spaces have a recursive structure and the state is observed, not inferred from information bearing observations. The tasks loop through their recursive state spaces over and over again. The choice is defined as a comparison between current and future stochastic offers. The current offer can be to stay in the patch and collect an approximately known, decreasing reward, or take the current pair of gambles that have been offered. The future stochastic offer can be explicitly calculated from the information given. It is either the value of a new patch, given the current travel time, or the expected value of the decision stage for the set of gambles that can be drawn from the current set. These average values are fixed with a sufficient time horizon. Therefore the strategy is to either stay in the current patch until the reward value drops below the value of leaving, or to sample gambles until the sampled gamble is worth more than the expected value of future samples. In foraging tasks there is generally no updating of distribution estimates, and therefore foraging differs fundamentally from exploration, in this respect.
There are two important factors that drive choice preferences across these tasks. The first is uncertainty, and the second is the time horizon. Uncertainty affects these models in two ways. First, in bandit tasks when novel options are available, or equivalently when non-stationary options have not been explored for some time, the distribution of possible reward values is broad and uncertainty is high. Therefore, sampling the options a few times to learn about them is valuable, given a sufficient time horizon. The value of this uncertainty is driven by the future expected value. The sampling itself, however, decreases uncertainty about the options. When one learns that an option either is, or is not valuable, then one can act accordingly. Thus, increased uncertainty drives value through the FEV. Because we used models that maximize expected reward, uncertainty does not affect IEV. However, as uncertainty can lead to a larger FEV, decreasing uncertainty and therefore decreasing FEV increases the relative importance of IEV on the total action value. The same reasoning applies to the information sampling tasks. As long as uncertainty is high, the FEV is high. When uncertainty is decreased, however, the IEV of guessing an urn becomes larger. Interestingly, and in contrast to this, increasing uncertainty in temporal-discounting tasks actually decreases preference for delayed, larger rewards [5]. (Temporal-discounting tasks are tasks in which subjects are offered a choice between an immediate smaller reward and a delayed larger reward.) This is because of the state space of temporal-discounting tasks. One can model temporal discounting tasks using an MDP which, at each time step, includes the possibility of exiting the path to the reward and terminating in a state with no reward, with some probability. If this probability of terminating in a no reward state increases, it becomes less likely that one will get to the reward, for a fixed delay to the reward. Interestingly, this is thought to be a fundamental factor that drives crime [39].
Time horizon is also important. In infinite horizon problems the time horizon is controlled by the discount parameter. In bandit problems, the time horizon affects the relative value of exploration. In stationary problems the time-horizon affects the relative value of exploring novel or unknown options. Longer time-horizons, or discount parameters closer to 1, increase the value of exploration. In non-stationary environments this relationship is more complex, as the non-stationarity limits the effective time-horizon of any policy. In foraging tasks, however, time horizon is also important. In undiscounted infinite horizon problems, travel times are irrelevant. If one has an infinite, undiscounted time horizon, then any finite travel time does not affect value. In the non-stationary bandit task, when the discount parameter approaches one, the algorithm samples options with lower means more often. As another example, consider the simplified MDP shown in Fig. 8. The undiscounted, infinite horizon solution to this problem does not favor action 1 over action 2 [40] because the relative value of this initial transient reward will be zero in the infinite time limit. Methods such as sensitive discount optimality exist to deal with such situations, although these can only be applied to tractable state spaces [40]. However, a discounted MDP favors action 1, in this case. This suggests that temporal discounting, in some form, is ubiquitous because it is always biologically (or computationally) relevant. Whether discounting is specifically exponential or hyperbolic, or takes on some other form is less the issue. More important is some sort of monotonic decrease in the value of future rewards with distance into the future.
