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.pntd.0001336 | Vegetation and the Importance of Insecticide-Treated Target Siting for Control of Glossina fuscipes fuscipes | Control of tsetse flies using insecticide-treated targets is often hampered by vegetation re-growth and encroachment which obscures a target and renders it less effective. Potentially this is of particular concern for the newly developed small targets (0.25 high × 0.5 m wide) which show promise for cost-efficient control of Palpalis group tsetse flies. Consequently the performance of a small target was investigated for Glossina fuscipes fuscipes in Kenya, when the target was obscured following the placement of vegetation to simulate various degrees of natural bush encroachment. Catches decreased significantly only when the target was obscured by more than 80%. Even if a small target is underneath a very low overhanging bush (0.5 m above ground), the numbers of G. f. fuscipes decreased by only about 30% compared to a target in the open. We show that the efficiency of the small targets, even in small (1 m diameter) clearings, is largely uncompromised by vegetation re-growth because G. f. fuscipes readily enter between and under vegetation. The essential characteristic is that there should be some openings between vegetation.
This implies that for this important vector of HAT, and possibly other Palpalis group flies, a smaller initial clearance zone around targets can be made and longer interval between site maintenance visits is possible both of which will result in cost savings for large scale operations. We also investigated and discuss other site features e.g. large solid objects and position in relation to the water's edge in terms of the efficacy of the small targets.
| Sleeping Sickness (Human African Trypanosomiasis) is a serious threat to health and development in sub-Saharan Africa. Due to lack of vaccines and prophylactic drugs, vector control is the only method of disease prevention. Small (0.25×0.5 m) insecticide-treated targets have been shown to be cost-efficient for several Palpalis group tsetse flies, but there are concerns that they may become obscured by vegetation with a subsequent reduction in efficiency. We showed that the efficiency of the small targets was largely uncompromised by vegetation encroachment because G. f. fuscipes readily enter between and under vegetation to locate a small target, e.g. into small (1 m diameter) site clearings and underneath a very low (0.5 m) canopy. This implies that the dense vegetation, typical of the riverine habitats of Palpalis group tsetse, will not compromise the performance of tiny targets, as long as there are adequate openings of >30 cm between vegetation. Moreover, the maintanence of cleared areas around targets seems less important for the control of G. f. fuscipes with consequent savings in costs for control operations.
| The major vectors of Human African Trypanosomiasis (HAT) are in the Palpalis group tsetse flies, especially the G. fuscipes subspecies, which are responsible for transmission of >90% of reported HAT cases [1], [2]. In the present situation with limited drug and no vaccine availability, vector control remains an important addition to current efforts against HAT. Tsetse control with insecticide-treated blue/black cloth panels (c. 1–2 m wide ×1 m high), called targets [3], have been used successfully for several Morsitans group tsetse fly species, but only to a limited extent for Palpalis group tsetse [4]. Control of Palpalis group flies is costly and requires high densities of 10–30+ targets to be deployed per km2. In contrast, Morsitans group tsetse can be controlled with odour-baited targets at densities as low as 4per km2 [5], [6], [7]. It is clear from published studies that factors such as the vegetation, the coverage of the habitat achieved with deployed targets and the correct siting and maintenance of targets play a very important role in efficient control [8]. Targets or traps have to be deployed in sites which allow for the maximum number of tsetse flies available in the range of attraction to locate them. If an odour is used with the device for control of Morsitans group flies, this range is about 5–150 m plus, while an unscented target or trap has a range of about 5–30 m [9]. Limited artificial odours exist at present for Palpalis group flies [10], [11] so the trap or target's efficacy relies heavily on its visibility.
The accepted principle for identifying a suitable site for a trap or target for tsetse species is that the site has open access and visibility in most directions with no large bushes nearby and no low overhanging canopy. For example, optimal sites for the Morsitans group flies G. m. morsitans and G. pallidipes are open and well away from trees and bushes [8]. For G. austeni (also a Morsitans group fly) sites inside the shaded forest, but still ‘open’ due to a high tree canopy and little undergrowth, is best [12]. The optimal trapping sites reported for the Palpalis group fly, G. f. fuscipes, are open sites close to the water's edge [13], or an open site outside the forest but not more than 5 m away from the forest edge [14]. Optimal sites for G. tachinoides and G. p. gambiensis are on the river's edge in direct sunshine [15]. In practice the best available site in the chosen control area, or the next best potential site, will be selected and improved by cutting back vegetation and clearing undergrowth to increase visibility of the target or trap. However, the majority of sites will also include some other features such as large tree trunks, thick bushes, large rocks etc. This immediate arrangement of vegetation and solid objects around the site, i.e. the site morphology, can significantly affect tsetse catches [8]. For example, if a leafy bush with overhanging canopy grows within 1 m of a target catches of G. m. morsitans and G. pallidipes decreased by 70–80%, while if encroaching vegetation reduced the site clearing to 2 m diameter and covered about 66% of the perimeter catches also decreased by 70% [8].
Despite the importance of the Palpalis group tsetse in disease transmission there is limited information available on the effects of site morphology on target or trap efficiency for these flies, apart from the general description of what is believed to be a good site mentioned above. Understanding the impact of site morphology, especially vegetation encroachment, is imperative following the newly developed cost-efficient small targets (c. 0.125 m2) for control of five major HAT vectors namely, G. fuscipes fuscipes, G. f. quanzensis, G. f. martinii, G. palpalis gambiensis and G. tachinoides [16], [17], [18]. These small targets, as much as 8× smaller than the standard 1×1 m target and using 24× less material than the biconical trap, show great potential for economic savings in control of Palpalis group tsetse. However, the effectiveness of such small targets might be severely and rapidly compromised in the field if vegetation re-growth is as serious a problem as it is with Morsitans group flies as described above. Potentially this factor could rapidly negate the economic savings of using small targets. To address these concerns we have evaluated the performance of small targets for G. f. fuscipes in different scenarios of site morphology and vegetation encroachment as may be typically encountered in the tropical environment. The better understanding of the behaviour of G. f. fuscipes in relation to site features will contribute to effective and efficient deployment of control and monitoring devices in large scale control of G. f. fuscipes.
Studies were performed from May to December 2010 on two small islands (each c. 0.5 km2), called Big and Small Chamaunga (0° 25′ S, 34°13′ E), off Mbita point in Lake Victoria, Kenya. See [10], [16] for detailed description.
The standard sampling device was a 25×25 cm target made from blue cotton cloth with an adjacent flanking net (25×25 cm) of fine black netting. Henceforth, the term ‘target’ refers to this combination of cloth and netting. Electrocuting grids fitted in a frame covered both the cloth and netting and killed flies on impact, which then fell into trays of water below the grids. See [16] for detailed description. Experiments ran for 12 days each during the peak activity time of G. f. fuscipes, from 09:00–12:00 hours. The standard experimental design was a series of Latin-squares of treatments x days x sites, with sites at least 50 m apart. Analysis of variance was performed after transforming the daily catches (n) to log (n+1). Only detransformed catches are discussed in the text, while the transformed standard errors of the difference (SED) are provided in Tables 1 and 2. The term ‘significant’ denotes that means are different at the P<0.05 level of probability or less.
We investigated the following four aspects of site morphology and scenarios for vegetation encroachment; diagrams of the arrangements of targets and surrounding vegetation and other objects are shown in Fig. 1. All treatments were compared to a standard target without any surrounding bushes or other objects in a clearing c. 5 m in diameter.
1. Vegetation encroachment from the sides, for example when a target site is not maintained and vegetation re-growth results in: a) obstruction of the perimeter and b) decreasing the diameter of the site's clearing. In situations such as these the visibility of the small target and access for tsetse to it and around it, becomes restricted. To simulate bushes, we fixed leafy branches to stick frameworks to form hedges (Fig. 2) which we placed in various arrangements around the target (Figs. 1A–C), as described below. Similar hedges used to simulate site effects for the Morsitans group tsetse G. m. morsitans and G. pallidipes showed that there was no significant difference in the responses of tsetse to artificial bushes and real ones [8].
The first experiments studied the effect of percentage obstruction of the perimeter of a target site. The target was either completely unobstructed (100% visibility, control treatment) or (A) bushes (1.5 m long, 1 m from the target) were placed on all four sides (0% visibility,) or on two sides (50% visibility) with the hedges being placed either (B) orthogonally or (C) in parallel to the long axis of the target.
The next experiments looked at the effect of surrounding the targets with an incomplete ring of bushes as follows:
2. Vegetation encroachment from above; e.g. when a target is deployed under a tree or shrub with overhanging branches. Metal poles of appropriate length were used to support a framework of green sticks with interwoven leafy branches which formed a canopy above the target (Fig 3). Canopies were 1.5×1.5 m in diameter and 2 m, 1 m or 0.5 m above ground level (Fig. 1E, with overhead vegetation only). A subsequent experiment then investigated a combination of a canopy above a target and a bush next to it, for example when a large bush grew next to as well as over the target. The canopy was 1 m above the target and either (A) one or (B) hedges were placed orthogonally c. 0.75 from the target (Fig 1E).
3. Proximity to solid objects; e.g. large rocks which may obscure a target, or a thick tree trunk next to the target. Due to the great variety in size, colours, shapes and combinations of site morphology in nature, it is not possible to duplicate these exhaustively or change these features between sites. A partial simulation of large rocks could be achieved by placing drums horizontally on the ground, or vertically on top of each other to simulate these large objects (Fig 1, diagrams D&F). The drums were made of plastic (50 cm diameter × 80 cm high, volume = 160 L) covered with matt black cotton cloth and placed either next to, or in front of a target.
In addition, we also looked at the responses of G. f. fuscipes to a small target next to a real tree bole (a paw-paw tree bole 30 cm diameter, 1.8 m high) and whether the orientation of the target to the tree was of importance, i.e. with the blue cloth or the black netting panel closest to the bole (Fig 1G).
4. Catches of G. f. fuscipes at different distances from the water's edge. This was done because standard field procedure is to place the device close to the water's edge [19], [20] partly to increase visibility, but also because casual field observations show flies apparently move along the water's edge. A standard small target was deployed in a randomized block design in four sites. The control site was the water's edge, with the other three sites at 2 m or 4 m inland or 2 m into the water. For the latter, the target and collection tray were fixed to a floating platform of sticks and the electric cables lengthened to reach the power supply on the shore. The same set-up was repeated but using a standard Biconical trap as collection device.
Table 1 lists all the experiments on site morphology and their results.
Following the vegetation encroachment experiments, we looked at the effect on catches of large solid objects next to a small target. As described in Material and Methods, we used drums as artificial rocks and tree trunks for this study, due to the difficulty in otherwise simulating these objects in the field. Our data showed that when either the ‘rock’ or ‘tree trunk’ was placed next to the target there was no significant difference compared to catches from the control target (Table 1, experiment 7). In fact, the catches of female G. f. fuscipes increased in both cases, by 1.2× when the rock (Treatment A) was used (6.7 tsetse/day, s.e.d. = 0.1) and by 1.1× when the tree was used (Treatment B, 6.1 tsetse/day). Tsetse flies are attracted to large black objects and black drums and flat black cloth panels are used routinely in experiments to increase visual attraction [8], [16]. Therefore the observed increase in catches may be expected, but the more interesting question is what happens when such large black objects obscure the visibility of the small target, e.g. when a large rock is directly obscuring a small target. We found (Experiment 8) that the unobscured target (9.9 flies/day, s.e.d. = 0.08) caught 80% (P<0.001) more females than the target with one drum in front (Treatment A, 2.0 flies/day) and 98% more females than with a drum on each side (Treatment B, 0.03 flies/day). Catches of male G. f. fuscipes showed no significant difference (P = 0.2) between the target in the open and either of the treatments, although 20% less flies were caught with one drum in front of the target (Treatment A, 3.2 flies/day) and 80% less with a drum on each side of the target (Treatment B, 0.2 flies/day), completely obscuring the frontal views. When placing a target next a real tree trunk (Table 1, exp. 9) there was a doubling in female catches with both the blue cloth closest to the trunk (Treatment B, 2.2 flies/day) and with the netting closest to trunk (Treatment C, 2.8 flies/day) although this was not significant (P = 0.26).
Finally, we investigated the effect of the position of a small target and biconical trap in relation to the water's edge (Table 2, experiment 1 & 2). For G. f. fuscipes, the trapping sites usually used in control campaigns are open and close to, or right on the water's edge [13]. Casual field observations indicate that flies may use the water's edge as a movement ‘corridor’, perhaps due to more abundant green vegetation for shelter, higher humidity and higher chance of finding a host, particularly monitor lizards which inhabit these aquatic margins. However, our results show that a small target placed on the water's edge did not catch significantly more female flies than targets placed 2 m (7.2 flies/day, s.e.d. = 0.15), 4 m inland (6.0 flies/day), or floating 2 m into the water (5.8 flies/day). When a biconical trap was used as collection device (Table 2, experiment 2), female catches on the water's edge were slightly better (5.7 flies/day, s.e.d. = 0.2) than that at 2 m inland (2.5 flies/day), 4 m inland (4.8 flies/day) and 2 m in the water (1.6 flies/day). However, the differences were not significant. Although deployment at the water's edge may be desirable, it appears not to be essential because target efficiency does not decrease significantly over a few meters at least. This is important as it means targets can be sited to minimise losses due to flooding.
Vegetation encroachment around a small target, from the sides and above, does not significantly affect its killing efficiency for G. f. fuscipes as long as there are some openings between adjacent bushes, wider than 30 cm. These results are intriguing because the rapid re-growth potential of the tropical vegetation in the habitat of G. f. fuscipes and other Palpalis group tsetse, combined with the small size of the targets, could make it seem improbable that these targets will remain effective. Indeed, our results show that only one such scenario, grass regrowth very close to the target, poses a serious threat to their performance. Our simulation of grass height corresponds roughly to between c. 15 (15 cm high) and 60 days (60 cm high) as observed in the rainy season in the field. As expected, the small diameter clearings (0.75 m) created by the proximity of surrounding grass significantly decreased target catches. However, this represents severe and complete grass regrowth around a target, something which does not happen frequently in nature because grass rarely grows uniformly and there always remain some openings between clumps of grass to allow visibility and access to a target. In addition and as matter of routine, this scenario is easily prevented by the proper initial clearing of target sites. In some circumstances this can aided by the subsequent use of systemic herbicides such as glyphosate which can inhibit grass regrowth for several months afterwards. For example, application of glyphosate maintained reduced grass cover for up to 26 weeks on a rainforest edge [21]. Limited studies have been done on the effect of vegetation encroachment on the efficiency of a target or trap for Palpalis group tsetse species. The most relevant studies are from Morsitans group flies [8] where the effect of vegetation close to a trap dramatically and significantly reduced catches of G.m. morsitans and G. pallidipes. For example, one bush with an overhanging canopy next to a trap, decreased catches of both Morsitans group species by more than 80%, while a decrease in diameter of clearing size from 12 m to 2 m led to about 65% decrease in catches. In contrast, our data for G. f. fuscipes showed no significant difference between control and treatment catches in both scenarios, even with only 1 m diameter clearings.
The presence of a few bushes surrounding a target site, not obscuring more than about 70% visibility, may in fact be slightly beneficial. An apparently similar situation was evident with G. m. morsitans and G. pallidipes, where trap catches increased if 2–6 bushes were within 2–12 m from the target [8]. However, the smallest clearing size used for the G. m. morsitans and G. pallidipes experiments was 2 m radius (4 m diameter), at which catches of both species were 65% less than the open trap. As the clearing diameter was increased to 12 m, the catches increased. For G. f. fuscipes a remarkably small clearing of even 1 m diameter remained effective. The importance of an opening about 50 cm wide between adjacent bushes around a target was evident as catches reduced significantly (by 68% for females) if this opening was 30 cm and less. This was also found for G. m. morsitans and G. pallidipes, with catches of both species increasing significantly when the opening size is widened from 25 cm to 50 cm and more [8].
We showed that G. f. fuscipes readily enters between and through leafy vegetation to locate a small target. This behaviour corresponds with the habitat along the islands and shore of Lake Victoria, where their main hosts are monitor lizards. G. f. fuscipes have to locate these medium to small-sized reptiles between the leafy vegetation and rocks. Other site features such as large rocks or tree boles close to the target also affect catches of G. f. fuscipes, e.g. a single large solid object to the side of a small target, whether this was an artificial rock or tree bole, or a natural tree bole, actually increased catches. On the other hand, if one or more such objects obscured the frontal view of the target, catches decrease significantly. In addition, it would seem that the waters edge is not a required trap or target site for G. f. fuscipes. This is important as changes in water height can easily sweep away control devices with much cost to control programmes. The priority should be given to visibility rather than proximity to waters edge (at least within the 4 m investigated here), because target efficiency does not decrease significantly over just a few meters between the water's edge and inland.
As illustrated in this work, the small targets retain their killing efficacy in several situations of vegetation encroachment, even in small clearings of 1 m diameter and with leafy bushes close-by and above. Nevertheless, in practice, we recommend that sites be cleared to at least 2–3 m in diameter during initial deployment and that overhanging or intruding vegetation be cut back. This will allow for maximum visibility of the target during the first months after deployment. Maintenance intervals will vary between locations depending on vegetation regrowth rates, but under conditions in the study area we expect the small targets to remain efficient for 3–6 months after initial deployment, with no maintenance visits required in-between. If possible the use of a systemic herbicide applied on the site will prevent the regrowth of grass and other vegetation. The possible herbicides available for use next to watercourses are very limited; for example glyphosate is the only product registered for such use in the U.K.
The data presented here demonstrates the potential for less frequent maintenance visits to cut back and control vegetation, which is a major financial constraint in tsetse control operations [22] where targets have to be serviced regularly to maintain efficiency. Another reason for maintenance visits is to ensure that the target is still in its correct position, is upright, the cloth is in good condition and that the moving parts are free. When using large targets, this maintenance has to be carried out regularly and irrespective of whether the vegetation needs clearing. This will be largely unnecessary when using the small targets because they will be more stable and not blown over or bent by strong winds as frequently as large targets.
Clearly, there is potential for low-cost, low-maintenance control of G. f. fuscipes, and there is a necessity of these types of studies on other Palpalis group tsetse species in other tropical environments, to allow for better understanding and control of these major vectors of HAT.
|
10.1371/journal.pbio.0060068 | Functional Adaptation of a Plant Receptor- Kinase Paved the Way for the Evolution of Intracellular Root Symbioses with Bacteria | Nitrogen-fixing root nodule symbioses (RNS) occur in two major forms—Actinorhiza and legume-rhizobium symbiosis—which differ in bacterial partner, intracellular infection pattern, and morphogenesis. The phylogenetic restriction of nodulation to eurosid angiosperms indicates a common and recent evolutionary invention, but the molecular steps involved are still obscure. In legumes, at least seven genes—including the symbiosis receptor-kinase gene SYMRK—are essential for the interaction with rhizobia bacteria and for the Arbuscular Mycorrhiza (AM) symbiosis with phosphate-acquiring fungi, which is widespread in occurrence and believed to date back to the earliest land plants. We show that SYMRK is also required for Actinorhiza symbiosis of the cucurbit Datisca glomerata with actinobacteria of the genus Frankia, revealing a common genetic basis for both forms of RNS. We found that SYMRK exists in at least three different structural versions, of which the shorter forms from rice and tomato are sufficient for AM, but not for functional endosymbiosis with bacteria in the legume Lotus japonicus. Our data support the idea that SYMRK sequence evolution was involved in the recruitment of a pre-existing signalling network from AM, paving the way for the evolution of intracellular root symbioses with nitrogen-fixing bacteria.
| As an adaptation to nutrient limitations in terrestrial ecosystems, most plants form Arbuscular Mycorrhiza (AM), which is a symbiotic relationship between phosphate-delivering fungi and plant roots that dates back to the earliest land plants. More recently, a small group including the legumes and close relatives has evolved the ability to accommodate nitrogen-fixing bacteria intracellularly. The resulting symbiosis is manifested by the formation of specialized root organs, the nodules, and comes in two forms: the interaction of legumes with rhizobia, and the more widespread Actinorhiza symbiosis of mostly woody plants with Frankia bacteria. The symbiosis receptor kinase SYMRK acts in a signalling pathway that legume plants require to trigger the development of nodules and the uptake of fungi or bacteria into their root cells. Here we show that the induction of Actinorhiza nodulation also relies on SYMRK, consistent with the idea that both types of nodulation evolved by recruiting common signalling genes from the pre-existing AM program. We observed that SYMRK from different land plant lineages differs significantly in exon composition, with a “full-length” version in the nodulating clade and shorter SYMRK genes in plants outside this lineage. Only the most complete SYMRK version was fully functional in nodulation, suggesting this gene played a central role in the recruitment event associated with the evolution of intracellular root symbioses with bacteria.
| Nitrogen limits plant growth in many terrestrial ecosystems. Evolutionary adaptations to this constraint include symbiotic associations with bacteria that are capable of converting atmospheric nitrogen into ammonium. Extracellular associations of plants with diverse groups of nitrogen-fixing bacteria are phylogenetically widespread, but only a small group evolved the ability to accommodate bacteria endosymbiotically inside cell wall boundaries. Bacterial symbionts are confined within tubular structures called infection threads, which are surrounded by a host-derived membrane that is continuous with the plasma membrane, and bound by plant cell wall–like material [1,2]. The bulk of host plants including all actinorhizal species retain the bacterial symbionts within these structures during the nitrogen-fixing stage of the symbiosis [1,3]. In the most advanced forms found exclusively among legumes (Fabales) and Gunnera [4], symbiotic bacteria are delimited from the host cell cytoplasm only by a plant-derived membrane in the mature stage of the symbioses. In the respective legumes, they develop into bacteroids contained in organelle-like symbiosomes, where nitrogen fixation takes place (for a recent review, see [5]). Bacterial endosymbioses in both legumes and actinorhizal plants are typically associated with the formation of novel plant organs, so-called nodules, which are root-derived in the majority of cases [6]. Nitrogen-fixing root nodule symbiosis (RNS) occurs in two major forms. Actinorhiza hosts belong to three eurosid orders (Figure 1) and nodulate with Gram-positive actinobacteria of the genus Frankia [7]. Legumes, on the contrary, enter specific interactions with members of a diverse group of Gram-negative bacteria, termed rhizobia. For almost a century, the extreme diversity in organ structure, infection mechanisms, and bacterial symbionts among nodulating plants obscured the fact that the nodulating clade is monophyletic, which was revealed by molecular phylogeny relatively recently [8]. The restriction of endosymbiotic root nodulation to a monophyletic group of four angiosperm orders (Figure 1) is coincident with a patchy occurrence within this clade. These observations led to the hypothesis that a genetic change acquired by a common ancestor may predispose members of this lineage to evolve nodulation endosymbiosis [8].
The molecular adaptations underlying the evolution of plant-bacterial endosymbioses are still a mystery, despite a substantial biotechnological interest in understanding the genetic differences between nodulating and non-nodulating plants. While the molecular communication between legumes and rhizobia has been studied in some detail, important clues are expected from the genetic analysis of the yet underexplored Actinorhiza.
Bacterial signalling molecules and corresponding plant receptors involved in RNS are known only for the legume–rhizobium interaction. Frankia signals may be biochemically distinct from rhizobial chito-oligosaccharide nodulation factors [9], which would suggest an independent mechanism of host–symbiont recognition in Actinorhiza.
Phenotypic analysis of legume mutants has revealed a genetic link between RNS and Arbuscular Mycorrhiza (AM), which is a phosphate-scavenging association between plant roots and fungi belonging to the phylum Glomeromycota [10]. AM is widespread among land plants, where forms of AM are found in representatives of all major lineages. Fossil evidence for ancient AM-like associations [11] suggests a role of this symbiosis in the colonization of land about 450 million years ago.
The link of plant–fungal and plant–bacterial endosymbioses in legumes, which involves at least seven genes [12–16] termed “common symbiosis genes” [17], inspired the idea that during the evolution of bacterial endosymbiosis, genes were recruited from the pre-existing AM genetic program [18]. However, the molecular steps involved are not clear.
To gain insight into the evolution of nitrogen-fixing root nodulation, we analysed common symbiosis genes across angiosperm lineages with different symbiotic abilities. Many, including the calcium/calmodulin kinase gene CCaMK [14,19], or genes encoding the predicted cation channels CASTOR and POLLUX [12,20,21], are conserved in overall domain structure. We discovered exceptional diversification among genes encoding the symbiosis receptor kinase SYMRK in different species (Figure 1). While putative SYMRK kinase domains are conserved and contain characteristic sequence motifs discriminating them from related kinases (Figure S1), the predicted extracellular portion of SYMRK occurs in at least three versions of domain composition (Figure 1 and Table 1). The longest SYMRK version is present in all tested eurosids, including nodulating and non-nodulating lineages. Comprising 15 exons, it encodes three leucine-rich repeat (LRR) motifs and an extended N-terminal domain of unknown function (NEC-domain, Figure 1 and Table 1). Outside of the eurosid clade, which encompasses all nodulating groups, one or more exons are absent from SYMRK coding sequences (Figure 1 and Table 1).
Genetic evidence indicates that SYMRK acts near a point of molecular convergence of AM and legume-rhizobium signalling [16,22]. The presumed ability of its diverged extracellular domain to perceive symbiosis-related signals [16] renders it a prime target for investigating the molecular adaptations underlying the evolution of RNS.
The homogenous occurrence of “full-length” SYMRK genes among legumes, actinorhizal plants, and non-nodulating eurosids raises the intriguing possibility that SYMRK is involved in the proposed genetic predisposition [8] of this clade to evolve nodulation. An important prediction following from this hypothesis is the common requirement of a full-length SYMRK version for all types of RNS. Furthermore, also non-nodulating members of this monophyletic clade may carry nodulation-competent versions of SYMRK. To test this concept, we analysed the functional capabilities of “full-length” SYMRK genes from symbiotically diverse eurosids.
To investigate SYMRK function in Actinorhiza, we reduced root mRNA levels of D. glomerata (Datisca) SYMRK (DgSYMRK) via RNA-interference (RNAi). Quantitative PCR following reverse transcription showed a 36%–99% reduction of DgSYMRK transcript levels in knockdown roots (n = 16) compared with vector control roots (n = 16). Eight weeks after inoculation with Frankia bacteria, no nodules were detected on DgSYMRK RNAi roots (Figure 2A and 2B), except for small, primordial swellings on 16% of independent transformed roots (9/55). Nonsilenced control roots of the same plants and roots transformed with a binary vector lacking the silencing cassette (transgenic control roots) showed wild type–like nodules with lobed structure typical for Datisca (Figure 2A and 2B). This result demonstrates that SYMRK is essential for Actinorhiza development in Datisca. In conjunction with the well-documented role of legume SYMRK in the interaction with rhizobia [16,23], SYMRK thus represents a common genetic requirement for the two types of bacterial root endosymbiosis.
To test whether DgSYMRK is also required for AM, we inspected DgSYMRK RNAi roots for AM formation with the fungus Glomus intraradices (Glomus). Datisca wild-type roots of the same plants used for hairy root induction and independent transgenic control roots formed AM, with dense arbuscular colonization of inner cortical cells (Figure 2C–2F). In contrast, symbiotic development in DgSYMRK RNAi roots was strongly impaired. In 82% of independent transformed roots, no fungal infection was observed, despite the presence of extensive extraradical mycelium (Figure 2G), with those roots exhibiting strong reduction levels of DgSYMRK being nonsymbiotic concerning both nodulation and AM formation. Occasional infection attempts occurred but typically were aborted in the outer cell layers (Figure 2G and 2H). We conclude that similar to the situation in legumes, SYMRK of the actinorhizal plant Datisca is involved in both bacterial and fungal endosymbioses.
To determine whether SYMRK plays a role in the specific recognition of rhizobia by legume hosts, we tested whether L. japonicus (Lotus) SYMRK (LjSYMRK) can mediate nodulation in another legume, which interacts with a different rhizobial partner. The specific symbiont of Lotus is Mesorhizobium loti, whereas Medicago truncatula (Medicago) interacts with Sinorhizobium meliloti. Medicago dmi2 5P mutants exhibit a deletion in exon three of the SYMRK ortholog DMI2, leading to a frameshift and premature stop codon. Dmi2 5P plants form no infection threads or nodules upon inoculation with either rhizobial strain. Transgenic roots of these plants, and of wild-type control plants carrying LjSYMRK, formed infection threads and indeterminate, pink nodules typical for Medicago [24] with S. meliloti (Figure S2 and Table 2). LjSYMRK can therefore fully restore nodulation of Medicago with S. meliloti, indicating that SYMRK is not directly involved in determining legume–rhizobium specificity.
Medicago dmi2 5P mutants are also impaired in AM. No arbuscules were observed within 2 wk of co-cultivation, with fungal infection being aborted at the root surface or after entry into epidermal cells (Figure S2 and Table 2). LjSYMRK restored the AM defect in transgenic roots of this line (Figure S2 and Table 2), demonstrating that SYMRK is sufficiently similar to DMI2 to support both fungal and bacterial endosymbioses in Medicago.
To analyze the symbiotic capabilities of “full-length” eurosid SYMRK genes from a legume (MtDMI2), an actinorhizal plant (DgSYMRK), and the non-nodulating, AM-forming Tropaeolum majus (Tropaeolum; Brassicales) (TmSYMRK), we tested their potential to function in the Lotus symbiosis signalling context. We introduced these genes, under the control of the Lotus SYMRK promoter region, into roots of Lotus line SL1951–6 (symrk-10), which carries a symrk mutant allele encoding a kinase-dead SYMRK version [25,26]. Upon inoculation with Glomus, symrk-10 roots form no AM, and fungal infections are typically associated with aberrant hyphal swellings and are aborted after entry into epidermal cells (Figure 3A and 3B, and Table 2). Interaction with M. loti is blocked at an early stage, and no infection threads or nodules form (Figure 3C and 3D, and Table 2). In symrk-10 roots transformed with MtDMI2, DgSYMRK or TmSYMRK both AM and nodulation were restored, the latter involving the formation of infection threads and pink, bacteria-containing nodules (Table 2 and Figure 3) that were indistinguishable from wild-type nodules. In conclusion, consistent with a role of SYMRK in the predisposition to evolve RNS, we could not detect a functional diversification of the eurosid SYMRK version linked to features differentiating actinorhizal or legume nodulation, or to the specific recognition of bacterial symbionts. Thus, other factors, such as nod factor receptor kinases [22,27,28] or yet-unknown additional components, are likely accountable for the fine-tuning of recognition specificity in plant–bacterial endosymbioses within the eurosids.
SYMRK from the non-nodulating eudicots Papaver rhoeas (poppy) and Lycopersicon esculentum (tomato) represent intermediate length and domain composition (Figure 1 and Table 1). To explore the symbiotic capabilities of this version, we introduced the two-LRR encoding LeSYMRK genomic sequence fused to the LjSYMRK promoter into Lotus symrk-10 transgenic roots. LeSYMRK selectively restored AM symbiosis, whereas neither infection threads nor nodules developed upon inoculation with M. loti (Table 2 and Figure S3E–S3L).
A reduced SYMRK version is present in the AM-forming, non-nodulating monocots Oryza sativa (rice) and Zea maize (maize), encoding two LRRs only and a short N-terminal region, mainly covered by a single exon aligning with exon four of LjSYMRK (Figure 1 and Figure S1). To determine whether the short SYMRK version of rice, OsSYMRK, is sufficient for endosymbiosis formation in Lotus, we introduced the OsSYMRK genomic sequence controlled by the LjSYMRK promoter into symrk-10 roots. AM formation was fully restored in these roots, whereas nodulation with M. loti was impaired (Table 2 and Figure S3M–S3Z). OsSYMRK-containing symrk-10 roots inoculated with M. loti exhibited primordial swellings generally devoid of bacteria (Figure S3U–S3X). In rare cases, infection threads and small round nodules were observed, which contained bacterial colonies (Figure S3Y and S3Z). Hence, similar to LeSYMRK, OsSYMRK is compromised in supporting bacterial endosymbiosis in Lotus.
In legumes, SYMRK is indispensable for root endosymbiosis with rhizobia and AM fungi [16,23]. We show here that this endosymbiosis gene is also required for nodulation in the actinorhizal plant Datisca. SYMRK, which is likewise essential for Actinorhiza formation of the tree species Casuarina glauca (Fagales) [29], represents the first known plant gene required for Actinorhiza, indicating a shared genetic basis of the two different types of RNS. A future task will be to determine whether further endosymbiosis genes acting in concert with SYMRK in legumes are also required for Actinorhiza.
The ability of different SYMRK versions from both dicot and monocot lineages to restore AM in Lotus indicates a homologous nature of the AM genetic program in angiosperms. This is consistent with the observation that loss-of-function mutations in the rice version of the legume symbiosis gene CCaMK results in loss of AM symbiosis [30]. In Arabidopsis, the absence of root symbiotic capability is accompanied by a deletion of several symbiosis genes, including SYMRK and CCaMK [21,31,32].
Our survey of SYMRK sequences across angiosperms revealed at least three structurally distinct versions, and we could show that this polymorphism is functionally related to the root symbiotic capabilities of host plants. The variation in SYMRK domain composition is exceptional among the known common symbiosis genes. The congruence between the phylogenetic distribution of the “full-length” SYMRK version with the nodulating clade strongly suggests a link between SYMRK sequence evolution and the acquisition of endosymbiotic root nodulation with bacteria. An attractive hypothesis is that SYMRK sequence divergence was a critical step in mediating the recruitment of the otherwise conserved common symbiosis pathway from the pre-existing AM genetic program. Recruitment was proposed to account for the genetic link of AM and nodulation in legumes [17,18] and would make root–bacterial endosymbiosis as a whole a fascinating example for novel traits evolving on the basis of pre-existing genetic patterns.
A common feature associated with endosymbiotic bacterial infection in both actinorhizal [33] and legume hosts [34] is the formation of intracellular pre-infection threads (PITs) in host cells. These cytoplasmic structures resemble the pre-penetration apparatus (PPA) preceding fungal infection during AM formation [35]. Forming in anticipation of bacterial symbionts, PITs are thought to coordinate the uptake of bacteria and determine the spatial progression of infection through the host cell [33,34]. A similar role in guiding fungal transition through host cells in AM has been demonstrated for PPAs [35]. These developmental similarities in AM, Actinorhiza, and legume-rhizobium infection may reflect a common genetic program for endosymbiosis establishment and symbiont uptake in all three types of interactions. In AM, PPAs are not formed in mutants that are defective in certain common symbiosis genes [35]. It is therefore possible that a recruitment of AM symbiosis genes during the evolution of RNS facilitated the induction of intracellular accommodation structures in response to bacteria.
Repetitive LRR modules have been implicated in the determination and evolution of novel recognition specificities of receptor proteins [36–38]. Interestingly, adaptive changes reflecting positive selective constraints can be traced in LRR– and NEC–encoding regions of SYMRK genes from different Medicago species, but these do not correlate with shifts in rhizobial specificity [39]. Our functional comparison of eurosid SYMRK versions indicates that SYMRK is not involved in determining recognition specificity in nodulation. However, an extended SYMRK version containing a set of three LRR motifs, as present in eurosid SYMRK genes, is required for fully supporting nodulation symbiosis of Lotus with M. loti. Shorter SYMRK versions from tomato or rice only suffice for AM. These functional differences may be caused by individual amino acid sequence polymorphisms, or alternatively, exons that are specifically required for bacterial endosymbiosis may be lacking in rice and tomato SYMRK versions.
At an overall structural level, exon acquisition from other genes encoding LRR or NEC-like domains [23,40] or, alternatively, retainment of exons in eurosid SYMRK genes, may have been an integral genetic factor in the evolution of bacterial endosymbiosis in angiosperms. The observation of small nodule-like structures on Lotus symrk mutant roots transformed with the OsSYMRK construct is counterintuitive, considering that the LeSYMRK version, which resembles the legume version more closely, does not support such developmental responses. One possible explanation may be that the nonmatching NEC region of LeSYMRK negatively interferes with nodulation, but not AM signalling in Lotus.
The NEC domain encoded by Lotus SYMRK exons two and three, upstream of the conserved LRR flanking region (CEC), is present across eudicot plants (Figure 1). Its function outside the nodulating group is unknown. The proposed involvement of SYMRK in processes such as reduction of the touch sensitivity of root hairs [41] may rely on this domain thereby imposing selective constraints. The NEC domain shows possible overall relatedness but only a low level of similarity to sequences present in the rice genome, and to sequences other than SYMRK candidates in genomes of dicots like Arabidopsis [23]. The apparent divergence observed among these potentially homologous sequences of yet unknown function is consistent with a hypothetical role as a receptor domain.
It will be a future challenge to determine the contribution of individual SYMRK LRR units as well as of the NEC domain and to resolve at the amino acid level the features of SYMRK proteins involved in conferring endosymbiotic nodulation capacity.
The diversity and scattered occurrence of nodulation symbioses within the eurosid lineage suggest multiple independent origins [42]. Only a subset of the plant species carrying the “full-length” version of SYMRK develop root nodules, yet SYMRK of the non-nodulating Tropaeolum proved competent to support nodulation in Lotus. Hence, there must be additional genetic features distinguishing the nodulators. Candidate genes include those that express the legume LysM receptor kinases NFR1 and NFR5 [22,27,43], which are required for responsiveness to rhizobial lipo-chito-oligosaccharide nodulation factors, but not for AM formation. A potential relevance of LysM receptors in Actinorhiza, or the identity of alternative receptors perceiving yet unknown Frankia signals, remains to be determined.
We used a PCR strategy employing degenerate primers to obtain SYMRK sequence information from diverse angiosperms, for which no genome or root-derived expressed sequence tag sequences were available. Degenerate primers for the isolation of SYMRK genes were positioned in regions of the coding sequence conserved among SYMRK candidates, but not in other similar O. sativa (rice) and A. thaliana (Arabidopsis) sequences. For primer sequences, see Table S1.
λ Zap cDNA libraries were available for isolation of Ly. esculentum (tomato) and Alnus glutinosa (alder) SYMRK. A cosmid clone carrying the LeSYMRK genomic region was isolated from a pooled tomato Cf2/9 library (kind gift of J.D.G. Jones, The Sainsbury Laboratory, United Kingdom) and shotgun sequenced.
For rapid amplification of cDNA ends (RACE) reactions, total RNA was extracted from roots of uninoculated seedlings or young plants and DNaseI treated. RT and 5′/3′RACE reactions were done using the SMART RACE kit (Clontech), following nested degenerate PCR reactions ([10 s 94 °C, 10 s 52 °C, 30 s 72 °C] × 35, 5 min 72 °C) to obtain initial sequence information.
For hairy root complementation assays, SYMRK cDNAs were amplified from root cDNA preparations (Superscript II, Invitrogen) of the respective species. Binary transformation vectors were pCAMBIA 1302 or pK7WG2D,1 [44] derivatives. L. japonicus (Lotus), Me. truncatula (Medicago), D. glomerata (Datisca), and T. majus (Tropaeolum) SYMRK coding sequences were amplified from complementary DNA using primers LjSYMRK_EC_f with LjSYMRK_PK_r, MtSYMRK_EC_f with MtSYMRK_PK_r, DgSYMRK_EC_f with DgSYMRK_PK_r and TmSYMRK_EC_f with TmSYMRK_PK_r (Table S1), respectively. An MtSYMRK genomic segment containing intron one was amplified from total DNA using primers MtSYMRK_EC_f with MtSYMRK_01_r (Table S1) and ligated to the 3′ fragment of the MtSYMRK cDNA following BstB1 digestion of both. The genomic sequences of tomato and rice SYMRK were amplified from total DNA using primers LeSYMRK_EC_f with LeSYMRK_PK_r and OsSYMRK_EC_f with OsSYMRK_PK_r (Table S1), respectively. A 285-bp fragment amplified with primers polyA_NOS_f and polyA_NOS_r (Table S1) from pJawohl8 RNAi (kind gift of P. Schulze-Lefert, Max Planck Institute for Plant Breeding Research, Cologne, Germany) was used as terminater in all constructs. SYMRK genes or coding sequences were under control of 4,970 bp of genomic sequence preceding the LjSYMRK translation start site. For pK7WG2D constructs, the cauliflower mosaic virus (CaMV) 35S promoter driving the Gateway-cassette was deleted Sal1(10124)-Sal1(38).
For Datisca RNAi experiments, a pRedRoot- [45] based binary vector encoding DsRED1 for visible selection was equipped with a pKANNIBAL [46] CaMV 35S promoter-driven hairpin construct containing 367-bp of 3′ coding and untranslated sequence of DgSYMRK in forward and reverse orientation, amplified with primers DgSYMRK_RNAi_f and DgSYMRK_RNAi_r (Table S1).
L. japonicus wild-type ecotype B-129 Gifu and derived mutant line symrk-10 [25] were inoculated with M. loti R7A as described [16]. Growth conditions were 24 °C constant at 16-h-light/8-h-dark cycles. For infection of Agrobacterium rhizogenes–induced transgenic roots in Lotus, M. loti MAFF 303099 expressing DsRED was applied at a final optical density at 600 nm (OD600) of 0.02 in translucent plastic boxes containing 300 ml Seramis (Mars) and 150 ml liquid Fahraeus Plant (FP) medium [47]. Ag. rhizogenes-transformed Me. truncatula A17 wild type and dmi2 5P (kind gift of G. Oldroyd, John Innes Centre, United Kingdom) plants were inoculated with S. meliloti 1021 carrying pBHR-mRFP [48] (OD600 of 0.02) in planting pots containing Seramis, and fertilized with FP medium supplemented with 50 μM KNO3 two times per week. For nodulation assays and simultaneous observation of infection thread (IT) formation in Lotus and Medicago, plants were harvested 4 and 5 wk post inoculation, respectively. Prior to inoculation, roots showing no eGFP fluorescence were removed. For AM phenotyping of Lotus and Medicago, plants were co-cultivated with G. intraradices BEG195 and harvested after 3 or 2 wk of co-cultivation, respectively. Root systems were stained with acidic ink as described [49]. Prior to staining of Ag. rhizogenes–induced root systems, roots showing no eGFP fluorescence were removed. Roots were scored AM-positive (AM+) if symbiotic structures (arbuscules and vesicles) were present, as AM-negative (AM–) if no arbuscules were present. Occasional Lotus symrk mutant roots showing vesicles not accompanied by arbuscules were scored AM–. Where complemented Lotus symrk mutant roots exhibited aborted infection sites in co-occurrence with successful infection and colonization events involving arbuscule and vesicle formation, roots were scored AM+. Datisca seeds and Frankia inoculum were a kind gift from K. Pawlowski (Department of Botany, Stockholm University, Sweden). Datisca was inoculated with compatible Frankia by potting in substrate with ∼1 g/l crushed nodules and with G. intraradices BEG195 by adding substrate extracted from pots of inoculated Allium schoenoprasum plants. Growth conditions were 16 h light/8 h dark at 22 °C and 60% relative humidity. Seeds of T. majus and P. rhoeas were purchased at Notcutts Garden Centres (UK). The ability to develop AM with G. intraradices was confirmed for all species involved in the study.
Transgenic roots on Lotus symrk-10 mutants were induced using Ag. rhizogenes strains AR1193 [50] and LBA1334 [51] as described by Díaz et al. [52] (modified).
Medicago seedlings were transformed as described at http://www.isv.cnrs-gif.fr/embo01/manuels/index.html (modified), using strain Ag. rhizogenes AR1193 [50].
Twelve-wk-old Datisca plants were inoculated with Ag. rhizogenes strain LBA1334 [51] carrying the silencing construct by stem injection, and roots emerging at infection sites were covered with substrate. Three-wk post inoculation roots were inspected for DsRED1 fluorescence. Nonfluorescent roots were removed, and plants were repotted and grown for 8 wk. After determination of the nodulation phenotype, individual fluorescent roots were divided into two halves. One half was stained for mycorrhiza visualization, the second used for total RNA extraction (RNeasy Plant Kit, Qiagen). Quantitative RT-PCR was performed with GeneAmp5700 (Applied Biosystems) using the SuperScript III Platinum Two-Step qRT-PCR-Kit (Invitrogen). A 123-bp DgSYMRK fragment was amplified using primers DgqPCR_SYMRK_f with DgqPCR_SYMRK_r (Table S1). As control, polyubiquitin cDNA was amplified using primers DgqPCR_Ubi_f with DgqPCR_Ubi_r (Table S1). Representative fragments were sequenced for identity confirmation.
Databases used for BLAST sequence search and analysis included http://www.ncbi.nlm.nih.gov/BLAST/, http://www.arabidopsis.org/Blast/, http://www.gramene.org/Multi/blastview, and http://genome.jgi-psf.org/Poptr1/Poptr1.home.html.
Sequences of SYMRK homologs were deposited at the EMBL Nucleotide Sequence Database (http://www.ebi.ac.uk/embl/) under accession numbers AY935263 (Al. glutinosa); AM271000, AM931079 (D. glomerata coding and genomic sequence, respectively); AY935267 (Lupinus albus); AY935265 (T. majus); AY935266, AY940041 (Ly. esculentum coding and genomic sequence, respectively); AM270999 (P. rhoeas); AM851092 (Po. trichocarpa). The GenBank (http://www.ncbi.nlm.nih.gov/Genbank/index.html) accession number for pCAMBIA 1302 is AF234298.
|
10.1371/journal.pgen.1006072 | Keap1-Independent Regulation of Nrf2 Activity by Protein Acetylation and a BET Bromodomain Protein | Mammalian BET proteins comprise a family of bromodomain-containing epigenetic regulators with complex functions in chromatin organization and gene regulation. We identified the sole member of the BET protein family in Drosophila, Fs(1)h, as an inhibitor of the stress responsive transcription factor CncC, the fly ortholog of Nrf2. Fs(1)h physically interacts with CncC in a manner that requires the function of its bromodomains and the acetylation of CncC. Treatment of cultured Drosophila cells or adult flies with fs(1)h RNAi or with the BET protein inhibitor JQ1 de-represses CncC transcriptional activity and engages protective gene expression programs. The mechanism by which Fs(1)h inhibits CncC function is distinct from the canonical mechanism that stimulates Nrf2 function by abrogating Keap1-dependent proteasomal degradation. Consistent with the independent modes of CncC regulation by Keap1 and Fs(1)h, combinations of drugs that can specifically target these pathways cause a strong synergistic and specific activation of protective CncC- dependent gene expression and boosts oxidative stress resistance. This synergism might be exploitable for the design of combinatorial therapies to target diseases associated with oxidative stress or inflammation.
| Nrf2-related transcription factors regulate gene expression programs that protect organisms against chemical or oxidative stress. Nrf2-activating drugs hold promise for the treatment of diseases that are connected to oxidative stress or inflammation. We identified Fs(1)h, a bromodomain-containing BET protein, as a negative regulator of Nrf2 function in Drosophila. BET proteins are involved in transcription regulation and chromatin organization and have been implicated in several diseases, including cancer. Fs(1)h interacts with acetylated lysines on CncC, the homolog of Nrf2 in Drosophila, and thereby prevents target gene activation. Nrf2 can be released from this inhibitory effect by small molecules that specifically interfere with the binding of BET proteins to acetylated targets. Fs(1)h regulates Nrf2 independently of Keap1, a well-studied Nrf2 regulator. Consequently, chemical inhibitors of Keap1 and of Fs(1)h can be combined to achieve synergistic activation of Nrf2 target genes and strongly boost oxidative stress tolerance in Drosophila. The Keap1-independent mechanism of Nrf2 regulation is conserved in mammals. We suggest that the synergistic effect of combinatorial Nrf2 targeting drugs may be effective for the treatment of different oxidative stress and inflammation-related diseases.
| Nrf2 transcription factors are critically important for the health, homeostasis and longevity of multicellular organisms [1–3]. When cells are confronted with oxidative or chemical stress, Nrf2 stimulates the expression of gene products that protect cell integrity including antioxidants, redox regulators, phase II detoxification enzymes, and factors that maintain proteostasis. The Kelch domain protein Keap1 has been identified as a key mediator of the acute activation of Nrf2 in response to oxidative stress [4, 5]. In unstressed cells Keap1 assembles a Cullin 3-containing ubiquitin ligase complex that targets Nrf2 for proteolysis by associating with its NH2-terminally located NEH2 domain. This Keap1-mediated degradation of Nrf2 is relieved upon stress exposure, so that the transcription factor can accumulate in the nucleus and bind to so-called ARE (antioxidant response element) sequences in target gene promoters.
Over the last few years Nrf2 has been implicated in a range biological processes in addition to stress responses. Examples include the regulation of energy metabolism [1, 6], stem cell maintenance [7] and aging [8, 9]. These functions are probably regulated by signals other than toxic insults and presumably require a different transcriptional response, in terms of kinetics and target gene profile. In order to explore how this expanded range of Nrf2 functions might be regulated we conducted a large-scale screen for gene products that are involved in Nrf2 target gene activation.
The experimental model used in our studies is Drosophila melanogaster. The fruit fly has an Nrf2 signaling system, which resembles that of mammals [3, 10]. The Drosophila ortholog of Nrf2, CncC is encoded by a long splice product of the cap’n’collar gene [8]. The conservation of the Nrf2 pathway genes and its powerful genetic tools make Drosophila an excellent model to study this important signaling system.
The Drosophila CncC signal transduction pathway can, like its mammalian counterpart, mediate transcriptional responses to various types of chemical or oxidative insults and protect the organism from ensuing damage. CncC can also be activated by dietary dosing with cancer chemo-preventive agents such as oltipraz and sulforaphane [8, 11]. These drugs cause Nrf2 activation without harmful stress and negligible side effects to exposed cells or organisms. Animal experiments have shown these compounds to protect against chemical carcinogens in an Nrf2-dependent manner [12]. Oltipraz and similar drugs exert their effect by interfering with the inhibitory function of Keap1 [13–15].
In a high throughput RNAi screen using Drosophila S2 cells [16] we found the gene fs(1)h to encode a negative regulator of the Drosophila Nrf2 homolog, CncC. Multiple independent double stranded RNAs that target fs(1)h mRNA, caused a significant and specific increase in the activity of an ARE luciferase reporter gene [11], identifying the fs(1)h gene as a potential inhibitor of Nrf2 function. The product of the fs(1)h gene, Fs(1)h, short for female sterile (on the first chromosome) homeotic, is counted as a member of the heterogeneous group of Trithorax proteins which generally function as epigenetic regulators [17, 18]. Fs(1)h is the sole member of the BET protein family in Drosophila [19, 20]. BET proteins are characterized by the presence of two bromodomains, adjacent to a so-called extra terminal, or ET, domain [21]. Through the bromodomains, BET proteins specifically bind to polypeptides carrying acetylated lysine residues, including acetyl-histones [22]. Mammalian BET proteins, notably BRD4, have been implicated in the regulation of gene expression. They are known to bind to chromatin and to interact with components of the transcriptional machinery such as P-TEF B and RNA polymerase II [23, 24]. In addition, the activities of specific transcription factors such as NF-κB and Twist can be regulated by direct interaction with BRD4 [25–27]. Similarly, functional experiments and genome-wide ChIP mapping studies suggest that the Drosophila BET protein Fs(1)h functions in the regulation of gene activity [18]. Fs(1)h gene products have been found associated with transcription control region and genomic insulator elements [28, 29]. Our experiments show that Fs(1)h can physically interact with CncC to inhibit its transcriptional function. This mechanism is independent of Keap1-mediated Nrf2 regulation.
A previously described high throughput RNAi screen in Drosophila S2 cells had identified Fs(1)h as a possible negative regulator of CncC target gene activity [16]. The function of Fs(1)h as a CncC inhibitor, as suggested by this screen was confirmed by performing transient transfection assays in Drosophila S2 cells. These experiments demonstrated that knock down of Fs(1)h caused an increase in ARE reporter gene activity of a similar magnitude as seen upon knock-down of the canonical CncC repressor Keap1 (Fig 1B). When, in addition to Keap1 or Fs(1)h, CncC was knocked down, the activation of the ARE reporter was significantly reduced (Fig 1B). In addition, over-expression of Keap1 to specifically inhibit Nrf2 signaling by limiting its nuclear accumulation, abrogated the induction of the ARE-luciferase reporter in response to Fs(1)h knock down (S5A Fig). Taken together, these indicate that the stimulatory effect of Fs(1)h on ARE reporter is dependent on CncC function. Quantitative mRNA measurements in vivo by RT-qPCR supported this conclusion. Flies in which Fs(1)h was knocked down by dsRNA expression under the control of the tub-GS-Gal4 driver showed that endogenous CncC target genes (gcl-C, gstD1, keap1) were up-regulated in fs(1)h knock down conditions (Fig 1C).
Like some mammalian BET protein genes, the Drosophila fs(1)h locus yields alternatively spliced transcripts that give rise to two different polypeptides, one of 120 and one of 210 kDa molecular mass (Fig 1A) [18, 30]. In the following we will refer to the short and the long isoform as Fs(1)h-S and Fs(1)h-L, respectively. Both alternative splice products contain the two bromodomains and the ET domain. The unique peptide sequences that extend the Fs(1)h-L isoform comprise a C-terminal motif (CTM) (Fig 1A). Genome-wide ChIP experiments have shown that the Fs(1)h-S and L differ substantially in their genomic binding patterns and presumably in their function [20, 28]. To test whether the two splice forms might also differ in their effect on CncC-regulated transcription, we designed dsRNAs that would specifically target only one or the other splice variant (see Materials and Methods). Selective knock down of the long isoform Fs(1)h-L induced ARE reporter activity. However, knock down of the short isoform, Fs(1)h-S failed to do so (Fig 1D). We conclude that the repression of CncC activity is a specific function of Fs(1)h-L. Western blot experiments confirmed the efficient and selective knock down of individual isoforms by the respective dsRNAs (S1A Fig).
The inhibitory function of Fs(1)h on CncC reporter gene activity can also be observed in vivo. We conducted experiments with Drosophila stocks carrying an ARE-GFP reporter gene in which GFP expression is controlled by four consensus ARE sequences (ARE-GFP, [11]). A UAS fs(1)h-RNAi construct was expressed under the control of the RU486-inducible tubulin-GS-Gal4 driver to ubiquitously knock down endogenous fs(1)h transcripts in adult flies. This resulted in robust activation of the ARE-GFP reporter transgene in the animals, recapitulating the effect seen in S2 cells. To rule out that the activation of ARE reporter activity was the consequence of an off-target effect, we conducted this experiments with two independent RNAi expression lines and saw similar results (Figs 2A, S1B and S2B).
To examine the in vivo effect of fs(1)h loss-of-function on the ARE-GFP reporter at the cell level, we conducted knock down experiments in groups of cells using actin-flipout-Gal4, a driver that can be clonally activated by a short heat-shock induced pulse of flp recombinase expression. Clones of RNAi-expressing cells can be distinguished by the expression of an RFP marker. The resulting RFP-labeled fs(1)hRNAi-expressing cells showed increased ARE-GFP reporter activity, consistent with the presumed function of Fs(1)h as a CncC inhibitor. Fig 2B shows such a clone in the crop, a part of the foregut that, together with the anterior midgut, functions as the stomach in Drosophila [31]. We found that the epithelial cells of the crop have low basal and high inducible ARE activity. The stimulatory effect of fs(1)h knock down was restricted to the cells of the clone, demonstrating that the regulatory function of the protein acts by a cell-autonomous mechanism.
The results of the loss-of-function experiments described above suggested that Fs(1)h can suppress gene activation by CncC. However, an alternative interpretation would be that the observed increase of CncC activity might be an indirect consequence of possible stress or damage in fs(1)h loss-of-function conditions. To rule out the latter mechanism, and to demonstrate the repressive function of Fs(1)h directly, we asked whether its over-expression could decrease CncC target gene activity in adult flies. Fs(1)h over-expression was achieved by two alternative strategies. First, we used an EP-line (P(EP)fs(1)h[EP439]), a fly stock in which a Gal4 responsive enhancer was integrated in the 5’ genomic region of the fs(1)h gene. If combined with the ubiquitously expressed tub-GS-Gal4 driver, expression of endogenous fs(1)h mRNA was stimulated when flies were exposed to dietary RU486 (S1B Fig). This resulted in a marked reduction of ARE reporter activity (Fig 2C). Second, we generated a transgenic fly line that expresses a cDNA encoding Fs(1)h-L, the isoform that showed CncC repressor activity in S2 cell-based studies. Inducible expression of the Fs(1)-L under the control of tub-GS-Gal4 caused a clear reduction of ARE-GFP reporter activity (S2A Fig). In addition, over-expression of Fs(1)h-L in RFP-marked hsFlp-induced clones in the ejaculatory bulb of adult Drosophila, a tissue with a high basal level of ARE reporter activity, had a negative effect, confirming that Fs(1)h represses CncC in a cell-autonomous fashion (Fig 2D). Thus, Fs(1)h functions as a repressor of CncC’s transcriptional activity both in cell culture and in adult flies.
The finding that Fs(1)h acts as a repressor of CncC’s transcriptional output, predicted that it should modulate the biological functions of the Drosophila Nrf2 ortholog. Nrf2-induced gene expression programs can protect organisms against acute oxidative stress. Experimentally such a stress can be generated by dietary exposure to diethyl maleate (DEM), a glutathione-depleting agent. We have demonstrated previously that genetically increasing CncC activity can boost the resistance of flies against oxidative insults like DEM exposure [8]. To test if Fs(1)h might similarly affect oxidative stress resistance, we measured the DEM sensitivity of flies, in which the activity of Fs(1)h was either suppressed or increased. For loss-of-function experiments, fs(1)h mRNA was knocked down in cohorts of young adults flies for 4 days by RU486-induced expression of a corresponding RNAi construct. Subsequently, the animals were transferred to vials containing filter paper laced with a lethal concentration of DEM and the time course of survival was recorded. Cohorts in which fs(1)h was depleted displayed a significantly increased survival time as compared to controls (Figs 2E and S3C). For gain-of-function experiments Fs(1)h was over-expressed by feeding EP-fs(1)h; tub-GS-Gal4 stocks with RU486 for 4 days. DEM exposure resulted in a more rapid demise of the flies in these cohorts compared to matching controls (Figs 2F and S3B). Similarly, inducible over-expression in the UAS-Fs(1)h-L fly line also resulted in sensitization to DEM stress (S3A Fig). These DEM exposure experiments support the conclusion that Fs(1)h is a significant regulator of stress defense mechanisms, presumably through its effect on CncC.
Through their double bromodomains BET proteins like Fs(1)h can bind to polypeptides that carry acetylated lysine residues. To assess whether acetyl lysine binding might contribute to the repressive function of the protein towards CncC’s transcriptional activity, we conducted experiments with JQ1, a specific inhibitor of BET-domain interactions with acetylated substrates [32]. Treatment of S2 cells with JQ1 caused a strong activation of ARE-luciferase reporter activity that was markedly reduced by knock down of either CncC or its obligate heterodimerization partner MafS (Fig 3A). Consistent with the observation that dsRNA-mediated knock down of either CncC or MafS, individually does not completely eliminate the targeted transcripts, a residual response of the reporter to JQ1 could be observed. To confirm the conclusion that the active form of Drosophila Nrf2, the MafS-CncC heterodimer, specifically mediates the effect of JQ1 we combined CncC and MafS knock down conditions. Under these conditions, the JQ1-mediated induction of the ARE reporter was almost completely abrogated. This observation confirms the specificity of the JQ1-effect and further supports our conclusion that Fs(1)h affects ARE activity by interfering with CncC function. Similarly, dietary JQ1 exposure (Fig 3B) resulted in a strong enhancement of ARE reporter activity also in flies. This stimulatory effect does not seem to be a stress response, as we did not detect any increased mortality in flies or cells after treatment with JQ1 as they would result from treatment with Nrf2-inducing stressors such as paraquat or DEM.
The experiment described above suggests that JQ1 can relieve the inhibition of CncC by Fs(1)h. Since JQ1 has been shown to block the interaction of BET bromodomains with acetyl lysine residues, we proposed that Fs(1)h would interact with acetylated CncC and that this interaction would be disrupted by JQ1, causing an increase in CncC activity. Indeed, specific acetylated forms of mammalian Nrf2 have been described [33, 34]. To assess whether CncC might also be a substrate for acetylation and could therefore represent a potential binding partner for the Fs(1)h bromodomains, we performed immune precipitation/western blot experiments. S2 cells expressing a Flag-epitope tagged version of CncC were processed for immuno-precipitation with an anti-Flag antibody. Staining western blots of the resulting immuno-precipitates with a generic anti-acetyl lysine antibody revealed that, like its mammalian counterpart, CncC protein is acetylated (Fig 3C). Treatment of the cells with the broad-spectrum HDAC inhibitor LBH589 increases the acetylation of the immuno-precipitated CncC, confirming that the protein is a substrate for acetylation / deacetylation reactions in vivo.
Next, we tested whether Fs(1)h might physically interact with CncC and, if so, whether Fs(1)h binding correlates with CncC acetylation status. Confirming this idea, we found that Flag-tagged CncC expressed in Schneider cells could be co-immuno-precipitated with endogenous Fs(1)h-L. The yield of the recovered Fs(1)h-L increased in the presence of HDAC inhibitor and decreased in the presence of JQ1 (Fig 3C). These results indicated that Fs(1)h-L and CncC bind to each other through a BET bromodomain–acetyl lysine interaction.
Next we wanted to investigate whether the stimulatory effect of Fs(1)h inhibition on CncC target gene expression might be mediated by changes in the expression of CncC or some other gene. Therefore, we treated S2 cells with the protein synthesis blocker cycloheximide prior to JQ1 exposure. Under those conditions, protein synthesis, as measured by luciferase expression (Fig 4A), was completely inhibited, but the JQ1-mediated induction of the CncC target genes gstD1 and keap1 was unaffected (Fig 4B and 4C). We conclude that Fs(1)h regulates CncC at the post-translational level to affect the expression of its target genes. Such a model fits our hypothesis that Fs(1)h regulates CncC protein function by physically interacting with it.
The cnc gene generates several different protein products by alternative splicing [37, 38]. The three initially described Cnc splice forms CncA, CncB and CncC all share a common C-terminal region that comprises the bZIP dimerization and DNA-binding domain. Moving from the A to the C isoform the polypeptides include progressively more N-terminal sequences. Of these three gene products only CncC contains the NEH2 domain, which mediates binding to and repression by Keap1 [8]. To assess which sequences are required for binding to Fs(1)h we performed co-immuno-precipitation experiments with different Cnc protein isoforms. Epitope-tagged versions of CncA, B, and C were expressed in S2 cells and immuno-precipitated with an anti-Flag antibody. The immuno-precipitates were analyzed by western blot probed with an anti Fs(1)h antibody that recognizes both the short and the long isoforms of Fs(1)h. Endogenous Fs(1)h-L protein was only detected in the immuno-precipitated material from S2 cells that expressed the longest isoform, CncC. However, it was not co-precipitated with CncA or CncB (S4 Fig). Notably this behavior is shared by Keap1, which also binds to CncC only but not to the two shorter isoforms. No interaction between CncC and Fs(1)h-S was identified in this experiment. This is consistent with our finding that only Fs(1)h-L is involved in the negative regulation of CncC. We conclude that sequences within the region of amino acids 1–578 which are unique to CncC, the only isoform with recognized functions in stress response, are required for both Fs(1)h-L and Keap1 binding. Further experiments are required to test if the binding of the two inhibitory proteins, Keap1 and Fs(1)h can happen simultaneously or is mutually exclusive [25–27]. Interestingly, we did detect the short isoform Fs(1)h-S as in the co-precipitate with CncB (S4 Fig). CncB is not involved in stress responses, but functions in embryonic patterning and development. It has been suggested to act as a negative regulator of transcription [10, 37]. The functional relevance of a potential interaction between CncB and Fs(1)h-S awaits experimental analysis.
The discovery that Fs(1)h physically interacts with and inhibits CncC raised the question of how this activity relates to the regulatory function of Keap1, the canonical inhibitor of Nrf2 proteins. Specifically, we wanted to determine whether Fs(1)h acts via Keap1-mediated CncC degradation, or through an independent mechanism to control target gene expression. To address this question we conducted transfection experiments in S2 cells with drugs and RNAi’s that specifically interfere with Keap1 or Fs(1)h functions (Fig 5A). The experiments described above indicated that JQ1 prevents the inhibitory effect of Fs(1)h on CncC, thereby stimulating ARE-dependent gene expression. Consistent with this interpretation, fs(1)h knock down and JQ1 exposure stimulate ARE-luciferase reporter activity to a similar degree (Fig 5A, compare lane 1 with 2 and 3), and combining the fs(1)h RNAi with JQ1 treatment does not further enhance this activity (Fig 5A, lane 4). Likewise, oltipraz exposure and Keap1 knock down activated the reporter to comparable levels, yet the effect of the two treatments is non-additive, confirming that oltipraz acts specifically through inhibition of the Keap1-CncC interaction (lanes 5–7). Interestingly, however, eliminating the function of both Fs(1)h and Keap1 simultaneously by combined RNAi and drug treatments caused a synergistic activation of the reporter (lanes 8 and 9). Similarly, combining JQ1 and Oltipraz treatment resulted in a synergistic activation of the ARE reporter in S2 cells over a range of concentrations (Fig 5B). Confirming the results from S2 cells, the combined treatment of adult flies with oltipraz and JQ1 lead to stronger induction of the gstD-GFP reporter relative to the effects of either oltipraz or JQ1 alone (Fig 5C). gstD1 is a target gene of CncC and a gstD-GFP reporter had previously been generated by placing a GFP transgene under the control of a gstD1 promoter fragment [8].
The induction of gstD-GFP reporter activity in flies in response to JQ1 treatment also indicated that natural CncC regulated enhancers, and not only synthetic ARE reporters are responsive to Fs(1)h inhibition in adult Drosophila. Mirroring the measurements of luciferase reporter constructs, the expression of four conserved Nrf2 target genes (gstD1, gstE9, keap1, gclc) was inducible by treatment with JQ1 or oltipraz and combined treatment had a more than additive effect relative to the response to the individual drugs (Fig 5D).
In further experiments we found that fs(1)h knock down also led to a synergistic activation of ARE luciferase reporter activity when combined with either Keap1 knock down or CncC over-expression. Presumably, CncC over expression overwhelms the capacity of Keap1 to target the protein for degradation. Thus, both in conditions of CncC over-expression and Keap1 knock down, CncC can accumulate in the nucleus. Gene activation by this increased nuclear pool of CncC protein would then be under negative control by Fs(1)h (S5A Fig). Taken together, these results indicate that Fs(1)h and Keap1 regulate Nrf2 function through independent mechanisms.
The effective and specific pharmacological activation of the Nrf2 response by combined JQ1 and oltipraz treatment might be therapeutically beneficial, provided that the mechanisms that we have characterized in Drosophila are conserved and operational in mammals. Interestingly, it has been shown in recent publications [39, 40] that mammalian Nrf2 can be stimulated by JQ1. We therefore tested if mammalian cells displayed cooperative activation of Nrf2 target genes in response to the combinatorial treatment that we had found effective in Drosophila. Confirming this idea, combined treatment of human HEK293 cells with sulforaphane, a Keap1 inhibitor functionally similar to oltipraz, and JQ1 results in a more than additive activation when compared to individual treatments (Fig 5E). In addition, JQ1 but not sulforaphane caused synergistic activation of NQO1-luciferase reporter when combined with over-expressed Nrf2, indicating that the Keap1-independent regulation of Nrf2 by BET protein is conserved in mammals (S5B Fig). Combined dosing of these two drugs might therefore open new avenues for treatment of diseases where increased Nrf2 activity could be beneficial.
If a combinatorial treatment with the Keap1 inhibitor oltipraz, and the BET bromodomain inhibitor JQ1 activates Nrf2-dependent stress defense and antioxidant gene expression programs more strongly than oltipraz or JQ1 alone, we would expect a further increase in the resistance of flies to acute oxidative challenges. To confirm this prediction we exposed adults to a lethal dose of DEM after they had been kept on food containing oltipraz, JQ1 or both drugs for 4 days. Monitoring the time course of mortality showed that pretreatment with oltipraz and JQ1 alone enhances the oxidative stress tolerance when compared to the control group. However, pretreatment with a combination of oltipraz and JQ1 increased survival times even further (Figs 6A and S3D). Thus, combinatorial treatment with oltipraz and JQ1 can yield effective protective effects against oxidative challenges and may also be useful for therapeutic applications in which Nrf2 function is expected to be beneficial.
BET proteins including Drosophila Fs(1)h and the members of the mammalian Brd family function as epigenetic readers and transcriptional regulators [28, 29, 41, 42]. They can affect the function of specific transcription factors such as NF-kB, FosL and Twist, but also interact with components of the general transcription machinery such as P-TEFb [23–27, 43]. Mammalian BET proteins, most notably Brd4, have raised broad biomedical interest as well, because they can support several types of cancer, including NUT midline carcinoma, multiple myeloma and acute leukemia [44–47]. The implication of BET proteins in these malignancies could have clinical impact as the availability of specific inhibitors makes them potentially promising drug targets. In order to evaluate and exploit these opportunities, it is critical to gain a better understanding of the complex functions and molecular targets of BET proteins.
We have identified the Drosophila BET protein Fs(1)h as a novel inhibitor of the Drosophila Nrf2 homolog CncC. The fs(1)h gene encodes two different isoforms, Fs(1)h-S and Fs(1)h-L. Fs(1)h-S has previously been reported to transcriptionally activate the ultrabithorax (ubx) gene, a function that is not shared by the larger Fs(1)h-L isoform [18]. Conversely, our study showed that only Fs(1)h-L, but not Fs(1)h-S, has the potential to inhibit CncC target gene activation. The divergence of function between Fs(1)h-S and L is also reflected in their respective genomic binding patterns: A recent genome-wide ChIP-seq study showed that Fs(1)h-S and Fs(1)h-L are present at different locations throughout the Drosophila genome [28]. Whereas Fs(1)h-S was enriched in promoters and enhancers, Fs(1)h-L was predominantly localized to the insulators. Interestingly, our analysis of the same ChIP-seq data revealed that Fs(1)h-L also localizes to the promoter region of many CncC target genes such as gstD1, gclC, and keap1. It is possible that Fs(1)h-L interacts with promoter-bound CncC to prevent the induction of its target genes.
Nrf2 is an attractive drug target. Compounds such as oltipraz and sulforaphane which interfere with the Keap1-Nrf2 interaction to activate target genes without causing cell stress, are effective as cancer chemo-preventive agents in animal experiments [48–51]. Nrf2-activating drugs can also be beneficial in the treatment of other diseases that are associated with oxidative stress, including neurodegenerative and inflammatory conditions [52, 53]. Indeed, an Nrf2 inducer, dimethyl fumarate, marketed under the brand name Tecfidera, has recently gained FDA approval for treatment of multiple sclerosis [54, 55]. The identification of JQ1 as a CncC activator and, the discovery that JQ1 and oltipraz synergistically activate Nrf2 dependent gene expression programs in Drosophila and mammalian cells suggest innovative therapeutic options. Conventional Nrf2 activating drugs like sulforaphane and oltipraz, though effective in some cases, show moderate induction of Nrf2 and often have off-target effects [49, 52]. Combinations of JQ1 and oltipraz-like drugs might stimulate Nrf2 dependent protective gene expression more efficiently than either compound alone, or could achieve beneficial effects at lower doses, thereby decreasing the risk of unwanted side effects.
Based on the known properties of Fs(1)h and Keap1 and our cell culture and biochemical experiments we suggest a mechanisms by which Fs(1)h represses CncC function. As opposed to the effect of Keap1, which targets CncC for cullin 3-mediated proteasomal degradation in the cytoplasm, the inhibitory, acetylation-dependent interaction between Fs(1)h and CncC presumably occurs in the nucleus where the BET protein resides (Fig 6B). The independent mechanisms of CncC inhibition by Keap1 and Fs(1)h are consistent with the synergistic effect of combined JQ1 and oltipraz treatment on Nrf2 target gene expression.
fs(1)h is the only BET protein encoding gene in Drosophila and Brd4 is the closest ortholog of Fs(1)h in mammals [28]. Recent publications by Michaeloudes et al and Hussong et al have implicated Brd4 in the regulation of mammalian Nrf2 [39, 40]. In agreement with our conclusions, Michaeloudes and colleagues have found that Brd4 and other mammalian BET proteins namely Brd2 and Brd3, physically interact with Nrf2. Hussong and colleagues, on the other hand, suggested a somewhat different model of Nrf2 regulation by Brd4. They reported that JQ1-treatment of human cells activates Nrf2 target genes, but suppresses Keap1 expression. Accordingly, they reason that the activation of Nrf2 target genes by JQ1 might be mediated by a loss of Keap1 repression [40]. The data we generated in Drosophila S2 cells, on the other hand, show that both Fs(1)h knock down and JQ1 treatment leads to increased keap1 transcription, which is consistent with the previous identification of the Keap1 gene as a target of CncC mediated feedback regulation. It is possible that species- or cell type-specific differences in JQ1-Keap1 cross talk exist and account for these divergent results. In any case, our finding that JQ1 and sulforaphane synergistically activate an ARE reporter in human cells supports the notion that combinatorial treatment with these two drug types might be therapeutically beneficial.
Our data show that the suppression of CncC activity by Fs(1)h relies on a bromodomain–acetyl lysine interaction. This finding implies a repressive function of acetylation on CncC’s transactivation potential. Such a conclusion appears to be at odds with reports showing acetylation of Nrf2 by CBP to increase its binding to target DNA and to enhance target gene transcription [33, 34]. However, Mercado et al. have reported that inhibition of HDAC2 and the resulting increase in Nrf2 acetylation can suppress Nrf2-mediated target gene induction and antioxidant defense in chronic obstructive pulmonary disease (COPD) [56]. It seems therefore that Nrf2 acetylation can have positive as well as negative effects on the transcriptional function of Nrf2 proteins. Further studies are needed to uncover whether specific acetylation marks on Nrf2 lead to either activation or suppression of its transcriptional activity and whether BET proteins can selectively interact with the inhibitory acetylation marks on Nrf2.
The identification of BET proteins as Nrf2 repressors adds another facet to an increasingly complex picture of Nrf2 signaling and biology. Over the last few years several other Keap1-independent pathways of Nrf2 regulation have been described based on experiments in mammalian cell culture: for example it was shown that PKC-δ mediated phosphorylation of Ser40 of mammalian Nrf2 promotes its stabilization and nuclear translocation [57–59]. The Src-family tyrosine kinase Fyn can phosphorylate Tyr568 of Nrf2 causing its nuclear export and degradation [60]. Glycogen synthase kinase 3 beta (GSK-3β) can phosphorylate Fyn and increase its nuclear accumulation and thereby promotes nuclear export of Nrf2 and inhibition of Nrf2 signaling [61]. In addition, GSK-3β can phosphorylate the serine residues at the β-TRCP-binding motif (DSGIS338) in the Neh6 domain of Nrf2 and promote Cullin 1 dependent proteasomal degradation by β-TRCP [62]. Future experiments will help to evaluate whether the Fs(1)h and CncC interaction cooperates with any of these more recently described mechanisms of Nrf2 regulation.
All plasmids were generated by standard recombinant DNA and PCR methods. The pUAS-HA-attB plasmid was generated by cloning attB sequence amplified with PattB-F and PattB-R primers into pUAS-HA plasmid [8]. UAS-Fs(1)h-L plasmid was generated in two steps. First the Fs(1)h short isoform cDNA was amplified from LD26482 cDNA clone with PFs(1)hS-UAST-HA-F and PFs(1)hS-UAS-HA-R primers and was cloned into pUAS-HA-attB plasmid to generate pUAS-Fs(1)h-S-HA-attB plasmid. Then the cDNA for the C-terminal motif of Fs(1)h-L was amplified from total Drosophila cDNA with PUAST-Fs(1)h-L-int-F and PUAST-Fs(1)hL-int-R primers and was cloned into pUAS-Fs(1)h-S-HA-attB plasmid to generate pUAS-Fs(1)hL-HA-attB plasmid. The primer sequences are provided in Table A in S1 Text.
UAS-Fs(1)h-L transgenic fly strain was generated by Genetic Services Inc, MA using ΦC31 recombinase-mediated site-directed transformation [63]. EP-Fs(1)h fly line #10097 (Bloomington Stock Center) was used to over-express Fs(1)h whereas v51227 and v108662 fly lines (Vienna Stock Center) were used to knock down Fs(1)h.
Drosophila S2 cells and HEK293 cells were transiently transfected with plasmid constructs using the calcium phosphate method [11].
To study the effect of JQ1 on reporter gene expression in vivo, 5-day-old flies that were mated for one day and then separated into males and females, were fed food supplemented with 0.25 mM JQ1 (APExBIO) for 48 hours. 15–20 flies were used in each group for these experiments and 3–5 representative flies were chosen randomly for imaging. To assess the effect of JQ1 on the cell-based reporters, S2 cells were transiently transfected with the reporter plasmids by the calcium phosphate method. 8 hours after the PBS wash and medium change, the transfected cells were transferred to 96-well plates and treated with 1 μM JQ1 and were incubated at 25°C for 24 hrs. Treatment with oltipraz was done following the same protocol [11]. In order to study the effect of different chemicals on the ARE reporter in mammalian cells, transfected 293T cells were treated with 0.5μM JQ1 and 10μM sulforaphane (Enzo Life Sciences) and were incubated at 37°C for 18hrs. DMSO was used as the solvent control in all the drug treatments.
dsRNAs (200–700 bp) were synthesized and purified following the protocol provided as described [64] using ‘T7 RiboMAX Express RNAi System’ kit (Promega) and ‘RNeasy Kit’ (QIAGEN). Briefly, 1X106 S2 cells were bathed with 8μg dsRNAs and these cells were transfected with luciferase reporter plasmids using the calcium phosphate method 3 days after dsRNA treatment. Predesigned dsRNAs obtained through the E-RNAi webservice [65] were used to knock down Fs(1)h, CncC, Keap1 whereas Fs(1)h-L was knocked down using the dsRNA described by Kockmann et al. [29]. dsRNA targeting the unique 3’ UTR region of Fs(1)h-S was designed by the ‘SnapDragon’ webservice. The sequences of primers used to generate amplicons for dsRNA synthesis are provided in Table B in S1 Text.
The ‘Dual Glo Luciferase Assay System’ kit (Promega) was used to measure the activities of cell-based firefly and renilla luciferase reporters.
Oxidative stress resistance of adult flies of different genotypes was assessed as previously described [8]. Newly emerged flies of the specified genotypes were mated for one day, separated into females and males, and at 5 days of age were transferred to RU486-containing (300μM) or control food containing the solvent (ethanol). After 4 days 4 groups of 20 flies from each condition were starved for 3 hours in empty vials, and then fed a solution of 5% sucrose ± a semi-lethal dose of DEM (20mM). Survivors were scored after 36 hours and Log-rank tests were performed on the survivorship data using ‘GraphPad Prism’ software. In order to study the synergistic effect of oltipraz and JQ1 on stress sensitivity, flies raised on food supplemented with 0.4mM oltipraz and/or 0.1mM JQ1 for four days were exposed to similar DEM treatment and the survivorship data was collected and analyzed as before.
To study the effect of Fs(1)h knock down and Fs(1)h-L over-expression on ARE activity in clones of tissue hsFlp; Act>Y>Gal4, UAS-RFP, ARE-GFP; UAS-Fs(1)h-RNAi and hsFlp; Act>Y>Gal4, UAS-RFP, ARE-GFP; UAS-Fs(1)h-L flies were used. The embryos and larvae were incubated at 18°C before L2 larvae were heat treated at 37°C for 30 minutes and then returned to 18°C. Adult flies were dissected in PBS and crops and ejaculatory bulbs were fixed at room temperature for 30 minutes in 100mM glutamic acid, 25mM KCl, 20mM MgSO4, 4mM sodium phosphate, 1mM MgCl2, and 4% formaldehyde (pH 7.5). DNA was stained with Hoechst dye. Confocal images were collected using a Leica TCS SP5 system and were processed using Adobe Photoshop.
S2 cells transfected with pAct-Gal4 and pUAS-CncC-FLAG plasmids were harvested and washed with cold 1XPBS. The cells were then re-suspended in lysis buffer (50mM Tris-HCl, 200mM NaCl, 5mM EDTA, 5% Glycerol, 0.2% NP-40) that contained protease inhibitor complex (Roche) and kept on a nutator at 4°C for 1 hour. The cell debris was removed by centrifugation and the lysate was passed through hypodermic syringe to shear DNA. The protein concentration was estimated using Bradford’s reagent and 20–30 μg of protein from each sample was set aside as input. The lysate was pre-cleared with Protein G beads (GE Biosciences) for 1 hour at 4°C. The beads were separated by centrifugation and the cleared lysate was transferred into the fresh chilled tube. Anti-FLAG (Sigma-Aldrich Co) antibody was added at 2 μg per mg lysate protein concentration and was incubated at 4°C for 2 hours. Then 20 μl of 50% Protein G bead slurry was added to the lysate and incubated overnight on rotating disc at 4°C. The beads were spun down and were washed 5 times with 750 ul of cold lysis buffer. After the final wash, 7.5 μl of sample buffer and 7.5 μl of lysis buffer were added to the beads and incubated at 95°C along with the input samples for 10 minutes. The beads were spun down at the top speed for 5 minutes and the supernatant was loaded for electrophoresis. For LBH and JQ1 treatment, the cells were treated with 2μM of LBH (APExBIO) and 10 μM JQ1 for 6 hours prior to the cell lysis. JQ1 was also added to the lysate from cells treated with JQ1 at 10 μM concentration after the cell lysis. Primary antibody specific to Fs(1)h-L isoform (1:2000 dilution) (Kindly provided by Dr. Victor Corces) [28] and secondary anti-rabbit antibody (1:5000 dilution) (BioRad) were used to probe the western blot membrane to examine the binding of Fs(1)h-L protein to CncC-FLAG protein. Antibody that recognizes both isoforms of Fs(1)h (Kindly provided by Dr. Igor Dawid) was used in 1:2000 dilution in western blot experiments to validate the selective knock down of Fs(1)h-S and Fs(1)h-L with isoform-specific dsRNAs. Antibody against acetylated lysine (Cell Signaling Technology) was used in (1:1000) concentration.
mRNA was prepared from whole flies or from S2 cells using Trizol reagent (Invitrogen). DNaseI (NEB) was used to remove contaminating DNA before phenol: chloroform: isoamyl alcohol (25:24:1, Amresco) extraction. Maxima reverse transcriptase (Fermentas) and oligo dT primers were used to generate cDNA. cDNA was diluted 1:50 to serve as template for quantitative real time PCR (qPCR). qPCR reactions were done in triplicates using qPCR super-mix (BioRad) on a BioRad MyIQ thermal cycler. ‘Delta-delta Ct’ method was used for normalization to actin5c transcript levels. Data shown are averages and standard deviations from at least three biological replicates. The sequences of primers used to qPCR are provided in Table C in S1 Text.
|
10.1371/journal.pntd.0000618 | Complete In Vitro Life Cycle of Trypanosoma congolense: Development of Genetic Tools | Animal African trypanosomosis, a disease mainly caused by the protozoan parasite Trypanosoma congolense, is a major constraint to livestock productivity and has a significant impact in the developing countries of Africa. RNA interference (RNAi) has been used to study gene function and identify drug and vaccine targets in a variety of organisms including trypanosomes. However, trypanosome RNAi studies have mainly been conducted in T. brucei, as a model for human infection, largely ignoring livestock parasites of economical importance such as T. congolense, which displays different pathogenesis profiles. The whole T. congolense life cycle can be completed in vitro, but this attractive model displayed important limitations: (i) genetic tools were currently limited to insect forms and production of modified infectious BSF through differentiation was never achieved, (ii) in vitro differentiation techniques lasted several months, (iii) absence of long-term bloodstream forms (BSF) in vitro culture prevented genomic analyses.
We optimized culture conditions for each developmental stage and secured the differentiation steps. Specifically, we devised a medium adapted for the strenuous development of stable long-term BSF culture. Using Amaxa nucleofection technology, we greatly improved the transfection rate of the insect form and designed an inducible transgene expression system using the IL3000 reference strain. We tested it by expression of reporter genes and through RNAi. Subsequently, we achieved the complete in vitro life cycle with dramatically shortened time requirements for various wild type and transgenic strains. Finally, we established the use of modified strains for experimental infections and underlined a host adaptation phase requirement.
We devised an improved T. congolense model, which offers the opportunity to perform functional genomics analyses throughout the whole life cycle. It represents a very useful tool to understand pathogenesis mechanisms and to study potential therapeutic targets either in vitro or in vivo using a mouse model.
| Trypanosoma congolense is a parasite responsible for severe disease of African livestock. Its life cycle is complex and divided into two phases, one in the tsetse fly vector and one in the bloodstream of the mammalian host. Molecular tools for gene function analyses in parasitic organisms are essential. Previous studies described the possibility of completing the entire T. congolense life cycle in vitro. However, the model showed major flaws including the absence of stable long-term culture of the infectious bloodstream forms, a laborious time-consuming period to perform the cycle and a lack of genetic tools. We therefore aimed to develop a standardized model convenient for genetic engineering. We succeeded in producing long-term cultures of all the developmental stages on long-term, to define all the differentiation steps and to finally complete the whole cycle in vitro. This improved model offers the opportunity to conduct phenotype analyses of genetically modified strains throughout the in vitro cycle and also during experimental infections.
| Animal African Trypanosomosis (AAT) is a severe disease caused by several species of African trypanosomes, flagellate protozoan parasites transmitted by an insect fly vector. It affects domestic livestock in sub-Saharan Africa and consequently has a severe economic impact [1]–[3]. AAT is widely spread affecting 40 countries situated in regions that could potentially be the most productive. The main pathological symptoms of animal trypanosomosis are weight loss, anaemia, and immunosuppression, but the mechanisms involved are poorly understood. It is estimated that 50 million cattle and 70 million small ruminants are at risk, costing the continent between 1.5 and 5 billion US dollars per annum. In terms of severity and consequences for productivity, T. congolense is the main causative agent of AAT. This parasite has a complex life cycle: bloodstream forms (BSF) proliferate in the blood of the infected mammalian host and are ingested by tsetse during the blood meal, procyclic forms (PCF) differentiate in the insect midgut, migrate to the proboscis (mouth parts) where they attach as epimastigote forms (EMF) and finally differentiate into infective metacyclic forms (MCF) that are transmitted to a new mammalian host during the next blood meal.
This parasite strongly hinders the agricultural development of the sub-Saharan region and an understanding of the virulence mechanisms is essential in order to block the parasite and/or the associated pathogenesis. Such studies require functional characterization of virulence genes based on genetic engineering. Genetic tools are well established in the T. brucei parasite with wide use of inducible transgene expression [4],[5]. More specifically, gene expression inactivation using RNA interference (RNAi) system greatly contributed to the functional analysis of T. brucei genes to such an extent that T. brucei has become the preferred model for trypanosomiasis studies [6],[7]. However, T. congolense is a strictly intravascular parasite whereas T. brucei can leave blood vessels and invade tissues implying differences in virulence mechanisms, host/pathogens relationships and pathogenic effects between the two species. Furthermore, while for T. brucei only the PCF and BSF stages can be cultivated in vitro and differentiation is limited to one step (from BSF to PCF), the whole life cycle of T. congolense can be reproduced in vitro (cultivation of all the developmental stages accomplishing all the differentiation steps) [8],[9]. Especially, the in vitro differentiation of the tsetse vector stages (from PCF to MCF) was described much earlier, this process is termed metacyclogenesis [10]. Therefore, it will be very informative and useful to use T. congolense not only as a model for AAT studies but also for deciphering the mechanisms underlying the different differentiation steps.
The major drawback we aim to overcome is that all cultivation techniques currently available are limited even for the metacyclogenesis process (which could take several months [10]) resulting in the quasi-absence of genetic tools. A single strain has been developed for the inducible RNAi system in PCF (TRUM183 strain) [11], and only one recent study described genetic modification of EMF (IL3000 strain) by transfection with a GFP vector and subsequent differentiation into MCF [12]. However, production of infectious BSF through differentiation of genetically modified parasites has never been achieved. In addition, there is an important heterogeneity between the T. congolense isolates used despite the inclination to refer to the IL3000 clone since the beginning of its genome sequencing project (Sanger Center, UK). And, there are disparities regarding the in vitro differentiation especially during metacyclogenesis and transfection efficiency [13] (our data). Finally, the question of the infectivity/virulence of wild type or genetically modified MCF differentiated in vitro has not been addressed until now.
Another severe restriction in this research field is the lack of stable long-term BSF culture. Some authors described BSF culture systems [14],[15] that turned out to be usable only for short-term applications eliminating the possibility of performing drug trials (requiring continuously dividing cultures) and above all, transfection assays, consequently resulting in the absence of functional genomics for T. congolense BSF.
Because the T. congolense model displayed significant limitations, we decided to undertake a systematic analysis of all the developmental stages regarding culture methods, differentiation efficiency, and transfection success rate by using 5 different wild type isolates and 2 genetically modified cell-lines. In this work we describe the success in developing an improved in vitro model and standardized culture methods.
Primary cultures of bovine aortic endothelial cells (BAE) were routinely maintained in EGM2V medium (Lonza) in 24 well plates in a 5% CO2 humidified atmosphere at 37°C as described by the manufacturer. Cells were used between passages 9 and 20. Co-culture with parasites was conducted in the parasite medium.
Six Trypanosoma congolense strains of the Savannah type were used in this study. IL1180 originated from the Serengeti in Tanzania [16]. ILC-49 was originally isolated from a cow from the Transmara, Kenya, and was passaged in rodents [17]. This strain and its derivative, clone IL3000, were kindly provided by the International Livestock Research Institute, Nairobi, Kenya. STIB910 is a cloned derivative of STIB249 originally isolated in 1971 from a lion in Tanzania [16], and was kindly provided by R. Brun (Basel, Switzerland). TREU1457 was derived from a stock originally isolated in Nigeria (Zaria/67/LUMP/69) [18] kindly provided by C.A. Ross. TRUM183:29-13 is a genetically modified version of the TRUM183 strain, an uncloned serodeme, kindly provided by J.E. Donelson [11].
Details on culture media used are listed in Table 1. Procyclic forms (PCF) were routinely maintained in TcPCF-3 medium. Cells were maintained in log-phase culture at 27°C in 25 cm2 flasks by changing the medium every 3 days.
Epimastigote adherent forms (EMF) appeared in this medium after several weeks of culture by maintaining the culture in stationary phase (change of half of the medium every 4 days). The entire culture supernatant was replaced with fresh medium every 2–3 days, after the colonies of EMF were well established. Alternatively, EMF could be cultivated in a few hours using a starvation medium named TcEMF-1. Briefly, 5 ml of PCF culture (≈107 cells/ml) was collected by centrifugation (1600×g, 10 min), placed in the same volume of TcEMF-1 medium and incubated at 27°C in 25 cm2 flasks for 2 hours. Individual adherent EMF or colonies were observed after only 1 hour. Heat-inactivated fetal calf serum (10%, Adgenix) was added to the flask and this medium (named TcEMF-2) was used to maintain the obtained EMF culture as described above.
Metacyclic forms (MCF) developed in the EMF cultures after few days up to few weeks depending on the strain. MCF could be purified by DE52 anion-exchange column chromatography (Whatman Plc., Brentford, U.K.) [19]. Furthermore, the number of MCF produced per ml of EMF culture were quantified using a haemocytometer.
For bloodstream forms (BSF) cultures, endothelial cells were used as a feeder cell layer. BAE were seeded in 24-well plates. After one day, the medium was removed, cells were washed with PBS (phosphate-buffered saline: 137 mM NaCl, 10 mM Phosphate, 2.7 mM KCl, pH 7.4) and incubated in 1 ml of BSF specific medium. BSF were cultured in 2 different media, the TcBSF-3 was used during the adaptation phase from the infected mouse blood or from metacyclic differentiation and the TcBSF-1 was used for well-established cultures. BSF were always maintained at 34°C in a humidified atmosphere containing 5% (vol/vol) CO2.
Parasites collected from infected mouse blood could be either differentiated in PCF or cultivated as BSF. Cultures were initiated with tail blood of infected mice when parasitaemia reached a minimum of 108 parasites/ml. 5 drops of tail blood were collected in 1 ml of TcPCF-1 or TCPCF-2 media for differentiation in PCF or in 1 ml of TcBSF-3 for BSF culture. A few derivatives of the TcBSF-3 medium were also used during the adaptation medium development phase by using horse, foal or lamb sera (Invitrogen) instead of goat serum. Centrifugation (200×g, 1 min) was used to remove the majority of red blood cells. Parasites present in the supernatants were subsequently incubated in 24 well plates at 27°C for PCF, and at 34°C in a humidified atmosphere containing 5% (vol/vol) CO2 for BSF.
For PCF, half of the medium was replaced as soon as the well-shaped parasite density in the upper part of the well reached the middle log phase (≈5.106 cells/ml). This process could take a few days to a few weeks depending on the strain. Following this adaptation phase, half of the medium was changed every 2 days.
pLEW13, pLEW29 with a modified integration site (here named pLEW29c) and p2T7Ti/αTUB were kindly provided by J.E. Donelson [11]. pLEW20 is a typical expression vector for T. brucei [20], which allows inducible expression of proteins under the control of a strong pol I promoter regulated by two tetracycline operators. To facilitate the integration and the subsequent expression of reporter proteins in T. congolense, the integration site, the rRNA spacer region (1 to 121), was replaced by the βTUB region (1 to 134) of the pLEW13 vector. A PCR fragment from the pLEW13 was produced using specific primers containing ScaI and NdeI restriction sites respectively, sub-cloned in pGEM-T vector (Promega), and inserted in the pLEW20 ScaI-NdeI restricted plasmid. This new vector, named pLEW20c, was used to produce the enhanced green fluorescent protein (EGFP: optimized for expression and fluorescence in mammalian cells) or the Renilla luciferase (RLuc). PCR amplified fragments corresponding to EGFP and RLuc were inserted in the HindIII-BamHI restriction sites of the pLEW20c vector to produce pLEW20c-EGFP and pLEW20c-RLuc.
Stably transfected strains were obtained with the Amaxa nucleofection method. The Amaxa Nucleofector® system (Lonza, Levallois-Perret, France) was used as described by the manufacturer. Briefly, a pellet of 107 parasites was resuspended in 100 µl of Basic Parasite Nucleofector® solution, mixed with 15 µg of NotI linearized plasmid and subjected to nucleofection with a specific program. Amaxa solutions and programs used are listed in Table 2. Stably transfected trypanosomes were selected using BSF or PCF media supplemented with bleomycin for pLEW20c and p2T7Ti/αTUB, G418 for pLEW29c and hygromycin for pLEW13. Increasing concentrations of antibiotics were used for BSF selection (from 0.5 µg/ml to 2 µg/ml Bleomycin; 0.25 µg/ml to 1.25 µg/ml Hygromycin and 0.1 µg/ml to 0.5 µg/ml G418), while a fixed concentration was used for PCF (2 µg/ml Bleomycin, 2.5 µg/ml G418 and 6.25 µg/ml Hygromycin). During BSF selection, half of the medium was changed every two days by carefully removing it without disturbing the adherent parasites. For PCF, the medium was not changed until the parasite density reached at least 5.106 cells/ml. After the selection phase, single clones were generated by limiting dilution, expanded and assessed by either PCR on genomic DNA (using primers specific of the vector amplifying the resistance gene) or luciferase activity quantification or GFP fluorescence depending on the vector used.
Transgene induction was achieved by growing the cells in the presence of 1 µg/ml of tetracycline.
The Renilla Luciferase Assay System (Promega, Charbonnières-les-bains, France) was used to measure in vitro luciferase activity as described by the manufacturer. Briefly, non-transformed and pLEW20c-Rluc-transfected T. congolense clones were grown to a total of 5.106 parasites/ml and centrifuged (1500×g, 10 min). The pellet was washed with PBS, resuspended in 20 µl lysis buffer and added to 100 µl of the reaction mix. Renilla luciferase activity was monitored over 5 min every 10 sec after substrate addition by using an Optima microplate reader (BMG Labtech, Germany) and expressed as relative light units (RLU) per µg protein.
PCF were collected by centrifugation (1600×g, 10 min) of culture supernatants. Individual EMF were collected by brief washing of the adherent cells with PBS and gentle scraping of the colonies with a cell scraper followed by centrifugation (1600×g, 10 min). MCF were purified from culture supernatants by DE52 anion-exchange column chromatography (Whatman Plc., Brentford, U.K.) [19]. BSF were collected by extensive flushing of the culture to remove adherent cells followed by centrifugation (1600×g, 10 min). To observe EMF colonies, glass coverslips were placed in 24-well plates and incubated with EMF until formation of colonies (the process can take a few days). To analyze BSF/BAE interactions, BAE cells were first incubated for at least six hours on glass coverslips placed in 24-well plates to allow them to adhere before addition of the BSF and co-culturing for several hours. Cells were fixed with formaldehyde as described [21]. Slides were then incubated with mouse anti-paraflagellar rod (PFR) antibodies (anti-PFR2 L8C4 monoclonal antibody diluted 1∶100), kindly provided by K. Gull, (Oxford, U.K.)) in PBS 0.1% (vol/vol) Triton X-100 0.1% (wt/vol), BSA for 30 min, washed three times with PBS and incubated for 30 min with Alexa Fluor 488 goat anti-mouse IgG secondary antibody (diluted 1∶100) (Molecular probes). Finally, cells were incubated for 5 min with 1 mg/ml 4′,6′-diamidino-2-phenylindle (DAPI) and mounted in Vectashield (Vector Laboratories). Cells were observed with a Zeiss UV microscope and images were captured using a MicroMax-1300Y/HS camera (Princeton Instruments) and Metaview software (Universal Imaging Corporation).
Direct observations of parasites in culture were also conducted: cells were observed with an Axiovert (Zeiss) UV microscope and images were captured by an Axiocan camera (Zeiss) and Axiovision 3.0 software (Zeiss).
Total protein preparations of trypanosomes were obtained by lysis of live parasites with 2% (wt/v) SDS (10 µl per 106 parasites) and heating at 100°C in the presence of a protease inhibitor cocktail (complete Mini, EDTA-free; Roche Diagnostics, GmbH). 10 µl of this lysate was loaded per well and separated by SDS-PAGE (12%) before transfer onto polyvinyl difluoride (PVDF) membranes (Immobilon-P, Millipore) and processed for Western blotting as described previously [22]. Monoclonal antibodies (mAb) #491 and #3C6 were kindly provided by T. W. Pearson and B. Loveless (Victoria, Canada). mAb #491 recognizes a carbohydrate epitope present on the protease-resistant surface molecule of T. congolense and on other surface molecules of T. congolense and T. simiae [12],[23]. mAb #3C6 was made by B. Loveless against recombinant CESP procured by N. Inoue (B. Loveless and T. W. Pearson, unpublished data). Membranes were incubated for 2 hours with dilutions of hybridome culture supernatant (1∶10) or mouse anti-T. brucei tubulin (1∶2,000). Antigen-antibody interactions were revealed with the Immobilon Western chemiluminescent horseradish peroxidase substrate (Millipore).
Eight-week-old female Balb/c mice were purchased from Charles River Laboratories (L'arbresle Cedex France), NOD/SCID (NOD.Cg-Prkdescid Il2rgtm1 Wjl/Szj) were bred locally in specific pathogen-free conditions and used for experiments from five to six weeks of age. All animal studies adhered to protocols approved by the University of Bordeaux 2 animal care and use committee and the commission de genie genetique (Direction Generale de la Recherche et de l'Innovation). Balb/c mice either treated or not with the immunosuppressant cyclophosphamide (200 mg/kg, Sigma-Aldrich) or NOG mice were inoculated by intraperitoneal injection with variable amounts of infectious trypanosomes. Inoculated parasites were: (i) purified MCF, (ii) a mix of EMF and MCF from cell culture supernatants, (iii) BSF cultivated in vitro or from infected mouse blood. Parasitaemia was subsequently monitored daily by microscope observation. Blood samples were collected three times a week by tail bleed in 100 µl capillary tubes coated with Na-heparin. To determine the packed red blood cell volume percentage (PCV), capillary haematocrit tubes were sealed, centrifuged (15,000 rpm, 12 min) and analyzed with the haematocrit reader supplied by the manufacturer (Heraeus).
T. congolense PCF adaptation to an axenic culture system directly from infected mice blood was described previously [24]. However, with described methods and media, for most of the isolates, differentiation efficiency was low and required several months to attain stable culture. Furthermore, some strains could not undergo the adaptation process to axenic culture. Therefore, we aimed to standardize a simple method for rapid initiation from infected blood and subsequent stabilization of PCF in continuous culture.
Blood samples were collected from mice experimentally infected with various strains (IL3000, ILC-49, IL1180, TREU1457, TRUM183:13–29, STIB910). When the parasitaemia reached the mid log phase (∼5.106 cells/ml), a few drops of tail blood were placed in 1 ml of media (TcPCF-1, TcPCF-2 or TcPCF-3; Table 1) and inoculated in 24-wells plates at 27°C after discarding most of the red blood cells by centrifugation. TcPCF media are derived from the 109-c medium [24] with variation in serum nature and percentage, and presence of cis-aconitate and citrate.
BSF were always transformed into PCF within 2 days but degenerated cells (abnormal shape, giant or clustered cells), unable to divide, appeared soon after and settled at the bottom of the well. These abnormal cells constituted the major part of the culture during the first 3–10 weeks before giving way to well-shaped dividing PCF. This adaptation phase is the critical step. When the culture reached at least 5.106 cells/ml, the medium was changed with 1 ml of fresh medium. After a few days, the doubling time reached its maximum (8–15 hours depending on the strain) and the culture was transferred to 25 cm2 flasks and standard cultivation methods in TcPCF-3 medium were applied. The presence of 3 mM of cis-aconitate and citrate shortened the adaptation phase to at least 2 weeks, probably facilitating the differentiation process as described for T. brucei [25],[26]. Furthermore, this adaptation phase could only be conducted in TcPCF-2 (goat serum instead of fetal calf serum) for ILC-49 and TRUM183:13–29. In contrast, all the other tested strains preferred the TcPCF-1 medium. For the multiplication step, all strains grew faster in TcPCF-1. The length of time required to achieve a stabilized PCF culture in flask is reported in Table 3 for all the strains tested. We observed that even with this improved system, the time period required to achieve the process was strain dependant.
Transgene expression and RNA interference (RNAi) in T. congolense PCF have been reported previously [11],[27]. Specifically, the tetracycline inducible system first used in T. brucei (constitutive expression of RNA polymerase and tetracycline repressor, [5]) has been adapted for T. congolense by Inoue and colleagues in the TRUM183 uncloned serodeme [11]. Nevertheless, classical electroporation methods are not sufficient to ensure an acceptable success rate of transfection in several other strains (our data). For example, the most used and currently sequenced (Sanger Center) IL3000 strain is awkward for genetic manipulation (upon our data: less than 30% of electroporation assays lead to selection of stable transfectants). Because transfection efficiency could be greatly enhanced in the parasite T. brucei using the Amaxa nucleofection method [28], we applied this technology to T. congolense.
To test this method on T. congolense PCF, we tried 3 different nucleofection programs and 3 different Amaxa solutions recommended by the manufacturer with different vectors as described in material and methods: p2T7Ti/αTUB as RNAi control experiment, pLEW20c-EGFP and pLEW20c-Rluc allowing inducible transgene expression of EGFP, Renilla luciferase, pLEW13 and pLEW29c inserted successively in the IL3000 genome to devise a tetracycline inducible system. The new cell-line was named IL3000:13–29. Stable transfection achievement has been assessed by PCR on genomic DNA and by GFP fluorescence or luciferase activity measurement. Results are reported in Table 2. The best results were obtained with the Parasitic 2 Amaxa solution and the X-001 program: independently of the T. congolense strain used, nearly 100% of the transfection assays led to selection of stable transfectants. With the other programs, electroporation led to cell death at much higher levels. The inducible operating system was also valued as shown in Figure 1, 2 and Table 4. Despite the absence of tetracycline, a basal expression level of GFP fluorescence and luciferase activity was observed; this is probably the result of a leakage of the tetracycline inducible system, which has already been described previously [20],[29]. Upon addition of tetracycline, transgene expression was enhanced as indicated by the rise in fluorescence intensity (Figure 1) and the 4-fold increase of luciferase activity (Table 4). For the RNAi experiment, as expected, the α-tubulin gene expression generated an altered cell morphology called the FAT phenotype [30] (Figure 2). Without tetracycline, this phenotype is absent, after tetracycline addition, the entire population displayed the lethal FAT phenotype within 48 hours. The absence of apparent leakage in this experiment is probably due to the natural enrichment of the culture with more tightly repressed cells due to the lethality associated with the FAT phenotype. This hypothesis is reinforced by prolonged selection time for p2T7Ti/αTUB transfectants and decrease of the doubling time in the course of the selection process.
Cultivation of all the life cycle stages from the tsetse vector has been described previously [8],[9]. However, using these protocols, production of infective MCF from insect form cultures could take few weeks to several months depending on individual strains of T. congolense [13]. The differences lay in cytoadherence timing, interval between attachment, infective MCF appearance and in the number of MCF produced. In order to shorten the metacyclogenesis process. We aimed to improve the differentiation efficiency in terms of timing and standardization.
First, to asses timing differences among strains, we tried to perform metacyclogenesis using the described methods with the wild type strains IL3000, ILC-49, IL1180, TREU1457, STIB 910 and the genetically modified cell lines TRUM183:13–29 and IL3000:13–29 transfected with either pLEW20c-Rluc, pLEW20c-EGFP or p2T7Ti/αTUB. PCF were inoculated with 5.106 cells/ml in TcPCF-3 medium containing the optimal concentrations of glutamine and proline (2 mM glutamine and 10 mM proline) for metacyclogenesis as described by C. Ross [10]. These conditions were verified to be optimal (data not shown). After inoculation, half of the medium was changed every 2 days. To appreciate metacyclogenesis development, we considered several criteria to characterize the EMF and the MCF. (i) The first observed event was the EMF typical cytoadherence: parasites attached to the bottom of the flasks, grew and formed adherent bundles. (ii) Cell shape, size and kinetoplast and its attached flagella localization in relation to the nucleus change during metacyclogenesis process. These main morphological features were examined microscopically. For kinetoplast positioning, cells were labeled with a flagellar marker (anti-PFR antiserum) and stained with DAPI for orientation. (iii) MCF emergence was highlighted by purification on DE52 anion exchange chromatography. MCF infectivity assay in mice is described bellow. (iv) Stage specific markers were analyzed by western-blot, using two specific monoclonal antibodies. According to the literature, mAb #491 recognizes carbohydrate epitope shared by different surface molecules (such as PRS, GARP and CESP) differentially expressed during metacyclogenesis and mAb #3C6 binds to the EMF specific protein CESP [12],[23]. Results are presented in Figure 3 and Table S1. We confirmed that metacyclogenesis was successful as we chiefly observed the kinetoplast repositioning through the cycle (kinetoplast is always anterior to the nucleus in the EMF contrary to being on the posterior end in PCF and MCF) and MCF appearance within 1–2 weeks after cytoadherence. Furthermore, the expression pattern of stage specific markers is in agreement with published data on in vitro differentiated EMF: CESP is specifically induced in EMF stage and carbohydrate epitope presence is strongly increased during the PCF to EMF differentiation [12]. Nevertheless, we observed a large discrepancy among wild type strains with, for example, a required period of 3 to 10 weeks to observe cytoadherence for wild type strains. Moreover, differentiation was more difficult for genetically modified parasites since MCF getting required several months. For all the strains, the critical step seemed to be EMF differentiation.
Subsequently, we focused on the method to optimize this step by using the IL3000 strain. First, we attempted to influence the physiological state of the parasite by modifying the cell density of the culture. Unexpectedly, we observed a significant increase in the rate of cytoadherence (few days instead of weeks) when a density higher than 2.107 cells/ml was used, corresponding to the stationary phase. All the strains were subjected to this treatment and EMF differentiation followed by MCF differentiation was noted (results are reported in Table 3). With this treatment, all the strains were able to pass through metacyclogenesis and parasites fit all the stage specific features described above, but variations in EMF differentiation timing were observed. Conversely MCF always appeared within 2 weeks. One of the features of stationary phase is depletion of metabolites in the medium. Taking this into consideration, we used a medium containing only base powder, glutamine, proline and no serum, TcEMF-1. Surprisingly, in this medium, individual cells and bundles of adherent IL3000 cells appeared after 30 minutes and reached a maximum after 2 hours. Prolonged exposure time in the absence of serum resulted in cell death. Therefore, after 1–2 hours, we added 10% FCS in this medium to produce TcEMF-2. Half of the medium was then changed every two days and it was found that MCF always appeared after 1–2 weeks. Adherent cells induced by serum depletion (named iEMF for induced EMF) displayed the main EMF criteria (cytoadherence, shape, kinetoplast positioning and ability to differentiate in MCF) but exhibited differences in molecular markers (Figure 3B). Indeed, no signal was observed either with mAb #491 or #3C6, which implies that CESP is not expressed and the carbohydrate epitope is absent.
This rapid method was tested on all the strains and worked even for transgenic strains: iEMF always appeared within 1 hour and MCF within 1–2 weeks. To confirm the potential use of the process to conduct functional analysis through the life cycle, GFP fluorescence, luciferase activity and FAT phenotype were determined in transfected strains either induced or non induced (Figure 1, 2 and Table 4). As expected, the results are similar to those of PCF. This process provides the opportunity to test phenotypes rapidly and efficiently throughout the parasite's developmental stages of the tsetse fly.
To assess infectivity and virulence of in vitro MCF, we injected 107 parasites into Balb-c mice and measured the level of parasitaemia. The different strains were not equally infective for mice (Table 3). IL3000, TREU1457 and STIB910 produced an acute parasitaemia with a single peak, a high parasite load, a strong anaemia and death within 2 weeks (Figure S1). IL1180 resulted in a chronic infection with a fluctuating blood parasitaemia leading to a survival time greater than one month (Figure S1). As for ILC-49 and transgenic cell lines, MCF were not infective and consequently no parasites were observed in the blood.
Previous parasitological studies associated T. brucei virulence differences with the capacity of sub-cellular membrane fractions to induce immunosuppression [31]. For this reason, we used immunosuppressed mice (cyclophosphamide pretreated Balb-c) and immunodeficient mice (NOD/SCID) in order to by-pass the lack of infectivity and allow these strains to develop in the murine host (Figure 4). MCF resulted in parasitaemia only in immunodeficient mice with a long prepatent period (at least 1 month). On the other hand, once parasites appeared, we observed a high parasite load, development of anaemia and death within two weeks. Subsequently, 107 BSF containing blood was injected into new Balb-c mice pretreated or not with cyclophosphamide or into NOD/SCID mice. We obtained similar results as the parasitaemia only developed in NOD/SCID mice albeit with a shorter prepatent period (after 2 weeks). During the second passage, BSF were observed not only in NOD/SCID mice but also in immunosuppressed Balb-c mice. Following the third passage, blood samples were infective in immunocompetent Balb-c mice. For the fourth passage we observed the same infectivity and virulence through passages resulting in the development of acute parasitaemia with parameters comparable to those obtained with blood BSF stocks never cultivated in vitro (field isolates directly adapted to rodents) (Table 3).
To gain a better understanding of this phenomenon, we examined the possibility of a requirement of a host adaptation step for these non-infective MCF. Indeed, it has been described that MCF of the same T. congolense clone produced by tsetse fed with various host blood displayed virulence differences in mice and required an adaptation period to develop in mice [32]. Furthermore some T. vivax field isolates have been adapted to rodents using intraperitoneal inoculation of goat serum [33]. We therefore, injected 500 µl of goat serum intraperitoneally in immunosuppressed Balb-c mice one day before parasite infection and every subsequent day until BSF were observed in the blood. We observed parasites as soon as one week after MCF inoculation and mice eventually developed an acute parasitaemia. This protocol did not work with immunocompetent mice. The same treatment was required for the first passage in mice. From the second passage, goat serum injection was not essential and blood samples were infective in immunocompetent Balb-c mice. These data confirmed the existence in some strains of an adaptation process to the murine host.
We aimed to develop long-term cultures of T. congolense BSF in order to conduct drug trials and genetic modifications. We followed the described procedures to obtain BSF from infected mice blood [14]. Dividing BSF were easily obtained but parasites soon began to degenerate and the culture collapsed after 3 to 10 days depending on the strain. We investigated the consequences of varying media components such as serum and base powder. We inoculated standard T. brucei BSF medium containing either IMDM or MEM base powder and derived media containing 20% of various sera (horse, foal, lamb, goat) as described earlier [34]. The results showed a preference for goat serum and MEM base powder (data not shown). Supplementation of these media with “serum plus” increased cell density and attachment (TcBSF-1). We also examined temperature effect by cultivating parasites between 34°C and 37°C and observed higher longevity at 34°C. But despite these improvements, the culture conditions were not sufficient for long-term culture.
To go one step further with the culture system, we used endothelial cells as feeder cell layer. Indeed, strong interactions between vascular endothelium have been demonstrated in vivo [35],[36] and in vitro with BAE [37],[38]. Addition of BAE greatly improved the BSF condition as cells promptly interacted with BAE, proliferated faster and could be cultivated for 3–4 weeks. Nevertheless, this was not sufficient for the selection process of stable transfectants since some major drawbacks remained: cultures were very sensitive and preservation through freezing was not possible. Eventually, we managed to settle BSF adaptation by adding RBC lysate to the culture (TcBSF-2), and by using fresh goat serum provided directly after sampling thus avoiding alteration through preservation or treatment (chemical or irradiation) of the commercial serum (TcBSF-3). Therefore, the adaptation process had to be carried out in TcBSF-3, this phase lasted few weeks before stabilization of the culture. All the strains tested could be cultivated for at least a month but stabilization has been achieved only for IL3000, IL3000:13–29 and STIB 910. When the culture is stable, commercial goat serum can be used and RBC lysate and serum plus were no longer essential to sustain growth.
To test our BSF culture system, we compared growth of T. brucei and T. congolense in different media (Figure 5, Table S2). RBC lysate and haemoglobin slightly improved growth rates for both species and the absence of reducing agents affected T. brucei to a greater extent than T. congolense. The absence of BAE in T. congolense culture had a great impact on growth rate, and furthermore, the culture was more sensitive as it was not possible to inoculate with less than 2.105 cells/ml. It was also observed that in the presence of BAE T. congolense reached the stationary phase later and supported a higher cell density than T. brucei (2.107 cells/ml vs. 4.106 cells/ml).
In addition, we subjected T. congolense MCF to differentiation with this adaptation system in order to directly obtain BSF in vitro avoiding the murine step. We obtained similar results: all the strains differentiated in BSF and were cultivated for one month and we completely stabilized IL3000, IL3000:13–29 and STIB910 for long-term culture. Infectivity and virulence of in vitro cultured BSF were tested by inoculating mice with parasites cultured for several months. Identical results were obtained as for MCF (Table 3). To complete the in vitro cycle, the cultivated BSF were subjected to PCF differentiation as described earlier. All the strains tested passed through this step and gave rise to a stable PCF culture.
Finally, we cultivated transfectant strains and managed to observe the corresponding phenotype: GFP fluorescence, luciferase activity and FAT phenotype (Figure1, 2 and Table 4). This experiment proved unambiguously the ability to fulfill phenotype analysis of genetically modified cell lines during the complete life cycle of T. congolense.
In the same way as PCF, we used the Amaxa nucleofection technology to transfect BSF. It turned out that BSF transfection was very difficult. Results are summarized in Table 2. We tried to construct the inducible system based on the tetracycline repressor directly in BSF and managed to achieve the first step, which was the integration of the pLEW13 vector into the genome. However, we failed to insert the second vector, pLEW29c, despite numerous endeavors. We then attempted to perform a constitutive RNAi experiment or reporter gene expression. We obtained successful results with the pLEW20-Rluc vector and highlighted luciferase activity (Table 4). We also achieved inducible expression of luciferase by directly transfecting the IL3000:13–29 cell line produced from the in vitro cycle. Direct transfection of T. congolense BSF is therefore possible but remains arduous as it took several weeks and the success rate was low.
The development of a continuous culture system represents a major breakthrough in T. congolense research. Indeed, parasite biology studies imply understanding of its complex life cycle including differentiation, environment adaptation, signaling, virulence and pathogenesis, which become accessible once powerful genetic tools such as RNAi are exploitable throughout the life cycle. For the first time, we have conducted a systematic study, which allowed us to define the best in vitro culture conditions for each stage in terms of differentiation efficiency followed by long-term in vitro culture design. In this way, we succeeded in reproducing the complete in vitro life cycle allowing the study of genetically modified cell-line phenotypes throughout the cycle (Figure 6).
Firstly, we designed a new enriched adaptation medium which enabled the isolation of BSF directly from the blood of infected mice and their long-term in vitro culture, which makes them amenable to in vitro drug trials, preservation through freezing and transfection assays. However, transfection of BSF remains a difficult task as we were unable to perform direct RNAi experiments but did succeed in overexpressing the luciferase reporter gene. Secondly, we improved the efficiency of PCF transfections and set up a tetracycline inducible system for the T. congolense reference strain (IL3000, the strain used by the Sanger Center for the genome sequencing project). We validated this system by the inducible expression of reporter genes (GFP and Luciferase) and an inducible RNAi experiment targeting the essential tubulin gene as a control. Thirdly, we standardized metacyclogenesis techniques, not only on wild type isolates, but also on genetically modified parasites (either for transgene expression or RNAi) and the time needed to achieve the differentiation has been dramatically shortened to ∼10 days instead of months. Finally, BSF differentiation could either be achieved by infecting mice or directly in vitro. Then, PCF transfection followed by metacyclogenesis became an efficient, rapid and reliable way to study gene function in all the developmental stages as well as during the differentiation steps. For the first time, powerful tools such as reverse genetics (gene expression inactivation and characterization of resulting phenotypes), becomes now available for this devastating parasite.
Adaptation of trypanosomatids to rodents for experimental infections was essential to analyze the host-pathogen interaction [39]. Specifically the use of inbred mouse models have proven to be a valuable tool in pathology research [40]. Distinctness in mouse susceptibility to trypanosomiasis among the mouse strains is always observed. For example, for T. congolense and other trypanosome species, BALB/c mice (compared to C57Bl/6) are the most susceptible [41]–[43]. Discrepancy in virulence among T. congolense isolates has also been reported in ruminant livestock as well as in laboratory animals [44],[45]. To go one step further in the understanding of virulence mechanisms, it is essential to analyze the phenotype of genetically modified parasites during experimental infection. This is the reason why we assessed the impact of long-term in vitro culture and of minor genome modification on parasite infectivity and virulence. Results showed that BALB/c mice are less susceptible to transfected cell-lines as well as to 1 of 5 tested wild type savannah strains after a complete in vitro life cycle. This deficiency in infectivity can be overcome by using immunodeficient animals or goat serum injected immunosupresssed mice. Furthermore, infectivity levels return to normal after few passages in the murine host. The survival time of infected mice is always correlated with the parasitaemia peak arising and since the parasites developed, we concluded that disease parameters like anemia are not influenced by in vitro passage. These data indicate that while in some cases the ability of the parasite to multiply and be maintained in the host is lowered, parasite virulence is not affected. We propose that infectivity decrease relies on a host adaptation step. Development of parasitaemia in the mammalian host implies for the parasite to fit the new environment in terms of energy supplying and trypanocides conditions. Impairment of the immune system could allow a time period sufficient for the parasite to adapt to its new physiological state. Goat serum injections might provide the parasite with essential factors like nutriments and create a microenvironment promoting the emergence of host adapted parasites that are able to develop in mice blood. Such phenomenon has been suggested for T. vivax adaptation to laboratory rodents using multiple passages in irradiated rats and intraperitoneal injection of goat serum [33].
This idea is in harmony with the requirement of goat serum supplemented medium for T. congolense BSF in vitro culture. Composition of this serum must suit the specific needs for parasite growth. The concentrations of vitamins and related metabolites vary from one sera to another. Goat serum is one of the most commonly used to enhance cell multiplication [46],[47]. The greatest efficiency of fresh goat serum (not treated) reinforced the hypothesis that the presence of labile components is required for the parasite well-being. One can also notice that the 3 strains impaired for infectivity (ILC-49, TRUM183:13–29 and IL3000:13–29) are those which need goat serum to promote their in vitro differentiation in PCF. These findings illustrated that even for closely related strains (all are of the Savannah genetic group), subtle differences in physiological requirements might exist.
Another important factor to promote BSF in vitro growth is the presence of the erythrocyte lysate, which became dispensable after some time. Haemoglobin is a component of red blood cells and we showed that its presence in culture medium enhances the growth rate of the parasite cultures. The first culture media described for Trypanosoma spp. were always supplemented with blood lysate, then replaced by hemin [48]. However, neither haemoglobin nor hemin, are sufficient to ensure BSF adaptation from infected mice blood (data not shown). Hemin is also essential for trypanosomatids PCF differentiation and growth [24],[49]. The parasite's inability to synthesize heme explains this requirement, since heme containing proteins like cytochrome c [50], are essential for parasite viability. However, erythrocyte lysate might provide the parasite with other unknown components, which are essential during the adaptation phase.
Generally, T. congolense culture has to mimic the host environment. Thus, a high amino acid concentration in insect forms media reflects the high proline content in tsetse hemolymph subsequently used for energy metabolism of the parasite. An even more striking similarity is the use of BAE layer to cultivate BSF. Indeed, in infected animals, T. congolense BSF are found adhering to erythrocytes and to endothelial cells of the microvasculature by their flagellum [35],[36],[51]. This requirement is no longer essential after an adaptation phase. Finally, production of infective MCF requires strong adherence of EMF mimicking the bundles formed in the fly proboscis. However, parasite journey in fly midgut then in proboscis is complex and can't be fully reproduced in vitro. Especially coat antigens expression is refined and follows a defined kinetic in the different insect stages. For example, one of the major coat protein, GARP, is weakly expressed in early PCF, present in the midgut then downregulated before being strongly expressed in EMF isolated from proboscis [23]. In culture, this marker is expressed in both stages PCF and EMF, reflecting an artefactual absence of regulation during in vitro EMF differentiation [23]. Furthermore, Butikofer et al. [23] did not find reactivity with mAb #491 on EMF from tsetse, although Sakurai et al. [12] did find increased reactivity with EMF compared to PCF within in vitro culture. Similarly, the iEMF phenotype (absence of CESP expression and of the carbohydrate marker) may represent another illustration of the importance of the environmental factors on differentiation. Indeed, the difference observed in surface markers may result from the absence of serum at the time of differentiation or from the stress induced by the sudden environment change. Nevertheless, despite the absence of those markers iEMF are able to differentiate into infective MCF suggesting that those surface markers expression is not essential for in vitro EMF differentiation but must be involved in the in vivo process where parasites have to attach themselves to the wall of the labrum in the food canal to transform into EMF. The iEMF stage could also represent a transitory stage in the insect. Furthermore the absence of CESP expression does not hinder in vitro iEMF adhesion implying that this protein is not essential for cytoadherence and that other proteins might be involved. Successful culture and differentiation of iEMF and EMF combined with targeted gene deletion through RNAi represent powerful tools to understand the elaborated mechanism of the metacyclogenesis process encountered in the insect.
Parasites display a great ability to adapt but the required time period for this adaptation is the limiting step in the initiation of stable culture of the different developmental stages as well as for experimental infections. We succeeded to go through this limiting step to secure cultures of all the T. congolense developmental stages. Standardization of media and methods using various strains, more or less demanding, allowed the design of the complete in vitro lifecycle of all the tested strains. Definition of these optimal conditions also greatly improved the success rate of PCF nucleofection, that should be very useful for routine genetics based analysis.
To conclude, the ability to dominate the in vitro metacyclogenesis in combination with the transgenic PCF technique provides an essential tool to investigate the functional role of T. congolense genes throughout the cycle as well as in vivo during experimental infections in mice.
|
10.1371/journal.ppat.1001047 | Bacteriophage Lysin Mediates the Binding of Streptococcus mitis to Human Platelets through Interaction with Fibrinogen | The binding of bacteria to human platelets is a likely central mechanism in the pathogenesis of infective endocarditis. We have previously found that platelet binding by Streptococcus mitis SF100 is mediated by surface components encoded by a lysogenic bacteriophage, SM1. We now demonstrate that SM1-encoded lysin contributes to platelet binding via its direct interaction with fibrinogen. Far Western blotting of platelets revealed that fibrinogen was the major membrane-associated protein bound by lysin. Analysis of lysin binding with purified fibrinogen in vitro confirmed that these proteins could bind directly, and that this interaction was both saturable and inhibitable. Lysin bound both the Aα and Bβ chains of fibrinogen, but not the γ subunit. Binding of lysin to the Bβ chain was further localized to a region within the fibrinogen D fragment. Disruption of the SF100 lysin gene resulted in an 83±3.1% reduction (mean ± SD) in binding to immobilized fibrinogen by this mutant strain (PS1006). Preincubation of this isogenic mutant with purified lysin restored fibrinogen binding to wild type levels. When tested in a co-infection model of endocarditis, loss of lysin expression resulted in a significant reduction in virulence, as measured by achievable bacterial densities (CFU/g) within vegetations, kidneys, and spleens. These results indicate that bacteriophage-encoded lysin is a multifunctional protein, representing a new class of fibrinogen-binding proteins. Lysin appears to be cell wall-associated through its interaction with choline. Once on the bacterial surface, lysin can bind fibrinogen directly, which appears to be an important interaction for the pathogenesis of endocarditis.
| The binding of bacteria to human platelets is thought to be a central event in the development of endocarditis (a life-threatening cardiovascular infection). We have previously found that platelet binding by Streptococcus mitis is mediated by surface components encoded by a bacteriophage contained within the host bacterium. We now show that lysin (an enzyme of bacteriophage origin) contributes to platelet binding via its direct interaction with fibrinogen on the platelet surface. Lysin bound to purified fibrinogen in vitro, and this interaction specifically involved the Aα and Bβ chains of fibrinogen. Binding of lysin to the Bβ chain was further localized to a region within the fibrinogen D fragment. Disruption of the gene encoding lysin gene resulted in a significant reduction in binding to fibrinogen by S. mitis, as well as a major reduction in virulence, as measured by a rat model of endocarditis. These results indicate that lysin is a multifunctional protein, representing a new class of fibrinogen-binding molecules. Lysin is localized to the bacterial surface via its interaction with cell wall choline, where it then can bind fibrinogen directly. Cell surface lysin apparently also contributes to the development of endovascular infections via its previously unrecognized fibrinogen binding activity.
| The pathogenesis of infective endocarditis is a complex process, involving numerous host-pathogen interactions [1], [2]. A key interaction for disease establishment and progression is the binding of microbes to human components, including platelets, fibrinogen, fibrin, and fibronectin [3], [4], [5], [6], [7], [8]. Although this binding appears to be a central requirement for virulence, only a limited number of endocarditis-related adhesins has been identified [7], [8], [9].
Among the viridans group streptococci, Streptococcus mitis is a leading cause of endovascular infection [10], [11], [12], [13], [14]. Despite its increasing importance as a human pathogen, relatively little is known about the virulence determinants of this organism, particularly with regard to its interaction with platelets or other host components. Our previous studies identified two surface proteins (PblA and PblB) encoded by a lysogenic bacteriophage (SM1) that mediate the binding of S. mitis to human platelets, through their interaction with the membrane ganglioside GD3 [15], [16], [17]. Disruption of the genes encoding PblA and PblB results in a significant decrease in platelet binding in vitro, as well as a marked reduction in virulence, as measured by an animal model of endocarditis [16], [17].
Expression of these proteins on the bacterial surface is dependent upon the activities of phage holin and lysin, which permeabilize the cell envelope, thereby permitting the transport of PblA and PblB to the cell wall, where they attach to phosphocholine (PC) residues [16]. Of note, disruption of the gene encoding lysin (lys) resulted in a profound reduction in platelet binding, to levels that were significantly lower than those seen with either the parent strain, or a pblA/plbB double knock-out mutant [16]. These findings suggested that lysin mediates platelet binding in part through a mechanism independent of its role in the export of PblA and PblB.
For these reasons, we investigated the mechanisms by which lysin mediates binding to platelets, and whether this interaction contributes to the pathogenesis of streptococcal endocarditis. Our studies indicate that phage lysin can be localized on the bacterial surface through its interaction with PC residues. Surface-bound lysin can subsequently bind both free and platelet-associated fibrinogen, through its specific interaction with the Aα and Bβ chains of the protein. Loss of lysin expression is associated with reduced virulence in the setting of endocarditis, indicating that the binding of lysin to fibrinogen is an important factor in the pathogenesis of this infection.
Using the NCBI Conserved Domain Database (CDD) search system [18], bioinformatic analysis of the predicted amino acid sequence of lysinSM1 (Accession number Q9AF60), revealed that an amidase-5 domain (Pfam05382; amino acids 4–146) is present at the amino terminus, and a putative choline-binding domain is found at the carboxyl terminus (COG5263; amino acids 128–271; Fig. 1). The N-terminal domain of lysinSM1 (N-lysinSM1) exhibits 75% amino acid identity to the Pal lysin (accession number O03979) of the pneumococcal bacteriophage Dp-1, and 74% identity to the lysin (accession number Q8E0W3) of the prophage lambdaSa1 of Streptococcus agalactiae [19], [20]. The C terminus of lysinSM1 (C-lysinSM1) contains a choline-binding domain homologous to that found in the pneumococcal LytA autolytic enzyme (62% identity), which anchors the protein to PC residues present in LTA or teichoic acids [19].
To assess whether lysinSMl demonstrated its predicted biological activities, we first examined its binding to DEAE-cellulose, a property that is a hallmark of choline-binding proteins [21]. LysinSM1, N-lysinSM1, and C-lysinSM1 were expressed individually in Escherichia coli, and lysates from these strains were applied to a DEAE-cellulose column [16]. N-lysinSM1 failed to bind the matrix, with the protein being detected in the wash volumes (Fig. 2A). In contrast, lysinSM1 and C-lysinSM1 were only eluted with ™ buffer containing 2% choline chloride. Thus, lysinSM1 appears to be a choline-binding protein, with this interaction being mediated by the C terminus.
We then examined whether purified lysinSM1 bound directly to PC residues of LTA purified from S. pneumoniae HS0001 and S. mitis SF100. Purified FLAGlysinSM1 was incubated with immobilized LTAs, and binding was assessed by ELISA with anti-FLAG antibody. As shown in Fig. 2B, lysinSM1 bound LTA from S. mitis SF100 and from S. pneumoniae HS0001, both of which contain PC. Of note, binding levels of lysinSM1 to these LTAs were comparable and concentration-dependent. In contrast, little or no binding to LTA was detected from strain HS0001-EA, which has no PC. C-lysinSM1 also bound LTA from SF100 and HS0001, whereas N-lysinSM1 did not (Figure S1). These results confirm that lysinSM1 interacts with the PC residues of LTA, and that binding is mediated by the predicted choline-binding domain within the C terminus.
As mentioned above, analysis of the predicted amino terminus of lysinSM1 indicated that it encodes an amidase with g-D-glutaminyl-L-lysin endopeptidase activity [20]. To assess its lytic activity, we tested the bactericidal properties of lysinSM1 in vitro (Fig. 2C). When compared with organisms treated with buffer alone, exposure of S. pneumoniae HS0001 to lysinSM1 resulted in a mean (± S.D.) reduction of 5.07±1.28 log10 CFU per ml (P<0.05). LysinSM1 was also active against the SM1 host strain SF100 and its isogenic variant, PS1006, though it only reduced mean titers by 0.8±0.04 (P<0.05) and 1.65±0.21 log10 CFU per ml (P<0.05), respectively. No bactericidal activity was seen when tested against Staphylococcus aureus, Streptococcus sanguinis, Streptococcus pyogenes, or E. coli. Of note, neither purified N-lysinSM1 nor C-lysinSM1 had bactericidal activity against strains HS0001 or SF100 (data not shown). Thus, lysinSM1 has lytic activity against PC positive strains, such as S. mitis and S. pneumoniae, though the latter species is considerably more sensitive to the enzyme. Moreover, lysinSM1 requires its choline-binding domain, in addition to its predicted amidase domain, for this activity.
We have previously observed that disruption of the gene encoding lysinSM1 resulted in a significant reduction in platelet binding by S. mitis [16]. To assess whether lysinSM1 could interact directly with human platelets, we evaluated the binding of FLAGlysinSM1 to immobilized human platelets and to isolated platelet membranes (Fig. 3A). FLAGlysinSM1 was incubated with platelet monolayers or platelet membranes, and bound FLAGlysinSM1 was then detected with anti-FLAG antibody. When tested by this approach, we found that FLAGlysinSM1 strongly interacted with both whole platelets and platelet membranes in a concentration-dependent manner. In contrast, no binding of FLAGlysinSM1was seen to wells coated with only a casein-based blocking reagent (Western Blocking Reagent; Roche).
To identify the membrane receptor for lysinSM1, we assessed by far Western blotting the binding of FLAGlysinSM1 to platelet membranes that had undergone SDS-PAGE and transfer to nitrocellulose (Fig. 3B). Although the platelet membrane extracts contained numerous proteins, ranging in mass from 50 to 250 kDa, FLAGlysinSM1 bound only a small number of proteins. The highest levels of binding were seen with two proteins of MW 65 kDa and 55 kDa, which were similar to the molecular masses of the Aα and Bβ chains of human fibrinogen (64 and 56 kDa), respectively. To confirm that platelet membrane extracts contained fibrinogen, the preparations were probed with antibodies directed against the three major chains of fibrinogen (Aα, Bβ and γ). As shown Fig. 3B, each subunit of fibrinogen was present in the membrane extracts. To directly confirm that lysinSM1 bound fibrinogen on platelet membranes, we assessed whether lysinSM1 binding to immobilized platelet membranes was inhibited by anti-human fibrinogen IgG. As shown in Fig. 3C, pre-treatment of membranes with 30 or 100 µg/ml of anti-fibrinogen antibodies significantly reduced subsequent lysin binding. These results further indicate that fibrinogen is the principal component on platelet membranes and that lysinSM1 is bound to platelet membranes through it interaction with fibrinogen.
Since fibrinogen is a key factor in the pathogenesis of infective endocarditis, and because it is a receptor for some bacterial adhesins [7], [22], [23], [24], [25], [26], we further investigated the interaction of this protein with lysinSM1. We first assessed the binding of the increasing concentrations of FLAGlysinSM1 to human fibrinogen (3 µg/ml) immobilized in microtiter wells. In control studies, no significant binding of fibrinogen by FLAG-tagged alkaline phosphatase (FLAGAP) was detected (Figure S2). In contrast, FLAGlysinSM1 showed significant binding to immobilized fibrinogen, which increased in direct proportion to the amount of protein applied (Fig. 4A). At concentrations above 125 µg/ml FLAGlysinSM1, binding reached a plateau, indicating that it was saturated. In addition, the binding of FLAGlysinSM1 to immobilized fibrinogen was effectively blocked by both unlabeled lysinSM1 and fibrinogen (Fig. 4B). Similar levels of lysin binding were seen with rat fibrinogen, the host used for our subsequent virulence assays (Figure S3).
For some bacteria, binding to fibrinogen is dependent on whether the protein is in solution or immobilized on a surface. For example, Group A and G streptococci can bind both soluble and immobilized forms of fibrinogen, whereas several oral streptococci appear to bind only immobilized fibrinogen [27], [28], [29]. To assess whether fibrinogen binding by lysinSM1 was phase-dependent, we reversed the binding conditions, such that FLAGlysinSM1, untagged lysinSM1, and FLAGAP (all at 10 µg/ml) were immobilized in microtiter wells, and probed with the increasing concentration of fibrinogen in solution. Under these conditions, fibrinogen was still found to bind lysinSM1 and FLAGlysinSM1 comparably, whereas no significant binding to FLAGAP was detected (Fig. 4C).
As noted above, we found that lysinSM1 bound two proteins associated with platelet membranes that corresponded to the Aα and Bβ chains of fibrinogen. To confirm that lysinSM1 binds specifically to these subunits, we assessed by far Western blotting the interaction of lysinSM1 with purified human fibrinogen (Fig. 4D). When separated by SDS-PAGE under reducing conditions, fibrinogen appeared as three bands, having the expected masses. When transferred to nitrocellulose and probed with FLAGlysinSM1, binding could be detected to the Aα and Bβ chains only, confirming the results seen with platelet membranes.
The fibrinogen molecule is comprised of two subunits, each containing three polypeptide chains (Aα, Bβ, and γ; Fig. 5A). Cleavage of fibrinogen with plasmin produces a series of fragments, most notably the E fragment containing the central part of the molecule, and the D fragment containing the terminal regions. To further identify the domains of fibrinogen bound by lysinSM1, we examined the interaction of FLAGlysinSM1 to the D and E fragments. When assessed by ELISA, FLAGlysinSM1 showed high levels of binding to immobilized fragment D, which were comparable to those seen with whole fibrinogen (Fig. 5B). In contrast, no significant binding to the E fragment was seen.
Purified fibrinogen fragment D contains three subunits, each representing a part of the three major chains (α chain fragment = 15 kDa, β chain fragment = 44.5 kDa, and γ chain fragment = 42 kDa) (Fig. 5A). To further identify the subdomains of fibrinogen bound by lysinSM1, purified fragment D was separated under reducing conditions and transferred to a nitrocellulose membrane. When assessed by far Western blotting, binding by FLAGlysinSM1, was limited to the Bβ chain component of fragment D with no binding detected to the Aα chain (Fig. 5C). These data indicate lysinSM1 binds a region contained within AA 134–461 of the Bβ chain. Of note, lysinSM1 bound the full-length Aα chain (Fig. 3B), but not its D or E fragments (Fig. 5C), suggesting that the lysinSM1 binding to the Aαchain requires the C terminus (AA 197–610).
To assess the impact of lysin expression on bacterial binding to fibrinogen, we compared the adherence of SF100 (WT) and PS1006 (Δlysin) to fibrinogen immobilized in microtiter wells. As shown in Fig. 6A, SF100 had high levels of binding to immobilized fibrinogen, which increased in proportion to the amount of fibrinogen in the wells. PS1006 showed markedly reduced levels of binding, as compared with the parent strain. For example, when tested with wells coated with 30 µg/ml of fibrinogen, PS1006 had only 18.8±4.7% (mean ± SD) of maximal binding, as compared with 89.7±12.8% for SF100 (P<0.05, unpaired t-test). Complementation of the lysin mutation in trans restored fibrinogen binding by PS1006 (Fig. 6B), thereby demonstrating that the loss of binding observed with lysin disruption was not due to polar or pleiotropic effects.
The above results suggested that the binding to immobilized fibrinogen by SF100 is mediated by lysinSM1 expressed on the bacterial surface. To confirm that lysin was sufficient to mediate fibrinogen binding, we next examined whether exogenous lysinSM1 could attach to the cell wall of PS1006 and restore binding. The PC-negative strain SK598 served as a negative control. Each strain was incubated with purified FLAGlysinSM1 at RT for 30 min. After washing to remove nonspecifically bound protein, cell wall bound FLAGlysinSM1 was extracted with 2% choline, and the amount of FLAGlysinSM1 recovered was assessed by Western blotting. As shown in Fig. 6C, exogenous FLAGlysinSM1 could readily be detected in the cell wall extracts of PS1006, whereas no binding of FLAGlysinSM1was observed with SK598.
We then assessed whether this interaction was sufficient to enhance the binding of PS1006 to fibrinogen (Fig. 6D). PS1006 was suspended in PBS containing a range of concentrations of purified lysinSM1 and then tested for its binding to immobilized fibrinogen, as described above. As expected, PS1006 incubated in PBS alone showed minimal levels of binding to fibrinogen. This was not due to a loss of PblA and PblB expression, since the pblA/pblB negative strain PS344 had levels of fibrinogen binding that were similar to those of the parent strain. Exposure of PS1006 to FLAGlysinSM1 increased fibrinogen binding in a concentration-dependent manner. Indeed, 10 µg per ml of FLAGlysinSM1 was sufficient to restore PS1006 binding to levels comparable to those seen with SF100.
To assess the impact of lysinSM1 on pathogenesis, we compared the relative virulence of SF100, PS344 and PS1006 in a rat co-infection model of infective endocarditis [16], [55]. We first compared SF100 with PS344 to confirm previous results obtained in a rabbit model of infection [16]. As was observed with rabbits, disruption of pblA and pblB was also associated with attenuated virulence in rats, with PS344 having significantly reduced levels of bacteria within all tissues (Table 1). Disruption of lysinSM1 also produced a significant reduction in virulence. Rats co-infected with SF100 and PS1006 had significantly lower densities of the lysin mutant strain in vegetations (mean ± SD = 5.07±1.50 log10 CFU/g) as compared with the parent strain (6.91±1.35 log10 CFU/g; n = 8, P = 0.009). Densities of PS1006 were also significantly reduced in kidneys (P = 0.008) and spleens (P<0.001) as compared with SF100. We then examined the relative impact on virulence of abrogated PblA and PblB expression, versus loss of lysin (Table 1). In animals co-infected with PS344 and PS1006, titers of the latter mutant were significantly reduced in all tissues examined, as compared with the former. In particular, the mean densities of PS1006 within vegetations (6.59±1.45 log10 CFU/g) were significantly lower than those of PS344 (8.32±0.76; n = 8; P = 0.008), as were densities within kidneys (P = 0.027) and spleens (P = 0.006). We then re-analyzed these data by comparing the ratio of PS344 to PS1006 within tissues, with the CFU of each strain normalized to the number of CFU within the inoculum (competition index) (Figure S4). When assessed by this approach, the levels of the lysinSM1 mutant PS1006 remained significantly reduced in all tissues, as compared with PS344. Thus, lysinSM1 appears to be a significant virulence determinant in the setting of infective endocarditis. Moreover, its role in pathogenesis is not due solely to any effect it may have on PblA and PblB expression. Instead, it appears to have an impact upon the development of infective endocarditis independent of these other phage-encoded proteins.
The binding of pathogenic bacteria to platelets is thought to play a key role in the pathogenesis of infective endocarditis. This interaction may be important both for the initial attachment of bacteria to the endocardial surface, and for the subsequent formation of vegetations. Numerous endocarditis-associated pathogens have been shown to bind platelets directly in vitro, through a variety of mechanisms [3], [4], [7], [8], [25], [30]. The ability to bind platelets in vitro has been linked to virulence for several of the most common endocarditis-associated species, including Staphylococcus aureus, Streptococcus gordonii, and Streptococcus sanguinis [5], [31], [32], [33]. Previous work from our laboratory has shown that platelet binding by S. mitis strain SF100 is mediated in part by two proteins (PblA and PblB) encoded by the lysogenic bacteriophage SM1 [16]. The functional localization of these proteins to the cell surface requires the phage lysin (lysinSM1), which permeabilizes the host organism, thereby permitting the transport of PblA and PblB from the cytoplasm to the bacterial surface, and their subsequent attachment to the cell wall [16]. During the course of these studies, we noted that disruption of the gene encoding lysinSM1 reduced platelet binding in vitro more profoundly than the loss of PblA and PblB localization, indicating that lysin had a role in platelet binding beyond facilitating PblA and PblB transport. It was unknown, however, whether lysin itself could directly mediate binding, or rather, the effects of lysin on bacterial permeability led to the surface expression of other proteins (either phage or bacterial) that could enhance platelet binding.
Our current results demonstrate that lysin can bind human platelets directly through its interaction with fibrinogen. Purified lysin was found to bind fibrinogen, regardless of whether the proteins were in solution or immobilized. The binding of lysin with fibrinogen also was saturable, consistent with a receptor-ligand interaction. Lysin binding was restricted to the D fragment of the Aα and Aβ chains, further indicating that this is a specific process. This interaction appears to be important for the binding of S. mitis to fibrinogen, since disruption of the gene encoding lysin markedly reduced fibrinogen binding by bacteria in vitro. The addition of exogenous purified lysin to these mutants restored binding to WT levels, confirming that lysin can directly mediate the interaction of S. mitis with fibrinogen.
A number of other bacterial proteins have been shown to bind fibrinogen, including the M protein and serum opacity factor of Streptococcus pyogenes, FbsA of Streptococcus agalactiae, SdrG of Staphylococcus epidermidis, and several proteins of Staphylococcus aureus (clumping factors A and B, fibronectin binding protein A) [4], [22], [25], [30], [34]. However, none of these proteins exhibit any primary sequence homology with lysinSM1. The staphylococcal autolysins Aaa and Aae do resemble lysinSM1, in that they appear to have both enzymatic and fibrinogen binding activities in vitro [35], [36]. A search against the SMART and Pfam databases indicates that collectively these proteins belong to the NlpC/P60 superfamily of proteins, containing their catalytic domain that are characteristic of this group of proteins. However, the predicted catalytic activity of lysinSM1 (amidase 5) is different from that autolysins Aaa and Aae. LysinSM1 has no sequence similarity to either the Aaa or Aae protein, and unlike these other proteins, it is a choline-binding protein. Thus, lysin appears to be a multi-functional protein that can mediate S. mitis binding to fibrinogen, in addition to its role in the transit of the PblA and PblB proteins to the cell surface.
LysinSM1 was also associated with increased virulence in a rat model of infective endocarditis. When animals were co-infected with the parent strain SF100 and the lysinSM1 mutant PS1006, densities of the lysin mutant were significantly reduced within vegetations, kidneys, and spleens, as compared with the parent strain. Moreover, the virulence of PS1006 was also attenuated, when compared with its pblA and pblB-deficient isogenic variant, PS344. These results indicate that, beyond its importance for PblA and PblB expression, lysin contributes to virulence through a mechanism beyond its role in the transport of these bacteriophage-encoded adhesins. It is possible that there are other, unrecognized phage-encoded virulence factors that require lysin for export or localization. However, in view of the ability of lysin to bind fibrinogen directly (both human and rat), and that fibrinogen binding has been associated with virulence for several other adhesins [37], [38], [39], [40], [41], it is likely that this interaction of lysin with fibrinogen contributes to the pathogenesis of infective endocarditis by S. mitis. Given that lysin-fibrinogen binding enhances bacterial adherence to platelets in vitro, and that bacterium-platelet binding has been linked to virulence, it is likely that lysin-mediated binding to platelets via fibrinogen is an important pathogenetic interaction. However, it is also possible that lysin mediates streptococcal binding to fibrinogen on other surfaces, such as damaged endothelium. Finally, it is conceivable that lysin contributes to virulence through other, as yet unidentified interactions.
In summary, lysin is a novel fibrinogen-binding protein encoded by a lysogenic bacteriophage of S. mitis. In addition to its expected role in cell wall degradation, lysin also appears to be an adhesin mediating the attachment of this organism to human platelets, through its interaction with cell wall PC, fibrinogen, and the platelet membrane receptor for fibrinogen, glycoprotein IIb/IIIa (Fig. 7). Lysin also appears to contribute significantly to virulence, which could explain the persistence of certain bacteriophages within their host organisms. Although induction of the phage lytic cycle extracts a toll on host viability, in vivo this may be more than offset by the enhanced virulence resulting from lysin expression. Since fibrinogen is also present within gingival crevicular fluid, lysin-fibrinogen binding may also contribute to the colonization of oral surfaces by S. mitis [42], [43]. Although we do not know the exact prevalence of lysinSM1 homologs in other organisms, recent studies of S. pneumoniae and Enterococcus faecalis indicate that lysogenic bacteriophages encoding homologs of PblA and PblB are often present within these species [44], [45]. Since lysins are required for the phage life cycle, these findings suggest that homologs of lysinSM1 may also be encoded by such prophages. If so, then lysin binding to fibrinogen could prove to be an important interaction for a range of Gram-positive pathogens.
Blood was obtained from healthy human volunteers, using a protocol approved by the Committee on Human Research at the University of California, San Francisco. All human studies were conducted according to the principles expressed in the Declaration of Helsinki. Written informed consent was obtained from all study participants prior to their participation. All procedures involving rats were approved by the Los Angeles Biomedical Research Institute animal use and care committee, following the National Institutes of Health guidelines for animal housing and care.
N terminal Met-FLAG-alkaline phosphatase (FLAGAP) and purified rat fibrinogen were purchased from Sigma-Aldrich. Purified human fibrinogen and the fibrinogen fragment D and E (produced by cleavage of fibrinogen with plasmin) were obtained from Haematologic Technologies. Rabbit anti-human fibrinogen polyclonal IgG was purchased from Innovative Research.
Genomic DNA was isolated from SF100, using Wizard Genomic DNA purification kits (Promega), according to the manufacturer's instructions. Polymerase chain reaction (PCR) was performed with the primers listed in Table S2. To clone lys gene into E. coli expression vector, PCR products were purified, digested, and ligated into pET28FLAG to express FLAG-tagged versions of full length lysinSM1 (amino acids [AA] 1–295), the amino terminus of lysinSM1 (AA 1–158; N-lysinSM1), or the carboxy terminus of lysinSM1 (AA 141–295; C- lysinSM1). Untagged lysinSM1, C-lysinSM1, and His-tagged N-lysinSM1 (HisN-lysinSM1) were cloned into pET22b(+) (Novagen). The plasmids were then introduced to E. coli BL21(DE3) by transformation. LysinSM1, FLAGlysinSM1, C-lysinSM1 and FLAGC-lysinSM1 were purified with DEAE-cellulose columns, as described previously [16]. FLAGN-lysinSM1 and HisN-lysinSM1 were purified by either Ni-NTA (Promega) or anti-FLAG M2 agarose affinity chromatography (Sigma-Aldrich), according to the manufacturers' instructions.
A gene replacement cassette was constructed by cloning the chromosomal regions flanking lys upstream and downstream of the cat gene in pC326 [16]. A 339 bp upstream segment was amplified using primers KO4F and KO4R, and then digested with XhoI and HindIII. A 513 bp downstream segment was amplified with primers KO6F and KO6R, and then digested with EcoRI. The upstream and downstream fragments were cloned sequentially into the corresponding sites of pC326. The resulting plasmid, pKO-lys, was introduced into SF100 by natural transformation as previously described [17]. In brief, overnight SF100 cultures were diluted 100-fold in fresh THB supplemented with 20% heat-inactivated horse serum, 200 ng/ml competence-stimulating peptide (CSP; DWRISETIRNLIFPRRK), and 1 µg/ml of plasmid. Transformation mixtures were incubated 4 h at 37°C and then plated on blood agar containing 5 µg chloramphenicol per ml. To complement in trans the lys mutation in PS1006, lys was amplified using primers 3206-XbaI and 5206-EcoRI and then cloned into the streptococcal expression vector pDE123. This plasmid was derived from pDC123 by replacing the chloramphenicol resistance marker with an erythromycin resistance marker [46]. The resulting plasmid, pDE-lys, was introduced into PS1006 by natural transformation.
The bacteria and plasmids used in this study are listed in Table S1. S. mitis strains were grown in Todd-Hewitt broth (Difco) supplemented with 0.5% yeast extract (THY). PS344 (ΔORF47-PblB::pVA891) and PS1006 (ΔlysinSM1) are isogenic variants of S. mitis SF100, which is an endocarditis-associated clinical isolate [16]. All three strains grow comparably well in vitro. S. pneumoniae strains were grown in either a chemically defined medium (CDM; JRH bioscience) [47] supplemented with 0.1% choline chloride, or THY. S. pneumoniae HS0001 is a nonencapsulated pneumococcal strain derived from the TIGR4 strain by deleting the capsule synthesis locus as described previously [48]. S. pneumoniae HS0001-EA is a PC-negative strain derived from HS0001 as described previously [49]. Escherichia coli DH5α and BL21(DE3) strains were grown at 37°C under aeration in Luria broth (LB; Difco). Appropriate concentrations of antibiotics were added to the media, if required.
Transformed E. coli BL21(DE3) cells were harvested by centrifugation, washed and suspended in 50 mM Tris-maleate (™) buffer (Sigma-Aldrich), pH 6.3. Cells were disrupted by treatment with B-PER lysis solution (Pierce, Rockford, IL) and the debris was removed by centrifugation at 4,000 rpm for 10 min at 4°C. Supernatants were loaded on a 2 ml DEAE-cellulose (Sigma-Aldrich) column equilibrated with 50 mM ™ buffer, pH 6.3. The column was washed with at least 3 volumes of 50 mM ™ buffer, pH 6.3, containing 1.5 M NaCl and 0.1% choline chloride, until no protein was detected in the eluent. The retained proteins were then eluted with 50 mM ™ buffer, pH 6.3, containing 1.5 M NaCl and 2% choline chloride. Recombinant protein was dialyzed against PBS and then stored at −70°C.
Early log phage (A600 = 0.5) bacteria were harvested by centrifugation and suspended in PBS at approximately 108–109 CFU/ml. Bacteria samples were then incubated with or without 30 µg/ml of purified lysinSM1 at 37°C for 30 min. Samples were serially diluted in PBS and plated onto blood agar, to determine the number of surviving bacteria.
Platelet membranes were prepared by glycerol lysis and gradient centrifugation, as previously described [50]. In brief, isolated human platelets were lysed in 5 volumes of lysis buffer (8.5 mM Tris-Cl, 96.5 mM NaCl, 85.7 mM glucose, 1 mM EDTA, 10 mM EGTA [pH 7.4]) containing Complete Protease Inhibitor Cocktail (Roche). The sample was centrifuged (5,900× g, 10 min) to remove unlysed platelets, and the supernatant was applied to a sucrose step gradient (10 ml of 33% sucrose on 5 ml of 66% sucrose in buffer). After ultracentrifugation (90 min, 63,000× g, 4°C), the membranes were removed, dialyzed against PBS containing 10% glycerol, and stored at −70°C.
Samples were separated by electrophoresis through 4–12% NuPAGE Bis-Tris gels (Invitrogen) and transferred onto nitrocellulose membranes. The membrane were treated with a casein-based blocking solution (Western Blocking Reagent; Roche) at room temperature, and then incubated for 1 h with FLAGlysinSM1 (5 µg/ml) or purified human fibrinogen (1 µg/ml) suspended in PBS-0.05% Tween 20 (PBS-T). The membranes were then washed three times for 15 min in PBS-T, and bound probe proteins were detected with mouse anti-FLAG monoclonal antibody (Sigma-Aldrich) or rabbit anti-fibrinogen polyclonal IgG antibody.
Washed, fixed human platelets or purified platelet membranes were immobilized in 96 well microtiter plates as described previously [51]. To reduce non-specific adherence, the wells were then treated with the casein-based blocking reagent for 1 h at room temperature. The blocking solution was removed by aspiration, and the wells were incubated with 0 to 100 µg of FLAGlysinSM1 in PBS for 1 h, at RT, followed by washing to remove unbound protein. Bound FLAGlysinSM1 was detected by ELISA with anti-FLAG antibody. For some studies, the wells containing platelet membranes were pretreated with 0 to 100 µg/ml of rabbit anti-fibrinogen antibody for 30 min, followed by washing to remove unbound antibody. Binding by FLAGlysinSM1 (5 µg/ml) was then assessed as described above.
Rat fibrinogen (10 µg/ml), human fibrinogen, or human fibrinogen D or E fragments (all 15 nM in PBS), were immobilized in 96-well microtiter dishes by overnight incubation at 4°C. The wells were washed twice with PBS and blocked with 300 µl of casein-based blocking solution for 1 h at room temperature. The plates were washed three times with PBS, and a range of FLAGlysinSM1 concentrations in PBS with Tween 20 (0.05%) were added. The plates were then incubated for 2 h at 37°C. Unbound protein was removed by washing with PBS, and plates were incubated with mouse anti-FLAG antibodies for 1 h at 37°C. Binding was assessed by ELISA, using HRP-conjugated rabbit anti-mouse IgG, for 1 h at 37°C. FLAGAP (25–100 µg/ml) served as a control for nonspecific binding.
To examine the binding of fibrinogen to immobilized FLAGlysinSM1, untagged lysinSM1, or FLAGAP (10 µg/ml in PBS) were immobilized in 96 well microtiter plates, followed by blocking of the wells with the casein blocking solution. The wells were incubated with a range of human fibrinogen for 1 h at room temperature, followed by washing. Bound fibrinogen was detected by ELISA, using anti-human fibrinogen IgG.
Cultures of PS1006 and S. mitis SK598 in the early log phage of growth (A600 = 0.5) were harvested by centrifugation and suspended in PBS. The bacteria were incubated with purified FLAGlysinSM1 (0 to 10 µg/ml) for 30 min at room temperature. The samples were washed twice with PBS to remove unbound FLAGlysinSM1 and incubated with PBS-2% choline chloride to elute choline-binding proteins from the cell walls, as described previously [21]. Eluted cell wall proteins were harvested by centrifugation and loaded onto SDS-PAGE. Cell wall bound FLAGlysinSM1 was detected by western blotting with anti-FLAG antibody.
Overnight cultures of S. mitis SF100 or its isogenic mutants (PS1006 and PS344) were diluted 1∶10 in fresh THY broth, incubated for 1 h at 37°C, and then exposed to UV light (λ = 312 nm) for 3 min, to induce the expression of the lysogenic bacteriophage SM1. The cultures were then incubated at 37°C for an additional 2 h, followed by harvesting by centrifugation. The pellets were suspended in PBS, and adjusted to a concentration of 106 CFU/ml. One hundred microliters of each suspension were added to wells that had been coated overnight with 30 µg/well of fibrinogen in carbonate buffer. The plates were incubated at room temperature for 1 h, and the wells were washed three times with PBS to remove nonadherent bacteria. The wells were then treated with 50 µl of trypsin (2.5 mg/ml) for 30 min at 37°C to release the bound bacteria. The number of bound bacteria was determined by plating serial dilutions of the recovered bacteria onto blood agar.
LTA was prepared from S. pneumoniae HS0001 and S. mitis strains by organic solvent extraction and octyl-Sepharose chromatography, as previously described [52]. In brief, bacteria were cultured at 37°C for 10 h in CDM with 0.1% choline chloride (Fisher scientific Inc.). To purify PC negative LTA, S. pneumoniae HS0001-EA was cultured for 16 h in CDM supplemented with 2% ethanolamine. Pelleted bacteria were suspended in 0.05 M sodium acetate (pH 4.0) and lysed by sonication. After extraction from the lysate with a chloroform and methanol mixture (1∶0.9), the LTA was adsorbed onto an octyl-Sepharose CL-4B (Sigma-Aldrich) equilibrated in a mixture of 15% n-propanol and 0.05 M sodium acetate (pH 4.7). The absorbed LTA was then eluted with 35% n-propanol in 0.05 M sodium acetate (pH 4.7).
Purified LTA was analyzed by matrix-assisted laser desorption ionization-time of flight (MALDI-TOF) mass spectrometry [52] (Figure S5). In brief, 1 µl of a sample (1 µg/ml) and 1 µl of matrix solution (0.5 M 2, 5-dihydroxybenzoic acid and 0.1% trifluoroacetic acid in methanol) were applied to a sample plate. After drying, the sample was analyzed with a mass spectrometer (Voyager Biospectrometry DE Pro workstation; PerSeptive Biosystems). Purified LTA showed three major peaks that corresponded to LTA with five, six, and seven repeating units, respectively. The mass difference between the major peaks was 1299 or 1100 amu, corresponding to an oligosaccharide repeating unit with two PC groups or two phosphoethanolamine groups [52]. In addition, PC expression by strains HS0001 and SF100 was directly assessed by western blotting with anti-PC antibody (TEPC-15; Sigma-Aldrich) [53] (Figure S6).
The relative virulence of SF100 and its isogenic variants was compared in a competition model of infective endocarditis, as described previously [16], [54]. In brief, Sprague-Dawley female rats (250 to 300 g each) were first anesthetized with ketamine (35 mg/kg) and xylazine (10 mg/kg). A sterile polyethylene catheter was surgically placed across the aortic valve of each animal, such that the tip was positioned in the left ventricle, to induce the formation of sterile vegetations (nonbacterial thrombotic endocarditis). The catheters were left in place throughout the study. Seven days post-catheterization, the animals were infected intravenously with an inoculum of 105 CFU containing a 1∶1 mixture of a) SF100 and PS344, b) SF100 and PS1006, or c) PS344 and PS1006. At 72 hr post-infection, the rats were euthanized with thiopental (100 mg IP). Animals were included in the final analysis only if the catheters were correctly positioned across the aortic valve at the time of sacrifice, and if macroscopic vegetations were visible. All cardiac vegetations, as well as samples of the kidneys and spleens, were harvested, weighed, homogenized in saline, serially diluted, and plated onto 8% Todd Hewitt agar (±2.5 µg/ml of chloramphenicol or 5 µg/ml of erythromycin) for quantitative culture. The plates were cultured for 48 h at 37°C, and bacterial densities were expressed as the log10 CFU per gram of tissue. Differences in means were compared for statistical significance by the paired t-test. The data were also analyzed by calculating a “competition index,” which was defined as the ratio of the paired strains within tissues for each animal, normalized by the ratio of organisms in the inoculum. The mean of the log10 normalized ratios was tested against the hypothesized ‘no effect’ mean value of 0, as described previously, using a paired t-test, with P<0.05 as the threshold for statistical significance [55].
Data expressed as means ± standard deviations were compared for statistical significance by the paired or unpaired t test, as indicated.
|
10.1371/journal.pgen.1002604 | Gene Reactivation by 5-Aza-2′-Deoxycytidine–Induced Demethylation Requires SRCAP–Mediated H2A.Z Insertion to Establish Nucleosome Depleted Regions | 5-Aza-2′-deoxycytidine, approved by the FDA for the treatment of myelodysplastic syndrome (MDS), is incorporated into the DNA of dividing cells where it specifically inhibits DNA methylation by forming covalent complexes with the DNA methyltransferases (DNMTs). In an effort to study the correlations between DNA methylation, nucleosome remodeling, and gene reactivation, we investigate the integrated epigenetic events that worked coordinately to reprogram the methylated and closed promoters back to permissive chromatin configurations after 5-Aza-2′-deoxycytidine treatment. The ChIP results indicate that H2A.Z is deposited at promoter regions by the Snf2-related CBP activator protein (SRCAP) complex following DNA demethylation. According to our genome-wide expression and DNA methylation profiles, we find that the complete re-activation of silenced genes requires the insertion of the histone variant H2A.Z, which facilitates the acquisition of regions fully depleted of nucleosome as demonstrated by NOMe–seq (Nucleosome Occupancy Methylome–sequencing) assay. In contrast, SRCAP–mediated H2A.Z deposition is not required for maintaining the active status of constitutively expressed genes. By combining Hpa II digestion with NOMe–seq assay, we show that hemimethylated DNA, which is generated following drug incorporation, remains occupied by nucleosomes. Our data highlight H2A.Z as a novel and essential factor involved in 5-Aza-2′-deoxycytidine–induced gene reactivation. Furthermore, we elucidate that chromatin remodeling translates the demethylation ability of DNMT inhibitors to their downstream efficacies, suggesting future therapeutic implications for chromatin remodelers.
| Epigenetic changes, which include chemical modifications to the DNA and changes in the proteins that package DNA to fit into a cell, play an important role in gene expression regulation. The fact that a number of abnormal epigenetic changes that lead to the silencing of genes occur during tumorigenesis has prompted the design of epigenetic therapies. The ultimate goal of these therapies is to reverse the aberrant epigenetic modifications observed in cancer cells, thereby restoring cells to a “normal” state. 5-Aza-CdDR, a FDA approved drug for MDS treatment, reverses a chemical modification of the DNA resulting in gene reactivation. The data presented here show the importance of H2A.Z, a special DNA packaging protein variant, in the gene reactivation process induced by 5-Aza-CdR. The presence of H2A.Z facilitates the access of proteins at gene regulatory regions, which is a necessary step for gene re-expression. A better understanding of the events that follow 5-Aza-CdR treatment is a necessary step towards the design of combination and/or personalized epigenetic therapies.
| The eukaryotic genome is compacted into chromatin and associated proteins. The fundamental repeating unit of chromatin is the nucleosome, which contains ∼147 bp of DNA wrapped around a histone protein octamer [1]. However, chromatin conformations change during various cellular processes, such as the cell cycle, transcription or DNA damage [2]. During gene activation, transcription factors compete with chromatin packaging proteins in order to gain access to the DNA sequence and read the genetic information accurately. Accumulated evidence shows that the chromatin architecture of gene promoter regions strongly regulates gene transcription [3]. This chromatin environment might be altered by DNA methylation, post-translational modifications of histone proteins, histone variants and nucleosome positioning [4].
In mammalian cells, ∼60% of gene promoters are located within CpG islands, where cytosine methylation of CpG dinucleotides impairs gene expression. Histone modifications and histone variants are also strongly correlated with transcriptional status [3]. Nucleosome positioning plays an essential role in gene transcriptional regulation according to recent genome wide studies, which show that the majority of active or poised promoters have decreased nucleosome density [5]. Furthermore, the histone variants H2A.Z and H3.3, which are located at specific genome regions such as promoters, enhancers and insulators, work coordinately to destabilize nucleosomes [6]–[8]. The ATP dependent nucleosome remodelers catalyzing H2A.Z incorporation, namely SRCAP and p400 complexes in mammalian cells, have been suggested to be involved in transcriptional regulation, however, the role of H2A.Z remains controversial [9]–[15].
Abnormalities in epigenetic modifications play an essential role in tumorigenesis [16], and the reversal of them is the basic concept of epigenetic therapy for cancer. DNA methyltransferases (DNMT inhibitors), such as 5-azacytidine (5-Aza-CR) and 5-Aza-2′-deoxycytidine (5-Aza-CdR), are approved by the FDA for the treatment of MDS [17]–[18]. Although CpG demethylation is the direct and immediate consequence of treatment with DNMT inhibitors (5-Aza-CR and 5-Aza-CdR) [19], the level of demethylation in tumor suppressor genes does not predict clinical outcome, which suggests that unknown biological processes link the demethylation effects of DNMT inhibitors to their clinical benefits [20]. Several reports have already shown that, in addition to CpG demethylation, DNMT inhibitors indirectly reduce some repressive histone marks, increase acetylation of histone H3 and promote nucleosome depletion upstream of the transcription start sites (TSS) [21]–[24].
Here, taking advantage of a high resolution nucleosome positioning assay developed by our laboratory, we further study the integrated epigenetic changes following 5-Aza-CdR induced demethylation. In addition to the rapid enrichment of H3K4me3 at promoter regions, we find that H2A.Z incorporation increases in response to demethylation. Notably, CpG demethylation induced enrichment of H2A.Z and H3K4me3, as well as nucleosome depletion coordinately constitute a “permissive” chromatin architecture independently of histone acetylation levels. Inhibiting H2A.Z deposition by SRCAP knockdown lessens the establishment of “permissive” promoter environments and ultimately reduces the levels of gene reactivation after 5-Aza-CdR treatment. Genome-wide gene expression and DNA methylation studies further confirm that SRCAP-mediated H2A.Z insertion promotes DNA demethylation induced gene re-expression but has minimal effects on constitutively active genes. Our study reveals an important function of SRCAP/H2A.Z in promoting the reactivation process induced by 5-Aza-CdR but not in maintaining the expression of constitutively active genes and provides an insight to the chromatin structure of hemimethylated DNA.
To investigate the effects of DNA demethylation on chromatin architecture and gene expression, we treated the RKO colon cancer cell line, with 1 uM 5-Aza-CdR for 24 hours and followed the sequential changes in mRNA expression, DNA methylation and histone marks at the promoters of the MLH1, CDKN2A and MYOD1, which are methylated and silenced in RKO cells (Figure 1).
MLH1 expression began to rise at D2 after 5-Aza-CdR treatment, reaching a maximal level at D3 and remained constant for 4 days (Figure 1A). We performed a quantitative Methylation-sensitive Single-Nuleotide Primer Extension (Ms-SNuPE) assay to detect the DNA methylation changes at the indicated days (Figure 1B). A striking decrease in DNA methylation was observed at D2 (∼40%). The methylation level at the MLH1 promoter remained nearly constant from D2 to D9. We then used ChIP to monitor the changes of histone marks after 5-Aza-CdR treatment (Figure 1C). The enrichments of H2A.Z, H3K4me3 and H3K9/14 acetylation (acH3K9/14) were normalized to histone H3 levels to eliminate the potential influence of nucleosome depletion after drug treatment, and the ChIP primers were designed to amplify stable nucleosome regions which located just downstream of TSSs [5]. Interestingly, our results showed that the H2A.Z enrichment significantly changed after 5-Aza-CdR treatment (p<0.001), and could be observed as early as D2 when DNA methylation was substantially decreased (Figure 1C). H3K4me3 increased immediately after treatment and the enrichment of acH3K9/14 modestly increased at D2 and peaked by D3 displaying a similar pattern to the levels of MLH1 expression (Figure 1A).
The mRNA levels of CDKN2A rose steadily from D2 after 5-Aza-CdR treatment and peaked at D7 but abruptly dropped at D9 (Figure 1D). Although the active histone marks increased at the promoter of CDKN2A in a manner similar to MLH1, the H2A.Z level (p<0.05) diminished nearly to the baseline at D9 along with a rapid decline of acH3K9/14 from D5 to D9 (Figure 1F).
The mRNA level of MYOD1, a self-regulated gene expressed exclusively in muscle cells, remained undetectable and showed no RNA polymerase II (pol II) enrichment after 5-Aza-CdR treatment even though it showed a demethylation pattern similar to that of MLH1 (Figure 1G, 1H; Figure S1A). Most interestingly, we observed modest changes in H2A.Z and H3K4me3 at the MYOD1 promoter, whereas acH3K9/14 remained extremely low, in agreement with its low expression level (Figure 1I). Therefore the MYOD1 promoter acquired a “permissive” state for expression after 5-Aza-CdR treatment but the gene was not expressed.
In addition to analyzing DNA methylation changes at the single strand level (Figure 1B, 1E, 1H), we detected and quantified asymmetrically methylated DNA after 5-Aza-CdR treatment. We performed a hemimethylation Ms-SNuPE assay based on Hpa II digestion as we have developed previously and used the CDKN2A gene as a model (Figure S1B-S1D) [25]. The majority of DNA molecules were hemimethylated DNA duplexes at D2 (∼60%). Even at D5, ∼5% of the DNA duplexes were composed of hemimethylated DNA molecules. There was a small portion of fully demethylated DNA duplexes at D2, however, the maximal levels of double strand demethylation (∼40%) was detected at D3–5 when hemimethylated DNA levels dropped.
Collectively, our results demonstrate that 5-Aza-CdR treatment eventually produces two types of demethylated DNA duplexes, hemimethylated and fully demethylated DNA. After demethylation H2A.Z and H3K4me3 are deposited in all three promoters. Interestingly, only acH3K9/14 shows a correlation between its enrichment and mRNA expression.
We next investigated the effects of 5-Aza-CdR on nucleosome occupancy (Figure 2, Figure 3). We previously examined the nucleosome occupancy status of the MLH1 promoter in RKO and LD419 cells using MNase-ChIP assay and Methylase-based Single-Promoter Analysis assay (MSPA) [22]. We confirmed our previous findings using a recently developed high resolution assay, which uses the GpC methyltransferase (M.CviPI) instead of CpG methyltransferase (M.SssI) to methylate GpC sites that are not occupied by nucleosomes or tightly bound transcription factors [26]. By analyzing the methylation status of GpC sites, this NOMe-seq (Nucleosome Occupancy Methylome-sequencing) assay provides a digital footprint of nucleosome occupancy and allows the study of nucleosome positioning in both CpG islands and CpG poor regions regardless of their CpG methylation status.
Nucleosome occupancy at the MLH1 and CDKN2A promoters in LD419 and RKO cells were analyzed by NOMe-seq using PCR primers lacking CpG or GpC sites to avoid complications due to cytosine methylation in the parental molecules (Figure 2). Both promoters were unmethylated in LD419 cells and had clear nucleosome depleted regions (NDRs) which were accessible to the exogenous M.CviPI. In contrast, the promoters of MLH1 and CDKN2A were methylated in RKO cells and were inaccessible to the GpC methyltransferase, indicating that hypermethylated promoters were fully occupied by nucleosomes.
To investigate the nucleosome occupancy changes accompanying drug induced DNA demethylation, we used primers which were specifically designed to amplify the DNA strands which had become demethylated at CpG sites and studied the accessibility of these demethylated molecules to M.CviPI (Figure 3; Figure S2). The maximum demethylation of CpG sites was observed at D2 (Figure 1B), but only one out of twenty-five demethylated DNA strands (4%) had a nucleosome depleted region larger than 146 bp around the TSS (green bar on the graph)(Figure 3A). At D3, a proportion of unmethylated DNA strands at the gene promoters (24%) showed a nucleosome depleted area large enough to accommodate at least one nucleosome, consistent with the presence of active histone marks and increased gene expression. Extensive nucleosome depletion (40%), the H2A.Z enrichment and gene reactivation reached a maximal level at D5 (Figure 3A). NOMe-seq analysis of the CDKN2A promoter yielded similar results and showed depletion of the −1 nucleosome at D5 (12%) (Figure 3B). Interestingly, the MYOD1 promoter showed drug-induced enrichment of H2A.Z and H3K4me3 as well as nucleosome depletion around the TSS (20%) without MYOD1 expression (Figure 3C). Therefore, changes in histone modifications and nucleosome depletion were the direct consequences of DNA demethylation and did not require transcriptional activation for some genes, such as MYOD1.
Although the nucleosome occupancy at the MLH1 promoter was dramatically decreased on the demethylated DNA strands at D5 after 5-Aza-CdR treatment, a portion of demethylated DNA strands remained inaccessible to M.CviPI (Figure 3A). As shown in Figure S1C, the majority of demethylated DNA strands were associated with hemimethylated DNA duplexes at D2. And the demethylated DNA strands at the D2 were highly occupied by nucleosomes. To further investigate the nucleosome occupancy on hemimethylated DNA, we pre-digested the NOMe-seq DNA samples before bisulfite treatment with an excess of Hpa II. Demethylated DNA strands associated with symmetrically demethylated DNA duplexes are destroyed by Hpa II digestion, whereas the demethylated strands in hemimethylated DNA duplexes are resistant to digestion (Figure S1B). Next, we used the same PCR primers as shown previously to amplify the remaining demethylated DNA strands. The NOMe-seq results from Hpa II digested DNA clearly showed that the promoters of MLH1 and CDKN2A were occupied by nucleosomes when the underlying DNA was hemimethylated (Figure 3D, 3E). Our results show that DNA demethylation at promoter regions induces substantial changes in nucleosome occupancy which only occurs on symmetrically demethylated but not hemimethylated DNA.
The enrichment of H3K4me3 after demethylation has been well studied [27], however the role of H2A.Z insertion in gene reactivation is unclear. Thus, we explored the potential role of H2A.Z in 5-Aza-CdR induced gene reactivation by knocking down SRCAP (Figure S3A), which catalyzed H2A.Z deposition in a cell cycle independent manner [28], and subsequently treating the cells with 5-Aza-CdR (Figure 4A).
The expression of MLH1 and CDKN2A was strongly attenuated by SRCAP knockdown, concomitant with a reduction of H2A.Z levels at the promoters (Figure 4A). The enrichment of acH3K9/14 at the reactivated promoters was reduced after SRCAP knockdown as well. Interestingly, the 5-Aza-CdR induced H2A.Z deposition was also inhibited by SRCAP knockdown at the MYOD1 promoter, but the gene expression and acH3K9/14 levels remained undetectable as expected. Of note, SRCAP knockdown showed minimal effects on the H3K4me3 levels at all three promoters, suggesting that the H3K4me3 mark was independent of H2A.Z levels. In addition, knockdown of SRCAP did not affect DNA methylation levels at promoters of these examined genes (Figure S3B).
In contrast, the mRNA level of GRP78, which is usually over-expressed in cancer cells and has enriched H2A.Z around its promoter [2], was not reduced and even modestly increased after SRCAP knockdown. The H2A.Z level at the GRP78 promoter dramatically dropped by nearly 90% compared with non-target (NC) siRNA treated cells. Meanwhile, the levels of H3K4me3 and acH3K9/14 remained high showing that these marks were independent of H2A.Z levels. In addition to GRP78, we analyzed two more genes, LAMB3 and G3BP, both of which were unmethylated and expressed in RKO cells (Figure S3D, S3E). The enrichment of H2A.Z, which has been identified at both promoters [10], [29], was reduced by SRCAP knockdown, but neither the mRNA expression nor the histone marks were significantly affected. Remarkably, knockdown of SRCAP in LD419 cells did not inhibit the expression of MLH1 and CDKN2A, though the H2A.Z enrichment at the promoter regions had been reduced. Again, no difference in H3K4me3 and acH3K9/14 levels was detected after SRCAP siRNA treatment (Figure S4A).
We next investigated the function of SRCAP-mediated H2A.Z deposition on 5-Aza-CdR induced nucleosome occupancy changes. Substantial nucleosome depletion at the MLH1 promoter was observed on the demethylated DNA strands in the NC siRNA treated cells as previously shown (Figure 4B). When SRCAP-mediated H2A.Z incorporation was inhibited, nucleosome depletion was much curtailed at the promoter (32% to 20%). Similarly, we observed that inhibiting SRCAP-mediated H2A.Z deposition prevented the depletion of nucleosome induced by the 5-Aza-CdR in the vicinity of the CDKN2A (20% to 4% ) and MYOD1 (16% to 8%) TSSs. In contrast, the NDRs at the GRP78, LAMB3 and G3BP promoters were not reduced and even showed modest increases (Figure S3C, S3F, and S3G). Similarly, the NDRs upstream of the TSS of MLH1 did not change after SRCAP knockdown in LD419 cells (Figure S4B).
Taken together, these results demonstrate that SRCAP-mediated H2A.Z deposition and associated nucleosome depletion play a key role in re-constructing a poised chromatin architecture around demethylated promoters. In contrast, continued H2A.Z presence is not critical in maintaining an open chromatin environment of actively transcribed genes.
To elucidate the importance of H2A.Z deposition for gene reactivation following 5-Aza-CdR induced demethylation, we conducted genome-wide studies to assay global DNA methylation and gene expression changes after drug treatment.
We interrogated global promoter DNA methylation patterns using the Infinium HumanMethylation27 platform, which includes 27,578 CpG dinucleotides spanning 14,495 well-annotated, unique gene promoter and/or 5′ gene regions (from −1,500 to +1,500 from the TSS). The DNA methylation level for each interrogated CpG site is reported as a beta value, ranging from zero (low DNA methylation) to one (high DNA methylation). In NC siRNA treated cells, the CpG sites could be roughly separated into two groups based on the bimodal distribution of the beta values: a hypomethylated group (beta value<0.2) and a hypermethylated group (beta value>0.8) [30] (Figure 5A). Consistent with the Ms-SNuPE results (Figure 1B, 1E, 1H), the CpG sites within the promoters of MLH1, CDKN2A and MYOD1 had beta values higher than 0.8. The peak representing hypermethylated probes was notably shifted towards the left after 5-Aza-CdR treatment. To further expand this observation, CpG probes were plotted between 5-Aza-CdR and PBS treatment (control) in NC siRNA treated cells (Figure 5B). Using a beta value difference of 0.25 as a threshold for differential DNA methylation and separating CpG probes based on beta values>0.8 associated with control treatment, 2,638 CpG probes (1278 genes) were identified to be demethylated by 5-Aza-CdR in NC siRNA treated cells and knockdown of SRCAP only showed little effect on DNA methylation patterns, as 2,515 CpG probes (1208 genes) were found to be demethylated in SRCAP siRNA treated cells (Figure 5C; Figure S5A, S5B).
To evaluate gene expression changes, we performed a permutation analysis (1,000 permutations) using Significance Analysis of Microarrays (SAM) in NC or SRCAP siRNA transfected cells after 5-Aza-CdR treatment [31]. The identification of differentially expressed genes was performed among the indicated groups (Figure 5D–5F; Figure S5C, S5D). We found that SRCAP knockdown had minimal impact on global gene expression (Figure 5E), while 5-Aza-CdR treatment significantly up-regulated 97 genes (representing 130 different transcripts) in NC treated cells and 86 genes (representing 105 different transcripts) in SRCAP siRNA treated cells, with an 81% overlap between two groups (Figure 5D; Figure S5E, S5F). We did not observe any gene significantly down-regulated by 5-Aza-CdR treatment in either NC or SRCAP siRNA treated cells.
To visualize the global gene expression difference between SRCAP and NC siRNA, we plotted the observed log2 fold change for all the interrogated transcripts on the platform (Figure 5F). After extracting the promoter DNA methylation beta values of 97 genes, which were reactivated by 5-Aza-CdR in NC siRNA treated cells from the Infinium array, we found that 44 genes (representing 92 different CpG loci), including MLH1 and CDKN2A, had beta value differences greater than 0.2 as shown in the heatmap and box plot (Figure 5G, 5H). We next concentrated on these 44 genes which were demethylated and subsequently reactivated. Within this group of genes, knockdown of SRCAP significantly inhibited the reactivation of some transcripts such as EPM2AIP1 from up-regulated to non-responsive. Although the majority of genes that were up-regulated by 5-Aza-CdR in NC siRNA treated cells were still induced in SRCAP siRNA treated cells, the reactivation levels were strikingly decreased (red circles). The fold changes of the 44 genes induced by demethylation were calculated in the inserted box plot, further showing the significant effects of SRCAP knockdown (p<2×10−16). To validate this genome-wide analysis, we randomly selected four candidate genes (CHFR, CTCFL, SYCP3 and EPM2AIP1) from the pool of 44 genes and analyzed the expression changes (Figure S6A–S6C). The reactivation levels of four methylated genes were significantly suppressed by SRCAP knockdown. We then validated the histone marks changes on the promoters of CHFR and SYCP3. We found that SRCAP knockdown prevented H2A.Z deposition as well as diminished gene reactivation, which was consistent with our results from MLH1 and CDKN2A. To confirm that the H2A.Z insertion was causing the observed effects, we knocked down p400, which is a homolog of SWR1 and has been identified as another key player in H2A.Z deposition [13]. We found that inhibiting p400 could also reduce the reactivation of MLH1 and CDKN2A in RKO cells (Figure S6D). In addition to knocking down the two catalytic subunits of SRCAP and p400 complexes, inhibition of YL-1, the binding partner of H2A.Z in the SRCAP complex [32]–[33], also suppressed 5-Aza-CdR induced gene reactivation (6E) and depletion of the −1 nucleosome (Figure S6F).
Our integrated study reveals that 5-Aza-CdR robustly reduces global promoter DNA methylation levels, and subsequently reactivates gene expression. Decreasing SRCAP expression inhibits global gene reactivation but has no effect on DNA methylation at promoter. However the maintenance of active gene expression might not require highly enriched H2A.Z.
Although recent studies have begun to explore the epigenetic factors involved in the 5-Aza-CdR mediated demethylation process [21], [24], our study focuses on the dynamic changes in chromatin architecture immediately after 5-Aza-CdR treatment (Figure 6). We demonstrate that removing DNA methylation rapidly induces H2A.Z incorporation, which confirms the antagonistic relationship between H2A.Z and DNA methylation observed in genome-scale studies of arabidopsis, human breast tissue and tumorigenesis of a B-cell lymphoma model in mouse [34]–[36]. Although some reports show that loss of pie1 in Arabidopsis or H2A.Z in mammals increases DNA methylation levels at gene body regions [34], [37], our data demonstrate that DNA methylation levels at promoter regions are not affected by transiently inhibiting H2A.Z insertion. Previous reports have showed positive correlations between H2A.Z insertion and the expression of CDKN1A, estrogen receptor target genes, muscle differentiation-specific genes and PcG protein target genes in ES cells [11], [13], [33]. Our genome-wide expression results demonstrate that SRCAP-mediated H2A.Z deposition at promoter regions is necessary for complete gene reactivation induced by DNA demethylation. We show that inhibition of SRCAP-mediated H2A.Z insertion prevents nucleosome depletion at the promoters of MLH1 and CDKN2A after 5-Aza-CdR treatment. In addition, knockdown of YL-1, the binding partner of H2A.Z in the SRCAP complex, also reduces the CDKN2A gene reactivation and nucleosome depletion around the TSS region that are induced by 5-Aza-CdR treatment. Collectively, our data provides evidence for the hypothesis that SRCAP/H2A.Z directly promotes transcription by reducing nucleosome occupancy at promoter regions [38]. Nevertheless, H2A.Z enrichment is necessary but not sufficient for gene reactivation according to our data. As shown at the MYOD1 promoter, the modest enrichment of H2A.Z contributes to the establishment of a “permissive” environment regardless of the subsequent gene reactivation status. The reduction of M.CviPI accessibility at the MYOD1 promoter after SRCAP knockdown suggests that H2A.Z mediated nucleosome depletion is not the consequence of gene expression and might be an early event in transcription initiation. In addition, Hardy et al [15] showed that H2A.Z was recruited to the promoter regions prior to pol II binding. We observed pol II enrichment at the MLH1 promoter but not at the MYOD1 promoter, though both promoters have been remodeled structurally, which suggested the formation of “permissive” promoter regions might not require the presence of pol II at an early stage.
Unlike H2A.Z, histone H3 acetylation occurs concomitantly with gene expression, and is not required for establishing this early “permissive” promoter. Furthermore, the data from the SRCAP knockdown experiments strongly indicate that H2A.Z incorporation, especially SRCAP mediated deposition, is independent of H3K9/K14 acetylation. Functional studies of the Tip48/49 complex , which shares some components of the SRCAP complex, showed that acetylation of H2A enhanced H2A.Z insertion [39]. However, a recently published report demonstrated that inhibiting NF-Y, one of the proteins with highly similarity to core histones, prevented H2A.Z deposition at promoter regions but had no observable effect on histone H3K9/14 acetylation, indirectly supporting our conclusions [40].
Similarly to the behavior of H2A.Z, H3K4me3 enriches at promoters following demethylation, in agreement with the reported mutually exclusive relationship between H3K4me3 and DNA methylation [41]. Although many reports show that H3K4me3 levels are correlated with gene expression status, Thomson et al [42] demonstrated that artificially inserted promoter-less DNA sequences containing unmethylated CpG sites were sufficient to acquire H3K4me3. In our study, DNA demethylation associated H3K4me3 enrichment creates a “permissive” promoter configuration; however, such a configuration is not sufficient for gene activation. The presence of key transcription factors is also necessary [29]. In yeast, Set1, an H3K4 methyltransferase, and H2A.Z have redundant functions in preventing the spread of Sir-mediated silencing, indicating that the presence of H3K4 methylation marks and H2A.Z are not dependent on each other [43]. Interestingly, our results show that the enrichment of H3K4me3 is not affected after inhibiting SRCAP-mediated H2A.Z deposition and suggest that in mammalian cells the regulation of these histone marks might not depend on each other. Therefore, it would be interesting to study the specific effects of H3K4me3 on chromatin remodeling after DNA demethylation in the future.
It has been reported that certain DNA sequences and the binding of pol II or transcription factors influence nucleosome occupancy by different mechanisms [44]–[45]. Recent reports have shown that methylated DNA facilitates nucleosome assembly in vitro, and reciprocally, stable nucleosomes contribute to the establishment and maintenance of DNA methylation [46]–[48]. Here, we apply the now well-established NOMe-seq assay to investigate the correlation between nucleosome positioning and DNA methylation after drug treatment [45], [49]. Our data demonstrate that promoter regions are highly occupied by nucleosomes when DNA duplexes are either symmetrically methylated or hemimethylated in living cells, suggesting the dominant role of DNA methylation in maintaining stable nucleosomes. Using reconstituted histone octamers and single-stranded M13 constructs, Deobagkar et al show that hemimethylated DNA prevents chromatin expression [50]. Thus complete nucleosome depletion takes place only on symmetrically demethylated DNA after 5-Aza-CdR treatment, which is probably required for full gene activation. Furthermore, our data have confirmed the feasibility of utilizing NOMe-seq in future to investigate the drug induced nucleosome remodeling globally.
According to our genome-wide analysis, 5-Aza-CdR induces global demethylation yet only a limited number of genes are significantly reactivated, indicating that CpG demethylation and the subsequent establishment of open chromatin architectures are essential but not sufficient to induce gene reactivation. The “permissive” promoters created by 5-Aza-CdR treatment, such as the MYOD1 promoter, have active histone marks and NDRs, which are needed for assembling of the transcriptional machinery, but other regulatory factors are required to fully reactivate these genes. Therefore, epigenetic modulators which regulate histone marks and nucleosome positioning have strong abilities to promote or impede the pharmacological functions of 5-Aza-CdR. Our results provide a rationale to design clinical trials combining DNMT inhibitors with other anticancer drugs, especially histone deacetylase inhibitors, which facilitate histone acetylation. A comprehensive understanding of the coordinated interplay between epigenetic regulators and 5-Aza-CdR will help explain the drug's clinical outcomes as well as promote the discovery of novel therapeutic targets.
RKO, a colon cancer cell line, was purchased from ATCC and was maintained in MEM medium with 10% FBS. LD419, a normal human bladder fibroblasts was generated by Dr Louis Dubeau and was maintained in McCoy's 5A supplemented with 20% FBS.
RKO cells were plated at 2×106 cells/100-mm dish and treated with 1 µM of 5-Aza-CdR (Sigma Chemical Co., St. Louis, MO) for 24 hours. The NC pool siRNA (D-001810-10-05) and the ON-TARGET plus siRNA targeting SRCAP (L-004830-00-0005), p400 (L-021272-01-0005) and YL-1 (VPS72 L-020097-00-0005) (Thermo Fisher Scientific Inc.) were transfected into RKO cells 24 hours before 5-Aza-CdR treatment using DharmaFECT™ siRNA transfection reagents (Thermo Fisher Scientific Inc.).
Total RNA was extracted using RNeasy kit (Qiagen) and was converted to cDNA by M-MLV Reverse Transcriptase (Promega) using random primer (Promega). The sequences of gene specific primers and taqman-probs are available upon request. With each set of PCR primers, titrations of known amounts of DNA were included as a standard for quantization.
DNA methylation level was determined by Ms-SNuPE as described previously [51]. Briefly, CpG sites were interrogated for each promoter. The methylation level of each gene is the average of the three CpG sites examined by Ms-SNuPE.
ChIP was performed as described previously [52]. Ten µg of the following antibodies were used: anti-Histone H3 (Abcam), anti-acetylated Histone H3K9/14 (Milipore), anti-H2A.Z (Abcam) and anti-RNA polymerase II(Abcam). Five µl of anti-H3K4me3 (Active Motif) antibody was used. Ten µg of Rabbit IgG (Millipore) was used as a non-specific antibody control. PCR primers are available upon request.
Nuclei preparation and GpC Methyltransferase treatment were performed as described previously [26]. Briefly, freshly extracted nuclei were treated with 200 U of GpC methyltransferase for 15 min at 37°C. An equal volume of stopping solution (20 nM Tris-HCl, 600 mM NaCl, 1% SDS, 10 mM EDTA) was added to stop the reaction. The final mixture was incubated at 55°C overnight with 400 µg/ml proteinase K. DNA was isolated and bisulfite converted. The regions of interest were amplified and cloned into pCR 2.1-TOPO vector (Invitrogen) for DNA sequencing.
Hemimethylation analysis was performed as described previously. Undigested or Hpa II-digested DNA from RKO cells before and after treatment was subjected to bisulfite modification. Hpa II digests unmethylated DNA but does not cut a fully or hemimethylated configuration of its CCGG target sequence. Bisulfite-treated DNA was then amplified by PCR using Ms-SNuPE primers that flanked one Hpa II site at the CDKN2A promoter. The equations used to determine hemimethylation levels used were as described previously [25].
The Illumina Infinium DNA methylation assay technology has been described previously [30]. The Infinium DNA methylation assay was performed at the USC Epigenome Center according to the manufacturer's specifications (Illumina, San Diego, CA). The Illumina Infinium DNA methylation assay (HumanMethylation27_270596_v.1.2) examines DNA methylation status of 27,578 CpG sites located at promoter regions of 14,495 protein-coding genes and 110 microRNAs. Downstream processing and beta value calculations were done as previously described [53].
Expression analysis was performed using the Illumina whole-genome expression BeadChip (HumanWG-6 v3.0, 48,803 transcripts) (Illumina). The hybridized chips were stained and scanned using the Illumina HD BeadArray scanner (Illumina). Scanned image and bead-level data processing were performed using the BeadStudio 3.0.1 software (Illumina). The summarized data for each bead type were then processed using the lumi package in Bioconductor [54]. The data were log2 transformed and normalized using Robust Spline Normalization (RSN) as implemented in the lumi package.
All statistical tests were done using R software (R version 2.12.1, 2010-12-16, R Development Core Team, 2009). ‘lumi’ package was used to normalize and process gene expression data. ‘samr’ (version 1.28) package was used for all permutation tests to access significance of gene expression changes. Differential gene expression (significance) change was established for each application by setting the cut-off on a FDR of q = 0.05 after applying 1000 permutation. The following CRAN packages were used to generate plots: ‘ggplot2’ and ‘LSD’ (version 1.0). The H2A.Z ChIP results from three biological experiments were analyzed by one way ANOVA using Prism 3(GraphPad).
All summarized probe profile data and processed expression data and DNA methylation data which are used in this study have been deposited to Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/projects/geo/) under accession Number GSE26685.
|
10.1371/journal.pntd.0006248 | Host preferences support the prominent role of Hyalomma ticks in the ecology of Crimean-Congo hemorrhagic fever | Crimean-Congo hemorrhagic fever virus (CCHFV) is a tick-borne zoonotic agent that is maintained in nature in an enzootic vertebrate-tick-vertebrate cycle. Hyalomma genus ticks have been implicated as the main CCHFV vector and are key in maintaining silent endemic foci. However, what contributes to their central role in CCHFV ecology is unclear. To assess the significance of host preferences of ticks in CCHFV ecology, we performed comparative analyses of hosts exploited by 133 species of ticks; these species represent 5 genera with reported geographical distribution over the range of CCHFV. We found that the composition of vertebrate hosts on which Hyalomma spp. feed is different than for other tick genera. Immatures of the genus Hyalomma feed preferentially on species of the orders Rodentia, Lagomorpha, and the class Aves, while adults concentrate mainly on the family Bovidae. With the exception of Aves, these hosts include the majority of the vertebrates consistently reported to be viremic upon CCHFV infection. While other tick genera also feed on these hosts, Hyalomma spp. almost completely concentrate their populations on them. Hyalomma spp. feed on less phylogenetically diverse hosts than any other tick genus, implying that this network of hosts has a low resilience. Indeed, removing the most prominent hosts quickly collapsed the network of parasitic interactions. These results support the intermittent activity of CCHFV foci: likely, populations of infected Hyalomma spp. ticks exceed the threshold of contact with humans only when these critical hosts reach adequate population density, accounting for the sporadic occurence of clinical tick-transmitted cases. Our data describe the association of vertebrate host preferences with the role of Hyalomma spp. ticks in maintaining endemic CCHFV foci, and highlight the importance of host-tick dynamics in pathogen ecology.
| Crimean-Congo hemorrhagic fever virus (CCHFV), a cause of severe hemorragic disease in humans, is maintained in nature in a tick-vertebrate-tick enzootic cycle characterized by silent persistence of endemic foci with intermittent epidemics. Most studies support ticks of the genus Hyalomma as the main CCHFV vectors, but some laboratory reports and field studies also allude to other tick species that may be responsible for virus circulation. Here we converted the tick-host interactions of 133 species of ticks with reported geographical distribution over the range of CCHFV into a network of relationships. By a series of network analyses, we found that immatures of the genus Hyalomma are unique in their hosts preferences among the examined tick genera. Immatures of the genus Hyalomma concentrate on rodent and rabbit hosts, which most efficiently support CCHFV maintenance and transmission by ticks in nature. Based on these data, we formulate the hypothesis that the ecological relationships between Hyalomma spp. and their hosts form a delicate equilibrum that differentiates human epidemics from periods of silent CCHFV maintenance.
| Crimean-Congo hemorrhagic fever (CCHF) is a tick-borne zoonotic disease seen exclusively in humans that can progress from mild, non-specific signs to a severe and fatal hemorrhagic disease. The etiologic agent, Crimean-Congo hemorrhagic fever virus (CCHFV; family Nairoviridae, genus Orthonairovirus), is transmitted to humans predominantly via tick bites, but may also be transmitted nosocomially or by handling tissues from viremic animals (e.g., in abattoirs). As non-human vertebrate hosts do not develop clinical signs [1], maintenance in nature is largely silent. Recent reviews summarize the current knowledge about serology in animals [2], routes of transmission [3], and the tick species unambiguously involved in CCHFV circulation in natural and permanent foci [4].