The explore-exploit trade-off is often modeled with heuristics. A strong criticism of heuristics is that they explain no more than they assume, and tell us no more than the data does [28]. Heuristics, however, can provide reasonable solutions to engineering problems, often provide insight into patterns in the data, and may better approximate behavior than normative models [15]. For example, recent work has explicitly examined the role of noisy vs. directed exploration, and found that human subjects use both directed and noisy exploration strategies [31]. In some cases, however, heuristics can be difficult to interpret. For example, the beta or inverse temperature parameter in delta-rule reinforcement learning (DRRL) is often thought to control the “explore-exploit” trade-off. This parameter can only control noise in choice processes, however, and standard implementations of DRRL do not turn this noise down as reward values are learned. Therefore, exploration cannot be differentiated from noise in the choice process using this parameter and poor learning looks like exploration. Several more sophisticated variants including Thompson sampling [41,42] and related algorithms [43], however, decrease exploration with learning and can achieve minimal regret.
In an MDP framework exploration need not be undirected or noisy. Exploration can be an intentional, directed, normative strategy if there is sufficient knowledge of the environment and the agent has sufficient computational resources. One does not necessarily explore so much as one learns or accumulates information (bandit or information sampling tasks) until the additional information indicates that an alternative choice is better. Every choice delivers some information, because one is always transitioning through states as choices deliver information in these tasks. Equivalently, leaving the current patch in a foraging task is an explicit calculation of the relative value of traveling to a new patch, the expected value of which is characterized by some probability distribution over patch values. It is possible that animals have relatively unsophisticated strategies for dealing with these issues. It seems likely, however, that they have developed at least a good approximation to the underlying normative utilities, at least in tasks that match the animal’s ecological nitch or on which the animals have extensive experience.
We modeled the tasks using markov decision processes with either observable (MDP) or partially observable (POMDP) states. Tasks were modeled as finite or infinite horizon, discrete time, and discounted (i.e. with a discount parameter γ < 1) or undiscounted (i.e. with a discount parameter γ = 1) as indicated in the manuscript. Some models also included a cost-to-sample. For discrete state models the utility, u, of a state, s, at time t is
ut(st) = maxa∈Ast{ r(st,a)+C(st,a)+γ∑j∈Sp(j|st,a)ut+1(j) }
where Ast is the set of available actions in state s at time t, r(st,α) is the reward that will be obtained in state s at time t if action a is taken. The variable C(st,α) is a cost-to-sample, which may be zero. The summation on j is taken over the set of possible subsequent states, S at time t+1. It is the expected future utility, taken across the transition probability distribution p(j|st,α). The transition probability is the probability of transitioning into each state j from the current state, st if one takes action a. The γ term represents a discount factor. The terms inside the curly brackets are the action value, Q(st,α) = r(st,α) + C(st,α) + γΣj∈sp(j|st,α)ut+1(j), for each available action. For continuous state models the utility is
utst = maxa∈Astrst,a+C(st,a)+γ∫Sp(j|st,a)ut+1(j)dj
All state integrals over continuous states were calculated with discrete approximations. Equations 1 and 2 assume a reward maximizing agent, through the max operator.
For discrete state, finite horizon models with tractable state spaces, we used the backward induction algorithm to calculate utilities and action values [28]. This was done for the 2-armed stationary bandit, beads and sampling foraging tasks. With a finite horizon the final state delivers a reward, but no further actions are possible. Therefore, if we start by defining the utilities of the final states, we can work backwards and define the utilities of all previous states. Specifically, the algorithm proceeds as follows [40], where N is the final state.
1. Set t = N
uN(sN)=r(sN) for all sN ϵ N.
2. Substitute t-1 for t and compute
utst = maxa∈Astrst,a+C(st,a) +γ∑j∈Sp(j|st,a)ut+1(j)
Set
Ast,t* = argmaxa∈Astrst,a+C(st,a) +γ∑j∈Sp(j|st,a)ut+1(j)
3. If t = 1 stop, otherwise return to 2.
The non-stationary 2-armed bandit, novelty and patch-leaving foraging tasks were modeled as infinite horizon POMDPs or MDPs. The utilities were fit using the value iteration algorithm [40]. This algorithm proceeds as follows. First, the vector of utilities across states, v0, was initialized to random values. We set the iteration index, n = 0. Then computed:
vn+1 = maxa∈Ast{ r(s,a)+γ∑j∈Sp(j|st,a)vn(j) }.