Ticks are vectors and reservoirs for CCHFV; vertebrates act as a bridge, transmitting the virus to new generations of ticks. Infected vertebrates develop a short viremia [1], and virus is transmitted to ticks feeding on viremic hosts, or through co-feeding with infected ticks (demonstrated by [5,6]), which release the virus into the feeding cavity where other uninfected ticks feed. Most reports support ticks of the genus Hyalomma as the main CCHFV vector, but laboratory and field studies allude to other tick species that may also be responsible for virus circulation [4]. To date, interactions between CCHFV and the tick are not well known, including hypothetical molecular factors that could regulate infection and viral dissemination through a variety of physiological and anatomical barriers in the tick [7]. While molecular interactions are an obvious research target to explain the prominence of Hyalomma spp. in CCHFV maintenance and transmission, other non-molecular relationships may also be involved or even predominate. Some ticks of other genera may act as efficient CCHFV vectors under adequate laboratory conditions (e.g., as reviewed in [4]), raising the question about the importance of intimate molecular relationships between tick and virus versus the simple ecological interactions of ticks and key hosts in supporting silent CCHFV foci. In other words, the dynamics of CCHFV transmission may be driven by purely ecological factors and not depend on molecular compatibility.
Most tick species are not restricted in range by their hosts; rather, climate is the main driver of their distribution patterns [8]. With the exception of some monoxenic species, ticks regularly feed on a wide range of hosts. Permanent foci of some tick-transmitted pathogens are restricted to the range of key reservoirs or vectors. For example, Borrelia burgdorferi s.l. is intimately linked with the tick genus Ixodes; Babesia bovis is transmitted exclusively by boophilid ticks; and tick-borne encephalitis virus is restricted to rodents reservoirs [9–11]. Likewise, CCHFV could conceivably circulate only in areas that support a delicate equilibrium of abundance and composition of appropriate hosts for the ticks.
We hypothesized that a key factor for maintaining CCHFV foci is a precise combination of host species that feed the ticks, thereby amplifying infection in the tick population. A network-based analysis, comprising relationships among ticks and hosts connected by pairwise relations, enables the reconstruction of associations between ticks and key vertebrates in the circulation of a pathogen. This kind of ecological modeling is made possible by an extensive toolbox developed for network research (see, for example, [12–14]). The structural properties of tick-vertebrate networks reveal new insights into the associations linking ticks and hosts that are key for supporting permanent CCHFV foci.
Here we aimed to compare the combinations of hosts parasitized by ticks colonizing the reported range of CCHFV, focusing on explicit relationships between ticks and the hosts reported to support viremia. We explicitly tested the phylogenetic diversity and the centrality (i.e., the relative importance in the network of connections) of the groups of hosts used by these ticks, the resilience of the networks to the removal of hosts, and the existence of clusters of tick-host interactions. We used these findings to elaborate on the specific relationships of Hyalomma spp. with their hosts and determine if these relationships differ from those of other ticks. We also attempted to identify the key factors shaping the circulation of CCHFV mainly by Hyalomma spp. ticks, and ascertain how these specific combinations drive unstable foci of the virus.
A species-by-species analysis of the tick-host relationships is not possible, because i) some ticks species are underreported (e.g., prevalence, host preference) in the literature and therefore an evident bias in the number of hosts is expected; and ii) the immatures of some ticks are difficult to identify, leading to the reporting of host species that support improperly identified ticks. Following the same reasoning, the analysis of the relationship between ticks and specific host species is not possible; some vertebrates may be very poorly surveyed (because they are rare, difficult to trap, or protected, etc.), which would undoubtedly bias the holistic approach. We therefore used the data on families of hosts for each species of tick, as reported by [15]. The geographical range refers to the complete Palearctic and Afrotropical regions, which are the territories in which CCHFV circulation has been described. Then, data were summarized at the level of tick genera.
An estimation of the relative importance of each family of vertebrates was developed in the context of a network of tick-host relationships, at the level of tick genera and life stages (larvae, nymphs, or adults), similar to network approaches commonly performed in other scientific fields [16–18]. A network is a construct that reflects organisms (nodes) that interact in any way (links). In our approach, nodes are ticks and vertebrates, and links display the reported finding of a given tick and life stage on a vertebrate of a given family. It is thus a directed network, since ticks have been recorded on hosts. The basic index of a network is the weighted degree (WD), defined as the weighted number of times a group of hosts is recorded for the complete set of tick genera [19]. We calculated the betweenness centrality (BNC, [20]), an index that explains how important a node is in linking several other nodes of either ticks or vertebrates. The ecological significance of the index in our application is immediate: BNC is higher for families of vertebrates that are predominantly used as hosts by several genera of ticks. Separate calculations of BNC for each genus and life stage of ticks give a comparitive overview of the relative importance of the hosts. Clusters of the network were calculated using the algorithm of Neumann [21]. A cluster is a group of nodes that interact more among themselves than with other nodes in the network. Clusters have importance in this context because they reflect groups of vertebrates among which ticks interact more commonly, thus displaying an ecological relationship. Network calculations were done for the complete dataset (to capture the structure of the complete network of interactions), as well as separately for every genus and stage of ticks (to understand the ecological relationships of every tick genus independently of the rest).
For each genus and stage of ticks, we calculated the genetic richness of the exploited hosts using Faith’s phylogenetic distance (PD). This metric is based on the sum of distances of the branches that link any pair of families in the phylogenetic tree [22], and is used to determine if the different genera and stages of ticks parasitize phylogenetically narrow or wide host ranges. PD is an adequate estimate in this context, and supersedes simple measures of host variability based solely on the number of different taxa that serve as tick hosts. We first obtained the phylogenetic tree of the families of hosts, as available in the Open Tree of Life (OTL, https://tree.opentreeoflife.org). The OTL is a repository of phylogenetic trees and produces synthetic trees for a broad range of organisms. It can be accessed through its API to obtain portions of the complete phylogenetic tree stored in the repository. We used a script in the R programming environment to query OTL for the phylogenetic pattern of the families of hosts used by the ticks examined in this study. The resulting tree (see S1 Fig) contained data on the phylogenetic relationships of 92 families of vertebrates and was suitable for obtaining estimates of the relative branching of the vertebrates utilized by each genus of tick, but had no calibrated date because it was a synthetic tree. The remaining 40 families of hosts had no information in OTL, and the available information in GenBank was too fragmented to be combined with the already built tree.
We evaluated the resilience of the network of tick-vertebrate relationships separately for each tick genus and life stage to understand how random or directed attacks could affect its stability. Resilience of a host-parasite network is an important feature emanating from the network approach, and can be evaluated by removing the hosts either randomly or based on their BNC order in the network. Resilience is measured in terms of the probability of network collapse; the removal of key hosts may lead the network to break down without further links of the parasites to the remaining hosts. We built and obtained indexes of the network structure on the R programming environment [23] using the igraph [24], bipartite [25], and picante [26] packages. The resilience of the networks of each genus and stage of ticks after recursive removal of host nodes was evaluated with the package NetSwan for R [27]. Visualization of the networks was done in Gephi v0.91 [28].
To analyze the relationships between ticks and hosts with evidence of a potential role in CCHFV circulation, we compiled a list of tick species distributed over the geographical range of CCHFV. This range includes the complete Afrotropical region, the Mediterranean Palaearctic region, and most of Central Asia from the Turkish steppes to India. We focused on 5 tick genera: Amblyomma, Dermacentor, Hyalomma, Ixodes, and Rhipicephalus, which contain species that have been implicated in CCHFV transmission through either studies of natural foci or in the laboratory. All species in the genera were included in the analysis, but not every species included has been reliably linked to CCHFV transmission. Our dataset contained 22 species of genus Amblyomma, 2 of Dermacentor, 18 of Hyalomma, 44 of Ixodes, and 47 of Rhipicephalus, with a total of 1591 pairs of reported associations between the 133 tick species and 132 families of hosts. Phylogenetic calculations were performed on 92 vertebrate families. The network construct had a total of 147 nodes (genera and stages of ticks, families of hosts) and 553 links. The complete list of ticks and hosts is included as S1 Table. Values of BNC for each family of hosts (converted to the range 0–100 to improve comparisons) are included in S1 Data.
The network construct provides a representation of the tick-host relationships, enabling research on the principles behind complex interactions. Fig 1 displays the network and its clusters, explicitly describing the sets of nodes that interact more among themselves than with others (see also S2 Fig). Up to 5 groups or clusters of organisms can be detected, denoting dominant interactions between sets of ticks and vertebrates. Interestingly, each cluster was formed by the 3 life stages of the same genus of ticks, except for Dermacentor and the adults of genus Rhipicephalus. The adults of genus Dermacentor appeared in the same cluster as adults of genus Hyalomma, and the immatures clustered with immatures of the genus Ixodes. The adults of Rhipicephalus formed their own cluster of interacting organisms.
The overview of the relationships between ticks and vertebrates (Fig 1) shows 54 families of hosts (36.73% of total) that are only slightly important, in terms of BNC, for supporting the network of ticks; these families appear at the periphery of the network. Most Amblyomma species use hosts of the classes Reptilia and Aves, and, in a large proportion, the family Bovidae. Patterns of diversity for Amblyomma spp. are high, with 36, 35, and 37 vertebrate families used by larvae, nymphs, and adults, respectively. A similar pattern is seen for Hyalomma spp.: the larvae, nymphs, and adults use 41, 39, and 37 vertebrate families. Most of the vertebrates families used as hosts by Hyalomma spp. are small endotherms, including members of the orders Rodentia, Lagomorpha, and Artiodactyla, and the class Aves.
Leporidae and Muridae had BNC values of 1222 and 583 for larvae of the genus Hyalomma, and 1332 and 581 for nymphs, demonstrating the importance of these vertebrate families in supporting ticks of this genus. In contrast, Muridae had a BNC value of only 76 for larvae of the genus Amblyomma, and neither Muridae nor Leporidae was parasitized by Amblyomma nymphs. Adults of both Hyalomma and Amblyomma spp. feed on ungulates (family Bovidae, BNC = 460 for Hyalomma and 1544 for Amblyomma). Interestingly, adults of the genus Hyalomma also utilized Leporidae (BNC = 248).
The pattern was completely different for Ixodes spp., the ticks that feed on the widest variety of vertebrates, including 51 families exploited by larvae, 59 by nymphs, and 56 by adults. Every stage of this tick genus was widely and loosely distributed over a wider variety of hosts than other tick genera.
Rhipicephalus spp. were highly eclectic in host preference, and the 3 life stages of this genus do not cluster in a discrete group, preferring Bovidae, Canidae, and a variety of Carnivora hosts during the larval, nymphal, and adult stages, respectively. Results for the genus Dermacentor, with only 2 species included in analysis, showed that immatures mainly parasitize Muridae, while adults concentrated on Suidae.
To evaluate whether ticks of each genus were restricted to a wide or narrow host range according to PD of vertebrate families, we aimed to capture the PD of the hosts supporting every genus and stage of the ticks examined. For example, a tick genus may use several vertebrate families that are phylogenetically very related and thus uses a narrow range of hosts, or utilize a few vertebrate host families that are phylogenetically distant, thus covering a broad range of phylogenetic diversity. Analysis resulted in a tree containing data on the phylogenetic relationships of 92 vertebrate families (S1 Fig). We found that the PD of the different genera and stages of ticks varied highly (summarized in Table 1). With the exception of the genus Dermacentor, the genus Hyalomma showed the lowest PD, even though Hyalomma spp. can parasitizes a number of host families similar to other tick genera. Although the low PD for genus Dermacentor was notable, only 2 species of this genus were included in the study, compared to 18 species of Hyalomma ticks.
From the results in Table 1 and Fig 1, we summarized both BNC and PD for each host family, tick genus, and tick life stage (Figs 2 and 3). Hyalomma was i) the genus with lowest phylogenetic diversity of hosts during all 3 developmental stages; and ii) the only genus the immatures of which concentrate on Leporidae and Muridae hosts while the adults fed mainly on Bovidae. This was demonstrated by high BNC values of these vertebrate families for Hyalomma spp., which was not seen for other tick genera. Most important in this context, adult Hyalomma spp. ticks have also been found associated with Leporidae hosts. Additionally, a few Aves species have a relatively high importance as hosts for Hyalomma spp., and are involved mainly in circulating immature ticks.
Many complex systems display a surprising degree of error tolerance. However, networks with prominent hubs have low resilience and are extremely vulnerable to attacks (that is, to the selection and removal of a few nodes that play a vital role in maintaining the connectivity of the network) [29]. We aimed to capture the behavior of the networks of each tick genus and stage after recursive removal of host nodes. Resilience is an important feature of the ecological networks in which some organisms (ticks) depend on the presence of others (hosts) that may be key for the circulation of the parasite. After removing each node, the complete network was recalculated and its connectivity was re-evaluated. The percent of connectivity loss was the key measure of the resilience of the network to attack (random attack, or removing in order of decreasing BNC or decreasing WD). These calculations could not be done for the genus Dermacentor, because the connectivity dropped unrealistically after the removal of only a few nodes due to the limited number of species studied.
The networks of every stage and genus of tick analyzed were very resilient to random removal of hosts, all of them resulting in a loss of ~75% of connectivity after the removal of 50% of the nodes according to decreasing BNC of hosts (Figs 4 and 5). Random removal of host nodes promoted higher loss of connectivity in every network. However, lowest resilience of the networks was obtained when they were subjected to removal of hosts according to their WD; Hyalomma larvae and nymph results were deeply affected by removing as few as 2% of hosts with the highest WD. Removing Bovidae, Muridae, and Leporidae, host families that have been reported to develop consistent viremia upon CCHFV infection, resulted in almost complete collapse of the Hyalomma spp. larvae and nymph networks. Removing Bovidae, Leporidae, and Suidae promoted a ~50% loss of connectivity for Hyalomma spp. adults. Although immatures of the genus Amblyomma were also affected by the removal of hosts according to their WD, these ticks were most affected by removing lizards and amphibians, which are not known to be involved in the CCHFV lifecycle. The ecological significance of these findings is that Amblyomma spp. larvae and Hyalomma spp. immatures depend on key vertebrate families as hosts. Notably, the hosts on which Hyalomma spp. mostly depend are of pivotal importance for CCHFV circulation.
The main vectors of CCHFV are considered to be ticks of the genus Hyalomma. However, viral transmission has been confirmed under laboratory conditions in ticks of other genera co-occurring with Hyalomma. What differentiates CCHFV vector capacity of Hyalomma spp. from that of other ticks, including those capable of virus transmission, is not clear. We examined ecological factors to investigate whether special characteristics of the communities of hosts used by each tick genus could have a role in CCHFV epidemiology and distribution. The main aim was to describe the ecological relationships among ticks and vertebrates, and to discern if distinct interactions could capture the prominent role of the tick genus Hyalomma in CCHFV circulation.
CCHFV is well known to circulate through the 3 stages of the tick developmental cycle. The virus persists in ticks through the developmental stages by transstadial survival, and is maintained in new tick generations by transovarial passage [5,30]. Brief viremia in vertebrate tick hosts is the bridge by which the virus accesses other ticks. CCHFV can also infect ticks by transmission among co-feeding ticks, a process by which uninfected ticks feeding in close proximity with infected ticks on non-viremic hosts become infected [31]. Field surveys systematically report clumped distributions of ticks on vertebrates [32]: a few hosts carry large numbers of ticks aggregated in close proximity, while most of the hosts carry few or no ticks. This is of particular interest for the co-feeding mechanism, since the ticks concentrate highly on small mammals, which develop longer viremia and serve as hosts for immature ticks (reviewed by [1]). Since these groups of hosts are important for CCHFV transmission, the preferences for them would concentrate most of the tick populations on key carriers for viral circulation, and increase the probability of CCHFV transmission by co-feeding.
Here, based on network analyses of CCHFV associated tick-host relationships, we found that every genus of ticks examined, except Dermacentor, had its own set of preferred hosts; the exception was probably because only 2 Dermacentor species were included in the dataset. Also, our results suggest that Rhipicephalus adults prefer a set of hosts completely different from those exploited by immatures of the same genus. Most importantly, we identified unequivocal host relationships of Hyalomma genus ticks. Immatures of this tick feed on only a few members of Rodentia, Lagomorpha, and Aves, while adults are tightly associated with large ungulates, with lower but still prominent preferences for Lagomorpha and Suidae.
Rodents, lagomorphs, and ungulates, but not birds (with the exception of ostriches [33]), have been demonstrated to develop viremia for a variable, but brief, period of time [1,34,35]. Values of centrality of these hosts show that they are of major importance for Hyalomma spp. Other genera of ticks may use similar groups of hosts, but are also widely distributed over many other host families. In other words, the immatures of the genus Hyalomma tend to concentrate and over-aggregate on vertebrates that have been shown to be important in CCHFV transmission, even if viremia in these hosts is transient. Other genera of ticks feed on these same hosts, but also on a wide array of alternative hosts that are not known to circulate the virus. This feature has been called the dilution effect for other tick-transmitted pathogens, like B. burgdorferi [36,37]. While the effect seems not to be universal [38,39], an adequate balance of carrier and non-carrier hosts that ticks can use would attenuate the prevalence of a pathogen in ticks. We explicitly propose that the particular associations of Hyalomma spp. with their hosts are responsible for the prominent role of this tick genus in CCHFV circulation. It must to be noted that, depending on abundance of various host species in a given region, ticks of other genera could theoretically circulate the virus also, as demonstrated in laboratory protocols [4].
This scenario of immature Hyalomma ticks over-aggregating on some key vertebrates while adults infest large ungulates (which may be also viremic) has been reported as the main driver of CCHF epidemics [40]. Abandoned agricultural areas become populated by large patches of natural flora, facilitating shelter for rodents, birds, wild suids, and wild ungulates. We hypothesize that the overpopulation of these key hosts increases the abundance of Hyalomma spp. ticks, which could fuel CCHFV prevalence rates in the vectors in a feedback mechanism. As more ticks become infected, the probability of infecting reservoirs and naïve ticks and of transmittng the virus to humans increases.
Further analyses suggest that Hyalomma spp. ticks are associated with a phylogenetically narrow spectrum of hosts, accounting for the low resilience of the network of hosts for this tick genus. The lowest PD value was obtained for genus Hyalomma together with Dermacentor, but the result for the latter was considered biased, since only 2 Dermacentor species were included in this study. Other genera of ticks, including Ixodes, Amblyomma, and Rhipicephalus, had significantly higher values of host PD, suggesting that these ticks feed on a much wider range of hosts. In other words, the lack of significant host preference for the immatures of these tick genera would result in greater stochasticity in tick abundance, affecting the circulation of the virus.
CCHFV epidemiology is characterized by silent persistence of virus foci with intermittent epidemics. When abundance of key hosts is low, Hyalomma tick populations could also remain low, enough to further circulate CCHFV (probably only by transovarial passage [31]), but well below the threshold (R0) necessary to break the barrier of contact with humans. Small changes in the composition of vertebrate hosts could slighlty increase the value of R0, leading to the few CCHF cases reported annually in endemic countries [41]. Expansion of key host populations would lead to CCHF epidemics. It is necessary to stress that the only data about CCHFV distribution come from the detection of human clinical cases. Therefore, no information exists about tick densities or serology in hosts for areas where the virus circulates at levels below the epidemic threshold. When the key hosts for CCHFV circulation are absent, immature Hyalomma spp. ticks would use other hosts that are not viremic, deeply affecting the prevalence of the virus in these vectors.
The approach of this study is purely ecological and is based on the tenets of the network theory, which has deep roots in social behavior [42], links among computers [43], or mutualism between plants and pollinators [44,45]. An unbalanced representation of the tick-host interactions could constrain the results of this development, since poorly collected species could introduce a bias in the total number of records. We, however, evaluated the strength of associations between partners using a purposely inclusive systematic division of ticks (genera) and hosts (families) to prevent the noise generated by undersampled species, together with robust markers representing the relationships in directed networks [46]. This approach guarantees a minimum bias in indexes of the network but not in PD estimations. Furthermore, this approach reduces potential for biogeographical bias based on co-distribution of both tick and hosts in cluster analysis, and supports that the reported cluster formation is derived from an acual preference towards particular vertebrates.
A species-by-species analysis of tick-host relationships is not possible here, as detailed earlier. Our broad approach does not account for CCHFV strains (e.g., AP92, lineage Europe 2) that, in addition to circulation by Hyalomma spp. ticks [47], have been suggested to be circulated predominantly by other species such as those of the genus Rhipicephalus [48]. It should be noted, however, that no proof of the vectoral capacity of Rhipicephalus spp. ticks for strain AP92 has ever been obtained under adequate laboratory conditions [2]. A broad approach also prevents the ability to delineate diverse viral lineages circulated by different species of Hyalomma in our analyses. However, these relationships, which are likely due to overlapping geographical ranges of both ticks and viral strains, do not affect the observations detailed here. Analyses are based on inclusion of all tick reports, irrespective of Hyalomma species, and serves as a broad investigation on what differentiates Hyalomma ticks ecologically from other genera of potential CCHFV tick vectors.
While our approach is validated by field data supporting the importance of a critical combination of hosts that coexist during the life cycle of Hyalomma ticks, we must take into account the complete lack of data regarding the intimate molecular relationships of CCHFV with the tick. The tick gut represents the first barrier against pathogens, and the gut cell membrane is the key to dissemination of the virus into the body of the vector. The need to conduct these studies under high biocontainment has precluded the understanding of basic mechanisms that CCHFV uses to enter the tick gut and to disseminate to salivary glands for further circulation. While the ecological hypothesis that we outlined here provides a meanignful interpretation of CCHFV dynamics in the field, the relationships between the molecular machinery of the virus and the tick as an environment must be understood.
Our current knowledge on CCHFV distribution has been gathered from reported human clinical cases, which provide a fragmented picture of the much wider geographical range of the virus. Surveying the virus in wild hosts and questing ticks, together with an extensive record of tick-host relationships, is urgently needed to update exisiting data about this potentially lethal agent. Viral foci must also be associated with adequate definitions of the environmnetal niche to make sense of the elusive behaviour of the so-called silent foci. These studies, together with a deeper knowledge of the molecular mechanisms shaping the virus-vector interactions, are esential to identify the routes of CCHFV circulation and the exposed populations, and to outline adequate preventive mesaures.
|
10.1371/journal.pcbi.1003408 | Searching for Collective Behavior in a Large Network of Sensory Neurons | Maximum entropy models are the least structured probability distributions that exactly reproduce a chosen set of statistics measured in an interacting network. Here we use this principle to construct probabilistic models which describe the correlated spiking activity of populations of up to 120 neurons in the salamander retina as it responds to natural movies. Already in groups as small as 10 neurons, interactions between spikes can no longer be regarded as small perturbations in an otherwise independent system; for 40 or more neurons pairwise interactions need to be supplemented by a global interaction that controls the distribution of synchrony in the population. Here we show that such “K-pairwise” models—being systematic extensions of the previously used pairwise Ising models—provide an excellent account of the data. We explore the properties of the neural vocabulary by: 1) estimating its entropy, which constrains the population's capacity to represent visual information; 2) classifying activity patterns into a small set of metastable collective modes; 3) showing that the neural codeword ensembles are extremely inhomogenous; 4) demonstrating that the state of individual neurons is highly predictable from the rest of the population, allowing the capacity for error correction.
| Sensory neurons encode information about the world into sequences of spiking and silence. Multi-electrode array recordings have enabled us to move from single units to measuring the responses of many neurons simultaneously, and thus to ask questions about how populations of neurons as a whole represent their input signals. Here we build on previous work that has shown that in the salamander retina, pairs of retinal ganglion cells are only weakly correlated, yet the population spiking activity exhibits large departures from a model where the neurons would be independent. We analyze data from more than a hundred salamander retinal ganglion cells and characterize their collective response using maximum entropy models of statistical physics. With these models in hand, we can put bounds on the amount of information encoded by the neural population, constructively demonstrate that the code has error correcting redundancy, and advance two hypotheses about the neural code: that collective states of the network could carry stimulus information, and that the distribution of neural activity patterns has very nontrivial statistical properties, possibly related to critical systems in statistical physics.
| Physicists have long hoped that the functional behavior of large, highly interconnected neural networks could be described by statistical mechanics [1]–[3]. The goal of this effort has been not to simulate the details of particular networks, but to understand how interesting functions can emerge, collectively, from large populations of neurons. The hope, inspired by our quantitative understanding of collective behavior in systems near thermal equilibrium, is that such emergent phenomena will have some degree of universality, and hence that one can make progress without knowing all of the microscopic details of each system. A classic example of work in this spirit is the Hopfield model of associative or content–addressable memory [1], which is able to recover the correct memory from any of its subparts of sufficient size. Because the computational substrate of neural states in these models are binary “spins,” and the memories are realized as locally stable states of the network dynamics, methods of statistical physics could be brought to bear on theoretically challenging issues such as the storage capacity of the network or its reliability in the presence of noise [2], [3]. On the other hand, precisely because of these abstractions, it has not always been clear how to bring the predictions of the models into contact with experiment.
Recently it has been suggested that the analogy between statistical physics models and neural networks can be turned into a precise mapping, and connected to experimental data, using the maximum entropy framework [4]. In a sense, the maximum entropy approach is the opposite of what we usually do in making models or theories. The conventional approach is to hypothesize some dynamics for the network we are studying, and then calculate the consequences of these assumptions; inevitably, the assumptions we make will be wrong in detail. In the maximum entropy method, however, we are trying to strip away all our assumptions, and find models of the system that have as little structure as possible while still reproducing some set of experimental observations.
The starting point of the maximum entropy method for neural networks is that the network could, if we don't know anything about its function, wander at random among all possible states. We then take measured, average properties of the network activity as constraints, and each constraint defines some minimal level of structure. Thus, in a completely random system neurons would generate action potentials (spikes) or remain silent with equal probability, but once we measure the mean spike rate for each neuron we know that there must be some departure from such complete randomness. Similarly, absent any data beyond the mean spike rates, the maximum entropy model of the network is one in which each neuron spikes independently of all the others, but once we measure the correlations in spiking between pairs of neurons, an additional layer of structure is required to account for these data. The central idea of the maximum entropy method is that, for each experimental observation that we want to reproduce, we add only the minimum amount of structure required.
An important feature of the maximum entropy approach is that the mathematical form of a maximum entropy model is exactly equivalent to a problem in statistical mechanics. That is, the maximum entropy construction defines an “effective energy” for every possible state of the network, and the probability that the system will be found in a particular state is given by the Boltzmann distribution in this energy landscape. Further, the energy function is built out of terms that are related to the experimental observables that we are trying to reproduce. Thus, for example, if we try to reproduce the correlations among spiking in pairs of neurons, the energy function will have terms describing effective interactions among pairs of neurons. As explained in more detail below, these connections are not analogies or metaphors, but precise mathematical equivalencies.
Minimally structured models are attractive, both because of the connection to statistical mechanics and because they represent the absence of modeling assumptions about data beyond the choice of experimental constraints. Of course, these features do not guarantee that such models will provide an accurate description of a real system. They do, however, give us a framework for starting with simple models and systematically increasing their complexity without worrying that the choice of model class itself has excluded the “correct” model or biased our results. Interest in maximum entropy approaches to networks of real neurons was triggered by the observation that, for groups of up to 10 ganglion cells in the vertebrate retina, maximum entropy models based on the mean spike probabilities of individual neurons and correlations between pairs of cells indeed generate successful predictions for the probabilities of all the combinatorial patterns of spiking and silence in the network as it responds to naturalistic sensory inputs [4]. In particular, the maximum entropy approach made clear that genuinely collective behavior in the network can be consistent with relatively weak correlations among pairs of neurons, so long as these correlations are widespread, shared among most pairs of cells in the system. This approach has now been used to analyze the activity in a variety of neural systems [5]–[15], the statistics of natural visual scenes [16]–[18], the structure and activity of biochemical and genetic networks [19], [20], the statistics of amino acid substitutions in protein families [21]–[27], the rules of spelling in English words [28], the directional ordering in flocks of birds [29], and configurations of groups of mice in naturalistic habitats [30].
One of the lessons of statistical mechanics is that systems with many degrees of freedom can behave in qualitatively different ways from systems with just a few degrees of freedom. If we can study only a handful of neurons (e.g., N∼10 as in Ref [4]), we can try to extrapolate based on the hypothesis that the group of neurons that we analyze is typical of a larger population. These extrapolations can be made more convincing by looking at a population of N = 40 neurons, and within such larger groups one can also try to test more explicitly whether the hypothesis of homogeneity or typicality is reliable [6], [9]. All these analyses suggest that, in the salamander retina, the roughly 200 interconnected neurons that represent a small patch of the visual world should exhibit dramatically collective behavior. In particular, the states of these large networks should cluster around local minima of the energy landscape, much as for the attractors in the Hopfield model of associative memory [1]. Further, this collective behavior means that responses will be substantially redundant, with the behavior of one neuron largely predictable from the state of other neurons in the network; stated more positively, this collective response allows for pattern completion and error correction. Finally, the collective behavior suggested by these extrapolations is a very special one, in which the probability of particular network states, or equivalently the degree to which we should be surprised by the occurrence of any particular state, has an anomalously large dynamic range [31]. If correct, these predictions would have a substantial impact on how we think about coding in the retina, and about neural network function more generally. Correspondingly, there is some controversy about all these issues [32]–[35].
Here we return to the salamander retina, in experiments that exploit a new generation of multi–electrode arrays and associated spike–sorting algorithms [36]. As schematized in Figure 1, these methods make it possible to record from ganglion cells in the relevant densely interconnected patch, while projecting natural movies onto the retina. Access to these large populations poses new problems for the inference of maximum entropy models, both in principle and in practice. What we find is that, with extensions of algorithms developed previously [37], it is possible to infer maximum entropy models for more than one hundred neurons, and that with nearly two hours of data there are no signs of “overfitting” (cf. [15]). We have built models that match the mean probability of spiking for individual neurons, the correlations between spiking in pairs of neurons, and the distribution of summed activity in the network (i.e., the probability that K out of the N neurons spike in the same small window of time [38]–[40]). We will see that models which satisfy all these experimental constraints provide a strikingly accurate description of the states taken on by the network as a whole, that these states are collective, and that the collective behavior predicted by our models has implications for how the retina encodes visual information.
The idea of maximizing entropy has its origin in thermodynamics and statistical mechanics. The idea that we can use this principle to build models of systems that are not in thermal equilibrium is more recent, but still more than fifty years old [41]; in the past few years, there has been a new surge of interest in the formal aspects of maximum entropy constructions for (out-of-equilibrium) spike rasters (see, e.g., [42]). Here we provide a description of this approach which we hope makes the ideas accessible to a broad audience.
We imagine a neural system exposed to a stationary stimulus ensemble, in which simultaneous recordings from N neurons can be made. In small windows of time, as we see in Figure 1, a single neuron i either does () or does not () generate an action potential or spike [43]; the state of the entire network in that time bin is therefore described by a “binary word” . As the system responds to its inputs, it visits each of these states with some probability . Even before we ask what the different states mean, for example as codewords in a representation of the sensory world, specifying this distribution requires us to determine the probability of each of possible states. Once N increases beyond ∼20, brute force sampling from data is no longer a general strategy for “measuring” the underlying distribution.
Even when there are many, many possible states of the network, experiments of reasonable size can be sufficient to estimate the averages or expectation values of various functions of the state of the system, , where the averages are taken across data collected over the course of the experiment. The goal of the maximum entropy construction is to search for the probability distribution that matches these experimental measurements but otherwise is as unstructured as possible. Minimizing structure means maximizing entropy [41], and for any set of moments or statistics that we want to match, the form of the maximum entropy distribution can be found analytically:(1)(2)(3)where is the effective “energy” function or the Hamiltonian of the system, and the partition function ensures that the distribution is normalized. The couplings must be set such that the expectation values of all constraint functions , , over the distribution P match those measured in the experiment:(4)These equations might be hard to solve, but they are guaranteed to have exactly one solution for the couplings given any set of measured expectation values [44].
Why should we study the neural vocabulary, , at all? In much previous work on neural coding, the focus has been on constructing models for a “codebook” which can predict the response of the neurons to arbitrary stimuli, [14], [45], or on building a “dictionary” that describes the stimuli consistent with particular patterns of activity, [43]. In a natural setting, stimuli are drawn from a space of very high dimensionality, so constructing these “encoding” and “decoding” mappings between the stimuli and responses is very challenging and often involves making strong assumptions about how stimuli drive neural spiking (e.g. through linear filtering of the stimulus) [45]–[48]. While the maximum entropy framework itself can be extended to build stimulus-dependent maximum entropy models for and study detailed encoding and decoding mappings [14], [49]–[51], we choose to focus here directly on the total distribution of responses, , thus taking a very different approach.
Already when we study the smallest possible network, i.e., a pair of interacting neurons, the usual approach is to measure the correlation between spikes generated in the two cells, and to dissect this correlation into contributions which are intrinsic to the network and those which are ascribed to common, stimulus driven inputs. The idea of decomposing correlations dates back to a time when it was hoped that correlations among spikes could be used to map the synaptic connections between neurons [52]. In fact, in a highly interconnected system, the dominant source of correlations between two neurons—even if they are entirely intrinsic to the network—will always be through the multitude of indirect paths involving other neurons [53]. Regardless of the source of these correlations, however, the question of whether they are driven by the stimulus or are intrinsic to the network is unlikely a question that the brain could answer. We, as external observers, can repeat the stimulus exactly, and search for correlations conditional on the stimulus, but this is not accessible to the organism, unless the brain could build a “noise model” of spontaneous activity of the retina in the absence of any stimuli and this model also generalized to stimulus-driven activity. The brain has access only to the output of the retina: the patterns of activity which are drawn from the distribution , rather than activity conditional on the stimulus, so the neural mechanism by which the correlations could be split into signal and noise components is unclear. If the responses are codewords for the visual stimulus, then the entropy of this distribution sets the capacity of the code to carry information. Word by word, determines how surprised the brain should be by each particular pattern of response, including the possibility that the response was corrupted by noise in the retinal circuit and thus should be corrected or ignored [54]. In a very real sense, what the brain “sees” are sequences of states drawn from . In the same spirit that many groups have studied the statistical structures of natural scenes [55]–[60], we would like to understand the statistical structure of the codewords that represent these scenes.
The maximum entropy method is not a model for network activity. Rather it is a framework for building models, and to implement this framework we have to choose which functions of the network state we think are interesting. The hope is that while there are states of the system as a whole, there is a much smaller number of measurements, , with and , which will be sufficient to capture the essential structure of the collective behavior in the system. We emphasize that this is a hypothesis, and must be tested. How should we choose the functions ? In this work we consider three classes of possibilities:
It is important that the mapping between maximum entropy models and a Boltzmann distribution with some effective energy function is not an analogy, but rather a mathematical equivalence. In using the maximum entropy approach we are not assuming that the system of interest is in some thermal equilibrium state (note that there is no explicit temperature in Eq (1)), nor are we assuming that there is some mysterious force which drives the system to a state of maximum entropy. We are also not assuming that the temporal dynamics of the network is described by Newton's laws or Brownian motion on the energy landscape. What we are doing is making models that are consistent with certain measured quantities, but otherwise have as little structure as possible. As noted above, this is the opposite of what we usually do in building models or theories—rather than trying to impose some hypothesized structure on the world, we are trying to remove all structures that are not explicitly contained within the chosen set of experimental constraints.
The mapping to a Boltzmann distribution is not an analogy, but if we take the energy function more literally we are making use of analogies. Thus, the term that emerges from constraining the mean spike probabilities of every neuron is analogous to a magnetic field being applied to each spin, where spin “up” () marks a spike and spin “down” () denotes silence. Similarly, the term that emerges from constraining the pairwise correlations among neurons corresponds to a “spin–spin” interaction which tends to favor neurons firing together () or not (). Finally, the constraint on the overall distribution of activity generates a term which we can interpret as resulting from the interaction between all the spins/neurons in the system and one other, hidden degree of freedom, such as an inhibitory interneuron. These analogies can be useful, but need not be taken literally.
We have applied the maximum entropy framework to the analysis of one large experimental data set on the responses of ganglion cells in the salamander retina to a repeated, naturalistic movie. These data are collected using a new generation of multi–electrode arrays that allow us to record from a large fraction of the neurons in a 450×450 µm patch, which contains a total of ∼200 ganglion cells [36], as in Figure 1. In the present data set, we have selected 160 neurons that pass standard tests for the stability of spike waveforms, the lack of refractory period violations, and the stability of firing across the duration of the experiment (see Methods and Ref [36]). The visual stimulus is a greyscale movie of swimming fish and swaying water plants in a tank; the analyzed chunk of movie is 19 s long, and the recording was stable through 297 repeats, for a total of more than 1.5 hrs of data. As has been found in previous experiments in the retinas of multiple species [4], [61]–[64], we found that correlations among neurons are most prominent on the ∼20 ms time scale, and so we chose to discretize the spike train into Δτ = 20 ms bins.
Maximum entropy models have a simple form [Eq (1)] that connects precisely with statistical physics. But to complete the construction of a maximum entropy model, we need to impose the condition that averages in the maximum entropy distribution match the experimental measurements, as in Eq (4). This amounts to finding all the coupling constants in Eq (2). This is, in general, a hard problem. We need not only to solve this problem, but also to convince ourselves that our solution is meaningful, and that it does not reflect overfitting to the limited set of data at our disposal. A detailed account of the numerical solution to this inverse problem is given in Methods: Learning maximum entropy models from data.
In Figure 2 we show an example of N = 100 neurons from a small patch of the salamander retina, responding to naturalistic movies. We notice that correlations are weak, but widespread, as in previous experiments on smaller groups of neurons [4], [6], [9], [65], [66]. Because the data set is very large, the threshold for reliable detection of correlations is very low; if we shuffle the data completely by permuting time and repeat indices independently for each neuron, the standard deviation of correlation coefficients,(18)is , as shown in Figure 2C, vastly smaller than the typical correlations that we observe (median , 90% of values between and ). More subtly, this means that only ∼6.3% percent of the correlation coefficients are within error bars of zero, and there is no sign that there is a large excess fraction of pairs that have truly zero correlation—the distribution of correlations across the population seems continuous. Note that, as customary, we report normalized correlation coefficients (, between −1 and 1), while maximum entropy formally constrains an equivalent set of unnormalized second order moments, [Eq (6)].
We began by constructing maximum entropy models that match the mean spike rates and pairwise correlations, i.e. “pairwise models,” whose distribution is, from Eqs (5, 7),(19)When we reconstruct the coupling constants of the maximum entropy model, we see that the “interactions” among neurons are widespread, and almost symmetrically divided between positive and negative values; for more details see Methods: Learning maximum entropy models from data. Figure 3 shows that the model we construct really does satisfy the constraints, so that the differences, for example, between the measured and predicted correlations among pairs of neurons are within the experimental errors in the measurements.
With N = 100 neurons, measuring the mean spike probabilities and all the pairwise correlations means that we estimate separate quantities. This is a large number, and it is not clear that we are safe in taking all these measurements at face value. It is possible, for example, that with a finite data set the errors in the different elements of the correlation matrix are sufficiently strongly correlated that we don't really know the matrix as a whole with high precision, even though the individual elements are measured very accurately. This is a question about overfitting: is it possible that the parameters are being finely tuned to match even the statistical errors in our data?
To test for overfitting (Figure 4), we exploit the fact that the stimuli consist of a short movie repeated many times. We can choose a random 90% of these repeats from which to learn the parameters of the maximum entropy model, and then check that the probability of the data in the other 10% of the experiment is predicted to be the same, within errors. We see in Figure 4 that this is true, and that it remains true as we expand from N = 10 neurons (for which we surely have enough data) out to N = 120, where we might have started to worry. Taken together, Figures 2, 3, and 4 suggest strongly that our data and algorithms are sufficient to construct maximum entropy models, reliably, for networks of more than one hundred neurons.
How well do our maximum entropy models describe the behavior of large networks of neurons? The models predict the probability of occurrence for all possible combinations of spiking and silence in the network, and it seems natural to use this huge predictive power to test the models. In small networks, this is a useful approach. Indeed, much of the interest in the maximum entropy approach derives from the success of models based on mean spike rates and pairwise correlations, as in Eq (19), in reproducing the probability distribution over states in networks of size [4], [5]. With N = 10, there are possible combinations of spiking and silence, and reasonable experiments are sufficiently long to estimate the probabilities of all of these individual states. But with N = 100, there are possible states, and so it is not possible to “just measure” all the probabilities. Thus, we need another strategy for testing our models.
Striking (and model–independent) evidence for nontrivial collective behavior in these networks is obtained by asking for the probability that K out of the N neurons generate a spike in the same small window of time, as shown in Figure 5. This distribution, , should become Gaussian at large N if the neurons are independent, or nearly so, and we have noted that the correlations between pairs of cells are weak. Thus is very well approximated by an independent model, with fractional errors on the order of the correlation coefficients, typically less than ∼10%. But, even in groups of N = 10 cells, there are substantial departures from the predictions of an independent model (Figure 5A). In groups of N = 40 cells, we see K = 10 cells spiking synchronously with probability ∼104 times larger than expected from an independent model (Figure 5B), and the departure from independence is even larger at N = 100 (Figure 5C) [12], [15].
Maximum entropy models that match the mean spike rate and pairwise correlations in a network make an unambiguous, quantitative prediction for , with no adjustable parameters. In smaller groups of neurons, certainly for N = 10, this prediction is quite accurate, and accounts for most of the difference between the data and the expectations from an independent model, as shown in Figure 5. But even at N = 40 we see small deviations between the data and the predictions of the pairwise model. Because the silent state is highly probable, we can measure very accurately, and the pairwise models make errors of nearly a factor of three at N = 100, and independent models are off by a factor of about twenty. The pairwise model errors in are negligible when compared to the many orders of magnitude differences from an independent model, but they are highly significant. The pattern of errors also is important, since in the real networks silence persists as being highly probable even at N = 120—with indications that this surprising trend might continue towards larger N [39] —and the pairwise model doesn't quite capture this.
If a model based on pairwise correlations doesn't quite account for the data, it is tempting to try and include correlations among triplets of neurons. But at N = 100 there are of these triplets, so a model that includes these correlations is much more complex than one that stops with pairs. An alternative is to use itself as a constraint on our models, as explained above in relation to Eq (17). This defines the “K-pairwise model,”(20)where the “potential” V is chosen to match the observed distribution . As noted above, we can think of this potential as providing a global regulation of the network activity, such as might be implemented by inhibitory interneurons with (near) global connectivity. Whatever the mechanistic interpretation of this model, it is important that it is not much more complex than the pairwise model: matching adds only ∼N parameters to our model, while the pairwise model already has parameters. All of the tests given in the previous section can be redone in this case, and again we find that we can learn the K-pairwise models from the available data with no signs of overfitting. Figure 6 shows the parameters of the K-pairwise model for the same group of N = 100 neurons shown in Figure 2. Notice that the pairwise interaction terms remain roughly the same; the local fields are also similar but have a shift towards more negative values.
Since we didn't make explicit use of the triplet correlations in constructing the K-pairwise model, we can test the model by predicting these correlations. In Figure 7A we show(21)as computed from the real data and from the models, for a single group of N = 100 neurons. We see that pairwise models capture the rankings of the different triplets, so that more strongly correlated triplets are predicted to be more strongly correlated, but these models miss quantitatively, overestimating the positive correlations and failing to predict significantly negative correlations. These errors are largely corrected in the K-pairwise model, despite the fact that adding a constraint on doesn't add any information about the identity of the neurons in the different triplets. Specifically, Figure 7A shows that the biases of the pairwise model in the prediction of three-point correlations have been largely removed (with some residual deviations at large absolute values of the three-point correlation) by adding the K-spike constraint; on the other hand, the variance of predictions across bins containing three-point correlations of approximately the same magnitude did not decrease substantially. It is also interesting that this improvement in our predictions (as well as that in Figure 8 below) occurs even though the numerical value of the effective potential is quite small, as shown in Figure 6E (quantitatively, in an example group of N = 100 neurons, the variance in energy associated with the V(K) potential accounts roughly for only 5% of the total variance in energy). Fixing the distribution of global activity thus seems to capture something about the network that individual spike probabilities and pairwise correlations have missed.
An interesting effect is shown in Figure 7B, where we look at the average absolute deviation between predicted and measured , as a function of the group size N. With increasing N the ratio between the total number of (predicted) three-point correlations and (fitted) model parameters is increasing (from ≈2 at N = 10 to ≈40 for N = 120), leading us to believe that predictions will grow progressively worse. Nevertheless, the average error in three-point prediction stays constant with network size, for both pairwise and K-pairwise models. An attractive explanation is that, as N increases, the models encompass larger and larger fractions of the interacting neural patch and thus decrease the effects of “hidden” units, neurons that are present but not included in the model; such unobserved units, even if they only interacted with other units in a pairwise fashion, could introduce effective higher-order interactions between observed units, thereby causing three-point correlation predictions to deviate from those of the pairwise model [67]. The accuracy of the K-pairwise predictions is not quite as good as the errors in our measurements (dashed line in Figure 7B), but still very good, improving by a factor of ∼2 relative to the pairwise model to well below 10−3.
Maximum entropy models assign an effective energy to every possible combination of spiking and silence in the network, from Eq (20). Learning the model means specifying all the parameters in this expression, so that the mapping from states to energies is completely determined. The energy determines the probability of the state, and while we can't estimate the probabilities of all possible states, we can ask whether the distribution of energies that we see in the data agrees with the predictions of the model. Thus, if we have a set of states drawn out of a distribution , we can count the number of states that have energies lower than E,(22)where is the Heaviside step function,(23)Similarly, we can count the number of states that have energy larger than E,(24)Now we can take the distribution to be the distribution of states that we actually see in the experiment, or we can take it to be the distribution predicted by the model, and if the model is accurate we should find that the cumulative distributions are similar in these two cases. Results are shown in Figure 8A (analogous results for the pairwise model are shown in Figure S5). Figure 8B focuses on the agreement between the first two moments of the distribution of energies, i.e., the mean and variance , as a function of the network size N, showing that the K-pairwise model is significantly better at matching the variance of the energies relative to the pairwise model.
We see that the distributions of energies in the data and the model are very similar. There is an excellent match in the “low energy” (high probability) region, and then as we look at the high energy tail () we see that theory and experiment match out to probabilities of better than . Thus the distribution of energies, which is an essential construct of the model, seems to match the data across >90% of the states that we see.
The successful prediction of the cumulative distribution is especially striking because it extends to E∼25. At these energies, the probability of any single state is predicted to be , which means that these states should occur roughly once per fifty years (!). This seems ridiculous—what are such rare states doing in our analysis, much less as part of the claim that theory and experiment are in quantitative agreement? The key is that there are many, many of these rare states—so many, in fact, that the theory is predicting that ∼10% of the all the states we observe will be (at least) this rare: individually surprising events are, as a group, quite common. In fact, of the combinations of spiking and silence ( distinct ones) that we see in subnetworks of N = 120 neurons, of these occur only once, which means we really don't know anything about their probability of occurrence. We can't say that the probability of any one of these rare states is being predicted correctly by the model, since we can't measure it, but we can say that the distribution of (log) probabilities—that is, the distribution of energies—across the set of observed states is correct, down to the ∼10% level. The model thus is predicting things far beyond what can be inferred directly from the frequencies with which common patterns are observed to occur in realistic experiments.
Finally, the structure of the models we are considering is that the state of each neuron—an Ising spin—experiences an “effective field” from all the other spins, determining the probability of spiking vs. silence. This effective field consists of an intrinsic bias for each neuron, plus the effects of interactions with all the other neurons:(25)If the model is correct, then the probability of spiking is simply related to the effective field,(26)To test this relationship, we can choose one neuron, compute the effective field from the states of all the other neurons, at every moment in time, then collect all those moments when is in some narrow range, and see how often the neuron spikes. We can then repeat this for every neuron, in turn. If the model is correct, spiking probability should depend on the effective field according to Eq (26). We emphasize that there are no new parameters to be fit, but rather a parameter–free relationship to be tested. The results are shown in Figure 9. We see that, throughout the range of fields that are well sampled in the experiment, there is good agreement between the data and Eq (26). As we go into the tails of the distribution, we see some deviations, but error bars also are (much) larger.
We have seen that it is possible to construct maximum entropy models which match the mean spike probabilities of each cell, the pairwise correlations, and the distribution of summed activity in the network, and that our data are sufficient to insure that all the parameters of these models are well determined, even when we consider groups of N = 100 neurons or more. Figures 7 through 9 indicate that these models give a fairly accurate description of the distribution of states—the myriad combinations of spiking and silence—taken on by the network as a whole. In effect we have constructed a statistical mechanics for these networks, not by analogy or metaphor but in quantitative detail. We now have to ask what we can learn about neural function from this description.
It is widely agreed that neural activity in the brain is more than the sum of its parts—coherent percepts, thoughts, and actions require the coordinated activity of many neurons in a network, not the independent activity of many individual neurons. It is not so clear, however, how to build bridges between this intuition about collective behavior and the activity of individual neurons.
One set of ideas is that the activity of the network as a whole may be confined to some very low dimensional trajectory, such as a global, coherent oscillation. Such oscillatory activity is observable in the summed electrical activity of large numbers of neurons—the EEG—and should be reflected as oscillations in the (auto–)correlation functions of spike trains from individual neurons. On a more refined level, dimensionality reduction techniques like PCA allow the activity patterns of a neural network to be viewed on a low-dimensional manifold, facilitating visualization and intuition [85]–[88]. A very different idea is provided by the Hopfield model, in which collective behavior is expressed in the stabilization of many discrete patterns of activity, combinations of spiking and silence across the entire network [1], [2]. Taken together, these many patterns can span a large fraction of the full space of possibilities, so that there need be no dramatic dimensionality reduction in the usual sense of this term.
The claim that a network of neurons exhibits collective behavior is really the claim that the distribution of states taken on by the network has some nontrivial structure that cannot be factorized into contributions from individual cells or perhaps even smaller subnetworks. Our goal in this work has been to build a model of this distribution, and to explore the structure of that model. We emphasize that building a model is, in this view, the first step rather than the last step. But building a model is challenging, because the space of states is very large and data are limited.
An essential step in searching for collective behavior has been to develop experimental techniques that allow us to record not just from a large number of neurons, but from a large fraction of the neurons in a densely interconnected region of the retina [36], [89]. In large networks, even measuring the correlations among pairs of neurons can become problematic: individual elements of the correlation matrix might be well determined from small data sets, but much larger data sets are required to be confident that the matrix as a whole is well determined. Thus, long, stable recordings are even more crucial than usual.
To use the maximum entropy approach, we have to be sure that we can actually find the models that reproduce the observed expectation values (Figure 2, 3) and that we have not, in the process, fit to spurious correlations that arise from the finite size of our data set (Figure 4). Once these tests are passed, we can start to assess the accuracy of the model as a description of the network as a whole. In particular, we found that the pairwise model began to break down at a network size (Figure 5). However, by adding the K-spike constraint that reproduces the probability of K out of N neurons spiking synchronously (Figure 6), which is a statistic that is well-sampled and does not greatly increase the model's complexity, we could again recover good performance (Figures 7–9). Although the primary goal of this work was to examine the responses of the retina under naturalistic stimulation, we also checked that the K-pairwise models are able to capture the joint behavior of retinal ganglion cells under a very different, random checkerboard stimulation (Figure S7). Despite a significantly smaller amount of total correlation (and a complete lack of long-range correlation) in the checkerboard stimuli compared to natural scenes, the pairwise model still deviated significantly from data at large N and the K-spike constraint proved necessary; this happened even though the total amount of correlation in the codewords is smaller for the checkerboard stimulus. Characterizing more completely the dependence of the statistical properties of the neural code on the stimulus type therefore seems like one of the interesting avenues for future work.
How can we interpret the meaning of the K-spike constraint and its biological relevance? One possibility would be to view it as a global modulatory effect of, e.g., inhibitory interneurons with dense connectivity. Alternatively, might be an important feature of the neural code for downstream neurons. For example, if a downstream neuron sums over its inputs and fires whenever the sum exceeds a threshold, will be informative about the rate of such threshold crossings. We note that is a special case of a more general linear function, , where are arbitrary weights and θ is a threshold ( and for K). An interesting question to explore in the future would be to ask if the K-statistic really is the most informative choice that maximally reduces the entropy of the K-pairwise model relative to the pairwise model, or whether additional modeling power could be gained by optimizing the weights , perhaps even by adding several such projection vectors as constraints. In any case, the K-pairwise model should not be regarded as “the ultimate model” of the retinal code: it is a model that emerged from pairwise constructions in a data-driven attempt to account for systematic deficiencies of Ising-like models on large populations. Similarly, systematic deviations that remain, e.g., at the low and high ends of the effective field as in Figure 9, might inform us about further useful constraints that could improve the model. These could include either new higher-order interactions, global constraints, or couplings across time bins, as suggested by Refs [12], [90].
Perhaps the most useful global test of our models is to ask about the distribution of state probabilities: how often should we see combinations of spiking and silence that occur with probability P? This has the same flavor as asking for the probability of every state, but does not suffer from the curse of dimensionality. Since maximum entropy models are mathematically identical to the Boltzmann distribution in statistical mechanics, this question about the frequency of states with probability P is the same as asking how many states have a given energy E; we can avoid binning along the E axis by asking for the number of states with energies smaller (higher probability) or larger (lower probability) than E. Figure 8 shows that these cumulative distributions computed from the model agree with experiment far into the tail of low probability states. These states are so rare that, individually, they almost never occur, but there are so many of these rare states that, in aggregate, they make a measurable contribution to the distribution of energies. Indeed, most of the states that we see in the data are rare in this sense, and their statistical weight is correctly predicted by the model.
The maximum entropy models that we construct from the data do not appear to simplify along any conventional axes. The matrix of correlations among spikes in different cells (Figure 1A) is of full rank, so that principal component analysis does not yield significant dimensionality reduction. The matrix of “interactions” in the model (Figure 1D) is neither very sparse nor of low rank, perhaps because we are considering a group of neurons all located (approximately) within the radius of the typical dendritic arbor, so that all cells have a chance to interact with one another. Most importantly, the interactions that we find are not weak (Figure 1F), and together with being widespread this means that their impact is strong. Technically, we cannot capture the impact within low orders of perturbation theory (Methods: Are the networks in the perturbative regime?), but qualitatively this means that the behavior of the network as a whole is not in any sense “close” to the behavior of non–interacting neurons. Thus, the reason that our models work is likely not because the correlations are weak, as had been suggested [34].
Having convinced ourselves that we can build models which give an accurate description of the probability distribution over the states of spiking and silence in the network, we can ask what these models teach us about function. As emphasized in Ref [4], one corollary of collective behavior is the possibility of error correction or pattern completion—we can predict the spiking or silence of one neuron by knowing the activity of all the other neurons. With a population of N = 100 cells, the quality of these predictions becomes quite high (Figure 15). The natural way of testing these predictions is to look at the probability of spiking vs. time in the stimulus movie. Although we make no reference to the stimulus, we reproduce the sharp peaks of activity and extended silences that are so characteristic of the response to naturalistic inputs, and so difficult to capture in conventional models where each individual neuron responds to the visual stimulus as seen through its receptive field [78].
One of the dominant concepts in thinking about the retina has been the idea that the structure of receptive fields serves to reduce the redundancy of natural images and enhance the efficiency of information transmission to the brain [91]–[94] (but see [65], [95]). While one could argue that the observed redundancy among neurons is less than expected from the structure of natural images or movies, none of what we have described here would happen if the retina truly “decorrelated” its inputs. Far from being almost independent, the activity of single neurons is predicted very well by the state of the remaining network, and the combinations of spiking and silence in different cells cluster into basins of attraction defined by the local minima of energy in our models. While it is intriguing to think about the neural code as being organized around the “collective metastable states,” some of which we have identified using the maximum entropy model, further work is necessary to explore this idea in detail. Unlike our other results, where we could either compare parameter-free predictions to data, or put a bound on the entropy of the code, it is harder to compare the model's energy landscape (and its local minima) to the true energy landscape, for which we would need to be able to estimate all pattern probabilities directly from data. It is therefore difficult to assess how dependent the identified collective states are on the form of the model. Nevertheless, for any particular model assignment of activity patterns to collective states, one could ask how well those collective modes capture the information about the stimuli, and use that as a direct measure of model performance. We believe this to be a promising avenue for future research.
With N = 120 neurons, our best estimate of the entropy corresponds to significant occupancy of roughly one million distinct combinations of spiking and silence. Each state could occur with a different probability, and (aside from normalization) there are no constraints—each of these probabilities could be seen as a separate parameter describing the network activity. It is appealing to think that there must be some simplification, that we won't need a million parameters, but it is not obvious that any particular simplification strategy will work. Indeed, there has been the explicit claim that maximum entropy approach has been successful on small () groups of neurons simply because a low-order maximum model will generically approximate well any probability distribution that is sufficiently sparse, and that we should thus not expect it to work for large networks [34]. Thus, it may seem surprising that we can write down a relatively simple model, with parameters that number less than a percent of the number of effectively occupied states ( parameters for effective states at N = 120) and whose values are directly determined by measurable observables, and have this model predict so much of the structure in the distribution of states. Surprising or not, it certainly is important that, as the community contemplates monitoring the activity of ever larger number of neurons [96], we can identify theoretical approaches that have the potential to tame the complexity of these large systems.
Some cautionary remarks about the interpretation of our models seem in order. Using the maximum entropy method does not mean there is some hidden force maximizing the entropy of neural activity, or that we are describing neural activity as being in something like thermal equilibrium; all we are doing is building maximally agnostic models of the probability distribution over states. Even in the context of statistical mechanics, there are infinitely many models for the dynamics of the system that will be consistent with the equilibrium distribution, so we should not take the success of our models to mean that the dynamics of the network corresponds to something like Monte Carlo dynamics on the energy landscape. It is tempting to look at the couplings between different neurons as reflecting genuine, mechanistic interactions, but even in the context of statistical physics we know that this interpretation need not be so precise: we can achieve a very accurate description of the collective behavior in large systems even if we do not capture every microscopic detail, and the interactions that we do describe in the most successful of models often are effective interactions mediated by degrees of freedom that we need not treat explicitly. Finally, the fact that a maximum entropy model which matches a particular set of experimental observations is successful does not mean that this choice of observables (e.g., pairwise correlations) is unique or minimal. For all these reasons, we do not think about our models in terms of their parameters, but rather as a description of the probability distribution itself, which encodes the collective behavior of the system.
The striking feature of the distribution over states is its extreme inhomogeneity. The entropy of the distribution is not that much smaller than it would be if the neurons made independent decisions to spike or be silent, but the shape of the distribution is very different; the network builds considerable structure into the space of states, without sacrificing much capacity. The probability of the same state repeating is many orders of magnitude larger than expected for independent neurons, and this really is quite startling (Figure 14). If we extrapolate to the full population of ∼250 neurons in this correlated, interconnected patch of the retina, the probability that two randomly chosen states of the system are the same is roughly one percent. Thus, some combination of spiking and silence across this huge population should repeat exactly every few seconds. This is true despite the fact that we are looking at the entire visual representation of a small patch of the world, and the visual stimuli are fully naturalistic. Although complete silence repeats more frequently, a wide range of other states also recur, so that many different combinations of spikes and silence occur often enough that we (or the brain) can simply count them to estimate their probability. This would be absolutely impossible in a population of nearly independent neurons, and it has been suggested that these repeated patterns provide an anchor for learning [12]. It is also possible that the detailed structure of the distribution, including its inhomogeneity, is matched to the statistical structure of visual inputs in a way that goes beyond the idea of redundancy reduction, occupying a regime in which strongly correlated activity is an optimal code [17], [18], [49], [97].
Building a precise model of activity patterns required us to match the statistics of global activity (the probability that K out of N neurons spike in the same small window of time). Several recent works suggested alternative means of capturing the higher-order correlations [12], [98]–[103]. Particularly promising and computationally tractable amongst these models is the dichotomized Gaussian (DG) model [100] that could explain correctly the distribution of synchrony in the monkey cortex [104]. While DG does well when compared with pairwise models on our data, it is significantly less successful than the full K-pairwise models that we have explored here. In particular, the DG predictions of three-neuron correlations are much less accurate than in our model, and the probability of coincidences is underestimated by an amount that grows with increasing N (Figure S6). Elsewhere we have explored a very simple model in which we ignore the identity of the neurons and match only the global behavior [39]. This model already has a lot of structure, including the extreme inhomogeneity that we have emphasized here. In the simpler model we can exploit the equivalence between maximum entropy models and statistical mechanics to argue that this inhomogeneity is equivalent to the statement that the population of neurons is poised near a critical surface in its parameter space, and we have seen hints of this from analyses of smaller populations as well [6], [9]. The idea that biological networks might organize themselves to critical points has a long history, and several different notions of criticality have been suggested [31]. A sharp question, then, is whether the full probability distributions that we have described here correspond to a critical system in the sense of statistical physics, and whether we can find more direct evidence for criticality in the data, perhaps without the models as intermediaries.
Finally, we note that our approach to building models for the activity of the retinal ganglion cell population is entirely unsupervised: we are making use only of structure in the spike trains themselves, with no reference to the visual stimulus. In this sense, the structures that we discover here are structures that could be discovered by the brain, which has no access to the visual stimulus beyond that provided by these neurons. While there are more structures that we could use—notably, the correlations across time—we find it remarkable that so much is learnable from just an afternoon's worth of data. As it becomes more routine to record the activity of such (nearly) complete sensory representations, it will be interesting to take the organism's point of view [43] more fully, and try to extract meaning from the spike trains in an unsupervised fashion.
This study was performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee (IACUC) of Princeton University (Protocol 1827 for guinea pigs and 1828 for salamanders).
We analyzed the recordings from the tiger salamander (Ambystoma tigrinum) retinal ganglion cells responding to naturalistic movie clips, as in the experiments of Refs. [4], [36], [65]. In brief, animals were euthanized according to institutional animal care standards. The retina was isolated from the eye under dim illumination and transferred as quickly as possible into oxygenated Ringer's medium, in order to optimize the long-term stability of recordings. Tissue was flattened and attached to a dialysis membrane using polylysine. The retina was then lowered with the ganglion cell side against a multi-electrode array. Arrays were first fabricated in university cleanroom facilities [105]. Subsequently, production was contracted out to a commercial MEMS foundry for higher volume production (Innovative Micro Technologies, Santa Barbara, CA). Raw voltage traces were digitized and stored for off-line analysis using a 252-channel preamplifier (MultiChannel Systems, Germany). The recordings were sorted using custom spike sorting software developed specifically for the new dense array [36]. 234 neurons passed the standard tests for the waveform stability and the lack of refractory period violations. Of those, 160 cells whose firing rates were most stable across stimulus repeats were selected for further analysis. Within this group, the mean fraction of interspike intervals (ISI) shorter than 2 ms (i.e., possible refractory violations) was .
The stimulus consisted of a short (t = 19 s) grayscale movie clip of swimming fish and water plants in a fish tank, which was repeated 297 times. The stimulus was presented using standard optics, at a rate of 30 frames per second, and gamma corrected for the display.
We randomly selected 30 subgroups of cells for analysis from the total of 160 sorted cells. In sum, we analyzed groups of neurons, which we denote by , where N denotes the subgroup size, and indexes the chosen subgroup of that size. Time was discretized into Δτ = 20 ms time bins, as in our previous work [4], [6], [9]. The state of the retina was represented by if the neuron i spiked at least once (was silent) in a given time bin t. This binary description is incomplete only in ∼0.5% of the time bins that contain more than one spike; we treat these bins as . Across the entire experiment, the mean probability for a single neuron to make a spike in a timebin (that is, ) is ∼3.1%. Time discretization resulted in 953 time bins per stimulus repeat; 297 presented repeats yielded a total of N-bit binary samples during the course of the experiment for each subgroup.
We used a modified version of our previously published learning procedure to compute the maximum entropy models given measured constraints [37]; the proof of convergence for the core of this L1-regularized maximum entropy algorithm is given in Ref. [106]. Our new algorithm can use as constraints arbitrary functions, not only single and pairwise marginals as before. Parameters of the Hamiltonian are learned sequentially in an order which greedily optimizes a bound on the log likelihood, and we use a variant of histogram Monte Carlo to estimate the values of constrained statistics during learning steps [107]. Monte Carlo induces sampling errors on our estimates of these statistics, which provide an implicit regularization for the parameters of the Hamiltonian [106]. We verified the correctness of the algorithm explicitly for groups of 10 and 20 neurons where exact numerical solutions are feasible. We also verified that our MC sampling had a long enough “burn-in” time to equilibrate, even for groups of maximal size (N = 120), by starting the sampling repeatedly from same vs different random initial conditions (100 runs each) and comparing the constrained statistics, as well as the average and variance of the energy and magnetization, across these runs; all statistics were not significantly dependent on the initial state (two–sample Kolmogorov-Smirnov test at significance level 0.05).
Supplementary Figure S1 provides a summary of the models we have learned for populations of different sizes. In small networks there is a systematic bias to the distribution of parameters, but as we look to larger networks this vanishes and the distribution of becomes symmetric. Importantly, the distribution remains quite broad, with the standard deviation of across all pairs declining only slightly. In particular, the typical coupling does not decline as , as would be expected in conventional spin glass models [70]. This implies, as emphasized previously [9], that the “thermodynamic limit” (very large N) for these systems will be different from what we might expect based on traditional physics examples.
We withheld a random selection of 20 stimulus repeats (test set) for model validation, while training the model on the remaining 277 repeats. On training data, we computed the constrained statistics (mean firing rates, covariances, and the K-spike distribution), and used bootstrapping to estimate the error bars on each of these quantities; the constraints were the only input to the learning algorithm. Figure 1 shows an example reconstruction for a pairwise model for N = 100 neurons; the precision of the learning algorithm is shown in Figure 2.
The dataset consists of a total of binary pattern samples, but he number of statistically independent samples must be smaller: while the repeats are plausibly statistically independent, the samples within each repeat are not. The variance for a binary variable given its mean, , is ; with R independent repeats, the error on the estimate in the average should decrease as . By repeatedly estimating the statistical errors with different subsets of repeats and comparing the expected scaling of the error in the original data set with the data set where we shuffle time bins randomly, thereby destroying the repeat structure, we can estimate the effective number of independent samples; we find this to be , about 37% of the total number of samples, T.
We note that our largest models have constrained statistics that are estimated from at least 15× as many statistically independent samples. Moreover, the vast majority of these statistics are pairwise correlation coefficients that can be estimated extremely tightly from the data, often with relative errors below 1%, so we do not expect overfitting on general grounds. Nevertheless, we explicitly checked that there is no overfitting by comparing the log likelihood of the data under the learned maximum entropy model, for each of the 360 subgroups , on the training and testing set, as shown in Figure 3.
The parametrization of the K-pairwise Hamiltonian of Eq (20) is degenerate, that is, there are multiple sets of coupling constants that specify mathematically identical models. This is because adjusting all local fields by a constant offset adds a term linear in K to ; similarly, adjusting all pairwise couplings by a constant offset adds a quadratic term to . For comparing model predictions (i.e., observables, entropy, the structure of the energy landscape etc) this is inconsequential, but when model parameters are compared directly in Figure 5, one must choose a gauge that will make the comparison of the pairwise and K-pairwise parameters unbiased. Since there is no in the pairwise model, we extract from the of the K-pairwise model all those components that can be equivalently parametrized by offsets to local fields and pairwise couplings. In detail, we subtract best linear and quadratic fits from the reconstructed , such that the remaining only constrains multi-point correlations that cannot be accounted for by a choice of fields and pairwise interactions; the linear and quadratic fit then give us adjustments to local fields and pairwise interactions.
To find the metastable (MS) states, we start with a pattern that appears in the data, and attempt to flip spins from their current state into , in order of increasing i. A flip is retained if the energy of the new configuration is smaller than before the flip. When none of the spins can be flipped, the resulting pattern is recorded as the MS state. The set of MS states found can depend on the manner in which descent is performed, in particular when some of the states visited during descent are on the “ridges” between multiple basins of attraction. Note that whether a pattern is a MS state or not is independent of the descent method; what depends on the method is which MS states are found by starting from the data patterns. To explore the structure of the energy landscape in Figure 11, we started 1000 Metropolis MC simulations repeatedly in each of the 10 most common MS states of the model; after each attempted spin-flip, we checked whether the resulting state is still in the basin of attraction of the starting MS state (by invoking the descent method above), or whether it has crossed the energy barrier into another basin. We histogrammed the transition probabilities into other MS basins of attraction and, for particular transitions, we tracked the transition paths to extract the number of spin-flip attempts and the energy barriers. The “basin size” of a given MS state is the number of patterns in the recorded data from which the given MS state is reached by descending on the energy landscape. The results presented in Figure 11 are typical of the transitions we observe across multiple subnetworks of 120 neurons.
Entropy estimation is a challenging problem. As explained in the text, the usual approach of counting samples and identifying frequencies with probabilities will fail catastrophically in all the cases of interest here, even if we are free to draw samples from our model rather than from real data. Within the framework of maximum entropy models, however, the equivalence to statistical mechanics gives us several tools. Here we summarize the evidence that these multiple tools lead to consistent answers, so that we can be confident in our estimates.
Our first try at entropy estimation is based on the heat capacity integration in Eq. (29). To begin, with neurons, we can enumerate all states of the network and hence we can find the maximum entropy distributions exactly (with no Monte Carlo sampling). From these distributions we can also compute the entropy exactly, and it agrees with the results of the heat capacity integration. Indeed, there is good agreement for the entire distribution, with Jensen-Shannon divergence between exact maximum entropy solutions and solutions using our reconstruction procedure at ∼10−6. As a second check, now usable for all N, we note that the entropy is zero at T = 0, but bits at . Thus we can do the heat capacity integration from T = 1 to instead of T = 0 to T = 1, and we get essentially the same result for the entropy (mean relative difference of across 30 networks at N = 100 and N = 120).
Leaning further on the mapping to statistical physics, we realize that the heat capacity is a summary statistic for the density of states. There are Monte Carlo sampling methods, due to Wang and Landau [73] (WL), that aim specifically at estimating this density, and those allow us to compute the entropy from a single simulation run. Based on the benchmarks of the WL method that we performed (convergence of the result with histogram refinement) we believe that the entropy estimate from the WL MC has a fractional bias that is at or below . The results, in Figure S2A, are in excellent agreement with the heat capacity integration.
K-pairwise models have the attractive feature that, by construction, they match exactly the probability of the all-silent pattern, , seen in the data. As explained in the main text, this means that we can “measure” the partition function, Z, of our model directly from the probability of silence. Then we can compute the average energy from a single MC sampling run, and find the entropy for each network. As shown in Figures S2B and C, the results agree both with the heat capacity integration and with the Wang–Landau method, to an accuracy of better than 1%.
The error on entropy estimation from the probability of silence has two contributions: the first has to do with the error in that contributes to error in Z by Eq (30), and the second with the estimate of the mean energy, , of the model. By construction of the model, needs to be matched to data, but in fact that match is limited by the error bar on itself estimated from data, and on how well the model reproduces this observable; these two errors combine to give a fractional error of a few tenths of a percent. From this error one may then compute the fractional error in Z; for N = 120 groups of neurons, this is on average . For the entropy estimation, we also need the average energy; this itself can be estimated through a long Metropolis MC sampling. The sampling is unbiased, but with an error of typically between half and a percent, for N = 120 sets. Together, these errors combine into a conservative error estimate of ∼1% for the entropy computed from the silence and from the average energy, although the true error might in fact be smaller.
Finally, there are methods that allow us to estimate entropy by counting samples even in cases where the number of samples is much smaller than the number of states [71] (NSB). The NSB method is not guaranteed to work in all cases, but the comparison with the entropy estimates from heat capacity integration (Figure S3A) suggests that so long as N<50, NSB estimates are reliable (see also [108]). Supplementary Figure S3B shows that the NSB estimate of the entropy does not depend on the sample size for N<50; if we draw from our models a number of samples equal to the number found in the data, and then ten times more, we see that the estimated entropy changes by just a few percent, within the error bars. This is another signature of the accuracy of the NSB estimator for N<50. As N increases, these direct estimates of entropy become significantly dependent on the sample size, and start to disagree with the heat capacity integration. The magnitude of these systematic errors depends on the structure of the underlying distribution, and it is thus interesting that NSB estimates of the entropy from our model and from the real data agree with one another up to N = 120, as shown in Figure S3C.
The pairwise correlations between neurons in this system are quite weak. Thus, if we make a model for the activity of just two neurons, treating them as independent is a very good approximation. It might seem that this statement is invariant to the number of neurons that we consider—either correlations are weak, or they are strong. But this misses the fact that weak but widespread correlations can have a non–perturbative effect on the structure of the probability distribution. Nonetheless, it has been suggested that maximum entropy methods are successful only because correlations are weak, and hence that we can't really capture non–trivial collective behaviors with this approach [34].
While independent models fail to explain the behavior of even small groups of neurons [4], it is possible that groups of neurons might be in a weak perturbative regime, where the contribution of pairwise interactions could be treated as a small perturbation to the independent Hamiltonian, if the expansion was carried out in the correct representation [34]. Of course, with finite N, all quantities must be analytic functions of the coupling constants, and so we expect that, if carried to sufficiently high order, any perturbative scheme will converge—although this convergence may become much slower at larger N, signaling genuinely collective behavior in large networks.
To make the question of whether correlations are weak or strong precise, we ask whether we can approximate the maximum entropy distribution with the leading orders of perturbation theory. There are a number of reasons to think that this won't work [109]–[112], but in light of the suggestion from Ref [34] we wanted to explore this explicitly. If correlations are weak, there is a simple relationship between the correlations and the corresponding interactions [34], [113]. We see in Figure S4A that this relationship is violated, and the consequence is that models built by assuming this perturbative relationship are easily distinguishable from the data even at N = 15 (Figure S4B). We conclude that treating correlations as a small perturbation is inconsistent with the data. Indeed, if we try to compute the entropy itself, it can be shown that even going out to fourth order in perturbation theory is not enough once N>10 [111], [112].
|
10.1371/journal.ppat.1000062 | A Cysteine Protease Is Critical for Babesia spp. Transmission in Haemaphysalis Ticks | Vector ticks possess a unique system that enables them to digest large amounts of host blood and to transmit various animal and human pathogens, suggesting the existence of evolutionally acquired proteolytic mechanisms. We report here the molecular and reverse genetic characterization of a multifunctional cysteine protease, longipain, from the babesial parasite vector tick Haemaphysalis longicornis. Longipain shares structural similarity with papain-family cysteine proteases obtained from invertebrates and vertebrates. Endogenous longipain was mainly expressed in the midgut epithelium and was specifically localized at lysosomal vacuoles and possibly released into the lumen. Its expression was up-regulated by host blood feeding. Enzymatic functional assays using in vitro and in vivo substrates revealed that longipain hydrolysis occurs over a broad range of pH and temperature. Haemoparasiticidal assays showed that longipain dose-dependently killed tick-borne Babesia parasites, and its babesiacidal effect occurred via specific adherence to the parasite membranes. Disruption of endogenous longipain by RNA interference revealed that longipain is involved in the digestion of the host blood meal. In addition, the knockdown ticks contained an increased number of parasites, suggesting that longipain exerts a killing effect against the midgut-stage Babesia parasites in ticks. Our results suggest that longipain is essential for tick survival, and may have a role in controlling the transmission of tick-transmittable Babesia parasites.
| Ticks are important ectoparasites among the blood-feeding arthropods and serve as vectors of many deadly diseases of humans and animals. Of tick-transmitted pathogens, Babesia, an intracellular haemoprotozoan parasite causing a malaria-like disease, called babesiosis, gain increasing interest due to its zoonotic significance. When vector ticks acquire the protozoa via blood-meals, they invade midgut and undergo several developmental stages prior to exit through salivary glands. It has long been conceived that midguts of these ticks evolve diverse innate immune mechanisms and perform blood digestion critical for tick survival. A cysteine proteinase, longipain, was identified from the three-host tick Haemaphysalis longicornis, which shows potent parasiticidal activity. Longipain is localized in midgut epithelium and its expression is induced by blood feeding. This protein is passively secreted into midgut lumen where it exerts enzymatic degradation of blood-meals. A series of experiments unveil that longipain-knockdown ticks when fed on Babesia-infected dog, exhibited a significantly increased numbers of parasites compared with controls. Longipain has shown to interact on the surface of Babesia parasites in vitro and in vivo, and is thought to mediate direct killing of the parasites, suggesting that longipain may be a potential chemotherapeutic target against babesiosis and ticks themselves.
| The ixodid ticks are obligate hematophagous organisms that belong to the phylum Arthropda, and are classified with spiders and scorpions in the class Arachnida [1]. Ticks are long-term blood-pool feeders, while mosquitoes are short-term vessel-feeders, and the process of blood digestion in ticks differs mechanistically from that in haematophagous insects [2],[3]. After blood feeding, ticks can increase more than 50 times in body weight compared with their original weight due to the acquired host blood meal, which mainly consists of red blood cells. Blood digestion in ticks is a slow intracellular process that takes place via phagocytosis by desquamated epithelial cells in the midgut [4],[5]. The tick midgut is considered to contain evolutionally acquired molecules involved in host blood digestion [6],[7]. The existence of secreted proteolytic enzymes in blood-sucking ticks suggests that they are required for various functions necessary for survival via successful blood-feeding behavior, which includes continuous feeding for days, or even weeks [8],[9], However, the precise mechanism responsible for the host blood digestion is unknown.
Ticks as vectors are the most important ecto-parasites of domestic animals and are the second-most important vector next to mosquitoes among arthropods that transmit infectious diseases in human [10]. Pathogens, including viruses, bacteria and parasites, are taken up in the blood meal and exposed to a potentially hostile environment in the tick's midgut before invading the epithelium, where they subsequently multiply. Recent studies have shown that the midgut is involved in diverse arthropod innate immune responses against pathogens [11]–[14]. Thus, tick proteolytic enzymes in the midgut may play critical roles in host blood meal digestion and pathogen transmission. Such products may become increasingly important as drug targets and vaccine candidates for both tick control and tick-borne diseases [15]–[17].
The ixodid tick Haemaphysalis longicornis is an important disease vector for human and animal pathogens, including the causative agents of babesiosis, Q fever and Russian encephalitis. H. longicornis is the primary vector of the pathogens causing babesiosis of humans and domestic animals in Japan [18]. Babesiosis is a human malaria-like disease that has recently been considered as an emerging zoonosis [19]–[21]. Recent reports have shown that endo- and ecto-parasite cysteine proteases play numerous indispensable roles in the survival of the parasites [22]. We hypothesized that longipain, a tick cysteine protease, would play a specific physiological role in blood feeding and Babesia parasite transmission. We report here the characterization of longipain isolated from H. longicornis.
We isolated a longipain cDNA that was 1,352 bp long and included a start codon at nucleotides 159–161 with a consensus lower eukaryote initiation sequence (AXXATGG) and a stop codon at nucleotides 1182–1184. The ORF extending from position 159 to position 1184 codes for 341 amino acid residues. The 3′ untranslated region contained 149 bp and ended with an 18-bp poly (A)+ tail that began 10 bp downstream from AATAAA, which is the eukaryotic consensus polyadenylation signal. The predicted sequence showed the presence of a signal peptide of 17 amino acid residues, suggesting that longipain is secreted. The mature protein had a predicted molecular mass of 36,3335.2 Da and pI 4.33. Longipain belongs to the papain family of cysteine proteases [23] and possesses each of the conserved motifs identified in the active site of lysosomal cathepsins (Fig. 1A). Longipain also possesses an occluding loop motif that is thought to be responsible for the exopeptidase activity of cathepsin B-like proteases [23]. The nine amino acid peptide loop (residues 48–56) is suggestive of mannose-6-phosphate-independent trafficking [24]. The presence of three predicted N-glycosylation sites suggests that longipain may be a glycosylated protein. An NCBI search revealed that longipain possessed a conserved cathepsin L domain in addition to a cathepsin B domain. However, the conserved inter-space motif of cathepsin L that is structurally different from that of cathepsin B was not found in the pro-region of longipain [25]. The three-dimensional structures of vertebrate cathepsin B and L have revealed the important amino acid residues that contribute to the substrate specificity [26]. In these cathepsins, the substrate preference is primarily determined by the S2 subsite of the active site pockets [27]. The S2 pocket of longipain contains Asp332, like the human malaria parasite cathepsin L. Longipain was found to be most similar in sequence architecture to the known spider cathepsin B from Araneus ventricosus, with 64 % similarity (Fig. 1B). Longipain also shares similarity to some other GenBank™ sequences, including the cathepsin B-like protease of the amphioxus (Branchiostoma belcheri tsingtaunesehuman), the kissing bug (Triatoma sordida) and humans, to which the similarities were 57%, 56%, 55%, respectively.
To identify the endogenous longipain in H. longicornis, we examined the localization of endogenous longipain by immunohistochemistry (Fig. 2A). The endogenous antigen bound to anti-longipain antibody was detected in club-shaped midgut epithelial cells protruding into the gut lumen of a partially fed adult [28],[29]. A positive reaction was not detected within the salivary gland or other tissues, suggesting that the endogenous longipain specifically expressed in the midgut. No positive reaction was detectable in either the internal or external tissues with preimmune mouse serum. Next, we performed two-dimensional immunoblot analysis (Fig. 2B). More than 200 visible protein spots appeared on silver-stained two-dimensional gels. Mouse longipain antibody strongly reacted with a protein of molecular mass 39 kDa and a pI of 4.5. To identify the internal amino acid residues, a protein spot was excised from extracts of adult female ticks using two-dimensional gel electrophoresis and processed as described previously [30]. After in-gel digestion with lysyl endopeptidase, peptides were collected by reverse phase high pressure liquid chromatography (RP-HPLC), and several peptides were analyzed to determine the internal amino acid sequence. The resultant sequences of the 39 kDa RP-HPLC-purified peaks were identical to those of the deduced amino acid sequence encoded by longipain cDNA. The sequencing confirmed that endogenous longipain has a molecular weight of 39 kDa with a pI of 4.5. To examine the expression pattern of longipain during blood-feeding, we evaluated the level of endogenous longipain in unfed and fed nymphs. Figure 2C shows that the expression of endogenous longipain was significantly increased after blood-sucking.
To determine the precise subcellular localization of longipain in the midgut, we performed immunoblotting, immunofluorescence and immunoelectron microscopy with anti-longipain. We were able to detect a positive band at 39 kDa in the lumen contents (Fig. 3A). Immunofluorescence analysis showed that endogenous longipain was expressed in the shedding cells (Fig. 3B). Prior studies suggested that shedding cells of the midgut are produced by passive release into the lumen [2],[31]. Furthermore, we detected endogenous longipain both on the surface and in the lysosomes of the midgut epithelial cells (Fig. 3C).
Mammalian cathepsin B and L are widely expressed cysteine proteases involved in both intracellular proteolysis and extracellular matrix remodeling [32],[33]. Recent reports demonstrated that parasite proteases function in a broader chemical environment than the homologous host environment [34]. We examined the hydrolysis activities using synthetic substrates to analyze the enzymatic function of yeast-expressed longipain (longipain, Fig. 4A). The activity assay using the purified longipain showed hydrolysis of Z-Arg-Arg-MCA and Z-Phe-Arg-MCA substrates (Table 1). The specific activity of longipain for Z-Arg-Arg-MCA was optimal at pH 5 and at temperature 15°C, and that for Z-Phe-Arg-MCA was optimal at pH 8 and at 35°C. In the assay of inhibition of cysteine protease activity, 10 µM E64 caused strong inhibition of longipain activity for the hydrolysis of Z-Phe-Arg-MCA (Table 2), but other inhibitors had little or no effect on the activity. E64 and PMSF had inhibitory effects on the ability of longipain to hydrolyze Z-Arg-Arg-MCA. Figure 4B shows the catalytic efficiencies of longipain from pH 3 to 9. The peak of activity with cathepsin B as substrate occurred at pH 3.6. Hydrolysis of cathepsin L as substrate showed dramatically different features, exhibiting a broad peak of catalytic efficiency from pH 7 to 8. Next, we examined the temperature dependency of longipain activity and found that the catalytic efficiency also showed different optimal temperatures between the two substrates. Longipain had good activity against cathepsin B at 15°C, while cathepsin L was strongly preferred at 37°C. Our results demonstrate that longipain functions not only over a broad range of pH conditions, but also over a wide range of temperature conditions.
The contents of the blood meal in the midgut are composed of a variety of native proteins derived from the hosts. To identify the endogenous substrates in the midgut, several components of host blood were tested for hydrolysis by rlongipain. Although a proteolytic effect was not shown against most of the components, longipain was shown to successfully cleave spectrin, a major component of erythrocyte membranes. Erythrocyte spectrin is a heterodimer composed of a 280-kDa α subunit and a 246-kDa β subunit which associate in a side-to-side, antiparallel configuration to form a 100-nm rod-like structure. Spectrin in other tissues may be composed of distinct but homologous α and β subunits, and is sometimes referred to as fodrin [35]. Figure 5 shows the hydrolysis of host erythrocyte membrane components by longipain. Spectrin α and β subunits reacted strongly with anti-human spectrin antibody (arrowheads). Fragments with smaller molecular mass, corresponding to hydrolyzed spectrin, were detected below the intact form on the immunoblot. The results obtained with various amounts of longipain indicated that the hydrolysis of spectrin seemed to be a dose-dependent reaction. The longipain showed strong hydrolysis of spectrin at pH 4.0–8.5, 20–50°C. Ticks can feed only on blood to obtain nutrients. Thus, spectrin may not be the only endogenous substrate for longipain in the tick midgut. Hemoglobin was used as a substrate and was not hydrolyzed by longipain.
Prior reports suggested that vector proteases facilitate the invasion of pathogens [36],[37]. It was therefore of interest to test the hypothesis that longipain would enhance the spread of H. longicornis-bearing Babesia parasites after their release from the infected erythrocytes. We incubated the equine Babesia parasite B. equi in medium supplemented with longipain. Longipain inhibited merozoite proliferation at a concentration of 0.125 µmol (Fig. 6A). The inhibition was dose dependent. Giemsa-stained smears of the culture medium showed abnormal multidividing forms and pyknosis of B. equi parasites (Fig. 6B). Free merozoites were also found in the culture medium. No morphological changes of erythrocytes were seen in the presence of longipain at any concentration. This finding suggests that longipain may react with intraerythrocytic substrates rather than membrane-bound components. Next, B. equi were incubated with biotin-labeled longipain in culture medium to explore how longipain interacted with the parasites. Fluorescence microscopy showed positive reactions at the surface of the free-merozoites, but not erythrocytes (Fig. 6C). We therefore assume that the killing of Babesia parasites in the midgut of H. longicornis tick is meditated by several midgut-derived proteolytic enzymes.
We hypothesized that endogenous longipain may promote host blood digestion and at the same time decrease Babesia parasite survival in H. longicornis. To test these possibilities, we used RNA interference (RNAi) to knockdown longipain mRNA by dsRNA in adult H. longicornis. RNAi, a phenomenon in which double-strand RNA (dsRNA) silences gene expression through specific degradation of the cognate mRNA, is a direct and efficient way of producing and identifying the loss-of-function of targeted genes as a reverse genetic tool. In this study, the dsRNA-treated ticks were attached to a dog preinfected with the canine Babesia parasite B. gibsoni. We then assessed whether the transmission of Babesia parasites was affected by the endogenous longipain. Ticks injected with phosphate buffered saline (PBS) alone or with dsRNA generated from E. coli MalE maltose-binding protein (MBP) gene were used as control. Although longipain dsRNA-treated (mean±standard deviation; 84.0±10.9 mg, n = 6), PBS-treated and MBP dsRNA-treated ticks (PBS; 85.2±14.7 mg, n = 6, MBP; 86.2±12.4, n = 6) showed similar behavior of blood-feeding by day 3, longipain depression clearly impaired tick blood feeding after day 4. Ticks feed rapidly on a host before engorgement [2] although the underlying mechanism responsible for the developmental effect is unclear. The engorged adult ticks in the longipain dsRNA-treated group had a smaller and rounder appearance than those in the PBS-treated and MBP dsRNA-treated groups, and none had cuticular wrinkles on the dorsum, as were found in ticks in control groups (Fig. 7A). Significant differences in body-weight at engorgement were observed between the knockdown (mean ±standard deviation; 108.5±34.1 mg, n = 10) and control groups (PBS; 322.1±49.8 mg, n = 16: MBP; 322.9±44.0, n = 10). Reverse transcription polymerase chain reaction (RT-PCR) analysis revealed that injection of longipain dsRNA caused complete loss of longipain mRNA (Fig. 7B). This indicates that the significant reduction in longipain mRNA expression was due to a gene-specific dsRNAi effect. A decrease of endogenous longipain was also seen by immunofluorescence and immunoblot analyses (Fig. 7C, D). The midguts of longipain dsRNA-treated, PBS-treated and MBP dsRNA-treated ticks expressed almost equal amounts of H. longicornis serine protease (HlSP, data not shown). HlSP was previously identified from the midgut of H. longicornis and was shown to be associated with host blood feeding and digestion by our reverse genetic studies [38]. The present results strongly indicate that longipain dsRNA was efficiently delivered to the midgut of ticks via injection at the fourth coxae. H. longicornis is a three-host tick whose larvae, nymph, and adults all engorge on animals [18]. Thus, H. longicornis can transmit Babesia parasites interstadially through both larva to nymph and nymph to adult transition. Babesia parasites in the midgut lumen of the adult H. longicornis after blood feeding invade the epithelium, move to the ovary and finally arrive at the eggs [19]. We assessed whether Babesia parasites were affected by the suppression of endogenous longipain. Immunoflurorescence analysis revealed an increase in the number of B. gibsoni in the midgut lumen and in the epithelium of longipain dsRNA-treated ticks as compared to those of PBS-injected and MBP dsRNA-treated ticks on day 6 after injection (Fig. 8 A, upper panel). We then examined the localization of endogenous longipain and B. gibosoni using anti-longipain and anti-B. gibosoni antibodies by double immunostaining, in order to obtain more visible interaction between longipain and Babesia parasies. Double-immunostaining data clearly revealed the complete absence of longipain-mediated endogenous interaction in the midgut microenvironment, showing only an increased number of B. gibsoni in the longipain-knokdown ticks compared to those served as PBS-injected and MBP-treated controls (Fig. 8A, lower panel). This evidence suggests that longipain exerts its effect directly on the B. gibsoni parasites and mediates killing of the parasites. Actually, quantitative PCR analysis clearly demonstrated increased number of B. gibsoni in the midgut, consistent with the results of immunofluorecence staining (Fig. 8 B). In the ovary, B. gibsoni was detected in longipain-knockdown ticks, but not in control ticks (Fig. 8 C). The longipain dsRNA-treated ticks showed a significant increase in the number of B. gibsoni in the ovary and the hatched larvae as compared to those of PBS-injected and MBP dsRNA-trated ticks by quantitative PCR analysis (Fig. 8 D). Longipain-knockdown ticks showed approximately a 3-fold increase in the ability to transmit Babesia parasites. Together, these results confirm that longipain-mediated killing may regulate the number of Babesia parasites in the tick midgut.
Tick-borne disease is a major public health issue in many parts of the world, where the increasing prevalence of drug resistance underscores the need to identify new drug targets [39],[40]. Examination of metabolic pathways such as those involved in digesting the host blood meal has provided numerous attractive candidates for chemotherapeutic development, since blood-feeding is essential for tick survival [41],[42]. The molecular basis of how the digestion is maintained in nature during the complex life cycle of ticks is poorly understood. We speculate that ticks possess specific gene-products that have been acquired during the process of their evolution for host blood feeding. Our results obtained in the present study demonstrate that longipain is a multifunctional cysteine protease that functions in host blood meal digestion and in the regulation of the vectorial capacity for tick-borne Babesia parasites.
We confirmed here that longipain is passively secreted into the lumen of the midgut. Thus, it is likely that the process of digestion of the host blood meal occurs in the midgut lumen [43]. As the results of histological studies showed that longipain may function in two different physiological locations (the lysosomes of the epithelium and the midgut lumen), we concentrated our subsequent enzymatic studies on longipain. In vitro enzymatic functional assays revealed distinct pH and temperature preferences of longipain for the activity against cathepsin B and L substrates, indicating that longipain may be involved in anatomically specific activity in the midgut. In the present study, longipain was found to be released into the lumen and localized in lysosomes, suggesting that the activity in these locations is dependent on the pH conditions in the lumen and in the epithelium. Interestingly, the optimal temperatures for the synthetic cathepsin B and L substrates were also distinct. These results prompted us to examine whether longipain possesses a strong ability to cleave endogenous substrates derived from the host blood meal. Furthermore, we hypothesized that major protein components of the blood meal are targeted for hydrolysis over a wide range of pHs and temperatures. Interestingly, we were able to detect cleaved forms of spectrin upon hydrolysis by longipain. The developmental cycle of ticks involves two distinct drastic patterns [1]: a blood-feeding phase that occurs upon attachment to the host and a non-invasive phase that occurs during the period from engorgement until a subsequent attachment to the next host. The ambient temperature in vector ticks varies from the host body temperature to lower temperatures. The broad enzymatic properties of longipain might be related to the behavioral features of blood-feeding ecto-parasite ticks, i.e., whether feed on the host or remain off the host. Longipain functions as a protease virtually both at acidic and at neutral pH, in addition to functioning over a wide range of temperatures. Longipain may substitute for the functions of cathepsin B and L in ticks [44]. Taken together, our results using in vitro and in vivo substrates strongly suggest that longipain may have an expanded role beyond host blood digestion, as demonstrated by its pH preference and location in the midgut. Longipain may function in a broader range of environments in order to promote the survival of ticks.
The ixodid ticks have four developmental stages (egg, larva, nymph and adult) in their life cycle. Prior studies showed that the processes of digestion of the host blood meal in the midgut are distinct at different developmental stages of H. longicornis [2],[29]; pinocytosis occurs only at the nymph stages, while both phagocytosis and pinocytosis occur at the adult stage. It is likely that endogenous longipain may function stage-dependently against host blood components in the tick midgut. Ticks must acquire nutrients from the host blood meal and metabolize these nutrients via catabolism and anabolism. The unique life cycle and resulting microenvironment of ticks has led to the evolution of metabolic pathways which differ from those in mammalian hosts. An enzyme cascade of proteolysis of host blood components has been elucidated in blood-feeding helminths, hookworms and schistosomes, and human malaria parasites [45]–[48]. Our group has identified various proteases involved in host blood digestion from the midgut of H. longicornis [38],[41]. Intriguingly, to digest the meal, the unique haematophagus physiology of ticks relies predominantly on proteases that are distinct from those used by insects as well as from those in endo-parasites [2],[49]. Elucidation of the molecular mechanisms by which midgut proteases digest the host blood meal may make it possible to exploit the unique pathways and enzymes in the design of control strategies.
The interaction between ticks and pathogens in the midgut constitutes a critical aspect of disease transmission and a potential target for efforts to control tick-borne diseases [50],[51]. At the vector stage in H. longicornis, Babesia parasites must complete a complex developmental cycle in the tick in order for transmission to occur. After the release of Babesia via erythrocyte rupture, the parasites must invade into the epithelial cells of the midgut. Their development depends on the balance between the ability of the tick to establish a defense response against the parasite and the ability of the parasite to escape the tick's immune response [52]. Proteolytic enzymes from the midgut of ticks may inhibit the proliferation of the midgut-stage parasites. We hypothesized that H. longicornis possesses a specific gene product that exerts a partial protective response against Babesia in the midgut. Our results in vitro demonstrated that longipain kills the merozoite stage parasites released from erythrocytes in the case of equine Babesia parasites. Intriguingly, a cysteine protease purified from the venomous protein in the midgut of the social aphid exerts a killing effect against enemies, and its insecticidal activity may play a role in colony defense [53]. This biological function is suggestive of the evolutionary route of the aphid-specific molecules. H. longicornis might have evolutionally acquired babesiacidal activity in relation to becoming a vector of Babesia parasites. The killing effect points to a possible link with the structural features of longipain, which remains to be explored in the future.
Our reverse genetic analysis revealed that longipain regulates the blood meal digestion, demonstrating that loss of function causes a macroscopically detectable delay of blood-sucking speed, followed by a gain of body weight, suggesting that the effect of longipain dsRNA is specific to the protein being targeted. Longipain might function as a proteolytic enzyme in the process of blood-meal digestion in the lumen and the epithelium. Prior studies showed that a disease-bearing vector transmits only a limited number of parasites during blood-feeding, suggesting the existence of a partially successful natural defense mechanism against the parasite [54]. In the malaria vector (the mosquito), midgut serine protease is involved in regulating the parasite burden and the ability of Plasmodium parasites to invade into the midgut epithelium [55],[56]. An excessive number of parasites might destroy the midgut epithelium, resulting in the haemolymph flowing into the lumen, causing the death of the tick. A decreased level of proteolytic enzymes in the midgut of disease vectors might have a major impact on vectors and pathogens. Thus, we hypothesize that longipain acts as a defense molecule against invading Babesia parasites in H. longicornis. We found that longipain was highly expressed in the midgut, where it was localized in extracellular and intracellular parts of the epithelium. Knockdown of longipain induced a significant increase in the number of parasites in the lumen, suggesting that longipain directly kills Babesia parasites in the midgut. The recruitment of similar receptor and ligand interactions in both vector ticks and mammals in the fight against infection suggests that they have developed similar mechanisms and molecular pathways to recognize and eliminate the invaders [14],[57],[58]. These longipain knockdown studies revealed a substantial reduction in the proportion of host blood feeding and a corresponding increase in the proportion of the number of parasites in the midgut as well as the organs subsequently infected during the parasite life cycle. The fact that longipain is involved in the killing of the dog Babesia parasites suggests the existence of cross-talk between the tick immune response of H. longicornis and Babesia parasites. It is likely that diverse arthropod innate immune responses against Babesia parasites may be conserved and may contribute to controlling the tick vectorial capacity [59],[60].
In summary, we have identified longipain, a multifunctional cysteine protease from a babesial parasite vector tick. The present findings strongly suggest that longipain plays dual roles in the host blood meal digestion and in the control of transmission of the Babesia parasite. Our data demonstrate the pivotal role of longipain in the maintenance of the babesial vector ticks. Understanding the function of the midgut molecules that participate in the interaction between the Babesia parasite and vector ticks will lead to novel approaches to the control of animal and human babesiosis and provide a model for tick-borne diseases.
Haemaphysalis longicornis (Okayama strain) were maintained at the Laboratory of Parasitic Diseases, National Institute of Animal Health, Tsukuba, Ibaraki, Japan, on rabbits as described previously [31].
The Babesia parasites used in this study were as follows: a horse Babesia parasite, Babesia equi, and a dog Babesia parasite, B. gibsoni. The U.S. Department of Agriculture strain of B. equi was maintained by in vitro culture at the National Research Center for Protozoan Diseases (NRCPD) [61]. The NRCPD strain of B. gibsoni was maintained in chronically infected dogs at NRCPD [62].
All animals used in this study were acclimatized for 2 weeks prior to experiments. Animal experiments performed at the National Institute of Animal Health (NIAH) were conducted in accordance with the protocols approved by the NIAH Animal Care and Use Committee (Approval nos. 441, 508, 578). Animal experiments carried out at Obihiro University of Agriculture and Veterinary Medicine (OUAVM) were conducted in accordance with the Guiding Principles for the Care and Use of Research Animals promulgated by OUAVM (Approval nos. 6–42, C-2).
Prior studies have shown that animals can be rendered immune to tick infection by repeated infection with ticks, suggesting that tick-secreted proteins play a role in immunity against challenge infection [43],[63]. An H. longicornis cDNA expression library was immunoscreened using antibodies to H. longicornis generated in rabbits by repeated infection of adult ticks [31]. We thereby found a cDNA encoding the putative longipain among several clones that were immunoreactive with the H. longicornis immunized rabbit serum. The nucleotide sequences of the cDNAs were determined by the Sanger dideoxy chain termination method, using a PRISMTM Ready Dye Terminator Cycle Sequencing Kit (Perkin-Elmer, http://las.perkinelmer.com). GENETYX-WINTM sequence analysis software and the BLAST network server of the National Center for Biotechnology Information (NCBI) were used to analyze the nucleotides and deduce the amino acid sequences for determining homologies with previously reported sequences in GenBank. The SignalP 3.0 program (Center for Biological Sequence Analysis Biocentrum-DTU, http://www.cbs.dtu.dk/services/SignalP) was used for the prediction of the cleavage site for the signal peptide [64]. Potential N-glycosylation sites were analyzed with the ScanProsite program (Alexandre Gattiker and the Swiss Institute of Bioinformatics) [65]. The amino acid sequence of longipain was aligned with the sequences of known cysteine protease family members using the alignment program ClustalW (http://www.ddbj.nig.ac.jp./E-mail/clustalw-j.html).
A specific antibody against longipain was generated in mice (Japan SLC, http://www.jslc.co.jp) immunized with E. coli-expressed longipain. The entire coding region of longipain except the signal sequence was subcloned into a plasmid expression vector, pTrcHisB (Invitrogen, http://www.invitrogen.com), as described [29]. The plasmid was transformed into E. coli strain TOP10F' (Invitrogen) and the purification process was monitored by SDS-PAGE using a T7 Taq® monoclonal antibody (Novagen, http://www.merckbiosciences.com). The recombinant protein was purified using AKTA equipped with a HiTrap chlelating HP column (GE Healthcare, https://www1.gelifesciences.com). Mice were immunized with 50 µg of the longipain using TiterMax Gold (CytRx, http://www.titermax.com) and boosted two more times as described previously [66]. Serum was prepared from blood collected 2 weeks after the final immunization.
The partial coding region of longipain was amplified by PCR using a sense primer, longipain PICZC 5′/EcoRI (CCG AAT TCT AAT GTC TGA CCG CTA TTT GGT TCC CGT CGA CAT G), which contains an EcoRI site (shown in bold) and an antisense primer, longipain 3′/XhoI (CCG AGC TCG AGA TAT CTA GGT ATT CCA GCT ACA AC), which contains an XhoI site (shown in bold) and a yeast initiation consensus sequence (underlined). Competent yeast cells of strain GS115 (Mut+, His-) were prepared according to the protocol of the EasySelect Pichia Expression kit (Invitrogen) and transformed with pPICZC-longipain. Mut+ transformants were detected after growth on agar minimal methanol+histidine plates (1.34% yeast nitrogen base, 4×10−5% biotin, 0.5% methanol, 1.5% agar) at 30°C for 2 days and subsequently cultured on agar minimal dextrose+histidine plates. For large scale expression of longipain, 50 ml of minimal glycerol+histidine was inoculated with transformed yeast cells and the cells were grown for 2 days on a rotary shaker. Expression was induced by exchanging the medium for 250 ml of minimal methanol+histidine medium and shaking for 24 h at 30°C. Yeast cells were collected by centrifugation at 4°C, 1,600 g and resuspended in an equal volume of breaking buffer (50 mM sodium phosphate, pH 7.4, 1 mM EDTA, 5% glycerol) containing 1 mM PMSF (phenylmethylsulfonyl fluoride), and homogenized with an equal volume of acid-washed glass beads (0.5 mm), and the supernatant was obtained by centrifugation at 4°C, 26,400 g. The longipain was purified with the use of AKTA equipped with a HiTrap chlelating HP column and was dialyzed in an Slide-A-Lyzer Dialysis Cassette (Pierce, http://www.piercenet.com)
One- and two-dimensional immunoblotting of H. longicornis was performed as described previously [29]. Adult female tick protein extract was subjected to one- (SDS-PAGE) or two-dimensional (IEF/SDS-PAGE) electrophoresis, and the proteins were transferred onto a nitrocellulose membrane, and then incubated with mouse anti-longipain antibody (diluted 1∶400). Tick immunohistochemistry was performed with mouse anti-longipain antibody as described previously [66]. The sections on glass slides were incubated with mouse anti-longipain (1∶200). After color development, the sections were observed under a microscope (Axiophot; Carl Zeiss, http://www.zeiss.com).
The tick midgut was processed as described previously [67]. Thin sections (approximately 80 nm thick) were cut on a Leica UCT ultramicrotome and were mounted on glow-discharged nickel grids and stored on 2% gelatin until labeled. Immunolabeling was performed using mouse anti-longipain (diluted 1∶500) with anti-mouse IgG 12-nm colloidal gold-conjugated secondary antibody. Samples were stained with uranyl acetate and lead citrate and then examined with a Hitachi H-7500 electron microscope.
Tick immunofluorescence analysis was performed as described [31]. Bound mouse anti-longipain was detected using anti-mouse IgG Alexa 488 (Invitrogen). The sections were mounted in Vectashield ® (Vector) with 4′,6-diamino-2-phenylindole (DAPI) and photographed with a fluorescence microscope (Leica, http://www.leica-microsystems.com) using appropriate filter sets. Images were collected by using Leica FW4000 software.
The midguts of 72-h-fed female adults were dissected under stereomicroscopic examination and transferred to a tube on ice. Midguts were opened with a needle to release the lumen contents and agitated gently in 50 µl of phosphate buffered saline containing a cocktail of protease inhibitors (Roche, http://www.rocheusa.com). A tube containing the opened midgut and lumen contents was centrifuged at 4°C, 1,600 g. The supernatant was collected and used to perform immunoblot analysis for detecting endogenous longipain.
Assays of the activity of the longipain expressed in yeast were performed with fluorogenic substrates (Peptide Institute Inc., http://www.peptide.co.jp) in a final volume of 100 µl containing 25 mM citric acid sodium phosphate and 5 mM dithiothreitol (DTT) [68],[69]. The hydrolysis of Z-Arg-Arg-MCA as a substrate for cathepsin B, Z-Phe-Arg-MCA as a substrate for cathepsin B/L and Suc-Leu-Leu-Val-Tyr-MCA as a substrate for chymotrypsin was measured (Table 1). Each concentration was tested in triplicate. Km and kcat values were determined by fitting initial rate data obtained using multiple substrate concentrations to the Michaelis-Menten equation.
Fluorogenic substrate assays were done with 50 µg/ml of enzyme. Fluorogenic assays were monitored by fluorescence spectrophotometry at 380 nm excitation and 460 nm emission (TECAN, http://www.tecan.com). The optimum pH for the enzyme activity was determined using citrate-sodium phosphate buffer with a pH range of 2.5–8.5 and the optimum temperature was determined using citrate-sodium phosphate buffer with a pH of 5 or 8. The ability of protease inhibitors to inhibit longipain activity against each substrate was also investigated (Table 2).
Haemoglobin-free rabbit erythrocyte membranes were prepared as described below [70],[71]. Two milliliters of freshly drawn rabbit blood were centrifuged at 800 g, 4°C and the supernatant was removed. The pellet of erythrocytes was washed with PBS three times and erythrocyte ghosts were prepared by hypotonic lysis in 1 mM phosphate buffer (PB), pH 7.0. Erythrocyte ghosts were washed in 1 mM PB several times to remove internal contents and collected by centrifugation at 12,000 g, 4°C, and stored in 1 ml of PBS at 4°C. The membrane stock suspension was diluted 1∶100 in PBS and 15-µl aliquots were incubated with longipain in a final volume of 100 µl containing 25 mM citric acid/sodium phosphate and 5 mM DTT for 6 h at 37°C. The reaction mixtures were subjected to SDS-PAGE and immunoblot analysis using rabbit anti-human spectrin serum (Sigma, http://www.sigmaaldrich.com) at a dilution of 1∶400.
B. equi merozoites were grown in horse erythrocytes in vitro as previously described [62] and incubated in the presence of longipain at different concentrations. Parasitemia was assessed daily by microscopic observation of Giemsa-stained blood smears.
Longipain was labeled with biotin (Biotin labeling kit-NH2, Dojindo, http://www.dojindo.co.jp) according to the manufacturer's protocol. The labeled longipain was added to the culture medium, and the cells were washed several times with PBS. The cells were smeared on slide glasses and fixed in methanol. Biotin-labeled longipain was detected using Alexa 594 streptavidin at a dilution of 1∶500 (Invitrogen). The samples were mounted in Vectashield with DAPI and photographed with a camera-equipped fluorescence microscope as described above.
The RNAi procedure in ticks was carried out using dsRNA as described previously [72],[73]. The coding sequence of mature longipain was cloned into pBluescript II SK+ plasmid and the inserted sequence was amplified by PCR using the oligonucleotides T7 (5′-GTAATACGACTCACTATAGGGC-3′) and CMo422 primers (5′-GCGTAATACGACTCACTATAGGGAACAAAAGCTGGAGCT-3′) to attach T7 promoter recognition sites at both the 5′ and 3′ ends. The MBP gene was cloned into pBluescript II SK+ plasmid to generate control dsRNA [74]. The inserted sequence was amplified by PCR using a forward primer (5′-TTATGAAAATAAAAACAGGTGCA-3′) and a reverse primer (5′-CTTGTCCTGGAACGCTTTGTC-3′). The PCR products were purified using a gel extraction kit (QIAGEN, http://www.qiagen.com). dsRNA complementary to the DNA insert was synthesized by in vitro transcription using T7 RNA polymerase (Promega, http://www.promega.com) according to the manufacturer's protocol. Two micrograms of dsDNA were used as a template and 50–100 µg of dsRNA were synthesized. One microgram of longipain dsRNA in 0.5 µl of PBS was injected from the fourth coxae into the haemocoel of unfed adult H. longicornis females fixed on a glass slide with adhesive tape. The injections were carried out by using 50-µl microcapillaries (MICROCAP®, Drummond Scientific, http://www.drummondsci.com) drawn to fine-point needles by heating. The needles were connected to an air compressor. Control ticks were injected with 0.5 µl of PBS alone or with 1 µg of MBP dsRNA in 0.5 µl of PBS. The ticks were allowed to rest for 1 day at 25°C. No mortality resulted from the injection alone, as both control and longipain dsRNA-treated ticks survived after injection while being kept in an incubator prior to placement on the host.
The longipain dsRNA-injected ticks were placed on the ears of an 8-month-old female beagle (Oriental Bio service, http://www.oyc.co.jp) that was infected with the dog Babesia parasite B. gibsoni. During attachment, the dog maintained 12% intra-erythrocytic parasitemia in the peripheral blood. The pattern of the control ticks injected with buffer alone was comparable to that of the uninjected ticks used simultaneously to infect the same host. On day 6, ticks were recovered from the dog. The individual organs of ticks were dissected after removal of the midgut contents under a microscope. To verify the gene silencing by longipain dsRNA, RT-PCR was performed as described previously [75]. Total mRNA was isolated using a QuickPrep™ Micro mRNA Purification Kit (GE Healthcare) as described in the supplier's protocol. cDNA was then synthesized from 30 µg of mRNA using an RNA PCR Kit (AVM) Ver.3.0 (Takara) following the manufacturer's instructions. PCR was performed using longipain-specific oligonucleotides and oligonucleotides specific for H. longicornis with 500 ng of cDNA as template in a final volume of 50 µl. PCR products were resolved by 1.8 % agarose gel electrophoresis.
Tick sections were prepared as described above. Samples were immunolabeled with mouse anti-B. gibsoni (1∶250) and rabbit anti-longipain (1∶150). Bound antibodies were detected using Alexa Fluor 488 goat anti-mouse IgG(H+L) (1∶500) and Alexa Fluor 594 goat anti-rabbit IgG(H+L) (1∶250). The sections were mounted with DAPI and photographed with a fluorescence microscope as described above.
The prevalence and intensity of B. gibsoni infection in the dissected organs were evaluated using a real-time quantitative PCR assay. Initially we standardized the PCR protocol using B. gibsoni P18 gene-specific primers (D3:5′-TCCGTTCCCACAACACCAGC-3′, D4:5′-TCCTCCTCATCATCCTCATTCG-3′) and purified B. gibsoni genomic DNA. B. gibsoni P18 which encodes a major surface protein, is a well-known gene that has been demonstrated to be useful as a diagnostic tool for B. gibsoni infection [24]). The PCR reaction was performed using a LightCycler 1.5 (Roche) and DNA master SYBR Green I (Roche) with 4 mM MgCl2. Standard curves used to quantify relative gene concentrations were made from tenfold serial dilutions of the B. gibsoni parasites (genomic DNA) with the following incubation conditions: 95°C×600 s denaturing step, 45 cycles of 95°C×15 s, and 55°C×10 s, and 72°C×15 s, using the Fit Point Method of Light Cycle Software 3.5.3. This protocol resulted in highly specific amplification, with no amplification of dog, tick or a range of other B. gibsoni DNAs. Evaluation of the number of B. gibsoni using DNA extracted from excised organs was performed according to the established PCR protocol. DNA extraction and determinations of the concentration were performed as described.
Sequence data reported in this manuscript are available from GenBank (http://www.ncbi.nlm. nih.gov/Genbank) under accession number AB255051. |
10.1371/journal.pcbi.1004010 | Exploring O2 Diffusion in A-Type Cytochrome c Oxidases: Molecular Dynamics Simulations Uncover Two Alternative Channels towards the Binuclear Site | Cytochrome c oxidases (Ccoxs) are the terminal enzymes of the respiratory chain in mitochondria and most bacteria. These enzymes couple dioxygen (O2) reduction to the generation of a transmembrane electrochemical proton gradient. Despite decades of research and the availability of a large amount of structural and biochemical data available for the A-type Ccox family, little is known about the channel(s) used by O2 to travel from the solvent/membrane to the heme a3-CuB binuclear center (BNC). Moreover, the identification of all possible O2 channels as well as the atomic details of O2 diffusion is essential for the understanding of the working mechanisms of the A-type Ccox. In this work, we determined the O2 distribution within Ccox from Rhodobacter sphaeroides, in the fully reduced state, in order to identify and characterize all the putative O2 channels leading towards the BNC. For that, we use an integrated strategy combining atomistic molecular dynamics (MD) simulations (with and without explicit O2 molecules) and implicit ligand sampling (ILS) calculations. Based on the 3D free energy map for O2 inside Ccox, three channels were identified, all starting in the membrane hydrophobic region and connecting the surface of the protein to the BNC. One of these channels corresponds to the pathway inferred from the X-ray data available, whereas the other two are alternative routes for O2 to reach the BNC. Both alternative O2 channels start in the membrane spanning region and terminate close to Y288I. These channels are a combination of multiple transiently interconnected hydrophobic cavities, whose opening and closure is regulated by the thermal fluctuations of the lining residues. Furthermore, our results show that, in this Ccox, the most likely (energetically preferred) routes for O2 to reach the BNC are the alternative channels, rather than the X-ray inferred pathway.
| Cytochrome c oxidases (Ccoxs), the terminal enzymes of the respiratory electron transport chain in eukaryotes and many prokaryotes, are key enzymes in aerobic respiration. These proteins couple the reduction of molecular dioxygen to water with the creation of a transmembrane electrochemical proton gradient. Over the last decades, most of the Ccoxs research focused on the mechanisms and energetics of reduction and/or proton pumping, and little emphasis has been given to the pathways used by dioxygen to reach the binuclear center, where dioxygen reduction takes place. In particular, the existence and the characteristics of the channel(s) used by O2 to travel from the solvent/membrane to the binuclear site are still unclear. In this work, we combine all-atom molecular dynamics simulations and implicit ligand sampling calculations in order to identify and characterize the O2 delivery channels in the Ccox from Rhodobacter sphaeroides. Altogether, our results suggest that, in this Ccox, O2 can diffuse via three well-defined channels that start in membrane region (where O2 solubility is higher than in the water). One of these channels corresponds to the pathway inferred from the X-ray data available, whereas the other two are alternative routes for O2 to reach the binuclear center.
| Cytochrome c oxidases (Ccoxs) are the terminal enzymes of the respiratory chain in eukaryotes and in aerobic prokaryotes (reviewed in [1]). These integral membrane proteins belong to the heme-copper oxidases superfamily and couple dioxygen (O2) reduction to the translocation of protons across the membrane. Ccox takes up four electrons from cytochrome c (cyt c) in the positively charged side of the membrane (the inter-membrane space in mitochondria or the periplasm in bacteria) and eight protons from the negatively charged side (eq. 1) [2], [3]:(1)where the subscripts P and N refer to the positive and negative sides of the membrane, respectively.
Four of the eight protons reported in equation 1 are used to reduce one O2 molecule and form two water molecules [2], [3], whereas the remaining protons are pumped from the negative to the positive side of the membrane. This overall process contributes to the generation and maintenance of a transmembrane electrochemical proton gradient, which can be further utilized for several energy-requiring processes, such as ATP synthesis [4].
Based on structural and phylogenetic analysis, the heme-copper oxidases superfamily is currently divided into three major subfamilies [5]: A, B and C. The main differences between the three families are the pathways and mechanisms of proton transfer/pumping. The A-type Ccoxs, which are the subject of this work, are widespread through all kingdoms of life [5] and among them are the most thoroughly explored Ccoxs [3], [6], such as the bovine heart mitochondria, the Paracoccus (P.) denitrificans and the Rhodobacter (R.) sphaeroides enzymes. These Ccoxs contain, in the catalytic subunit (subunit I), a low spin heme a and a heterodinuclear center named binuclear center, BNC (Fig. 1A). The BNC is deeply buried in the core of the protein and it is formed by a high-spin heme a3 and a copper ion (CuB). In subunit II, these Ccoxs contain only one redox center, a binuclear copper center named CuA, which accepts electrons from the soluble cyt c and then transfers them to the BNC via heme a.
It is believed that protons (both chemical and pumped) are transported from the N-side of the membrane to the BNC via two special proton conducting pathways [3]: the D- and K-pathways (Fig. 1A). A third putative proton-conducting pathway, the H-pathway, was proposed for the mammalian Ccox only [6], [7], and it was suggested to be exclusively used for the transfer of the pumped protons [8].
Several high-resolution crystallographic structures of the A-type family are nowadays available in the literature (e.g. mammalian [7], [9]–[11] and bacterial Ccoxs [12]–[15]) and, based on these structures, it is known that all A-type members share a remarkable structural similarity of the core functional unit formed by subunits I and II (Fig. 1A). Subunit I consists of twelve transmembrane α-helices and contains the BNC and the heme a center. Subunit II is formed by a solvent exposed globular β-sheet domain (which functions as a docking surface for cyt c) and two transmembrane α-helices. It contains only one redox center, the binuclear copper center (CuA). Moreover, at the interface between subunits I and II, Ccox has one Mg+2 ion whose function is still not well understood, but it was suggested to be part of the exit pathway for the pumped protons and for water formed in the BNC [16], [17]. Subunit III, although not considered to be part of the core functional unit, is also highly conserved among the A-type subfamily. Nevertheless, its absence significantly increases the probability of suicide inactivation [18], [19] and thereby reduces the catalytic lifespan of Ccox (in 600-fold or more) [19].
Based in the X-ray data available (eg. [7], [12], [13]), a putative O2 channel for the A-type family was proposed (Fig. 1B). Iwata and co-workers, after pressurizing R. sphaeroides Ccox crystals with xenon, were able to identify a continuous hydrophobic channel that starts in the membrane region of subunit I [13]. This putative O2 channel has two possible entrances that merge together in a region close to the proton-gating residue, E286I (the residues are numbered according to the R. sphaeroides Ccox sequence and the subscript indicates the subunit number). This pathway presents a constriction point which does not allow the access of O2 to the BNC, at least without the occurrence of some conformational change in the protein. Unfortunately, until now, none of the mutagenesis and biochemical studies performed in this channel [20]–[22] was able to clearly demonstrate that it serves as an O2 route into the BNC. All the tested mutations were located too close to the BNC [20], [21], which made the interpretation of the results difficult and did not allow to unambiguously distinguish between the structural obstruction of the O2 channel and the perturbation of the BNC binding kinetics. However, and contrary to the A-type family, in the B-type family the channel used by O2 to reach the BNC is nowadays considered to be well established. The crystallographic studies (with xenon pressurization) performed in the Thermus (T.) thermophilus ba3 enzyme [23]–[25], lead to the identification of a “Y-shaped” hydrophobic channel that runs from the membrane region towards the BNC. This channel, although located roughly at the same position of the putative O2 channel in the A-type Ccox, does not possess a constriction point close to the BNC. In the A-type Ccoxs, the narrowing of the O2 channel is mainly caused by two conserved bulky residues (W172I and F282I in R. sphaeroides [13]), whereas in the B-type Ccox, smaller residues occupy these positions (Y133I and T231I in T. thermophilus [23], [24]). The differences between the A- and B-type regarding the O2 channel are thought to reflect the different functional environments of each type of Ccox.
Although the static crystal structures have been a valuable tool for providing insights into the O2 diffusion and for identifying potential O2 channels in Ccox, the elucidation of the molecular basis of O2 diffusion requires the knowledge of the Ccox conformational dynamics. Transiently formed cavities and openings inside the protein (frequently regulated by side chain rotation or by water movements) are not visible in the static X-ray structures, but have already been shown to be very relevant for ligand diffusion (see for example [26]). In this context, molecular dynamics (MD) simulation techniques (with sufficient simulation time and conformational sampling) appear as an alternative for studying the dynamic behavior of proteins and to determine their ligand occupation probabilities inside the protein. In the last decade, computational methods have been widely used to study gas migration in a number of proteins and MD simulations have successfully allowed the identification of several alternative routes for ligand diffusion (e.g. hydrogenase [26]–[28], myoglobins [29], [30], oxidases [31], [32] and laccases [33]). Moreover, the combination of MD simulations with Implicit Ligand Sampling (ILS) [29] calculations allows the calculation of the energy cost of transferring any small, apolar molecule (like O2 or H2) from the solvent to the protein and consequently to compute a 3D free-energy landscape for that specific ligand molecule (e.g [29], [33], [34]).
Over the last decades, most of the Ccox research using computational methods focused on the mechanisms and energetics of reduction and/or proton pumping (e.g [17], [35]–[56]). In the A-type Ccox, little emphasis has been given to the identification of the routes used by O2 to move from the solvent towards the BNC, a question only addressed, to our knowledge, by Hofacker and Schulten [31] and by Farantos and co-workers [31]. In the first work, Hofacker and Schulten [31] used MD simulations to study O2 diffusion in the vicinity of the BNC in a bacterial Ccox from P. denitrificans and in the bovine CcOx enzyme. Their simulations revealed a unique, well-defined O2 diffusion channel starting in the membrane-spanning surface of subunit I, close to the interface with subunit III. More recently, Farantos and co-workers [32] have applied the ILS method in order to study the binding of several small gas molecules around the BNC region in the A-type Ccox from P. denitrificans and in the B-type Ccox enzyme from T thermophilus. From these calculations, the authors were able to identify several cavities around the heme a3 region that are conserved in both the A-type and B-type enzymes. This study is however limited to the BNC region, not including other parts of the protein and, consequently, not allowing the analysis of the whole O2 permeation process.
The main objective of this work is to identify the O2 channels in the fully reduced Ccox from R. sphaeroides [15] using a combination of MD simulations (with and without explicit O2) and ILS calculations. Our results revealed the existence of three putative O2 diffusion channels. One of channels correlates very well with the channel inferred from the X-ray data available, whereas the other two are alternative routes for O2 to reach the BNC, and were not observed in the X-ray structures pressurized with xenon. Both alternative channels start in the membrane phase and terminate close to Y288I.
In this work, we investigate the diffusion of O2 molecules from the solvent to Ccox using MD simulations. We started our study by performing simulations with the enzyme (in the fully reduced state) with explicit O2 molecules, as described in detail in the Materials and Methods. Although all O2 molecules were initially placed randomly in the solvent, as time progresses, the gas enters the membrane and concentrates in the lipid tails region (see S8 Figure in Text S1), similarly to what has been described experimentally [57], [58]. After 100 ns of simulation, 46 O2 molecules were internalized in the membrane, which corresponds to more than 50% of the O2 placed originally inside the simulation box.
Furthermore, during the simulation time, some O2 molecules move from the membrane into the protein. The number of O2 inside Ccox increased slowly during the first 30 ns of simulation until it stabilizes at ∼8 molecules on average (see S9 Figure in S1 Text). In general, before entering Ccox, the O2 molecules explore the protein's surface and bind briefly to the cavities and niches formed mainly by hydrophobic residues. However, after 100 ns, none of the internalized O2 molecules was able to reach the BNC in any replicate. Nonetheless, and in order to determine which regions of the protein are more populated by the O2 molecules during our simulations, we calculated the O2 probability density maps [59] over the 100 ns (for all five replicates) and the results are depicted in Fig. 2.
As it can be seen, most of the high-affinity regions are located in the membrane phase (Fig. 2A) and some of them show a good correlation with the putative O2 channel inferred from the X-ray data pressurized with xenon [13] (Fig. 2B). This channel starts at the membrane spanning region in direct contact with the hydrophobic tails of the lipids and it is formed predominately by uncharged and aromatic residues. It has a Y shape with the two entrances located between helices 5 and 8 and helices 11 and 13 in subunit I. The side chains of I104I, L105I, A153I, L157I, F108I, L174I, V194I, L246I, and I250I contribute to form this channel. The two possible entrances merge together into a constriction point located close to E286I. This reduction in the diameter of the channel is caused by the phenyl ring of F282I and the indole ring of W172I (see Fig. 2B). It has been suggested that during the O2 reduction cycle, some protein rearrangement or side chain rotation (of F282I and/or W172I) is required in order to open the constriction point and allow O2 access to the BNC [7], [23]. In our O2 simulations, the side chains of F282I and W172I, although showing some flexibility and being able to slightly change their conformation, did not open an entrance point large enough to allow the passage of O2 into the BNC. For this reason, it was impossible to directly observe O2 diffusion towards the catalytic site.
Recent computational studies have suggested that the hydration level of the internal hydrophobic cavity located after the constriction point (formed by F282I, W172I and E286I) could be important for the regulation of proton transfer in Ccoxs (e.g [55], [56]). It was also suggested that the water distribution inside this cavity is modulated by the protonation of the heme a3 D-propionate and that these hydration changes strongly affect the E286I proton affinity [56]. This hydrophobic cavity bridges between the end of the D-pathway and the BNC region and it has been predicted to contain several water molecules, at least transiently (e.g. [38], [60]–[62]). However, in our simulations, no water molecules were observed in this region and no.direct and visible connection between E286I and the BNC was identified.
Since the simulations with explicit O2 were not able to properly sample the diffusion events (probably due to the simulation timescales), and in order to identify possible alternative O2 diffusion channels in fully reduced Ccox, the ILS method [29] was used. This method uses the protein conformations obtained from a ligand-free MD trajectory and calculates, for all the positions inside the protein, the potential of mean force for placing a small, apolar, low-interacting molecule at that point. The 3D energy map generated in this case represents the Gibbs free energy of moving a O2 molecule from water into any place inside the protein (ΔGwat→prot(O2)) and it can be used to infer about the energetically favourable diffusion paths inside a given protein. In these maps, the regions where ΔGwat→prot(O2) is low represent the positions where O2 has a high probability of residing, meaning that the O2 affinity in that position is high. The ILS method reduces significantly the sampling problems and it takes into account the dynamic conformational changes of the protein and all the transiently formed cavities, which can be combined to form transitory diffusion pathways.
From the 3D affinity map for Ccox obtained from the ILS calculations (Fig. 3A), we can see that Ccox possesses a complex free energy landscape with several possible O2 binding cavities, most of them located at the protein surface and that do not directly connect to the BNC. Moreover, the calculated affinity map from ILS not only correlates well with some of the high probability regions found in the explicit O2 simulations (see Fig. 3B) but also provides a more complete description of the free energy landscape for O2 inside Ccox due to the sampling of lower affinity zones, not sufficiently sampled by the normal MD simulations. Indeed, the ILS approach allowed us to identified three major routes (named channel 1, channel 2 and channel 3) with high probability of O2 occupancy, interconnecting the protein surface to the BNC. Two of the channels (channel 1 and channel 2) approach the BNC from the subunit I side whereas the third one (channel 3) approaches the BNC from the subunit II side (Fig. 3A). Interestingly all three channels start in the membrane spanning region where the O2 concentration is higher than in aqueous phase, which makes physical sense.
Furthermore and in order to identify the energetically preferred routes for O2 to access the BNC, we extracted, from the ILS affinity map, the lowest free energy pathways connecting the exterior with the O2 high affinity sites (basins in the O2 free energy landscape) inside Ccox, using the same methodology as Damas et al [33] (for details, see the Data Analysis section of Materials and Methods). From the analysis of the O2 free energy landscape (Fig. 3C), we observe that there are many energy minima and low free energy pathways connecting the solvent/membrane region to the interior of the protein. Nonetheless, all these low free energy entrance channels converge into three distinct pathways as we approach the BNC.
The O2 channel 1 approaches the BNC from the subunit I side and corresponds to the channel inferred from the X-ray data (for R. sphaeroides [13] or for bovine [7]). This pathway has two entry points that are fused together in a free energy minimum located in the constriction point just before the BNC (M10 in Fig. 4A). The free energy profile for this pathway (Fig. 4A) is characterized by a very high permeation free energy barrier in the constriction point (associated to the bulky side chain of W172I) and three deep local minima (M8, M9 and M10 in Fig. 4A) located between W172I, F282I and E286I for M10, L243I, F282I and L283I for M8, and I250I, V194I and F108I for M9. The local minima observed just below the Ccox surface (M1 and M6) can probably act as scavengers for the O2 freely diffusing in the membrane and help to create an O2 reservoir inside the protein. In this pathway, W172I and E286I seem to act as the gateway residues that control O2 access to the catalytic site. Nevertheless, the passage from the constriction point to the BNC (moving from M10 to M11 in Fig. 4A) implies the overcoming of a very high free energy barrier of 39.5 kJ·mol−1 (see Fig. 4A and Fig. 5A), which makes O2 diffusion via this pathway very slow and difficult.
Over the last 20 years, several biochemical studies [20]–[22] have tried to demonstrate that this pathway is the one used for O2 diffusion in the A-type Ccoxs, but, until now, no direct measurement proved this unequivocally. All the tested mutations (e.g V287II [20], [21] and G283IV [22]) were located too close to the two metals (Fe in heme a3 and Cu in CuB) in the BNC, which made the interpretation of the results very difficult. It was impossible to unambiguously distinguish between the steric hindrance of the O2 channel (introduced by the mutations) and the perturbation in the structure and local binding kinetics at the BNC. Moreover, until now, only one X-ray structure of the A-type Ccoxs has been pressurized with xenon (considered to be an O2 analog) [13]. In this structure [13], two hydrophobic xenon binding sites were identified, none of them located beyond the constriction point of channel 1. One xenon was observed at the entrance of the X-ray inferred channel while the other was close to E286I. Since no xenon was observed after the constriction region, there is still no consensus whether this pathway is indeed a functional channel for O2 diffusion in these Ccoxs. Additionally, Hofacker and Schulten [31], with the objective of studying O2 diffusion towards the BNC, performed several MD simulations of Ccoxs with several explicit O2 molecules. The authors used the locally enhanced sampling (LES) technique [63], [64] to identify the pathway used by O2 both in a bacterial (P. denitrificans) and in the bovine Ccox enzyme. In their simulations, some of the O2 molecules reached the BNC via a unique and well-defined channel located in the same region as the X-ray inferred channel from [13]. Given the very short duration of their simulations (picoseconds), the observation of such events was only possible due to the lowered free energy barriers experienced by the O2 molecules when using the LES technique. Nevertheless, a single protein trajectory is still effectively employed in the LES approach, making the protein conformational sampling of that study more limited than the one in the present study. Finally, Farantos et al. [32] by performing ILS calculations in the A-type Ccox from P. denitrificans and in the B-type Ccox enzyme from T thermophilus, were able to calculate the free energy surfaces for the interaction of several small polar (CO, NO) and apolar (O2 and Xe) around the BNC region. Regarding the X-ray inferred channel, their results correlate very well with the ones reported in the present study and the high affinity cavity Xe1 observed by the authors in [32] coincides with minima M11 in the O2 channel 1.
The channel 2 (Fig. 3A, Fig. 4C and 4D) also starts at the membrane region and has only one possible entrance point located between the transmembrane helices 13 and 16 of subunit I, around residues V234I, L238I and V325I. Moreover, this channel terminates close to Y288I, a tyrosine residue covalently bonded to H284I, one of the CuB-coordinating histidines. Y288I is located in the end of the K proton-conducting pathway and it has been suggested to be an essential residue for the catalytic mechanism, supplying the proton used in the cleavage of the O-O bond (see [65] for a review). The energy profile for this alternative channel (Fig. 4C and 4D) is characterized by one very low local free energy minimum (M19 in Fig. 4C) located between Y288I and T359I and by a high free energy minimum (M13 in Fig. 4C) around L292I and I323I. The highest energy barriers for this small pathway are 20.4 kJ·mol−1 (Fig. 5B), which are substantially smaller than the energy barrier in the constriction point of the O2 channel 1 (39.5 kJ·mol−1). However, the high energy minimum M13 located in the middle of the channel (see Fig. 4C) may hinder O2 diffusion via this pathway.
Lastly, the O2 channel 3 approaches the BNC from the subunit II side (Fig. 3A, Fig. 4C and 4D) and its entrance is located between the transmembrane helices 28 and 30 of subunit II, around residues F71II and V167II. This channel runs parallel to the heme a3 hydroxylethylfarnesyl tail and also terminates just bellow Y288I. It is generally formed by hydrophobic and aromatic residues, such as I363I, F391I and V272II (see Fig. 4D). The energy profile for this alternative channel (Fig. 4C) is characterized by only one deep local energy minimum (M19 in Fig. 4B) located between Y288I and T359I and by several smaller (in comparison with the O2 channel 1) energy barriers, which suggest a higher probability of O2 passage. In this pathway, the largest energy barriers (19.9–22.3 kJ·mol−1 in Fig. 5C) are related with the zone of the hydroxylethylfarnesyl tail of heme a3 and the side chains of I363I and F391I. This alternative O2 channel was firstly suggested by Tsukihara et al. [7], based on the X-ray structure of the bovine heart Ccox, as one of the three possible O2 entrance points in the A-type oxidases. More recently, Farantos et al. [32] were also able to observe the final part of this channel (located close to heme a3) in the A-type Ccox from P. denitrificans.
Even though only subunits I and II were considered in this work, the entry points for all the channels described above are not obstructed by the missing subunits.
Moreover, it is also interesting to notice that all O2 channels overlap in space with canonical proton pathways. The O2 channel 1, which corresponds to the X-ray inferred channel, overlies with the end of the D-pathway, whereas the two alternative channels (O2 channel 2 and channel 3), which end close to Y288I, overlap with the final part of the K-pathway.
For the two alternative channels identified here, the O2 pathway is not permanently open and by this reason escaped detection when the static X-ray structures were inspected (e.g [7], [12], [13]). These channels resemble a chain of separate, multiple and transiently formed hydrophobic cavities that are interconnected with each other via the fluctuations of the residues lining the pathway. During the diffusion process, O2 enters the protein and probably due to the thermal fluctuations of the heme a3 tail and residues lining the pathways, it is allowed to diffuse further into the protein core until it reaches the BNC.
Deciphering the actual role of these O2 channels will require further mutational and biochemical analysis. Several mutations sites can be suggested (see Table 1), on the three O2 channels, in order to clarify the role of each pathway in O2 diffusion towards the BNC in the A-type Ccoxs.
For the X-ray inferred channel, it is clear that alternative mutation sites, located farther away from the metal ions in the BNC, need to be constructed and studied in order to clarify the actual role of this pathway in O2 diffusion. W156I and F282I are excellent candidates for mutation experiments and their mutation for less bulkier residues, such as tyrosine and threonine (similar to what is observed in the T. thermophilus ba3 enzyme [23]–[25]) would give important insights into how O2 diffusion occurs via this channel. Furthermore, for the alternative channels, the residues lining both pathways are also good candidates for mutational studies, because longer and/or bulkier side-chains could, in principle, obstruct the alternative channels and thus clarify the question of whether these channel are, indeed, used to supply O2 to the BNC in the A-type Ccoxs. Nevertheless, until new experimental data is available, we cannot rule out the hypothesis that all three channels may be working under physiological conditions.
Although A-type Ccoxs have been widely studied during the last four decades, the details of the O2 diffusion mechanism are still very incomplete. In particular, the existence and the characteristics of the channel(s) used by O2 to travel from the solvent/membrane to the BNC are still unclear. In this study, we have used an integrated strategy of all-atom MD simulations (with and without explicit O2 molecules) and ILS calculations, designed to examine and characterize the O2 delivery channels in fully reduced Ccox from R. sphaeroides. Altogether, our results suggest that O2 does not diffuse unspecifically inside this protein and instead, uses three well-defined channels running from the interior of the membrane (where O2 solubility is higher than in the aqueous phase) towards the Ccox core. The first pathway has two entrance points, located between helices 5 and 8 and helices 11 and 13 of subunit I, which converges into the constriction point just before the BNC. This channel correlates very well with the channel inferred from the available X-ray structures. The second pathway has only one entry located between the transmembrane helices 13 and 16 of subunit I and it terminates close to Y288I. The third identified pathway approaches the BNC from the subunit II side. This channel runs parallel to the heme a3 hydroxylethylfarnesyl tail and also terminates just below Y288I. According to our observations, the hydrophobic channel detected in the X-ray structures does not constitute the most likely (energetically preferred) entrance point for the O2 molecules in this Ccox. From the O2 affinity map, O2 accesses the BNC via the alternative dynamic channels formed by transient hydrophobic cavities, whose opening and closure is regulated by the thermal fluctuations of the protein. This may be the reason why these channels were not visible in the static X-ray structures.
In summary, our results suggest that the original hypothesis (based on static X-ray structures and mutational studies on A-type Ccox) that proposed, that O2 permeation occurs via a unique, continuous and permanently open channel, is indeed a simplification. Our current work does not rule out the role of the X-ray inferred channel, but suggests other alternative routes to the BNC. Furthermore, it emphasizes the need to take into account the dynamic behavior of the protein in order to obtain a more complete description of the O2 putative channels and a more detailed picture of the mechanisms underlying O2 diffusion in these Ccoxs.
The 2.15 Å resolution crystal structure of the fully reduced Ccox from R. sphaeroides (pdb code: 3FYE) [15] was used as the starting point for this work. This X-ray structure only contains the minimum functional unit (subunits I and II) for Ccox. Only the water molecules with a relative accessibility to the solvent lower than 50% were kept. The relative accessibility of water was computed using the program ASC [66], [67], resulting in the selection of 240 water molecules.
Since the GROMOS 54A7 force-field [68] lacks the proper parameterization for the Ccox redox centers, the atomic partial charges for reduced CuA, heme a and BNC centers were calculated using quantum mechanical calculations with the software Gaussian09 [69] and RESP fitting [70], as described in detail in S1 Text in section 1. The van der Waals parameters for the iron atom (located in the two heme groups) were taken from the universal force field [71] whereas the remaining bonded and van der Waals parameters for the metal centers were adapted from the GROMOS 54A7 force field [68].
The protonation state of each individual protonable group at pH 7.0 was determined using a combination of Poisson-Boltzmann calculations, performed with the package MEAD (version 2.2.5) [72]–[74], and Metropolis Monte Carlo simulations, using the program PETIT (version 1.3) [75]. These calculations were performed using the methodologies described in [75], [76]. For details related with the determination of the protonation state of the protonatable residues, see section 2 in S1 Text.
Subunits I and II of Ccox were inserted in a pre-equilibrated dimysristoylphosphatidylcholine (DMPC) lipid membrane (for details related with the membrane construction, equilibration and characterization see [77]). The optimal position of the protein relative to the membrane was determined using the location of the charged residues in the transmembrane helices as a reference. After Ccox insertion into the membrane, all the DMPC molecules located within a cut-off distance of 1.2 Å from the protein atoms were removed, as described in detail elsewhere [77], [78]. Subsequently, the system (protein, membrane and crystallographic waters) was hydrated in a orthorhombic box using a pre-equilibrated box of SPC water molecules [79]. The water molecules misplaced in the center of the membrane (formed by the highly hydrophobic lipid tails), were removed upon visual inspection. The final system contained the reduced Ccox embedded in a 175 DMPC lipid membrane surrounded by 19,645 water molecules, in a total of 75,178 atoms.
All MD simulations were performed using the software package GROMACS 4.0.4 [80] together with the united atom GROMOS 54A7 force-field [68] for the protein and lipids and the previously described atomic partial charges and parameters for the redox centers. The simple point charge (SPC) water model was used [79]. Periodic boundary conditions were applied to all simulations. Non-bonded interactions were calculated using a twin range method [81] with short and long-range cut-offs of 8 and 14 Å, respectively. A reaction field correction [82], [83] was applied for the truncated electrostatic interactions, considering a dielectric constant of 62 [84]. The SETTLE algorithm [85] was used to constraint the bond lengths and angle in water molecules, while the LINCS algorithm [86] was used to keep all remaining bonds constrained. The time step for integrating the equations of motion was 0.002 ps and the neighbor list was updated every 5 steps. The simulations were performed at the constant temperature of 310 K, which is above the phase transition temperature for the DMPC lipids (Tm = 296–297 K) in order to ensure that the membrane is in the liquid crystalline state [87]. A Berendsen heat bath [88] was used, with separate couplings for the protein, membrane and solvent, using a relaxation time constant of 0.1 ps. The pressure was coupled semi-isotropically (coupling constant of 5.0 ps and isothermal compressibility of 4.6×10−5 bar−1 [84]), resulting in an independent coupling of the lateral (Px+y) and perpendicular (Pz) pressures. For all simulations, the x+y and z pressure components were kept at 1 atm and no surface tension was applied [84]. These simulation conditions were shown by Poger et al. [84], [89] to correctly reproduce several experimental measurements for this type of membranes.
The system was energy minimized with the steepest-descent method in order to remove excessive strain by performing 5000 steps of steepest-descent minimization with harmonic restraints applied to all non-hydrogen atoms (protein and lipids), followed by further 5000 steps restraining the non-hydrogen atoms of the protein, ending with 5000 steps with restraints applied to the Cα atoms only. After the minimization procedure, and in order to allow proper repacking of the lipids around the protein, a 20 ns MD relaxation was executed in three steps. First, a 0.5 ns simulation was performed with position restraints to all non-hydrogen atoms of the protein and solvent, at constant temperature and pressure. Afterwards, an additional 0.5 ns simulation was performed, with position restraints applied to the non-hydrogen atoms of the protein only. Finally, only the Cα atoms were restrained for a period of 19 ns. A force constant of 1000 kJ mol−1nm−2 was used for all the steps that included harmonic position restraints. The unrestrained simulations started after these 20 ns of restrained simulation.
In order to reduce the well known sampling problems in membrane-protein simulations, five MD simulations, 100 ns each, were performed, resulting in 0.5 µs of total simulation time. All replicates were initiated with different sets of random velocities. These simulations will be hereafter designated as O2-free simulations.
After 20 ns of restrained simulations, we randomly added 84 molecules of dioxygen (O2) in the solvent zone of each system. No O2 was placed inside the protein nor inside the hydrophobic core of the membrane (see S4 Figure in S1 Text). This new set of simulations will be, hereafter, designated as O2 simulations. The water molecules within a 2 Å distance from the O2 molecules were deleted, similarly to the procedure described in [33].
In order to allow the solvent to adapt to the newly added O2 molecules, a 0.5 ns MD simulation with position restraints on all non-hydrogen atoms (force constant of 1000 kJ mol−1nm−2) was performed. After this initialization procedure, unrestrained MD simulations were carried out and the simulation conditions and parameters were similar to the ones described previously for the MD simulations without O2, except for the temperature coupling groups used. In this set of simulations, the O2 molecules were included in the same group as the protein. 5 MD simulations, 100 ns each, were performed. The parameters for the O2 molecules were taken from the previously published work of Victor et al [90].
The 84 O2 molecules added to the system corresponds to an O2 concentration of ∼0.235 M, which is higher than the experimental solubility of this gas in water. However, this high O2 concentration does not affect the structural properties of the protein as shown in S5 Figure in S1 Text.
Moreover, the use of this high number of O2 molecules is necessary to obtain reliable statistics within a reasonable simulation time.
Sites with high O2 affinity were determined using the ILS method, as previously described in [29]. In this method, the potential of mean force for placing an O2 molecule in any position inside the protein is calculated according to:(2)
In equation 2, the implicit ligand potential of mean force, , is an average over a finite number of protein and solvent configurations () and over a number of different equally probable orientations of the ligand (). Moreover, is the Boltzmann constant, is the absolute temperature, and is the interaction energy between the protein and solvent configuration () with the ligand located at position with the orientation .
In our case, the O2 free energy map was constructed using the last 50 ns (for each replicate) of the O2-free simulations. For the calculations, all 50005 conformations ( = 10001 conformations x 5 replicates in equation 2) were fitted to the X-ray structure using the Cα atoms. A grid of 51×55×87 dimensions was used with a grid spacing of 1 Å and 400 O2 insertions were performed per grid point ( = 400 in equation 2). All calculations were carried out using a version of GROMACS 4.0.4 Widom TPI algorithm, modified almost in the same way as described in [33]. The only difference is that the ligand insertions here were made within the whole space of the grid cube (the grid cube is centered at the insertion point and with edge length equal to the grid spacing), while in the previous work (described in [33]) the insertions were only possible within the inscribed sphere on the grid cube.
The 3D free energy map obtained describes the Gibbs free energy of moving an O2 molecule from vacuum to a given position in the system, ΔGvac→prot(O2). This map was then converted into the ΔGwat→prot(O2) map of interest using a ΔGwat→prot(O2) calculated as described in [33].
The secondary structure assignment was performed with the program DSSP [91]. To determine the percentage of secondary structure loss relative to the X-ray structure, the secondary structure classes considered were: α-helix, 310-helix, 5-helix, β-sheet and β-bridge.
For the energy landscape analysis, we used the method described in [33]. In short, this method classifies the energy landscape into energy basins through a steepest-descent tessellation and, afterwards, identifies the lowest-energy point within the boundaries between each pair of neighboring basins, i.e. the saddle point between those basins. After this procedure, a network of paths between all energy minima of the landscape can be constructed using the steepest-descent paths from the saddle points to the minima. A cutoff of 20 kJ·mol−1 was used for the network construction.
The errors of the free energy profiles were calculated using two blocks: the first block corresponds to the frames ranging from 50 ns to 75 ns (for all five replicates) whereas the second block contains all the frames ranging from 75 ns to 100 ns. The errors were determined as half the difference of the energies observed between the two blocks for each minina and for each transition. The method used for error calculation assumes that similar minima and pathways can be identified in the two blocks. However, for the O2 channel 1, the pathways connecting M6 to M8 and M10 to M11 were not visible in one of the blocks and by this reason their associated errors could not be calculated.
|
10.1371/journal.ppat.1001220 | HIV Capsid is a Tractable Target for Small Molecule Therapeutic Intervention | Despite a high current standard of care in antiretroviral therapy for HIV, multidrug-resistant strains continue to emerge, underscoring the need for additional novel mechanism inhibitors that will offer expanded therapeutic options in the clinic. We report a new class of small molecule antiretroviral compounds that directly target HIV-1 capsid (CA) via a novel mechanism of action. The compounds exhibit potent antiviral activity against HIV-1 laboratory strains, clinical isolates, and HIV-2, and inhibit both early and late events in the viral replication cycle. We present mechanistic studies indicating that these early and late activities result from the compound affecting viral uncoating and assembly, respectively. We show that amino acid substitutions in the N-terminal domain of HIV-1 CA are sufficient to confer resistance to this class of compounds, identifying CA as the target in infected cells. A high-resolution co-crystal structure of the compound bound to HIV-1 CA reveals a novel binding pocket in the N-terminal domain of the protein. Our data demonstrate that broad-spectrum antiviral activity can be achieved by targeting this new binding site and reveal HIV CA as a tractable drug target for HIV therapy.
| Although the current standard of care for Human Immunodeficiency Virus (HIV) is high, viral resistance has emerged to every drug currently in the clinic, in some cases rendering the entire class ineffective for patients. A new class of antiretroviral drugs would be effective against strains of HIV-1 that are resistant to any existing drug and would expand the therapeutic options available to patients. Capsid is the primary structural protein of HIV and a critical part of the viral replication cycle, both in the assembly of viral particles and in the infection of host cells. We report a new class of antiretrovirals that targets HIV-1 capsid and demonstrate that it is active at two critical stages in the viral replication cycle. These compounds were consistently effective against a range of clinical strains of HIV-1, from various sub-types, as well as HIV-2. Finally, the compounds bind in a unique pocket on capsid that has not previously been highlighted as a drug binding site. We believe this new class of antiretrovirals can serve as a starting point for the development of a new generation of HIV-1 therapeutics and, more generally, underscores the potential of capsid as a therapeutic target.
| Highly active antiretroviral therapies (HAART) against human immunodeficiency virus type 1 (HIV-1) have proven in recent years to be extremely effective at reducing viral load and significantly delaying disease progression [1]. However, there remains a pressing need to discover and develop new classes of HIV inhibitors. The virus continues to acquire resistance to currently administered antiretroviral drugs and the rate of transmitted resistance is increasing [2], [3]. The discovery of compounds that inhibit the replication of HIV-1 via new mechanisms offers the best hope of generating drugs that are active against all HIV-1 variants in the clinic. The potency of these compounds would not be affected by mutations that confer resistance to existing therapies [4].
The capsid protein (CA) of HIV-1 plays critical roles in both late and early stages of the viral replication cycle and is widely viewed as an important unexploited therapeutic target [4], [5], [6]. At the earliest stages of particle assembly, the interactions between CA domains of the Gag polyprotein help drive the formation of immature particles at the membrane of host cells [7]. After the release of immature particles from infected cells, proteolytic processing of the Gag polyprotein is completed, leading to capsid assembly and formation of the mature virus. During assembly, the viral RNA genome is packaged into a capsid particle composed of a lattice of CA protein hexamers that form a distinct fullerene cone shaped particle [8]. After virus fusion with a target cell, the core is released into the cytoplasm and CA is thought to undergo a controlled disassembly reaction in order for reverse transcription of the viral genome to occur properly [9].
The HIV-1 CA protein has attracted increased interest as a drug discovery target in recent years. A small molecule, CAP-1, and two versions of a peptide inhibitor, CAI and NYAD-1, have been described that target HIV-1 CA in vitro and appear to interfere with CA function in infected cells [10], [11], [12]. In addition, high resolution structural data on the hexameric lattice that forms the full core structure has been reported [13], [14]. These structures illustrate the distinct roles and importance of inter-subunit interfaces in the CA complex and have shed some light on the potential mechanisms of previously reported CA assembly inhibitors.
Here we describe a novel series of antiviral compounds that target HIV-1 CA in infected cells and appear to interfere with both the viral uncoating process and the formation of infectious particles. Mechanism-of-action and in vitro resistance studies of this series are described. A high resolution co-crystal structure has been determined and illustrates a novel binding pocket in the N-terminal domain (NTD) of HIV-1 CA that is distinct from any previously described. We demonstrate that targeting this new binding pocket with small molecules results in broad-spectrum antiviral activity. This study provides the starting point, where structure-based drug design is a viable option, for the development of a new class of HIV therapeutics.
PF-1385801 was identified as a hit in a high throughput screen for inhibitors of HIV replication [15]. Several analogs of this compound demonstrated activity in antiviral assays using the MT-2 T-cell line and HIV-1 NL4-3. PF-1385801 inhibited HIV-1 replication with a 50% effective concentration (EC50) of 4.5 µM and exhibited a 50% cytotoxic concentration (CC50) of 61 µM, resulting in a therapeutic index (TI, CC50/EC50) of 14 (Fig. 1). More potent analogs were subsequently designed and synthesized, as detailed in Fig. 1.
A key issue in the development of novel HIV drugs is the potential therapeutic spectrum (consistency of activity across multiple strains of HIV). To address this, two of the compounds were evaluated in antiviral assays using peripheral blood mononuclear cells (PBMCs) infected with a diverse set of HIV-1 lab strains and clinical isolates covering 6 clades and both X4 and R5 tropic viruses (Fig. 2). PF-3450074 and PF-3759857 were active against all strains of HIV-1 tested with median EC50 values of 0.207 (range 0.113 to 0.362 µM) and 1.17 (range 0.51 to 3.17) µM, respectively (Fig. 2 and Tables S1 and S2). In addition, PF-3759857 was active against HIV-2 with an EC50 of 4.7 µM. This tight spectrum of activities against HIV-1 compared well with the marketed drugs, AZT, a nucleoside analog, and efavirenz (EFV), a non-nucleoside reverse transcriptase inhibitor (NNRTI) (Fig. 2 and Tables S3 and S4).
To positively identify the antiviral target of this class of inhibitors, PF-1385801 resistant viral variants were selected in in vitro serial passage experiments. Sequence analysis of cDNAs derived from resistant viral variants selected in the presence of PF-1385801 revealed a single mutation, T107N, located in the NTD of the CA protein. No mutations associated with resistance selection were identified in the integrase or reverse transcriptase coding sequences. Recombinant HIV-1 NL4-3 virus encoding the T107N substitution in HIV-1 CA exhibited an 11-fold reduction in susceptibility to PF-1385801 when compared to wild-type NL4-3 (Table 1). Similar reductions in susceptibility were observed for PF-3450071 (6-fold) and PF-3759857 (10-fold), while no shift was observed for EFV (NNRTI) or the integrase inhibitor AG-110079. In subsequent in vitro resistant virus selection experiments using a compound with a larger TI, PF-3759857, additional substitutions were identified in CA (T107N, H87P, Q67H, K70R and L111I), after 53 days in culture. Virus isolated from the final passage of this experiment showed a >60-fold reduction in susceptibility to PF-3759857 and recombinant NL4-3 virus encoding the five substitutions (5M) exhibited a >40-fold reduction in susceptibility to PF-3450074 (up to the cytotoxicity limit of the compound). Wild-type NL4-3, the T107N mutant, and the 5M mutant virus all exhibited similar levels of infectious virus production from transfected cells and comparable levels of viral replication in reporter gene-based infection assays (in the absence of compound), suggesting that the substitutions do not significantly impair replication capacity. However, more detailed studies, such as direct competition assays are required to accurately determine if the substitutions affect replication fitness.
These results demonstrate that mutation(s) in HIV-1 CA (e.g. T107N) are sufficient to confer resistance to this new class of inhibitors and demonstrate that HIV-1 CA is the target in infected cells. Although additional work is needed to determine the individual contributions of the mutations identified in addition to T107N, it appears that more than one mutation is required to generate high levels of resistance to this inhibitor class. While it is a definitive tool for target identification, it should be noted that in vitro selection can be used to generate resistant mutants to any known antiviral and does not accurately predict the clinical barrier to resistance, which is a function of several in vivo parameters, including the pharmacokinetics of the compound.
To determine the stage of the replication cycle targeted by this new class of compounds, PF-3450071 and PF-3450074 were evaluated in single cycle infection assays. Such assays monitor the early steps of infection up to the integration of the viral cDNA into the host cell chromosome and expression of that sequence. PF-3450071 and PF-3450074 were tested in parallel infections with single-cycle HIV-1 virus packaged with either wild type HIV-1 envelope (NL4-3 pseudovirus) or vesicular stomatitis virus glycoprotein (VSVG), an envelope that allows viral entry by an alternative mechanism (VSV pseudovirus). In addition, three compounds with known mechanisms, AMD3100 (HIV CXCR4/entry inhibitor), EFV (NNRTI/early mechanism) and NFV (protease inhibitor (PI)/late mechanism) were evaluated in the assays. PF-3450071 and PF-3450074 were active against both the NL4-3 and VSV pseudoviruses (Table 2). This inhibition profile was similar to EFV. The profile was distinct from that of AMD3100, which inhibited the HIV-1 envelope NL4-3 pseudovirus, but not the VSVG pseudovirus, and NFV which was not active in the assay (Table 2). These data indicate that PF-3450071 and PF-3450074 act early in the HIV-1 replication cycle at a step following HIV-1 envelope-mediated entry.
To begin dissecting the early events of infection, DNA was isolated from single cycle infections of MT-2 cells in the presence or absence of inhibitors and analyzed by quantitative PCR. Primer sets were used that detect total viral cDNA produced, 2-LTR circles (a nuclear episomal form of the viral cDNA), or provirus (the viral genome integrated into host cell chromosomes) as described previously [16]. PF-3450074 inhibited the accumulation of total viral cDNAs (9% of control) and as result inhibited the levels of integrated provirus DNA measured (2% of control) (Fig. 3a). This profile was similar to that observed for the RT inhibitor EFV, indicating that PF-3450074 either inhibits reverse transcription or a step prior to it. In contrast, the integrase inhibitor, AG-110079, resulted in a 4-fold increase in 2-LTR circle accumulation and minimal reduction in total viral cDNA (78% of control) (Fig. 3a). As the profile for PF-3450074 did not match that of the integrase inhibitor control, integrase inhibition can be ruled out as a possible mechanism.
To further analyze the mechanism of action, the inhibition profile of PF-3450071 was compared to that of AMD3100 and EFV in synchronized time-of-addition experiments, where inhibitors are added a various time points following infection to monitor when the susceptible step has occurred. The viral entry inhibitor, AMD3100, showed a dramatic early loss in activity, even within the first two hours of infection. The NNRTI, EFV, retained the majority of inhibitory activity (>78%) when added up to 4 hours after infection and lost activity thereafter, with a mid-point around 6 hours after infection. PF-3450071 displayed a profile that was distinct from both comparators. The compound maintained the majority of its inhibitory activity (>84%) when added up to 2 hours after infection, and lost significant levels of activity when added 3 or more hours after infection (Fig. 3b). These results suggest that this new class of HIV-1 inhibitor targets an early step in the virus replication cycle, after viral entry but before reverse transcription, possibly viral uncoating. Consistent with these data, PF-3450074 did not inhibit recombinant HIV-1 RT in standard biochemical assays (IC50>100 µM).
Virus production-infectivity assays were conducted to determine if compounds affect the late stages of viral replication. Infectious virus was expressed from transfected producer cells in the presence or absence of compound and the infectivity of the resulting virus was tested by diluting the supernatant on to an indicator cell line. As infectious virus is produced from a transfected DNA, bypassing the early stages of the replication cycle, only inhibitors targeting the later stages of viral replication (i.e. post-integration) should show activity in the assay. We tested two compounds (PF-3450071 and PF-3450074) in the viral production-infectivity assay using NFV (PI) and EFV (NNRTI) as late stage and early stage inhibitor controls. Both PF-3450071 and PF-3450074 inhibited the production of infectious virus with EC50 values of 0.78 and 0.33 µM, respectively (Table 2). As expected the HIV-1 PI, NFV, inhibited infectious virus production while the RT inhibitor, EFV was not active in the assay (Table 2). Notably, the overall levels of Gag proteins released into the supernatant of transfected cells were not affected by PF-3450074, as measured by p24 ELISA (Fig. 3c). Thus, this series does not inhibit HIV-1 particle production but renders the nascent particles noninfectious.
Western blot analysis of viral supernatants with antibodies directed against the HIV-1 CA protein (p24) showed that PF-3450071 did not affect HIV-1 Gag proteolytic processing (Fig. S1), which demonstrates that the compounds do not inhibit HIV-1 protease or virion maturation in the manner described for inhibitors such as Bevirimat (PA-457) [17] and PF-46396 [18]. Collectively, these data suggest an effect of the series on proper assembly of fully processed CA protein in nascent viral particles.
To observe the effects of this class of compound on the morphology of nascent viral particles, PBMCs were infected with HIV NL4-3 either in the absence or presence of PF-3450074 at 7 µM (∼10× EC50). The cells and resulting viral products were fixed and treated for transmission electron microscopy (TEM). In untreated infections, approximately 66% of particles (78/118 observed) were native-like, as defined by being ∼100nm in diameter and having a distinct central density representative of a mature capsid core). A close in view of a native-like particle from an untreated culture (Fig. 4a) clearly illustrates the conical capsid core in the center of the virion. A wider view shows the high level of uniformity among the particles produced in the untreated infections (Fig. 4b). Both native-like and apparent immature particles are present at consistent shape and size. In contrast, native-like particles were not observed in PF-3450074-treated infections. The particles produced in PF-3450074-treated infections lack a clear central density representative of a mature capsid (Fig. 4c). Consistent with the observation that these compounds do not affect p24 levels, the number of particles in PF-3450074-treated infections appears similar to untreated, however the morphology and size of particles is highly variable (Fig. 4d).
The simplest explanation to reconcile the data above involves a model where this new class of inhibitors targets both early (viral core uncoating) and late (viral core assembly) events in the replication cycle by affecting CA – CA interactions and thus core stability. To study the effects on capsid assembly, we evaluated the inhibitors in in vitro CA multimerization assays [10]. Such assays can be used to measure the effect of compounds on the rate of formation of higher order CA multimers or tubes that are widely thought to represent many aspects of native core structure. Addition of PF-3450074 results in a significant increase in the rate of CA multimerization (Fig. 3d). In contrast, a structural analogue with no antiviral activity, PF-4159193 (Fig. S2), did not affect the kinetics of CA multimerization, indicating that this profound effect is correlated with antiviral activity and not a general physical effect of this series. As we have reproduced with the CAI peptide (Fig. 3d), all of the previously reported HIV CA assembly inhibitors decreased the rate of multimerization in this assay [10], [11], [12]. These data demonstrate indirectly that PF-3450074 interacts with HIV-1 CA and further suggest a mechanism that is fundamentally distinct from previously reported HIV CA inhibitors.
To further understand the mode of action of this novel class of compounds, we determined the crystal structure of HIV-1 CA NTD protein in complex with PF-3450074 using a CA protein construct that contained a single glycine residue in place of the cyclophilin binding loop (residues 87–99). Although the cyclophilin binding loop is important for viral infection, it is not required for proper HIV CA protein folding and in vitro multimerization function [8]. Binding affinity of PF-3450074 to the crystallographic construct (Kd = 3.42 µM) is similar to that observed for both full length wild type CA (Kd = 2.79 µM) and isolated wild type NTD (Kd = 2.24 µM), as measured by isothermal titration calorimetry. Using this construct, the co-crystal structure of PF-3450074 was solved with the NTD of HIV-1 CA to 1.8 Å resolution. The structure of the complex showed that the overall fold of the CA protein is the same as previously described CA structures [19], [20], [21] and illustrates that neither the compound, PF-3450074, nor the loop deletion caused any significant shifts in the protein structure (Fig. 5a).
PF-3450074 occupies a preformed pocket in the HIV-1 CA NTD bounded by helices 3, 4, 5 and 7 (Fig. 5a). The R1 and R2 aromatic moieties of the compound occupy two hydrophobic sub-pockets and provide most of the key interactions which anchor the compound to the NTD (Fig. 5b). The indole substituent protrudes from the NTD R3 sub-pocket close to Lys70. The binding site for PF-3450074 is distinct from the sites targeted by CAP-1 and CAI/NYAD-1 [10], [11], [22], [23], [24]. These results confirm that PF-3450074 directly binds HIV CA and are consistent with this class of inhibitors acting via a unique mechanism. Coordinates are stored in the Protein Data Bank (PDB ID code 2XDE).
We describe a novel class of inhibitors that target HIV CA by a unique mechanism that interferes with both early and late events in the viral replication cycle. HIV CA plays an essential role in several stages of viral replication and is viewed as an important, yet unexploited target for therapeutic intervention [4], [5], [6]. This new series demonstrates that small molecules targeting HIV CA can have potent broad-spectrum antiviral activity. We demonstrate directly that HIV CA is the antiviral target of these inhibitors in infected cells by showing that mutations in HIV CA confer resistance to several members of the series. EM analysis shows that the series profoundly affects the morphology of nascent HIV particles. We demonstrate that the compounds affect CA protein multimerization in vitro and we have elucidated details of the novel compound binding site on HIV CA by solving a co-crystal structure of a compound from the series bound to the NTD of the protein. Our data strongly suggest that this new series of inhibitors targets HIV CA function during both the virion uncoating and viral core assembly processes.
Previous studies have described two molecules that target HIV-1 CA assembly in vitro, CAP-1 (a small molecule) and CAI (a dodecapeptide) [10], [11], [12]. CAP-1 acts in the late stage of viral replication and does not inhibit HIV-1 infection when added to pre-formed HIV-1 particles. A cell-permeable derivative of CAI, NYAD-1, inhibits formation of both immature and mature HIV-1 virus particles as well as early events in the replication cycle at low micromolar concentrations. The properties of the new class of CA inhibitors described in this study are clearly distinguished from those of other HIV-1 inhibitors, including previously described CA inhibitors. Unlike CAP-1, the small molecules described here inhibit both early and late events in the HIV replication cycle. In addition, PF-3450074 did not inhibit Gag particle production from HIV-1 transfected cells, suggesting that the compound series does not affect immature particle assembly. This is in contrast to the effects on immature particle assembly reported for NYAD-1. A co-crystal structure of a representative compound (PF-3450074) demonstrated a new binding site on HIV-1 CA distinct from those described for CAP-1 or CAI. Furthermore, PF-3450074 increased the rate of HIV-1 CA multimerization in vitro, while CAI and CAP-1 decreased the rate of CA multimerization in the same assay. While this does not necessarily define the action of these compounds on replicating virus, it does suggest a fundamentally different mechanism of inhibition from that of previously described CA inhibitors.
The proposed mechanism of action for both the early and late stage activities of this new class of inhibitors involves a direct effect on higher-order structures of HIV CA, in assembly and uncoating. Although the present data do not indicate whether the compounds enhance or inhibit the uncoating process, either effect is likely to interfere with proper reverse transcription [25]. HIV-1 capsid mutations proximal to the PF-3450074 binding pocket have been described that either destabilize or enhance the stability of viral cores and result in specific postentry defects in virus replication [24]. It is possible that such mutations and the compounds described in this study have analogous effects on inter-subunit capsid interactions. To gain further insights into the mode of action of PF-3450074, we generated a model of an assembled capsid hexamer in complex with PF-3450074 (Fig. 6a) based on superpositioning of published assembled capsid structures [13], [14] with the structure of the PF-3450074/CA complex. In the model, the R3 indole group which protrudes from the NTD in our structure localizes to the interface between capsid monomers in an assembled capsid and sits directly between the NTD of one capsid monomer and the C-terminal domain of another, making contacts to Tyr-169, Leu-172, Arg-173, Gln-179, and Lys-182 (Fig. 6b). This suggests the R3 indole group of PF-3450074 could play a critical role in modulating inter-subunit interactions. Both the CA NTD contact residues described by the co-crystal structure and these putative C-terminal contacts are well conserved across viral strains (Tables S5 and S6). This is consistent with the broad-spectrum antiviral activity observed for this series.
Although the sum of our results suggests a mechanism that affects interactions between capsid monomers, the early stage activity is consistent with other models. Cyclophilin A plays a critical role in the early stages of HIV-1 replication through interactions with the viral capsid [26]. Also, capsid-binding restriction factors such as the tripartite motif containing (TRIM) proteins prevent the infection of many primate cells with HIV or SIVs from other species [5], [9]. Thus, based on our data, we cannot dismiss the possibility that, during early infection, this new series might affect specific capsid-host protein interactions that mediate the viral uncoating process. A detailed study of the early stage activity of PF-3450074 has demonstrated direct destabilization of the HIV-1 capsid and a dependence on cyclophilin A, indicating that the compound induces premature uncoating of the virus, potentially through a mechanism similar to that of TRIM restriction [27].
In this study, we identify a new binding site on HIV-1 CA that can be targeted by small molecule inhibitors resulting in broad-spectrum antiviral activity. In addition, we describe the discovery and characterization of a novel series of compounds that act at this site and inhibit the virus at two points in the replication cycle. This series should serve as a good starting point for the development of a new class of HIV therapeutics through structure-based drug design or other approaches. The broad spectrum activity of this series is particularly exciting and highlights this novel mechanism as a significant therapeutic opportunity.
HeLa CD4 LTR/beta-Gal, MT-2, PM1, CEM-SS and HEK 293 cells as well as pNL4-3 HIV-1 infectious clone, HIV-1 IIIB, HIV-1 RF, HIV-1 BaL, and all primary isolates were obtained through the National Institutes of Health (NIH) AIDS Research and Reference Reagent Program, Bethesda, MD. JC53BL cells were sourced from Tranzyme. Efavirenz (EFV) was kindly provided by DuPont Merck (Wilmington, DE). Nelfinavir (NFV), 135137, PF-3450074, and PF-3759857 were synthesized by Pfizer Inc.
As described [18], host cells were infected with HIV-1 NL4-3, HIV-1 IIIB, HIV-1 RF. The cytopathic effect was measured using XTT reagent and the therapeutic index (TI) calculated by dividing the CC50 (mock infected cells) value by the EC50.
PHA-stimulated PBMC's were incubated for 1 hour with virus at an moi of 0.001–0.01. They were plated in 96-well plates at 5×105 cells/ml and incubated for 5 days at 37°C.10 µL of the supernatant was then transferred to a plate containing 40 µL of JC53BL cells at 0.5×105 cells/ml. After 2 days, the cells were processed for β-galactosidase activity using the FluorAce kit (Bio-Rad). PF-3794231 was tested against clinical isolates at the Southern Research Institute (Frederick, MD) as previously described [18].
The single-cycle infectious HIV-1 reporter viruses were generated as previously described [28] by co-transfecting HEK293 cells with an HIV-1 NL4-3 single-cycle infectious cDNA (pNL4-3deltaEnv) and an NL4-3 or VSV-G envelope expression vector. Half-log dilutions of test compounds were added to HeLa CD4 LTR/beta-Gal target cells, seeded in 96-well plates at a cell density of 1×104 cells per well in DMEM containing 10% FBS. Compound-treated or compound-free target cells were then infected with the HIV-1 single-cycle infectious virus and after 72 hours measured for the induction of beta-galactosidase Data were expressed as the percent of reporter gene activity in infected compound-treated cells relative to that of infected, compound-free cells.
Envelope-deleted NL4-3 cDNA (pNL4-3deltaEnv) was co-transfected into HEK 293 cells with an HIV envelope expression vector as previously described [29]. Half-log dilutions of test compounds were then added to transfected cell cultures 3 hrs after transfection. The supernatants were harvested 72 hrs after transfection and infectious virus production was subsequently quantified by measuring the induction of the beta-galactosidase reporter gene after 100-fold dilution into fresh medium and incubation in the presence of HeLa CD4 LTR/beta-Gal target cells for 72 hours.
Viral variants resistant to PF-1385801 or PF-3759857 were selected as described previously [18]. To construct NL4-3 recombinant virus containing the Q67H, K70R, Q87P, T107N and L111I amino acid substitutions identified in the serial passage studies, viral cDNAs that had been TOPO cloned for sequence analysis were digested with BssHII and ApaI and ligated back into a wild type pNL4-3 background. The effect of amino acid substitutions in HIV-1 Gag sequences on PF-3759857 susceptibility was measured in HIV Replication assays as described previously [15].
Proteins were expressed in E. Coli BL21(DE3) in 2YT containing kanamycin. Cells were harvested following induction with IPTG and growth overnight at 20°C. CA1-231 was purified as described [30]. For CA1-146/Δ87-99G, cells were resuspended and lysed in buffer A (50 mM TrisHCl pH 7.4, 150 mM NaCl, 10 mM imidazole, and 2 mM β-mercaptoethanol). The lysate was clarified by centrifugation, filtered then applied to a His trap nickel affinity column (Sigma), eluting with buffer A supplemented with 300 mM imidazole. The eluted material was further purified on a Sephacryl 100HR column in 50 mM TrisHCl pH 7.4, 150 mM NaCl and 2 mM β-mercaptoethanol.
Multimerisation assays were performed as previously described [10]. Compound was added to 30 µg of full-length CA protein in 50 mM sodium phosphate buffer, pH 8.0 in a volume of 30 µl. Capsid assembly was initiated by addition of a concentrated NaCl solution (50 µl 5 M NaCl in 50 mM sodium phosphate, pH 8.0). Optical density was monitored on a Molecular Devices SpectraMax spectrophotometer at 350 nm every 20 s for 1 h.
PBMCs were incubated with 7 µM PF-3450074 (approx. 10× IC50), for 30 minutes before being infected with mock virus (medium alone), or with NL4-3 at an MOI of 1. At 72 hours post-infection the supernatants were removed and the cells were fixed for 60 min in MacDowell's fixative (4% (v/v) paraformaldehyde, 1% glutaraldehyde in 0.1M phosphate Milliong buffer, pH 7.3) at 4 C. The cells were rinsed in 0.15M Phosphate Sörensen buffer with 0.2% (v/v) NaCl (pH 7.3) and suspended in warm 3.5% agar in water. The agar blocks were cooled to 4°C and washed three times for 10 minutes in 0.15M Phosphate Sörensen buffer with 0.2% (v/v) NaCl (pH 7.3). The agar blocks with the cells were stained using 2% Osmium tetroxide in 0.3M phosphate Sörensen buffer (pH 7.4) for 1 h at 4°C. The blocks were washed 4 times for 10 minutes in sterilized water at room temperature before they were dehydrated in graded ethanol (50%, 70%, 90%, 100%, 100%) for 15 minutes at each concentration at room temperature. The blocks were treated with propylene oxide three times for 15 minutes in room temperature before the resin was infiltrated for 1 hour at room temperature with a 1∶1 propylene oxide∶Epon mix (Epon contains epoxy embedding medium 20 mL (Epon 812 resin)) and 12 ml MNA (methyl nadic anhydride) and 9 ml DDSA (dodecenyl succinic anhydride)) before the resin was treated in epon overnight at room temperature. The resin was then embedded with Epon and BDMA 3% for 1 hour at room temperature. The resin was polymerised for 48h at 58–60°C. Approximately 1 micron thick semi-thin sections were cut and stained in toluidine blue and observed in light microscopy. Subsequently, ultra-thin sections of 70–90 nm were cut on an Ultracut E Reichert microtone, and stained with uranyl acetate and lead citrate. Observations were made using a Jeol 1200 EX II electron microscope.
Isothermal titration calorimetry (ITC) was performed using the VP-ITC calorimetric system (GE Healthcare). Protein solutions for ITC were dialyzed against buffer A (50 mM TrisHCl pH 7.5/150 mM NaCl/2 mM β-mercaptoethanol). The dialyzed protein solution (15 µM) in the calorimetric cell (1.4274 ml) was titrated at 25°C with ligand (200 µM) in the buffer A using 1×2 µL, followed by 25×10 µL injections. Heat evolved was obtained from the integral of the calorimetric signal and heat of dilution was negligible in titrations of the ligand into buffer only. Analysis was carried out with Origin 5.0 software (GE Healthcare). Binding parameters such as the number of binding sites (n), the binding constant (Ka, M−1), and the binding enthalpy (ΔHa, kcal/mol of bound ligand) were determined by fitting the experimental binding isotherms.
Protein was concentrated to 30 mg/ml and inhibitor added to 5 mM from a 100 mM DMSO stock solution. After standing for 2 hours, hanging drop crystallizations were set up consisting of 2 µl of protein and 1 µl of well solution (20% PEG 8000, 100 mM phosphate-citrate pH 4.2 and 200 mM sodium chloride). Crystals grew overnight at room temperature and were frozen for X-ray data collection following addition of 60% NDSB containing 40% ethylene glycol (4 µl). Data was collected at the ESRF (ID14-4) on an ADSC Q315 detector, integrated with mosflm [31] and scaled with CCP4 package SCALA [32]. The structure was solved by molecular replacement using MOLREP [32] with a truncated model of the N-terminal CA. The structure was refitted using QUANTA version 2000.1 (Accelrys Inc., San Diego, CA) and refinement was carried out using REFMAC [33]. Data collection and refinement statistics are shown in Table S7. Coordinates are stored in the Protein Data Bank (PDB ID code 2XDE).
|
10.1371/journal.ppat.0030077 | Epigenomic Modifications Predict Active Promoters and Gene Structure in Toxoplasma gondii | Mechanisms of gene regulation are poorly understood in Apicomplexa, a phylum that encompasses deadly human pathogens like Plasmodium and Toxoplasma. Initial studies suggest that epigenetic phenomena, including histone modifications and chromatin remodeling, have a profound effect upon gene expression and expression of virulence traits. Using the model organism Toxoplasma gondii, we characterized the epigenetic organization and transcription patterns of a contiguous 1% of the T. gondii genome using custom oligonucleotide microarrays. We show that methylation and acetylation of histones H3 and H4 are landmarks of active promoters in T. gondii that allow us to deduce the position and directionality of gene promoters with >95% accuracy. These histone methylation and acetylation “activation” marks are strongly associated with gene expression. We also demonstrate that the pattern of histone H3 arginine methylation distinguishes certain promoters, illustrating the complexity of the histone modification machinery in Toxoplasma. By integrating epigenetic data, gene prediction analysis, and gene expression data from the tachyzoite stage, we illustrate feasibility of creating an epigenomic map of T. gondii tachyzoite gene expression. Further, we illustrate the utility of the epigenomic map to empirically and biologically annotate the genome and show that this approach enables identification of previously unknown genes. Thus, our epigenomics approach provides novel insights into regulation of gene expression in the Apicomplexa. In addition, with its compact genome, genetic tractability, and discrete life cycle stages, T. gondii provides an important new model to study the evolutionarily conserved components of the histone code.
| Apicomplexan parasites, including Toxoplasma gondii, are responsible for a variety of deadly infections, but little is understood about how these important pathogens regulate gene expression. Initial studies suggest that alterations in chromatin structure regulate expression of virulence traits. To understand the relationship of chromatin remodeling and transcriptional regulation in T. gondii, we characterized the histone modifications and gene expression of a contiguous 1% of the T. gondii genome using custom DNA oligonucleotide microarrays. We found that active promoters have a characteristic pattern of histone modifications that correlates strongly with active gene expression in tachyzoites. These data, integrated with prior gene predictions, enable more accurate annotation of the genome and discovery of new genes. Further, these studies illustrate the power of an integrated epigenomic approach to illuminate the role of the “histone code” in regulation of gene expression in the Apicomplexa.
| Toxoplasma gondii is an obligate intracellular apicomplexan parasite responsible for encephalitis in immunocompromised individuals and birth defects when a fetus is exposed in utero [1,2]. The life cycle of T. gondii is complex, with multiple differentiation steps that are critical to survival of the parasite in its human and feline hosts [3]. The genetic tractability of T. gondii has caused it to emerge as a model for the study of apicomplexan parasites [3], and the recent sequencing of the T. gondii genome (http://www.toxodb.org) is adding to our appreciation of the unusual nature of apicomplexan genomes [4,5].
A remarkable finding is the relative paucity of genes encoding proteins with motifs that indicate transcription factor function in apicomplexan genomes [6,7]. This has led to the proposal that gene regulation in apicomplexan parasites is controlled mainly via RNA stability [6], despite the tightly regulated patterns of gene expression observed in different stages of the life cycle of T. gondii [8] and Plasmodium falciparum [9]. However, that certain DNA motifs are recurrent in the promoters of these organisms and bind to nuclear factors [10−14] suggests that unrecognized transcription factors may exist, but are not encoded by genes with recognizable structural features. On the other hand, the RNA polymerase II machinery [7,15] and genes with motifs indicating potential chromatin remodeling and modification functions [6,16] are conserved within the Apicomplexa. Epigenetic processes have significant clinical relevance in light of studies that implicate the histone deacetylase Sir2 homolog in regulation of antigenic variation in P. falciparum [17,18].
To obtain a genome-wide view of gene expression in T. gondii tachyzoites, we examined the epigenetic organization and transcription patterns of a contiguous 1% of the T. gondii genome using custom microarrays. Histone modifications—including acetylation of histone H4 (H4ac), acetylation of lysine 9 (H3K9ac), and trimethylation of lysine 4 of histone H3 (H3K4me3)—have been identified at certain individual active loci in T. gondii [19], suggesting a role in gene expression. We hybridized the tiled genomic microarrays with material derived from chromatin immunoprecipitations using antibodies to modified histones. By simultaneously hybridizing the microarray to tachyzoite-derived cDNA, we tested the genome-wide association of specific histone modifications with gene expression.
We generated a custom oligonucleotide microarray containing 12,995 50-mer features tiling a 650-kb region of Chromosome 1b, with an average resolution of one oligonucleotide every 50 bp (Figure 1). Chromosome 1b of the RH strain of the 63-Mb T. gondii genome has been extensively annotated and has a single nucleotide polymorphism frequency comparable with the rest of the genome, an average of 5.7 exons per coding sequence (CDS), and a gene density of one gene per 7.4 kb [20]. Currently, 91 genes are predicted within the 650-kb region of the RH strain of T. gondii.
The amino acid sequences of the tails of eukaryotic histones H3 and H4 are strongly conserved (Figure S1), allowing us to use a panel of commercial antibodies for chromatin immunoprecipitation (ChIP) in T. gondii. After screening antibodies to modified histones for T. gondii nuclear localization (Table 1; Figure S2), we performed ChIP using DNA isolated from the intracellular tachyzoite stage of T. gondii. As a control, we used an antibody against a T. gondii kinase with no DNA-binding potential (M. Gissot and K. Kim, unpublished data). The immunoprecipitated DNA was amplified, tested to ensure enrichment for control loci was maintained, and co-hybridized to the 650-kb tiling array with input DNA.
We studied the distribution of three modified histones (H4ac, H3K9ac, and H3K4me3) previously described as activation marks in other eukaryotes [21,22]. The ChIP material applied to the microarray (ChIP-chip) generated strong focal peaks of enrichment for the three different modified histones. Signal was readily discriminated from noise, even looking at the raw ChIP/input DNA ratio (Figure 1), a finding confirmed by the p-values derived from the use of the ChIPOTle analytical approach [23], which approached zero for each locus. We observed 52 clear, discrete, and coincident H4ac, H3K9ac, and H3K4me3 peaks within the 650-kb region tiled on the microarray (Figure 1B; Table S1). The H3K9ac and H4ac peaks have a median size of 1,550 bp, whereas the H3K4me3 peaks are relatively smaller (median size of 1,300 bp; Table 1).
As previously observed for other eukaryotes [21,22], these modifications co-localize and associate to form a complex pattern at focused loci in the T. gondii genome (Figure 1). More than 96% of the H3K4me3, H3K9ac, and H4ac (Figure 1B) peaks are placed in the predicted intergenic regions. Moreover, the identified peaks for the three modifications are located close to the 5′ of predicted genes. Indeed, the distance between the identified peak and the start codon of the closest gene is less than 1,000 bp for more than 85% (45/52) of the H3K9ac and H4ac peaks (Table 1), and less than 1,500 bp for more than 90% (48/52). Similarly, for more than 90% (49/52) of H3K4me3 peaks, the end of the peak and the first predicted initiation codon was within 1,000 bp (Table 1).
We also performed ChIP using antibodies against histone H3 dimethylated at arginine 17 (H3R17me2), another putative general activation mark in T. gondii [19] and other eukaryotes [24,25]. Recently, it was suggested that this histone modification is present at all active promoters in T. gondii [19] based upon PCR examination of selected promoters after ChIP with anti-H3R17me2. Using the same antibody, we show that this modification is restricted to a subset of promoters (Figure 1B). This histone mark overlapped with only four of the 52 modified histone peaks identified (4.5% of the genes present on the microarray). All four genes have expressed sequence tags (ESTs) for both tachyzoite and bradyzoite stages.
The H3K4me1 and H3K4me2 marks were also investigated using ChIP-chip and were not specifically enriched, as determined by analysis of hybridizations by the ChIPOTLe software (unpublished data). We also verified that the modified histone peaks identified were not due to local core histone enrichment by performing ChIP-chip with an antibody specific to the C-terminus histone H3.
The ChIP-chip results were validated using quantitative single-locus PCR (Figure 2 is representative of eight loci validated). Using real-time quantitative PCR on ChIP samples, we amplified regions within the predicted gene (primer set 1) or in intergenic sequences (primer sets 2−6). We found enrichment of the three modified histones in regions identified as peaks in the ChIP-chip experiments (primer sets 2, 3, 5, and 6). In contrast, the three activation marks tested were not significantly enriched in a region located within the predicted gene (primer set 1) and a region between the two identified peaks (primer set 4). We also verified that there was no significant local enrichment of the core histone H3 (positive PCR for all primer sets).
To verify the link between the modified histone peaks and transcription of the nearest predicted gene, we hybridized cDNA made from intracellular tachyzoites to the tiled microarray (Figure 3). Using three different analytical approaches, we identified regions on the tiled portion of the genome with significant gene expression (Table S1). Overall, 51 of the 52 regions with a cluster of H3K9ac, H4ac, and H3K4me3 peaks had a significant cDNA hybridization signal adjacent to them. These data were consistent with EST studies, with 46 of the 49 genes represented by at least one EST for the tachyzoite stage expressed in our dataset. In our study, 31% (21/67) of the genes expressed were not represented by any EST data, demonstrating the limits of the EST mapping approach for identifying expressed genes.
Two transcribed loci did not correspond to a predicted gene (Table S1). One locus had associated H3K9ac, H4ac, and H3K4me3 peaks characteristic of active chromatin and corresponded to a transcription unit represented by two overlapping tachyzoite ESTs (CN197705 and CK737836). A partial open reading frame (ORF) was discovered after alignment of those ESTs. After comparing this sequence with the nr database (http://www.ncbi.nlm.nih.gov/BLAST), we found the ORF had homology with the cytochrome oxidase subunit III (COX3) gene of Plasmodium (highest p-value = 5e-06). This gene is not annotated in the current version of the T. gondii genome (http://www.toxodb.org). The other locus was also represented by an EST (BG659482) and appears to be driven by a promoter that displays promoter activity in both directions (Table 2). However, this transcribed locus does not have an ORF and appears to represent a non-coding RNA.
We also found two regions of clustered genes with stage-specific expression based on EST data. One region predicts a set of five tandemly arrayed kinases with ESTs primarily from the oocyst stage (Figure 3A). Another region is characterized by five genes predicted as BSR4 homologues with ESTs primarily from the bradyzoite stage (Figure 3B). No significant expression during the tachyzoite stage could be detected, and neither region had any of the three histone modification peaks characteristic of active chromatin. For the bradyzoite-specific locus, two ESTs were recovered from a Type III strain (VEG) tachyzoite cDNA library. These ESTs could reflect differences in gene expression between strains or represent the low level of bradyzoite forms frequently present in Type II and Type III tachyzoite cultures.
To test that the clustered peaks were located at active promoters, we performed luciferase reporter assays (Table 2; Figure 4A and 4B). We cloned regions spanned by the H3K9ac and H4ac peaks and tested their ability to drive the expression of the luciferase in transient transfection assays. Of the 12 loci tested, 11 were able to drive expression of luciferase (Table 2; Figures 4 and S3).
Regions 5′ of non-expressed genes that lacked clustered peaks of modified histones or regions spanned by a predicted ORF are not able to drive the expression of luciferase (Table 2; Figure S3). However, the two loci with overlapping H3K4me3, H3K9ac, and H4ac peaks located within rather than 5′ to annotated gene coding regions were both able to drive the expression of a reporter gene (Table 2; Figure 4B).
Of the 52 activation peaks identified, only one lacked evidence of mRNA expression in its vicinity. This peak is located 5′ to a predicted gene (Tg1b.2420), a locus with the characteristics of a DNA-repair protein, but not associated with any EST in the T. gondii database at any stage of the life cycle. The promoter of this gene yielded a background activity as low as the untransfected parasites (Table 2; Figure S3).
H3K4me3 peak distribution is consistently shifted toward the 5′ end of genes in comparison with H3K9ac (unpublished data) and H4ac peaks (Figure 4C and 4D). PCR studies confirmed that the shift of the H3K4me3 peak predicts the orientation of transcription. (Four genes were tested with two represented in Figure 2.) As predicted by these data, most of the sequences tested have directional activity, as would be expected for genuine promoters (Table 2; Figure 4C and 4D). However, seven of the 52 peaks are located in regions where two genes are transcribed in opposite directions, providing biological evidence for sequences in T. gondii that have promoter activity in both directions as shown in other Apicomplexa [26].
As illustrated, H3K4me3, H3K9ac, and H4ac peaks identify promoters. We also found seven predicted genes (as defined in [20]) that were expressed but lacked modified histone peaks at their predicted promoter. In all such cases, these genes are preceded within 1,000 bp by an expressed gene that is transcribed in the same direction and bears histone activation peaks at its 5′ end. These genes likely represent gene prediction errors, since RT-PCR in two cases confirmed a single transcription unit with the adjacent gene (Figure S4).
We have employed an integrative approach to epigenomics, combining simultaneous analysis of ChIP-on-chip and gene expression on a tiling array encompassing a 0.65-Mb contiguous portion of the T. gondii Chromosome 1b. The H3K9ac, H4ac, and H3K4me3 modifications co-localize at focused loci in the T. gondii genome and correlate with significant gene expression. We confirmed that the enrichment observed was not due to local enrichment of the H3 core histone by performing ChIP with an antibody directed against the C-terminus of the histone H3. In contrast, in T. gondii, the H3K4me1 and the H3K4me2 modifications are present at equal amounts in active and inactive chromatin as previously shown for human promoters [22] and in contrast to Saccharomyces cerevisiae [23].
To our knowledge, this study is the first to explore the distribution of the H3R17me2 modification on a genomic scale. Surprisingly, this modification is enriched only at a subset of active promoters. Thus, T. gondii uses its histone modification machinery not only as a general landmark of activated promoters but also to specifically attribute a distinctive mark to certain promoters. ESTs have been sequenced from both tachyzoite and bradyzoite stages for those four genes, whereas only 26 of the 91 predicted genes on our chip (28%) have ESTs in both tachyzoite and bradyzoite stages. The H3R17me2 mark may have significance during the tachyzoite to bradyzoite differentiation process, but the number of loci discovered in this study are too limited to speculate further upon the specificity conferred by this trait. The recent discovery of the importance of arginine methylation during early development of mouse embryo indicates a specific role for the H3R17me2 during differentiation [27].
The H3K9ac and H4ac peaks in T. gondii are larger than those previously observed in human (approximately 700 nucleotides [nt]) [28] but similar in size to those found in yeast [23]. It appears that the number of modified nucleosomes is in the same range for these three organisms despite their difference in genome compaction. Such similarity in the size of the peaks may have functional implications for RNA polymerase II.
The placement of the three “gene activation” modifications coincides, but H3K4me3 peaks are shifted toward the 5′ end of expressed genes. This difference has been documented in human cell lines [29] and predicts the directionality of promoters in T. gondii. Although most promoters appear to be orientation-specific, the tiled region of the T. gondii genome encodes several regions that exhibit promoter activity in both directions. Further mapping studies are needed to determine whether these are true bi-directional promoters or two separate promoters facing in opposite directions.
We observed an exceptional correlation between gene expression and the presence of co-localized modified histone peaks. The few discrepancies between the EST database and our gene expression data are likely due to differences in gene expression between the strain we used (RH, Type I) and the strains used to generate “tachyzoite” cDNA libraries. Type II and III tachyzoite cultures, in contrast to Type I strains, frequently have a low level of basal bradyzoite forms.
One region represented on our array had a cluster of H3K9ac, H4ac, and H3K4me3 peaks but was unable to drive luciferase expression. Interestingly, these peaks are located 5′ to a gene (Tg1b.2420) predicted to encode a protein similar to DNA-repair protein XRCC3, a protein essential for ultraviolet radiation–induced double-strand break repair from bacteria to mammals [30]. Expression of this gene was not detected by reverse transcriptase−PCR (RT-PCR; Table S1) and there were no associated ESTs in the T. gondii database at any stage of the life cycle, which could be explained by rapid processing or degradation of the mRNA for this gene. Alternatively, the promoter could be in a poised state waiting for activation or for the release of a repression, as observed in a study of rapidly induced genes in human T cells [31]. As suggested for T cells, the activation marks associated with this promoter could signify the presence of epigenetic memory in T. gondii. In a study of human promoters, 20% of those genes with overlapping H3ac and H3K4me3 marks lacked evidence of mRNA expression [21].
Prior microarray gene expression studies in T. gondii have been based upon cDNAs [32] rather than tiled genomic microarrays. Our survey of tachyzoite gene expression for this contiguous 1% of the T. gondii genome enabled us to identify new tachyzoite-expressed genes and discover transcripts in regions where genes have not been predicted. For example, a cluster of modified histone “activation” peaks helped us to identify a gene coding for a cytochrome oxidase subunit III, which is not annotated in the current version of the T. gondii genome, and a possible non-coding RNA. Moreover, our study illustrates the power of empirical annotation of the genome in terms of promoters and their transcriptional orientation, enhancing gene prediction approaches beyond what is currently possible using DNA sequence-based approaches alone.
In conclusion, we have performed the first mapping to our knowledge of the epigenome of an apicomplexan parasite. Taken together, the data indicate that T. gondii uses a multipart histone modification system to assign a functional role to certain DNA sequences and underscores the ability of this unicellular apicomplexan parasite to employ a complex set of tools to control its gene expression. These data are consistent with the extensive repertoire of proteins predicted to modify histones in the T. gondii genome [16]. Moreover, our study illustrates the power of empirical annotation of the genome in terms of promoters and their transcriptional orientation, enhancing gene prediction approaches beyond what is currently possible using DNA sequence-based approaches alone.
T. gondii is a medically important pathogen and is genetically tractable. It is a powerful model for studying the gene regulation of apicomplexan parasites and may now represent a new model system for understanding evolutionarily conserved components of the “histone code.” Further, epigenetic regulators may represent potential therapeutic targets and provide new tools to fight toxoplasmosis and other parasitic diseases like malaria.
T. gondii RH strain was maintained in confluent monolayers of human foreskin fibroblasts (HFF). Parasites were harvested 24 h after invasion of HFF cells and purified as previously described in [33].
ChIP was performed as described [34] with slight modifications. Briefly, chromatin from intracellular tachyzoites grown in HFF for 24 h was cross-linked for 10 min with 1% formaldehyde at room temperature and purified after a sonication step yielding fragments of 500−1,000 bp. Immunoprecipitations were performed with the appropriate rabbit serum (Table 1) at 4 °C overnight and washed extensively as published previously [34]. DNA was further subjected to a treatment with proteinase K for 2 h and then purified using the Qiagen PCR purification kit (http://www.qiagen.com). As a negative control, we used rabbit antiserum to PKA2, a kinase that is not present in the nucleus (M. Gissot and K. Kim, unpublished data).