(3)
After each iteration we calculated the change in the value estimate, Δv = vn+1-vn, and examined either ||Δv||<∈ or span(Δv) <∈. The span is defined as.span(v) = maxs∈S v(s)—mins∈S v(s) For infinite horizon undiscounted models the value continues to grow with iterations of equation 3, but the spans converge [40]. This is because the final state values are the average costs per stage plus a differential. This only applied to the foraging patch leaving example with discount parameter equal to 1. We only examined differential values, in that case, so the average cost per stage is subtracted out, because it is added to all states.
We also used approximate methods for the non-stationary 2-armed bandit and the novelty task, as their state spaces were intractable over relevant time horizons. For these POMDPs we defined a basis, and then approximated the utility with
v^(s)=∑i=1Maiϕi(s).
(4)
In all cases we used fixed basis functions so we could calculate the basis coefficients, ai using least squares techniques. We assembled a matrix Φi,j = ϕi(sj), which contains the values of the basis functions for specific states, sj. We then calculated a projection matrix
H = Φ(Φ'Φ)-1Φ'
(5)
And calculated the approximation
v^ = Hv.
(6)
The bold indicates the vector over states, or the sampled states at which we computed the approximation. When using the approximation in the value iteration algorithm, we first compute the approximation, v^. We then plug the approximation into the right hand side of equation 3, vn+1 = maxa∈Astr(s,a)+γ∑j∈Sp(j|st,a)v^n(j). We then calculate approximations to the new values v^n+1 = Hvn+1. This is repeated until convergence.
For basis functions we used piece-wise polynomials and/or b-splines [44]. For b-splines see [44]. For piecewise polynomials, the first basis functions are given by hi(x) = xi-1. For an order K spline (i.e. for cubic K = 3), i goes from 1 to K+1. In addition to these global polynomials, we also add hj(x) = (x-tj)K for the J knots, tj. Because all of the state spaces were multidimensional and the piece-wise polynomial basis varied between knots, we also had to compute products of the basis functions across dimensions. Computing the full tensor product basis space was usually intractable. It created a projection matrix that either could not be stored in memory or iteration over the very large projection matrix was so slow that the algorithm would not converge in a reasonable amount of time. Therefore we started with linear terms and added interaction terms of increasing order (i.e. second order, third order, …) until the approximation stopped improving. We did not find an improvement by going beyond the quadratic terms.
Knot locations were explored systematically to find locations that led to good approximations. Approximations were checked in several ways. First, we plotted vn+1 vs. v^n to see that they were consistent after convergence, as well as checking the variance of the residual. Second, we added knots to see if the fit was improved. Third, we increased the order of the polynomial to see if the fit was improved. Cubic polynomials (i.e. K = 3) were used in all cases. When the order was increased beyond cubic the value iteration often diverged. Finally, performance of the approximate inference MDP for the novelty task could be compared to a corresponding finite horizon model, at least for short time horizons to see if they made consistent predictions.
For the novelty task, the numerics were easier to implement if we approximated the number of samples for each option (N) and the probability that it was rewarded (p). We used a 3rd order B-spline basis. Knot locations for N were 0 and 150, and the algorithm was optimized at (using Matlab colon operator notation) N = e0:54:5 and p = 0: 0.25: 1. The N values were not integers, but this does not affect evaluation of the value function. Interactions up to second order were included. For the non-stationary two-armed bandit, the means were fit with a 3rd order B-spline, and the standard deviations were fit with a 2nd order piece-wise polynomial. This approach gave well-behaved value functions. The node locations for the means were given by -30 50 and 130. The means were evaluated at 0, 12.5, 25, 37.5, 50, 62.5, 75, 87.5, 100. The node locations for the standard deviations were given by 0.25, 1, 3, 5, and 15. The standard deviations were evaluated at 0.5, 1, 2, 3, 4, 5, 7 and 14. Interactions between all basis functions up to second order were included in the model.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.