We generated a tiled array of 50-bp oligonucleotides with 12,295 oligos encompassing 650,000 bp (1,230,000−1,880,000) of Chromosome 1b [20] with a spacing of 50 bp between each oligonucleotide. The array was fabricated in the NimbleGen Systems (http://www.nimblegen.com) 12-plex format, which allows simultaneous hybridization of 12 identical arrays on a single slide. Amplification of immunoprecipitated DNA and 100-ng input DNA was performed using the ligation-mediated PCR technique [35]. After amplification, the immunoprecipitated DNA was tested for enrichment of control loci by qPCR and co-hybridized to the 650-kb tiling array with input DNA. DNA was labeled using random primers coupled to a fluorochrome and hybridized according to NimbleGen Systems procedures. At least two biological replicates were performed.
Real-time quantitative PCR was performed on the 7300 ABI apparatus using the Power Sybr (ABI, http://www.appliedbiosystems.com) mastermix in a 20-μL volume according to the manufacturer's instructions. PCR primers were designed using the Primerexpress software (ABI) to amplify regions of 100−150 nt. A 10-fold dilution of input was compared with 0.5 ng of immunoprecipitated DNA. Each experiment was performed at least three times in duplicate.
The RNA from three replicate flasks containing RH strain–infected HFFs and one control flask containing only HFFs was purified using TRIzol. RNA integrity was verified on the Agilent Bioanalyzer (http://www.agilent.com). Ten micrograms of total RNA was retrotranscribed using the BD Sprint Powerscript kit (http://www.bdbiosciences.com) and random hexamers and made double-stranded cDNA (dscDNA) using Escherichia coli polymerase I. dscDNA labeling with fluorochrome-coupled random hexamer and hybridization to the array was performed following NimbleGen protocols. NimbleGen scanning and spot finding software were used.
Significant peaks for ChIP-on-chip were identified with the ChIPOTle software [23] using a permutation simulation to estimate the background distribution (with a window size of 500 bp, 300 permutations, and a p-value of 0.001). Peaks with a p-value of less than 10−10 (which produces about 50 times more significant regions than false regions) and with a peak height cut-off of 2 were considered significant. The false discovery rate was 0.1%.
After background correction using random probes, gene expression was calculated as the average of the log2 ratio of the intensity given by the HFF plus parasite dscDNA to the intensity given by the HFF-alone dscDNA. With ChIPOTle, expression was considered significant with a p-value < 0.05 and a high average ratio above 1 or 0.6. Peaks of significant expression were also identified using the detection peaks tool in SignalMap software with a sliding window of 150 bp and a log2 ratio threshold of 1 or 0.8. A peak is identified when there are at least four data points within a window above the threshold value. The height of the peak is the maximum of the data points within the window. In addition, the raw log2 ratios were normalized using loess regression to remove the dependence of the variance on the mean and partitioned into segments along the chromosome with the function segmentation within the Bioconductor package “tilingArray” (http://www.bioconductor.org) [36], using 300 and 3,000 for the maxseg and maxk arguments, respectively. Since a one-to-one correspondence between the segments and the gene annotations does not exist (e.g., when several adjacent genes are not transcribed), tests of significance were carried out using the means of the probes that mapped fully to each annotated gene. The intensity threshold between the untranscribed and transcribed segments was determined by fitting a mixture model to the segment means using the “mclust” package from Bioconductor [36]. The significance of expression for each annotation was calculated using the binomial test on the signs of the differences between the probe intensities and the threshold [37]. The p-values were adjusted for multiplicity using the Benjamini–Yekutieli procedure from the “multtest” package of Bioconductor [36] with a false discovery rate of 0.1%.
Regions of T. gondii DNA (Tables 2 and S1) were subcloned into pCR8-GW vector (Invitrogen, http://www.invitrogen.com), sequenced, and cloned into a Gateway vector expressing Firefly luciferase. Plasmid (50 μg) was co-transfected with 20 μg plasmid expressing Renilla luciferase under the control of the Tubulin promoter (both plasmids gift of M. W. White, Montana State University) following standard transfection protocols [38]. Luciferase assay was performed after 24 h using the Promega Dual-Luciferase Kit (http://www.promega.com) according to manufacturer instructions. Each assay was repeated three times in duplicate.
Gene predictions were as described in Khan et al. [20]. The sequences corresponding to the CDS were extracted for a 650-kb region of the RH strain Chromosome 1b [20] and were set up as a BLAST database using the BLAST program downloaded from NCBI (http://www.ncbi.nlm.nih.gov/BLAST). We then used a perl script to blast the 88,535 EST sequences downloaded from the ToxoDB Web site (http://www.toxodb.org/download/release-3.3/EST/nuc) against the BLAST database. The e-value cut off of e-25 was considered significant.
The European Bioinformatics Institute (http://www.ebi.ac.uk) accession numbers of genes and proteins used in this study are TgIb.1560c; TgIb.1570; TgIb.1580c; TgIb.1590; TgIb.1600; TgIb.1610c; TgIb.1620; TgIb.1630c; TgIb.1640; TgIb.1650c; TgIb.1660; TgIb.1670; TgIb.1680c; TgIb.1690c; TgIb.1700c; TgIb.1710; TgIb.1720c; TgIb.1730c; TgIb.1740; TgIb.1750c; TgIb.1760c; TgIb.1770c; TgIb.1780; TgIb.1790c; TgIb.1800c; TgIb.1810c; TgIb.1820; TgIb.1830; TgIb.1840; TgIb.1850c; TgIb.1860; TgIb.1870c; TgIb.1880; TgIb.1890; TgIb.1900c; TgIb.1910; TgIb.1920; TgIb.1930c; TgIb.1940; TgIb.1950c; TgIb.1960c; TgIb.1970; TgIb.1980c; TgIb.1990; TgIb.2000; TgIb.2010c; TgIb.2020; TgIb.2030; TgIb.2040; TgIb.2050; TgIb.2060; TgIb.2070c; TgIb.2071; TgIb.2080; TgIb.2090; TgIb.2100c; TgIb.2110; TgIb.2120c; TgIb.2130c; TgIb.2140c; TgIb.2150c; TgIb.2160c; TgIb.2170c; TgIb.2180c; TgIb.2190; TgIb.2200; TgIb.2210c; TgIb.2220c; TgIb.2230; TgIb.2240; TgIb.2250; TgIb.2260; TgIb.2270; TgIb.2280c; TgIb.2290; TgIb.2291; TgIb.2300; TgIb.2310; TgIb.2320c; TgIb.2330; TgIb.2340; TgIb.2350c; TgIb.2360c; TgIb.2370c; TgIb.2380c; TgIb.2390c; TgIb.2400; TgIb.2410; TgIb.2420; TgIb.2430; and TgIb.2440c.
GenBank dbEST (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=Nucleotide) accession numbers of ESTs used in this study are BG659482, CN197705, and CK737836.
Microarray data have been submitted to the Gene Expression Omnibus database (http://www.ncbi.nlm.nih.gov/projects/geo) under accession numbers GSM139203–GSM139216 and GSM139134–GSM139136; the number for the complete series is GSE7262. |
10.1371/journal.pgen.1003962 | Translation Initiation Factors eIF3 and HCR1 Control Translation Termination and Stop Codon Read-Through in Yeast Cells | Translation is divided into initiation, elongation, termination and ribosome recycling. Earlier work implicated several eukaryotic initiation factors (eIFs) in ribosomal recycling in vitro. Here, we uncover roles for HCR1 and eIF3 in translation termination in vivo. A substantial proportion of eIF3, HCR1 and eukaryotic release factor 3 (eRF3) but not eIF5 (a well-defined “initiation-specific” binding partner of eIF3) specifically co-sediments with 80S couples isolated from RNase-treated heavy polysomes in an eRF1-dependent manner, indicating the presence of eIF3 and HCR1 on terminating ribosomes. eIF3 and HCR1 also occur in ribosome- and RNA-free complexes with both eRFs and the recycling factor ABCE1/RLI1. Several eIF3 mutations reduce rates of stop codon read-through and genetically interact with mutant eRFs. In contrast, a slow growing deletion of hcr1 increases read-through and accumulates eRF3 in heavy polysomes in a manner suppressible by overexpressed ABCE1/RLI1. Based on these and other findings we propose that upon stop codon recognition, HCR1 promotes eRF3·GDP ejection from the post-termination complexes to allow binding of its interacting partner ABCE1/RLI1. Furthermore, the fact that high dosage of ABCE1/RLI1 fully suppresses the slow growth phenotype of hcr1Δ as well as its termination but not initiation defects implies that the termination function of HCR1 is more critical for optimal proliferation than its function in translation initiation. Based on these and other observations we suggest that the assignment of HCR1 as a bona fide eIF3 subunit should be reconsidered. Together our work characterizes novel roles of eIF3 and HCR1 in stop codon recognition, defining a communication bridge between the initiation and termination/recycling phases of translation.
| Protein synthesis (translation) utilizes genetic information carriers, mRNAs, as templates for the production of proteins of various cellular functions. Typically it is divided into four phases: initiation, elongation, termination and ribosomal recycling. In this article we argue that the strict mechanistic separation of translation into its individual phases should be reconsidered in the light of “multitasking” of initiation factors eIF3, HCR1 and ABCE1/RLI1. In detail, we show that eIF3 and HCR1 not only promote the initiation phase but also specifically act at the other end of the translational cycle during termination. We present genetic and biochemical data linking eIF3 and HCR1 with both eukaryotic release factors (eRF1 and eRF3) and the ribosomal recycling factor ABCE1/RLI1, and propose a model for how all these factors co-operate with each other to ensure stringent selection of the stop codon. Collectively, our findings suggest that changes in one phase of translation are promptly communicated to and coordinated with changes in the other phases to maintain cellular homeostasis of all ongoing processes.
| Protein synthesis or mRNA translation is a complex and highly conserved process that can be separated into initiation, elongation, termination and ribosome recycling phases. Although these four phases are distinct in time, there is a longstanding notion for some form of communication among them. Notably, several initiation factors and related proteins have been proposed to function in more than one phase. These include ABCE1/RLI1 and GLE1, which are believed to promote both the initiation and termination phases by a mechanism that remains to be elucidated [1]–[3], eIF5A proposed to stimulate all three major phases [4], and the bona fide translation initiation factor eIF3, which has been recently suggested to promote the recycling phase, at least in a mammalian in vitro reconstituted system [5], [6].
The beginning of a translational cycle involves a series of steps that culminate in the assembly of the 80S initiation complex (IC) on the AUG start codon (reviewed in [7]). These steps include 1) Met-tRNAiMet recruitment to the 40S subunit to form the 43S pre-initiation complex (PIC), 2) mRNA recruitment to the 43S PIC to form the 48S PIC, 3) scanning of the 48S PIC to the first recognized start codon, and 4) joining of the 60 subunit to commit the resulting 80S IC to the elongation phase. The translation initiation factor eIF3, which in yeast consists of five essential core subunits (eIF3a/TIF32, b/PRT1, c/NIP1, i/TIF34 and g/TIF35) and one transiently associated, non-essential subunit (eIF3j/HCR1), is actively involved in regulation of the first three of these steps [7]. In the PIC assembly steps, the action of eIF3 is further stimulated by one of its interacting partners, the ATP-binding cassette protein ABCE1/RLI1, by an unknown mechanism [1]. In contrast to the most of eIFs, eIF3 interacts with the solvent-exposed side of the small ribosomal subunit [7] and as such it was proposed to be able to interact with active 80S ribosomes post-initiation [8]–[10].
The end of a translational cycle involves another series of steps that culminate in the release of a newly synthesized polypeptide from the translating ribosome (the termination phase), and in the dissolution of the ribosome:tRNA:mRNA complex (the recycling phase). Termination begins when a stop codon enters the ribosomal A-site, forming a pre-termination complex (pre-TC) [11]. In eukaryotes, all three stop codons are decoded by the eukaryotic release factor 1 (eRF1). According to recent models [12], [13], eRF1 enters the ribosomal A-site in complex with a second release factor, eRF3, in its GTP bound form. Recognition of a stop codon triggers GTPase activity of eRF3, which leads to its dissociation from the complex in its GDP bound form. eRF1 is then free to activate the ribosomal peptidyl transferase centre (PTC), which hydrolyses the bond between the P-site tRNA and the nascent polypeptide. Importantly, these steps are promoted by RLI1 in an ATP-independent manner; i.e. by the same factor that also somehow stimulates the eIF3 function in the initiation phase. Molecular details of this RLI1 role in termination are similarly not known, nevertheless, the proposed active role of RLI1 in stop codon recognition is consistent with observations that conditional down regulation of RLI1 protein levels increases stop codon read-through in yeast [2]. Based on the most recent structural model, RLI1 binds to the same site on the terminating ribosome as eRF3 (thus their binding is mutually exclusive), and its 4Fe-4S domain interacts with the C-terminal domain of eRF1 to push the conserved GGQ motif in the middle domain of eRF1 to the PTC next to the acceptor stem of the P-site tRNA to trigger polypeptide release [13].
Recycling of eRF1-associated post-termination complexes (post-TCs) is also mediated by ABCE1/RLI1, this time, however, in an ATP-dependent manner [6], [12]. It was hypothesized that RLI1, upon binding and hydrolyzing ATP, switches its conformation into a closed state, and the mechanochemical work generated by this switch splits post-TCs into free 60S subunits and deacylated tRNA- and mRNA-bound 40S subunits (40S-post-TC) [13]. Finally, Pisarev et al. showed that the release of tRNA and mRNA from the 40S-post-TCs is in vitro ensured by the bona fide initiation factors eIF1, eIF1A and eIF3 [5], [6]. eIF3, and in particular its j subunit (HCR1 in yeast), were suggested to play the key role in mRNA dissociation.
Since the implication of eIF3 in the recycling process was deduced only from experiments carried out with 11-codon long model mRNAs in mammalian in vitro reconstituted systems, we decided to investigate whether or not eIF3 also plays a direct role in translation termination and/or ribosomal recycling in the living cell. Here we show that the five core eIF3 subunits and HCR1 control translation termination and stop codon read-through in yeast, although in the opposite manner. HCR1 specifically cooperates with eRF3 and RLI1 and based on our and previous findings we propose that HCR1 and its mammalian orthologue should no longer be considered as bona fide subunits of eIF3. In any case, involvement of the canonical translation initiation factor eIF3 in termination strongly supports the idea that there is a highly coordinated communication between individual translational phases.
eIF3 and the eIF3-core-associated factors like HCR1 and RLI1 play a role in ribosome recycling – at least in vitro [5], [6], while only RLI1 is to date known to somehow promote also the preceding translation termination step [2]. In order to address whether eIF3 itself is likewise functionally involved in translation termination, we first measured the frequency of stop codon read-through in a collection of eIF3 mutants using an established dual-luciferase reporter assay, specifically designed to be independent of mRNA levels [14]. This reporter system is similar to the one which was also used to demonstrate increased stop codon read-through upon conditional down-regulation of RLI1 [2]. The [psi−] strain background used in these initial experiments contains a genome-encoded UGA suppressor tRNA leading to unusually high basal UGA read-through levels of 3–4%, which is however ideal for studying stop-codon read-through effects. Importantly, as shown below, the results we obtained are independent of the presence of this suppressor tRNA.
The eIF3 mutants that were chosen for read-through analysis were previously shown to affect multiple initiation steps, from the 43 PIC assembly (due to reduced 40S-binding affinity of eIF3) to scanning for AUG recognition (with wild-type 40S-binding affinity of eIF3); summarized in Table S1. Strikingly, the majority of mutations in the core eIF3 subunits that we tested showed a significant reduction in stop-codon read-through (Figure 1) that thus could not be simply attributed to the reduced eIF3 association with ribosomes. Also, since the effect of eIF3 mutations on translation initiation does not correlate well with the observed translation termination defect (such as, for example, in case of prt1-W674A vs. tif34-DD/KK or tif35-TKMQ vs. tif35-RLFT mutants), we conclude that the impact of these mutations on translation initiation vs. termination is genetically separable. Importantly, in contrast to all core eIF3 subunits, deletion of the non-essential hcr1 gene encoding eIF3j resulted in significantly increased stop-codon read-through (Figure 1), similar to that reported for RLI1 down-regulation [2]. Neither eIF3 mutations nor hcr1Δ have any impact on eRF1, eRF3 and RLI1 protein levels.
In order to confirm this unexpected result and to explore whether the observed effect on translation termination was specific to eIF3 or common to all other members of the Multifactor complex (MFC; composed of eIF1, eIF2, eIF3 and eIF5) and their closely co-operating factor eIF1A, we used partial depletion alleles (DaMP alleles) for these essential factors from the genome-wide DaMP collection [15]. DaMP alleles contain a selectable marker cassette inserted into the 3′-UTR of a gene, leading to destabilization of the respective mRNA via the nonsense mediated decay pathway (NMD).
By their nature, DaMP alleles show varying degrees of depletion for different genes, and data obtained with DaMP alleles have to be interpreted with this in mind. Since the depleted genes are all essential, loss of the corresponding gene product below a critical level will affect growth, and demonstration of reduced growth for an individual strain can thus be taken as a reliable indicator for depletion below a critical threshold. In contrast, an absence of growth phenotypes cannot be unambiguously interpreted, as depletion of the gene product may have occurred but may have remained above a level where fitness is detectably affected. It should be noted that growth phenotypes are a better indicator of functional depletion than assessment of physical expression levels by Western blotting, since the relationship between translation factor abundance and translation rates is non-linear and generally not predictable [16]. Finally, since the DaMP strain background does not contain a UGA suppressor tRNA in the genome, these data also exclude possible suppressor tRNA effects on the eIF3 mutants in the initial experiments.
When we compared growth rates of the MFC DaMP alleles to the corresponding wild type (wt) strain (Figure S1), we observed that most non-eIF3 MFC factors but only one of the eIF3 strains (TIF35) showed a growth phenotype indicative of a significant depletion. When we proceeded to assess stop codon read-through in these strains, we observed that the one eIF3 strain for which the growth assay indicated significant depletion (about 2.5-fold as determined by Western blotting) also showed significantly reduced stop codon read-through. In addition, the NIP1 depletion strain also showed modest but significant reduction in stop-codon read-through, which may be caused by depletion of the gene product to a level that does not yet affect growth rates. In contrast, none of the depletion alleles for the non-eIF3 MFC component showed a reduction in read-through, although eIF1A and eIF2α showed small but significant increases in stop codon read-through and, interestingly, eIF2γ showed even higher increase (∼2-fold) that is similar to hcr1Δ. While the mechanism behind the increased read-through in the eIF1A and eIF2 alleles is yet to be explored, these observations demonstrate that i) reductions in eIF3 activity reliably lead to reductions in stop-codon read-through levels, whether this reduction is caused by point mutations or other gene ablation alleles, ii) this effect is specific to core eIF3 subunits, whereas other MFC components and HCR1 display either none or the opposite phenotype, and iii) general reduction in the initiation rates does not automatically affect the precision of translation termination – see for example the read-through data for eIF1 (Figure S1), the protein levels of which were depleted by ∼5-fold – suggesting that the reduced levels of read-through displayed by eIF3 mutants do not necessarily occur as an indirect consequence of an overall compromised protein synthesis. Interestingly, a similar phenotype (reduced read-through) was also observed for overexpression of eRF1 in otherwise wt cells (Figure S2; combined overexpression of eRF1 and 3 led to further, modest exacerbation), indicating that the eIF3 mutants may somehow affect the availability of eRFs for the stop codon in the A-site.
Next we wished to explore a molecular mechanism of the increased stop-codon read-through phenotype displayed by deletion of the non-essential hcr1 gene (Figure 1), which sharply contrasted with the opposite termination phenotype of mutant core eIF3 subunits (Figure 1). As discussed above, recent reports suggested that the ABCE1/RLI1 protein not only critically promotes ribosomal recycling [6], [12], [17] but is also somehow involved in translation termination, as its conditional depletion produced an increased read-through defect [2]. Moreover, RLI1 was also implicated in biogenesis and transport of pre-ribosomes from the nucleolus [18] and in stimulating translation initiation by promoting assembly of 43S PICs together with eIF3 [1]. The striking resemblance of the latter effects with the previously reported functions of HCR1 [18]–[24] plus the earlier observations that RLI1 directly interacts with HCR1 via its ABC2 domain [2] and that combination of hcr1Δ with the TAP-tagged RLI1 results in synthetic lethality [18] prompted us to test a potential functional redundancy between these two proteins.
Strikingly, we found that overexpression of RLI1 (about 2.5-fold) fully suppressed the slow growth defect of an hcr1Δ strain (Figure 2A). Moreover, high copy (hc) RLI1 also fully suppressed the increased read-through phenotype of this strain (Figure 2B). By way of control, we overexpressed elongation factor eEF3 (encoded by YEF3) as an independent ABC cassette-containing protein engaged in translation, which had no effect on the growth or read-through phenotypes of the hcr1Δ strain (Figure S3A and B). In addition, hc RLI1 did not suppress the increased read-through defect of the DaMP eIF2γ mutant (data not shown), further underscoring the novelty of the proposed role for HCR1 and eIF3 in termination.
Importantly, no hc suppression was observed when either the formation of the RLI1 N-terminal 4Fe-4S clusters (C25S and C61S mutants) or the ATP binding by its ABC cassettes (K116L, K391L, G224D G225D and G470D G471D mutants) were compromised (Figure 2B and data not shown). The integrity of the crucial N-terminal region of RLI1 as well as its intact ATPase activity are therefore critically required for a functional replacement of HCR1. In the opposite arrangement, hc HCR1 suppressed neither the slow growth nor the increased read-through phenotype of the Tet::RLI1 conditional depletion strain (Figure S4). It is noteworthy that in agreement with earlier results [2], [12], [17], the intact N-terminal 4Fe-4S clusters and the ability of RLI1 to bind and hydrolyze ATP were absolutely essential for restoration of the read-through defect in the Tet::RLI1 cells (Figure S4).
In order to find out if HCR1 acts independently of eIF3 in the termination process, we examined the read-through phenotype of hcr1 mutations, which are known to eliminate binding of full length HCR1 to eIF3 [21], [22]. As shown in Figure 2B (mutations hcr1-Box-NTA, -Box6 and -Box6+R/I), no effect was observed implying that the HCR1 function in termination does not require its physical association with eIF3.
One of the major initiation phenotypes of hcr1Δ is a leaky scanning defect (a decreased ability to recognize AUG as the translational start site resulting in increased scanning past it), which can be suppressed by hc eIF1A [21]. As can be seen in Figure 2B, neither hc eIF1A nor eIF1 suppressed the read-though defect of hcr1Δ. Similarly, hc RLI1 did not suppress the leaky scanning defect of hcr1Δ (Figure S5). Hence, these findings clearly suggest that the hcr1Δ defects in initiation and termination are genetically separable and that RLI1 cannot replace HCR1 in all of its functions. Importantly, however, since hc eIF1A suppressed the hcr1Δ growth defect only partially [21], as opposed to the full suppression by hc RLI1 (Figure 2A), we propose that the major contributor to the hcr1Δ slow growth phenotype is not a defect in initiation, as previously believed, but a defect related to translation termination. Altogether it seems that this substoichiometric “subunit” of eIF3 works more independently of the core eIF3 than previously thought and therefore we suggest not considering HCR1 as a bona fide eIF3 subunit anymore (see Discussion for more details).
If eIF3 and HCR1 are indeed involved in translation termination as our read-through data indicate (Figure 1), it should be possible to detect a complex between these molecules and the release factors in vivo. We therefore carried out a series of in vivo pull down experiments using Myc-tagged RLI1, or TAP-tagged HCR1, a/TIF32 or eRF3 as baits. To stabilize presumably only transient interactions between eIF3 and termination/recycling factors, the TAP-tag experiments were performed after modest (1%) pre-treatment of growing cells with formaldehyde as described in [25]. As shown in Figure 3A, Myc-tagged RLI1 specifically pulled down selected eIF3 subunits (∼51±1% of the input for a/TIF32 and ∼43±2.6% for g/TIF35) and HCR1, as shown before [1]. In addition and in contrast to the latter study, we also observed significant co-precipitation of both release factors (∼22±1.7% for eRF1 and ∼13±3.9% for eRF3). The TAP-tagged HCR1 repeatedly co-purified with selected eIF3 subunits, as expected, but also with RLI1 (∼58±5.8%) and small but specific amounts of eRF3 (9±0.2%) and eRF1 (∼2±0.2%) (Figure 3B; eRF1 is indicated by an asterisk). eRF3 also co-precipitated with TAP-tagged a/TIF32 (∼16±4.7%), and, importantly, TAP-tagged eRF3 reproducibly and specifically brought down small but significant amounts of core eIF3 subunits (∼6.4±0.7% for a/TIF32 and ∼18±5.8% for g/TIF35) but no other MFC-members such as eIF1 (Figure 3C and D; note that the mobility of a/TIF32 and eRF3 vary between Input and Elution lanes due to a TEV protease-mediated cleavage of the TAP tag). We also tested the TAP-tagged eRF1 strain, however, no proteins were recovered – not even the TAP-eRF1 by itself – indicating that this particular fusion allele is not functional. Importantly, the yield of neither of these experiments was affected by RNase A treatment (Figure S6) and no ribosomes were present in the purified complexes (see RPS0A strips in panels A–D) strongly suggesting that the ribosome- and RNA-free complexes of eIF3, HCR1, eRF1, eRF3 and RLI1 do exist in the cytoplasm. More specifically, these experiments show that eIF3 and HCR1 contacts all critical termination players discussed in this study, though we cannot conclude whether all these factors occur in one single super-complex, or whether we are pulling down their partial subcomplexes.
The fact that eIF3 and HCR1 associate with eRFs in vivo and that their mutations affect fidelity of the termination process prompted us to test protein-protein interactions between eIF3 subunits, HCR1 and both eRFs. We fused individual eIF3 subunits and HCR1 to a GST moiety and used these fusions in pull-down assays with in vitro synthesized, radiolabeled, well-defined domains of eRF1 and eRF3. As shown in Figure 3E, the N-terminal and Middle (N-M) domains but not the middle and C-terminal (M-C) domains of eRF1 specifically interacted with GST-g/TIF35 and weakly also with GST-i/TIF34, in contrast to GST-HCR1 and a negative control of the GST protein alone. (The N domain of eRF1 carries determinants of the stop codon recognition; the M domain contains the conserved GGQ motif required for peptide release; and the C domain interacts with eRF3.) Interestingly, DOM34/Pelota, the release-like factor closely related in sequence and structure to eRF1, also binds eIF3g in human cell lines [26], albeit in this case via Pelota's C-terminal domain. No interactions between eIF3 subunits and eRF3 were observed. g/TIF35 can be divided into the N-terminal Zn-finger and C-terminal RRM domains and our GST pull down experiments revealed that eRF1-NM binds specifically to the N-terminal Zn-finger domain of g/TIF35 (Figure 3F). Hence we propose that eIF3 and eRF1 are in a direct contact via two small eIF3 subunits and the NTD of eRF1, which requires further support from the M domain to get fully engaged in these interactions.
To provide more solid evidence implicating eIF3 and HCR1 in the process of termination in vivo, we tested whether or not both factors associate with polysomal 80S ribosomes by separating the formaldehyde cross-linked whole cell extracts (WCEs) on sucrose gradients by high velocity sedimentation into two major polysomal pools; the first containing light polysomes (disomes and trisomes) and the other heavy polysomes (from pentasomes up). These two pools were then treated with RNase I (Invitrogen) to chop polysomal mRNAs into segments containing either initiating 43S-48S PICs or elongating/terminating 80S ribosomes. The second round of sucrose gradient centrifugation (so called resedimentation; [25]) was employed to separate the 43-48S PICs from 80S couples in each polysomal pool into two fractions, which were then subjected to Western blotting. In both pools, the 80S fractions contained more than 50% of total eIF3 in comparison with the 40S fractions (Figure 4A), clearly demonstrating that only the lesser proportion of eIF3 occuring in polysomes is associated with initiation complexes. Strikingly, in case of HCR1, the 80S fractions contained even more, ∼90% of this protein from the overall pool. In contrast, the “polysomal” fraction of eIF5, which is known to tightly interact with eIF3 during translation initiation, was predominantly associated with 40S species. These results are thus consistent with a role of eIF3 and HCR1 in other translational phases than just initiation.
In the resedimented light and heavy polysomes different ratios of terminating versus initiating plus elongating ribosomes can be expected based on the following arguments. Under nutrient-rich conditions, the disome/trisome fraction will contain short mRNAs that cannot accommodate more than two/three ribosomes, poorly translated mRNAs, as well as recently transcribed mRNAs, which are not yet in a steady state phase with regards to their ribosome occupancy. Since mRNAs shorter than 60 codons make up only 2% of the yeast transcriptome [27], we expect that a majority of mRNAs in this pool are newly synthetized species with a standard/average mRNA length. Hence we posit that the light polysomal mRNAs contain a smaller proportion of terminating ribosomes than mRNAs isolated from heavy polysomes, since the likelihood of having a stop codon occupied by a terminating ribosome increases with the increasing number of elongating ribosomes per mRNA. In support of this rationale, we observed more than 3-fold increase in the amounts of eRF3 associated with 80S ribosomes isolated from heavy versus light polysomes. By the same token, if eIF3 was involved specifically in translation termination events, we would expect stronger association of eIF3 and HCR1 with 80S couples originating from heavy polysomes. We do indeed observe that the 80S/40S ratio of eIF3 and HCR1 association is 2- and 2.5-fold higher, respectively, for heavy polysomes than for the light ones, in contrast to eIF5 where it remains the same (Figure 4A). These results thus strongly suggest that eIF3 and HCR1 are present at 80S ribosomes during the terminating process.
To further support this conclusion, we employed the Tet::SUP45 conditional depletion strain. We rationalized that if eIF3 specifically associates with terminating ribosomes, depletion of eRF1 should significantly reduce the presence of eIF3 subunits (as well as the presence of eRF3) in the 80S couples isolated from heavy polysomes. To test this, we formaldehyde cross-linked the Tet::SUP45 cells grown in the presence or absence of 1 µg/ml doxycycline for six hours before harvesting, resolved the WCEs on sucrose gradients, collected only the fractions containing heavy polysomes, treated these fractions with RNase I and separated the resulting 43-48S PICs and 80S species by the second round of centrifugation (resedimentation). Thus isolated 80S couples from Dox− versus Dox+ Tet::SUP45 cells were loaded in six serial two-fold dilutions to the SDS-PAGE gel and the amounts of RPS0A (as a loading control) and associated eIFs and eRFs were analyzed by quantitative western blotting. Depletion of the key termination factor had to be rapid to avoid disassembly of stalled post-TCs in polysomes as well as secondary defects of dying cells. In our set-up we achieved ∼70% depletion of eRF1 and observed no changes in polysome profiles of Dox− versus Dox+ Tet::SUP45 cells (data not shown). As predicted, whereas the 80S-associated amounts of eIF1 remained unchanged (small), the amounts of eRF3 and two eIF3 subunits were significantly reduced (by ∼40%) in Dox+ versus Dox− Tet::SUP45 cells (Figure 4B). Note that while the overall levels of eRF1 were depleted by ∼70%, polysome-associated eRF1 was only depleted by ∼30%, which is consistent with the quantitatively similar reduction in polysome association observed for eRF3 and eIF3.
In order to examine how the network of interactions between translation initiation and termination factors affects their functions, we investigated the distribution of selected translation factors in wt cells and cells mutated for either of the factors under study using formaldehyde cross-linking of living cells by sucrose density gradients analysis of WCEs [25].
Figures 5A and S7A show a typical distribution of the selected proteins across all gradient fractions obtained from wt WCEs and divided into several separable groups: “Top” (fractions 1–4), “40S” (5–6), “60S” (7–8), “80S+light polysomes” (9–13) and “heavy polysomes” (14–18). For technical reasons, several fractions from individual groups were pooled together to fit all samples on a single SDS-PAGE gel. Whereas eRF3 is clearly enriched in the polysome-containing fractions and practically lacking in the Top fractions, eRF1 is more or less evenly distributed across the entire gradient, and RLI1 predominantly sediments in the Top fractions and partially also in the 40S-containing fractions. Importantly, all strains that we worked with in this study are [psi−] and hence the observed sedimentation of eRF3 into heavier fractions cannot be attributed to SUP35 aggregation. In agreement with the aforementioned analysis, eIF3 (represented by a/TIF32) and HCR1 show a robust enrichment in polysomal fractions, similar to eRF3, whereas eIF5 occurs mainly in the Top and 40S fractions.
As shown in Figures 5B, E and S7A, deletion of hcr1 significantly shifted the amounts of “initiating” eIF3 from the 40S fractions to the Top, as observed before [20], whereas it had no effect on polysomal distribution of eRF1 and RLI1. However, it led to a statistically significant accumulation of eRF3 in heavy polysomes with a commensurate reduction in lighter fractions (Figures 5B, D and S7A). We interpret the accumulation of eRF3 in heavy polysomes as an increased number of post-TCs bound by eRF3 in a less-productive manner; perhaps with a decreased dissociation rate. Most importantly, overexpression of RLI1, which suppresses both the read-through and Slg− phenotypes of hcr1Δ (Figure 2), partially but significantly restored the eRF3 distribution in polysomes to wt (Figures 5C, D and S7A). The fact that the 40S-binding by the “initiating” eIF3 was not restored (Figures 5C, E and S7A) underscores a specificity of the RLI1 suppressor effect on the HCR1 role in termination versus initiation. One way to explain these observations is that HCR1 may promote the release of eRF3·GDP from the post-TCs upon stop codon recognition and GTP hydrolysis on eRF3, which serves as a prerequisite for the subsequent binding of RLI1 as well as the eRF1-stimulated hydrolysis of the bond between the P-site tRNA and the nascent polypeptide (see our model in Figure 6). Inability to complete this step may lead to a reduced stop codon recognition resulting in an increased read-through, which was observed. Hence the suppression effect of RLI1 on the molecular level could be explained by proposing that increased dosage of RLI1 forces dissociation of eRF3·GDP from the post-TCs by mass action and thus eliminates a need for HCR1.
If our model is correct, one can predict that the polysomal levels of HCR1 could be reduced without any functional defect, if the interaction between eRF1 and eRF3 was impaired. Hence we next analyzed changes in polysomal distribution of factors of interest in the sup45Y410S mutant, which disrupts the eRF1–eRF3 interaction [28]. As expected, the sup45Y410S mutant significantly shifted the amounts of eRF3 from polysomal to the Top fractions when compared to wt (Figures 7A–C and S7B). Importantly, in accord with our proposed model, a similar, significant change was also observed for HCR1 but not for a/TIF32 and RLI1. We interpret these data by proposing that HCR1 readily dissociates along with eRF3 when eRF3 binding to the post-TCs is weakened by a mutation. This effect could be either direct or indirect/allosteric. The fact that we could not detect any direct binding between HCR1 and eRF3 using conventional in vitro protein-protein binding techniques may speak for the latter option; however, it is also possible that HCR1–eRF3 binding does occur but only in the context of the post-TCs.
To further support our model, we analyzed genetic interactions between the hcr1 deletion strain and selected mutations in both release factors. The temperature sensitive eRF mutants we used are all known to cause termination defects including stop codon read-through strong enough to suppress the ade1-14 nonsense allele [29]. They include a sup35N536T mutant located in a region near the C-terminus of eRF3 that disrupts termination by an unknown mechanism, a sup45M48I mutant that interferes with stop codon decoding [30], and the aforementioned sup45Y410S mutant that directly disrupts the eRF1–eRF3 interaction [28].
It could be proposed that if the sup45Y410S Ts− mutant reduced or even eliminated a need for HCR1 functioning in termination, an epistatic interaction should be observed when this mutant is combined with hcr1Δ. Consistently, at the permissive temperature the absence of HCR1 further increased read-through of this sup45 mutant (Figure 8A, 30°C; compare open and grey bars with the black one) and also exacerbated its slow growth (Figure 8B). However, at the higher temperature, where the eRF1:eRF3 interaction is more severely disrupted by the sup45Y410 mutation [28], as evidenced by its increased termination defect (Figure 8A; compare grey bars between 30 and 34°C), the absence of HCR1 had only a little additional effect on the sup45Y410S read-through (Figure 8A, 34°C; black vs. grey bars). Moreover, the sup45Y410S mutation also completely eliminated the negative impact of hcr1Δ on growth rates at this temperature (Figure 8B). The specificity of this epistatic interaction is further underscored by the fact that neither sup45M48I (eRF1) nor sup35N536T (eRF3) mutants showed any synthetic effects in the background of the hcr1 deletion (Figure S7C).
Finally, to obtain further genetic evidence supporting our findings implicating eIF3 in regulation of the termination process, we combined two selected mutations in the a/TIF32 subunit of eIF3 (Δ8 and Box17), both reducing the stop codon read-through in otherwise wt cells (Figure 1), with the sup35N536T and sup45Y410S mutants. When combined, the double mutants show a stop codon read-through frequency that is clearly reduced compared to either release factor mutant (Figure 8C), demonstrating that the tif32 mutations partially rescue the read-through phenotype of the latter. In contrast, when we investigated slow growth (Slg−) and temperature sensitive (Ts−) phenotypes, we observed synthetic exacerbation of these phenotypes (Figure 8D). This demonstrates that i) the release factors and the core eIF3 complex have antagonistic functions in the same stage of the termination phase and losses in their functions can thus partially compensate for each other in terms of the stop codon read-through efficiency; and ii) that the degree of stop codon read-through per se is not the major source of the fitness defects in these strains. This latter notion is consistent with earlier quantitative trait analyses, which showed that the termination defects are unlinked from growth defects in many eRF1 mutants [31]. Hence, synthetic exacerbation of growth could be explained by proposing that besides the stop codon recognition step (which is assessed in the dual luciferase assay), also other aspects of termination are impaired in the eRF and eIF3 mutants, which, in combination with the initiation defects of eIF3 mutants, reduce the growth rate as a compound effect.
It is becoming increasingly apparent that factors involved in regulating various steps of gene expression may have multiple functions and that this multitasking may integrate transcription, mRNA export, translation and mRNA decay into a delicately regulated higher-order process. For example, translation initiation factor eIF3 links translation initiation to transcription [32], to mRNA export [3] and to the NMD pathway [33], [34]. Here we show for the first time that eIF3 and HCR1 critically connects initiation of translation with its termination.
In particular, deletion of hcr1 increases stop codon read-through, independently of its association with the rest of eIF3, and results in accumulation of eRF3 in heavy polysomes. Increased dosage of ribosomal recycling factor RLI1 then substitutes for the HCR1 roles in termination (but not in initiation) and in enabling efficient cell growth, together implying that the HCR1 function in termination is more critical for optimal cell proliferation than its function in translation initiation. This is consistent with the fact that yeast HCR1 is only loosely associated with the core eIF3 complex [19] and that it was shown to interact with both sides of the 40S mRNA entry channel on its own [21]. Similarly, its mammalian ortholog also appears to be the most loosely associated subunit of all 13 eIF3 subunits that is, in addition, often missing from the purified 12-subunit complex [35]–[37]. Moreover, it was also shown to associate with the 40S ribosome completely independently of the rest of eIF3 and promote several translational steps practically on its own (see for example [5], [35]–[40]). Taken together, we suggest considering eIF3j/HCR1 as an independent initiation factor (eIF) that associates and closely co-operates with eIF3 but it is not its integral part. We therefore propose to use the following designations for this old-new eIF: HCR1 for the yeast protein and hHcr1 for its mammalian counterpart.
In contrast to hcr1Δ, various mutants of core eIF3 subunits, but not of other initiation factors, decrease stop codon read-through in living cells and show synthetic phenotypes with mutant release factors eRF1 and 3. eIF3 also directly interacts with eRF1 and occurs in complex with eRF1, eRF3 and RLI1 in vivo. Since eIF3 and HCR1 were, based on in vitro experiments, previously implicated in promoting also the very final step of translation – ribosomal recycling [5], we propose that eIF3 – and to some extend also HCR1 – is one of the very few factors that connects various processes of mRNA life and integrates them into the ultimate translational output. Taking into account that the translation pathway is highly conserved among low and high eukaryotes, it is highly likely that this connecting role of eIF3 is also conserved.
Our observations that 1) eIF3 and eRF3 can be found enriched in 80S fractions isolated from RNase-treated heavy polysomes in an eRF1-dependent manner (Figure 4); 2) that a complex between eIF3, RLI1 and eRFs exists free of RNA and ribosomes in the cytoplasm (Figure 3A–D), and 3) that two small eIF3 subunits g/TIF35 (in particular its NTD) and i/TIF34 directly interact with the N and N-M domains of eRF1 (Figure 3E–F) together suggest that eIF3 does associate with terminating 80S couples and may come to the pre-TC in a pre-formed complex with eRFs. The alternative that they are ejected from post-TCs as a holocomplex upon completion of termination is highly unlikely considering that i) eRF3 must be ejected prior to RLI1 binding [12] and ii) that eIF3 is supposed to participate in the late steps of ribosomal recycling that should be devoid of eRF1 and RLI1 [5], [6]. The last scenario would be that eIF3 stays present on the elongating ribosome throughout the entire elongation cycle and promotes recruitment of eRF1·eRF3·GTP to the pre-TCs; there is, however, genetic evidence contradicting this possibility [9].
Our data show that eIF3 mutants specifically decrease stop codon read-through in otherwise wt cells (Figures 1 and S1) and that tif32 mutations partially compensate for the increased read-through in eRF mutants (Figure 8C). This clearly suggests that wt eIF3 modulates the precision of stop codon recognition by eRF1 in order to fine tune the termination process (see our model in Figure 6). During stop codon decoding, eRF1 was proposed to sit in the ribosomal A-site with a part of its N-domain contacting small ribosomal protein RPS3 and helix (h) 18 of 18S rRNA [41]. Strikingly, g/TIF35 also interacts with RPS3, in addition to RPS20 [42], and as both g/TIF35 and i/TIF34 are tightly bound to the extreme C-terminus of b/PRT1 [43], i/TIF34 is expected to occur nearby g/TIF35. Moreover, the C-terminal domain of a/TIF32 interacts with h16-18 of 18S rRNA [44] and RPS3 as well [22]. Taylor and colleagues further proposed that one of the conformational changes induced by eRF1–eRF3–GMPPNP binding to pre-TCs involves a movement of h16 of 18S rRNA and the N-terminal domain (NTD) of RPS3 toward each other, which results in the establishment of a new head–body connection on the solvent side of the 40S subunit and a constriction of the mRNA entrance. Hence, it is conceivable that eRF1 and eIF3, by contacting the same 40S binding partners, modulate these conformational changes in the termination complex in a way that influences a proper placement of eRF1 in the spatially restricted A-site. This scenario could provide a rational explanation for the antagonistic effect of eIF3 on translation termination.
For interpretation of these data it must be kept in mind that the reporter constructs we use essentially measure stop codon read-through on a premature termination codon. At present, we do not know whether the antagonistic influence of eIF3 on stop-codon read-through is restricted to such sites, or whether it also affects termination on stop codons located nearer to the poly(A) tail. However, our observation that the sup45Y410S mutant, which affects stop codon selection by disrupting the eRF1–eRF3 interaction, reduced the polysome-associated amounts of eRF3 and HCR1 (Figure 7) indicates that a delay or imperfection in the decoding of natural stop codons disrupts this “initiation-termination” complex, most probably to enable resumption of elongation. Investigation of the precise molecular mechanism of the eIF3 action in termination is a pressing task for our future research.
In contrast to mutations in core eIF3 subunits, deletion of hcr1 did not decrease but increased the stop codon read-through (Figure 1). The fact that mutations disrupting the HCR1 contact with eIF3 had no effect on read-through clearly suggests that the HCR1 role in termination is independent of its association with eIF3, as discussed above. Moreover, our findings that hcr1Δ results in accumulation of eRF3 in higher polysomal fractions (Figure 5) and that sup45Y410S (breaking the eRF1–eRF3 interaction) shifts HCR1 to the Top fractions (Figure 7) led to the model presented in Figure 6. We propose that following stop codon recognition and subsequent GTP hydrolysis on eRF3, HCR1 promotes eRF3·GDP ejection from the post-TCs to allow binding of its interacting partner RLI1 [2], which in turn stimulates polypeptide release – both eRF3 and RLI1 bind to the same site in the post-TC [13]. Inability to complete this step may lead to a reduced stop codon recognition resulting in an increased read-through. In support, eRF1 was shown to associate more firmly with post-TCs in the presence of eRF3 [6], which led the authors to propose that after GTP hydrolysis, eRF3 might not dissociate entirely from ribosomal complexes on its own and its release thus might require a stimulus by an additional factor; in our opinion by the HCR1 protein.
Based on the cryo-EM structures of DOM34:HBS1 (release factor-like proteins closely related in sequence and structure to eRF1:eRF3) on the yeast ribosome showing that the N-terminus of HBS1 extends away from the body of the protein and contacts the mRNA entry site, it was proposed that the N-terminus of eRF3 also occurs in the A-site area [13]. Since HCR1 was shown to occur in this area too [21], it could directly act upon this eRF3 domain to trigger the release of this factor in its GDP form from eRF1-bound post-TCs. In support, the N-terminal extension of S. pombe eRF3 was proposed to regulate eRF1 binding to eRF3 in a competitive manner [45]. Interestingly, both the N-terminus of eRF3 as well as the HCR1 protein as a whole are non-essential [45], [46], suggesting that they might act simply by shifting the equilibrium towards the loss of affinity between the eRF1 and eRF3·GDP binary complex. If true, the loss-of-function of both of them could be overcome by redundant mechanisms with slower reaction rates. In agreement, hc RLI1 fully suppressed the read-through effect of hcr1Δ in a manner dependent on its intact 4Fe-4S and ABC domains (Figure 2B). We propose that in the hcr1Δ cells, RLI1 makes its way to its binding site in the post-TCs by forcing dissociation of eRF3·GDP through mass action and thus eliminates a need for HCR1. These results are consistent with the aforementioned observation that the eRF1 mutation sup45Y410S, disrupting the eRF1–eRF3 interaction, shifts the amounts of eRF3 and also that of HCR1 from polysomes to the top of the gradient (Figure 7).
The model proposed in Figure 6 also explains the behavior of genetic interactions observed for the hcr1 deletion (Figure 8). Failure to eject eRF3·GDP can perceivably have two consequences. First, if peptidyl hydrolysis by eRF1 fails to be induced because RLI1 cannot bind to it, the eRF1·eRF3·GDP complex can dissociate from the ribosomal A-site, thus necessitating a renewed round of tRNA sampling with an ensuing risk of stop codon decoding by a near-cognate or suppressor tRNA. This is consistent with the increased stop codon read-through we observe experimentally in hcr1 deletion strains. Second, if peptidyl hydrolysis does take place (in vitro, eRF1 clearly has some release factor activity also in the absence of RLI1 [11]), a stalled ribosome complex would be formed in which eRF1 was still bound to eRF3, and in which RLI1 was thus not free to initiate the recycling step. Such stalled complexes would impede ribosome flow on the affected mRNA, reduce corresponding gene expression levels and potentially necessitate degradation by one of the surveillance pathways. If this occurred frequently, it would give rise to fitness defects, as we observe for hcr1 deletion strains. This is also consistent with the fact that deletion of hcr1 produces unexpectedly mild polysomal run-off with respect to its growth defect [23]. However, in the presence of eRF1 mutations, which accelerate spontaneous dissociation of eRF3·GDP from eRF1, timely RLI1 binding to eRF1 in the post-TCs would be re-enabled even in the absence of HCR1. This would explain why the sup45Y410S mutation, but not sup45M48I and sup35N536T mutations, eliminated the negative impact of hcr1Δ on growth rates at the semipermissive temperature (Figure 8B).
To further support our model, we wished to employ a recently established in vitro reconstituted yeast translation system, which has been used previously to monitor both the peptide release and ribosome recycling steps of the translation cycle [12]. HCR1 did not have an appreciable effect on ribosome recycling in this assay (unpublished observations). This is probably not surprising given that ribosome recycling is slow relative to the preceding steps. Thus, accelerating eRF3.GDP dissociation is unlikely to affect the observed rate of recycling. In contrast, the model predicts that the observed rate of peptide release by eRF1, eRF3 and RLI1 may accelerate in the presence of HCR1. Unfortunately, however, the rate of peptide release by eRF1, eRF3 and RLI1 is very rapid, such that further increases in rate (such as those that may occur in the presence of HCR1) are unable to be measured in this system. Since the former two assays are the only in vitro assays available to us at the moment and neither of them can either directly or indirectly monitor the rate of eRF3·GDP dissociation from the post-TCs, further efforts will be necessary to fully characterize the role of HCR1 in termination/recycling reactions biochemically.
Upon completion of the termination-specific reactions, eIF3, HCR1 and RLI1 further participate in the ribosomal recycling steps, as proposed by [6], and it is conceivable that all these factors remain bound to the small 40S subunit to promote the next round of initiation (Figure 6). Alternatively, the pre-occupation of the 40S·mRNA complex by the “initiation factors” that would not be recycled could ensure reinitiation on the same mRNA molecule as proposed by the mRNA closed-loop model [47]. An in vivo experimental evidence implicating eIF3, HCR1 and other eIFs in the recycling steps is, however, still lacking.
Taken together, we argue that strict mechanistic separation of translation into its individual, mutually independent phases should be reconsidered in the light of “multitasking” of eIF3, HCR1, RLI1 and most likely also eIF1 and eIF1A, for which evidence is presented here and elsewhere. Collectively, these findings suggest that changes in one phase of translation, evoked for example via cell signaling pathways, are promptly communicated to and coordinated with changes in the other phases to maintain cellular homeostasis of all ongoing processes. Without a doubt there is much to be learned about how all four phases of translation come together in one balanced system that rapidly and accurately responds to different needs of the cell exposed to constantly changing environmental conditions.
The lists and descriptions of plasmids and yeast strains used throughout this study can be found in the Supplemental Information (Tables S2, S3, S4 and Text S1).
Stop codon read-through assays were performed using a bicistronic reporter construct consisting of a Renilla luciferase gene followed by an in-frame firefly luciferase gene. Separating the two genes is either a tetranucleotide termination signal (e.g., UGA C) [plasmids pTH477 (URA3) or YEp-R/T-UGAC-L (LEU2)] or, for control purposes, a similar sequence containing a sense codon (e.g., CAA C) [plasmids pTH460 (URA3) or YEp-R/T-CAAC-L (LEU2)]. It is noteworthy that this system avoids possible artifacts associated with changes in the efficiency of translation initiation associated with the function of the NMD machinery [48], because both the Renilla and firefly enzymes initiate translation from the same AUG codon. For further details, see [14]. Microtitre-plate based dual luciferase assays and analyses of the resulting data were as described [31]. Samples were processed in quintuplicate, and each experiment was repeated at three times.
Yeast whole cell extracts (WCEs) were prepared as described previously [49] except that buffer A (30 mM HEPES (pH 8.8), 20 mM KAc, 3 mM magnesium acetate,1 mM dithiothreitol, 1% Nonidet P-40 supplemented with Complete Protease Inhibitor Mix tablets (ROCHE), and protease inhibitors 1 µg/ml aprotinin, 1 µg/ml leupeptin, 1 µg/ml pepstatin and 100 µM phenylmethylsulfonyl fluoride (PMSF)) was used for lysis of the cells, and cell lysates were centrifuged at 3,000 r.p.m. for 10 min at 4°C. The co-immunoprecipitation analysis was performed as described elsewhere [50], using 500 µg of the total protein and 1 µl of mouse anti Myc-Tag IgG (CELL SIGNALING TECHNOLOGY).
Yeast cells expressing the TAP-tagged genes of interest were grown in YPD medium at 30°C to an OD600 of ∼1 and treated with 1% HCHO prior to harvesting for 60 mins. The WCEs was prepared as described above using buffer B (50 mM Tris-HCl (pH 7.6), 150 mM NaCl, 0.05% Tween 20) with all protease inhibitors in the presence or absence of 0.1 mg/ml RNase A. Samples containing 1 mg of total protein in a final volume of 600 µl were incubated for 2 h at 4°C with 50 µl of 1∶1 slurry of IgG Sepharose 6 Fast Flow beads in buffer B. Samples were centrifuged briefly and the supernatants were removed. The collected beads were then washed five times with 1 ml of ice cold buffer B, and incubated either with TEV protease (INVITROGEN) for 30 min at 30°C followed by boiling in the SDS-loading buffer for 5 min at 95°C, or directly boiled the SDS-loading buffer. Corresponding aliquots of input, eluate and wash (supernatant) were analyzed by SDS-PAGE followed by immunoblotting.
The 0.5% formaldehyde (HCHO) cross-linking followed by WCE preparation and fractionation of extracts for analysis of translational complexes were carried out as described previously [25] with the following exceptions. Cycloheximide was added at a concentration of 0.05 mg/ml 5 minutes before the HCHO treatment, after which the cells were broken by FastPrep Instrument (MP Biomedicals) at the intensity level of 5 in two 20 second cycles. The resulting WCEs were separated on 5–45% sucrose gradients.
GST pull-down experiments with GST fusions and in vitro-synthesized [35S]-labeled polypeptides (see Table S3 for vector descriptions) were conducted as follows. Individual GST-fusion proteins were expressed in E. coli, immobilized on glutathione-Sepharose beads and incubated with 10 µl of 35S-labeled potential binding partners at 4°C for 2 h. The beads were washed 3 times with 1 ml of phosphate-buffered saline, and bound proteins were separated by SDS-PAGE. Gels were first stained with Gelcode Blue Stain Reagent (Pierce) and then subjected to autoradiography. β-galactosidase assays were conducted as described previously [51].
The GST-RLI1 and GST-SUP45 fusion proteins encoded by pGEX-RLI1, pGEX-SUP45, respectively, were expressed in E. coli and purified from the WCE by incubation with Glutathione-Sepharose 4B beads (Pharmacia). The isolated proteins were resolved by SDS-PAGE (4–20% gels), excised from the gel, and washed with 1× PBS. Rabbits were injected with the purified protein and sera containing polyclonal antibodies against RLI1, SUP45, respectively, were obtained commercially by Apronex (Prague, the Czech Republic).
|
10.1371/journal.pgen.1000090 | Combined Analysis of Murine and Human Microarrays and ChIP Analysis Reveals Genes Associated with the Ability of MYC To Maintain Tumorigenesis | The MYC oncogene has been implicated in the regulation of up to thousands of genes involved in many cellular programs including proliferation, growth, differentiation, self-renewal, and apoptosis. MYC is thought to induce cancer through an exaggerated effect on these physiologic programs. Which of these genes are responsible for the ability of MYC to initiate and/or maintain tumorigenesis is not clear. Previously, we have shown that upon brief MYC inactivation, some tumors undergo sustained regression. Here we demonstrate that upon MYC inactivation there are global permanent changes in gene expression detected by microarray analysis. By applying StepMiner analysis, we identified genes whose expression most strongly correlated with the ability of MYC to induce a neoplastic state. Notably, genes were identified that exhibited permanent changes in mRNA expression upon MYC inactivation. Importantly, permanent changes in gene expression could be shown by chromatin immunoprecipitation (ChIP) to be associated with permanent changes in the ability of MYC to bind to the promoter regions. Our list of candidate genes associated with tumor maintenance was further refined by comparing our analysis with other published results to generate a gene signature associated with MYC-induced tumorigenesis in mice. To validate the role of gene signatures associated with MYC in human tumorigenesis, we examined the expression of human homologs in 273 published human lymphoma microarray datasets in Affymetrix U133A format. One large functional group of these genes included the ribosomal structural proteins. In addition, we identified a group of genes involved in a diverse array of cellular functions including: BZW2, H2AFY, SFRS3, NAP1L1, NOLA2, UBE2D2, CCNG1, LIFR, FABP3, and EDG1. Hence, through our analysis of gene expression in murine tumor models and human lymphomas, we have identified a novel gene signature correlated with the ability of MYC to maintain tumorigenesis.
| The targeted inactivation of oncogenes may be a specific and effective treatment of cancer. However, how oncogene inactivation leads to tumor regression is not clear. Previously, we have shown that even the brief inactivation of the MYC oncogene can result in the sustained regression of at least some tumors. To understand the mechanism, we have utilized several novel genomic analyses to define a set of genes that strongly correlate with the ability of the MYC oncogene to maintain tumorigenesis. First, we generated a novel data set from microarray analyses of murine tumors that we analyzed by StepMiner to identify discrete step changes in gene expression after the inactivation or the reactivation of the MYC oncogene. Second, we utilized Boolean Network Analysis to further define the subset of genes highly correlated with MYC in human tumorigenesis. Third, we utilized ChIP analysis to demonstrate that in many cases the permanent changes of gene expression we uncovered were associated with changes in the ability of MYC to occupy the promoter locus. Our general strategy could be similarly utilized in other experimental model systems to understand how specific oncogenes contribute to the maintenance of tumorigenesis.
| Overexpression of MYC is one of the most frequent events in human tumorigenesis [1]. MYC overexpression is thought to induce tumorigenesis by causing inappropriate gene expression resulting in autonomous cellular growth, proliferation, and the inhibition of cellular differentiation [2],[3]. Many laboratories have conditionally overexpressed c-MYC (MYC) utilizing conditional transgenic model systems [4]–[9]. In these models, the suppression of MYC led to permanent loss of tumorigenesis through proliferative arrest, differentiation and/or apoptosis [4]–[6],[10]. In some circumstances, even the brief suppression of MYC overexpression permanently prevents its ability to sustain tumorigenesis [6]. These and other observations have suggested the possibility that oncogenes such as MYC exhibit the phenomena of oncogene addiction [11]. However, the molecular basis of oncogene addiction is not clear. Recently, we have suggested that cellular senescence, which involves chromatin modifications and heterochromatin formation [12],[13], may be an important mechanism for sustained tumor regression upon MYC inactivation [14].
MYC is thought to play a role in the regulation of up to 15% of genes in the fly, mouse or human [3],[15],[16]. Thus, it seems likely that changes in gene expression programs, rather than individual genes, account for the phenotypic consequences of MYC inactivation. Consistent with this notion, MYC has recently been shown to globally influence chromatin structure through histone modifications [17]–[19]. Similarly, N-MYC was shown to globally regulate acetylation and methylation of histone molecules [20]. We have reported that MYC inactivation in tumors induces specific global changes in histone modification [14]. Although many MYC target genes have been identified in various cells or tissue contexts (summarized in http://www.myc-cancer-gene.org), it is hard to discern which of the many of MYC targets are associated with the ability of MYC to initiate and/or sustain tumorigenesis.
Many previous studies have examined changes in gene expression associated with the induction of MYC expression in cells [3], [15], [21]–[26]. Other groups have performed comparative analysis of gene expression profiles between murine constitutive MYC-induced tumors and human tumors in liver and prostate cancers [27],[28]. Both of these analyses identified similarities in gene expression between MYC-induced tumor models and human tumors. Although revealing, these studies would not necessarily identify gene products that are responsible for the ability of MYC to induce tumorigenesis. We speculated that by analyzing gene expression profiles in tumors generated from conditional transgenic models would allow us to identify gene expression signature specifically associated with the ability of MYC to initiate and maintain tumorigenesis. We performed microarrays on mRNA samples from a time-course experiment with MYC inactivated and then reactivated in osteosarcoma. The expression data was then examined using the StepMiner algorithm [29] to generate a list of genes associated with MYC-induced tumorigenesis in osteosarcomas (Figure 1). The StepMiner algorithm analyzes microarray time courses by identifying genes that undergo abrupt transitions in expression level, and the time at which the transitions occur. Importantly, by ChIP we were able to demonstrate that permanent changes in gene expression were frequently associated with measurable alterations in the ability of MYC to bind to the promoter regions of these genes in osteosarcomas. Furthermore, gene expression profiles were compared between osteosarcomas and the previously published MYC conditional pancreatic tumor [25] to generate a common gene signature associated with MYC-induced tumorigenesis in mice (Figure 1). Finally, Boolean analysis was used to further examine the correlation between levels of expression of this identified subset of genes among the published dataset of 7,171 human microarrays in U133A format. From this analysis, we were able to deduce a list of genes strongly correlated with the ability of MYC to maintain tumorigenesis.
The MYC induced osteosarcoma derived cell line, 1325, was grown in vitro [6] and treated with 20ng/ml of doxycycline in complete DMEM medium for various length of time to inactivate MYC expression. To reactivate MYC expression, doxycyline was removed by rinsing the bone tumor cells with an excess amount of PBS. mRNA was collected from bone tumor cells treated with doxycycline for 0, 4, 8, 12, 18, 24, 36, and 48 hours, and after removal of doxycycline for 4 , 8, 12, 18, 24, 36, and 48 hours. MYC levels were greatly reduced as early as 4 hours after doxycycline treatment (Figure 2A and Figure S1). We confirmed that the expression of MYC could be reactivated to a level similar to that of MYC-on tumors by thoroughly washing the cells with PBS (Figure 2A and Figure S1).
cDNA microarray analysis was performed on the RNA samples prepared from tumors in which MYC was inactivated and reactivated for different lengths of time. StepMiner analysis (Figure 2B, 2C) [29] was applied to this time-course microarray experiment to identify changes in gene expression at discrete time points before and after MYC inactivation and reactivation. StepMiner fits step functions to the data points using an adaptive regression scheme and identifies time points at which a gene is significantly induced or repressed. Examples of one-step expression pattern are illustrated in Figure 2C.
Recently, we have shown that MYC inactivation generally induces cellular senescence in several tumor models [14]. Therefore, we specifically examined if the expression of senescence associated genes changed upon MYC inactivation in osteosarcomas. Indeed, we did find that senescence associated genes such as p15INK4b, p21CIP, PCNA, MCM3, CYCLIN A [12],[30],[31] were up-regulated or down-regulated upon MYC inactivation (Figure S2 and Table S1). Thus, our results support the notion that MYC inactivation is inducing changes in gene expression that is associated with cellular senescence.
Generally, analysis of gene expression changes after StepMiner analysis revealed four discrete patterns of changes in gene expression upon MYC inactivation and reactivation: Permanently Repressed (PR), Permanently Induced (PI), Reversibly Repressed (RR) and Reversibly Induced (RI). For this analysis, we set p<0.01 as a cutoff for statistically significant changes in gene expression. We identified 1016 unique probes in the PR group, 1777 unique probes in the PI group, 1148 unique probes in the RI group, and 1167 unique probes in the RR group (Figure 3 and Tables S2 for lists of genes).
Based upon our previously published observation that even brief inactivation of MYC can result in the sustained loss of the neoplastic properties of MYC-induced osteosarcomas [6], we speculated that genes which are potentially important for sustained tumorigenesis would be permanently repressed or induced (e.g. the PR group or the PI group) upon MYC inactivation.
To identify associated functional activities associated with the PR and PI groups of genes, we applied Gene Ontology analysis (GO Term analysis) to the list of genes generated above. Biological functions that were identified for each step upon MYC inactivation are listed (Table S3, S4, and S5). Associated functions identified include gene products known to regulate metabolism, biosynthesis of nucleotides and proteins and genes involved in the regulation or function of ribonucleoprotein complexes.
Notably, MYC has been shown to regulate expression of ribosomal structure proteins and ribosomal RNAs [18],[32]. Hence, it is striking that the mRNA expression of 61 ribosomal structural proteins out of 82 ribosomal structural protein genes was decreased upon MYC inactivation and further decreased upon MYC reactivation in bone tumor (see Figure 4A and Table S3 for results of GO term analysis). To validate that these genes expression did change, we performed quantitative real-time PCR of 11 ribosomal structural proteins in osteosarcomas (Figure 4B). Moreover, we found that the same ribosomal structural proteins also changed upon MYC inactivation in our conditional model of lymphomas [4] (Figure 4B). We then examined if the decreased expression of ribosomal structural proteins associated with changes in rate of protein synthesis. We found that the protein synthesis rates were decreased in both bone tumor and lymphomas upon MYC inactivation (Figure 4C). Furthermore, the protein synthesis rate remained lowed upon MYC reactivation in bone tumor (Figure 4C).
MYC has been shown to regulate the gene expression of a multitude of genes [3], [15], [21]–[26], [33]–[35]. To examine if these genes changed in gene expression upon MYC inactivation and reactivation, we used two approaches. First, we retrieved the mouse homologs of MYC target genes listed in www.myc-cancer-gene.org, a collection of most of the published MYC target genes in different organisms and tissues (total of 1697 MYC targets)[16]. In osteosarcomas, 71 of the published MYC targets are permanently induced and 52 of the published MYC targets are permanently repressed upon MYC inactivation and reactivation (p<0.01) (Figure 5, see PR and PI). Second, we examined direct MYC target genes identified as defined by several recent publications [33]–[35]. Interestingly, only 7–11% of these identified direct MYC target genes exhibited sustained changes upon MYC inactivation in osteosarcoma (Figure 6 and Table S6).
An important recent report suggests that MYC binding to promoters is regulated by the chromatin structure at these gene loci [36]. Recently, we have shown that MYC inactivation is associated with global changes in chromatin structures [14]. Thus, it seemed that a possible explanation for the permanent changes in gene expression that we observed (Figure 3) is that the ability of MYC to bind to specific gene products is perturbed by changes in chromatin structure. To address this possibility directly, we used ChIP to examine MYC binding to E-box sequences of target genes in MYC activated and MYC reactivated conditions for osteosarcomas.
We specifically examine three groups of genes: the ribosomal structural proteins (Figure 4), the PR group (Figure 6, [35]) and the RR group genes that were identified previously as direct MYC targets before (Figure 6, [35]). A total of 168 E-box regions were examined by ChIP. As a control, we performed ChIP for osteosarcoma in the MYC OFF condition (Table S7). Binding of MYC to E-box regions is shown as the percentage of DNA brought down by ChIP for the MYC ON versus the MYC reactivated conditions (Figure 7). Note, that upon MYC reactivation the majority of ribosomal structural genes exhibited decreased MYC binding to E-boxes relative to the MYC ON condition (31 out of 41 data points fall below the line of X = Y, p-value = 4.34×10−4). Similarly, the majority of the genes with the PR pattern of gene expression exhibited a significant decrease of MYC binding to E-boxes relative to the MYC ON condition (42 out of 60 data points fall below the line of X = Y, p-value = 0.0016) when MYC was reactivated (Figure 7 and Table S7). In contrast, the group of genes that exhibited the RR pattern of gene expression exhibited no particular increase or decrease in MYC binding to E-boxes compared with the MYC ON condition (33 out of 67 data points fall below the line of X = Y, p-value = 0.4). Our results support the possibility that the permanent changes in gene expression upon MYC inactivation can be explained in many cases because of a change in the ability of MYC to bind to specific promoter loci.
To determine if the gene signature we identified would also be seen in another tumor model system, we compared our microarray data from MYC-induced osteosarcoma with a previously reported microarray data set from a MYC-induced pancreatic tumor model to identify a common expression signature for MYC-induced tumorigenesis [25]. In the published report, MYC-ERTAM was expressed specifically in β-cell pancreatic tissues with MYC-on for 2, 4, 8, 24 hours, and 21 days (referred as tumorigenesis arrays in the published paper), and MYC off in pancreatic tumors for 2, 4, and 6 days (referred as tumor regression arrays in the published paper). MYC activation induced pancreatic tumors and MYC inactivation resulted in tumor regression through apoptosis [7]. cDNA from these samples was applied to oligo arrays from Affymetrix [25]. As previously suggested in the paper, we assumed that genes were induced (repressed) upon MYC activation and repressed (induced) upon MYC inactivation were potentially important for MYC induced tumorigenesis.
We first used the StepMiner algorithm was applied to the raw data generated from these published experiments to obtain lists of genes that increase (or decrease) in expression upon tumorigenesis and decrease (or increase) in expression upon tumor regression (Figure 8 and Table S8). After StepMiner analysis, 196 and 65 unique probes were identified as induced and repressed genes respectively, which are associated with MYC-induced tumorigenesis. The osteosarcoma data set was filtered via the induced gene list or the repressed gene list generated from the pancreatic tumors. Then, we applied StepMiner analysis to identify genes that are permanently repressed or permanently induced with a p-value<0.01. By comparing microarray data from two independent MYC conditional tumor models, we found a common gene signature with 42 genes associated with MYC-induced tumorigenesis (Figure 9). Among the list of genes, there are 34 unique genes positively correlating with MYC-induced tumorigenesis and 8 unique genes negatively correlating with MYC-induced tumorigenesis in mice (Figure 9).
MYC overexpression has been implicated in the pathogenesis of many types of human cancer, in particular, hematopoietic tumors [1]. To see if the gene signature we defined in murine tumor models was predictive of genes whose expression was strongly correlated with MYC between MYC and human homologs in human lymphomas, we retrieved all publicly available human microarrays (n = 7,171) in Affymatrix U133A platform. Then, we classified the expression level of each gene on each array as “low” or “high” relative to a threshold using Boolean analysis ([29] and Sahoo et al. RECOMB 2007 in press, see Figure 10A). We found that MYC expression is “high” in human lymphomas (204 out of 221 lymphoma cases ignoring the “intermediate” values, see Figure 10B). Figure 10B shows the gene expression scatter plot of MYC and RPS2, which are both highly expressed in lymphoma arrays (total of 273 lymphoma microarrays are highlighted with red color). We then examined to see if the expression of MYC-associated genes identified above (Figure 4 and 9) are “high” or “low” in more than 95% of the lymphoma microarrays.
The Boolean analysis identified that the expression of both small and large ribosomal structural proteins is high in human lymphomas (Figures S3 and S4) as was observed in murine osteosarcomas and lymphomas (Figure 4). We further investigated if the expression of human homologs of the common gene signature from the murine microarray data is “high” or “low” in human lymphomas. 63 unique probes from the induced list (Figure 9) and 9 probes from the repressed list (Figure 9) were found in the U133A format (see Table S9). We found 14 out of 63 probes correlated with the human arrays. Genes whose expression was “high” in more than 95% of human lymphomas, whose gene names include: BZW2, H2AFY, SFRS3, NAP1L1, NOLA2, UBE2D2 and CCNG1 (p = 4.07×10−5, Figure 11). From the repressed list of genes 4 out of 9 probes had low expression in more than 95% of the human lymphomas, whose gene names include LIFR, FABP3 and EDG1/HEXIM1 (p = 0.03, Figure 11). We have listed al the genes identified and their associated functions (listed in the Swiss-Prot data base) (Figure 11). Many of these genes have functions that could account for MYC activity. Notably, CCNG1, LIFR and EDG1/HEXIM1 are involved in cell cycle or signaling pathways. H2AFY and NAP1l1 are involved in modulating chromatin structures. SFRS3 and NOLA2 are involved in mRNA and rRNA processing. BZW2, UBE2D2 and FABP3 are involved in metabolism such as protein or fatty acid synthesis.
Finally, we validated our results obtained by microarray analysis through quantitative real-time PCR (Figure S5). Moreover, we found that these identified genes exhibited similar patterns of changes in gene expression upon MYC inactivation in our model of MYC-induced lymphoma (Figure S6). Therefore, we have identified a subset of MYC regulated gene products that are highly correlated with the ability of MYC to maintain tumorigenesis.
MYC target genes have been implicated in a multitude of biological functions [16]. Many additional potential MYC targets have been identified through microarray analysis [3], [15], [21]–[26],[37]. However, it has not been easy to discern which if any of these genes are involved in the ability of MYC to initiate or maintain tumorigenesis. We have combined microarray analysis of two conditional transgenic model systems and a human comparative Boolean analysis to determine which of these identified genes most strongly correlated with MYC expression from total of 273 datasets of human lymphoma microarrays in U133A format. We also utilized ChIP to demonstrate that a large number of the genes that were permanently suppressed upon MYC inactivation exhibited changes in the ability of MYC to bind to their promoter loci. Thus, we identified a gene signature strongly correlated with the ability of MYC to maintain tumorigenesis. Our results have possible implications for why MYC induces tumorigenesis in specific cellular contexts.
To identify this gene signature, we utilized our conditional transgenic model system of MYC-induced osteosarcoma in which we have previously shown that upon MYC inactivation tumors permanently lost the ability of MYC to induce tumorigenesis [6]. Thereby, we defined an initial gene signature consisting of 2,793 unique probe sets of genes that included genes whose expression was permanently changed (Figure 3). This gene signature includes gene products that have been already implicated as MYC targets (Figure 5). Most notably, ribosomal structural proteins were strongly correlated with MYC-induced tumorigenesis in murine osteosarcomas, lymphomas (Shachaf CM et. al. submitted) and in human lymphomas. These results suggest that the ability of MYC to induce ribosomal gene products is important to its ability to initiate and maintain tumorigenesis.
Our results are consistent with a multitude of evidence suggesting that MYC can regulate ribosomal gene expression [15]. In Drosophila, the biological connection of MYC and ribosomal structural proteins can also be seen in the small cell-size phenotypes of both MYC mutants and ribosomal structural protein genes mutants [38]–[40]. MYC globally regulates protein synthesis through regulating expression of ribosomal RNAs, tRNAs, RNA helicases, and translation elongation factors [18],[41]. Notably, it had been shown that rate of protein synthesis was increased 3-fold in MYC-overexpressing fibroblasts compared to MYC knockout fibroblasts [42]. We confirmed that the inactivation of MYC in tumor cells resulted in a reduction of both ribosomal protein gene expression and rate of protein synthesis in murine tumor models (Figure 4). Ribosomal genes could play important function in influencing protein translation and thus in this manner influence the ability of MYC to function as an oncogene. In this regard, it is notable that a recent study in Zebra fish identified some ribosomal protein genes as tumor-suppressors [43]. Nevertheless, it is not clear how ribosomal structural protein genes function as tumor-suppressors during tumorigenesis.
Interestingly, changes in the gene expression of ribosomal structural proteins, although observed in both our model of MYC induced osteosarcoma and lymphoma, were not seen in a model of pancreatic islet cell tumors (Figure 4, 8, and [25]). Thus, it is possible that ribosomal protein genes expression play a role MYC-induced tumorigenesis only in specific types of cancer. We are reassured of the likely importance of ribosomal gene products in MYC associated tumorigenesis for we were able to confirm that MYC and ribosomal structural proteins are highly correlated in human lymphomas (Figure S4 and S5). It remains to be directly determined if these ribosomal genes are playing a role in MYC induced tumorigenesis.
Genes that we identified as most strongly correlated with MYC-induced tumorigenesis (Figure 9) in mice are involved in diverse biological processes such as transcription regulation, RNA processing, proliferation, fatty acid transport and cell signaling (Figure 11). Furthermore, some of the genes identified have been previously implicated in tumors or oncogenic signaling pathways. BLMH has been previously shown to be a MYC target [44]. UBE2d2 has been implicated as a target of the WNT signaling pathway in a microarray experiment [45]. NAP1l1 has been shown to be a tumor marker for colon cancer [46]. TRIP13 expression was highly elevated in tumor tissues [47]. Altered regulation of CCNG1 has been observed in breast cancer [48]. High expression of NOLA2 has been seen in squamous cell lung cancer [49]. Interestingly, anti-tumor effects have been observed for genes with expression reversely correlated with MYC. FABP3 has been proposed as tumor suppressor in breast cancer [50]. EDG1 has been shown to be an inhibitor for breast cancer growth [51]. Our data now suggest that BZW2, H2AFY and SFRS3, which function in translation initiation [52], chromatin structure [53], and mRNA splicing [54], respectively, may also be involved in tumorigenesis.
We were able to utilize our MYC conditional tumor models as tools to uncover genes that are strongly correlated with tumor maintenance. However, we recognize that it is very unlikely that any of the individual genes we identified are sufficient alone to explain the ability of MYC to initiate or maintain tumorigenesis. Rather it is highly likely that it is a constellation of gene expression changes that are responsible for the ability of MYC to maintain tumorigenesis.
We can now offer a possible explanation for why the brief inactivation of MYC can result in the permanent loss of the ability of MYC to sustain tumorigenesis [6]. MYC inactivation appears to result in permanent changes in the ability of MYC to function as a transcription factor (Figure 12). Recently, we have shown that MYC inactivation induced chromatin modifications associated with cellular senescence [14]. The particular structural state of chromatin has been shown to influence the ability of MYC to bind to specific promoter loci [36]. Indeed, our results illustrate that upon MYC inactivation there were permanent changes in the ability of MYC to bind to the promoters of specific gene loci (Figure 7). It remains to be determined the mechanism of these changes in chromatin structure. One possibility is that MYC itself is contributing to changes in chromatin structure through global changes in chromatin modifications, which seems an attractive possibility based upon the work from many laboratories [14],[20],[55]. Regardless of the mechanism, our results point to the fact that the genes that MYC can regulate are different in different cellular contexts and that this appears to have a direct bearing on when MYC overexpression results in a neoplastic phenotype. We note that we could not explain all of the permanent changes in gene expression based upon differences in MYC binding to promoter loci. Thus, it is likely there are additional mechanisms by which MYC's ability to regulate gene expression has been altered.
One of the biggest challenges in understanding how MYC contributes to tumorigenesis has been to address the conundrum that MYC has both direct and indirect influence on the expression of so many different genes and these genes are involved in a multitude of biologic functions. Many of these genes may not be relevant to how MYC overexpression contributes to tumorigenesis. Here we have illustrated by using a defined transgenic mouse model that exhibits conditional tumorigenesis such that upon MYC inactivation tumor cells permanently loses a neoplastic phenotype that we can define a specific gene list that is specifically correlated with MYC's ability to maintain tumorigenesis. To perform this analysis we combined two novel methods of gene expression analysis, the StepMiner and the Boolean analysis, as a powerful strategy to perform an unbiased comparative analysis of microarray data from conditional MYC-induced tumor models and all the available published human data with Affymetrix U133A format. Our strategy may be generally useful for the identification of gene signatures associated with the ability of specific oncogenes to initiate and sustain tumorigenesis and the identification of potential new therapeutic targets for the treatment of cancer.
Osteosarcoma-derived cell line 1325 [6] were cultured with DMEM medium supplemented with 10% FBS, 1% Pen/Strep, L-Glutamine, and non-essential amino acids (Invitrogen). Lymphomas were cultured with RPMI medium supplemented with 10% FBS, 1% Pen/Strep, L-Glutamine and 3.96×10−4% of 2-mercaptoethanol (Sigma). 20ng/ml of doxycycline was added to the medium for inactivating MYC expression. Seven times, each time with 20 mls of PBS, was applied to cells to completely remove doxycycline in the medium.
For rate of protein synthesis, lymphoma-derived cell line 6780 [14] or bone tumor cell line 1325 grown in complete medium with or without doxycycline were rinsed with PBS and then replenished with DMEM (with or without doxycyline) without methionine and cysteine (Invitrogen), containing 10% dialyzed fetal calf serum (Invitrogen), 1% Pen/Strep, L-Glutamine. One hour later, cells were labeled with 30 μCi of EXPRE35S35S (PerkinElmer) per plate for 60 minutes and then washed with PBS. Cells were lysed and TCA precipitation was applied to determine the incorporation of radiolabeled amino acids. Aliquots of cell lysate were used for protein determination by DC Protein Assay (Bio-Rad). The protein synthesis rate was calculated as TCA-precipitable counts per minute divided by micrograms of protein in the same sample.
cDNA were synthesized by Superscript II (invitrogen) followed by manufacture's protocol. Real-time PCR for human c-MYC (probes and primers from Applied Biosystems) and mouse GAPDH [56] were performed in ABI PRIZM analyzer. Sequences for primers for quantitative real-time are listed in Table S8.
Mouse cDNA microarrays were produced at Stanford Functional Genomic Facility. cDNA labeling and hybridization were followed as previously described [57]. Briefly, mRNA from bone tumor cells were extracted by Trizol (Invitrogen) based on the protocol provided by the manufacturer. 30 μg of total RNA from bone tumor and reference RNA generated by pooling RNA from various mouse tissues were used for each microarray experiment. cDNA from bone tumor cells was labeled with Cy5-dUTP and reference cDNA was labeled with by Cy3-dUTP (Amershan) after reverse-transcription. Labeled cDNAs were concentrated by Microcon YM-30 (Millipore) before hybridizing with microarrays for 16 hours at 65°C. After hybridization, microarrays were washed and spin dry before scanned on the GenePix 40000B Array Scanner (Axon). Raw array images were analyzed using the GenePix 5 software (Axon). Microarray data was then submitted to the Stanford Microarray Database (SMD) for normalization. Data after normalization was then applied with the StepMiner algorithm to identify changes in gene expression.
The StepMiner fits step functions to time-course microarray data and provides a statistical measure of the goodness of fit [29]. The steps are placed between time points at the sharpest change between low expression and high expression levels, which gives insight into the timing of the gene expression-switching event. Mathematically, steps are placed at a position that minimizes the sum of square error and an F-statistic with appropriate degrees of freedom is used to produce a p-value for the goodness of fit. The StepMiner automatically characterizes the genes in to five different groups: Up, Down, Up-Down, Down-Up and Other [29]. The genes are primarily sorted in ascending order according to the timing of their change and secondarily sorted in ascending order according to their p-values.
ChIP was performed based on the protocol provided in the kit with some modifications (ChIP assay kit by Upstate Biotech). Briefly, bone tumor cells were grown on the condition described above with (MYC OFF and MYC reactivated conditions) or without (MYC ON condition) doxycycline (20ng/ml). 48 hours treated with doxycycline, cells were either harvested (as MYC OFF condition) or extensively washed with PBS (see above) to remove doxycycline in the medium. 48 hours after washing, cells were harvested (as MYC reactivated).
Formaldehyde (Fisher) was added to the medium to a final concentration of 1% for cross-linking at 37°C for 10 minutes. Cross-linking was stopped by adding glycine to a final concentration of 0.125M. Cells were washed with cold PBS containing protease inhibitors (1mM PMSF, 1 μg/ml aprotinin and 1 μg/ml pepstatin A) and pelleted by centrifugation. Cell pellets were then lysed in SDS lysis buffer (1% SDS, 10mM EDTA, 50mM Tris, pH 8.1, with proteases inhibitors mentioned above). Cells were sonicated with a Branson 250 sonicator at a power setting of 3 for 3 times with 10 sec for each sonication and the cells were cooled down with ice for 1 min between each sonication. This condition of sonication yielded genomic DNA fragments with a size about 100–600 base pairs. Samples were then immunoprecipitated with c-MYC antibody (2 μg of N262 from Santa Cruz Biotech) followed the protocol provided by the kit (Upstate Biotech). DNA samples from the ChIP experiments were applied for quantification by Real-time PCR (ABI PRISM 7900 HT) with SYBR green. Promoter sequences (−2000 to +2000 relative to the transcription start sites) of murine MYC targets were retrieved from UCSC genome browser and primers flanking the E-box were designed by Primer3 (http://frodo.wi.mit.edu/) (Table S10).
Data from 7,171 publicly available raw Affymetrix U133A human microarrays were collected from the Gene Expression Omnibus (GEO) [58] and normalized together using the RMA algorithm [59],[60]. Thresholds were assigned for each probe set by first sorting the expression values for that probe set on all arrays in ascending order, and then fitting a step function to the data using the StepMiner. This approach places the threshold cutoff at the largest jump from low values to high values. In the case where the gene expression levels are evenly distributed from low to high, the threshold cutoff tends to be near the mean expression level. If the assigned cutoff for a gene is t, expression levels above t + 0.5 are classified as “high,” expression levels below t−0.5 are classified as “low,” and values between t −0.5 and t+0.5 are classified as “intermediate” (Sahoo et al. RECOMB 2007 in press). Two hundred and seventy three different human Lymphoma microarray experiments were identified using a simple string search “Lymphoma” in the GEO description of the experiment. Genes that are “high” or “low” in more than 95% of the Lymphoma experiments were automatically discovered. Human homologs of genes which were associated with MYC-induced tumorigenesis in mice were selected for this manuscript. |
10.1371/journal.pntd.0003587 | Head and Neck Mycetoma: The Mycetoma Research Centre Experience | Mycetoma is a unique neglected tropical disease which is endemic in what is known as the “mycetoma belt”. The disease has many devastating impacts on patients and communities in endemic area and is characterised by massive deformity, destruction and disability. Mycetoma is commonly seen in the foot and hand and less frequent in other parts of the body. Mycetoma of the head and neck is a rarity and is associated with high morbidity and even mortality if not treated early. In this communication we report on 49 patients with head and neck mycetoma followed up at the Mycetoma Research Centre in Khartoum. Most of the reported patients had actinomycetoma and the majority were young adult males from mycetoma endemic areas in the Sudan. Most of them were students, farmers and workers. Prior to presentation the majority had long disease duration and the cause was multifactorial. Advanced disease with massive lesion, deformity and disability was the common presentation. There was no obvious history of local trauma, familial tendency or other predisposing factor identified in this group of patients. MRI and CT scan were the most accurate diagnostic tools to determine the disease extent. The treatment outcome was rather poor and characterised by low cure rate, poor outcome and high follows-up dropout. Such a gloomy outcome calls for structured and objective health education programs.
| Although head and neck mycetoma is a rare disease entity yet it is a dreadful disease for the patient, the family and the treating physician. It is potentially fatal and the most challenging to treat. The current study highlighted that, most patients were young adult males, from rural areas of the Sudan and of low socioeconomic status. The lack of medical and health facilities, financial constrains and compounded by poor health education in endemic areas meant that most of the studied patients presented late with advanced disease. The diagnosis of mycetoma in the studied population was confirmed by several imaging techniques; MRI, CT scan and radiography and tissue diagnosis by histopathology and cytology techniques. The treatment outcome was rather unsatisfactory. The cure rate was low, the dropout rate was high and the disease was associated with high morbidity. Structured and objective health education programmes in the endemic areas is important to encourage patients to seek medical advice early in the course of the disease particularly that there is no clear preventive or control measurement in mycetoma.
| Mycetoma is one of the neglected tropical diseases, characterised by massive deformity, disability and can be fatal if not managed properly and timely [1–3]. It is a chronic, specific, granulomatous, progressive subcutaneous inflammatory disease that spreads to involve the skin, deep structures and bones [4,5]. The disease is caused by true fungi or by certain bacteria and hence it is usually classified into eumycetoma and actinomycetoma, respectively [6,7]. Madurella mycetomatis is the commonest eumycetoma causative agent while Streptomyces somaliensis and Nocardiae are the common causative organisms for actinomycetoma [8–10]. Mycetoma has a definite geographic distribution and it is endemic in what is known as the “Mycetoma Belt” that includes Sudan, Senegal, Somalia, South India, South America and Mexico; however, it is reported in many other countries [11–17]. The infection usually progresses slowly over many years and it is commonly painless and that may contribute to the late presentation of many patients [2,18]. The painless subcutaneous mass, multiple sinuses and discharge with grains is distinctive of this infection [1] Young adult males in the age range 20–40 years are more frequently affected [2,4]. Farmers, workers and students are affected most but no occupation is exempted [2,5].
The diagnosis of mycetoma is tedious and several tools are required to reach a proper diagnosis. These tools include imaging techniques such as radiography, ultrasonography, CT, MRI [19–21], molecular techniques such as PCR [22], serodiagnosis as ELISA, CIE [23, 24] as well the classical grain culture and histopathological diagnosis [23]. Although the disease can be diagnosed clinically this is not accurate and can be misleading.
Early lesions are amenable to medical and surgical treatment with good prognosis [24,25]. Generally, actinomycetoma responds to medical treatment in the form of combined antibiotics while eumycetoma requires both antifungal and surgical excision [26,27]. Late and advanced disease is difficult to treat, has poor prognosis and is associated with high recurrence and amputation rates [28]. Currently there is no preventive or control measurements as the route of infection, susceptibility and resistance to the infection are still an enigma and hence health education is essential to avoid the disease and its high morbidity and complications. Mycetoma of the head and neck region is a rarity, patients commonly present with massive lesions and is associated with poor prognosis and can be fatal. In this communication, the Mycetoma Research Centre of the University of Khartoum experience of managing 49 patients with mycetoma of the head and neck region is presented.
This descriptive, cross-sectional hospital based study was conducted at the Mycetoma Research Centre (MRC), University of Khartoum, Khartoum, Sudan. The study included 49 patients with confirmed head and neck mycetoma seen in the period January 1991 and October 2014. The diagnosis of mycetoma was confirmed by careful interview, meticulous clinical examinations and standard investigations.
The investigations included fine needle aspiration for cytology (FNA), histopathological examination of surgical biopsies using different staining techniques and grains culture in various media. The common sero-diagnostic test used was counter-immuno-electrophoresis. Different imaging techniques were used and that included radiography of the affected sites in at least in two views: anterio-posterior and lateral, lesion ultrasound examination, and in some patients MRI and CT scan. The electronic patients’ notes were carefully and meticulously reviewed.
The study ethical clearance was obtained from Soba University Hospital Ethical Committee, it waived the need for consent.
Statistical analysis was conducted using SPSS computer programme. Data was summarized as percentages for categorical variables and mean ± standard error of the mean (SEM) and median for continuous variables.
The 49 studied patients with confirmed head & neck mycetoma constituted 0.76% of the total MRC patient population seen during the study duration. In the present study, 33 patients (67.3%) had actinomycetoma and 16 (32.7%) had eumycetoma. There were 39 males (79.6%) and 10 females (20.4%). Their ages ranged between 9 and 67 years with a mean age of 27.9 ± 14.7 years. 31(63.3%) of the patients were under 40 years-old at presentation and 23(46.9%) were in the age group 1–20 years. Only five patients (10.2%) were more than 50 years of age at presentation.
In this study, there were 14 students (28.6%), 11 workers (22.4%) and 10 farmers (20.4%). Due to prolonged illness and disability, seven patients (14.3%) were unemployed. There were five housewives (10.2%) in this population.
The majority of the patients, 27(55.1%) were from central Sudan; AL Jazeera State 10(20.4%) and Sinnar State 7(14.3%). There were eight patients (16.3%) from Kordofan States, seven patients (14.3%) from Khartoum State, five patients (10.2%) from Kassala State, four patients (8.1%) form Darfour States and three patients (6.1%) from the White Nile State.
The disease duration at presentation ranged between one and 40 years with a mean duration of 11.23 ± 19.7 years. The majority of the patients 33(67.3%) had mycetoma for less than 10 year and 16(32.7%) of them had the disease for less than one year. Only three patients (6.1%) had the disease for more than 30 years.
Thirty patients (61.2%) had history of discharge contained grains and the colour of the grains was yellow (38.8%), black (24.5%) or white (8.2%). Pain at the mycetoma site, was not a frequent symptom among the study population; documented in only 11 patients (22.4%). Only fourteen patients (28.6%), recalled history of local trauma at the mycetoma site and three patients could not recall such a history. Concomitant other illness was documented in only two patients (4.1%). Four patients (8.2%) had family history of mycetoma.
The majority of patients, 36(73.5%) had recurrent disease and underwent previous surgical excisions; 23 patients (46.9%) had one surgical excision, five patients (10.2%) underwent two surgical excisions, six patients (12.2%) had three previous surgical excisions while two patients (4.1%) had more than three surgical excisions. The type of anaesthesia used ranged between general (41.8%) and local (58.2%).
Different parts of the head and neck were involved which included the frontal (n = 12), occipital (n = 5), parietal (n = 1) and temporal region (n = 1). Multiple skull bones involvement was documented in 12 patients, (Table 1). Five patients had combined frontal and parietal and/ or temporal bone involvement. Two patients had combined occipito-temporo-parietal bones affection. One patient had massive sphenoid, ethmoid, maxillary, nasal bones, anterior cranial fossa, temporal, frontal and occipital bones and supra-orbital areas. One patient had infra-temporal fossa mycetoma extending to the nasopharynx involvement. Two patients had base of the skull and occipital mycetoma with cervical region extension. The orbit was involved in two patients. The upper eye lid, buccal cavity and cheek were affected in one each. Ten patients (20%) had cervical mycetoma. Four patients (4%) had intracranial lesions (Figs. 1, 2, 3).
The mycetoma lesions were classified according to their sizes into small (less than 5 cm), moderate lesion (5–10 cm) and massive lesion (>10cm). The study showed that, 20 patients (40.8%) had massive lesions at presentation while 12 patients (24.5%) had small lesions. At presentation, 30 patients (61.2%) had lesions with sinuses; they were active in 15 patients (30.6%), healed in six patients (12.2%) and nine patients (18.3%) had both active and healed sinuses.
Grains discharged through the sinuses were not detected on clinical examination in 37 patients (75.5%) while in 12 patients (24.5%) grains were detected. Local hyper-hydrosis at and around the mycetoma lesion was detected in one patient (2%).
Regional lymph nodes enlargement was detected in six patients (12.2%). Dilated tortuous veins proximal to the mycetoma lesions were not detected in the present. One patient presented with massive intracranial eumycetoma with minimal skin and subcutaneous affection.
At presentation 30 patients had skull and cervical X-Ray examination in at least two views and that showed normal findings in eight patients (16.3%), soft tissue mass in 10(20.4%), periosteal reaction in one patient (2%), bone destruction in five patients (10.2%) and in six patients (12.2%) a combination of these findings were detected (Fig. 4).
Ultrasound examination of the mycetoma lesion was performed in 11 patients (22.4%). This showed evidence of eumycetoma in five patients (10.2%), actinomycetoma in four patients (8.2%) while in two patients (4.1%) no diagnosis was established.
Most of the patients had MRI examination and it showed the skin, subcutaneous, skull and intracranial disease spread with the typical dot-in-circle sign in most of them, (Figs. 5, 6).
FNA for cytology was performed in 21 patients to confirm the diagnosis and it showed evidence of Actinomadura madurae in eight patients (16.3%), M. mycetomatis in six patients (12.2%), Streptomyces somaliensis in four patients (8.1%) and in three patients (6.1%) no grains were not detected.
In this series, 24 patients (49%) had histopathological examinations of surgical biopsies. The diagnosis of M. mycetomatis was established in nine patients (18.4%), Streptomyces somaliensis in 11 patients (22.4%) and Actinomadura madurae in three patients (6.1%). In one patient (2%) no diagnosis was established due to grains absence and the diagnosis was established by FNA.
For actinomycetoma a combination of antimicrobial agents was given and that included streptomycin sulphate and dapsone, or streptomycin and trimethoprim-sulfamethoxazole. More recently, trimethoprim-sulfamethoxazole 8/40 mg/kg/day in cycles for 5 weeks and amikacin 15 mg/kg/day in a divided dose every 12 hours for 3 weeks were administered. The two week interval of amikacin in the five-week cycle is used for renal and audiometric monitoring.
For eumycetoma several antifungal agents combined with various surgical excisions were performed. The common antifungal agents used were ketoconazole and Itraconazole.
All patients were offered follow up appointments but due to various reasons 14 patients (28.5%) were subsequently lost for follow up, five patients (10.2%) were completely cured, and 30 patients (61.2%) had partial cure.
The incidence of mycetoma of the head and neck region is infrequent. Review of the medical literature revealed only few reports on mycetoma in this site [10,29–31], and although Sudan is considered the mycetoma homeland, only few reports on head and neck mycetoma were reported. Lynch in 1964, reported on 1860 mycetoma patients and of these only 18 patients (0.96%) had head and neck mycetoma [15]. Mahgoub in 1977 reported an incidence of 3% of head and neck mycetoma [32]. In 1986, Gumaa and her associates reported on 15 out 400 patients with mycetoma (3.75%) involving the head and neck region. This communication is in line with the fact that, mycetoma at this region is a rarity.
In agreement with the previously reported series, actinomycetoma was the prevalent type of mycetoma in our series and the explanation for this prevalence remains unclear [10,29]. It is possible that the actinomycetes are resilient and able to survive in the extra-paedal areas more than eumyceteces.
Males were predominantly affected in our series and this is in accordance with previous reports from the Sudan [10,11,29]. Again the explanation for this is unclear; however there is suggestion that sex hormones play a role in this predominance [33]. The majority of the reported patients were young adults with a mean age of 27.9 ± 14.7 years and this is a typical age in mycetoma patients [4,10,34]. Students were affected most, and this may be explained by the fact that, young age groups of patients contract the disease more. The study showed that 44.8% of the affected patients were farmers and workers. This is an important finding as the nature of their work puts them in direct contact with the soil on a daily basis and it has been postulated that the soil harbours the causative organisms and these patients are constantly exposed to minor injuries which facilitate the traumatic subcutaneous inoculation of the organisms.
The mean disease duration at presentation among the affected study population is quite long. This may be explained by the painless nature of the disease, the lack of health education, low socio-economic status of the affected patients and lack of medical and health facilities in the endemic regions.
The clinical presentation of patients in this series was typical and in agreement with other reports [1,2,35]. It started gradually at the subcutaneous tissue and progressed to affect the deep structures. It was painless in the majority of patients and that may be an important contributory factor for the late presentation in most patients.
The study showed that 73.5% of the patients had multiple surgical excisions and recurrence and most of them had surgery performed under local anaesthesia. It is well known that incomplete surgical excision performed under local anaesthesia is the major factor leading in recurrence.
At presentation almost half of the patients had massive lesions which is caused by their late presentation and the fact that, most of them had actinomycetoma which is known to be aggressive and can invade the deep structures and bone at an early disease stage [10].
Different skull parts were affected in our series, however, the frontal and occipital parts were affected most. The explanation for this is unclear however these parts are more prone to direct trauma and hence local inoculation of the causative organisms. Rare sites were encountered and this included the eye.
One patient presented with massive intracranial eumycetoma with minimal skin and subcutaneous involvement, again the explanation is unclear but deep inoculation of the infection may provide some explanation.
In the past, the disease extends in the head and neck area was assessed clinically and radiologically by skull and cervical X-rays or by cerebral angiography which is invasive and with many complications. Currently, the use of the MRI and CT scans provided an accurate assessment with minimal complications. Mycetoma has characteristic MRI features which are diagnostic. The MRI can delineate the involvement of the skin, subcutaneous, muscles and bones accurately and can grade the disease and help in planning patients’ management [21].
The present series showed poor treatment outcome, only five patients were cured and this is in line with previous reports [28–30,36]. This low cure rate necessitates the need for more efficient and safe novel drugs for the treatment of mycetoma. The dropout rate (28.5%) in our series is high. The reasons for the high dropout rate are multifactorial and to mention but a few, the patients’ dissatisfaction due to the high cost and the prolonged treatment duration which is commonly more than one year duration, the drug side effects and complications, the patients low socio-economic status, the lack of health education and difficulty to reach the MRC, particularly during rainy seasons. All these can contribute to the poor treatment outcome.
In conclusion, mycetoma of the head and neck region is a serious medical and health problem, is associated with serious complications, low cure rate and high follow-up dropout rate. The route of infection, susceptibility and resistance in mycetoma remains poorly understood and this is compounded by the lack of preventive and control measures. Hence health education may be the only tool to reduce the disease morbidity and mortality.
|
10.1371/journal.pgen.1000328 | Arabidopsis CULLIN3 Genes Regulate Primary Root Growth and Patterning by Ethylene-Dependent and -Independent Mechanisms | CULLIN3 (CUL3) together with BTB-domain proteins form a class of Cullin-RING ubiquitin ligases (called CRL3s) that control the rapid and selective degradation of important regulatory proteins in all eukaryotes. Here, we report that in the model plant Arabidopsis thaliana, CUL3 regulates plant growth and development, not only during embryogenesis but also at post-embryonic stages. First, we show that CUL3 modulates the emission of ethylene, a gaseous plant hormone that is an important growth regulator. A CUL3 hypomorphic mutant accumulates ACS5, the rate-limiting enzyme in ethylene biosynthesis and as a consequence exhibits a constitutive ethylene response. Second, we provide evidence that CUL3 regulates primary root growth by a novel ethylene-dependant pathway. In particular, we show that CUL3 knockdown inhibits primary root growth by reducing root meristem size and cell number. This phenotype is suppressed by ethylene-insensitive or resistant mutations. Finally, we identify a function of CUL3 in distal root patterning, by a mechanism that is independent of ethylene. Thus, our work highlights that CUL3 is essential for the normal division and organisation of the root stem cell niche and columella root cap cells.
| Ubiquitin-mediated proteolysis plays a central role in controlling intracellular levels of essential regulatory molecules in all eukaryotic organisms. This protein degradation pathway has a large number of components, including hundreds of ubiquitin protein ligases (E3s) that are predicted to have regulatory roles in cell homeostasis, cell cycle control, and development. Recent research revealed the molecular composition of CULLIN3 (CUL3)-based E3 ligases, which are essential enzymes in both metazoans and plants. Here, we report that in the model plant A. thaliana, CUL3 modulates the emission of ethylene, a gaseous plant hormone that controls a variety of processes such as fruit ripening and stress response. In particular, we provide evidence that CUL3 regulates root growth by a novel ethylene-dependant pathway. Thus, we showed that CUL3 knockdown inhibits primary root growth by reducing the root meristem size. Finally, we also identified a function of CUL3 in distal root patterning. Indeed, CUL3 function is required for normal division and organisation of the root stem cell niche and columella root cap cells. Overall, our results show that Arabidopsis CUL3 is essential for plant growth and development, not only during embryogenesis but also at post-embryonic stages.
| Regulation of protein stability through the ubiquitin proteasome system (UPS) is now considered as a major mechanism underlying many cellular and organismal processes, such as cell division, DNA repair, quality control of newly produced proteins, developmental and immune defense pathways, and in plants, light and phytohormone signal transduction [1]–[3]. Degradation via the UPS is a two-step process: the protein is first tagged by covalent attachment of ubiquitin and subsequently degraded by a multicatalytic protease complex called the 26S proteasome. The transfer of ubiquitin to a target protein substrate requires an ubiquitin protein-ligase (E3).
E3 enzymes act to specify the substrates and thus they play a key role in the ubiquitylation reaction. Several hundred different E3s have been identified in metazoan and plant genomes, based on specific, commonly shared structural motifs. However, the most intensively studied subclasses of E3s are those of the cullin-RING ligase (CRL) superfamily, which form multi-protein complexes [4]. CRL E3s can be viewed as two functional modules brought together by the CULLIN proteins, acting as molecular scaffolds. The first module forms the catalytic centre and is composed of a RING finger domain protein and an ubiquitin conjugating enzyme (E2). The second module can be considered as the substrate recognition module, in which a specific protein physically interacts with the target substrate.
A series of recent reports has shed light on the molecular composition and function of the CUL3-based CRL E3s (reviewed in [5]). Certain ‘Bric a brac, Tramtrack and Broad Complex/Pox virus and Zinc finger’ (BTB/POZ) domain proteins function as substrate specific receptors in Schizosaccharomyces pombe and Caenorhabditis elegans [6]–[9]. These BTB domain proteins bind CUL3, via the BTB domain and direct substrate specificity through an independent protein-protein interaction domain. The best-documented substrate for the CUL3-BTB ligases thus far, is the nematode MEI-1 protein, which regulates the meiosis-to-mitosis transition of fertilized embryos [6], [8]–[9]. In mammals, CUL3 function is essential and its loss-of-function in mouse produces an arrest during early embryogenesis [10]. Recent data have also implicated vertebrate CUL3 in cell cycle regulation [11] and signal transduction pathways, such as the Wnt-beta-catenin pathway [12]. In contrast to metazoans, the function of the CUL3 orthologs is not essential in budding and fission yeasts [7],[13].
The Arabidopsis genome encodes two CUL3-related proteins, called CUL3A and CUL3B [14]. Disruption of both genes causes embryo lethality [15]–[17], indicating that CRL3s play important functions during early steps of plant development. Moreover a genomic search revealed the existence of about 80 BTB-domain proteins in Arabidopsis belonging to different families depending on additional protein domains, either upstream or downstream of the BTB-domain, such as the meprin and TRAF homology (MATH) domain, the Armadillo repeats (ARM) and the tetratricopeptide repeats (TPR) [15]–[16], [18]–[19]. Protein interaction studies in yeast suggested that CUL3A and CUL3B may form many different CRL3 complexes, but their nature and substrates are still poorly documented in plants.
ETO1, an Arabidopsis BTB-domain protein, controls the stability of ACS5, a member of the 1-aminocyclo-propane-1-carboxylic acid synthases (ACS) that catalyse a rate-limiting step in ethylene biosynthesis [20]. Moreover, ETO1 was found to directly interact with CUL3A, which prompted the authors to propose ACS5 as the first reported substrate for a plant CUL3-based E3. However ETO1 is also able to inhibit ACS5 activity without CUL3, indicating a higher complexity in this regulation [20]. Supporting a function of CRLs in ethylene biosynthesis is the fact that cycles of (de)neddylation seem to play an important role in this process, because Arabidopsis RNAi lines in which the expression of two NEDD8-related proteins, RUB1 and RUB2, are reduced exhibit a triple response and overproduce ethylene [21].
To better characterize the function of CRL3 E3s in Arabidopsis, we identified a hypomorphic mutation in CUL3A, which, when combined with the cul3b null mutation strongly impairs overall CRL3 functions. Here we report a molecular and genetic characterization of this line, with a focus on ethylene biosynthesis and primary root growth.
Previously it was shown that the combined disruption of both Arabidopsis CUL3A and CUL3B genes causes embryo lethality [15]–[17]. To further investigate the function of CUL3 in plants, we searched for additional Arabidopsis T-DNA insertion lines. One line was of particular interest as the T-DNA was inserted at the very end of the CUL3A coding sequence. This mutant allele, called cul3a-3 was further characterised. The T-DNA insertion creates a mutation, in which the last two amino acids of the protein are replaced by an eight-residue peptide (Figure 1A). The cul3a-3 mutant line produces a lower abundant truncated transcript (Figure 1B). Interestingly, the CUL3A protein detected by a specific anti-peptide antibody [18] was not only less abundant, but was also hyper-rubylated (Figure 1C), suggesting that the truncated CUL3A protein is less prone to de-rubylation. It is noteworthy that cycles of rubylation/de-rubylation are important for CRLs activity (reviewed in [22]).
Homozygous cul3a-3 mutant plants are fertile and do not show morphological defects under normal growth conditions. As CUL3A and CUL3B genes are functionally redundant, we generated a double mutant using the previously characterized cul3b-1 knockout line [17]. The double homozygous cul3a-3 cul3b-1 mutant, hereafter called cul3 hypomorph (cul3hyp), was viable, but exhibited several developmental defects. Approximately 10% of homozygous cul3hyp seedlings displayed altered cotyledon phenotypes (Table 1). Some seedlings exhibited a single cotyledon while others, at a lower frequency, had three cotyledons (Figure 1D); seedlings with partially or totally fused cotyledons were also observed (not shown). In addition, the vascular patterning of cotyledons was often abnormal (Figure 1E). In particular, we observed interrupted and freely ending veins. In less than 1% of these seedlings we observed other abnormalities, such as root-less seedlings (not shown). The absence of root meristem was previously revealed in some of the Arabidopsis cul3a-1 cul3b-1 double null mutants that could complete their embryonic development [17].
The subset of cul3hyp seedlings that had normal organ patterning displayed epinastically curled cotyledons and shorter roots when grown in the light (Figure 1F). At a latter stage of development, the most striking phenotype was a reduced rosette size and a delay in flowering (Figure 1G). Consistently, a slight delay in flowering was observed in single cul3a-1 loss-of-function mutant [18].
Overall, our data indicate that Arabidopsis CUL3A and CUL3B are important for plant growth and development, both during embryogenesis and at post-embryonic stages.
Because ETO1 is involved in the ethylene biosynthetic pathway and physically interacts with Arabidopsis CUL3A [20], most likely through its BTB domain, we investigated whether the cul3hyp mutant is affected in ethylene-mediated processes. In accordance with this speculation, etiolated cul3hyp mutants displayed a typical triple response in the absence of ethylene, which is characterized by short hypocotyls, short roots, and exaggerated apical hooks (Figure 2). The phenotype was similar to that of eto1-1, though less severe than the constitutive triple response1 (ctr1-1) mutant. It is noteworthy that the single cul3a-3 mutant displayed a weak triple response. Moreover, when germinated in the presence 5 µM ACC, the cul3hyp mutant was still responsive to ethylene in a root elongation assay (Figure S1).
To better characterize whether the CUL3A and CUL3B genes are involved in the control of ethylene biosynthesis or might also play functions further downstream in the signalling cascade, we first used a pharmacological approach. We found that treatment with 2 µM aminoethoxyvinyl glycine (AVG), which inhibits ACC synthase and hence ethylene biosynthesis [23], significantly reversed the cul3hyp triple response (Figure S1). However, AVG has toxic effects and it inhibited root elongation of wild-type plants even at low concentrations (not shown).
Thus, we undertook a genetic approach and generated triple mutant combinations with ethylene-insensitive or resistant mutants in the ethylene-signalling cascade. In all triple mutants, etr1-1 cul3hyp, ein2-1 cul3hyp and ein3-1 cul3hyp, the triple response observed in the cul3hyp hypomorph was significantly, but not entirely suppressed (Figure 2). We conclude that the triple response phenotype of cul3hyp can be mostly explained by a function of CUL3A and CUL3B upstream of ethylene perception. However, the fact that both hypocotyl and root length were slightly reduced in all three triple mutants compared to their corresponding single ethylene insensitive or resistant mutants indicates that Arabidopsis CUL3 genes act also at other levels, which are ethylene independent.
We generated an eto1-1 cul3hyp triple mutant and found only a slight additive effect on the triple response regarding root growth (Figure S1). These data indicate that the triple response observed in cul3hyp is mainly attributed to a defect in the CUL3ETO1 E3 ligase, but does not exclude the possibility that the two ETO1-related proteins, EOL1 and/or EOL2 [20] may also play some minor roles in this process.
As a next step, we measured ethylene emission between day 3 and day 4 in the etiolated mutant seedlings. Consistently, the cul3hyp mutant accumulated two to three-fold more ethylene gas than did the wild type control (Figure 3A), but significantly less than eto1 seedlings. To further confirm that the triple response observed in the etiolated seedlings is the consequence of CUL3A/B knockdown, we transformed the cul3hyp double mutant with a binary vector spanning a CUL3A genomic fragment [18]. Several transformants were recovered, which suppressed the triple response (not shown) as well as ethylene overproduction (Figure 3A).
Because the triple response of the eto1-1 cul3hyp triple mutant was slightly more severe than in the eto1-1 single mutant (Figure S1), we measured ethylene production in the triple eto1-1 cul3hyp mutant (Figure 3A). There was an increase of about 3-fold in ethylene production as compared to the eto1-1 parent, suggesting that CUL3A/B controls ethylene production by both ETO1-dependent and independent mechanisms.
There was also an increase of about 3-fold in ethylene production in the etr1-1 cul3hyp mutant compared to etr1-1 (Figure 3A). Likewise, ethylene production was increased in the ein3-1 cul3hyp triple mutant compared to the cul3hypparent, despite the fact that ein3-1 does not overproduce ethylene. This data indicates that CUL3A/B acts additively with the autoinhibition control of ethylene biosynthesis.
ETO1 directly interacts with ACS5 to inhibit its activity, but also mediates ACS5 26S proteasome-dependent degradation, most likely via CUL3A/B [20]. Moreover, the cin5-3 mutation, which disrupts the ACS5 gene [24], significantly reduces ethylene production in eto1-1 and partially suppresses its constitutive triple response [25]. Thus, we speculated that the constitutive triple response observed in CUL3A/B knockdown was mainly the consequence of ACS5 protein stabilisation. To address this issue, we first introgressed a transgenic line expressing a Dex-inducible myc-tagged ACS5 [25] into the cul3hyp double mutant background. Due to partial silencing of the ACS5 reporter construct in the cul3 hypomorph, we could not compare directly the myc-ACS5 protein accumulation in cul3hyp and wild-type backgrounds at identical concentrations of dexamethasone (Dex). However, by increasing Dex levels, we could normalize the expression of myc-ACS5 in cul3hyp and compare myc-tagged ACS5 protein half-lives in both genetic backgrounds (Figure 3B). After Dex-induction and subsequent removal, seedlings were incubated in presence of cycloheximide, which blocks de novo protein synthesis, and the myc-tagged ACS5 protein levels were then determined by immunoblot analysis. Whereas ACS5 protein in the wild type background had a very short half-life of ±15 min as previously reported [25], there was no decrease in the level of ACS5 protein after 1 hour. Thus, Arabidopsis CUL3A and CUL3B are involved in the turnover of the ACS5 isoform. Furthermore, we produced the cin5-3 cul3hyp triple mutant and found that cin5-3 suppresses partially the triple response of cul3hyp (Figure 3C). Overall, we conclude that ACS5 is a primary target of CUL3A/B and ETO1 in seedlings, but other ACSs, in particular of the type-2 class, are most likely also degraded by this E3 ligase.
The cul3hyp mutant exhibits a shorter root (Figure 1F) and CUL3A/B genes are essential during embryogenesis for proper patterning of the hypophyseal lineage, important founders of the future root meristem [17]. Furthermore, the role of ethylene on primary root growth was recently emphasized by several reports [26]–[29]. To characterize the function of CUL3A/B during root development, the work focused on primary root growth.
At 11 days post-germination, the elongation of cul3hyp primary root showed a reduction of about 80% as compared to wild type (Figure 4A and Figure S2). The root growth defect was similar to ctr1. ein2-1 and ein3-1 (not shown) significantly suppressed this phenotype, but these mutations were unable to restore wild type root growth. Thus, CUL3A and CUL3B regulate primary root growth via both ethylene-dependent and independent mechanisms.
To better characterise this phenotype, we measured the length of cells in the cortex in the differentiation zone, because it was found that ethylene stimulates auxin biosynthesis and its basipetal transport to the root elongation zone, where auxin inhibits cell elongation [27],[29]. Consistent with such a scenario, we observed a 50% reduction in length of these cells in eto1-1 (Figure 4B). However to our surprise, cortical cell length in the differentiation zone was only marginally affected in the cul3hyp mutant. This suggests that the mechanism by which ethylene inhibits root growth in a CUL3-deficient mutant background is different from that reported for eto1 mutants and ACC-treated wild-type plants [27],[29].
Root growth depends on cell elongation and on cell production rates in the root apical meristem. Therefore we investigated whether CUL3A/B knockdown affects the meristem size and/or activity. Strikingly, we observed that in cul3hyp the meristem size and cell number was reduced compared to wild-type plants (Figure 4C–D). Interestingly, in the ctr1-1 mutant, there was also a significant reduction in both meristem size and cell number. Conversely, in eto1-1, the root meristem size and cell number were even slightly increased in comparison to wild-type plants, suggesting that ethylene overproduction in this mutant can have opposite effects in the root, negatively affecting cell expansion in the elongation zone but positively affecting root meristem size.
Importantly, the meristem phenotype of cul3hyp was suppressed by the ein2-1 and ein3-1 mutations and therefore is dependent on ethylene signalling. We investigated whether this phenotype in cul3hyp was the consequence of reduced cell cycle activity. To this end we introduced the pCYCB1;1::GUS reporter construct [30] into cul3hyp mutant background. The Destruction box of plant B-type mitotic cyclins targets proteins for degradation after mitosis [31] and thus the pCYCB1;1::GUS reporter is a suitable marker to identify cells in G2-to-early M phase. The number of CycB;1:GUS expressing cells was not significantly reduced in the root apical meristem of cul3hyp compared to wild-type plants (Figure 4E). Thus, the most likely explanation of the ethylene-dependent inhibition of root growth is that cells in the cul3hyp mutant prematurely exit the meristem and make the transition to cell expansion.
To get more insights into the mechanism(s) by with CUL3A/B regulate root growth, we determined the expression pattern of both genes in Arabidopsis roots by using promoter-GUS fusions. After short GUS staining, we observed the strongest histochemical pCUL3A::GUS localization in the stele, but also in the distal part of the root (Figure 5A). A similar expression pattern was also observed for the pCUL3B::GUS reporter, although the signal was weaker in the stele. In the cul3hyp mutant, this region revealed clear defects of cell division patterns in the quiescent center (QC), other cells of the root stem cell niche and the columella root cap (Figure 5B–E). Starch granule staining, which marks only differentiated columella cells, showed premature differentiation of the columella root cap initials (Figure 5B). To investigate whether this defect in root meristem patterning is associated with mis-specification of the QC or columella cells, we examined the expression of different markers in the cul3hyp mutant background. In cul3hyp plants expressing QC-specific marker QC46 (Figure 5B) or endodermis and QC marker pSCR::H2BYFP (Figure 5C), we could identify signal in cells of the QC region, even when their morphology or cell number was altered. However, when we used columella-specific markers (Figure 5D–E) and in particular Q1630, which is only expressed in layers C1 and C2 in wild type (n = 21), we observed a different pattern in the cul3hyp mutant plants. The marker was in general expressed in additional columella cell layers (50%, n = 32) or showed a patchy distribution (22%, n = 32).
It was recently reported that ethylene modulates cell division in the QC, which can eventually lead to additional columella cell layers [26]. Therefore we investigated whether the phenotype observed in cul3hyp is dependent on ethylene signalling. However, this was not the case as defects of cell division in the QC and columella remained in the ein3-1 cul3hyp triple mutant (Figure 5B). Moreover, contrarily to the report of Ortega-Martinez et al. [26], we did not observe deregulated QC cell divisions in the eto1-1 mutant. Thus, we conclude that the knockdown of CUL3A/B function in Arabidopsis disturbs distal root patterning by a mechanism that is ethylene independent.
Auxin is involved in distal pattern formation of Arabidopsis roots [32]. To investigate whether auxin signalling is affected in cul3hyp, we introduced the DR5rev-GFP reporter construct into the cul3hyp mutant and monitored GFP expression in the root tip (Figure 6A). The spatial distribution of DR5 expression in the mutant was more narrow and also reduced in intensity in comparison to wild type roots (Figure 6A), indicating that auxin signalling is reduced in the distal part of cul3hyp roots.
PIN-FORMED (PIN) proteins are rate-limiting factors catalysing polar auxin transport [33]–[34]. These proteins are crucial for auxin distribution and as such provide positional information to coordinate plant development. Because auxin signalling was reduced in the cul3hyp distal part of the root, one possibility is that the auxin gradient is disturbed in this mutant. Thus, we introduced into the cul3hyp mutant background different PIN::PIN-GFP reporters, consisting of their endogenous promoters and translational fusions between PIN1, PIN2 and PIN7 proteins and GFP. Whereas PIN1 and PIN2 expression patterns and levels were similar to wild type in cul3hyp (Figure 6B–C), we observed a higher expression level of PIN7 in columella cells (Figure 6D). The ratio of the GFP signal between the stele and the columella tissues was four-to-five times higher in cul3hyp roots in comparison to wild type (Figure 6E). PIN gene expression is regulated at the transcriptional, but also post-transcriptional levels [35]. Thus we performed quantitative RT-PCR assays on PIN7 gene expression on isolated wild type and cul3hyp root tips. In contrast to the PIN7 protein accumulation in columella cells, the PIN7 transcript level in cul3hyp was slightly reduced in comparison to wild type (Figure 7F). Our data indicate that CUL3A/B knockdown induces in the highest CUL3 expression domain PIN7 accumulation, most likely by a post-transcriptional mechanism.
ACSs are rate-limiting enzymes in ethylene biosynthesis (reviewed in [36]). Arabidopsis has nine ACSs, which are subdivided into three different types [37]. ETO1, a BTB domain-containing protein interacts with type-2 ACS proteins and mediates the degradation of at least one of them, ACS5 [20]. Because ETO1 also interacts in a yeast two-hybrid assay with Arabidopsis CUL3A, it was concluded that ACS5 becomes ubiquitylated by a CUL3ETO1 E3 ligase and subsequently degraded by the 26S proteasome [20],[37].
Our data indicate that ethylene production is induced in the cul3hyp mutant, which is consistent with such a scenario. In addition, we found that the half-life of ACS5 is prolonged in the cul3hyp mutant providing a molecular evidence for the involvement of CUL3A/B in the turnover of ACS5. However it is noteworthy that cul3hyp is a weak ethylene overproducer in comparison to eto1-1, cul3hyp produces about six-fold less ethylene than does eto1-1. Thus, it is possible that this difference in ethylene production is the consequence of a longer ACS5 half-life in eto1 in comparison to cul3hyp mutant. This is also consistent with the fact that cul3hyp still keeps some CRL3 activity, whereas eto1-1 is a null allele. However, we also cannot rule out that ETO1, which is still present in the cul3hyp mutant background, inhibits ACS5 protein activity by a process independent of protein degradation. Moreover, BTB-containing proteins themselves are also substrates for these CUL3 complexes [8],[38], which may even lead to a higher ETO1 protein accumulation in the cul3hyp mutant. This data would be consistent with the finding that ETO1 overexpression in Arabidopsis reduces kinetin-induced ethylene production and most importantly that ETO1 inhibits ACS5 activity via a direct interaction [20].
Arabidopsis ETO1 is part of a small gene subfamily containing two other closely related BTB-domain proteins, called EOL1 and EOL2 [20]. Genetic interactions showed only minor additive effects on the triple response morphology in eto1-1 cul3hyp triple mutant compared to eto1-1, but a strong increase in ethylene production was observed. This suggests that EOL1 and EOL2 are also involved in ethylene biosynthesis, though ETO1 is the main CUL3 receptor in this process. Although it was previously shown that the single eol1 and eol2 mutants grow like wild type in the dark [16], recent data demonstrated that both EOL1 and EOL2 negatively regulate ethylene biosynthesis by directing type-2 ACS proteins for degradation [39].
Is ACS5 the only ACS target of CUL3A/B in Arabidopsis? Based on our data we can say that ACS5 is a primary target, because ACS5 loss-of-function in cin5-3 significantly suppressed the triple response during early seedling development (Figure 3C) and the root growth inhibition (Figure 4A) of cul3hyp. Nevertheless, cin5-3 does not entirely revert the cul3hyp mutant; thus, it is likely that other ACSs are also targeted by CUL3A/BETO1 (Figure 7A). As for ETO1, EOL1 and EOL2 only interact with type-2 ACSs [37],[39],whereas type-1 and type-3 ACSs, which are also degraded by the 26S proteasome [40]–[41], are most likely recognized by another Arabidopsis E3 ligase.
It was previously found that ethylene-insensitive or resistant mutants, such as etr1-1, ein2-1 and ein4-1 produce increased amounts of ethylene [42]–[44], whereas ctr1-1 does not [45]. This suggested a negative feedback mechanism from ethylene perception and signal transduction to ethylene biosynthesis [43]. The finding that the cul3hyp mutation synergistically enhances ethylene production in etr1-1 indicates that CUL3A/B and ETR1 act in parallel and most likely independently in the regulation of ethylene biosynthesis (Figure 7A). This is consistent with previous studies showing that ETO1 also acts synergistically with etr1 to enhance ethylene production [46].
Arabidopsis primary root growth is reduced in a concentration-dependent manner when plants are exposed to ACC or to exogenously applied ethylene and this process is the consequence of down regulation of cell elongation ([47] and references therein). Recent results demonstrated that ACC treatment of wild-type plants positively regulates auxin biosynthesis and distribution in Arabidopsis roots [27]–[29]. Based on different approaches, a model was proposed in which ethylene-stimulated auxin is subsequently basipetally transported to the elongation zone where it inhibits cell elongation [27],[29].
In the present report we provide evidence that the ethylene pathway also acts on root development at a different level (Figure 7B). We show that the inhibition of CUL3A/B activity impairs primary root growth in an ethylene-dependent manner by reduction of root meristem size and cell number. In contrast to eto1-1 or ACC-treated plants [27],[29],[47], no significant effect was observed on cell elongation in the cul3hyp mutant. However, we found a similar root phenotype in ctr1-1 and the double ebf1-1 ebf2-1 [48] mutant backgrounds (this work and data not shown), both accumulating the transcription factor EIN3 leading to a constitutive ethylene response. It is noteworthy that the ctr1-1 mutation affects both cell elongation and root meristem size, indicating that ethylene signalling is involved in both mechanisms. Because the number of mitotic cells in the cul3hyp root meristem was not significantly different from wild type, we conclude that the reduced meristem size is the consequence of the premature transition of cells from the meristem to the cell elongation zone, rather than caused by differences in cell division activity.
An intriguing observation is that in eto1-1, which overproduces ethylene, or in ACC-treated roots ([27] and our data), no decrease on the root meristem size was observed. This suggests that despite the fact that ethylene is a volatile gas, depending on its sites of production and/or perception, ethylene can induce different local responses. Hence, this fits with the evidence that ACS genes in Arabidopsis display distinct expression patterns during plant development [49].
The strong CUL3A and CUL3B expression in the distal part of the root and their involvement in distal root patterning suggest that CUL3A/B proteins control the division and organisation of the stem cell niche and columella root cap cells. By which mechanism(s) do CUL3A/B genes maintain QC and root cap organisation? It was recently reported that ethylene modulates cell division in the stem cell niche [26]. Such an effect of ethylene would be consistent with the role of CUL3A/B in ethylene biosynthesis (see above). However, several lines of evidence argue against this possibility. First, 1 µM ACC treatment of wild type roots revealed no abnormalities in columella cell differentiation and/or tissue organisation [27]. Second, such a phenotype was not observed in the eto1-1 null mutant. Third and most importantly, the ethylene insensitive mutation ein3-1 did not suppress cell division defects in the QC and in columella cells. A possible explanation for the discrepancies between our data and [26] is that the two eto1-11 and eto1-12 alleles used in their studies, which are point mutations, are acting in a dominant negative way to inhibit CUL3A/B activity. Overall, even if a subtle effect of ethylene on stem cell division is not excluded, our data point to an ethylene-independent role of CUL3A/B in organizing stem cell niche and columella (Figure 7B).
Interestingly, we noticed that the DR5::GFP signal in the cul3hyp root tip was weaker and more restricted than in wild type roots, suggesting altered auxin signalling. In addition, we also observed a higher PIN7 protein accumulation in cul3hyp columella cells. One way to connect these observations is a scenario in which PIN7 missexpression leads to the depletion of intracellular auxin from the root tip, resulting in reduced auxin signalling and as a consequence patterning defects. Further experiments need to confirm such a model. Moreover, because PIN7 accumulation in columella cells is under post-transcriptional control and because CUL3A/B genes are strongly expressed in these cells, it will be interesting to address whether the PIN7 protein is a direct target of a still unknown CRL3 complex. Finally, a broader connection between CUL3A/B and auxin may exist, since several developmental abnormalities in the cul3hyp mutant, such as altered number of cotyledons, or defects in root and venation patterning and embryogenesis (this work and [17]) are reminiscent of auxin transport or signalling defects.
The Arabidopsis cul3a-3 (SALK 012973) mutant line has been identified using the web assisted program: http://signal.salk.edu/cgi-bin/tdnaexpress. The insertion site was confirmed by sequencing the T-DNA flanking srquences. The precise location of the T-DNA in the cul3a-3 mutant has been determined by sequencing, showing an insertion after nucleotide 2191 in the last exon of CUL3A. The last 8 nucleotides (TAGCCTAA) of CUL3A are replaced by 26 nucleotides from the left border of the T-DNA (ACATACGGTATCATATTGTGGTGTAA) leading to the addition of 8 amino acids (HIRYHIVV) to CUL3A protein sequence.
All Arabidopsis thaliana lines used are in the Columbia background except for the Q1630 reporter line, which is in the C24 background. The transgenic and mutant lines have been described elsewhere: ein3-1 [50], etr1-1 [51], eto1-1 [43], ctr1-1 [45], ein2-1 [43], cin5-3 [24], proPIN1-PIN1-GFP and proPIN2-PIN2:GFP [52], proPIN7-PIN7:GFP [53], CYCB1-GUSDB [54], Q1630 [55] and PET111 [56]. The CUL3A- and CUL3B-promoter GUS transgenic lines are described respectively in [18],[19].
Arabidopsis thaliana seeds were sterilized with chloral gas or with ethanol method, plated on 1/2 MS medium (1/2 MS salts [Gibco-BRL, Cleveland, OH], pH 5.8, 1% sucrose, and 0.8% agar), stored 2 to 3 days at 4°C in the dark, and then transferred to a plant growth room (21/25°C, 16-h photoperiod). For the triple response assay, surface-sterilized seeds were germinated in the dark on 1/2 MS medium. Plates with seeds were cold-treated at 4°C for 2 to 3 days, exposed to light at room temperature for 2 to 4 h to improve germination, then wrapped with aluminum foil and incubated at 22°C for 3 days in the dark. A minimum of 15 seedlings were scored per mutant by pulling them out of the growth medium, stretching them flat on the surface of another agar plate, taking in pictures, and then quantifying root and hypocotyl lengths using ImageJ (National Institutes of Health; http://rsb.info.nih.gov/ij). For seeds propagation, plants were grown to maturity at 22°C under 16-h photoperiod.
For Northern–blot analysis RNA was extracted from plant material using the Trizol reagent (Invitrogen, Paisley, UK). RNA gel blot analysis was performed with 20 µg of total RNA per lane. Northern blot procedure is described in [31]. 32P-labelled probes were synthesized with the Prime-a-Gene random prime labelling kit (Promega Corporation, Madison, WI, 8USA) using a 900-bp CUL3A cDNA fragment amplified with two gene-specific primers (c3as4 5′-ATGGATTTGGGTGAATCTGT-3′ and c3aS5 5′-CTCGGGGTGACTGCCATA-3′).
For qRT-PCR assays, RNA was extracted from 10 days old root tips (1cm of the root starting from the root tip) using the kit NucleoSpin RNA XS (Macherey Nagel). 1 µg of total RNA were reverse transcribed with High Capacity cDNA Reverse Transcription kit TM (Applied Biosystems). PCR was performed using gene-specific primers in a total volume of 15 µL SYBR Green Master mix (Roche) on a Lightcycler LC480 apparatus (Roche) according to the manufacturer's instructions. The TIP4l and At4g26410 genes were used as internal controls. The relative expression level of PIN7 gene in cul3hyp plants was compared with Col-0 control plants using GenEx Pro 4.3.5. software (MultiD Analyses) after normalization using the At4G26410 cDNA level and averaging over three replicates.
Primer list.
PIN7:TGGGCTCTTGTTGCTTTCA and TCACCCAAACTGAACATTGC
TIP4l: GTGAAAACTGTTGGAGAGAAGCAA and CAACTGGATACCCTTTCGCA
AT4G26410: GAGCTGAAGTGGCTTCAATGAC and GGTCCGACATACCCATGATCC
Transgenic lines used were dex-inducible myc-tagged ACS5 [25]. The coding region of ACS5 was amplified from cDNA of wild type and fused to a 6× myc cassette, and cloned into the binary GVG vector pTA7002 [57]. Plants were transformed with the plasmids by the floral dip method [58], and transformants were selected on MS medium containing hygromycin. T2 seedlings were grown on MS medium containing 10 nM dex and screened for lines that expressed the myc-tagged proteins at low levels in an inducible manner.
Tissues for protein analyses were ground in a denaturing buffer [59] followed by boiling for 5 min. After centrifugation, 20 µg of total protein extracts were fractionated on a 10% SDS-PAGE gel and blotted onto Immobilon-P membrane (Millipore, Bedford, MA, USA). The immunoreactive proteins were detected using peroxidase-conjugated goat anti-rabbit antibodies (Dianova, Chalfont St Giles, Bucks, UK) and ECL Western blot analysis system from Amersham.
Immunoblots were performed using a 5000-fold dilution of anti-Myc monoclonal antibodies from mouse (clone MYC-1A1 Euromedex) and a 5000-fold dilution of peroxidase-conjugated goat anti-mouse IgG (Molecular Probes). Signals were detected by film (within the linear range of detection) using the enhanced chemiluminescence protein gel blot analysis system (Amersham Biosciences). The blot was stained subsequently with Comassie blue to control the loading.
For ACS5 turnover assays, wild type and cul3hyp transgenic seedlings harbouring DEX-inducible myc-ACS5 were grown on MS medium for 8 days at 22°C. Twenty ml of liquid MS medium containing different DEX concentrations (see Figure 3) were poured onto the plates where seedlings were growing and incubated for 4 hours. Seedlings were incubated for different times in 1M cycloheximide after three washes of liquid MS medium. Total proteins were extracted and used for immunoblot analysis.
Seeds were sterilized using chlorine gas sterilization (100 mL bleach +3 mL HCl for ∼4 hours), and seeds were aliquoted in 0.4% top agar to 3 mL MS 1% sucrose media in 22-mL gas chromotography vials. Vials were maintained sterile using autoclaved aluminum foil, for 5 days in 4°C. Vials were incubated in the light for 4–6 hours, then capped and incubated at 22°C dark for 4 days. The vials were capped in the dark at day 3 and incubated to day 4. The accumulated ethylene was measured by gas chromatography as described in [24]. All genotypes are represented by 3 repetitions of 2–3 vials each. Ethylene measured in each vial was then divided by number of seedlings in the vial.
Root seedlings were photographed and their lengths were measured with ImageJ. At least 15 seedlings were processed, and at least three independent experiments were performed, giving the same statistically significant results. Root meristem lengths and epidermal cell lengths were measured on root mounted in chloral hydrate. Images were captured with a Zeiss Axioskop microscope (Carl Zeiss, New York, NY) equipped with a Nikon DXM1200 digital camera (Nikon Instruments Europe, Badhoevedorp, The Netherlands). The number of root meristematic cells was obtained by counting cortical cells showing no sign of vacuolisation. Root meristem length was assessed as the distance between the quiescent center and the first cell with a vacuole. ImageJ was used also for measurements of the length of root cortical cells. At least 10 seedlings were processed in at least three independent experiments giving similar results.
Histochemical GUS staining analyses of the CUL3A/B promoter, the CYCB-GUS and QC reporter lines were done as described in [60]. For the confocal microscopy, roots were visualized using a Leica (Wetzlar, Germany) MZ FLIII fluorescence stereomicroscope equipped with GFP and YFP filters. Propidium iodide (10 µg/mL in distilled water) was used to stain the cell walls of living root cells (red signal). For quantification of PIN7:GFP signal fluorescence ImageJ program was used. The ratio of GFP signal intensity between the stele and the columella was calculated for the wild type and cul3hyp roots. Approximately 15 seedlings/images were examined, in three independent experiments giving similar results.
|
10.1371/journal.pntd.0005708 | Chronic Trichuris muris infection causes neoplastic change in the intestine and exacerbates tumour formation in APC min/+ mice | Incidences of infection-related cancers are on the rise in developing countries where the prevalence of intestinal nematode worm infections are also high. Trichuris muris (T. muris) is a murine gut-dwelling nematode that is the direct model for human T. trichiura, one of the major soil-transmitted helminth infections of humans. In order to assess whether chronic infection with T. muris does indeed influence the development of cancer hallmarks, both wild type mice and colon cancer model (APC min/+) mice were infected with this parasite. Parasite infection in wild type mice led to the development of neoplastic change similar to that seen in mice that had been treated with the carcinogen azoxymethane. Additionally, both chronic and acute infection in the APCmin/+ mice led to an enhanced tumour development that was distinct to the site of infection suggesting systemic control. By blocking the parasite induced T regulatory response in these mice, the increase in the number of tumours following infection was abrogated. Thus T. muris infection alone causes an increase in gut pathologies that are known to be markers of cancer but also increases the incidence of tumour formation in a colon cancer model. The influence of parasitic worm infection on the development of cancer may therefore be significant.
| It is estimated that now 2 billion people currently live with chronic parasitic worm infections. As the incidences of cancer increase worldwide, the importance of these chronic inflammatory conditions on the development of cancer becomes more important. Several bacterial, viral and parasitic infections are already known to influence cancer development but as colon cancer is particularly prevalent worldwide, we wanted to assess the effect of a large intestinal dwelling worm, Trichuris muris (T. muris) on its aetiology. This whipworm is a natural infection of mice and has significant homology to human whipworm. From our studies, we showed that chronic infection alone induced changes in the caecum of the mouse that were comparable to those seen with a well-known carcinogen. In addition to this, T. muris infection was also able to increase the development of adenomas in the small intestine of mutant mice that spontaneously develop tumours. This change was abrogated if a T regulatory cell type was blocked during infection. The T regulatory cell type that arises during infection has been shown to play an important role in protecting the host from damage caused by the parasite and the immune response to it. The present study using the mouse model however, suggests that regulatory T cells can have negative effects, at least in terms of the development of bowel cancer. As so many people live with chronic, regulated parasitic infections, the importance of the parasites in cancer development may therefore be significant.
| Colon cancer is one of the leading causes of death within the western world and prevalence in developing countries has increased in the last decade [1]. There exists a strong link between inflammation and cancer [2]. This is emphasized in the colon where individuals with inflammatory bowel disease are predisposed to the development of colorectal cancer [3–5]. Furthermore, chronic infection and the resultant long-term exposure to inflammatory stimuli heighten the risk of neoplastic change. A number of chronic bacterial, viral and parasitic infections are associated with predisposition to neoplasia; Helicobacter pylori is associated with gastric cancer [6], Hepatitis B and C with liver cancer [7], Clonorchis sinensis with cholangiocarcinoma [8] and Schistosome infection with bladder cancer incidence [9].
Gastrointestinal worms, comprising Ascaris lumbricoides, Trichuris trichiura, Necator americanus and Ancylostoma duodenale species infect over 2 billion people worldwide and account for considerable morbidity and a loss of 5.2 million DALYS [10–12]. Individuals in endemic areas build up chronic infections due to repeated exposure with few people completely resolving infection. This chronic insult on the intestine is associated with intestinal inflammatory changes and it is now well understood that gut dwelling nematodes can manipulate the immune system (reviewed in [13,14]). Indeed, the therapeutic potential of worms in IBD [15,16], allergy [17,18] and inflammatory disease are apparent [19].
Trichuris muris (T. muris) is a natural parasite of mice and is extensively utilised as a laboratory model for the study of human whipworm infection, T. trichiura [20,21]. In susceptible hosts, the persistence of T. muris in the large intestine is characterised by the development of a strong type 1 (Th1) response, dysregulation of epithelial homeostasis and upregulation of inflammatory cytokines [22,23]. The generation of crypt cell hyperplasia is driven by an expansion of the proliferative compartment of the intestinal epithelium and is under immune control. The intestinal pathology associated with chronic T. muris infection closely resembles that seen in Crohn’s disease in humans and in human trichuriasis and is under the control of the regulatory cytokine IL-10 [24,25]. Given the heightened risk of IBD patients to colorectal carcinoma due to intestinal dysplasia and other genetic factors [26] and the global prevalence of intestinal helminth infection, it is clear that the nematode-neoplasia link warrants investigation. Here we assess the effects of a natural model of chronic intestinal helminth infection on the development of intestinal neoplasia.
Experiments were performed under the regulations of the Home Office Scientific Procedures Act (1986), Project licence 70/8127 and subject to review by the University of Manchester Animal Welfare and Ethical Review Body (AWERB). The experiments conform to the ARRIVE guidelines.
Male wild type (WT) C57BL/6 mice aged 6–8 weeks were purchased from Envigo, U.K. APCmin/+ mice were obtained from the Paterson Institute, Christie Hospital, Manchester, U.K. for initial d42 post infection (p.i.) studies, then from Birmingham University for all subsequent studies. C57BL/6 animals were used at 6–8 weeks of age and were housed for 7 days prior to experimentation. Both sexes of APCmin/+ C57BL/6 mice were housed in the same facilities and infected at 12 weeks of age with group sizes of 8–12 mice. All animals were euthanized using a rising concentration of CO2.
MLN and spleen cells were removed, cultured and restimulated for 24 hours under conditions previously described [27].We measured concentrations of TNF-α, IFN-γ, IL-6 and IL-10 in the culture supernatants using a cytokine bead assay (CBA, BD Biosciences, UK) performed according to the manufacturer’s instructions.
ELISA plates were coated with 5 μg/ml of overnight E/S in 0.05 M carbonate/bicarbonate buffer, pH 9.6 and incubated overnight at 4°C. Plates were blocked for 1 hour with 150 μl PBS/Tween-20 (PBST), 3% BSA at room temperature. Eight serial two-fold dilutions of sera in PBST were conducted from 1/20 to 1/2560 and transferred to the ELISA plates (50 μl/well) for 90 minutes at room temperature. Parasite specific IgG1 and IgG2a were detected using biotinylated rat-anti mouse antibodies (Pharmingen, UK and Serotec, UK respectively) diluted in PBST, 50 μl/well for 1 hour at room temperature. Streptavidin peroxidase was added at 75 μl/well for 1 hour and ABTS substrate was added at 100 μl/well. Plates were read after approximately 20 minutes at 405nm on a VersaMax microplate reader (Molecular devices, UK).
Proportions of CD4+, CD25+ and FoxP3+ cells in the MLN and spleen of WT animals were analysed using flow cytometry at day 80 p.i. FITC anti-CD3(ε) in combination with streptavidin-allophycocyanin, (Caltag Laboratories, Burlingame, CA) FoxP3 and a CD25 FL3 (PharMingen) were used for surface marker staining. Cells were analysed using CellQuest Pro software (BD Biosciences, UK). In subsequent APCmin/+ mice experiments, cells were analysed on a MACSQuant (Miltenyi Biotec, UK) using FITC-CD4, PE-CD25 and APC-FoxP3 (BD Biosciences, UK).
Both small and large intestine were removed from animals at autopsy, gently flushed out using saline, slit longitudinally and cut into 2 cm pieces. Sections were pinned out on wax coated petri dishes with the luminal face of the intestine facing upwards, and fixed with 4% formaldehyde for 24 hours. Sections of intestine were stained whilst pinned out with methylene blue to allow the visualisation of tumours (5 minutes at room temperature). For regional analysis of tumour burden in the small intestine, the intestine was divided into 3 equal length sections-from the duodenum to the ileum and labelled SA to SC respectively. Within each region, the intestine was divided into 2 cm pieces, which were assessed for tumour burden. The area of tissue, number and size of tumours was determined using a computer assisted Zeiss Axiohome™ microscope system under x 40 magnification.
Samples of cecum were removed and flushed out using saline. Samples were fixed intact in carnoy’s fixative for 30 minutes prior to storage in 70% ethanol. Tissues were prepared using the gut bundle technique [28]. Tissues were then paraffin embedded using standard histological techniques and 3μm sections were cut.
Sections were stained with haematoxylin and eosin (H&E) to allow the visualisation of apoptotic cells. Such cells are detected on the basis of their morphology using light microscopy, a method that has been used extensively [29–32]. Typically, apoptotic cells appear pink, circular, with crescent shaped nucleus, and are bubbled up out of the plane of focus. TUNEL labelling is another method which has can be used to detect apoptosis in the intestinal epithelium. However, this technique is prone to false positive and false negative results when compared with morphological assessment, as well as failing to distinguish between DNA cleaved by apoptosis and DNA fragments cleaved by other processes [33]. For the purpose of this investigation, therefore, morphological analysis was deemed the most reliable method to use.
Groups of 4 mice were treated by i.p. injection with 10mg BrdU (Sigma, Poole, U.K.) 40 minutes prior to sacrifice. All animals were killed at the same time- within and between experiments to minimise any differences in proliferation attributable to variation in circadian rhythm. Detection of nuclei that had incorporated BrdU was performed by immunohistochemistry, using a monoclonal anti-BrdU antibody (Mas 250b, Harlan Sera Laboratories, Loughborough, U.K.) as described [34]. Sections were analysed by scoring 50 caecal crypts per mouse, 4 mice per group.
Full-length longitudinal sections of crypts were selected for analysis. The blinded scoring commenced with the cell at the mid-point at the base of the crypt, which was designated as position 1 and continued until the crypt-crypt table was reached. This method of scoring allows the generation of statistically valid results [34] and was used to determine the levels of apoptotic and proliferating cells. In this way both the position and overall numbers of apoptopic or proliferating cells in the cecum can be determined.
Area of epithelium was assessed using the computer assisted Zeiss Axiohome™ Microscope system to mark around the area of interest. Analysis was performed on H&E stained sections, 4 mice per group, 3–4 circumferences per mouse. Circumference of the lumen was subtracted from circumference of the muscularis to give the area of epithelium. Individual crypt length and widths were determined using the same system, selecting well-orientated crypts and measuring from the base of the crypt to the lumen for crypt length, and the widest area of the crypt for width analysis. 50 crypts per mouse were measured, 4 mice per group.
Aberrant crypts were detected on the basis of their morphology on H&E stained cross sections of intestine as described [35]. A scoring system was devised to assess the severity of aberrant crypts, aberrant crypt foci [multiple aberrant crypts], epithelial hyperplasia and adenoma formation. Scores were assigned on the basis of number of aberrant crypts, the number aberrant crypt foci (clusters) per circumference as well as the degree of area of epithelium affected. A score of 0 indicates no detectable change and 4 the highest level of severity.
Statistical analysis was performed using Students t test. A value of p<0.05 was considered to be significant.
WT C57BL/6 mice develop a regulated Th1 driven intestinal inflammation during chronic infection with T. muris [36]. Given the strong association with the prolonged exposure to pro-inflammatory cytokines and the development of neoplasia, we were interested to see whether chronic infection was associated with neoplastic change in the intestine of WT mice during chronic infection. Gut pathology was assessed in WT C57BL/6 mice at day 80 p.i. This time point was selected as it ensured that animals had been exposed to inflammation for a long time period. An increase in lamina propria cell infiltrate was detected in infected animals over naive as well as epithelial hyperplasia (Fig 1A & 1B). Assessment of the levels of pro-inflammatory cytokine production revealed that the levels of IL-6, TNF-α and IFN-γ were increased in both the MLN (Fig 1C) and spleen (Fig 1D) at day 80 p.i. Chronic T. muris infection is also known to induce an anti-inflammatory IL-10 response in the host that acts to regulate pathology [25,37,38]. Interestingly, although the levels of the IL-10 increased significantly (p<0.05) in the spleen the levels were only minimally increased in the draining lymph node at the site of infection as reflected by antigen specific recall stimulation of cells from the MLN (Fig 1C & 1D).
Gut dysplasia was assessed using a number of histological markers, including the presence of epithelial cell hyperplasia, measurement of crypt width and the number of both aberrant crypts and aberrant crypt foci (clusters of aberrant crypts). Aberrant crypt foci were first described as lesions consisting of thick irregular (aberrant) crypts in methylene blue stained sections detectable following carcinogen treatment of mice [35,39]. These lesions have since been found to correlate with the degree of colonic neoplasia in humans [40,41], and are hypothesized to serve as biomarkers for colorectal adenoma and cancers [41]. At day 80 p.i. there was a significant increase in the development of crypt width and aberrant crypt formation in infected animals over naïve (Fig 1E & 1F). Additionally, the neoplasia score was increased in infected animals in all 3 categories, i.e. extent of hyperplasia, number of aberrant crypt foci and number of pre-adenomas (Fig 1G).
In order to determine the extent of neoplastic change caused by chronic T. muris infection, the effects seen with infection were compared to those seen when mice had been treated with a carcinogen. AOM is routinely used as a colon specific genotoxic carcinogen [42], therefore animals were treated with AOM at day 0 and 7 p.i. and infection was allowed to progress to chronicity. Using the histological markers for inflammation and neoplastic change used before, it can be seen that changes observed following infection (grey bars in PBS group) were as marked or greater (in the case of aberrant crypt score) as those seen with AOM treatment (black bar in AOM group) (Fig 2A–2D). Interestingly, AOM treatment had no effect on the worm burdens of mice (Fig 2E) and AOM-infected mice had similar histology scores in terms of crypt width, aberrant crypt score and neoplasia score to PBS-infected mice (Fig 2B–2D). AOM-infected groups however had significantly increased crypt width, aberrant crypt score and neoplasia scores when compared to AOM-naïve groups (Fig 2B–2D) further demonstrating the marked changes that T. muris is causing within the intestine. Cytokine production in the MLN of infected animals were similar between groups (S1 Fig) though, interestingly, naïve AOM treated animals produced increased amounts of the proinflammatory cytokines IFN-γ, TNF-α and IL-6.
As T. muris infection is clearly associated with neoplastic change in the large intestine and exacerbated compared to that seen with AOM treatment, we were interested to see whether T. muris had any effect on spontaneous adenoma formation in the well-established model of intestinal neoplasia, the APCmin/+ (Adenomatosis Polyposis Coli) mouse. Both humans and mice with a germ-line mutation in the APC gene have a predisposition to intestinal neoplasia [43]. In mice, Min (multiple intestinal neoplasia) is a dominant trait involving a nonsense mutation in codon 850 of the APC homologue. Loss of APC heterozygosity in these animals results in adenoma formation through out the GI tract [44]. APCmin/+ mice are routinely used as model for human Familial Adenomatous Polyposis (FAP). Although there is a clear genetic basis for adenoma formation in these animals, tumorigenesis is also influenced by a number of environmental factors, including bacterial infection, inflammation and toxic insult [45,46].
Mutant mice were given a low level T. muris infection to ensure chronicity and to mirror the typical parasite burden of naturally infected individuals. Patent infection developed in APCmin/+ mice, characterized by a Th1 dominated immune response and high levels of parasite specific IgG2a (S2 Fig). Animals were infected with T. muris at 12 weeks of age. This age was chosen, as this is the time when adenomas start to form in the colony of APCmin/+ animals used. It is important to note that both the onset and severity of adenoma formation varies between colonies of APCmin/+ mice and that for practical reasons, different colonies were used for day 18 and day 42 studies. In order to minimize variability between experiments, animals were housed in the same facility during experimentation and were age-matched. Analysis of tumour burden at 42 p.i. revealed that persistent T. muris did not promote tumour formation in the large intestine, the niche of the parasite (number of tumours in naïve animals 3.07±0.51 as compared to in infected animals 3.00±0.44). Surprisingly, T. muris infection did promote the development of adenoma formation throughout the small intestine. The number of tumours throughout the intestine was significantly increased at both day 18 and day 42 p.i. (Fig 3A & 3B). Total tumour area and mean tumour area however, were unchanged upon infection (Fig 3C & 3D). Enhanced adenoma formation in the small intestine was greatest in the lower small intestine (SC) (2.5 fold increase at day 42 p.i.), in comparison with the upper small intestine (SA). These changes were apparent at day 18 p.i. but more distinct by d42 p.i. (Fig 3E & 3F). Although no differences in total size and mean size of tumours was apparent on infection, there was a significant increase in the number of small tumours upon infection (Fig 3G).
It is clear from previous investigation that T. muris induces homeostatic dysregulation in the gut [22,23,47]. In an attempt to determine whether worm induced perturbation in epithelial cell cycle was playing a role in tumorgenesis in APCmin/+ mice, the levels of cell proliferation and apoptosis were assessed throughout the intestine. Both cell proliferation (Fig 3H) and cell death (Fig 3I) increased in the large intestine during chronic infection. An increase in the number of aopotopic cells was observed at the base of the crypts whilst more proliferating cells were found further up the crypt axis in infected animals. However, no effect of infection upon proliferation or apoptosis in the small intestine was seen despite its effect on adenoma formation at this site.
Chronic T. muris infection is known to drive a Treg response in mice [48]. Indeed an increase in the numbers of CD4+CD25+FoxP3+ cells is evident in the spleen of chronically infected mice (Fig 4A) and at day 18p.i. in APCmin/+ mice (Fig 4B). A strong correlation between the presence of Tregs and the inhibition of tumour immunosurveillance has been reported in a number of cancers (reviewed in [49]). To address whether Tregs were playing a role in the promotion of neoplasia in this study, APCmin/+ mice were treated with anti-CD25 monoclonal antibody throughout infection and the effects assessed at day 18 p.i. Antibody treatment significantly reduced the numbers of CD4+CD25+FoxP3+ cells in the MLN and spleen of treated animals (S3 Fig). As shown previously, numbers of tumours were increased upon infection in control isotype treated animals (Fig 4C). Moreover, in this experiment and cohort of APCmin/+ mice, an increase in the total tumour area (Fig 4D) and mean tumour area (Fig 4E) was seen upon infection in the isotype treated groups. Interestingly, anti-CD25 treatment increased the tumour number and area in naïve mice as compared to isotype treated animals although this was not significant. There was no change upon infection in number, total or mean tumour area in the anti-CD25 treated animals (Fig 4F–4H). This was further confirmed when assessing tumour size. Isotype treated mice showed a significant increase in the smaller sizes of tumours upon infection (Fig 4I) whilst there were no differences in any size category between anti-CD25 treated naïve and infected animals (Fig 4J). Importantly, antibody treatment did not affect worm burden (S2 Fig). Thus depression of CD4+CD25+FoxP3+ numbers was associated with the capacity to control tumours in the gastrointestinal tract of T. muris infected mice.
In order to confirm if our observations for T. muris extend to other GI nematode parasites, we used a different parasitic worm, Heligmosomoides polygyrus (H. polygyrus). This parasite resides in the small intestine of the mouse so also allows us to assess the effect of the physical damage caused by a large parasite on neoplastic change at the site of adenoma formation. APCmin/+ mice were infected at 12 weeks of age and the effects of infection assessed on day 28 p.i., a time by which T. muris infected APCmin/+ mice had significantly increased intestinal neoplastic change. All infected animals had multiple worms in the small intestine on autopsy. The numbers of Tregs were unchanged from naïve levels in both the MLN and spleen of infected mice (S3 Fig). In contrast to T. muris infection, there were no changes seen in tumour number, mean tumour area or total tumour area in H. polygyrus infected mice (Fig 5A–5C) at this time point. This was also apparent when examining the numbers of tumours in different locations of the small intestine, where no changes were seen with infection (Fig 5D).
Here we demonstrate that low-level chronic T. muris infection promoted the development of intestinal neoplasia to a level that is comparable to that induced by a chemical carcinogen. Moreover, the observation that T. muris increased the neoplastic change seen in AOM treated mice and promoted tumour formation in APCmin/+ mice identified that gut dwelling nematode infection can induce simultaneous activation of local and systemic dysplasia. We have also identified that the T. muris induced Treg response that accompanies infection may negatively influence the neoplastic change in WT mice and tumour development in APCmin/+ mice.
Genetic changes such as activated oncogenes or altered tumour suppressor genes (such as APC) in tumour cells are responsible for many aspects of neoplasia, indeed, over 80% of colorectal cancer cases are proposed to be due to the loss of APC [50]. It is now established that an inflammatory environment also plays a role [2,51,52]. During chronic T. muris infection in C57BL/6 animals there is intestinal inflammation, a large influx of inflammatory cells into the intestine and elevated levels of pro-inflammatory cytokine production in the MLN and spleen (Fig 1A–1D). This infection-induced inflammation may be promoting the development of epithelial neoplasia in these mice (Fig 1E–1G) and indeed, other intestinal infections have been shown to promote tumour formation in an inflammation dependent manner [53]. In order to quantify the neoplastic change seen with infection we used the AOM model of intestinal cancer. T. muris infection induced an increase in aberrant crypt foci and in hyperplasia as compared to AOM alone (Fig 2C and 2D). However, infection and AOM in combination did not show additional changes over infection alone. Thus we can conclude that T. muris infection initiates neoplastic changes in the gut that are significantly increased when compared to those seen with a commonly used chemical carcinogen.
To assess the effect of T. muris on a model of spontaneous neoplastic change rather than chemical induced tumours, we used the APCmin/+ mouse model of colon cancer. Interestingly, T. muris, a nematode that resides in the large intestine, was able to exacerbate intestinal neoplasia throughout the intestinal tract in these animals (Fig 3A–3F). This clearly demonstrates that a caecal nematode infection can potentiate neoplasia in both a localized and systemic manner. Even within 18 days of infection, significant changes were seen within the lower ileum and in the number of smaller tumours found. This progressed to significant changes seen throughout the small intestine by day 42 p.i. with even more size categories of tumours affected. The finding that the greatest increase in tumour number was in tumours of the smaller size category (Fig 3G) strongly suggested that T. muris infection was acting to promote new tumour formation rather than enhancing the growth of well differentiated preexisting tumours. There was no difference between mean tumour size in naïve and infected animals (Fig 3C & 3D), again supporting the hypothesis that infection does not significantly affect the growth of pre-existing tumours. T. muris is known to cause epithelial dysregulation in the intestine with increased epithelial proliferation and apoptosis [22,47], both mechanisms which could lead to tumour formation [54]. However, changes in these mechanisms were only found within the caecum, the parasite niche, and not in the small intestine, which is the site of most neoplastic change (Fig 3H & 3I). Therefore, although epithelial homeostasis may play an important role in the development of worm-induced dysplasia in the large intestine, it appears to have minimal impact in the small intestine.
Typical inflammatory cytokines associated with chronic T. muris infection were seen in both the MLN and spleen of infected APCmin/+ mice (S2 Fig). This complements studies by Rao et al [53] that demonstrates that H. hepaticus infection promotes tumour development both locally in the intestine as well as systemically in mammary tissue in APCmin/+ mice due to inflammatory cytokine production and studies on the cytokine microenvironment in these mice [55]. Furthermore, the administration of dextran sulphate sodium (DSS) to APCmin/+ mice exacerbates adenoma formation, highlighting the importance of intestinal inflammation in promoting adenoma formation in this system [56] and the ablation of inflammatory cytokines leads to a decrease in adenomas [57]. Additionally, an increased pro-inflammatory cytokine production seen in T. muris infected mice over naïve mice treated with AOM alone may explain the increased neoplastic change (S1 Fig). However, a T. muris infection also promotes a robust Treg response that protects the host from damage [48]. Indeed, a key cytokine produced by Treg cells, IL-10, is critical in host survival during T. muris infection [37]. The successful use of T. muris to counter allergy and to protect against colitis in mouse models has been postulated to be due to this induced Treg response and its ability to immune modulate. In other systems, immune suppression can promote cancer through the down regulation of the anti-tumour immune response [reviewed in [49]] although paradoxically, in colon cancer Tregs are found to play a protective role [58–60] and this may be down to the type of Tregs that are found [61] or indeed the balance of cytokine production and Tregs [62].
Using anti-CD25 monoclonal antibody treatment to depress the number of Treg cells in vivo during the course of T. muris infection significantly reduced the number of CD4+CD25+FoxP3+ cells in the MLN and spleen of treated APCmin/+ mice. In the isotype treated animals, infection increased the number of tumours (Fig 4C) in the mice as demonstrated previously (Fig 3A & 3B). However, the mean tumour area (Fig 4D) and total tumour area (Fig 4E) was also increased as compared to the previous study where it was unchanged after infection (Fig 3C & 3D) suggesting an effect of T. muris on the growth of the tumours rather than initiation. It is worthwhile to note that this colony developed a significantly higher number of tumours and showed clinical signs of disease much earlier than the previous colony suggesting an earlier advancement of the disease. This raises the interesting question of whether the timing of infection in the context of tumour development is important. It was clear however that T. muris infection still induced neoplastic change. In contrast, there were no differences in any neoplastic change readout between the infected and naïve groups of the anti-CD25 treated mice, supporting a role for Tregs in suppressing tumour control in infected mice. Anti-CD25 treatment of naïve animals did increase the numbers of tumours and the mean tumour area as compared to isotype treated animals suggesting a role for CD25+ cells in protection against spontaneous neoplastic change in the APCmin/+ mouse. The role of Tregs in APCmin/+ is complex and findings differ between studies [63–66]. This may in part be due to the phenotype of the T reg present and the microenvironment [61,62] and would certainly warrant further investigation in the context of T. muris infection. Regardless of the effects in naïve mice, the data here strongly supports a role for parasite-infection induced CD25+ T cells in suppressing anti-tumour immunity.
To confirm whether observations were specific to T. muris infection or a reflection of intestinal helminth infection in general, APCmin/+ mice were infected with H polygyrus. H. polygyrus is a small intestinal dwelling parasite that presents as a chronic primary infection and at day 28 p.i. does not induce a marked CD4+CD25+FoxP3+ response (S3 Fig) [67]. This parasite model had the added benefit of allowing assessment of any neoplastic change as a result of mechanical damage by the worm at the site where neoplastic changes were evident. H. polygyrus did not induce any significant changes in any of the parameters assessed. Additionally H. polygyrus did not elicit as strong a proinflammatory environment in the MLN or spleen as observed with T. muris (S4 Fig). Whether a more prolonged infection would result in such changes requires further investigation.
The importance of the T. muris induced Treg response being detrimental, in terms of tumour control, to the host is important as the regulatory response induced by the parasite has been suggested to be beneficial by controlling colitis and allergy in mice [15–17]. We propose that this infection induced Treg response actually has detrimental consequences for both WT mice and APCmin/+ mice. The promotion of neoplasia by T. muris has important connotations given that 600 million people [11] harbor chronic infection with this genus. Ultimately, the impact of such infections warrant further investigation, particularly when considering the rising trend of cancer prevalence and, in particular, infection-induced cancers in developing countries [68,69].
|
10.1371/journal.pgen.1000871 | Mislocalization of XPF-ERCC1 Nuclease Contributes to Reduced DNA Repair in XP-F Patients | Xeroderma pigmentosum (XP) is caused by defects in the nucleotide excision repair (NER) pathway. NER removes helix-distorting DNA lesions, such as UV–induced photodimers, from the genome. Patients suffering from XP exhibit exquisite sun sensitivity, high incidence of skin cancer, and in some cases neurodegeneration. The severity of XP varies tremendously depending upon which NER gene is mutated and how severely the mutation affects DNA repair capacity. XPF-ERCC1 is a structure-specific endonuclease essential for incising the damaged strand of DNA in NER. Missense mutations in XPF can result not only in XP, but also XPF-ERCC1 (XFE) progeroid syndrome, a disease of accelerated aging. In an attempt to determine how mutations in XPF can lead to such diverse symptoms, the effects of a progeria-causing mutation (XPFR153P) were compared to an XP–causing mutation (XPFR799W) in vitro and in vivo. Recombinant XPF harboring either mutation was purified in a complex with ERCC1 and tested for its ability to incise a stem-loop structure in vitro. Both mutant complexes nicked the substrate indicating that neither mutation obviates catalytic activity of the nuclease. Surprisingly, differential immunostaining and fractionation of cells from an XFE progeroid patient revealed that XPF-ERCC1 is abundant in the cytoplasm. This was confirmed by fluorescent detection of XPFR153P-YFP expressed in Xpf mutant cells. In addition, microinjection of XPFR153P-ERCC1 into the nucleus of XPF–deficient human cells restored nucleotide excision repair of UV–induced DNA damage. Intriguingly, in all XPF mutant cell lines examined, XPF-ERCC1 was detected in the cytoplasm of a fraction of cells. This demonstrates that at least part of the DNA repair defect and symptoms associated with mutations in XPF are due to mislocalization of XPF-ERCC1 into the cytoplasm of cells, likely due to protein misfolding. Analysis of these patient cells therefore reveals a novel mechanism to potentially regulate a cell's capacity for DNA repair: by manipulating nuclear localization of XPF-ERCC1.
| XPF-ERCC1 is a nuclease that plays a critical role in DNA repair. Mutations in XPF are linked to xeroderma pigmentosum, characterized by sun sensitivity, high incidence of skin cancer, and neurodegeneration, or XFE progeroid syndrome, a disease of accelerated aging. Herein we report the unexpected finding that mutations in XPF cause mislocalization of XPF-ERCC1 to the cytoplasm. Recombinant mutant XPF-ERCC1 derived from XP– and XFE–causing alleles are catalytically active and if delivered to the nucleus of cells restore DNA repair. This demonstrates that protein mislocalization contributes to defective DNA repair and disease arising as a consequence of mutations in XPF. It also illustrates a novel mechanism of regulating a cell's capacity for DNA repair: by manipulating nuclear localization of XPF-ERCC1 to enhance or inhibit repair and to prevent cancer or tumor resistance to chemotherapy, respectively.
| Xeroderma pigmentosum (XP) is a rare autosomal recessive disease characterized by photosensitivity and a greater than a 1000-fold increased risk of skin cancer in sun-exposed areas of the skin [1]. In approximately 20% of patients, there is also progressive neurodegeneration leading to loss of coordinated motion, vision and hearing [2],[3]. XP is caused by mutations in genes that encode proteins required for nucleotide excision repair (NER) of DNA. Eight complementation groups of XP have been identified based on fusion studies with XP patient cells. These complementation groups include XP-A through XP-G and a variant, XP-V. The severity of XP varies tremendously, with diagnosis occurring anywhere from infancy to adulthood [1]. The severity of the disease is determined largely by which gene is mutated and to what extent the mutation affects NER.
NER removes helix-distorting lesions in DNA, for example cyclobutane pyrimidine dimers (CPDs) and pyrimidine pyrimidone photoproducts (6–4PPs) caused by the ultraviolet (UV) component of sunlight [4]. There are two ways by which DNA damage is recognized in NER. Lesions anywhere in the genome can be recognized by the complex XPC-RAD23B [5],[6]. For some lesions, this is facilitated by a second complex XPE/DDB2-DDB1 [7]. Alternatively, lesions that occur in the coding strand of DNA, within transcribed regions, can trigger NER if they stall progression of RNA polymerase II [8],[9]. This requires CSA, CSB and XAB2 [10]–[12]. Once the damage is recognized, the subsequent steps of damage excision are believed to be uniform. The basal transcription factor TFIIH is recruited to the site of helix-distortion to unwind the DNA around the lesion, using two of its ten subunits, XPB and XPD [8],[13]. RPA and XPA bind the unwound repair intermediate to stabilize it and recruit subsequent factors. The damaged strand of DNA is then incised by two structure-specific endonucleases, the heterodimer of XPF-ERCC1 and XPG, which cut 5′ and 3′ of the lesion, respectively [14]–[17]. This leads to removal of the lesion as part of a 24–32 base oligonucleotide. The resultant gap is filled by the replication machinery including RPA, PCNA, RF-C, DNA polymerase δ/ε, and the backbone is sealed by DNA LIGI or LIGIIIα-XRCC3 [18]–[20].
XPF-ERCC1 is a highly conserved endonuclease that nicks double-stranded DNA 5′ to a junction with single-stranded DNA [21]. In addition to NER, XPF-ERCC1 is involved in the repair of DNA interstrand crosslinks (ICL) [22] and double-strand breaks [23]. XPF and ERCC1 are paralogs thought to have arisen by gene duplication of the conserved nuclease and helix-hairpin helix protein-interaction domains [24]. The proteins interact via their C-terminal helix-hairpin-helix domains and this interaction is required to stabilize both proteins in vivo [25],[26]. XPF contains the nuclease catalytic domain [27], whereas ERCC1 mediates the interaction with XPA and recruitment to NER complexes [28],[29]. Unlike NER-specific proteins XPA and XPC, XPF-ERCC1 appears to be essential for human development or viability since no patients have yet been identified who are homozygous for early nonsense or frameshift mutations in either gene.
XP-A and XP-C are among the most common complementation groups in XP [30]. XP-C patients have severe skin abnormalities but generally lack neurological symptoms [31]. In contrast, XP-A patients show profound neurodegeneration, in addition to cutaneous features [1]. XP-F patients typically have very mild cutaneous features of XP, including late onset of skin cancer, but often have complications due to neurodegeneration as adults. It was recently discovered that mutations in XPF can lead to a second disease, XFE progeroid syndrome (short for XPF-ERCC1), characterized by spontaneous, accelerated aging of multiple tissues, including the nervous system [32]. Herein, we attempted to understand the molecular basis for how mutations in XPF could lead to such diverse outcomes. This led to the surprising discovery that mutation of XPF promotes mislocalization of XPF-ERCC1 to the cytoplasm of cells.
We first asked if mutations in XPF that cause mild or severe disease differentially affect the biochemical properties of XPF-ERCC1. To answer this, we compared the biochemical properties of XPF-ERCC1 from two patients, XP42RO (a patient with mild XP, homozygous for a mutation causing an R799W substitution in XPF [33]) and XP51RO (a patient with XFE progeroid syndrome, homozygous for a mutation causing an R153P substitution in XPF [32]) to that of wild type XPF-ERCC1. Recombinant XPFWT-ERCC1, XPFR153P-ERCC1 and XPFR799W-ERCC1 were purified from baculovirus-infected Sf9 insect cells using a His6 tag on ERCC1. We previously reported [27] that our purified preparations of XPFWT-ERCC1 elute from a gel-filtration column in three fractions: (1) a minor fraction in the void volume (∼45 ml) containing aggregated, inactive protein; (2) active heterodimeric XPF-ERCC1 at ∼65 ml, which corresponds to a molecular weight of ∼200 kD, as expected, and (3) monomeric ERCC1, which peaks at ∼78 mL, which corresponds to ∼50 kD. Recombinant XPFWT-ERCC1 eluted as expected (Figure 1A). Both mutant protein complexes eluted with similar profiles that differed substantially from that of XPFWT-ERCC1. The majority of the mutant complexes eluted at ∼45 mL rather than at 65 ml, indicating that they were aggregated. The peak at 78 ml, corresponding to free ERCC1 was identical for both mutant and WT XPF-ERCC1 preps. These results suggest that the mutations in XPF that cause both mild and severe disease lead to protein misfolding that does not interfere with ERCC1 binding, but does lead to protein aggregation.
We were able to purify a small amount of XPFR153P-ERCC1 and XPFR799W-ERCC1 from the fractions eluting at 65 ml, indicating that at least some of the mutant proteins are likely to be properly folded. SDS-PAGE analysis of the complexes after an additional purification step over a heparin column revealed dramatically reduced yields of the complexes of XPFR153P-ERCC1 and XPFR799W-ERCC1 compared to XPFWT-ERCC1 (Figure 1B). Similarly, the amount of XPF protein detectable by immunoblot in whole cell extracts of human fibroblasts harboring the XPFR153P and XPFR799W mutations (XP51RO and XP42RO, respectively) was reduced compared to normal cells (C5RO) (Figure 1C).
The catalytic activity of the purified heterodimers was investigated by measuring their ability to incise a 32P-end-labeled stem–loop DNA substrate at the single-strand:double-strand DNA junction in the presence of 0.4 mM MnCl2 or 2 mM MgCl2 at a 2-fold molar excess of protein over substrate (Figure 1D). With XPFWT-ERCC1, >80% of the stem-loop substrate was cleaved. Both XPFR153P-ERCC1 and XPFR799W-ERCC1 also incised the DNA substrate, demonstrating that both mutant complexes retain catalytic activity (Figure 1D, lanes 5 & 7). Incision by both mutant complexes was reduced compared to the WT complex. This may simply reflect the fact that preparations of mutant heteroduplexes were less concentrated than XPFWT-ERCC1 (Figure 1B), inevitably leading to differences in the buffering conditions between incision reactions.
We previously observed that mutant XPF-ERCC1 complexes tend to be more active in the presence of Mn2+ than Mg2+ since this metal has less stringent requirements for the proper alignment of the active site residues [27]. Consistent with this, incision by XPFR153P-ERCC1 and XPFR799W-ERCC1 was increased ∼2-fold in the presence of Mn2+ compared to Mg2+, whereas the cation had no effect on incision by XPFWT-ERCC1 (Figure 1D). These data support the conclusion that even monomeric XPFR153P and XPFR799W are to some extent misfolded. Notably, there was not a dramatic difference in the enzymatic activity of XPFR153P-ERCC1 and XPFR799W-ERCC1 on stem-loop DNA substrates, indicating that the biochemical basis for the more severe phenotype associated with the R153P mutation is unlikely to be simply a loss of catalytic activity.
R153P is situated in a lysine-rich domain of XPF that might be part of a complex nuclear localization sequence (NLS). Thus we next asked if XPFR153P-ERCC1 is mislocalized in cells. Differential immunofluorescence was used to identify the subcellular localization of XPF-ERCC1 in XPF mutant cell lines that were co-cultured with normal cells as an internal control [34]. XPF was detected exclusively in the nucleus of normal fibroblasts (Figure 2A, upper panel). In the same sample, XPF was detected in the cytoplasm of XP51RO cells (harboring XPFR153P). In XP42RO cells (harboring XPFR799W), the XPF signal was pancellular so that the nucleus and cytoplasm could not be distinguished from one another. To explore this further, we used immunofluorescence to detect XPF in additional XP-F patient cell lines and to our surprise discovered that in all mutant cell lines XPF was frequently detected in the cytoplasm. In all cases, XPF mutant cell lines could be discriminated from normal fibroblasts by the staining pattern of the cell population: reduced nuclear XPF-ERCC1 and the presence of cells in which the heterodimer was exclusively cytoplasmic. This was true irrespective of the antibody used for analysis [monoclonal 3F2 (Cancer Research UK), monoclonal Ab-1 (Neomarkers), or polyclonal anti-XPF (Erasmus Medical Centre)] (data not shown). This indicates that mislocalization of XPF in the cytoplasm is a common consequence of XPF mutation.
This raised the possibility that misfolding of mutant XPF caused abnormal subcellular localization of the protein. Therefore, next we asked if these mutant XPF proteins still interacted with their obligate binding partner ERCC1. ERCC1 was detected in the nucleus of normal (wt) fibroblasts, as expected. However, like XPF, ERCC1 was frequently detected in the cytoplasm of XPF mutant cells (Figure 2B). All eight XPF mutant cells lines tested were readily distinguished from wt cells by their ERCC1 staining pattern. Furthermore, there was a strong correlation between the staining pattern for ERCC1 and XPF in all cell lines. This observation indicates that ERCC1 can interact with each of these mutant XPF proteins. Furthermore, the results suggest that normally ERCC1 enters the nucleus as a heterodimer with XPF, and is retained in the cytoplasm with XPF when XPF is misfolded.
To rule-out the possibility that the abnormal subcellular localization of XPF-ERCC1 was an artifact of immunofluorescence, wt and XPFR153P fibroblasts were fractionated and XPF-ERCC1 was detected by immunoblot in the nuclear and cytoplasmic fractions (Figure 2C). In normal cells (C5RO), XPF-ERCC1 is predominantly nuclear. XPF-ERCC1 was also detected in the cytoplasmic fraction, but to an extent that could be attributed to nuclear contamination, as determined by immunodetection of the nuclear protein nucleophosmin. In contrast, substantially more XPF and ERCC1 were detected in the cytoplasm than in the nucleus of XP51RO cells. Similar results were obtained for other XPF mutant cell lines (XP24BR, XP26BR and XP32BR) and using other antibodies against XPF and ERCC1 (data not shown). Degradation products of both ERCC1 and XPF are commonly detected on immunoblot, but were not overrepresented in the mutant cells. These data confirm the immunofluorescence data and support the conclusion that mislocalization of XPF-ERCC1 in the cytoplasm occurs in XPF mutant cells.
Remarkably, cytoplasmic XPF-ERCC1 is not detected by immunofluorescence in all XPF mutant cells within a population. To quantify the phenomena, the fractions of cells with exclusively nuclear, exclusively cytoplasmic, or pancellular XPF-ERCC1 were determined from immunofluorescence images (Figure 2D). In wt fibroblasts, 93% of cells have XPF-ERCC1 only in the nucleus. Seven percent of cells show pancellular XPF-ERCC1. But never is the complex seen exclusively in the cytoplasm. In all of the XPF mutant cell lines, XPF-ERCC1 was detected exclusively in the cytoplasm of a fraction of cells ranging from 3–46% of the total population. Thus all known XPF mutations lead to a reduction in nuclear XPF-ERCC1 and an increase in the amount of the complex detected in the cytoplasm.
To further rule out the possibility that the cytoplasmic XPF-ERCC1 detected was an artifact generated by non-specific antibodies, human XPFR153P and XPFWT were tagged with YFP and expressed in Xpf mutant hamster cells (UV41) for direct detection of XPF protein. The expression of fusion proteins was confirmed by immunoblot using antibodies against human XPF and GFP (Figure 3A). Immunodetection of XPF revealed overexpression of both fusion proteins relative to endogenous XPF protein levels in normal fibroblasts (C5RO). Numerous breakdown products of XPF were also observed, likely due to its overexpression. But only a single fusion protein migrating at the expected molecular mass of full length XPF-YFP was detected using an antibody that detects GFP. To determine if the fusion proteins were functional, transiently transfected cells were tested for their sensitivity to UV to measure NER and mitomycin C (MMC) to measure interstrand crosslink repair (Figure 3B). Wild-type XPF-YFP yielded near complete correction of the hypersensitivity of UV41 mutant cells to UV and MMC. By contrast, despite the fact that XPFR153P_YFP was overexpressed to the same extent as XPFWT-GFP, this protein was unable to correct either DNA repair defect (Figure 3B), as expected based on the hypersensitivity of the XP51RO patient cell lines [32]. To determine the sub-cellular localization of XPFR153P, cells expressing the YFP-tagged protein were plated on glass coverslips and the protein detected by fluorescence microscopy (Figure 3C). XPFWT-YFP was exclusively in the nucleus. However, XPFR153P_YFP was detected in the cytoplasm of 95% of the transfected cells. This confirms the immunodetection data indicating that mutant XPF is cytoplasmic.
Unscheduled DNA synthesis (UDS) measures the incorporation of radiolabeled nucleotides into the genome of non-S phase cells after exposure to UV radiation and is a direct measure of NER [35]. Previously, UDS in cells from patient XP42RO (XPFR799W) and XP51RO (XPFR153P) was reported to be 20% and <5% of that in normal fibroblasts, respectively (Table 1). UV-induced UDS was measurable in all of the mutant XPF cell lines except XP51RO. This demonstrates that all of the mutant XPF proteins, with the exception of XPFR153P, retain catalytic activity in vivo.
To ask if XPFR153P is also catalytically active in vivo, recombinant, purified XPFR153P-ERCC1 was microinjected into the nuclei of NER-deficient XP51RO primary fibroblasts to determine if UV-induced UDS could be restored. XP51RO cells were first fused on slides by treatment with inactivated Sendai virus to produce homopolykaryons (multinucleate cells). Only homopolykaryons were injected with protein, to permit identification of those cells that were injected with protein. The slides were irradiated with 10 J/m2 UV-C, cultured in the presence of 3H-thymidine and nuclear grains indicating sites of thymidine incorporation in non-S phase cells measured (Figure 4). As expected, delivery of XPFWT-ERCC1 to the nuclei of cells led to a significant increase in the number of grains detected in homopolykaryons relative to individual cells in the same culture (Figure 4A). A significant increase in UV-induced UDS was also detected in homopolykaryons injected with XPFR153P-ERCC1 (Figure 4C). This confirms the in vitro activity data and establishes that XPFR153P is catalytically active in vivo if it is delivered to the nucleus. UDS levels were not recovered to the same extent as when WT protein was injected (Figure 4D). But injection of XPFR153P-ERCC1 or XPFR799W-ERCC1 led to a similar increase in UDS. This suggests that the incomplete recovery of DNA repair is due to decreased concentration of the recombinant mutant proteins relative to XPFWT-ERCC1 (Figure 1). These data support the conclusion that mislocalization of XPF-ERCC1 contributes to the DNA repair defect and symptoms caused by XPF mutations.
To determine if the severity of disease associated with a particular mutation in XPF could be predicted by the amount of XPF-ERCC1 detected in cell nuclei, the results in Figure 2D were compared to the clinical information available about the patients from which the cells were derived. In Table 1, the cell lines are listed in order of decreasing UDS. Patients with mild disease tend to have greater UDS or DNA repair. Patients XP32BR and XP26BR could be exceptions, but they are too young to know the full extent of their disease. In Figure 2D, the cell lines are clustered into those from patients in which severe disease/neurodegeneration was documented (right) or not yet observed (left). There is a trend towards those with severe disease to have more cells with non-nuclear XPF-ERCC1, but this trend did not reach significance (p = 0.06, unpaired Student's t-test), likely due to the small sample size. Therefore, the detection of cells with cytoplasmic XPF-ERCC1, while maybe useful to screen for patients with XPF mutations is not sufficient to predict patient prognosis.
Classically, inherited mutations in a gene are associated with a single disease. However, mutations in several genes involved in the NER pathway can result in more than one disease. The most prominent example is XPD, which if mutated can cause the cancer-prone disease XP but also Cockayne Syndrome (CS) characterized by photosensitivity, growth retardation, developmental abnormalities and profound neurodegeneration, as well as trichothiodystrophy (TTD), which is similar to CS, but also involves the skin and nails [36]. Similarly, mutations in XPB can also cause XP, TTD and a combined XP-CS [37] and mutations in XPG can lead to XP or XP-CS [38]. Of all the genes whose products are required for NER, only XPB, XPD and XPG are required for the proper function and stability of the basal transcription factor TFIIH [39]. Thus the more severe symptoms of CS and TTD are attributed to a combined defect in NER as well as transcription [36], [40]–[42]. Mutations in XPF were recently linked to a second disease in addition to XP, a disease of systemic accelerated aging termed XFE progeroid syndrome [32]. In this study, we sought to determine how mutations in XPF can lead to such a wide variety of symptoms.
Since CS and TTD are attributed to defects in transcription [34],[43], the prediction is that mutations in XPB, XPD or XPG that cause CS or TTD should affect basal transcription in addition to NER, whereas mutations that affect only NER cause XP. Indeed, mutations in the catalytic domain of XPG, for example A792V, disrupt the endonuclease activity of XPG, but not its interaction with TFIIH and therefore causes classical XP [44]. Similarly, a mutation in the helicase domain of XPD, D234N, affects NER, but not basal transcription and therefore leads to XP [45]. By analogy, we examined the enzymatic activity of XPFR153P and XPFR799W, which cause XFE progeroid syndrome and XP respectively, and discovered that neither mutation ablates the catalytic activity of the protein. Recombinant protein complexes harboring either mutation are able to incise a stem-loop substrate in vitro (Figure 1) and to restore NER in vivo (Figure 4). This is in keeping with the fact that patients with XP-F have residual UDS or NER (Table 1). Intriguingly, XP-F patients tend to have much milder photosensitivity and later onset skin cancer than XP patients from other complementation groups with the same level of UDS [46]. One explanation for this is that NER occurs in XP-F cells but at a much slower rate [47], making UDS a relatively poor reflection of the true DNA repair capacity of a cell. In total, these data provide clear evidence that viable mutations in XPF do not ablate catalytic activity of the XPF-ERCC1 nuclease. Of note, all XP patients for which the mutation in XPF was confirmed by sequencing genomic DNA harbor one of three recurrent point mutations (R799W, R589W or P379S). The rarity and limited repertoire of only hypomorphic point mutations in patients strongly suggests that XPF-ERCC1 nuclease activity is essential for normal embryonic development.
Mutations in a single gene could lead to diverse clinical outcomes if mutations differentially affect the stability of the gene product. For example, mutations affecting the stability of the TFIIH complex are linked with TTD but not XP [34]. Total cellular XPF and ERCC1 are dramatically reduced in cells from a patient with XFE progeroid syndrome (Figure 2C). However, XPF levels are reduced to the same extent in whole cell extracts from a patient with mild disease and 20% of the normal level of NER (Figure 1C). Therefore, mutations in XPF clearly affect protein level, which undoubtedly contributes to reduced DNA repair and disease. However, the level of XPF-ERCC1 in patient cells, as detected by immunoblotting, is inadequate to explain the differences in the severity in the DNA repair defect and disease between patients with different mutations in XPF. Interestingly, in at least a subset of XP-F patients, XPF mRNA levels are normal, but XPF protein level is low [46], indicating that mutant XPF is unstable.
The novel and unexpected finding is that mutation in XPF leads to increased cytoplasmic localization of the XPF-ERCC1 nuclease complex (Figure 2 and Figure 3) and that this aberrant subcellular localization is what prevents XPF-ERCC1 from participating in DNA repair (Figure 4). This was demonstrated by immunofluorescence detection of the complex using multiple antibodies. The results were confirmed by examining the subcellular localization of fluorescently tagged recombinant XPF (Figure 2). In further support of this, ERCC1 is also mislocalized to the cytoplasm of cells from the one patient reported with a mutation in ERCC1 [48].
Cytoplasmic localization of XPF-ERCC1 is not observed in all cells harboring XPF or ERCC1 mutations, suggesting the possibility that mutations affect proper folding of XPF-ERCC1 and that misfolded proteins are preferentially sequestered in the cytoplasm through interactions with other proteins or preferentially degraded. Alternatively, there may be tremendous selection for cells with nuclear XPF-ERCC1. This is consistent with the notion that the repair complex is essential for viability. Indeed, continuous passaging of XP51RO cells over years leads to a striking increase in the fraction of cells with nuclear XPF-ERCC1 and reduced sensitivity to the crosslinking agent mitomycin C (Ahmad, Bhagwat and Niedernhofer, unpublished data). Thus the fraction of cells with cytoplasmic XPF-ERCC1 may be underrepresented in Figure 2D, although only early passage cells were used in this study.
Many human diseases are caused by misrouting or mislocalization of proteins, ranging from metabolic disorders to cancer. Mislocalization of the tumor suppressors p53 [49], FOXO [50], p27Kip1 [51] and β-catenin [52] into the cytoplasm rather than the nucleus, leads to a loss of protein function and is associated with cancer. In contrast, mislocalization of NF-κB [53], BRCA1 and BARD1 [53],[54] from the cytoplasm into the nucleus is also associated with a variety of tumors. A classic example of a disease caused by protein mislocalization is cystic fibrosis which is caused by retention of the cystic fibrosis transmembrane conductance regulator (CFTR) protein in the endoplasmic reticulum, instead of its localizing to the cell surface [55],[56]. In addition, nephrogenic diabetes inspidus, retinitis pigmentosa, emphysema and α1-antitrypsin deficiency liver disease are also caused by mislocalized proteins [55].
Mislocalization of proteins may result from a mutated nuclear localization sequence (NLS) or nuclear export sequence (NES). Remarkably, the majority of the missense mutations in XPF is at arginine residues and leads to conversion of the arginine to a noncharged residue. So these mutations could affect a complex NLS. All of the point mutations (R→P, R→W and P→S) are also predicted to alter protein structure, supporting the notion that XPF mutations affect protein folding and/or protein:protein interactions that are critical for nuclear localization.
Our data add XPF to the list of proteins that if mislocalized contribute to disease. While this leads to novel insight into the regulation of XPF-ERCC1 and DNA repair in cells, the extent of XPF-ERCC1 mislocalization, as measured by immunodetection, does not predict the level of NER (UDS) or disease severity (Table 1). This could be because each mutation differentially affects folding of the protein and thereby differentially affects protein expression, protein degradation and/or cellular localization. Another possibility is that there are modifier proteins that influence disease severity, in particular in patients with homozygous mutations. However, we believe the former is of primary importance based on the observation that titration the level of expression of ERCC1-XPF in mice directly impacts lifespan and the severity of symptoms [57],[58].
In the case of XPF, it is the absence of XPF and its binding partner ERCC1 in the nucleus leading to reduced repair of genomic DNA that is disease-causing, rather than toxicity of mislocalized protein. Our data illustrate a novel mechanism by which the DNA repair capacity of a cell is determined: by nuclear localization of XPF-ERCC1. The identification of proteins that regulate this could lead to novel targets for improving DNA repair to treat patients with mutations in XPF or reduce cancer risk after exposure to genotoxic agents. Alternatively, these proteins would be excellent targets for small molecule inhibitors that would reduce repair and thereby prevent tumor resistance to genotoxic cancer therapies.
Purification of recombinant XPF–ERCC1 was performed essentially as previously reported [27] from baculovirus-infected Sf9 insect cells using a His6 tag on ERCC1. In brief, plasmids pFastBac1-XPF and pFastBac1-ERCC1-His were used to transfect Sf9 insect cells, and to amplify the virus according to the manufacturer's instructions (BAC TO BAC system; Life Technologies). Cell extracts were prepared 65 hr after infection with an MOI of 5 and highly purified protein was obtained using chromatography on Ni−agarose, gel-filtration and heparin columns. Only XPF-ERCC1 eluting as proper heterodimer on the gel filtration column at ∼65 ml of eluant was collected. The aggregated protein, eluting in the void volume (∼40–50 ml), was not used in experiments.
The endonuclease activity of wild-type and mutant XPF-ERCC1 was performed using a stem−loop substrate also as previously described [27]. A stem12−loop22 oligonucleotide (GCCAGCGCTCGGT22CCGAGCGCTGGC) was 5′-32P end-labeled. Nuclease reactions were performed on 100 fmol of DNA substrate and 20–200 fmol of XPF-ERCC1 protein in a total volume of 15 µl in optimized nuclease buffer (25 mM HEPES pH 8.0, 40 mM NaCl, 10% glycerol, 0.5 mM β-mercaptoethanol, 0.1 mg/ml bovine serum albumin and 0.4 mM MnCl2 or 2 mM MgCl2). The reactions were incubated at 30°C for 2 h and stopped by adding 15 µl of 90% formamide/10 mM EDTA and heating at 95°C for 5 min. Samples were loaded onto 15% denaturing polyacrylamide gels and reaction products were visualized by autoradiography and quantified on a PhosphorImager (STORM860; Molecular Dynamics).
Human fibroblasts immortalized with hTert were cultured in Ham's F10 with 10% fetal calf serum and antibiotics and incubated at 3% oxygen as described previously [32]. Cell lines included those derived from a normal individual (C5RO) [59], the parent of a patient, heterozygous for a mutation in XPF [33], XP-F patients (XP42RO) [33], XP23OS [60], XP24KY [46], XP7NE [61], XP32BR, XP26BR, XP24BR [61], and XP62RO, and a patient with XFE progeroid syndrome caused by a mutation in XPF (XP51RO) [32]. Unscheduled DNA synthesis (UDS) in these cells lines was previously reported as referenced above and confirmed in mixed cultures (XP-F cells co-cultured with normal cells using a more accurate click-staining method, as recently described [62].
Cells were trypsinized, washed twice with PBS and lysed with 1 ml NETT buffer (100 mM NaCl, 50 mM Tris base pH 7.5, 5 mM EDTA pH 8.0, 0.5% Triton X-100) containing Complete™ mini protease inhibitor cocktail and phosphatase inhibitor cocktail (Roche Molecular Biochemicals). Then the lysates were freeze-thawed twice in liquid nitrogen to disrupt nuclear membranes. From each sample, 50 µg of protein was resolved on 10% SDS-PAGE gels after boiling for 10 min in the presence of loading buffer. XPF was detected using a human XPF monoclonal antibody (clone 219; Neomarkers, Fremont, CA) at a dilution of 1∶1000.
Cultures of primary human fibroblasts from patients with mutations in XPF or a normal individual were grown in the presence of different size beads (2 µm or 0.8 µm; Sigma). After 24 hr the cultures were trypsinized and washed extensively with phosphate-buffered saline to remove any extracellular beads. The cells were then mixed in various combinations and co-plated on glass coverslips to provide internal controls of normal XPF-ERCC1 protein levels [34]. After 16 hr, the cells were fixed with 2% paraformaldehyde in sodium phosphate buffer, pH 7.4, for 15 min then permeabilized with 0.1% Triton X-100 in PBS. The samples were immunostained with polyclonal anti-ERCC1 (1∶2000; [63]) or polyclonal anti-XPF (1∶1000; [16]) followed by goat anti-rabbit ALEXA 488 (1∶500; Molecular Probes). Samples were stained with Dapi to identify nuclei and examined by phase contrast microscopy to identify the genotype of the cells according to their bead content and by fluorescence microscopy for immunodetection of repair proteins.
XPF cDNA was cloned into pYFP-N1 (BD Biosciences Clontech, Palo Alto, CA) such that YFP was expressed as fusion protein at the C-terminus of XPF. This construct, pXPF-YFP-N1, was then used to create XPFR153P-YFP by QuickChangeR Site-Directed Mutagenesis Kit (Stratagene, Cedar Creek, TX) according to the manufacturer's instructions. The wild type and mutant constructs were transfected in XPF-deficient CHO cell lines UV41 or UV47 using lipofectamine 2000 (Invitrogen, Carlsbad, CA) following the manufacturer's instructions. Cells expressing YFP were flow sorted using Dako Cytomation MoFLo high-speed cell sorter (Dako North America, Carpinteria, CA) 24–48 hrs after transfection.
To study the subcellular localization of XPF, YFP-positive CHO cells were plated on glass coverslips and grown to 95% confluency. The next day, the samples were fixed with 2% paraformaldehyde in sodium phosphate buffer, pH 7.4, for 15 min. The cells were permeabilized with 0.1% Triton X-100 in phosphate-buffered and nuclei were stained with Dapi-vector shield (Vector Laboratories, Inc. Burlingame, CA). XPF-YFP was visualized using an Olympus BX51 fluorescent 4 microscope at 60–100X magnification.
Wild type (AA8), XPF-deficient (UV41), XPF-YFP and XPFR153P-YFP cells were seeded in 6 cm dishes in triplicate at 103–104 cells per plate, depending on the dose of genotoxin. After 16 hr, the cells were irradiated with UV-C or exposed to mitomycin C (MMC). After approximately one week, the cultures were fixed and stained with 50% methanol, 7% acetic acid and 0.1% Coomassie blue. Colonies, consisting of at least 10 cells, were counted using a Nikon SMZ 2B 15 stereomicroscope microscope with 10X eyepiece. The data were plotted as the number of colonies that grew on the treated plates relative to untreated plates ± the standard error of the mean for 2–3 independent experiments.
Whole cell extracts were prepared from C5RO and UV41 cells transfected with vectors expressing YFP, XPF-YFP or XPFR153P-YFP. Proteins were separated by SDS PAGE using a 10% gel. XPF was detected using a human XPF monoclonal antibody (clone 219; Neomarkers, Fremont, CA) at a dilution of 1∶1000. YFP was detected using a GFP monoclonal antibody (Clones 7.1 and 13.1; Roche, Indianapolis, IN) at a dilution of 1∶1000.
Microinjection of purified proteins was performed as previously described [65],[66]. Briefly, primary human fibroblasts from XP51RO were fused by treating cultures with inactivated Sendai virus and then plated on glass coverslips. Subsequently, purified, recombinant XPF-ERCC1 protein complex (wild type or containing the R799W or R153P substitution in XPF) was injected into the nuclei of homopolykaryons. The cultures were irradiated with 10 J/m2 UV-C and pulse labeled for 3 hrs with 3H-thymidine. Unscheduled DNA synthesis (UDS) was detected by autoradiography.
One to ten femtoliters of a 10–100 nM solution was injected into the nuclei of 10–20 homopolykaryons for each of the three recombinant proteins and the number of radiographic grains counted in at least 20 nuclei of the homopolykaryons and a similar number of nuclei of single cells in the same sample. The mean and standard deviation of the number of grains was calculated for each of the three proteins. An unpaired, two-tailed Student's t-test was used to determine if there was a significant difference in unscheduled DNA synthesis between cells that were injected with recombinant XPF-ERCC1 and cells that were not injected.
|
10.1371/journal.pbio.0060060 | Phenotypic Mismatches Reveal Escape from Arms-Race Coevolution | Because coevolution takes place across a broad scale of time and space, it is virtually impossible to understand its dynamics and trajectories by studying a single pair of interacting populations at one time. Comparing populations across a range of an interaction, especially for long-lived species, can provide insight into these features of coevolution by sampling across a diverse set of conditions and histories. We used measures of prey traits (tetrodotoxin toxicity in newts) and predator traits (tetrodotoxin resistance of snakes) to assess the degree of phenotypic mismatch across the range of their coevolutionary interaction. Geographic patterns of phenotypic exaggeration were similar in prey and predators, with most phenotypically elevated localities occurring along the central Oregon coast and central California. Contrary to expectations, however, these areas of elevated traits did not coincide with the most intense coevolutionary selection. Measures of functional trait mismatch revealed that over one-third of sampled localities were so mismatched that reciprocal selection could not occur given current trait distributions. Estimates of current locality-specific interaction selection gradients confirmed this interpretation. In every case of mismatch, predators were “ahead” of prey in the arms race; the converse escape of prey was never observed. The emergent pattern suggests a dynamic in which interacting species experience reciprocal selection that drives arms-race escalation of both prey and predator phenotypes at a subset of localities across the interaction. This coadaptation proceeds until the evolution of extreme phenotypes by predators, through genes of large effect, allows snakes to, at least temporarily, escape the arms race.
| Arms races between natural enemies can lead to the rapid evolution of extreme traits, high degrees of specialization, and the formation of new species. They also serve as the ecological model for the evolution of drug resistance by diseases and for host–pathogen interactions in general. Revealing who wins these arms races and how they do so is critical to our understanding of these processes. Capitalizing on the geographic mosaic of species interactions, we examined the dynamics of the arms race between snakes and their toxic newt prey. Garter snakes in some populations have evolved dramatic resistance to the tetrodotoxin defense of the their local prey. By evaluating the pattern of mismatches between toxicity and resistance, we discovered that predators sometimes escape the arms race through the evolution of extreme resistance, but that prey never come out ahead. The reason for this one-sided outcome appears to depend on the molecular genetic basis of resistance in snakes, wherein changes to a single amino acid residue can confer huge differences in resistance.
| The process of coevolution plays out on a wide spatial and temporal stage [1]. At any point in space and time, we might observe a given pair of interacting populations occupying any position within the range of their coevolutionary trajectory: from the early stages of escalation, to equilibrium, to cost-induced descalation. A variety of historical, geographical, and ecological factors influence the condition of a set of interacting populations [1–7], rendering it exceedingly difficult to draw inferences about the coevolutionary process through the examination of one or a few localities at a single time [8,9]. Some forces and dynamics, such as gene flow or selection mosaics, can be effectively revealed in experimental systems with short generation times [10–12]. However, for natural populations of long-lived organisms, it is only by extending the analysis of coevolution across the geographic range of an interaction that we can begin to elucidate the dynamics and history of the process [9,13–25].
Coevolution is fundamentally driven by reciprocal selection that is generated by the ecological interactions of coexisting species [1,2,26]. The phenotypic interface of coevolution is defined as the set of traits that mediate these interactions [27,28]. When interacting species have roughly matched abilities at the phenotypic interface, the potential for strong reciprocal selection exists, because interactions between individuals of the two species are expected to have variable fitness consequences for one another [9,16,27,29]. At high levels of mismatch in abilities, however, variable fitness is not associated with variation in trait values for either species—all hosts may be resistant to infection by a parasite or all prey may be too toxic for any predator to ingest. In such cases, reciprocal selection no longer occurs, and the coevolutionary process for this pair of populations is suspended until other forces such as gene flow or mutation introduce new variants in one or the other population of interactants. By comparing the degree of functional mismatch among localities, we can infer which populations currently have the potential to experience reciprocal selection at the phenotypic interface of coevolution.
Defining and recognizing a phenotypic match is nontrivial. Matches are usually considered coevolutionary “hot spots,” where current reciprocal selection is strong [1,2,4,21,22,30–32]. Despite apparent matched levels of traits based on functional inference, two interacting populations may not experience reciprocal selection for any number of unidentified ecological or genetic reasons that mediate fitness (e.g., interactions with other species, energetic costs associated with the production of phenotypic interface traits, abiotic factors that obscure variance in fitness [1,2,7,17–20,33–35]). Thus, functional phenotypic matches may not necessarily reflect true coevolutionary hot spots.
Mismatches, on the other hand, should be more explicitly recognizable. By definition, mismatches are localities wherein the distributions of traits at the phenotypic interface are functionally nonoverlapping; the most extreme traits in one species generate no ecological effect on the most extreme trait of the other species. In this case, individuals of each species will not experience variable fitness outcomes (and thus no selection or evolutionary response) related to the traits in question. Mismatches are sometimes identified by differences in mean phenotypes [6,9,16,31,34,36,37]. However, because interacting individuals invariably encompass a range of phenotypes, the more critical requirement is to identify nonoverlapping phenotypic distributions around some critical performance threshold [27,29]. An evaluation of mismatches as “cold spots” therefore requires a distributional evaluation—if means differ but distributions can produce pairings of individuals with variable fitness outcomes (e.g., the least resistant predator coupled with the most toxic prey), reciprocal selection might still result.
Across the range of an ecological interaction, the degree of performance mismatch is expected to vary broadly. Once considered a general piece of evidence for coevolution, tightly matched phenotypes between natural enemies may be relatively uncommon among [4,6,36,38–40] and within species [35,37,41], as suggested by recent theory. So many factors influence whether traits are matched or not that inferences of underlying processes that are based on pattern alone are tenuous [1–6,24,39,40,42]. Regardless of which factors influenced current phenotypic patterns, the distribution of phenotypic mismatches can provide insights into the current selection dynamics and coevolutionary trajectories. Mixed patterns of phenotypic match and mismatch are observed in a wide variety of empirical systems including the following: host–parasite systems [24,25,38,43–45], plants and seed predators [9,14–16,18], and chemically mediated coevolution between parsnips and insect herbivores [21,22]. In each of these examples, phenotypic mismatches have revealed a geographically variable pattern of coevolutionary selection.
We examined variation in the phenotypic interface of the arms race between the garter snake predator, Thamnophis sirtalis, and its deadly prey (newts of the genus Taricha) to assess patterns of functional mismatch across both the geographic and phenotypic ranges of this interaction. Our analysis spans the geographic range of sympatry of Taricha and T. sirtalis, over 2,000 km from British Columbia, Canada, to southern California (Figure 1). The 28 localities sampled encompass the full variation of predator and prey phenotypes known in this interaction. We used a detailed functional model of the phenotypic interaction between newts and snakes to identify localities where trait distributions are sufficiently mismatched as to obviate current reciprocal selection. The emergent patterns of phenotypic escalation and phenotype mismatch reported here reveal not only spatially variable present day coevolution, but also suggest historical dynamics that cannot be observed in more geographically limited analyses. Mismatches invariably occur in the direction of predators exhibiting traits that are more extreme than necessary to exploit local prey and, consequently, not experiencing selection as a result of prey toxicity. Within localities that display trait mismatches, the level of phenotypic escalation indicates different coevolutionary histories across the range of populations.
The phenotypic interface of the predator–prey interaction between garter snakes and newts of the genus Taricha revolves around tetrodotoxin (TTX). TTX is one of the most potent neurotoxins known, binding to the outer pore of voltage-gated sodium channels in nerve and muscle tissue, thereby blocking the propagation of action potentials [46,47]. Taricha have high levels of TTX in the skin and are lethal to a variety of potential predators [28,48–52]; individuals from some populations have up to 14 mg of toxin, which is enough TTX to kill thousands of mice or up to 10–20 humans. A growing body of evidence suggests that newts produce their own TTX, but the genetics and biosynthesis of this process are poorly understood [50,53–57]. Some garter snakes of the genus Thamnophis have evolved resistance to this prey toxin through modifications of the sodium channel structure in skeletal muscle [58,59] and are capable of ingesting whole adult newts without permanent adverse effects [60,61]. Resistance in snakes is heritable [62,63] and is associated with a cost of reduced locomotor performance [64]. The functional interactions and relationships between individual newt toxicity and effects on individual snakes have been worked out in detail [27,58–63,65,66].
Both toxicity of newts and resistance of snakes vary geographically [28,60,62,63]. Where newts are absent or nontoxic, T. sirtalis are not resistant to TTX [28,62]. Elevated TTX resistance in western T. sirtalis is clearly derived, reaching levels 10–1,000 times that of other members of the genus in some populations [65]. Population differences in resistance are correlated with functionally important differences in amino acid sequences of skeletal muscle sodium channels [59]. Although considerable effort has been devoted to understanding the evolution of geographic and genetic patterns of TTX resistance in snakes, similar information regarding variation in newt toxicity is lacking. Data from only a few localities in the Pacific Northwest of North America suggest tight matching between prey and predator phenotypes [60], but a comprehensive survey of newt toxicity has not been previously conducted.
Mean total skin TTX levels of newts ranged from no detectable TTX to 4.69 mg/newt and differed among populations (Figure 1; ANOVA: F28, 382 = 20.38, p < 0.0001). Across this phenotypic range, newt toxin levels were closely correlated with the resistance of sympatric snakes (Spearman ranked correlation, ρ = 0.71, p < 0.0001). Geographically, regions of highest newt toxicity also corresponded to regions of highest snake resistance, with extreme values of both traits found in the Willamette Valley of Oregon and the San Francisco Bay Area of California (Figure 1). Isocline maps of newt toxicity and snake resistance show generally similar spatial patterns of phenotypic variation (Figure 2A and 2B).
Despite the overall spatial concordance of predator and prey phenotypes, an analysis of functional interaction reveals that over one-third of the localities sampled may qualify as ecological mismatches (Figure 3, Table 1, and Figure S1). The isocline map of the degree of mismatch, d, indicates that most of the geographic range of the newt–snake interaction is best characterized as mismatched and that regions of close ecological match are small and spatially restricted (Figure 2C, yellow to red areas). Localities where phenotypes are closely matched do not uniformly coincide with areas of elevated predator and prey phenotypes, but instead, include both ends of the phenotypic distributions of newts and snakes (Figures 1 and 2 and Table 1).
The observed levels of phenotypic mismatch ranged from near zero to d > 2.5 (Table 1). Values of d > 0.6 indicate populations that lay outside the 15% and 85% lines in Figure 3 and values of d < 0.6 indicate populations that fell between the 15% and 85% lines. Estimates of 0 < |d| < 0.6 indicate that localities lie within the zone of potentially experiencing reciprocal selection. In every case of mismatch (ten localities), predator resistance was much greater than the effective level of toxicity of local prey (Figure 3, gray zone, Table 1, and Figure S1A).
Mismatches included six populations of newts (Parsnip Lake, Oregon; Bear Ridge, California; Inland Lake, British Columbia; Crescent City, California; Latah, Idaho; and Scott Lake, Oregon) with little or no TTX (Figure 3, purple symbols in gray zone, Table 1, and Figure S1A). Snake populations at the same localities all fall in the lowest level of TTX resistance for garter snakes, wherein ingestion of ≈0.1–0.5 mg of TTX would reduce performance to 50%. This level of resistance is equivalent to the ancestral level of TTX resistance for the genus Thamnophis (50% dose ≈ 0.11 mg), including mostly species that have never coevolved with tetrodotoxic newts [65]. Four additional localities not explicitly recognized as mismatched (Skagit River, Washington; Orick, California; Priest Lake, British Columbia; and Vandenburg, California), also exhibited predator resistance greater than toxicity of local prey as well as reduced levels of resistance and toxicity (Figure 3, blue symbols in gray zone and one blue symbol in the nonshaded zone, Table 1, and Figure S1B). These nearly mismatched localities had phenotypic distributions of newt toxicity wherein a small reduction in the performance of local snakes could only result from ingestion of the most toxic newts present. We did not observe a single case where prey levels were greater than comparable predator abilities.
The other four mismatches include four populations of moderately to highly toxic newts (San Mateo, California; East Bay, California; Willow Creek, California; and Omo, California) that co-occur with the most resistant snake populations known (Figure 3, light green and yellow symbols in gray zone, Table 1, and Figure S1A). All of these localities occur within two geographic regions (the Bay Area of California and central Sierra Nevada) and include three species of Taricha (Ta. torosa and Ta. granulosa in the Bay Area and Ta. sierrae in the central Sierra Nevada) and likely two lineages of T. sirtalis [67]. As with the other mismatched populations, snakes in these localities can ingest sympatric newts with no reduction in performance or fitness consequence (Figure 4 and Figure S1A), however both newt and snake phenotypes are highly elevated, compared with ancestral conditions and conspecific populations (Figures 1, 2A, 2B, and 3, and Table 1).
Population-specific interaction gradients confirm the interpretation of mismatch across the interaction (Figure 4, Figure S1, and Table S1). Regressions of the predicted performance of snakes after ingestion of co-occurring newts indicate that TTX levels observed in newts at mismatched localities do not have variable effects (Figure 4, Figure S1A, and Table S1). Resulting interaction gradients in these mismatched populations have an average slope that is substantially lower that that seen in matched populations, and no differences in expected fitness are associated with variation in either TTX levels and TTX resistance at these localities (Figure 4, Figure S1, and Table S1). Interaction gradient slopes (β) of the four nearly mismatched localities were an order of magnitude greater than in mismatched populations, but much less than in matched localities, suggesting that these nearly matched localities experience reduced potential for selection relative to other matched populations, but greater than our mismatched localities (Figure S1 and Table S1).
Geographic patterns of newt toxicity and snake resistance illustrate the complexity of coevolutionary conditions that can span an extant species interaction. At first glance, levels of prey defense and predator exploitative ability appear to be generally correlated across broad geographic and phenotypic ranges (Figures 1and 2). Functional evaluation of trait distributions, however, shows that over one-third of sampled localities represent substantial mismatch in predator and prey abilities, including regions with extreme phenotypic values for both species (Figures 1–3). In many cases, these mismatches are so great as to obviate any current reciprocal selection (Figure 4 and Figure S1). In ten localities, predators are sufficiently resistant to escape selection due to prey toxicity, but prey are never too toxic to be successfully ingested by local predators. The pattern suggests a dynamic in which interacting species experience reciprocal selection, driving escalation of both prey and predator phenotypes in an arms-race style interaction. In some cases, predators have “escaped” this race by evolving extreme resistance to prey toxins, but the converse escape by prey is not observed. Other populations of predators and prey appear not to have entered the arms race at all.
Our results suggest, contrary to previous analyses [60], that extreme trait mismatches are not uncommon in this predator–prey system. However, the absence of localities in which newt toxicity was high enough to kill or disable any sympatric snake suggests that it is possible for the predator, but not the prey, to evolutionarily escape the reciprocal selection of the arms race. This directional asymmetry appears to contradict theoretical predictions arising from equilibrium theory [68], as well as the so-called “Life-Dinner Principle” (i.e., that prey experience stronger selection than predators in an arms race) [69], which predict that arms-race coevolution should favor defensive adaptations in prey over offensive adaptations in predators. This pattern may reflect a reversal in selective inequity as predicted for systems with deadly prey [70], or it may be particular to the unique biology of the newt–garter snake interaction.
The adaptive changes in resistance and toxicity are mediated not only by the strength of selection, but also by the genetic architecture of the traits at the interface. There is reason to expect adaptive changes in mechanisms of TTX resistance to proceed in a less-than-gradual fashion. If phenotypic changes in resistance are due to one or a few genes, then fixation of such genes in snake populations could be rapid and lead to phenotypic mismatches in one (or few) evolutionary step(s). Much of the variation in TTX resistance in T. sirtalis results from the expression of TTX-resistant voltage-gated sodium channels in skeletal muscle [58,59]. Resistance in these sodium channels is conferred by a small number of nucleotide substitutions in the TTX binding site [59]. The extreme resistance in at least one of these mismatched snake populations (Willow Creek, California) results from the substitution of a single amino acid [59]. Rapid fixation of such a simple mutation could explain how some populations of predators have escaped the arms race with prey. On the prey side, little is known about the basis of differences in TTX toxicity in newts, but some [52,71,72] have suggested that constraints on toxicity due to limited exogenous factors (e.g., environmentally derived precursors of TTX) may be one factor allowing predators to outpace prey in the arms race. However, the extreme levels of toxicity found in some newt populations demonstrate that elevated levels of the trait are possible within Taricha.
Our results indicate that geographic regions of phenotypic escalation are not necessarily congruous with coevolutionary hot spots. Coevolutionary hot spots (where reciprocal selection is intense) and cold spots (where selection is absent) are defined on the nature of the interaction rather than the level of the phenotype [1,2,4,27,60]. Our data reveal that current cold spots exist at localities with upper and lower extremes of phenotypes in both predator and prey (Figures 1–3). These results contradict earlier assessments of the geographic mosaic of coevolutionary hot spots in this system, which assumed that elevated predator phenotypes coincided with intense coevolution [60]. Similarly, reciprocal selection is possible at localities previously identified as cold spots (e.g., Vancouver Island, British Columbia) where phenotype distributions overlap and appear well matched despite the low levels of prey toxicity and predator resistance (Figure 1). Despite the fact that both predator and prey phenotypes show similar geographic patterns of escalation, the distribution of phenotype mismatch (i.e., cold spots) is not concordant (Figure 2). Coevolutionary hot spots may not be unequivocally assessed on phenotypic data alone, but at least the potential for reciprocal selection is observed across the range of phenotypic values in both taxa.
The observed pattern of trait mismatches among localities suggests a general arms-race dynamic for the process of predator–prey coevolution between Thamnophis and Taricha. The majority of localities occupy a broad band of phenotypic values within which potential reciprocal selection might occur (Figure 3). This zone of possible matching includes linearly increasing values of both newt toxicity and snake resistance that range from ancestral levels and increase several orders of magnitude, consistent with a counter-escalating arms-race dynamic in which pairs of populations experience reciprocal selection and evolve ever-increasing trait values [1,2,37,69]. This phenotypic zone includes multiple lineages of Taricha as well as at least two lineages of snakes that have evolved extreme levels of TTX resistance, suggesting that escalating dynamics have occurred multiple times during this evolutionary interaction [60,67,73,74]. Because many factors might ameliorate reciprocal selection at these localities, it is not possible to be certain that each of these localities represents a currently coevolving pair of populations. One clear and testable prediction from this interpretation is that older interactions should represent the populations with elevated phenotypes if average realized selection and the genetic architecture of toxicity and resistance are stable across localities.
Mismatched localities fall into two distinct groups that likely have different explanations and implications. At one end of the phenotypic distribution are the four populations of moderately to highly toxic newts that co-occur with the most resistant snake populations known (Figures 1 and 3, light green and yellow symbols in gray zone, and Table 1). As with other mismatched populations, snakes in these localities can ingest sympatric newts with no or little reduction in performance or fitness consequence, however both newt and snake phenotypes are highly elevated compared to ancestral conditions and conspecific populations. This pattern suggests that these localities have undergone arms-race coevolution, but that predators have escaped the arms race through the rapid evolution of extreme TTX resistance (see above). The extreme levels of TTX present at other localities (e.g. Benton, Oregon ) suggest that there does not appear to be a physiological limit to toxicity that explains these mismatches. These localities occur in nearby geographic regions (the Bay Area of California and the central Sierra Nevada; Figure 1) and involve different species of Taricha (Ta. granulosa and Ta. torosa in the Bay Area, and Ta. sierrae in the Sierra Nevada) [75]. Phylogeographic evidence suggests these represent two related groups of snake populations [67], indicating that escape from the arms race has occurred once or possibly twice in this fashion.
At the opposite end of the phenotypic distribution, we see mismatched localities from multiple lineages that appear never to have engaged in the arms race (Figures 1 and 3, purple symbols in gray zone). Population-specific interaction gradients (Table S1 and Figure S1A) and values of d (Figure 3 and Table 1) confirm that the opportunity for reciprocal selection in these localities is negligible. Both prey and predator traits at these localities appear to be close to estimated ancestral levels, wherein snakes have the slight ecological advantage of some predisposition to TTX resistance [60]. Average TTX levels in these newt populations range from 0 (or below our measurable lower limit of 0.0001 mg) to a high of around 0.002 mg (Parsnip, Oregon). This level of TTX is at or below the concentrations detected in related salamandrid species. In Notophthalmus, the sister genus to Taricha, reported levels of TTX range from 0 to a high of ≈0.06 mg per animal (estimated from [76]). In Cynops pyrrohgaster, an Asian TTX-bearing newt, typical whole-animal TTX levels range from 0–0.2 mg, with most population means ≈0.002 mg (estimated from [77]). The highest reported TTX level in the European newt genus Triturus (sensu lato) is 0.017 mg TTX (Tr. cristatus), with TTX levels in other species of Triturus an order of magnitude lower [78]. These comparative data suggest that mismatched localities at the low end of the phenotypic distribution have not engaged in counter-escalating coevolution. Alternatively, these populations may have been coevolving in the past, but once reciprocal selection was alleviated, costs of toxicity and resistance drove levels of both traits back to reduced levels. The four nearly mismatched localities (Figure 3, blue symbols in gray zone and one blue symbol in nonshaded zone) suggest that populations in this lower range can move from disengaged to engaged or that the process may be cyclical. Multiple snake and newt lineages are represented at the localities with unelevated phenotypes, suggesting that the phenomenon is not merely a phylogenetic artifact [67,73–75].
The apparent dynamic of arms-race coevolution in the newt–snake system, then, includes three or more stages. First, we see localities with low levels of traits at the phenotypic interface. Some of these localities, for reasons not yet clear, do not experience reciprocal selection and appear never to have engaged in the arms race. All of these localities involve predators able to subdue toxic prey without ill effect, suggesting that if newt toxicity rose in these populations, reciprocal selection would follow. As populations of newts gain toxicity (through mutation, migration from more toxic locales, or some exogenous influences), counter-escalation ensues and can lead to up to three–order-of-magnitude increases in traits. Initial increases in toxicity might be promoted by selection from interactions with other species as in other systems [3,17,22], including predators on early life stages [50]. Some localities (e.g., Benton, Oregon) seem to persist in this escalation zone, while others (e.g., Omo, California) escape from the arms race due to rapid evolution of extreme resistance through simple genetic mechanisms. Such adaptive changes suspend reciprocal selection, and no counter escalation follows. The next step for these populations is unclear. If costs to either resistance or toxicity are high enough, we might expect to see such escaped populations eventually lose phenotypic value and return to the lower left of Figure 3, resulting in de-escalation and a long-term cyclical dynamic. This scenario is plausible and has been suggested as an important dynamic in the chemically mediated coevolution between parsnip webworms and their host plant [79]. However, de-escalation is not supported by the observed patterns in the newt–snake interaction, which do not reveal mismatched localities with intermediate levels of resistance or toxicity.
We sampled a total of 383 newts from 28 localities co-occurring with populations of garter snakes for which TTX resistance has been described [60]. This sampling regime included most of the geographic range of this interaction and included localities from the central coast of British Columbia to the central coast region of California (Figure 1). The number of individuals sampled for each locality ranged from two to a maximum of 57 (Table 1). Only sexually mature animals were assayed in order to minimize variation in toxin levels associated with ontogeny. We included both males and females in our analysis; sex ratios of specimens varied among localities. Although earlier work suggested that there might be minor gender differences in toxicity of Ta. granulosa [55], we detected no such differences in our data set (ANOVA: F1, 370 = 3.26, P = 0.0719). We sampled populations of Ta. granulosa, Ta. torosa, and Ta. sierrae. Because average toxicities of Ta. torosa and Ta. sierrae populations were completely within the range of Ta. granulosa populations, we included all three species in a single analysis for this study (Table 1).
We predicted and evaluated the distribution of expected performance outcomes for each population of snakes interacting with sympatric newts over the range of toxicity observed in newts from each given population. This model of the chemical ecology and physiology of the interaction is based on an extensive understanding of the functional interaction between newt toxicity and snake resistance [27,28,49,53,55,56,60–63,66]. For each locality, we estimated the toxicity (in mg of TTX) of newts, the doses of TTX (in mg) required to reduce performance of co-occurring snakes to 15%, 50%, and 85% of their baseline performance, and the degree of match or mismatch between newt and snake phenotypes (see below for details).
Newt toxicity estimates and quantification of skin TTX levels. The amount of TTX present in dorsal skin of individual newts was quantified with high-performance liquid chromatography–fluorescence detection, and estimates of total skin TTX (in mg) per animal were generated following previously published methods [28,49,50,53,55]. This methodology has been shown to be a highly repeatable and accurate method for measuring dorsal skin TTX levels [28,49,53,55] as well as for estimating the total skin TTX of individual animals [49].
Quantification of TTX-resistance in Th. sirtalis. Whole-animal resistance data (in mass-adjusted mouse units or MAMU) were taken from Brodie et al. [60], in which TTX-resistance was measured with a bioassay based on a reduction in organismal performance after an interperitoneal (IP) injection of TTX [60,62,63]. This bioassay provides a highly repeatable estimate of individual and population level differences in susceptibility to TTX that expresses resistance as a percentage of baseline locomotor performance. A measure of 50% resistance means that an individual (or population on average) could crawl at 50% of its baseline speed after an injection of a given amount of TTX. TTX-resistance estimates used here are based on data from a total of 2,449 snakes from 269 families from 28 populations. We used these published dose-response curves to interpolate the average 15%, 50%, and 85% IP resistance doses (in MAMU) for each locality.
Comparing TTX toxicity in newts with TTX resistance in Th. sirtalis. Because absolute levels of TTX resistance in snakes (i.e., doses in mg rather than in MAMU) are related to size [60,61,66], we adjusted population average TTX resistance with respect to post-partum female mass for each population. Adult females are the largest size class in a given population and therefore are the most likely to prey on newts. Additionally, because of asymptotic growth curves in snakes, adult females represent the best size class to compare across populations. The 15%, 50%, and 85% IP doses (in mg) of TTX for adult post-partum female snakes at each population of Th. sirtalis were thus estimated using the average mass of this demographic group at each locality (Table 1). In the case of one locality, East Bay, size data were unavailable and we used an estimate of the average female mass based on its nearest geographic neighbors. Because 1 MAMU = 0.01429 μg TTX per gram of snake [60,62,63,66,80], the IP dose of TTX (in mg) required to reduce performance to a given amount (e.g., 50%) for an adult female snake at any given locality is estimated as:
where θ is the performance reduction dose of interest (e.g., 15%, 50%, or 85% in MAMU) and snake mass is the mean post-partum weight of female snakes from a given population. We modeled the effect of oral consumption of newts by snakes by converting the above IP doses to oral dose. The relationship between oral and IP doses of TTX is linear for Th. sirtalis (as well as other vertebrates; e.g., mice). At all levels of resistance and doses of TTX, the oral dose required to achieve the same effect as an IP dose is 40× [61]. We converted the IP 15%, 50%, and 85% resistance doses (in mg) to oral doses (in mg) by multiplying each dose by 40.
Modeling mismatch. We defined a functional mismatch if ecological interactions between individuals of sympatric species do not result in variable fitness consequences for either taxa (i.e., all predators are able to subdue all prey without impairment, or all prey able to repel or kill all predators). We defined a given locality as “matched” if a sympatric interaction could potentially result in variable fitness outcomes for one or both taxa. This outcome was conservatively judged to occur if the average performance reduction of a local snake ingesting any sympatric newt fell between 15% and 85% of normal crawl speed. The phenotypic space referred to as “matched” is more properly the region wherein potential reciprocal selection could occur between TTX toxicity and resistance. At performance levels <15%, snakes that ingest newts are fully immobilized or killed and newts escape [81], whereas at performance levels >85% snakes are unaffected and all captured newts die. We visualized match and mismatch at individual populations by plotting total skin TTX of newts against the size adjusted, oral 50% dose of snakes at each locality along with 15% and 85% dose model lines (see below) on a log scale (Figure 3). The actual range of newt phenotypes at each locality was used to illustrate the distribution of prey phenotypes. Because predator phenotypes are based on an estimated asymptotic function, it was not possible to plot them as range and we used the 95% confidence interval around each localities 50% as an estimate of phenotypic range. Our data included populations of newts that had no measurable TTX; as a result we transformed all values (TTX in newts and 50% doses in snakes) by adding 0.0001 mg to each value. This adjustment maintained the overall relationship between newt phenotypes and snakes phenotypes but allowed zero values to be plotted. The 15%, 50%, and 85% model lines were plotted using the absolute (i.e., in mg) estimates of the 15%, 50%, and 85% resistance doses (see above for details) for each locality.
Quantification of phenotypic mismatch. We calculated (d) as the orthogonal distance from the joint mean of each locality to the predicted 50% performance line (Figure 3). This estimate of distance d from the best match provides a quantification of the degree of mismatch at a given locality. An analogous approach has been used to evaluate arms-races between the sexes within species [37]. Although the choice of 50% to express this mismatch metric is somewhat arbitrary, the model of performance was robust and returned similar results for a range of (40% to 60%) of hypothetical matches. Because of the extreme range and nonlinearity of snake 50% doses and the presence of newt populations that had TTX levels below our detectable levels, we used log-transformed values of the following—(newt total skin TTX + 0.0001) and (snake 50% dose + 0.0001)—to calculate d (see above). This method uses the equation for estimating the shortest distance from a point to a line:
where A and B are the respective components of the slope and C is the intercept of the line. Our model assumes that the best functional match of newt and snake phenotypes at a given locality is one in which ingestion of an average newt by an average adult female snake will result in a reduction of that snake's crawl speed to 50% of baseline. This assumption results in the prediction that the model line describing perfect match is:
Thus the line describing perfect phenotypic match has a slope and intercept of 1 and 0 respectively, and A = 1, B = −1, and C = 0, and our estimate of d simplifies to:
where xi = log (50 % dose + 0.0001) of snakes from a given locality, and yi = log (average total skin TTX + 0.0001) of co-occurring newts.
Population-specific interaction gradients. Interaction gradients were generated for each locality by estimating the performance reduction experienced by an average snake after ingesting any of the observed sympatric newts. Thus the gradients reflect the observed distributions of whole newt toxicity for each locality. Interaction gradients are estimated with simple linear regression (SNAKE PERFORMANCE = (NEWT TTX) * β + ERROR). to reveal the average slope of the fitness consequence analogous to directional selection gradients, regardless of the form of regression that best fits the data [82]. Snake performance values are calculated from population-specific dose-response curves (see above). For the purposes of plotting, we normalized newt TTX levels to range from 0–1, with the most toxic newts scaled to 1 for each locality.
Phenotypic distributions and functional matching. We used the quantitative estimate of mismatch d to visualize geographic patterns of mismatch. Isocline maps that included all sampled localities seen in Figure 1 were generated using inverse distance-weighted interpolation based on observed values (i.e., TTX levels, snake resistance, and d) and the latitude and longitude coordinates for each population. Because of nonlinearity in resistance values (see also [60]) oral 50% doses of >5 mg were entered as 5 mg. The function's power was set at two and the neighborhood at 500 km. Analyses were performed in ArcView GIS 3.3 with Spatial Analyst 2.0. Analysis of geographic patterns of TTX-resistance and justification for phenotype classes in snakes (Figure 2B) was performed as per [60]. We used multiple post-hoc comparisons to estimate phenotype classes for Figure 2A (newt total skin TTX). Populations with values of d > 0.6 (i.e., those that lie outside the range of the 15% and 85% dose lines and were considered mismatched) are colored in blue and purple (Figure 2C). Populations with values of d < 0.6 fell between the 15% and 85% lines and are colored in red, orange, yellow, and green (Figure 2C).
|
10.1371/journal.pgen.1002776 | The CCR4-NOT Complex Is Implicated in the Viability of Aneuploid Yeasts | To identify the genes required to sustain aneuploid viability, we screened a deletion library of non-essential genes in the fission yeast Schizosaccharomyces pombe, in which most types of aneuploidy are eventually lethal to the cell. Aneuploids remain viable for a period of time and can form colonies by reducing the extent of the aneuploidy. We hypothesized that a reduction in colony formation efficiency could be used to screen for gene deletions that compromise aneuploid viability. Deletion mutants were used to measure the effects on the viability of spores derived from triploid meiosis and from a chromosome instability mutant. We found that the CCR4-NOT complex, an evolutionarily conserved general regulator of mRNA turnover, and other related factors, including poly(A)-specific nuclease for mRNA decay, are involved in aneuploid viability. Defective mutations in CCR4-NOT complex components in the distantly related yeast Saccharomyces cerevisiae also affected the viability of spores produced from triploid cells, suggesting that this complex has a conserved role in aneuploids. In addition, our findings suggest that the genes required for homologous recombination repair are important for aneuploid viability.
| Aneuploidy is a major cause of abortive development and is implicated in tumorigenesis in humans. Recent studies revealed that the increased need for protein degradation might account for the detrimental effects of aneuploidy on a cell. Here, we investigated the genetic systems responsible for aneuploid viability. Using a collection of gene deletions in fission yeast, we isolated mutants that affect aneuploid viability. We found that an evolutionarily conserved transcription regulator, the CCR4-NOT complex, and its related factors are required for aneuploid viability, suggesting that regulation of mRNA turnover is required to tolerate aneuploidy. In addition, homologous recombination repair is important for aneuploid viability.
| Aneuploidy is defined as a deviation from a multiple of the basic chromosome number and is a major cause of developmental defects in animals and humans [1]. Aneuploidy is implicated in tumorigenesis [2]. Aneuploidy is caused by errors in chromosome transmission and generally occurs at a low rate, but rates increase when chromosome transmission fidelity is perturbed, e.g., by mutations at the spindle assembly checkpoint [2]. Polyploidy is related to aneuploid production; e.g., tetraploid cells generated by cell fusion are an efficient source of aneuploid cells [3]. Crosses between polyploids lead to aneuploid gametes in plants [4], [5].
Aneuploidy causes a range of phenotypic consequences and is usually detrimental to both cells and organisms (reviewed in [6]). For example, mouse embryonic fibroblast lines with an extra chromosome have cell proliferation defects [7], and in the yeasts Saccharomyces cerevisiae and Schizosaccharomyces pombe, aneuploid cells generally show defects in cell cycle progression and genome stability [8]–[10]. As the grade of aneuploidy increases, i.e., the number of chromosomes involved increases, aneuploidy becomes lethal to the cell [8], [9], [11]–[13]. In addition, certain types of aneuploids grow better in suboptimal conditions, e.g., under elevated genotoxic stress [13]. Aneuploidy affects development of the organism in various species across kingdoms [6]. In one model, aneuploid cells are proposed to contain excess proteins that do not participate in protein complexes because of a dosage imbalance in gene products [6]. This idea is consistent with the fact that many aneuploids are sensitive to proteasome inhibitors and to conditions that interfere with protein chaperone function [9], and that among mutations that improve the fitness of aneuploid cells, one is defective in a deubiquitinating enzyme [14].
In S. cerevisiae and S. pombe, the higher the grade of aneuploidy, the poorer the cell viability. S. cerevisiae (n = 16) generally does not tolerate aneuploidy if the number of extra chromosomes exceeds five [11], [12], while in S. pombe (n = 3) all six types of aneuploids between n and 2n are lethal or extremely unstable, except for cells disomic for chromosome 3, the smallest of its chromosomes [15]. Aneuploids with higher grades of aneuploidy do not necessarily die immediately; some sustain their viability for a period of time and may survive to form a colony. This can occur when the grade of aneuploidy is reduced, probably by incorrect mitotic chromosome segregation, the rate of which is increased in aneuploid cells [8], [10]–[12]. Given this, we reasoned that compromise of any gene that functions to sustain aneuploid viability will reduce the efficiency of colony formation from the aneuploid cells.
To identify such genes, we screened mutants in S. pombe that affect the viability of aneuploid cells using a collection of deletion mutants of non-essential genes. Our results suggest that an evolutionarily conserved protein complex, CCR4-NOT, which is central to the regulation of mRNA turnover, is necessary for aneuploid viability in both fission yeast and budding yeast. Further, other genes involved in mRNA decay and export were identified. We also show that homologous recombination repair is important for the survival of aneuploid cells.
To identify genes involved in the viability of aneuploid cells, we screened a collection of fission yeast deletion mutants (Materials and Methods) [16] by investigating either spores from triploid meiosis or mutants in which the γ-tubulin gene (gtb1) and a spindle checkpoint gene (mad2) are impaired, which are referred to as the “triploid meiosis method” and “gtub-mad2 method”, respectively, hereafter.
As indicated above, six mutants were selected from both screening methods. Three of the genes, not3, not2, and btf3, were orthologous to NOT3 (or NOT5), NOT2, and BTT1 in S. cerevisiae, respectively (http://old.genedb.org/genedb/pombe/) (see Figure 2 for the gtub-mad2 phenotype). These genes are components of the CCR4-NOT complex, which is a general transcription regulator [18]–[21]. The other three mutants were swi6 (chromodomain heterochromatin protein), clp1 (Cdc14-related protein phosphatase), and SPAC1B1.04c (predicted to be an ortholog of PAN3, a subunit of the poly(A)-specific ribonuclease complex) (Figure S1 and data not shown). In addition, another defective mutant in the poly(A) nuclease (PAN) complex, ppk26, had a similar effect on aneuploid viability: a low S/L = 0.52 and a weak but significant synergy with the gtb1 mad2 mutant (see Figure S1). We did not study these four mutants further.
The deletion collection contained five more mutants defective in CCR4-NOT complex components. Two of these, caf4 (CAF4/MDV1 in S. cerevisiae) and caf16 (CAF16), were indistinguishable from the wild-type, although the caf16 mutant had a mild effect on aneuploid viability in a subsequent study, as described below (Table 1). The three other deletion mutants, ccr4 (CCR4), caf1 (POP2), and rcd1 (CAF40), had synergistic effects with the gtb1 mad2 double mutant, but the ccr4 and, in particular, caf1 mutations, also had synergistic effects with the gtb1 mutation alone (Figure 2), suggesting that their effects are not specific to aneuploidy. These three mutants were not tested in the triploid meiosis screening because they generated unhealthy spores (Table S1), but subsequent examination showed that the aneuploid spores with the rcd1 mutant had poor viability in the triploid meiosis test, as described below (see Table 1).
We examined the mutants of the CCR4-NOT complex more closely. Individual spores from triploid meiosis were randomly separated using a micromanipulator and their growth profile was microscopically observed (Materials and Methods). Microcolony morphology was recorded 2 days after spore separation, and the formation of visible colonies was scored 3 to 4 d later (Table 1). We classified microcolonies/cells into six types [8]: type A, large microcolonies comprising normally-shaped cells; C1, microcolonies containing elongated cells with or without septa; C2, microcolonies mainly comprising short and aggregated cells (see Figure 1B for representative C1 and C2 types); D, one germinated cell or two apparently dead cells; E, no apparent germ tube formation or little morphologic change from spores; and others, microcolonies with fewer cells than those in A-type microcolonies, but different qualities from those in types C1 and C2. Types D and E cells in most cases did not divide or showed limited division after day 2. Previous tetrad analyses revealed that A-type microcolonies are produced from haploid or diploid spores, and C1 and C2-types from aneuploid spores [8]. The chromosome 3 disome (n+1) spores made up the C2-type ([8] and present study). As for the D and E types, many of them probably represented aneuploid cells, but some of them were likely euploid cells [8]. Therefore, to evaluate aneuploid viability, we focused on the C1 and C2 types.
As shown in Table 1, the frequencies of A-type microcolonies among the mutants did not significantly differ from wild-type (p<0.05), indicating that the mutations did not significantly affect the viability of euploid spores. In contrast, formation of visible colonies from C1-type microcolonies was reduced in the mutants with the exception of caf4. Results from other genetic studies indicate that at least two and possibly five types of aneuploids produce C1-type microcolonies ([8] and O. Niwa, unpublished results). It was not clear whether the CCR4-NOT mutations differentially affected aneuploid types.
Colony formation from C2-type microcolonies appeared to be reduced in the not3, not2, and particularly in the rcd1 mutants, while only slightly in the btf3 mutant. Our previous genetic study suggested that at least the majority of “C2-type” spores from triploid meiosis are chromosome 3 disomes, suggesting that the growth of the chromosome 3 disome is affected in these mutants. To directly address this possibility, we examined the viability of this aneuploid by crossing a “wild-type” disomic strain with a haploid strain carrying one of these mutations. For the not2 mutant, because the locus is mapped on chromosome 3, we used a different disomic strain, whose not2 locus was heterozygous with not2Δ and not2+ (see Materials and Methods). With the exception of rcd1, there was little bias against the mutations in meiotic segregants, indicating that these mutations did not appreciably affect the colony-forming efficiency of the disome (Table 2). Thus, it was not clear why the colony-forming efficiency of “C2-type” spores was reduced in some of the CCR4-NOT mutants in the triploid meiosis test. Comparison of the colony size of mutant disomes with that of the wild-type disome, however, revealed that not3 and not2 produced much smaller colonies on a selective plate when incubated at 36°C (Figure 3A), a temperature at which the growth of “wild-type” aneuploids was retarded (Figure 3A and Y. Tange and O. Niwa, unpublished results). The temperature sensitivity was more pronounced in the not3 mutant, so that colonies were barely visible even after prolonged incubation at 36°C. Chromosome 3 disome with not2/+ heterozygosity did not show the growth defect, indicating that the not2 deletion mutation was recessive to the wild-type with respect to the temperature-sensitive growth phenotype (data not shown), and that the presence of the G418-resistance gene did not interfere with aneuploid growth. We also noted that the mutant disomes were less stable so that haploid colonies tended to emerge at increased frequencies (see Figure 3B). Similarly, we compared the disomes of caf4 and caf16 mutants with the wild-type. No differences were noted in the colony-forming efficiency or the chromosome stability of these mutant disomes (data not shown). The not3 and not2 mutants showed no temperature sensitivity when they were haploid or diploid (see Figure 3A and Figure S2). These findings indicated that deficiency of not3 and not2, but not caf4 and caf16, affected the growth of the chromosome 3 disome, and indicated that the effects of the not3 and not2 mutants were not limited to the C1-type aneuploids. In addition, these data suggest that, among the CCR4-NOT genes investigated in this study, the rcd1 gene has the most important role in aneuploid viability.
To examine how the not3 mutation affects the growth of aneuploid cells, we compared the sizes of C1-type microcolonies. Photographs of the microcolonies were obtained after incubation for 52 h at 30°C. The mean area of C1-type microcolonies in the wild-type was approximately two times that of the not3 mutant (Figure 4). This was not due to a higher incidence of microcolonies containing euploid cells, because microcolonies containing euploid cells were not included in this analysis. The sizes of individual cells in the mutant microcolonies appeared to be smaller than the wild-type. In a control study using haploid spores, the size of the microcolonies after 16 h incubation was indistinguishable between mutant and wild-type. These findings suggest that the wild-type not3 gene is required to maintain growth of at least some types of aneuploid cells.
The primary function of the CCR4-NOT complex is thought to be the general regulation of mRNA levels for a wide range of genes. Accordingly, genome-wide gene expression analyses in S. cerevisiae revealed that some observed genes were either overexpressed or underexpressed by at least 2-fold in deletion mutants of the CCR4-NOT genes, although there are some inconsistencies between studies in the observed frequencies of the affected genes in each of the mutants [21], [22].
We introduced a not3, not2, or caf4 deletion mutation in a wild-type haploid fission yeast, and whole genome microarray analysis was performed for these mutants as well as for the parental wild-type strain to determine the effect of each mutation on the overall gene expression pattern in exponentially-growing cells. The expression profile of individual genes in a mutant was presented as the ratio to wild-type (see Materials and Methods). The number of genes with effective values for not3, not2 and caf4 mutants was 4940, 4926, and 4928, respectively. Among these, the number of genes whose expression was affected by at least 1.5-fold (p<0.05) was 141 (2.9%), 61 (1.2%), and 17 (0.3%), respectively (Table S2). Of these, 30 genes were affected in both not3 and not2, 10 genes in both not3 and caf4, and 4 genes in both not2 and caf4. Among the genes affected in both not3 and not2, 27 of 30 were either overexpressed or underexpressed in both of the mutants, suggesting that Not3 and Not2 components of the CCR4-NOT complex function in the same direction in the regulation of a subset of genes. There are, however, exceptions to this rule. The expression of urg1 (urg for uracil regulatable gene [23]) and urg2 (and, to a lesser degree, urg3 [data not shown]) was reduced in the not3 mutant, but increased in the not2 and caf4 mutants (Table S2). The expression profiles of SPAPB24D3.07c was opposite those of the urg genes (increased in not3 and decreased in not2).
Another feature of the gene expression profile was that many of the genes that were underexpressed in not3 and not2 mutants (Table S2) mapped within two subtelomeric regions of chromosome 2; one is a 110-kb region centered 120 kb from the left terminus and the other is a 70-kb region 90 kb from the right terminus. Several genes were in these regions in the not3 (28 genes) and not2 (16 genes) mutants, accounting for 36% and 41%, respectively, of the genes listed as underexpressed in Table S2. It should be noted that most of the genes that mapped to these regions but are not listed in Table S2 also tended to be underexpressed in these mutants (data not shown), suggesting that Not3 and Not2 are involved in the regional control of gene expression. Although we do not understand how these microarray results are relevant to aneuploid phenotypes, the numbers of genes affected in the not3, not2, and caf4 mutants roughly correlated with the severity of aneuploid phenotypes, such as the temperature-sensitivity of the chromosome 3 disome as well as the poor colony-forming efficiency of aneuploid spores (see Figure 3 and Table 1). Altered expression of some specific genes might be also relevant to aneuploid viability (see Discussion).
Because the CCR4-NOT complex is evolutionarily conserved, we examined whether deficiency of the complex in S. cerevisiae also affects aneuploid viability. We made triploid strains with not3 or caf4 deletion mutations in otherwise similar genetic backgrounds and tested the viability of the resulting spores (Materials and Methods). As a control, we separated spores produced in wild-type triploids either by tetrad dissection or by random spore analysis and scored the number of spores that formed visible colonies after incubation for 6 d at 30°C. Note that due to the large number of chromosomes in this yeast (n = 16), virtually all the spores were aneuploid. With a few exceptions, only aneuploids with fewer than six extra chromosomes are tolerated in this yeast [11]–[13], which comprise about 10% of the total spores produced in triploid yeast. As summarized in Table 3, the overall colony-formation rate was 54.3% (n = 1483), with only a slight difference between the two different triploid strains. This value is greater than those previously reported (18% in [12]) and 38.5% in [24]), which may be due to genetic variations among laboratory yeast strains. These colony-forming efficiency values indicate that a significant portion of aneuploid spores survive and produce colonies, most probably by reducing the number of extra chromosomes during cell proliferation. Experimental spores from the mutant triploids had reduced colony-forming efficiencies of around 34% (n = 987) and 36% (n = 960) for the not3Δ and caf4Δ mutants, respectively (p<0.01 for both). Because the mutants did not appreciably affect the viability of haploid spores (Table 3), the reduced viability of spores from triploid meioses suggests that these two genes have important roles in sustaining aneuploid viability.
In a separate experiment, we counted the number of cells in each microcolony grown from individual spores after incubation for 15.5 h (Figure 5A), and scored the number of visible colonies after 6 d of incubation. As summarized in Figure 5A, the number of spores that remained single cells or divided only once was significantly increased in both mutants compared to wild-type (p<0.01): 25.3% (wild-type) vs. 35.8% (not3Δ) and 45.6% (caf4Δ). In contrast, the number of spores that divided many times to produce nine or more cells comprised 35.9% (wild-type), 25.5% (not3Δ), and 20.6% (caf4Δ), indicating that the proliferation rate of many types of aneuploids was significantly reduced in the mutants (p<0.01). In this experiment, visible colony formation rates were 55.2% (wild-type, n = 384), 39.6% (not3Δ, n = 384), and 31.6% (caf4Δ, n = 384). For haploid spores, the timing of spore germination and subsequent cell divisions did not significant differ between wild-type and mutants (Figure 5A). Thus, it is likely that these CCR4-NOT mutants decreased the cell proliferation potential of aneuploid cells in S. cerevisiae. The caf4Δ mutation tended to have a greater effect on spore viability than the not3Δ mutation, which contrasts with the results for S. pombe, suggesting that the roles of individual components in the CCR4-NOT complex for aneuploid viability differ among these yeasts. More importantly, however, the CCR4-NOT complex appears to have a role in aneuploid viability, suggesting that its function in aneuploidy may be conserved in other eukaryotes.
To address how the CCR4-NOT defects in S. cerevisiae impact defined types of aneuploids, we examined the growth rate of several types of disomes that contain not3, caf4, or btt1 (btf3 in S. pombe) deletion mutants (see Materials and Methods). Disome XV has a significantly longer doubling time than the wild-type haploid [9]. We found that deficiency of the NOT3 gene further decreased the growth rate of this type of disome (Figure 5B), which was consistent with the smaller colonies produced by the mutant disome compared with disome XV carrying wild-type NOT3. Mild detrimental effects of the not3 and btt1 mutants might be seen in disome XIII and disome XV, respectively. As for disomes I and II, we detected no effects of any of the mutants on growth rate (data not shown).
Many types of aneuploids in S. cerevisiae are hypersensitive to genotoxic agents, including camptothecin (CPT), phleomycin, and hydroxyurea (HU), and some other types of aneuploids are rather resistant to some drugs such as rapamycin and bleomycin [10], [13]. We examined whether the CCR4-NOT mutations affected the sensitivity of the disomes to CPT, HU, and phleomycin. Based on their colony size, disome I (and disome XIII less clearly) became more resistant and disome II became more sensitive to CPT in the absence of NOT3 (Figure S3A). Also, disomes II became very weakly resistant to CPT in the caf4 mutant. In addition, haploids with the NOT3 defect were slightly sensitive to HU and the sensitivity became more conspicuous in disome I (Figure S3B). With regard to phleomycin, we noted no specific effect of the CCR4-NOT mutants on disomes I, II, XIII, and XV (data not shown). Thus, the CCR4-NOT defects did not have strong effects on the defined types of aneuploids with only one extra chromosome, yet the CCR4-NOT defects did have some specific interactions with the aneuploids.
In the course of the present study, we observed that a rad32 (a homolog of MRE11) mutant was defective in maintaining chromosome 3 disomy (Table 2). It was also synergistic with the gtb1 mad2 double mutant (Figure 6), suggesting that DNA recombination/repair is involved in aneuploid viability and/or maintenance. We tested whether deletion mutations in recombination/repair-related genes had a synergistic effect with the gtb1 mad2 double mutant. As shown in Figure 6, rhp51 (the RAD51 homolog), rhp55 (RAD55), rhp57 (RAD57), and eme1 (MMS4) showed synergistic interactions with the double mutant. The effect of rad55 and rad57 deletions seemed weaker than that of rad51, but some ambiguity remains due to the emergence of fast-growing colonies with unknown genetic properties (see Figure 6). All results from this and other repeated tests indicate that these two mutants had synergistic effects with the gtb1 mad2 double mutant. This finding is consistent with Rhp55 and Rhp57 functioning as a complex to stimulate Rhp51 activity [25]. The Mus81/Eme1 complex is a DNA structure-specific endonuclease that functions in a late stage of homologous recombination repair [26]. Because the eme1 mutant showed synergistic effects with the gtb1-93 mutation alone, it is possible that its effect was not specific to aneuploidy. Rad3 and Crb2 are involved in DNA damage checkpoint control [27], but these proteins do not appear to be involved in aneuploid viability, based on the negative results of both the gtub-mad2 assay (Figure 6) and the triploid meiosis test (Table S1). Overall, the results suggest that the core machinery for recombination repair, e.g., Rad32, Rhp51, Rhp55, and Rhp57, are involved in aneuploid viability. Because these gene products are, to a certain extent, required for sporulation, the triploid meiosis test could not be performed adequately. It should be noted that in these mutants, unlike in wild-type cells, the minichromosome Ch16 was not stable ([28] and data not shown). Thus a chromosome-destabilizing effect may explain why triple mutants became sicker than the parental double mutants. It is conceivable that recombination repair is required for aneuploid viability, however, because chromosome-destabilizing mutants did not necessarily have synergistic effects with the gtb1 mad2 double mutant (see Table S1), and particularly because many types of aneuploid cells have defective DNA damage repair [10]. In the segregation analysis of the chromosome 3 disome, rad32, rhp51, rhp55 and eme1, no or very few disome segregants were recovered (Table 2). This suggests that these homologous recombination repair genes are required for viability of the disome, maintenance of the extra chromosome, or both.
In addition, we fortuitously found that an mRNA transport mutant, mex67 (the ortholog of S. cerevisiae MEX67) [29], had suppressor activity for the gtb1 mad2 mutant. That is, the triple mutant produced slightly larger colonies than the parental strain on YE plates, a condition that enhances aneuploid production (Figure S1). We then tested other transport-related mutants by the gtub-mad2 method, including SPAC14C4.06c (S. cerevisiae NAB2), nup97 (NIC96), SPAC328.05 (HRB1/GBP2), and crp79 (no ortholog known in other species). Interestingly, crp79 had a similar effect on the gtb1 mad2 mutant, although the effect was weaker than that of mex67. Crp79 was identified as a multicopy suppressor of the essential transport mutant rae1 [30]. Among other tested mutants, SPAC14C4.06c made the gtb1 mad2 mutant sicker, while the others had little or no effect.
The two yeasts S. cerevisiae and S. pombe are in distantly related subgroups of the phylum Ascomycota [31], [32], thus comparisons of these yeasts should provide good insight into the operations of eukaryotic cells. The present results suggest that deletion mutations in at least some components of the CCR4-NOT complex affect the viability of aneuploids in both fission yeast and budding yeast. Genetic as well as biochemical studies in S. cerevisiae and other species revealed that the complex is involved in several aspects of mRNA metabolism, including negative and positive regulation of transcription initiation, mRNA elongation, RNA degradation in the nucleus, and deadenylation of the poly(A)-tail for mRNA decay, with its primary function being the regulation of mRNA level in response to different environmental conditions [18]–[20], [33]. In addition, Not4 has ubiquitin ligase activity [34], [35] and the CCR4-NOT complex interacts with a nascent-associated polypeptide complex [35], which suggests a protein metabolism function. In S. cerevisiae, a discrete form of the complex, approximately 1 MDa in size and containing 10 subunits, Cdc39 (also known as Not1), Cdc36 (Not2), Not3, Mot2 (Not4), Not5, Ccr4, Pop2 (Caf1), Caf40, Caf130, and Btt1, have been identified [19]–[21]. This “core” complex is associated with other components, including Caf4 and Caf16, probably in a loose manner to form a larger form of the CCR4-NOT complex. The CCR4-NOT complex is thought to be conserved in fission yeast [36], [37], although some structural and functional divergence in its evolution has been proposed [38].
The Ccr4-Pop2(Caf1) moiety of the complex in S. cerevisiae has poly(A)-specific deadenylase activity, which together with the PAN complex accounts for the cytoplasmic deadenylase required for mRNA degradation [39]. S. pombe Caf1 also has deadenylase activity [37], [40]. Although triploid meiosis data are missing for the ccr4 and caf1 mutants, these mutants had a strong synergistic effect in the gtub-mad2 assay (Figure 2). Further, we identified two genes, SPAC1B1.04c and ppk26 (presumed components of the PAN complex in S. pombe), whose deficiency was associated with synergistic effects with the gtb1 mad2 mutant and reduced viability of spores from triploid meioses. Thus, the decrease in cytoplasmic deadenylase activity appeared to be associated with reduced aneuploid viability. The deficiency in deadenylase activity should stabilize the mRNA, resulting in too much functional mRNA, which in turn leads to an increase in protein production that could result in an increased need for protein degradation.
We found that deletion mutants of mRNA export-related factors (Mex67 and Crp79) had opposite effects in the gtub-mad2 test, that is, these mutations partially rescued the poor colony formation of the gtb1 mad2 double mutant (Figure S1 and Y. Tange and O. Niwa, unpublished results). In these mutants, the amount of functional cytoplasmic mRNA might be decreased, which is opposite to the case in the deadenylase mutants. Our preliminary examination by the triploid meiosis test as well as by the segregation analysis for chromosome 3 disomy, however, indicated that the mex67 deletion mutation reduced, rather than increased, aneuploidy viability (Y. Tange and O. Niwa, unpublished results). More specifically, the chromosome 3 disome was extremely unstable and the other types of aneuploids had reduced colony-forming efficiency associated with retarded growth. Further studies are needed to understand why there is an apparent rescue of the poor colony growth of the gtb1 mad2 double mutant by the mex67 mutant. Nevertheless, it is interesting that a class of putative regulators of mRNA dynamics also probably affects the viability of aneuploids.
How do the other mutations in the CCR4-NOT components, that is, not3, not2, rcd1, and btf3, affect aneuploid viability? In the present study, we performed a gene expression analysis of the fission yeast not3, not2, and caf4 mutants. The results indicated that the numbers of genes whose expression is strongly affected in the deletion mutants of the CCR4-NOT complex is lower than that in corresponding mutants in S. cerevisiae [21], [22]. This finding suggests that components of the CCR4-NOT complex or the complex as a whole in S. pombe might have different functions in the gross regulation of mRNA metabolism from those in S. cerevisiae, or that fission yeast might have a system against perturbations in mRNA turnover to ensure mRNA homeostasis, at least in haploid cells. Several interesting points may be drawn from our microarray data. Firstly, as already mentioned, the numbers of genes affected in each of the mutants correlated with the severity of aneuploid phenotypes, e.g., the growth defect of the chromosome 3 disome is most severe in the not3 and least severe in the caf4 mutant. The larger number of genes affected in the not3/not2 mutants may be more detrimental to the gene expression imbalance occurs in aneuploid cells. Second, among genes whose expression is affected in the not3 and not2 mutants, a number of genes are involved in transport between the cell and its environment. This may be relevant to the fact that fission yeast aneuploids are generally sensitive to environmental changes, including temperature and nutrition ([8] and present study). Third, we observed that a kinetochore protein, CENP-C homolog (Cnp3), is underexpressed by 1.8- and 1.9-fold in the not3 and not2 mutants, respectively (Table S2). Fission yeast Cnp3 is required for correct chromosome segregation [41], but because the minichromosome Ch16 is not appreciably destabilized in either of these CCR4-NOT mutants (data not shown), this level of reduction in Cnp3 expression does not seem to interfere with chromosome segregation in the quasi-haploid situation. Also, this reduction may not readily explain the growth retardation observed in a type of aneuploid cells (Figure 4). Provided that chromosome stability is generally reduced in aneuploid yeasts [8], [10], the lower expression of CENP-C may bring about further chromosome destabilization, and thus reduced viability. It remains to be examined whether Cnp3 expression is also reduced in other CCR4-NOT mutants.
Another important point that must be considered is that mutants of the CCR4-NOT complex and its interacting factors are hypersensitive to DNA-damaging agents in both S. cerevisiae and S. pombe [42]–[45], suggesting that the complex is involved in DNA damage repair and/or checkpoint. In S. cerevisiae, CCR4 and DHH1 (an RNA helicase interacting with Ccr4/Pop2) are required for resistance to ionizing radiation and other DNA-damaging agents. POP2(CAF1), NOT3, NOT2, and some other interacting genes confer radiation hypersensitivity when deleted [43]. In S. pombe, caf1, ccr4, rcd1, and not2 mutants are sensitive to DNA replication stress and/or to an ultraviolet light mimetic agent [44], [45]. Provided that, in both fission and budding yeast, homologous recombination repair function for DNA double strand breaks may be generally impaired in aneuploid cells [10], it is conceivable that the DNA repair function of the CCR4-NOT complex is involved in aneuploid viability. This is consistent with our finding that the genes required for homologous recombination repair had a synergistic effect with a chromosome instability mutant that continuously produced aneuploid cells.
In summary, the present findings demonstrate that the CCR4-NOT complex and other factors involved in the regulation of cellular mRNA level as well as proteins that are required for DNA recombination/repair play a crucial role in determining the fate of aneuploid cells.
Culture media used in the study were YE and YPD (rich media), EMM and SD (synthetic media), MEA (for conjugation and sporulation in S. pombe), and Sporulation medium (for S. cerevisiae) [46], [47]. YE medium was prepared for fission yeast using Bacto Yeast Extract (Becton Dickinson, Franklin Lakes). YES medium contained five supplements (adenine, uracil, leucine, histidine, and lysine) in YE [47]. Phloxine B plates were prepared as described previously [47]. YPD was prepared with Bacto Yeast Extract, Bacto peptone, and dextrose, and used for S. cerevisiae. KYPD (K for Kyoto) was analogous to YPD medium, but Polypeptone (394-00115, Nihon Seiyaku, Tokyo) and Yeast Extract (42007000, Oriental-Yeast, Tokyo) were used instead of Bacto Peptone and Bacto Yeast Extract. KYPD was originally used as an optimal medium for the fission yeast aneuploid study, particularly for cultivating the chromosome 3 disome. EMM was another good medium for the chromosome 3 disome, when sodium glutamate (5 g/l) was used as the nitrogen source. NH4Cl was a very poor nitrogen source for aneuploid proliferation. KYPD was also used with 5 µg/ml of TBZ as a permissive incubation medium for the gtb1 mad2 double mutant. Malt Extract Broth was purchased from Oxoid (Basingstoke, UK) for MEA. For SD, Difco Yeast Nitrogen Base (without amino acids or without amino acids and ammonium sulfate) was used (Becton Dickinson).
The yeast collection we used in this study was an early version of a deletion library and consisted of 2663 deletion mutants, which covered approximately 73% of non-essential fission yeast genes (3630 genes according to Kim et al. [16]). Their genotype was h+ leu1-32 ura4-D18 ade6-M210 (or M216) orfΔ::kanMX4 (most of the open reading frame [ORF] of a gene was disrupted with the G418-resistance gene) [16]. For the triploid meiosis analysis, each strain was crossed with a wild-type h− strain, L972, to isolate h− orfΔ::kanMX4 and h+ leu1-32 orfΔ::kanMX4 segregants. G-418 (G5013, Sigma-Aldrich Inc, St. Louis, MO) at a concentration equivalent of 100 µg/ml was used for the selection. We failed to obtain the targeted segregants in crosses for 643 deletion mutants.
The h− segregant obtained was then treated with methyl 2-benzimidazole carbamate (MBC; Wako, Osaka) to induce diploidization. Briefly, MBC stock solution (7.5 mg/ml) in dimethyl sulfoxide was added to a logarithmic phase culture in YE medium at 1/300 volume of the medium, followed by incubation at 26°C for 4.5 h. After incubation, we separated the affected cells (elongated cells with swelling or a short protrusion near the middle of the cell) with a micromanipulator on a Phloxine B plate, and incubated them at 26°C to obtain diploid colonies. More than 50% of the separated wild-type cells formed diploid colonies. In some cases, we spread the MBC-treated cell culture directly onto Phloxine B plates and isolated dark-colored colonies as diploid colonies. Stability of the diploid cells was assessed by spreading the cells on Phloxine B plates. For those showing poor stability, we tested the stability of the Ch16 minichromosome. Table S1 provides semi-quantitative data regarding the stability of the diploid and the minichromosome.
The obtained h−/h− diploid strain was then crossed with a corresponding h+ haploid strain on MEA at 26°C for 2 to 3 d. To isolate spores, the cell mixture on the MEA plate was digested with 0.5% (v/v in water) β-glucuronidase (G7770, Sigma-Aldrich) at 36°C for 3 h or longer. The number of spores was counted microscopically with a counting chamber. Very few vegetative cells escaped from digestion, allowing subsequent experiments to be performed without purifying the spores.
A known number of spores was plated on YE plates and incubated at 30°C for 4 d. The numbers of small and large colonies were manually counted, and their ratio was calculated. In the initial phase of screening, we observed the plates after 2 d incubation, and visible colonies were counted and marked, followed by further incubation for 2 d, when newly appearing small colonies were counted. At this time, the previously marked colonies had generally grown to be “large” colonies. For a detailed analysis of the spores, individual spores were separated with a micromanipulator onto a YE plate. After 2 d incubation at 30°C, the morphology of each cell/microcolony was observed microscopically and classified into six classes according to Niwa et al. (2006) [8] (see text and Table 1 and Figure 1), followed by another 3 to 4 d of incubation after which we determined which microcolonies produced visible colonies.
To compare their size, we took photographs of microcolonies after 52 h incubation at 30°C. The photographs were printed with the images of each microcolony, cut out along the edge and weighed to determine their relative sizes. For a control experiment, spores produced by diploids were incubated for 16 h and the size of their microcolonies was determined.
A gtb1 mad2 double mutant, YT708 (h−), contained the hygromycin B-resistance gene, hph, which was inserted 600 bp upstream of the start codon of the gtb1-93 mutant gene, and the nourseothricin-resistance nat gene, which was used to disrupt the mad2 gene according to the previously described procedure [48], [49]. YT708 was crossed with h+ segregants as described above to introduce the G418-resistant deletion mutations to the gtb1, mad2 background. Hygromycin B (H0654, Sigma-Aldrich) and clonNAT (Werner BioAgents, Jena, Germany) were used for hph and nat gene selection at 50 µg/ml and 100 µg/ml, respectively, together with G418 to select triple-drug resistant recombinants on YES plates with or without 5 µg/ml TBZ (T8904, Sigma-Aldrich) at 33°C. In an initial screening, we selected deletion mutants that produced a reduced number of triple-drug resistant recombinant colonies on the TBZ-free YES plate, compared with wild-type. The rad32 deletion we used in this study was not obtained from the deletion library. Instead, it was made separately by replacing the whole ORF with the nat gene. We confirmed that all deletion mutants listed in Table 1, Table 2, Figure 6, and Figure S1 had the correct disruption.
Strain P219 was a chromosome 3 disome with the mating type of h−. Each chromosome 3 contained the ade6-M210 and ade6-M216 alleles. Because the ade6 mutations complement each other, the Ade+ phenotype was used to indicate chromosome 3 disomy. As anticipated based on a previous study [8], the Ade+ phenotype was associated with the C2-type microcolony morphology. P219 was crossed with an h+ haploid strain carrying a deletion mutation (kan) and ade6-M210 (or M216). In the rad32 mutation, the nat resistance gene was used for gene disruption. For the not2 mutation, we used strain 56-1 (h−), which was disomic for chromosome 3 and one of the chromosomes carried the not2 mutation. Random spores produced from these crossings were plated on EMM plates, followed by incubation at 30°C for 5 d to select for Ade+ colonies. Fifty (or 20) colonies were randomly chosen and tested for G-418 (or clonNAT) resistance. From the drug-resistant segregants as well as drug-sensitive segregants, up to 12 colonies were randomly selected and tested for instability of the Ade+ phenotype (a genetic characteristic of disomy) by streaking them out on YE plates on which the Ade+ (white) and Ade− (red) phenotypes could be discerned based on colony color. For the cross using the not2 mutation, we first selected Ade+ and G-418-resistant colonies. Each of them was streaked on YE plates containing G-418, and we determined whether they were homozygous or heterozygous for the not2 alleles based on the fact that each of the two ade6 alleles produce characteristic colony colors (ade6-M210; deep red; M216: pale red). Therefore, disomes that produced an even mixture of two different red colonies on the G-418 plate were judged to be homozygous for the not2 deletion allele, while if (almost) all of the Ade− colonies were one of the two red colors, we considered them heterozygous.
Fission yeast wild-type strain L972 was used as the parental strain. The whole ORF of not3, not2, or caf4 gene in the parent was replaced with the G418 resistance gene according to the standard procedure [48]. Gene expression analysis was performed independently twice for each of the mutant and parental strains as described below.
We used the Agilent DNA microarray (15k×8 format; Agilent Technologies, Santa Clara, CA) containing 15,208 probe spots in each array. The 5529 probes representing 5529 fission yeast genes from the S. pombe genome sequences [50] (GeneDB:: http://old.genedb.org/genedb/pombe/) were designed using the Agilent eArray platform. Each probe was spotted twice (1484 genes) or three times (4030 genes) to fill 15,058 spots in the array format. Probes for 15 genes selected as replicate probes were spotted 10 times.
PolyA-RNA targets for microarray were prepared as follows. A single colony of S. pombe cells on a YES plate was inoculated into YES liquid medium. Cells were incubated at 30°C and collected with filtration when they reached a density of 5×106 cells/ml. Total RNA was isolated by the acid phenol method [51] (http://www.sanger.ac.uk/PostGenomics/S_pombe/). Using the Low Input Quick Amp Labeling Kit, one-color including Cy3-CTP (Agilent Technologies), labeled targets were prepared with 200 ng of total RNA. The labeled targets were purified using an RNeasy Mini Kit (Qiagen Japan, Tokyo). Hybridization and washing were performed under the manufacturer recommended conditions (Agilent Technologies) with 50 ng of labeled targets.
Microarrays were scanned using an Agilent array scanner (G2505C). The fluorescence intensity of each spot was processed using the Feature Extraction software (ver. 10.7.3. as recommended in the manufacturer instructions (Agilent Technologies). All subsequent data processing and analyses were performed with the GeneSpring GX software (ver. 11.5; Agilent Technologies). A coefficient of variation of 50% was used as the cutoff value. Averaged values from the replicates were used to calculate fold-changes in the gene expression in the mutants compared with wild-type. Genes whose expression was changed by at least 1.5-fold are listed (Table S2: unpaired T-test, P<0.05). The sequences of the probes and original data from the microarray experiments were submitted to GEO (http://www.ncbi.nlm.nih.gov/geo/index.cgi; accession number GSE36454).
Four isogenic derivatives of SK1 [52] were kindly provided by A. Shinohara (Osaka University): MATα HO:::LYS2 lys2 ura3 leu2 trp1 (HM785), MATa HO:::LYS2 lys2 ura3 leu2 trp1 (HM786), MATa HO:::LYS2 lys2 ura3 leu2 his4B-LEU2 arg4-nsp (HM787), and MATα HO:::LYS2 lys2 ura3 leu2 his4X-LEU2-URA3 arg4-bgl (HM788). Diploid cells (nonmater) from conjugation between HM785 and HM786 were lightly irradiated with ultraviolet light and plated onto a rich medium plate. To select a mating-proficient diploid colony, colonies formed from this plate were replica plated onto the lawn of HM787 or HM788 cells placed on an appropriate minimal plate. Triploid cells were simultaneously isolated as Leu+ Trp+ His+ Arg+ (Ura+) colonies on the minimal plate and incubated overnight on a fresh minimal medium plate without prior purification, followed by inoculation onto the sporulation medium. For random spore analysis, asci were digested with 0.25 mg/ml Zymolyase-100T at 36°C for 2 h. When required, individual spores were separated with a micromanipulator on YPD plates and incubated for the indicated period. Diploid cells for control sporulation were made from HM785 and HM787. To obtain triploid cells with homozygous not3Δ or caf4Δ mutations, respective genes were disrupted individually in HM785, HM786, and HM787 by replacing the whole ORF with the G418 resistance kan gene. The resulting three mutant strains were used to construct triploid cells as described above.
Media used for disomic S. cerevisiae were as follows. SD (−His+G418) was a selective medium for all disomic strains (Table S3). Sodium glutamate (1%) was used as a nitrogen source. G418 was added to a final concentration of 200 µg/ml. All disomic strains and control strains were kindly constructed and provided as frozen stocks by J. Sheltzer (Massachusetts Institute of Technology, Cambridge, MA). Disomic strains were tested by CGH [9] to confirm correct whole chromosome disomy immediately before freezing the cultures. Frozen cells were inoculated and incubated on the selective plates at 26°C overnight. The resulting patches were scraped to inoculate liquid selective medium, followed by incubation at 22°C with vigorous shaking. To determine the doubling time of each strain, the OD600 was measured every 2 h and values between 0.15 and 1.0 were used to indicate an exponentially growing culture. To minimize possible overgrowth of an unwanted fast-growing cell population in the culture, the OD measurement was started within 24 h from the inoculation for liquid culture, with one interim dilution in fresh medium. At the end of the OD measurement, cultures were spotted on selective plates to ensure that the culture did not contain an abnormal number of fast-growing cells compared with the original frozen stock. Phleomycin (ant-ph-1, InvivoGen, San Diego, CA), camptothecin (208925, Calbiochem, Darmstadt, Germany), and hydroxyurea (H8627, Sigma-Aldrich) were added to YPD medium at the indicated concentrations for the test.
|
10.1371/journal.pgen.1002408 | Drosophila Ribosomal Protein Mutants Control Tissue Growth Non-Autonomously via Effects on the Prothoracic Gland and Ecdysone | The ribosome is critical for all aspects of cell growth due to its essential role in protein synthesis. Paradoxically, many Ribosomal proteins (Rps) act as tumour suppressors in Drosophila and vertebrates. To examine how reductions in Rps could lead to tissue overgrowth, we took advantage of the observation that an RpS6 mutant dominantly suppresses the small rough eye phenotype in a cyclin E hypomorphic mutant (cycEJP). We demonstrated that the suppression of cycEJP by the RpS6 mutant is not a consequence of restoring CycE protein levels or activity in the eye imaginal tissue. Rather, the use of UAS-RpS6 RNAi transgenics revealed that the suppression of cycEJP is exerted via a mechanism extrinsic to the eye, whereby reduced Rp levels in the prothoracic gland decreases the activity of ecdysone, the steroid hormone, delaying developmental timing and hence allowing time for tissue and organ overgrowth. These data provide for the first time a rationale to explain the counter-intuitive organ overgrowth phenotypes observed for certain members of the Minute class of Drosophila Rp mutants. They also demonstrate how Rp mutants can affect growth and development cell non-autonomously.
| Ribosomes are required for protein synthesis, which is essential for cell growth and division, thus mutations that reduce Rp expression would be expected to limit cell growth. Paradoxically, heterozygous deletion or mutation of certain Rps can actually promote growth and proliferation and in some cases bestow predisposition to cancer. The underlying mechanism(s) behind these unexpected overgrowth phenotypes despite impairment of ribosome biogenesis has remained obscure. We have addressed this question using the power of Drosophila genetics, taking advantage of our observation that four different Rp mutants, or Minutes, are able to suppress a small rough eye phenotype associated with a mutation of the essential controller of cell proliferation cyclin E (cycEJP). Our findings demonstrate that suppression of cycEJP by the RpS6 mutant is exerted via a tissue non-autonomous mechanism whereby reduced Rp in the prothoracic gland decreases activity of the steroid hormone ecdysone, delaying development and hence allowing time for compensatory growth. These data provide for the first time a rationale to explain the counter-intuitive organ overgrowth phenotypes observed for certain Drosophila Minutes. Our findings also have implications for the effect of Rp mutants on endocrine related control of tissue growth in higher organisms.
| One of the early phenotypic classes identified in Drosophila was the Minutes, which were classified based on the heterozygous adults having short slender bristles on the body, a generally smaller body size and a delay in the onset of metamorphosis [1]. It has long been considered that understanding the basis for these phenotypes will provide fundamental clues to the mechanisms underlying the control of cell growth and proliferation as well as of tissue and organ size [2]. In 1976 it became apparent that many Minute genes encode Ribosomal proteins (Rps) [3] and by 2007 most of the Minutes were confidently ascribed to the Rp genes [4]. In all organisms, Rps are essential for the assembly and optimal functioning of the ribosome and are, therefore, obligate for protein synthesis and cell growth (reviewed in [5]–[6]). Due to their essential role in ribosome biogenesis, mutations that reduce Rp expression would be expected to limit cell growth. This cell intrinsic requirement for Rps explains many aspects of the Minute phenotype, such as the thin bristles and reduced body size in some Minutes. In contrast, other aspects of the Minute phenotype have remained enigmatic.
Paradoxically, reduced levels of some Drosophila Rps result in overgrowth of specific tissues. For example, RpS6 mutant larvae have overgrown lymph glands, due to increased growth and over-proliferation of the lymph gland cells [7], and develop melanotic masses [8]–[9], a characteristic feature of over-proliferation of hemocytes [10]. Thus reduced RpS6 expression results in tissue overgrowth, consistent with RpS6 having a tumour suppressor like function. Similarly, we have shown that RpL5 or RpL38 heterozygous adult flies exhibit significant increases in the size of the wings due to increased cell growth [11]. Rps have also been implicated as tumour suppressors in the vertebrate zebrafish model, where a genetic screen identified a link between malignant peripheral nerve sheath tumours and heterozygosity for several loss-of-function Rp mutations [12].
In mammalian systems, there is also evidence that Rp heterozygosity is frequently associated with tissue overgrowth and predisposition to cancer. For example, mutations in RpS19, RpS17, RpS24, RpL35a, RpS7, RpL5, RpL11, RpS10 and RpS26 have been associated with the human disease Diamond Blackfan Anemia (DBA), a dominant autosomal bone marrow failure syndrome, characterised by hypoplastic anemia with a predisposition to leukemia [13]–[19]. Mutations in RpS14 are also associated with 5q- syndrome and predisposition to acute myeloid leukemia [20]–[21]. Although RpS19 heterozygosity disrupts ribosome biogenesis 22–24, how reduced levels of Rps promote the excessive proliferation associated with progression to leukemia remains unclear and whether the mechanism is related to tissue overgrowth of Minutes has not been investigated.
Defining the mechanisms by which Rp heterozygosity results in tissue overgrowth and how reduction in a certain Rp gene predisposes a specific tissue to overproliferation in Drosophila is critical to understanding the processes linking growth and proliferation with tissue homeostasis. Furthermore, the insight provided by the Drosophila system may provide important clues in understanding how Rp mutations can promote cancer in humans.
Development of the Drosophila eye has been extensively used to identify and characterise regulators of growth and proliferation [25]–[26]. The Drosophila eye is composed of a highly organised array of photoreceptor clusters or ommatidia, which develop from an epithelial monolayer known as the eye imaginal disc. Differentiation of the ommatidia occurs in a wave that moves from the posterior toward the anterior. The anterior cells divide asynchronously and are separated from the differentiated posterior compartment by the morphogenetic furrow (MF) [27]. Mitotic division cycles become synchronized in the MF where cells are paused in G1 and a subset of photoreceptor cells are specified. The remaining retinal cells synchronously re-enter the cell cycle in the “Second Mitotic Wave” (SMW), which is composed of a tight band of DNA synthesis and mitosis. These final cell divisions provide the cells required for differentiation of the ommatidial structures that form the adult eye [28].
A hypomorphic mutation of cycE, cycEJP [29], reduces cycE expression during eye imaginal disc development to result in decreased S phases and small, rough adult eyes due to fewer cells (Figure 1A, compare i with ii) [29]. cycEJP therefore provides a sensitised genetic background to identify modifiers of eye proliferation, with suppressors of the phenotype being classed as “tumour suppressors” and predicted to normally function as cell cycle inhibitors [26]. To examine the mechanism(s) underlying the overgrowth phenotypes exhibited by some Minutes we have taken advantage of the unexpected observation that mutant RpS6 suppresses the hypo-proliferative, small eye phenotype of cycEJP mutants [26]. The data presented here confirm that reduced function of RpS6 suppresses the cycEJP small eye phenotype and we further demonstrate that this is not associated with restored proliferation in the SMW. Suppression of the cycEJP adult eye phenotype was observed with Rp mutants for both the small subunit (RpS12 and RpS19) and the large subunit (RpL38), which suggests the ability to restore eye size may be a more general property of reduced Rp abundance. Further investigation revealed that reduced RpS6 does not, however, lead to increased levels of CycE protein in the eye and that reduction of RpS6 specifically in the eye does not suppress the cycEJP small eye phenotype. Instead we demonstrate that reduced Rp levels in the prothoracic gland in RpS6 mutants decreased the activity of steroid hormone ecdysone, delayed development and hence allowed additional time for restoration of growth in the cycEJP mutants.
Mammalian cyclin E (cycE) is a well-characterised oncogene and, like the Drosophila homolog, regulates G1- to S-phase progression [30]–[32]. The cycEJP hypomorphic mutant has reduced cycE expression predominantly in the developing eye imaginal disc and, as a result, fewer S phases and small, rough adult eyes (Figure 1A ii and [29]). Previously a genetic screen for modifiers of the cycEJP phenotype identified the RpS6 mutant RpS6air8, which reduces RpS6 expression, as a suppressor of the cycEJP small eye phenotype [26]. This observation is consistent with previous observations that reduced RpS6 expression can promote proliferation in RpS6 mutant larvae [7]–[9].
We utilised the cycEJP small eye phenotype to examine the mechanisms by which reducing Rp levels can result in tissue overgrowth. As the original RpS6air8 line was no longer available to confirm the previous findings [26], we demonstrated suppression of cycEJP using an alternate RpS6 mutation, RpS6WG1288 [8]–[9], which also exhibits the classic Minute phenotype of slender bristles (not shown) and a developmental delay (Figure 3C, red data points). RpS6WG1288/+ restored the eye size and reduced roughness in the cycEJP background to give adult eyes with a more wild-type appearance (Figure 1A, compare i and ii with iii). Thus, two independent RpS6 mutations (RpS6air8 and RpS6WG1288) suppress the cycE hypomorphic small eye phenotype, consistent with reduced RpS6 function leading to increased proliferation in the cycEJP mutant.
In order to test whether suppression of cycEJP was specific to mutation of RpS6 or was potentially a more general consequence of reducing Rp levels, we tested two other Rp mutants that give Minute phenotypes, RpS12s2783 and RpS19bEP3448. Reducing RpS12 and RpS19 levels, with the mutant alleles RpS12s2783 [33] and RpS19bEP3448 (http://flybase.org/reports/FBrf0104946.html) resulted in a moderate suppression of cycEJP (Figure 1A iv and v, respectively). The cycEJP eye phenotype was also suppressed with a large subunit Rp mutant, RpL382b1 [11] (Figure S1). The finding that mutations in four different Rps from both subunits suppress the cycEJP phenotype suggests that this may be a common feature of Minutes.
The majority of the suppressors examined in detail from the original cycEJP screen demonstrated the ability to restore CycE protein towards wild-type levels and an associated increase in S phase progression [26]. Thus we examined whether RpS6WG1288 might similarly restore CycE levels in the eye. However, examination of CycE levels in eye discs from 3rd instar larvae revealed that this was not the case (Figure 1B, compare i and iii with iv). As reported previously [29] and consistent with the reduced CycE levels, S phase cells were also reduced in eye discs of cycEJP (Figure 1C iii). In line with the finding that CycE was not altered, the reduced S phases in the SMW of cycEJP were not obviously increased by reducing RpS6 (Figure 1C iv). Thus suppression of the cycEJP phenotype occurs in the absence of obvious changes to CycE abundance and S phase progression.
To monitor whether there was an overall change to cell cycle progression in the eye, we carried out anti-phosphohistone H3 staining to identify cells in mitosis as an alternative measure of cell cycles in the SMW (Figure 1D and quantified in 1E). The SMW of cycEJP mutants exhibited a significant reduction in their mitotic index as expected (Figure 1D iii and 1E). Importantly however the mitotic index was not restored in cycEJP eyes by the RpS6 mutant (Figure 1D iv and 1E). Therefore in these animals there is not a significant increase in the rate of cell cycle progression in the SMW, which suggests that this is unlikely to be the mechanism underlying suppression of cycEJP by the RpS6 mutant.
The findings above suggested that the suppression of cycEJP by the RpS6 mutant was not associated with either restoration of CycE or with altered cell cycle progression. As the cycEJP hypomorph predominantly affects the eye, we sought to test whether specific reduction of RpS6 in the cycEJP eye could suppress the phenotype. Using the eye specific GMR-Gal4 to drive expression of a UAS-RpS6 RNAi transgene, in both the SMW and differentiated cells posterior to the morphogenetic furrow [34]–[35], resulted in a smaller eye with a glassy appearance and necrotic patches (Figure 2A, compare iii with iv) [36] and 50% reduction in RpS6 mRNA in eye-antennal discs (Figure 2B). We then tested whether specific reduction of RpS6 in the eye could suppress the cycEJP phenotype. Reducing RpS6 with GMR-Gal4, which results in a small eye phenotype alone, was unable to suppress the cycEJP phenotype, and rather resulted in an additive reduction in eye size (Figure 2C, compare ii with iv). Due to the severity of the GMR>RpS6 RNAi phenotype we also tested knockdown with an alternate eye driver Ey-Gal4, which is expressed in all eye cells [37]–[38]. This resulted in ∼20% reduction in RpS6 mRNA in eye-antennal discs (Figure 2B) and did not produce an obvious adult eye phenotype alone (Figure 2A, compare v with vi). Thus like heterozygous RpS6WG1288/+, Ey>RpS6 RNAi does not result in an obvious eye phenotype (Figure 2A, compare i with ii). However, in direct contrast to RpS6WG1288/+, Ey>RpS6 RNAi enhanced rather than suppressed the cycEJP rough eye phenotype (Figure 2D, compare ii with iv). Together these data demonstrated that reducing the abundance of RpS6 in the eye, either robustly or modestly, was unlikely to be the mechanism underlying suppression of the cycE hypomorphic phenotype by the RpS6 mutant.
Because specifically reducing RpS6 in the eye did not suppress the cycEJP small eye phenotype, we considered the possibility that the interaction between RpS6 and cycEJP might be mediated by a mechanism extrinsic to the eye. To test this we placed UAS-RpS6 RNAi expression under the control of a range of ubiquitous Gal4 drivers in an effort to replicate the environment of the RpS6 mutant, by reducing RpS6 in the whole fly. Knockdown of RpS6 with the strong ubiquitous drivers Daughterless-Gal4 or Tubulin-Gal4 resulted in either early larval or embryonic lethality (Table S1). This is likely to be a result of RpS6 levels dropping below the threshold required for sufficient ribosome assembly and thus protein synthesis to support cell growth and proliferation. Consistent with this observation, reduction of RpS6 mRNA levels with strong drivers expressed in specific embryonic segments or larval domains also resulted in lethality (Engrailed-Gal4, Patched-Gal4) or shrivelled, stumpy wings (MS1096-Gal4) (Table S1 and Figure S2, compare iii with iv).
In contrast to the strong Gal4 drivers, reducing RpS6 mRNA levels with the relatively weaker ubiquitous driver, Actin-Gal4, resulted in viable flies (Figure S2, compare i with ii), which had a reduction in RpS6 mRNA similar to the levels seen in RpS6WG1288/+ larvae (Figure 2B, compare striped black and striped green bars). Importantly, this low-level reduction of RpS6 throughout the fly resulted in suppression of the cycEJP eye phenotype (Figure 2E, compare ii with iv) and a significant increase in eye size (Figure 2F, green bars). These data suggested that factors extrinsic to the eye were essential for suppression of cycEJP by the RpS6 mutant, consistent with our inability to detect changes in CycE activity or protein levels in the eye in the RpS6 mutant background.
As Rp mutations are associated with a developmental delay, we considered the possibility that the cell non-autonomous mechanism by which mutant RpS6WG1288 and RpS6 RNAi suppressed cycEJP might involve, at least in part, the ecdysone pathway, which is known to control timing of development and thus the growth period of the larvae. Specifically, release of ecdysone from the prothoracic gland (PG) dictates the timing of the metamorphosis from larvae to pupae (reviewed in [39]). As adult fly size is determined by the size of the larva at the time of pupal molt, the timing of ecdysone release plays a vital role in the growth of the fly [40]. We therefore examined whether RpS6WG1288/+ might suppress the cycEJP eye phenotype via an ecdysone-dependent, cell non-autonomous mechanism.
Previous studies have reported a role for the PG as a size-assessment organ [41]–[43]. Inhibiting the growth of the PG causes an underestimation of body size and results in pupation at a larger size. Conversely, promoting the growth of the PG results in smaller flies [41]–[43]. For example, overexpression of a dominant negative isoform of PI3 Kinase (Dp110DN) specifically in the PG blocks insulin pathway signalling and PG growth [41]. The smaller PG and associated reduction in ecdysone levels in these animals results in larger pupae and adults due to an extended larval growth period [41]–[43].
We therefore tested if the RpS6 mutant might suppress the cycEJP phenotype by impairing PG growth and, as a consequence, affecting the level of ecdysone. During eye disc development the morphogenetic furrow moves forward by one row of ommatidia (3–4 cell rows) every 70 minutes [44] and the doubling time for cells in the proliferating, anterior portion of the eye disc is approximately 12 hours [45]. Thus a developmental delay would provide the anterior asynchronously dividing cells and the cells comprising the second mitotic wave of the eye imaginal disc extra time to grow and divide in order to compensate for the proliferation rate defect resulting from reduced CycE activity.
First, examination of heterozygous RpS6 (RpS6WG1288/+) PGs, marked by expression of GFP, revealed that the glands were 35% smaller than GFP marked control PGs at the same time after egg deposition (AED) (Figure 3A, compare i with ii and quantified in 2B). This is also consistent with reports of RpS6air8 mutant larvae having small, abnormal PGs [7]. As a direct consequence of reduced PG growth, it would also be expected that RpS6WG1288/+ larvae should be developmentally delayed. Examination of developmental timing in RpS6WG1288/+ heterozygotes revealed that reducing the levels of RpS6 resulted in a delay in eclosion of up to 18 hours, compared to wild type (Figure 3C, compare open black circle with open red triangle). Importantly, the delay associated with the RpS6 mutant is reduced by addition of the active form of ecdysone, 20-hydroxyecdysone (20E) (Figure 3C, red data points and statistical analysis shown in Table S3), which suggests the delay in the RpS6 mutant is dependent on ecdysone levels.
The observation that the number of SMW divisions in the RpS6WG1288/+; cycEJP/cycEJP eyes were not significantly different to cycEJP alone suggests that the developmental delay and associated extra time for more cell divisions might underlie suppression of cycEJP. To investigate this possibility we tested whether suppression of cycEJP by the RpS6 mutant was impaired when the developmental delay is reduced by addition of 20E (Figure 3D). First we demonstrated that the RpS6WG1288/+; cycEJP/cycEJP animals had a developmental delay comparable to that for the RpS6 mutant alone, which could be reduced by the addition of ecdysone (Figure 3C, blue data points and statistical analysis shown in Table S3). Importantly, acceleration of development by the addition of 20E to the RpS6WG1288/+; cycEJP/cycEJP larvae resulted in a failure to suppress the small eye phenotype (Figure 3D, compare iv with viii). Thus suppression of the cycEJP phenotype by the RpS6 mutant is dependent on a developmental delay, which is sensitive to the level of ecdysone.
To further test our hypothesis that reduced levels of individual Rps in the PG of Minute mutants might restore proliferation in the cycEJP eye by inducing a developmental delay, we sought to reduce Rp expression in the PG using AmnC651-Gal4 which drives expression in the PG [41] and UAS-Rp RNAi for RpS6, RpS13 or RpL38. We first demonstrated the RNAi was able to reduce RpS6 protein by knocking down specifically in the PG, and staining with an anti-RpS6 antibody (Figure S3A). Consistent with the importance of Rps for growth, reducing Rps in the PG resulted in much smaller PGs in these larvae compared with the control at the equivalent time point of 5 days AED (Figure 4A ii–iv). Moreover, reduction of RpS6 levels resulted in PGs that were smaller than for the RpS6WG1288/+ PGs, suggesting a greater reduction in RpS6 (compare Figure 4A ii to Figure 3A ii). Examination of the AmnC651>RpS6 RNAi PGs at 12 days AED revealed that the size of the gland was still considerably smaller than the control PG (data not shown). As a smaller PG would be predicted to result in less ecdysone synthesis and release, we examined if the reduction in PG size affected ecdysone activity in the larvae. qRT-PCR was performed on whole larvae to measure ecdysone activity indirectly by quantifying the mRNA levels of an ecdysone responsive gene, E74B [41]. E74B levels were normalised to Actin-5C, a non-ecdysone responsive gene. RNAi-mediated reduction of RpS6, RpS13 or RpL38 in the PG resulted in up to 90% decrease in E74B expression (Figure 4B), suggesting strongly reduced ecdysone activity, reflecting the small size of the PG.
Consistent with the robust reduction in PG size and reduced ecdysone activity, we observed an extreme developmental delay in the larvae with RNAi-mediated knockdown of RpS6, RpS13 or RpL38 in the PG. At day 5, these larvae were smaller in size compared with control larvae (Figure 4C, compare i with ii–v). While the control larvae underwent pupation as normal at day 5, larvae with reduction of RpS6, RpS13 or RpL38 specifically in the PG continued to feed and grow beyond day 10 to become giant larvae, which fail to pupate (Figure 4C, compare vi with vii–viii, x). The phenotype for the RpL5 knockdown in the PG was even more dramatic, being 2nd instar larval lethal (Figure 4C ix), suggesting that RpL5 was knocked down below the threshold required for cell intrinsic growth [36], [46]–[47] and, therefore, development of the PG gland. This is consistent with the lethality that results when strong drivers are used to express RNAi transgenes targeting the Rps investigated here (Table S2).
The AmnC561-Gal4 insertion is not expressed solely in the PG, being expressed throughout the ring gland early, in some cells in the ventral ganglion and in neurosecretory cells of the brain [41]. As the neurosecretory cells of the brain can also play a role in developmental timing and growth [48], we addressed the possibility that RpS6 knockdown in these cells might be responsible for the overgrowth by using another driver, P0206-Gal4 [43], that also expresses in the PG, but not in the neurosecretory cells. Consistent with the effect being mediated through defects in PG development, knockdown of either RpS6 or RpL38 using P0206-Gal4 also resulted in an extreme developmental delay whereby larvae continue to feed for greater than 20 days and fail to pupate, which was associated with a smaller PG (Figure S4A).
To assess whether the reduced ecdysone production was the cause of the developmental delay and larval overgrowth resulting from Rp knockdown in the PG, 20E was introduced to the food of AmnC651>RpS6 RNAi larvae (Figure 5A). The addition of 20E resulted in a variable restoration of pupariation, which ranged from progression towards cuticle darkening in larvae to cuticle development and early pupal morphology (Figure 5A, compare v with vi–vii and Figure 5B green bars). Although the AmnC651>RpS6 RNAi larvae were able to pupate, the ectopic addition of 20E was unable to initiate the final steps of metamorphosis, including the formation of adult structures. This suggests that ∼30% of the endogenous 20E activity achieved by feeding the larvae (Figure 5B) is sufficient to trigger pupariation, but is below the threshold required for adult metamorphosis. The failure of metamorphosis may be confounded by the fact that pupae, unlike larvae, can no longer take up 20E by feeding. Indeed, the largest peak of endogenous ecdysone release occurs after cuticle formation and is required for the formation of adult structures [39].
To confirm this failure to restore pupation was not due to insufficient 20E in the food we carried out a control rescue experiment with an alternate growth regulator, PI3K, which has previously been shown to modulate PG size and development [41]. Despite having a PG size similar to that of AmnC651>RpS6 RNAi (Figure 4A, compare ii with v) and associated extreme developmental delay, the AmnC651>Dp110DN (dominant negative PI3K) larvae were only moderately delayed and pupated, but eclosed as larger flies (Figure 5A, compare viii with x, and [41]). We demonstrated that feeding 20E to larvae overexpressing dominant negative PI3K in the PG (AmnC651>Dp110DN) restored the time of pupation back to day 5, the adults eclosed at a normal size (Figure 5A, compare iii and x with iv and xi), and E74B levels were significantly increased compared to that of control (Figure 5B, blue bars). This restoration of timing and size toward control suggested that the 20E was successfully taken up and processed by the AmnC651>Dp110DN larvae. The difference in the severity of the phenotypes in terms of developmental delay, strongly suggested that ecdysone levels are more sensitive to disruption of Rps and ribosome biogenesis than to disruption of insulin pathway-dependent growth in the PG
As RpS6 knockdown in the PG gland resulted in a failure to undergo pupation, in order to carry out further studies we examined whether we could reduce the severity of the phenotype and facilitate development into adult stages using a temperature sensitive isoform of the Gal4 repressor, Gal80 (Gal80TS [49]) that allows temporal control of the induction of RpS6 knockdown by RNAi in the PG. Thus, knockdown of RpS6 was delayed until late 2nd instar and although this still resulted in large, developmentally delayed larvae (Figure 6A, compare i with v), these larvae were able to undergo pupation and eclosed as large adults (Figure 6A, compare ii with vi). In addition, we observed increases in the eye size (Figure 6A, compare iii to vii) and statistically significant increase in the wing size (Figure 6A, compare iv to viii, quantified in 6B), in the AmnC651;Gal80TS>RpS6 RNAi adults compared with control.
We then tested whether we were able to alter this overgrowth by the addition of ecdysone. Indeed, addition of 20E to the AmnC651;Gal80TS>RpS6 RNAi restores the adults to a similar size to the AmnC651;Gal80TS control animals (Figure 6C, compare ii to iv). This suggests that the overgrowth also depends on reduced levels of ecdysone activity, as observed for the AmnC651>Dp110DN animals (shown in Figure 5A xi where body size is similar to control in AmnC651>Dp110DN +20E, Figure 5A ix). Thus the overgrowth phenotype resulting from reduction of RpS6 in the PG was sensitive to the level of 20E, which supports the hypothesis that the developmental delay associated with knockdown of RpS6 specifically in the PG is due to impaired ecdysone release and delayed metamorphosis.
Most importantly, reduction of RpS6 in the PG resulted in suppression of the cycEJP eye phenotype, with a statistically significant increase in adult eye size (Figure 6D, compare ii with iv, and quantified in 6E). Thus, the ability of the RpS6 mutant to suppress the cycEJP phenotype occurs, at least in part, through a defect in PG growth and the associated delay in development. The suppression by PG-driven RpS6 knockdown was not as strong as observed for the RpS6 mutant, which could be a consequence of the severe reduction in 20E activity in these animals (Figure 4B). As ecdysone release is required for proper morphogenetic furrow progression in eye discs [50], the drastic reduction in 20E levels in the PG-driven RpS6 RNAi animals, specifically in a background of diminished CycE levels, might also delay furrow progression. Thus, even though extra time is spent during the larval growth period, the suppression is incomplete because of the role of 20E in controlling the developmental signals required for furrow progression [50]–[51].
These data strongly support a model whereby RpS6WG1288/+ suppresses the small rough eye phenotype of cycEJP via a cell non-autonomous mechanism. Reduced abundance of RpS6 in the PG of cycEJP animals decreases PG size, ecdysone activity and consequently results in a developmental delay and time for additional growth of the eye. To definitively test this model, we examined the effect of restoring RpS6 expression in the PG of RpS6WG1288/+; cycEJP/cycEJP flies. According to the model above, if the decrease in RpS6 expression specifically in the PG is responsible for the ability of RpS6WG1288/+ to suppress the small cycEJP eye phenotype, then we would predict that restoring RpS6 expression specifically in the PG in the RpS6WG1288/+; cycEJP/cycEJP flies would prevent the developmental delay and inturn prevent the suppression of the small eye phenotype. Consistent with this, expression of RpS6 using the Phantom-Gal4 (Phm-Gal4) driver [43], a PG specific driver, resulted in ectopic expression of RpS6 in the PG (Figure S3B). Similar results were shown for enforced expression of RpS6 in the PG using PG driver AmnC651-Gal4. Restoration of expression of RpS6 in the PGs of RpS6WG1288/+; cycEJP/cycEJP flies using either the AmnC651-Gal4 (Figure 7A, compare iii with iv) or Phm-Gal4 driver (Figure 7B, compare iii with iv) prevented RpS6WG1288 from suppressing the cycEJP eye phenotype (quantified in Figure 7C). Subsequent studies demonstrated this was because enforced expression of RpS6 in the PG's of RpS6WG1288 animal prevented the developmental delay (Figure 7D, green data points and statistical analysis shown in Table S4). Together these data are consistent with the above model and unequivocally demonstrate that the ability of the RpS6WG1288/+ mutant to suppress the cycEJP phenotype is due to reduction of RpS6 abundance specifically in the PG.
In summary, these data strongly support the hypothesis that the ability of the RpS6 mutant to suppress the cycEJP small rough eye phenotype is, in large part, due to a reduction of PG size and an associated decrease in ecdysone activity, which results in an extended larval growth period that allows the eye discs extra time to grow. This model predicts that manipulation of other growth modulatory genes in the PG would also suppress the cycEJP phenotype. Indeed, consistent with this model, overexpression of UAS-Dp110DN in the PG was also able to suppress the cycEJP small rough eye phenotype (Figure 6F, compare ii with iv). As observed for the RpS6 mutant, CycE protein levels, BrdU and PH3 in the AmnC651>Dp110DN, cycEJP/cycEJP eye imaginal discs were not altered compared with cycEJP alone (Figure S5). As we do not see a significant increase in the SMW divisions in these animals, when compared with cycEJP alone, this further supports the idea that the increased time spent in the larval growth stage allows more time for division, which leads to suppression of the small eye phenotype.
Since the Minutes were first described in 1929 [2], geneticists have sought to understand the mechanisms underlying these phenotypes as an avenue toward elucidating the complex mechanisms controlling body size. More recently, heterozygous mutations in multiple Rp genes have been associated with overgrowth phenotypes [11]–[12], [20], but the underlying mechanism has remained poorly understood. We addressed this question here taking advantage of a genetic screen for modifiers of a cycE hypomorph, which identified an RpS6 mutant as a suppressor [26], to investigate possible mechanisms by which Rp mutations may contribute to overgrowth.
Our data demonstrate that Rp mutants suppress the cycE phenotype via a mechanism extrinsic to the eye, involving control of developmental timing though the PG. We propose the following model to explain this phenomenon (Figure 8). Firstly, reduced Rp levels in the PG of Rp mutant flies decreases ribosome biogenesis thus inhibiting PG growth, which in turn results in reduced ecdysone synthesis and a subsequent delay in development (Figure 8A). The extended growth period resulting from the developmental delay allows time for more cell divisions and growth in the eye, thereby allowing the eye imaginal disc to achieve normal size prior to pupation, thus suppressing the cycEJP small eye phenotype (Figure 8B). In support of the tissue extrinsic component of PG-ecdysone model, we have demonstrated that reducing RpS6 specifically in the PG suppresses cycEJP (Figure 6D), and conversely overexpression of RpS6 in the PG prevents suppression of the cycEJP by mutant RpS6 (Figure 7A–7B).
As a developmental delay is a consistent feature of Minutes, it was speculated by Brehme in 1939 that this aspect of the phenotype might be due to insufficient ecdysone (as reviewed in [47]). Our work confirms this hypothesis and importantly, also provides a framework for how the Rp Minute collection of mutants are associated with both impaired growth and, counter-intuitively, tissue overgrowth (Figure 8A). In essence final tissue/body size in a Minute fly is a product of interplay between the tissue intrinsic effect of altering Rp levels in the cells of individual tissues and the extrinsic effects of Rp mutants on hormone release (Figure 8A) and thus developmental timing. As Rps and the rRNAs are required in equimolar amounts to form functional ribosomes, the relative contribution of tissue intrinsic versus extrinsic growth requirements to final tissue/body size would be dependent on the expression level and stability of each Rp, which will dictate whether levels of the specific Rp are rate-limiting for ribosome biogenesis in a given tissue. Enlargement of tissues for any given Minute would only occur if reduction of the Rp in the affected tissue did not reduce levels below those required for tissue growth. If Rp levels were below the threshold in a particular tissue, its growth would be inhibited, effectively negating the effects of an increased larval growth period provided by the developmental delay. This is consistent with the observation that expression of a given Rp mRNA varies between tissues [52]–[54], indicating that a particular Rp may be rate limiting for proliferative growth in one tissue but not in another. For example, while all of the Minutes are developmentally delayed, wing overgrowth has not been widely described, suggesting that the reduced levels of the relevant Rp associated with the Minute in question are limiting in both the wing and PG. In contrast, RpL382b1/+ and RpL52d2/+ flies have overgrown wings [11] which suggests that the reduced level of RpL38 associated with RpL382b1/+ flies is not limiting for proliferative growth in wing discs but is limiting for PG growth, thus the extended growth period results in larger adult wings. Therefore the final size of the Minute and its individual tissues is the net effect of both the tissue extrinsic effects of reducing Rps in the PG, and the tissue intrinsic effects of reducing Rps in the cells of other tissues (Figure 8A).
The mechanisms behind maintaining body/organ size are complex, and in addition to intrinsic cellular growth rate and the time spent in the growth phase prior to pupation described above, recent studies of imaginal disc regeneration reveal that the final size of Drosophila imaginal tissues is sensitive to an overarching mechanism that slows the division rate of the non-regenerating compartments even in the event of developmental delay [55]. This may explain why the RpS6WG1288/+ mutant is able restore eye size back toward the wild type size in a background sensitised to impaired eye growth, ie., the cycEJP background, but does not normally lead to eye overgrowth or overgrowth of other tissue compartments, despite being associated with a developmental delay.
Clearly however, these final size constraints can be overridden or are not triggered in certain Minutes eg., RpL382b1/+ and RpL52d2/+ flies which have overgrown wings. In these cases the ongoing wing imaginal disc growth occurring during the extended larval period appears to be sufficient to overcome the normal size control checkpoints that normally restrict overgrowth of this tissue. Consistent with this model, knockdown of RpS6 or RpL38 specifically in the PG rather than the whole fly using the ring gland driver (P0206-Gal4) results in a smaller PG and developmental delay, which is associated with overgrown larvae (Figure S4A) and for RpL38 with significantly increased wing imaginal disc size (Figure S4B–S4C).
Together these findings demonstrate the complexities of the cell non-autonomous effects of Rp reduction on tissue growth, which has implications for many of the experimental manipulations carried out by Drosophila researchers. For example if mosaic clones are generated in the whole animal using the Minute technique to maximize size of mosaic clonal tissue, this might also impact on PG growth and have unforeseen cell non-autonomous effects on the tissue of interest, which will need to be taken into consideration.
Our studies also raise the interesting question of whether the cell non-autonomous mechanisms underlying tissue overgrowth phenotypes of Minutes described here are relevant to the mechanisms responsible for tissue-specific phenotypes associated with Rp mutations in vertebrates. These ribosomopathies [56] include abnormal erythrocyte maturation, thrombocytosis and a predisposition to leukemia, associated with Rp haploinsufficiency syndromes such as the 5q- syndrome and Diamond-Blackfan anaemia (DBA) in humans [13], [20] or nerve sheath tumours in fish [12]. We think the cell non-autonomous mechanism described herein is unlikely at least for the 5q- syndrome, as the pathogenesis of ribosomal protein-mediated bone marrow failure appears to be largely cell intrinsic involving ribosomal stress mediated activation of p53 and defective development of haematopoietic system [57]. This is not to say that cell extrinsic effects of ribosomopathies may not contribute to development defects and disease at some level in vertebrates, for example, through defective growth of tissues important for release of paracrine or endocrine acting hormones. Clearly additional studies are required to determine to what extent altered Rp gene dosage contributes to human disease other than bone marrow failure and whether they are mediated by cell intrinsic or extrinsic mechanism or, indeed both.
In summary, our findings establish that suppression of cycEJP by the RpS6 mutant is exerted via a mechanism wherein reduced Rp levels in the prothoracic gland decreases abundance of the steroid hormone ecdysone, delaying development and hence allowing additional time for tissue and organ overgrowth. These data provide for the first time a rationale to explain the counter-intuitive organ overgrowth phenotypes observed for certain Drosophila Rp mutants. Furthermore, they provide new insight into mechanisms governing tissue size homeostasis, suggesting that different tissues may exhibit distinct thresholds of expression of individual Rps. Thus, regulated expression of individual Rps could exert tissue specific effects on cell growth and organ size.
Unless otherwise stated the fly strains used were obtained from the Bloomington Stock Center and are described in FlyBase (http://flybase.org). The UAS-RpS6 transgenic lines for overexpression were generated by cloning the full-length RpS6 cDNA into pUAST and then injected into Drosophila embryos, as previously described in [58]. The following strains were described in: w−;cycEJP [29], AmnC651-Gal4 [41], Phm-Gal4 [43], P0206-Gal4 [42]–[43], UAS-Dp110DN [59], UAS-RasV12 [60], UAS-Cyclin E [61], UAS-p35 [62], GMR-p21 [63], UAS-cycD and UAS-cdk4 [64].
RpS6 RNAi construct: the longest open reading frame for RpS6 (654 bp) was PCR amplified with primers 5′-CTGCAGGAATTCGGACAGGTTGTGGAGGCCGAT-3′ and 5′-GGTACCGAATTCTTACTTCTTGTCGCTGGAGACAG-3′ (EcoRI sequence underlined) and PCR products were digested with EcoRI and ligated into the SYMpUAST vector [65].
RpS13, RpL5, RpL30 and RpL38 RNAi constructs: products were digsted with XbaI and inserted into pWIZ as inverted repeats in a two–step cloning process [66]. RpS13: the 302 bp coding region of the 3rd exon was PCR amplified with primers 5′-ATATTCTAGAGCATCATCCTGCGTGACTCGC-3′ and 5′-ATATTCTAGAGGCAACCAGGGCGGAGGC-3′ (XbaI sequence underlined). RpL5: the 264 bp coding region of the 2nd exon PCR amplified with primers 5′-GCGCTCTAGAGGTTTCGTTAAGGTAGTC-3′ and 5′-GCATTCTAGACTGGATCCCGTATTTGGG-3′. RpL30: the 199 bp 5′UTR and coding region of the 1st exon was PCR amplified with primers 5′-GCATTCTAGATCGCCTGCAGTGCTTTAACC-3′and 5′-ATATTCTAGACTCAGGGCGGGCGTGTTGC-3′. RpL38: the 213 bp coding region of the 2nd exon was PCR amplified with primers 5′-GCGCTCTAGAATGCCACGGGAAATTAAAG-3′ and 5′-GCGCTCTAGATTATTTCACCTCCTTTAC-3′. All constructs were injected into Drosophila embryos, as previously described in [58].
Conditional expression of UAS-RpS6 RNAi was carried out using a temperature sensitive isoform of Gal80, the repressor of Gal4 (Gal80TS [49]). Larvae were raised at the permissive temperature of 18°C and shifted at late 2nd instar to the restrictive temperature of 25°C.
For each experiment, forty 1st instar larvae were collected 24 hour AED (0–4 hour collections) from lay cages with grape agar plates. To measure time of eclosion, vials were checked for the number of eclosed adults every 2 hours from 10 days AED until adult flies no longer emerged. For 20-hydroxyecdysone treatment twenty 1st instar larvae were collected 24 hour AED (0–4 hour collections) and transferred into vials containing yeast paste supplemented daily with 0.75 mg/ml of 20-hydroxyecdsyone (Sigma).
Antibody staining, BrdU labelling and quantification were carried out as described previously [67]–[68]. Antibodies used were the anti-RpS6 polyclonal (raised in mice), anti-bromodeoxyuridine (Becton Dickinson), PH3 (Upstate) and anti-cycE (rat) (a gift from Helena Richardson). Serial sections of eye imaginal discs or prothoracic glands were taken on a Zeiss Imager Z1 using the LSM 510 Meta software. Image preparation and analysis were conducted in Adobe Photoshop CS2 v9.0 and ImageJ v1.37.
For light microscopy images were captured on an Olympus DP11 camera. Female adult eyes were imaged at 5.6× magnification, and larvae or adult flies were imaged at 1.6× magnification. All images were processed using Adobe Photoshop. Eye area was measured by tracing around the perimeter of the photoreceptor cells of cropped images using Metamorph Offline version 7.6.3.0 software.
For electron microscopy female adult flies were progressively fixed in 25% (v/v) acetone for 1 hour nutation at room temperature, 50% (v/v) acetone for 1 hour nutation at room temperature, 75% (v/v) acetone for 1 hour nutation at room temperature, and finally stored in 100% acetone. The sample was then critical point dried on a Balters CPD 030 Critical Point Dryer and coated with gold particles in an Edwards 6150B Gold Sputter Coater. Images were recorded on a Phillips XL30 FEG Field Emission Electron Microscope.
For measurements of prothoracic gland (PG) area size, confocal images of PGs taken at 40× magnification were quantified with BB Thermometer v1.1 Software (BenBritten.com).
Adult wings were mounted into Canada Balsam and xylene. Images were taken at 4.5× magnification. Whole wing area was measured using the magnetic lasso tool and record measurement function of Adobe Photoshop.
Total RNA was isolated from ten 3rd instar larvae or thirty 3rd instar eye imaginal discs with TRIzol (GibcoBRL) following manufacturer's instructions. cDNA was synthesised from 1 µg RNA using the Superscript First Strand synthesis system for RT-PCR (Invitrogen) following the manufacturer's guidelines. qRT-PCR was carried out with SYBR Green under standard conditions in the Step One Plus Real Time PCR system (Applied Biosystems)
Primer sequences were as follows:
RpS6 forward (TGTTCCAGCTCAACGTTTCCT)
RpS6 reverse (TCGTCGACCACTTCGAATAGC)
Actin 5C forward (CCCCAAGGCCAACCGTGAGA)
Actin 5C reverse (ACCCGAAGCGTACAGCGAGAGC)
E74B primers as published in [41].
Amplicon specificity was verified by melting curve analysis (65 to 90°C) after 40 cycles. An average Ct value for the three technical replicates was calculated for each sample. The fold change expression of RpS6 mRNA levels was normalised to equal RNA and determined using the 2−ΔΔCT method. E74B mRNA levels were normalised to Actin 5C mRNA levels of untreated control cells and determined using the 2−ΔΔCT method [69].
Statistical analysis was performed in GraphPad Prism software using either Unpaired t-test or One-way ANOVA, with Tukey's test for multiple comparisons, as stated in figure legends.
|
10.1371/journal.pntd.0003353 | Genetic Diversity and Population Structure of Trypanosoma brucei in Uganda: Implications for the Epidemiology of Sleeping Sickness and Nagana | While Human African Trypanosomiasis (HAT) is in decline on the continent of Africa, the disease still remains a major health problem in Uganda. There are recurrent sporadic outbreaks in the traditionally endemic areas in south-east Uganda, and continued spread to new unaffected areas in central Uganda. We evaluated the evolutionary dynamics underpinning the origin of new foci and the impact of host species on parasite genetic diversity in Uganda. We genotyped 269 Trypanosoma brucei isolates collected from different regions in Uganda and southwestern Kenya at 17 microsatellite loci, and checked for the presence of the SRA gene that confers human infectivity to T. b. rhodesiense.
Both Bayesian clustering methods and Discriminant Analysis of Principal Components partition Trypanosoma brucei isolates obtained from Uganda and southwestern Kenya into three distinct genetic clusters. Clusters 1 and 3 include isolates from central and southern Uganda, while cluster 2 contains mostly isolates from southwestern Kenya. These three clusters are not sorted by subspecies designation (T. b. brucei vs T. b. rhodesiense), host or date of collection. The analyses also show evidence of genetic admixture among the three genetic clusters and long-range dispersal, suggesting recent and possibly on-going gene flow between them.
Our results show that the expansion of the disease to the new foci in central Uganda occurred from the northward spread of T. b. rhodesiense (Tbr). They also confirm the emergence of the human infective strains (Tbr) from non-infective T. b. brucei (Tbb) strains of different genetic backgrounds, and the importance of cattle as Tbr reservoir, as confounders that shape the epidemiology of sleeping sickness in the region.
| Human African Trypanosomiasis (HAT) is a major health problem in Uganda, as there are recurrent sporadic outbreaks of the disease in traditionally endemic areas in south-east Uganda, and continued spread to new unaffected areas in central Uganda. In this study, we evaluate the evolutionary dynamics underpinning the origin of new disease foci and the impact of host species on parasite genetic diversity in Uganda. We found three distinct genetic clusters of T. brucei in Uganda and southwestern Kenya. Clusters 1 and 3 include isolates from central and southern Uganda, while cluster 2 contains mostly isolates from southwestern Kenya. These three clusters are not sorted by subspecies designation (T. b. brucei vs T. b. rhodesiense), host or date of collection. Our results show expansion of the disease to new foci in central Uganda occurred from the northward spread of T. b. rhodesiense. They also confirm the emergence of the human infective strains from non-infective T. b. brucei strains of different genetic backgrounds, and the importance of cattle as Tbr reservoir, as confounders that shape the epidemiology of sleeping sickness in the region.
| Trypanosoma brucei is a unicellular protozoan parasite, which causes human and animal trypanosomiasis in tropical Africa, transmitted by tsetse flies (Glossina spp). Trypanosoma brucei consists of three subspecies: T. b. brucei (Tbb), T. b. gambiense (Tbg), and T. b. rhodesiense (Tbr) that are morphologically indistinguishable and classified according to host specificity, type of disease, and geographical distribution [1–3]. Tbr and Tbg cause the acute and chronic forms of Human African Trypanosomiasis (HAT), respectively. Tbr is restricted to certain regions of East Africa, while Tbg is more widespread in West and Central Africa. Both forms of HAT have an overlapping distribution with the non-human infective Tbb, which infects a wide range of wild and domestic animals across the tsetse belt of tropical Africa and is one of the causative organisms of African Animal Trypanosomiasis (AAT) or Nagana. Both Tbr and Tbb can co-occur in the same non-human hosts as well as in the tsetse vector. However, recombination is known to happen only in the salivary glands of the tsetse [4]. Tbr is not a reproductively isolated taxon but regarded as a host-range variant of Tbb [5–7]. A single gene encoding the Serum Resistance Associated (SRA) protein allows Tbr to survive in humans [8]. This gene possesses two main alleles across the Tbr distribution [6–7] The human serum resistance associated gene is ubiquitous and conserved in Tbr throughout East Africa [6]and could potentially be spread naturally by genetic exchange between Tbr and Tbb [9].
While HAT is in decline on the continent of Africa [10], the disease still remains a major health problem in Uganda, characterized by recurrent sporadic outbreaks in the traditional endemic areas and spread to new unaffected areas in central Uganda [11]. Uganda is currently the only country in sub-Saharan Africa known to harbor all three subspecies of T. brucei. The locations of districts affected by HAT are shown in Fig. 1 [11–15]. During most of the 20th century, Tbr was limited to south-east Uganda in the old foci of Busoga (BS) and Bugiri (BG), and in areas bordering Tanzania and Kenya, such as Busia (BU), By the late 1980’s HAT appeared in Tororo (TR) and by 1998, HAT cases began to spread north and west being recorded in the Soroti (SR) district, north of Lake Kyoga in Central Uganda. From 2004 to date, all the districts in central Uganda—Kaberamaido (KA), Dokolo (DK), Lira (LR), Apac (AP), Kole (KO)—have reported HAT cases [15]. The affected areas increased in size from 13,820 to 34,843 km2, doubling the human population at risk [14]. Tbr and Tbg are now less than 120km apart. We refer to these foci in central Uganda as the new foci (Fig. 1). The epidemics in the new foci have been attributed to import of cattle carrying Tbr from disease endemic areas in the south [11], although recent work on the tsetse vector, Glossina fuscipes fuscipes, suggests that movement of susceptible flies from south to north could also be implicated in the emergence of disease in new foci [16–19]. Analyses of microsatellite and mitochondrial haplotype data show that the populations of G. f. fuscipes north and south of Lake Kyoga are genetically distinct and have identified long distance dispersal events [16, 17].
Population genetics studies have been carried out on T. brucei isolates across Africa, including HAT foci in Uganda and western Kenya. Analysis of Tbr isolates from the old foci in southeastern Uganda (BS, BG, BU, Fig. 1) by isoenzyme, RFLP, and microsatellite analyses show that they are relatively heterogeneous [20–25]. Genotype has been correlated with clinical presentation in patients and virulence in experimental mice [24]. Although it is assumed that Tbr spread from the old to the new foci, Tbr isolates from Soroti and Tororo (SR and TR respectively, Fig. 1) were genetically distinct from those in the old foci, but closely related to each other [25], which concurs with the idea that Tbr was introduced into Soroti via cattle from Tororo [10]. Microsatellite analysis (7 loci) of Tbr populations from Tororo/Soroti and Malawi showed that levels of genetic diversity were much higher in the Malawi focus, with evidence of recent genetic exchange between isolates [25]. The lack of genetic exchange and clonal, epidemic population structure of Tororo/Soroti Tbr agrees with the conclusions of previous population genetics studies [22, 23]. Thus, the local population structure of Tbr seems to depend on the relative amounts of clonal versus sexual reproduction, driven by transmission dynamics specific to the local conditions.
In this paper we used a set of 17 highly variable microsatellite loci [26–28] to investigate the patterns of genetic variation among 269 Tbb and Tbr isolates from Uganda and the neighbouring region of western Kenya in order to understand the extent of genetic exchange both within and between Tbb and Tbr and to investigate the origin and spread of HAT in Uganda. This is by far the most comprehensive study of genetic variation in Ugandan T. brucei yet undertaken. Understanding the population structure of T. brucei and the extent of genetic variation in both human infective and non-infective subspecies will reveal the potential for generation and spread of new human infective strains and is thus of critical relevance for disease control.
All 269 T. brucei isolate details are in Supplementary material (S1 Table). The T. brucei isolates were collected between 1959 and 2011 in 19 sites from the known parasite range in Uganda and western Kenya (Fig. 1). The isolates were obtained from various hosts (180 from humans, 57 from cattle, 1 from a sheep, 11 from pigs, 1 from a dog, 7 from wild animals and 12 from tsetse, S1 Table). Most of the samples (N = 194) were from archival cryopreserved collections, while 75 were collected in 2010 and 2011 mainly from Kole (KO) and Kaberamaido (KA). This is an important feature of this study, which aims to describe patterns of genetic variation and evolutionary processes of both Tbb and Tbr in all their potential hosts.
For these field samples, blood was collected on Whatman FTA (Fast Technology for Analysis of nucleic acids) cards (FTA is a registered trademark of GE Healthcare), which facilitates blood collection for nucleic DNA analysis. DNA extractions were carried out using DNeasy kits (Qiagen, Valencia, CA), following the manufacturers’ protocols. Other DNAs from isolates in the cryo collections were extracted by standard methods from cultured parasites (see S1 Table). For these isolates we chose material closest to the original field isolation to avoid selection bias through prolonged cell culture [29]. Trypanosome isolates from humans were collected for different studies according to local ethical guidelines and were treated anonymously.
All DNAs from the 2010 and 2011 field collections were screened using a diagnostic ITS based PCR test to separate T. brucei from other African trypanosomes [30]. All T. brucei samples were further tested for the presence of the SRA gene using the primer pairs SRA-R-SRA-F [31] and SRA H-SRA J [6]. Amplifications were carried out in a 25μl reaction volume containing 1X buffer (GoTaq colorless Promega), 1 mM each dNTP, 0.6 mM primers, 2 mM MgCl2, 0.5 mg/ml BSA and 0.5 U Go Taq polymerase. The amplification involved a denaturation step at 95°C for 2 min, followed by 50 cycles each at 95°C for 35 s, 56°C for 35 s, 72°C for 1 min, with a final extension step at 72°C for 7 min. PCR products were visualized on 2% agarose gels.
Fluorescently labelled forward primers for seventeen T. brucei microsatellite loci were used for microsatellite genotyping. Their sequence and chromosomal locations are in S2 Table [26–28]. PCR amplifications were carried out using Type-it microsatellite PCR kit (Qiagen, Germany). 1μl of genomic DNA diluted to approximately 100ng/μl was amplified using 5μl of Type-it Master Mix and 1μl each of forward and reverse primers in a total reaction volume of 15μl. PCR reactions were carried out using an Eppendorf Mastercycler Pro thermocycler (Eppendorf, Germany) under the following PCR cycling profile: initialization step of 95°C for 4 minutes, followed by twelve touch-down cycles of 95°C for 30 seconds, 60–50°C for 25 seconds and 72°C for 30 seconds, an additional 30 cycles of 95°C for 30 seconds, 50°C for 25 seconds and 72°C for 30 seconds, and a final extension step of 72°C for 20 minutes. As template concentration for DNA samples extracted from FTA cards varied, genotyping of the field samples was repeated 2–5 times and genotype calls accepted only where replicates were concordant.
PCR products were multiplexed in groups of two or three before fragment analysis and sizing by capillary electrophoresis using an automatic 3730xl DNA Analyzer (Applied Biosystems Inc.). Allele sizes were determined using Genescan ROX-500 internal size standard for loci; TB1/8, TB5/2, TB6/7, TB9/6, TB10/5, TB11/13, Tryp51, Tryp67, Tryp55, Tryp53 and Tryp59 and Liz-500 internal size standard for loci; Tryp66, Tryp54, Tryp62, Tryp59 and Tryp53. In a 96-well microtitre plate, 1 μl of PCR product was added to 9 μl formamide and 0.5ul of either ROX500 or Liz500 size standard.
Allele size calling was performed using GeneMarker version 2.4.0 (SoftGenetics, USA) and manually edited. Raw alleles were exported from GeneMarker to TANDEM version 1.0.9 [32] for allele binning. Genepop version 4.2 [33] was used to calculate number of alleles (Na), observed (Ho) and expected (He) heterozygosity levels under Hardy-Weinberg equilibrium (HWE) conditions. The same program was used to calculate allele richness (Ar; the number of alleles per locus, which is expected to be more sensitive to founder effects than is heterozygosity) and the inbreeding coefficient (Fis), one of the F statistics measuring genetic structure [34]. Fis measures the mean reduction in heterozygosity of an individual due to non-random mating in a population, thus the inbreeding within subpopulations, and ranges from -1 (all individuals heterozygotes) to +1 (no observed heterozygotes). Linkage disequilibrium (LD) was evaluated using the log likelihood ratio statistic (G—statistic) implemented in Genepop v4.2 [33].
Using the Bayesian clustering method implemented in STRUCTURE version 2.3.3 [35], patterns of population structure, individual assignment to
sampling localities, and levels of genetic admixture were tested by identifying genetic
clusters without using a priori sampling information on the number of genetic groups in the data set. Bayesian clustering implemented in STRUCTURE v2.3.3 [35] was used to assign isolates to genetic clusters (K) according to the allele frequencies at each locus. Five independent runs for K = 1–10 were carried out. For all runs, an admixture model and independent allele frequencies were used with a burn-in value of 250,000 steps followed by 1,000,000 iterations. The optimal value of K was determined using STRUCTURE HARVESTER v0.6 [36] to calculate the ad hoc statistic “ΔK” [37]. Assignment of individual strains to a given cluster and levels of genetic admixture within each individual were assessed using STRUCTURE membership coefficients (Q-values), which represent the fraction of the sampled genome that has ancestry in a given cluster.
Genetic clustering between T. brucei isolates was also determined using Discriminant Analysis of Principal Components (DAPC) implemented in the R [38] package Adegenet [39]. This method is not model based as the previous one, and thus does not make assumptions on HWE or LD. It also tends to perform better when hierarchical and clinal structure is present [40]. DAPC comprises two steps: 1) a principal component step, where the dimensionality of the multilocus allelic data is reduced to 15 principal components based on a-scores; and 2) a discriminant analysis step, where two discriminants are used to identify the linear combination of principal components from the first step that best distinguished prior groupings (populations) of individuals. The use of this multivariate approach is complementary to the STRUCTURE analysis, because of its ability to identify genetic structure in large databases without assumptions on the underlying genetic model. Thus, it is particularly suitable to identify variation between groups, while overlooking within-group variation. On the other hand, since DAPC does not specifically model for admixture, it is not suitable to identify individuals of mixed origin [40].
To measure the amount of genetic divergence among sampling localities, and the inferred genetic clusters and sampling sites, pairwise FST values and associated P values were calculated using ARLEQUIN v3.5 [41]. FST is another F-statistic measure (see above) and measures the proportion of the total genetic variance contained in a subpopulation. It ranges from 0 to 1, with high FST implying a considerable degree of differentiation among populations. Calculations to test for the statistical significance of the FST values were performed for 10,000 permutations. The same software was used to carry out a hierarchical analysis of molecular variance (AMOVA) to analyze the partitioning of the genetic variance (a) among and within the genetic clusters detected using previously described methods, (b) among and between three pre-defined groups within each genetic cluster: host (human, cattle, sheep, pig, dog, wild animals and tsetse flies), time of isolation, subspecies, and (c) among all samples based on date of collection. Samples were grouped at different time intervals (1 year, 5 years, 10 years) of collection to determine whether observed genetic variation could be attributed to temporal turnover. Each AMOVA analysis was run for 10,000 permutations with an allowable missing data level of 40%.
We used the LD bias correction method [42] implemented in LDNe [43] to estimate the effective population size (Ne) of each genetic cluster. We ran the analysis using a lowest allele frequency of 0.01.
Of the 269 T. brucei isolates analyzed, 210 (78%) were Tbr, as determined by the presence of the SRA gene. While the majority of SRA positive samples were found among human isolates, 32% (21/69) of isolates from non-human vertebrate hosts tested positive for the SRA gene (S3 Table), indicating that Tbr strains are circulating in these animals with cattle forming the largest proportion (16 of 21; 76%).
The final dataset for analysis included samples from 19 districts in Uganda and Kenya (Fig. 1), averaging 13 samples per district. The average amplification rate was 70.0% across the 17 microsatellite loci (S.E. 12.13%); the 2010/2011 field samples collected on FTA cards had variable template concentration, leading to non-amplification due to low template concentration [28]. Only two loci (Tryp66 and Tryp5_2) out of 136 pairwise comparisons showed significant values (p>0.5; S4 Table), thus suggesting that they are in linkage disequilibrium. However, as expected, due to clonal reproduction in T. brucei, all loci deviated from HWE in at least one district (S5 Table). Levels of genetic diversity were within the norm observed for diploid outbreeding organisms (Table 1). Allelic richness (AR) ranged between 2.24 and 7.35 (districts for which a single sample was collected were excluded; Table 1). Similarly, heterozygosity levels were within the norm (HE ranged from 0.34 to 0.70, HO from 0.27 to 0.57). FIS values were not high, ranging from -0.16 to 0.43 (Table 1), suggesting that inbreeding is not a major issue in this dataset. All genotypic data are submitted to Dryad (http://datadryad.org); DOI: doi:10.5061/dryad.m7q4c) [55].
Fig. 2, Table 1, and S1 Fig. show the results of the Bayesian clustering analyses as implemented in Structure; the 269 isolates are grouped in 3 genetic clusters (S1 Fig.). Clusters 1 and 3 as designated in Fig. 2 and S1 Table include isolates from mostly central and southeastern Uganda, while cluster 2 is mostly made up of isolates from Kenya. Besides geographic origin, Fig. 2 also shows the assignment of each isolate to one of the three clusters in relation to its host and taxonomic designation (Tbr vs Tbb, as assessed by the presence of the SRA gene). Tbb and Tbr samples are found together in clusters 1 and 3, indicating that Tbr strains are not genetically differentiated from the co-occurring Tbb strains; most isolates in cluster 2 were SRA positive. The results of the same analyses with samples grouped by collection date rather than geographic location is presented in S2 Fig. This Structure plot suggests that most of the early samples tend to belong to only two clusters (one and three), while samples from the early 1990’s mostly belong to the red and green cluster, although samples from the purple cluster still occur at these later dates. Interestingly the early samples were collected mostly from the Busia district in Uganda and Kenya. Temporal isolates from this region group in different clusters (Table 1), suggesting strain turnover in that region, although this analysis only shows a qualitative pattern (see results of the AMOVA analyses below). We also ran the same analyses omitting all the Kenyan samples to explore if without them we could detect additional subdivisions within the Uganda samples, but recovered only the same two clusters as in the analyses including all the samples (S3 Fig.). Note that the Structure results in Figs. 2 and 3 are not directly comparable, as the dataset and number of optimal clusters differ between the two analyses.
Within sampling sites, individuals with varying degrees of assignment to each of the three genetic clusters co-occur (S1 Table). This implies that, although the clusters are genetically distinct, genetic admixture is occurring. This is also evidence of recent long-range dispersal. An example of this phenomenon is the presence in a given sampling locality of (1) individuals with 100% assignment to a different genetic cluster than other samples from the same locality, and (2) genetically admixed individuals, likely the result of mating between local and immigrant strains. Importantly, localities in the southeastern (Busoga, BS, Busia, BU, Tororo, TR, Fig. 1), and central (Soroti, SR, Kaberamaido, KA, and Dokolo, DK, Fig. 1) Ugandan foci share strains from both cluster 1 and 3 (only one strain from cluster 2), implying that the strains from the old and new foci are not genetically distinct.
The southwestern Kenyan samples mostly belong to cluster 2 (Figs. 2 and 3), although a few individuals with genetic assignment to cluster 1 (blue bars in Fig. 2) can also be found in this region. Similarly, a few individuals from cluster 2 (both pure and admixed) can be found in central and southeastern Uganda, suggesting ongoing gene flow in both directions, even though most of the Kenyan and southern Uganda isolates belong to two different genetic clusters.
Fig. 3 shows the results of DAPC clustering of the same isolates, and confirms the identification of three distinct genetic clusters identified by the Bayesian based Structure analyses with the large majority of the individuals belonging to the same 3 clusters identified by structure. Table 1 reports the assignment of each isolate to the 3 clusters by both methods.
FST values between sampling sites ranged from 0 to 0.67 (S5 Table), and FST values between the three structure and DAPC inferred clusters ranged from 0.24 to 0.46 (S6 Table). The occurrence of statistically significant FST values among the three structure/DAPC inferred clusters confirms their genetic distinctiveness. The finding of relatively low and not statistically significant FST values among some of the isolates from different sampling sites and genetic clusters confirms the occurrence of genetic admixture also suggested by the Structure analysis (Fig. 2).
AMOVA results show the level of genetic diversity explained by the structure inferred genetic clusters and how much of the genetic variation is explained by collection date, species host, subspecies designation both among all the samples from the 19 sampling sites, regardless of their cluster assignation and within each of the three genetic clusters (Table 2). Most of the genetic variation was apportioned within (71.8%) rather than among the three Structure-defined clusters. Interestingly and contrary to the qualitative pattern shown in S2 Fig., very little of the observed genetic variation among the 19 sampling sites (8.49%) was explained by collection date (samples grouped in 10 year intervals), indicating that genetic variation in T. brucei is not explained by temporal turnover. This result was confirmed by carrying out the same analysis but grouping the samples at 1 and 5 year intervals (results not shown). Within the clusters, subspecies designation, date of collection, and host explained relatively little of the observed variation.
Effective population size estimates (Ne) calculated using LNDe [43] for the 3 clusters structure/DAPC inferred clusters are reported in Table 3 together with their confidence intervals. Ne were smaller in clusters 1 and 2 (13.1 and 8.1, respectively) than in cluster 3 (44.3; Table 3). As the confidence intervals around these estimates were relatively narrow, all clusters differed significantly (p<0.05) in effective population size. The small Ne suggests that clusters 1 and 2 represent clonal/rapid expansions, while the larger Ne observed for cluster 3 implies that isolates within this cluster have undergone more sexual reproduction and belong to an older established population than the isolates belonging to the other two clusters.
The aim of this study was to examine the pattern of genetic differentiation of Tbb and Tbr isolates in Uganda and western Kenya, to understand population structure and the modalities of parasite spread to help support sustainable control strategies for AAT and HAT in this region. Continent wide studies have already shown that Tbr and Tbb strains should not be treated as reproductively isolated taxa, as some Tbb strains are more closely related to Tbr strains than their conspecifics and vice versa [7]. The use of a larger number of highly variable microsatellite loci than in previous studies, coupled with a dense spatial and temporal sampling strategy, enabled us to identify three genetic partitions within the Uganda/Kenya T. brucei isolates that were not revealed by previous studies and the existence of ongoing gene flow between them (Figs. 2 and 3).
Two of the three clusters contain a mixture of Ugandan Tbr isolates from the old foci in the southeast and from the new foci in central districts, while the third cluster groups Tbr isolates from western Kenya. Thus, despite their geographic proximity and the widespread view that the Kenyan focus was an extension of that in southeast Uganda [44], the Ugandan and Kenyan Tbr populations seem to be genetically distinct, although there is evidence of genetic admixture likely via both long and short-range dispersal. From the earliest isoenzyme studies onwards, it has been clear that Tbr differs between geographically distant foci [2–4, 23–5], but more overlap might have been expected between these neighboring foci in Uganda and Kenya, which were in close contact via Lake Victoria [44]. One factor distinguishing HAT from the two areas is transmission by different tsetse species. HAT in the lakeshore region of southeast Uganda was originally transmitted by the morsitans group fly G. pallidipes [20, 44], and this fly was also the vector of HAT in South Nyanza, Kenya [3] and in Busia [44]; however, in the Alego outbreak of Tbr in Central Nyanza, Kenya, transmission was by the palpalis group fly G. f. fuscipes [45]. In Uganda, transmission of Tbr also switched to G. f. fuscipes as Tbr extended northwards into areas infested with this species from the mid 1970’s onwards [20] and G. f. fuscipes is regarded as the main HAT vector in Uganda [19, 46]. Therefore, the factor that led to genetic isolation of cluster 2 could be adaptation to transmission by a different tsetse vector, G. pallidipes. Our results clearly rule out the hypothesis that Tbr spread from its traditional focus in southeastern Uganda to western Kenya in the 1950’s along with G. pallidipes [44], and furthermore, G. pallidipes populations in Uganda and Kenya are genetically distinct [47].
Separate transmission cycles may also explain the partitioning of Ugandan Tbr isolates into two genetic clusters, despite the fact that they are now sympatric. Glossina f. fuscipes and G. pallidipes occupy different biomes, have different host-feeding preferences, and susceptibility to trypanosome infection. Therefore, a priori, divergence would be expected among the trypanosome populations adapted to transmission cycles involving either of these vectors, and a switch from transmission by G. pallidipes to G. f. fuscipes, as occurred in southeastern Uganda, would be expected to select for certain genotypes, while allowing the two divergent trypanosome populations to mix. It may also be significant that the G. f. fuscipes populations to the north and south of Lake Kyoga are genetically distinct [16, 17], implying that transmission cycles in the old and new foci were separate until trypanosomes were transferred via movement of infected humans and livestock.
In earlier studies, two Tbr genotypes circulating in the old foci were defined by isoenzyme profiles (zymodemes) and correlated with clinical presentation; the Zambezi zymodeme was associated with more chronic progression of HAT than the Busoga zymodeme [48]. Although Goodhead et al [24] found no simple correlation between zymodeme designation and clade based on 11 microsatellite loci, some population sub-structuring was evident in their analysis, and perhaps inaccuracy in zymodeme classification, which is based on relatively few informative isoenzyme loci, has obscured the relationship. Goodhead et al [24] also compared the genome sequences of one representative Busoga and Zambezi isolate and found that, although the genomes were >99.8% identical, they showed extensive chromosome-wide SNP variation. Comparison with Tbb or Tbg genomes revealed that some chromosomes were mosaics of shared alleles, suggesting that the Ugandan Tbr strains might have originated through a hybridization event between T. brucei of East and West African origin. Historically it is known that Tbg was present in the lakeshore region of southeast Uganda in the early 20th century, so it is indeed possible that introgression has occurred.
Previous studies showed that there is sub-structuring in trypanosome populations in relation to host and geography suggesting that both geography and host play a role in shaping the patterns of genetic differentiation among Tbb and Tbr isolates [23–4, 49]. Our study does not support this. Although the estimates of genetic differentiation among sampling sites are statistically significant for a number of pairwise comparisons (S5 Table), the biological significance of this result is questionable, given the AMOVA results from Table 2 which show that within each of the three genetic clusters taxonomy, date of collection, and host explain less than 16% of the overall observed genetic variation. However, it should also be noted that the results from the AMOVA analyses are somewhat weakened by the fact that the representation of time and space points or hosts is not uniform. To better address this aspect, a denser sampling scheme would have been appropriate. Unfortunately, while the spatial and host coverage could be improved by additional collections, which we plan to carry out, the temporal aspect of the study cannot be properly addressed, as additional collections are not available.
The finding of individuals of two genetic clusters in both the old and new Ugandan foci challenges previous studies [21, 24–5, 47, 49, 50], which suggested that Tbr isolates from the Ugandan old and new foci were genetically distinct. Our study, based on a much larger data set both in terms of loci and number of samples, and including both Tbb and Tbr co-occurring strains, suggests that the expansion of the disease to the new foci in central and western Uganda occurred from Tbr isolates spreading from the old to the new foci. This result is similar to what has been described to explain the spread of HAT in Tanzania, showing genetic homogeneity of Tbr isolates in the region [51]. In addition, estimates of Ne (Table 3) show that clusters 1 and 2 have much lower effective population sizes than cluster 3, indicating that clusters 1 and 2 experienced recent clonal expansion, whereas cluster 3 had a higher rate of sexual reproduction. This may also explain the discord between our results and those of others.
Our results concur with previous studies that identified Tbr epidemics involving multiple lineages [3, 20], since Tbr strains with different genetic background co-occur in both the new and the old foci (Figs 2 and 3). We found no evidence for temporal structure in Ugandan T. brucei, whereas Duffy et al [25] found evidence of genetic shifts in allelic frequencies between samples collected in 1970 and 1990, as well as very low genetic similarity between samples from the old and new Ugandan foci. Here, temporal variation does not explain the partitioning of the observed genetic variation as shown by the AMOVA (Table 2) analyses and by the occurrence in the same genetic clusters of samples collected at different time points from the same or different sampling sites (Fig. 2 and 3). Instead we found more evidence of geographic genetic structuring (S5 Table). In this sense our study parallels better the Duffy et al [25] result for the Malawi strains rather than Uganda ones, underscoring the importance of using highly variable markers for studies such as this, where genetic differentiation levels are expected to be small, given the narrow spatial and temporal scale of the study. The other important difference between these studies that may play a role in explaining the different results is that the Duffy et al [25] study was entirely focused on Tbr strains from human patients, while our study looked at the genetic differentiation of co-occurring Tbb and Tbr isolates and included 32% of Tbr strains from non-human isolates. Looking at the whole spectrum of circulating genotypes provides additional insights on the evolutionary origin of the strains and their level of genetic admixture, as this and other studies have clearly shown that Tbr strains originate from Tbb strains, when they acquire the SRA gene [7].
We agree with Duffy et al [25] that the clonal nature of T. brucei may play a very important role in shaping its population dynamics. Our data show clear evidence of linkage disequilibrium at most loci with striking differences in effective population size estimates between clusters 1 and 2 vs. cluster 3 (Ne; Table 3), which could be an example of the potential for rapid population contractions and expansion of different genotypes due to clonal reproduction. A phenomenon that can happen stochastically and could be responsible for the different Ne estimate for clusters 1 and 2 vs cluster 3.
Our results also confirm that cattle are an important reservoir for Tbr and thus are likely to fuel the epidemiology of sleeping sickness in Uganda as 28.1% of the T. brucei isolates found in cattle (16/57) this study were SRA positive (S1 Table). Importantly, in the Structure analyses they clustered together with the human isolates from the same geographic regions (Fig. 2), suggesting ongoing genetic exchange between T. brucei isolates from cattle and humans in the same area. This result provides the first empirical confirmation that cattle are an important intermediary in HAT transmission in the region [23], as suggested in connection with earlier Tbr outbreaks [3, 44] and more recently for the current epidemics in Soroti and Kaberamaido [11, 52]. As the distance separating the Tbr and Tbg foci in North western Uganda is less than 100 km [11, 14], understanding the role and impact of cattle in fueling movement of Tbr strains is of paramount importance, as these results suggest that continued cattle movement from southern districts can expedite the fusion of the two disease belts with unknown public health consequences. Similarly, increased livestock trade across southeastern Uganda and Western Kenya also poses a risk transferring Tbr from the old Uganda HAT foci in that region to Western Kenya, which has been reporting low HAT prevalence in the last decade [53–54].
In conclusion, this study shows that there is genetic structuring within T. brucei populations from Uganda and Kenya, separating the isolates into three groups. We found clear evidence of ongoing genetic admixture and long-range dispersal among Tbb and Tbr strains. The use of a dense sampling scheme and highly variable loci enabled us to detect genetic exchange between the old and new Uganda disease foci, possibly mediated by cattle movements across the region as both Tbb and Tbr strains were found circulating in cattle. These results have important implications for disease control, as they provide empirical evidence for the occurrence of genetic exchange between co-occurring human infective and non-infective strains, and the role of cattle in spreading the human disease. The study also emphasizes the importance of studying both Tbb and Tbr strains when attempting to understand the population dynamics of Tbr.
|
10.1371/journal.pgen.1007711 | Gata4 regulates hedgehog signaling and Gata6 expression for outflow tract development | Dominant mutations of Gata4, an essential cardiogenic transcription factor (TF), were known to cause outflow tract (OFT) defects in both human and mouse, but the underlying molecular mechanism was not clear. In this study, Gata4 haploinsufficiency in mice was found to result in OFT defects including double outlet right ventricle (DORV) and ventricular septum defects (VSDs). Gata4 was shown to be required for Hedgehog (Hh)-receiving progenitors within the second heart field (SHF) for normal OFT alignment. Restored cell proliferation in the SHF by knocking-down Pten failed to rescue OFT defects, suggesting that additional cell events under Gata4 regulation is important. SHF Hh-receiving cells failed to migrate properly into the proximal OFT cushion, which is associated with abnormal EMT and cell proliferation in Gata4 haploinsufficiency. The genetic interaction of Hh signaling and Gata4 is further demonstrated to be important for OFT development. Gata4 and Smo double heterozygotes displayed more severe OFT abnormalities including persistent truncus arteriosus (PTA). Restoration of Hedgehog signaling renormalized SHF cell proliferation and migration, and rescued OFT defects in Gata4 haploinsufficiency. In addition, there was enhanced Gata6 expression in the SHF of the Gata4 heterozygotes. The Gata4-responsive repressive sites were identified within 1kbp upstream of the transcription start site of Gata6 by both ChIP-qPCR and luciferase reporter assay. These results suggested a SHF regulatory network comprising of Gata4, Gata6 and Hh-signaling for OFT development.
| Gata4 is an important transcription factor that regulates the development of the heart. Human possessing a single copy of Gata4 mutation display congenital heart defects (CHD), including double outlet right ventricle (DORV). DORV is an alignment problem in which both the Aorta and Pulmonary Artery originate from the right ventricle, instead of originating from the left and the right ventricles, respectively. In this study, a Gata4 mutant mouse model was used to study how Gata4 mutations cause DORV. We showed that Gata4 is required in the cardiac precursor cells for the normal alignment of the great arteries. Although Gata4 mutations inhibit the rapid increase in the cardiac precursor cell numbers, resolving this problem does not recover the normal alignment of the great arteries. It indicates that there is a migratory issue of the cardiac precursor cells as they navigate to the great arteries during development. The study further showed that a specific molecular signaling, Hh-signaling and Gata6 are responsible to the Gata4 action in the cardiac precursor cells. Importantly, over-activation of the Hh-signaling pathways rescues the DORV in the Gata4 mutant embryos. This study provides a molecular model to explain the ontogeny of a subtype of CHD.
| Congenital Heart Defects (CHDs) occur in approximately 1% of live births [1] and are the most common serious birth defects in humans [2, 3]. Approximately one third of the CHDs involve malformations of the outflow tract (OFT), which leads to significant morbidity and mortality of children and adults [4]. Multiple OFT abnormalities involve the defective relationship between the Aorta and Pulmonary Artery to the underlying left and right ventricles. For example, double-outlet right ventricle (DORV) is an anomaly in which the Aorta and Pulmonary Artery originate from the right ventricle [4]. A key characteristic of DORV that distinguishes it from other OFT defects is that the aorta and pulmonary trunk are well separated but are improperly aligned over the right ventricle. The molecular basis of OFT misalignment has remained unclear.
SHF-derived cells migrate into the developing poles of the heart tube, to direct the morphogenesis of the cardiac inflow and outflow. The anterior SHF (aSHF) is essential for OFT and great artery development [5–9]. The failure of the aSHF-derived myocardial and endocardial contributions to the arterial pole of the heart causes a shortened OFT and arterial pole misalignment, resulting in inappropriate connections of the great arteries to the ventricular mass [10–12]. Deletion of genes responsible for SHF morphogenesis, such as Isl1, Mef2c, and Jagged1, leads to abnormal OFT formation including DORV [5, 6, 12–19]. These observations lay the groundwork for investigating the molecular pathways required for OFT development in SHF cardiac progenitor cells.
Gata4, a member of the GATA family of zinc finger transcription factors, is an essential cardiogenic transcriptional regulator implicated in many aspects of cardiac development and function [13–27]. Human genetic studies have implicated haploinsufficiency of GATA4 in human CHDs, such as atrial septal defects (ASD), ventral septal defects (VSD), and tetralogy of Fallot (TOF) [17, 28–32]. In mouse models, decreased expression of Gata4 results in the development of common atrioventricular canal (CAVC), DORV, and hypoplastic ventricular myocardium in a large proportion of mouse embryos [20, 33]. Multiple studies have demonstrated the molecular requirement of Gata4 in the endocardium for normal cardiac valve formation [15, 23, 34]. Furthermore, our previous study demonstrated that Gata4 is required in the posterior SHF (pSHF) for atrial septation. Both Hedgehog (Hh) signaling and Pten-mediated cell-cycle progression were shown to be downstream of Gata4 in atrial septation [35]. However, the mechanistic requirement for Gata4 in OFT development was not clear and seemed different as in atrial septation. For example, from the multiple Gata4 transcriptional targets that have been identified in the context of heart development, including Nppa, α-MHC, α-CA, B-type natriuretic peptide (BNP), Ccnd2, and Cyclin D2, Gli1, and Mef2c [13, 15, 16, 18, 36–38], only Mef2c has a known functional role in OFT development [12].
In this study, the mechanistic requirement for Gata4 in OFT development was investigated. Gata4-dependent pathways were revealed to be contributors to OFT development in Gata4 heterozygous mouse embryos.
Gata4 is strongly expressed in the heart, pSHF and OFT at E9.5 [20, 35, 39]. There is a gap in expression between the OFT and the pSHF at embryonic day 9.5 (Fig 1A, indicated by a “↓”). IHC staining for Gata4 at later stages during OFT development showed strong Gata4 expression in the heart, the developing OFT and the pSHF, but only in a limited subset of aSHF cells at E10.5 (Fig 1B, indicated by a “↓”). At E11.5, both the chamber myocardium and the developing OFT had strong Gata4 expression. However, Gata4 expression was absent from the cardiac neural crest (CNC)-derived distal OFT (Fig 1C, indicated by a “↓”).
Gata4 was previously reported to be required for OFT alignment [20]. To study the role of Gata4 in OFT development, Gata4 heterozygotes were examined for OFT defects. As described previously [35], Gata4 heterozygotes were generated by crossing Gata4fl/+ with Ella-Cre+/-, which drives Cre expression in the germline [40] to produce the germline Gata4 deletion. The Gata4 germline deletion was ensured by genotyping using the embryonic tail DNA (Fig 1S). Whereas Gata4fl/+ (n = 13) and EllaCre/+ (n = 12) embryos demonstrated normal heart at E14.5 (Fig 2A and 2A’, 2B and 2B’), 61.1% of Gata4+/-; Ella-Cre+/-embryos demonstrated VSD and DORV (Fig 2C’, 11/18, P = 0.0004). Consistent with prior work, primum ASDs with an absence of the DMP observed in 8 out of 18 Gata4+/; Ella-Cre+/-embryos [35] (Fig 2C). In addition, 37.5% of these embryos displayed A-V cushion defects (3 out of 8, Fig 2C vs. 2A) and 62.5% expressed right ventricle hypoplasia (possibly right ventricle non-compaction) (5 out of 8, Fig 2C vs. 2A).To determine the lineage requirement for Gata4 in AV septation, we analyzed mouse embryos haploinsufficient for Gata4 in the myocardium, CNC, endocardium or SHF (Fig 1S). We combined Tnt: Cre [41] with Gata4fl/+ to create Gata4 haploinsufficiency in the myocardium. Normal OFT alignment was observed in all Tnt-Cre+/-; Gata4fl/+ (12/12) and littermate control Gata4fl/+ embryos (9/9) at E14.5 (P = 1) (Fig 2E and 2E’ vs. 2D and 2D’, P = 1). Wnt1: CreERT/+ and Gata4fl/+ was combined to create Gata4 haploinsufficiency in the CNC induced by tamoxifen (TMX) administration at E8.5 and E9.5 [42, 43]. Normal OFT alignment was observed in all Wnt1-CreERT/+; Gata4fl/+ mutant embryos (24/24) and littermate control Gata4fl/+ embryos (16/16) at E14.5 (Fig 2F and 2F’ vs. 2D and 2D’, P = 1). Nfat1c: Cre [42, 43] and Gata4fl/+ were combined to create Gata4 haploinsufficiency in the endocardium. Normal OFT alignment was observed in 93.3% of the Nfatc1-Cre+/-; Gata4fl/+ mutant embryos (14/15) and all of littermate control Gata4fl/+ embryos (10/10) at E14.5 (Fig 2G and 2G’ vs. 2D and 2D’, P = 1). All embryos are viable at E14.5 and no other heart defects such as the atrial septal defects (ASDs), ventricle septal defects (VSDs), or malformations of the ventricular wall, were observed at this stage. These results demonstrated that Gata4 haploinsufficiency in the myocardium, CNC or endocardium did not contribute to abnormal OFT alignment.
We hypothesized that Gata4 is required in the aSHF for OFT alignment in aSHF-specific Gata4 heterozygous mice. This hypothesis was tested by combining Mef2cAHF: Cre with Gata4fl/+. Surprisingly, OFT misalignment with DORV was only observed in 1 out of 22 embryos and none of the littermate controls (Fig 2I and 2I’ vs. 2H and 2H’, P = 1). We next tested if Gata4 is required in the pSHF for OFT alignment in in pSHF-specific Gata4 heterozygous mice by crossing Osr1-CreERT/+ with Gata4fl/+ [44, 45]. CreERT was activated by TMX administration at E7.5 and E8.5 in Osr1-CreERT/+; Gata4fl/+ embryos to results in pSHF Gata4 happloinsufficiency [44]. Similarly, neither Gata4fl/+;Osr1-CreERT/+ embryos (0/5) nor littermate control Gata4fl/+ embryos (0/6) demonstrated OFT misalignments at E14.5 (Fig 2J and 2J’ vs. 2H and 2H’, P = 1). However, right ventricular hypoplasia was observed in 5 out of 8 embryos (62.5%, Fig 2J and 2J’ vs. 2H and 2H’). Nonetheless, these results demonstrated that Gata4 haploinsufficiency in either aSHF or pSHF supported normal OFT alignment.
Previous studies have shown that Hh signal-receiving progenitors localized in both the aSHF and the pSHF, and regulate the migration of SHF toward the OFT and inflow tract (IFT) to form the pulmonary artery and the atrial septum respectively [46–48]. We combined Gli1-CreERT2/+ with Gata4fl/+ to create Gata4 haploinsufficiency in SHF Hh signal-receiving progenitors. CreERT2 was activated by TMX administration at E7.5 and E8.5 in Gli1-CreERT2/+; Gata4fl/+ embryos. The reduced expression of Gata4 by the deletion of Gli1-CreERT2 recombination was confirmed by realtime-PCR using the SHF tissue of E9.5 embryos (Fig 1S–1G). With TMX administration at E7.5 and E8.5, 66.7% of Gli1-CreERT2/+; Gata4fl/+ embryos displayed DORV, while the littermate control Gata4fl/+ embryos displayed normal OFT alignment (Fig 2K and 2K’ vs. 2H, 2H’, 8/12 vs. 0/15, P = 0.0002). Blue ink was injected in the pulmonary artery of the Gata4fl/+ E14.5 embryos, resulting in the staining of the pulmonary artery and the right ventricle. However, the Gli1-CreERT2/+; Gata4fl/+ embryos showed staining in not only the right ventricle and the pulmonary artery, but also the aortic artery, which confirms the phenotype of DORV in these embryos (Fig 2L and 2M). In addition, when the embryos were given TMX at E8.5 and E9.5, normal OFT alignment was observed in all Gli1-CreERT2/+; Gata4fl/+ embryos (Table 1). To exclude the possibility that the phenotype might be due to the double heterozygosity for Gata4 and Gli1, the phenotype of the Gli1-CreERT2/+; Gata4fl/+ embryos without TMX treatment was examined. There were no heart defects observed in the embryos at E14.5 (Table 1). Considering that TMX activates expression 12 h after injection and that the action lasts for 36 hours [49, 50], we concluded that Gata4 is required in the SHF Hedgehog (Hh) signal-receiving progenitors from E8 to E10.5 for proper OFT alignment.
Our previous study demonstrated that Gata4 mutants disrupted cell cycle progression in the pSHF cardiac precursors resulting in atrial septal defects and genetically targeted downregulation of Pten rescued the proliferation defects in SHF of the Gata4 heterozygotes [35]. Would the defected cell cycle by Gata4 mutants lead to OFT alignment defects? In order to answer this question, the analysis was conducted to qualify if Pten downregulation (TMX at E7.5 and E8.5), could also rescue DORV in Hh-receiving cell-specific Gata4 heterozygotes. Decreased dosage of Pten caused DORV in only 1 of the 20 embryos, and none with ASD (Fig 3G, 3H and 3I). Consistent with the previous report, ASD in Gli1-CreERT2/+; Gata4fl/+ embryos was rescued by Pten downregulation (Fig 3G vs. 3D, 1/20 in Gli1-CreERT2/+; Gata4fl/+;Ptenfl/+ vs. 14/29 in Gli1Cre-ERT2/+;Gata4fl/+, P = 0.0013), but the Gli1-CreERT2/+; Gata4fl/+;Ptenfl/+ embryos displayed DORV, consistent with the incidence rate from Gli1-CreERT2/+; Gata4fl/+ embryos (Fig 3H vs. 3E, 12/29 vs. 6/20, Table 1, P = 0.5495). We next performed immunohistochemical (IHC) staining for H3S10 phosphorylation to assess the cell proliferation in the SHF at E10.5. This showed a significantly less percentile of H3S10+ cells in the SHF of the Gli1-CreERT2/+; Gata4fl/+ embryos versus the Gata4fl/+ (Fig 3R vs. 3L and Fig 3V, P = 0.013), suggesting a proliferation defect. However, this proliferation defect was restored in the Gli1-CreERT2/+; Gata4fl/+; Ptenfl/+ embryos (Fig 3U vs. 3L and Fig 3V, P = 0.500 vs. Gata4fl/+; P = 0.062 vs. Gli1-CreERT2/+; Ptenfl/+). Consistently, expression of the cell proliferation genes including Cdk2, Cdk4 and Ccnd2 was lower in the Gli1-CreERT2/+; Gata4fl/+ embryos but was restored to normal levels with a Pten knockdown (Fig 3W). This data suggested that correction of the SHF proliferation defects failed to rescue the OFT misalignment of the Gata4 mutant embryos, and thus different mechanisms were involved in the regulations of Gata4 for atrial septal and OFT.
We have previously reported that Gata4 acts upstream of Hh-signaling for atrial septation [35]. The requirement of Gata4 in Hh-receiving cells for OFT alignment suggested that Gata4 and Hh signaling might interact genetically in the SHF for OFT development. This hypothesis was tested in the Gata4 and Smo compound heterozygotes (Gli1-CreERT2/+; Gata4fl/+; Smofl/+) versus the littermate controls (Gli1-CreERT2/+; Gata4fl/+; or Gli1-CreERT2/+; Smofl/+) both induced by TMX administration at E7.5 and E8.5. Consistent OFT defects were observed in compound Gata4; Smo haploinsufficient embryos (Gli1-CreERT2/+; Gata4fl/+; Smofl/+) (5/9, Fig 4C–4E) whereas no OFT defects were observed in Gli1-CreERT2/+; Smofl/+ embryos (0/7, Fig 4B and 4B’; P = 0.0337). The total incidence of OFT defects occurred in the Gli1-CreERT2/+; Gata4fl/+; Smofl/+was not different from in the Gli1-CreERT2/+; Gata4fl/+ embryos (Fig 4C–4E, 5/9 vs. 4/6, P = 0.7326). However, a larger range of OFT defects was observed in Gli1-CreERT2/+; Gata4fl/+; Smofl/+ embryos, including DORV (3 out of 5, Fig 4C and 4C’), OA (1 out of 5, Fig 4D and 4D’) and persistent truncus arteriosus (PTA) (1 out of 5, Fig 4E and 4E’). PTA, caused by a combined defect of alignment and separation, was only observed in Gli1-CreERT2/+; Gata4fl/+; Smofl/+. This result suggested an interaction between Gata4 and Hh-signaling in OFT development.
We tested the hypothesis that Gata4 acts upstream of Hh-signaling for OFT development using a genetic epistasis study. The purpose was to understand if increased Hh-signaling via a constitutively activated Smo mutant, SmoM2 [35], induced by TMX administration at E7.5 and E8.5, could rescue the OFT misalignment in Gata4-heterozygotes. DORV was observed in 28.6% of littermate control Gli1-CreERT2/+; SmoM2fl/+embryos (2/7) (Fig 4G and 4G’) and 58.3% of littermate control Gli1-CreERT2/+; Gata4fl/+ embryos at E14.5 (7/12) (Fig 4H and 4H’). In contrast, none of Gli1-CreERT2/+; Gata4fl/+; SmoM2fl/+ embryos showed DORV (Fig 4I and 4I’), indicating significant rescue by Gli1-CreERT2/+; SmoM2fl/+(Fig 4I vs Fig 4H, P = 0.0071, Table 1). This results demonstrated rescue of DORV in Gata4-mutant embryos by constitutive Hh signaling.
Next, IHC staining for H3S10 phosphorylation in the SHF was performed to determine if the cell proliferation defects observed in the Gli1-CreERT2/+; Gata4fl/+ were rescued by overactivation of Hh-signaling. Clearly, the cell proliferation defects, indicated by less percentile of H3S10+ cells, observed in the SHF of the Gli1-CreERT2/+; Gata4fl/+ (Fig 4L” vs. 4J” and Fig 4N, P<0.01 vs. Gata4fl/+) were recovered by Gli1-CreERT2/+; SmoM2fl/+ (Fig 4M” vs. 4J” and Fig 4N, P>0.05 vs. Gata4fl/+). Consistently, gene expression was downregulated for multiple cell proliferation genes including the Cdk2, Cdk4, Cdk6 and Ccnd2 in the SHF of Gli1-CreERT2/+; Gata4fl/+ comparing to the Gata4fl/+ embryos, which was recovered in the Gli1-CreERT2/+; Gata4fl/+; SmoM2fl/+ embryos (Fig 4O). These results suggested that overactivating the Hh-signaling rescued proliferation defects in the SHF of Gata4 haploinsufficiency.
Hh signaling has been reported to regulate the migration of SHF Hh-receiving cells toward the arterial pole of the heart [46]. We therefore hypothesized that Gata4 drives SHF Hh-receiving cells migration toward the developing OFT. This hypothesis was tested by using genetic inducible fate mapping (GIFM) [51]. The Hh-receiving lineage cells were marked by TMX administration at E7.5 and E8.5 (Gli1-CreERT2/+; R26Rfl/+) and β-gal expression was evaluated at E11.5. We assessed if there was less migrating Hh-receiving SHF cells migrating through the distal OFT (dOFT) towards the proximal OFT (pOFT) in the Gata4 haploinsufficient embryos (Gli1-CreERT2/+; Gata4fl/+; R26Rfl/+) than the control embryos (Gli1-CreERT2/+; R26Rfl/+), and if this defects rescued in Gli1-CreERT2/+; R26Rfl/+; Gata4fl/+; SmoM2fl/+ embryos. Previous reports indicate a decreased number of Hh-receiving cells in the pSHF at E9.5 associated with developing defects of DMP in the Gata4fl/+; R26Rfl/+; Gli1Cre-ERT2/+embryos [35]. In concurrence, significantly less Hh-receiving cells within the aSHF region (Fig 5A vs. 5D and Fig 5J, 334.0 ± 1.4 vs. 186.7 ± 4.9, P = 0.001) of the Gli1-CreERT2/+; Gata4fl/+; R26Rfl/+ embryos was also observed. The cells of Hh-receiving lineage were analyzed in the developing OFT at this stage. By counting the number of β-galactosidase-expressing cells in the proximal half and the distal half of the OFT myocardium of the Gata4fl/+; R26Rfl/+ embryos, both of the regions of the Gli1-CreERT2/+; Gata4fl/+; R26Rfl/+ had less β-galactosidase-expressing cells than the littermate controls (Fig 5B vs. 5E and 5K, 91.7 ± 9.2 vs. 56.7 ± 1.4, P = 0.013 for dOFT; Fig 5C vs. 5F and 5L, 49.7 ± 10.6 vs. 26.7 ± 6.4, P = 0.091 for pOFT). Importantly, the lower number of Hh-receiving cells in the Gli1-CreERT2/+; Gata4fl/+ was partially recovered by overactivating the Hh-signaling in the aSHF (Fig 5A vs. 5G and 5J, Gli1-CreERT2/+; R26Rfl/+ vs. Gli1-CreERT2/+; R26Rfl/+; Gata4fl/+; SmoM2fl/+: 334.0 ± 1.4 vs. 258 ± 18.4, P = 0.028; Fig 5G vs. 5D and 5J, Gli1-CreERT2/+; R26Rfl/+; Gata4fl/+; SmoM2fl/+ vs. Gli1-CreERT2/+; R26Rfl/+: 258 ± 18.4 vs. 186.7 ± 4.9, P = 0.034). There was also complete restoration of the amount of β-galactosidase-expressing cells in the dOFT (Fig 5B vs. 5H and 5K, Gli1-CreERT2/+; R26Rfl/+ vs. Gli1-CreERT2/+; R26Rfl/+; Gata4fl/+; SmoM2fl/+: 91.7 ± 9.2 vs. 109.3 ± 19.8, P = 0.505) and the pOFT (Fig 5C vs. 5I and 5L, Gli1-CreERT2/+; R26Rfl/+ vs. Gli1-CreERT2/+; R26Rfl/+; Gata4fl/+; SmoM2fl/+: 49.7 ± 10.6 vs. 84.3 ± 15.7, P = 0.161).
To examine if a Gata4 heterozygous influenced the SHF cell recruitment within the proximal OFT, we analyzed the fate map of SHF lineage cells in the OFT of the Gata4 heterozygotes. Defined by Mef2cAHF::Cre driven β-galactosidase-expressing cells, the total number of the SHF lineage cells within the proximal and distal half of the OFT were compared between the Mef2cAHF::Cre; R26Rfl/+;Gata4+/-; and the Mef2cAHF::Cre;R24Rfl/+ embryos at E10.5. The number of SHF lineage cells populating the pOFT of the Mef2cAHF::Cre;Gata4+/-; R26Rfl/+ embryos was significantly less than those in the control Mef2cAHF::Cre; R26Rfl/+ embryos (Fig 5M vs. 5P); however, this decrement was not observed in the distal OFT (Fig 5N vs. 5Q). The distribution pattern of the SHF lineage was not different in the Mef2cAHF::Cre;Gata4+/-; R26Rfl/+ and the Mef2cAHF::Cre;R26Rfl/+embryos (Fig 5O vs. 5R). Fewer cells were observed to populate the developing dorsal mesocardium protrusion (DMP) in Mef2cAHF::Cre; Gata4+/-; R26Rfl/+(red arrow, Fig 5O vs. 5R). This was consistent with the previous report that Gata4 is required in the SHF for the DMP [35]. These results demonstrated the requirement of Gata4 for the SHF lineage cells populating in the developing OFT.
The role of Gata4 in the EMT for the endocardial cushion development has been well described previously [20, 23, 52]. Since less SHF lineage cells populating in the developing OFT myocardium were observed, we asked if this defect would affect the sequential events such as EMT, cell proliferation or cell survival in the developing OFT via a non-cell autonomous manner. Expression of mesenchymal marker N-cadherin was used to label the cushion cells undergoing EMT. LacZ staining was performed before the IHC staining, which indicated the active Cre recombination for specifically Gata4 knocking-down. The Hh-receiving cells were shown to populate at the conal OFT as early as E14.5. Significantly less N-cadherin staining of the conal OFT cushion cells in the Gli1-CreERT2/+; Gata4fl/+; R26Rfl/+ embryos versus the Gata4fl/+; R26Rfl/+ littermate control embryos at E10.5 (Fig 6A–6C vs. 6D–6E) was observed, suggesting that the EMT process of the OFT cushion was inhibited by the lower Gata4 expression in Hh-receiving cells. Cell proliferation was examined by BrdU incorporation at E11.5. Gli1-CreERT2/+; Gata4fl/+ embryos demonstrated 17% fewer BrdU-positive SHF cells in the OFT conal cushion (Fig 6G vs. 6J and 6I; P = 0.0134), but not the OFT truncal cushion (Fig 6H vs. 6K and 6L; P = 0.1998), when compared to the littermate Gata4fl/+embryos at E11.5. Cell death were assessed by TUNEL staining and no differences in either the conal or truncal cushion between Gli1-CreERT2/+; Gata4fl/+ and the Gata4fl/+embryos was observed (Fig 6M–6P). Together, these results demonstrated that Gata4 is required for normal cell EMT and proliferation in OFT conal cushion development, possibly through a non-cell autonomous manner.
Because Gata4 and Gata6 double mutant embryos display PTA [33], Gata6 expression in Gata4 mutants was examined. Gata6 was expressed in the heart, the OFT and strongly in the splanchnic mesoderm (Fig 7A, arrow), but not neural crest cell derivatives (Fig 7A, arrowhead) of the Gata4fl/+ embryo at E9.5. In Gata4 knockdown embryos specifically in the Hh-receiving cells, the Gata6 expression domain was strongly enhanced in the OFT and the splanchnic mesoderm. Consistently, enhanced expression of Gata6 in the OFT and the SHF of the Gata4fl/fl; Gli1-CreERT2/+ was further confirmed by the real-time PCR at the mRNA level (Fig 7B). The Gata6 expression in the SHF of Gata4fl/fl; Gli1-CreERT2/+ mouse embryo was increased by 1.7-fold comparing to the control Gata4fl/+ embryos (P<0.05). Gata6 expression in the OFT of the Gata4fl/fl; Gli1-CreERT2/+ mouse embryo was increased by 3.4-fold comparing to the littermate control (P<0.01). These results suggested a negative association between the expression of Gata4 and Gata6 in the SHF and developing OFT.
We tested the possibility of Gata4 regulating the expression of Gata6 in the SHF as a repressor by ChIP-qPCR using the microdissected SHF from the E9.5 wildtype mouse embryos. The Gata6 loci was bioinformatically interrogated for potential Gata4-responsive elements using the overlap of evolutionary conservation and Gata4 occupancy in HL-1 cells or embryonic mouse hearts [53, 54] (Fig 7C). Eight potential Gata4-binding regions for Gata6 were screened and our previously identified Gata4 responsive Gli1 loci [35] was used as the positive control (Gli1-ctrl). The results showed enrichments of the region Gli1-ctrl and the region Gata6-1b, but not the others (Fig 7D). This suggested that the region Gata6-1b, within 1 Kbp upstream of Gata6 start codon, was responsive to Gata4 binding. Luciferase reporter assay showed no changes of firefly luciferase activity under Gata6-1b expression in HEK293 cells (Fig 7E, Gata4+G6-luc vs. Gata4+pGL3, P>0.05). However, three mutant constructs for Gata6-1b, each ablating one Gata4-binding site, significantly enhanced the luciferase activity (Fig 7E, P<0.001 for all three mutants versus the Gata4+pGL3). Together, these results place Gata4 upstream of Gata6 as a repressor in the SHF.
The requirement of Gata4 for OFT development has been reported in mice and humans, and Gata4 mutations causing DORV have been reported in mice [19, 20, 33]. Here Gata4 is demonstrated required in the SHF Hh-receiving cells for OFT alignment. Our previous study has demonstrated that Gata4 is required for Hh signaling in the SHF for cell proliferation. However, the current study suggested that the cell proliferation defects in the SHF caused by Gata4 mutation may not be the only mechanism underlying the OFT misalignment. Another important contributing factor is the migration defect of the SHF cells, which were associated with disrupted Hh-signaling, as proved by the rescue of over-activating Hh-signaling. As subsequent events, Gata4 haploinsufficency in the Hh-receiving cells disrupted EMT process and cell proliferation in the conal OFT cushion, suggesting that both the cell-autonomous and non-cell autonomous effects of Gata4 drive OFT development. In addition, we demonstrated Gata4 as a repressor of Gata6 in the SHF by identifying Gata4-responsive binding sites in its promoter regions. This result provides a molecular explanation for the severity of OFT defects observed in Gata4 and Gata6 double mutant embryos. These data suggested that breaking down the threshold of GATA including Gata4 and Gata6, and Hh signaling tone might be associated with the severity of OFT defects.
The SHF was initially described as a progenitor field for the cardiac OFT and a rich literature has established the requirement of aSHF contributions for OFT development [5, 12, 35, 55–65]. More recently, the contribution of pSHF cardiac progenitors to the OFT and the future subpulmonary myocardium has been reported; however, the mechanistic requirement for this contribution is not well understood [46, 66–68]. The cell lineage in which Gata4 is required for OFT development has not been reported. Gata4 is expressed in both the aSHF and pSHF, although its expression is much stronger in the pSHF [35]. The decreased number of Mef2C-AHF::Cre positive cells in the proximal OFT cushion of E10.5 Gata4−/+ embryos demonstrated that Gata4 plays a role in adding the SHF progenitor cells to the developing OFT. Surprisingly, OFT defects were not observed in either aSHF-specific (Mef2c-AHF::Cre) or pSHF-specific (Osr1-CreERT) Gata4 haploinsufficiency. Instead, it was found that the severity of the OFT defects and incidence rate in embryos with Gata4 haploinsufficiency in Hh-receiving cells were identical to those in Gata4−/+ embryos. Because Hh-receiving cells are located throughout the SHF, these observations suggested that Gata4 is required in both pSHF and aSHF progenitor cells for OFT alignment.
Evidence was provided that Gata4 acts upstream of Hh-signaling in the SHF for OFT development. The Gata4−/+ embryos have combined phenotypes of ASD and DORV [35]. This study, adding new knowledge to the previous [35], have disclosure that Gata4-Hh-signaling plays active, but different, roles at the venous and atrial pole of the developing heart. In the SHF, Gata4-Hh-signaling controls cell cycle progression and thereby the proliferation of the cardiac progenitors. At the venous pole, diminished Gata4-Hh signaling for cell cycle regulation is balanced by Pten through transcriptional inhibition of Cyclin D4 and Cdk4 [13, 35], as DMP hypoplasia and SHF cell cycle defects are rescued by Pten knockdown [13, 35]. At the atrial pole, the Pten knockdown was able to rescue the cell cycle defects in the SHF, but failed to rescue DORV or OA defects in Gata4 heterozygous mutants. This observation suggests that correction of SHF cell proliferation is not sufficient to support a normal OFT development in Gata4 mutants. In this study, increased apoptosis was not observed in the SHF of Gata4 heterozygote mutant embryos [35]. Indeed, cell migration under the regulation of Gata4-Hh signaling is important for OFT development. However, fate mapping of the SHF using either Mef2c-AHF::Cre or the Gli1-CreERT2 disclosed less SHF-derived cells in the distal OFT in Gata4 mutant embryos. Specifically, there was a decreased number of SHF Hh-receiving cells throughout the migration route from the SHF into the OFT, which traveled from the dorsal mesocardium and continued through the rostral splanchnic mesoderm, past the distal OFT and reached the proximal OFT. Hh-receiving progenitors have been found to migrate from the aSHF to populate the pulmonary trunk between E9.5 to E11.5 [46], suggesting that Hh-signaling is required for SHF cell migration. The observation that DORV in Gata4 mutant embryos can be rescued by constitutive Hh-signaling is correlated with restored cell migration and cell proliferation in the SHF. Therefore, this data suggested that the normal Gata4 regulation of both the proliferation and the migration of the SHF cardiac progenitors are required for OFT development.
During development, the ventricular outlets are aligned to the ventricles by the fusion of conal cushions with the interventricular septum [69], and improper lengthening of the conal cushion may cause rotation problems resulting in misalignment. We demonstrated that Gata4 plays a non-cell autonomous role in the EMT process to give rise to the conal cushions mesenchyme. It is still unknown what specific signals the SHF-derived OFT Gli1+ myocardial cells provide for EMT. Future studies should aim to identify the ligands secreted by Hh-receiving cells. Moreover, a smaller percentage of BrdU+ cells in the conal cushion of the OFT was found at E11.5 of the Gata4fl/+; Gli1Cre-ERT2/+ embryos, suggesting Gata4 plays a role in regulating the OFT cushion cell proliferation. Inactivating N-Cadherin in the SHF resulted in hypoplastic OFT and right ventricle, associated with decreased proliferation [70, 71]. Thus, the lower number of proliferating cells might be associated with the lower expression of N-Cadherin. Overall, cellular, molecular and genetic evidence proved that Gata4-Hh signaling is required in OFT alignment via both the cell-autonomous and non-cell autonomous manners.
Although important Gata4 transcriptional targets in the heart have been identified [13, 18, 37], Gata4-dependent molecular pathways required for OFT development remain unknown. Gli1 was previously identified as a downstream target of Gata4 in the pSHF for atrial septation [35]. In the current study, it was further demonstrated that Gata4 controls Hh-signaling though Gli1 transcriptional regulation for cell migration and OFT alignment. In addition, Gata4 was demonstrated to be a transcriptional repressor of Gata6 in the SHF, and the Gata4-responsive sites in the Gata6 promoter region were identified. Previously, several downstream targets of Gata4 including the Mef2c and Gli1 have been recognized [13, 15, 16, 18, 36, 37], while none of them respond to the inhibitory effects. Gata6 is the first identified repressing target of Gata4, providing direct evidence that Gata4, as a transcription factor, is not only an activator but also a repressor. Enhanced Gata6 expression in Gata4 mutants might illustrate a compensatory feedback loop, given that Gata6 and Gata4 are redundant for cardiac myocyte differentiation [72, 73]. Gata4/Gata6 compound heterozygotes displayed persistent truncus arteriosus (PTA), a severe OFT defect caused by combined alignment and OFT septation defects [33]. This study shows that Gata4/Smo compound heterozygotes show a similar phenotype. Gata4 heterozygote alone does not display PTA, which might be due to the partial recovery of GATA function from enhanced Gata6 expression. Together with previous study [33], these data suggest a threshold of Gata4, Gata6, and Hh signaling and that is required for OFT development. This implies that GATA TFs may be essential for the quantitative regulation of Hh signaling, and diminished GATA function or reduced GATA and Hh signaling together may cause more severe OFT defects. Future studies will focus on the quantitative relationship between GATA tone and Hh signaling tone, as well as the Gata4 dependent gene regulatory network (GRN) [74] for OFT development.
All mouse experiments were performed in a mixed B6/129/SvEv background. Gata4fl/+, Gli1-CreERT2/+, Mef2cAHF::Cre, Smofl/+ mouse lines were kind gifts from Dr. Ivan Moskowitz lab (University of Chicago, Chicago). TnT-Cre+/- mouse line was from Dr. Yiping Chen lab (Tulane University, New Orleans). Nfat1c-Cre+/- mouse line was from Dr. Bin Zhou lab (Albert Einstein College of Medicine, Bronx, NY). The SmoM2fl/+, Osr1-CreERT/+ and EIIa-Cre+/- mouse lines were purchased from the Jackson Laboratory.
Mouse experiments were completed according to a protocol reviewed and approved by the Institutional Animal Care and Use Committee of Texas A&M University (#2015–0398), in compliance with the USA Public Health Service Policy on Humane Care and Use of Laboratory Animals.
TMX-induced activation of CreERT2 was accomplished by oral gavage with two doses of 75 mg/kg TMX at E7.5 and E8.5 [44, 46]. X-gal staining of embryos was performed as described [46]. The total number of β-gal positive cells was obtained by counting those on each individual sections and adding up all through the SHF and the OFT.
Standard procedures were used for histology and IHC. For BrdU incorporation, pregnant mice were given 100mg BrdU per kg bodyweight at 10mg/mL concentration solutions at E11.25 with two doses, 3 hours and 6 hours before sacrifice, respectively. The BrdU staining was performed using a BrdU In-Situ detection kit (EMD Millipore). For TUNNEL staining, an ApopTag plus peroxidase In-Situ apoptosis detection kit was used (EMD Millipore). IHC was performed using the following antibodies: anti-Gata4 (Abcam, #ab84593), anti-Gata6 (Abcam, #ab175349), and anti-N-cadherin (Abcam, #ab18203). After incubating with the first antibody, a VECTASTAIN ABC HRP Kit (LifeSpan BioSciences, Inc) was used for detecting the protein expression signal. For counting the ratio of proliferating cells, a total of 100 random cells within the SHF and the specific OFT regions per each section were counted using the Particle Analysis tool of ImageJ and the ratio of positively stained cells was recorded. For each sample, five equivalent serial sections were counted and the averages were taken for statistical analysis.
To obtain the pSHF splanchnic mesoderm for use in quantitative realtime-PCR, E9.5 embryos were dissected as described before [45, 75]. The heart, aSHF, and pSHF were collected separately in RNA-later, and then stored at −20°C until genotyping was completed.
Total RNA was extracted from the PSHF regions of mouse embryos hearts using RNeasy Mini Kit (QIAGEN), according to the manufacturer’s instructions. Two hundred ng of total RNA was reverse transcribed using a SuperScriptTM III Reverse Transcriptase kit from Invitrogen. qPCR was performed using a POWER SYBER Green PCR mater mix from Applied Biosystems. Results were analyzed using the delta-delta Ct method with GAPDH as a normalization control [76].
Chromatin Immunoprecipitation was performed as described previously [35]. The ChIP assay was performed using a Gata4 antibody (Santa Cruz, #sc-1237 X). Genomic regions with potential Gata4-binding sites and negative control sites are listed in Table 2. Primers used for evaluating the enrichment of the Gata4 pull-down fragments via realtime-PCR are listed in Table 3.
Dual Luciferase Reporter Assay was performed as described previously [35, 44, 45]. Genomic regions with potential Gata4-binding sites tested are listed in Table 2. Primers used for site-specific mutation and subclone are listed in Table 4.
|
10.1371/journal.pntd.0004387 | A Sero-epidemiological Approach to Explore Transmission of Mycobacterium ulcerans | The debilitating skin disease Buruli ulcer (BU) is caused by infection with Mycobacterium ulcerans. While various hypotheses on potential reservoirs and vectors of M. ulcerans exist, the mode of transmission has remained unclear. Epidemiological studies have indicated that children below the age of four are less exposed to the pathogen and at lower risk of developing BU than older children. In the present study we compared the age at which children begin to develop antibody responses against M. ulcerans with the age pattern of responses to other pathogens transmitted by various mechanisms. A total of 1,352 sera from individuals living in the BU endemic Offin river valley of Ghana were included in the study. While first serological responses to the mosquito transmitted malaria parasite Plasmodium falciparum and to soil transmitted Strongyloides helminths emerged around the age of one and two years, sero-conversion for M. ulcerans and for the water transmitted trematode Schistosoma mansoni occurred at around four and five years, respectively. Our data suggest that exposure to M. ulcerans intensifies strongly at the age when children start to have more intense contact with the environment, outside the small movement range of young children. Further results from our serological investigations in the Offin river valley also indicate ongoing transmission of Treponema pallidum, the causative agent of yaws.
| Buruli ulcer is a debilitating skin disease caused by Mycobacterium ulcerans. Although the understanding of this enigmatic pathogen has improved after decades of research, the mode of transmission is yet to be fully elucidated. Recent epidemiological studies have shown an underrepresentation of Buruli ulcer cases in children below the age of four as compared to older children. In order to investigate the different exposure of children to M. ulcerans and to several other pathogens with diverse modes of transmission, we conducted a sero-epidemiological study of 1,352 residents within a five kilometer radius along the Offin River of Ghana. While our results show early exposure of children to the mosquito transmitted malaria parasite Plasmodium falciparum as well as to soil transmitted helminths of the genus Strongyloides, a later onset of immune response was observed at an age of around four and five years for M. ulcerans and the trematode Schistosoma mansoni, that is transmitted by contact with infested water. Similarities in age-dependent exposure of these two pathogens suggest that transmission may take place outside of the very young children’s movement range when they come into contact with the environment at the periphery of their villages.
| Buruli ulcer (BU) is a neglected tropical skin disease presenting with a wide range of cutaneous manifestations, from non-ulcerated nodules, plaques or oedema to characteristic necrotizing ulcers [1]. While BU cases have been reported in more than 30 countries worldwide, most patients are from infection foci located in remote and rural tropical regions of West and Central Africa. BU is caused by infection with Mycobacterium ulcerans, a pathogen that has emerged from M. marinum by acquiring a plasmid conferring the capacity of producing the unique macrolide toxin mycolactone, accounting for much of the pathology of BU [2,3]. Until today, the mode of transmission of M. ulcerans has remained inconclusive, although proximity to aquatic habitats has long been identified as the major risk factor for contracting the disease [4]. Infection is thought to take place through either physical contact with undefined environmental reservoirs via skin abrasions or insect vectors [5–7].
It has long been recognized that in African BU endemic settings the majority of BU patients are children below 15 years of age [8]. However, a clear underrepresentation of children below the age of four becomes evident when the population age distribution is taken into account [9,10]. In line with this observation, our previous sero-epidemiological studies in Ghana and Cameroon have indicated that children below five years of age rarely develop antibody responses against the 18 kDa small heat shock protein (shsp) of M. ulcerans and thus seem to be considerably less exposed to the pathogen than older children [11]. While investigations of humoral immune responses against mycobacteria are complicated by a high degree of antigenic cross-reactivity between species, the immunodominant 18 kDa shsp overexpressed by M. ulcerans [12] represents a suitable marker for exposure to this pathogen [13]. It has no homologues in other prevalent pathogenic mycobacteria such as M. bovis and M. tuberculosis and additionally, sera from inhabitants of BU non-endemic regions generally showed no reactivity with this protein [13]. While populations living in BU non-endemic communities in proximity to the BU endemic regions seem to be similarly exposed to the 18 kDa shsp of M. ulcerans [14,15], we have observed for the three Ghanaian BU endemic river valleys Densu [15], Offin (this report) and Volta [15] an association between BU prevalence and the percentage of healthy individuals that have sero-converted.
Here we present a sero-epidemiological study carried out in the BU endemic Offin river valley of Ghana including 1,352 participants from 13 communities. The main objective was to compare the age-pattern of first humoral immune responses to the 18 kDa shsp of M. ulcerans with those to pathogens with various modes of transmission in order to contribute to our understanding of the transmission of M. ulcerans.
Ethical clearance for the study was obtained from the institutional review board of the Noguchi Memorial Institute for Medical Research (Federal-wide Assurance number FWA00001824). Written informed consent was received from all individuals involved in the study. Parents or guardians provided written consent on behalf of their children.
The Offin River is one of the major water bodies associated with BU in Ghana. It runs through the Ashanti and Central Regions, covering 11 health districts. In a nationwide active BU case search conducted in 1999, the Offin river valley was shown to be highly endemic for BU [16].
A total of 1,560 inhabitants of ten randomly selected communities and an additional three communities known to be BU endemic and located within a five kilometer radius along the Offin River (120 per community) (Fig 1) were randomly selected for sampling. Two milliliters of whole blood were drawn from 1,352 of 1,560 (87%) selected inhabitants in July 2013. In order to reduce dropout rates due to repetitive bleeding of participants, we randomly assigned each of the sampled individuals to one of three groups (A, B, and C) with each consisting of 450, 451, and 451 individuals to be followed up after 6 (January 2014), 12 (July 2014) or 18 (January 2015) months, respectively. Blood separation was achieved by centrifugation of the whole blood at 2’000 x g for 10 minutes. Serum was subsequently stored at -80˚C pending serological analysis. First serological analyses were performed between January and July 2014 allowing for a more thorough follow up of individual cases, such as sero-converters or sero-reverters.
ELISA was performed as described previously [11]. Briefly, 96-well Nunc-Immuno Maxisorp plates (Thermo Scientific) were coated with 0.25 μg recombinant M. ulcerans 18 kDa shsp per well, washed with washing buffer (dH2O, 0.01% Tween 20) and incubated with blocking buffer (5% non-fat dry milk in PBS). Subsequently, plates were incubated with human blood serum samples (1:100 diluted). After washing, horseradish peroxidase conjugated goat anti-human IgG (γ-chain specific, SouthernBiotech) was added. Plates were washed and developed with TMB Microwell Peroxidase Substrate (KPL). The reaction was stopped using 0.16 M sulfuric acid. The absorbance was measured at 450 nm in a Tecan Sunrise microplate reader. All samples were tested in duplicates and mean values were calculated.
The cut-off value for positivity (OD450 = 0.25) was determined by testing serum samples with a range of ODs in ELISA by Western Blot analysis (S1 Fig). Sero-conversion/reversion was defined as a change in OD (ΔOD) between baseline and follow up samples of at least ±0.3.
Western Blot analysis was performed as described [11]. Shortly, 15 μg of recombinant M. ulcerans 18 kDa shsp or P. falciparum AMA-1 were separated on NuPAGE Novex 4–12% Bis-Tris ZOOM Gels with 1.0 mm IPG well (Invitrogen) under reducing conditions. After electrophoresis, proteins were transferred onto nitrocellulose membranes using an iBlot Gel Transfer Device (Invitrogen). Membranes were blocked with 5% non-fat dry milk in PBS containing 0.1% Tween 20 and cut into strips. Strips were then incubated with human blood sera (1:1000 dilution), washed with 0.3 M PBS containing 1% Tween 20 and thereafter incubated with horseradish peroxidase conjugated goat anti-human IgG (γ-chain specific, Southern Biotech). After a second washing step, bands were visualized by chemiluminescence using ECL Western Blotting substrate (Pierce).
The contagious diseases syphilis and yaws are caused by closely related Treponema pallidum spp. Serological diagnosis requires detection of distinct antibodies against both a treponemal antigen and a non-treponemal antigen. Non-treponemal tests become reactive during the initial stage of infection and generally revert to negative after treatment. However, treponemal antigen-based confirmation is needed, since detectable antibodies can also be due to other inflammatory conditions. Treponemal serological tests on the other hand may remain reactive for life and thus require a positive non-treponemal test result to confirm active infection [17]. Here we used the Dual Path Platform (DPP) assay manufactured by Chembio Diagnostic Systems for the simultaneous detection of antibodies to treponemal and non-treponemal antigens. This serological test, which was developed for the point-of-care diagnosis of syphilis [18], was recently also applied to and evaluated for the screening and detection of patients with yaws [19]. We tested 5 μl of the serum samples in strict accordance with the manufacturer’s instructions.
ELISA results were analyzed using GraphPad Prism version 6.0 (GraphPad Software, San Diego California USA). The distribution of antibody titers is presented as box plots. These comprise a line for the median, edges for the 25th and 75th percentiles and traditional Tukey whiskers showing 1.5 times the interquartile distance. Dots on the graph represent individual points that lie outside that range. The overall difference and variation between samples tested in duplicates was estimated using the Bland-Altman method [25].
In order to assess exposure of the population living in a five kilometer radius along the Offin River to M. ulcerans, serum samples taken from 1,352 study participants of 13 communities across different age groups (Table 2) were tested in duplicates for the presence of anti- M. ulcerans 18 kDa shsp IgG by ELISA. Between the two duplicate test results, a negligible overall bias with a mean difference (OD1-OD2) of 0.004 was estimated. The variation in individual differences was very small, with 95% limits of agreement from −0.095 to 0.103. At baseline, 18% of the serum samples contained IgG titers against the M. ulcerans protein. In all, 3% (46/1,352) of the participants were BU cases with healed or active lesions and 13% (179/1,352) were household contacts of BU patients. Amongst the BU cases, 24% (11/46) had antibodies against the M. ulcerans 18 kDa shsp, while 15% (26/179) of the household contacts contained anti- M. ulcerans 18 kDa shsp titers in their sera.
The age distribution of the mean of duplicate anti- M. ulcerans 18 kDa shsp IgG titers for the study participants is shown in Fig 2A and Table 2. While in all age groups above nine years individuals with high IgG titers were detected, only three sera from children below the age of five years contained 18 kDa shsp specific antibodies and also most of the sera from children between five and seven years tested negative. The only child below four years of age with positive test result was a three year old resident of Mfantseman, a BU endemic community. We carefully followed up this particularly interesting individual case at two occasions. Before May 2013, he lived for his entire life in a BU non-endemic community (Wanpiem) not included in this study, but was transferred thereafter to the BU endemic community Mfantseman for schooling and was enrolled there in this study. Shortly after sampling his blood in July 2013, he was relocated to Wanpiem by his guardian. Scheduled for sampling after 12 months (July 2014), we followed him up at Wanpiem and observed that he had sero-reverted. His serum still tested negative for anti- M. ulcerans 18 kDa shsp IgG after 18 months of follow up (January 2015). Thus, his total length of stay in the BU endemic Mfantseman community was three months (May-July 2013); coinciding with the only sampling point at which we observed a high anti- M. ulcerans 18 kDa shsp IgG titer in his serum. Western blot results for the three time points reconfirmed the ELISA data (Fig 2B). Testing of sera from all other study participants younger than five years by Western Blot analysis revealed no specific bands (representing IgG titers) against the 18 kDa shsp for sera from children below four years of age (S3 Fig), while positive sera were detected in all age groups ≥4 years (S3 and S4 Figs), indicating that sero-conversion may start in some children around the age of four years.
Follow up samples were taken from 319 of 450 (71%), 329 of 451 (73%), and 356 of 451 (79%) participants assigned to groups A, B and C after 6, 12, or 18 months, respectively (Fig 3A). Testing of the serum samples for the presence of anti- M. ulcerans 18 kDa shsp IgG revealed that based on the defined value for the difference between paired samplings (ΔOD = 0.3), antibody titers remained stable for 98%, 97%, and 92% of the study participants, respectively. After 6, 12, and 18 months 1%, 0.3%, and 6% of the sero-positive individuals had sero-reverted, while 1%, 3% and 1.4% of the sero-negative participants had sero-converted (Fig 3B–3D).
Since only a minority of individuals exposed to M. ulcerans develops clinical BU, sero-epidemiological studies represent a valuable tool to assess the exposure of populations in BU endemic areas to the pathogen. In line with data obtained from previous sero-epidemiological investigations in BU endemic areas located in the Densu river basin of Ghana and in the Mapé river basin of Cameroon [11,15], we reconfirmed here for the population of the BU endemic Offin river valley that young children below four years of age are considerably less exposed to M. ulcerans than older children. In contrast, our serological data showed, as expected, that exposure to the mosquito transmitted malaria parasite P. falciparum and to soil-transmitted helminths of the genus Strongyloides takes place already in very young children, as indicated by an early development of humoral immune responses against these pathogens in some of the infants. The delay in exposure of M. ulcerans and the relatively abrupt onset are in stark contrast to the age-patterns for Plasmodium and Strongyloides. Contact with larvae-infested soil through faecal contamination is likely to be responsible for the observed early development of anti- Strongyloides serum IgG responses. Our results suggest that contact with M. ulcerans occurs outside the small movement range of infants, providing indirect evidence against mechanisms of transmission involving vectors or reservoirs present in the vicinity of the children’s homes. However, our data do not exclude an involvement of insect vectors commonly found at the periphery of villages close to water contact sites. We recognize that the age distribution of anti- P. falciparum serum antibodies depends not only on the mode of transmission, but also on the transmission intensity [27,28]. While there is some evidence that mosquitos may be involved in transmission of M. ulcerans in south-eastern Australia [29,30], it is overall unlikely that they play a major role as vectors in African BU endemic settings. This assumption is also supported by previous molecular epidemiological studies showing that transmission of newly emerging genetic variants of M. ulcerans is geographically highly clustered [31,32].
In Ghana schistosomiasis is mainly caused by S. haematobium and S. mansoni [33]. The egg forms of the parasite are shed into the environment via urine or faeces of an infected person. Through a complex life cycle involving an intermediate snail host, a healthy individual can be infected by coming into contact with water sources infested with invasive larvae [34]. Therefore, the risk of infection is related to water contact patterns. Peak prevalence is usually observed in school-aged children, but may be shifted to adulthood depending on the degree of endemicity in a setting [35]. In the present study we observed in the Offin river basin the typical age distribution of Schistosoma exposure with children below the age of eight being significantly less affected than older children and young adults. This pattern was similar to the age distribution of exposure to M. ulcerans in the same population. While the mechanisms of infection by schistosomes and M. ulcerans are likely to be different, similarities in the age-dependent patterns of exposure may be related to changes in water contact patterns. It remains to be elucidated whether infection with M. ulcerans from an environmental reservoir takes place through skin lesions or via invertebrate vectors, such as aquatic insects. In addition, our data indicate that anti- M. ulcerans 18 kDa shsp IgG titers are relatively stable with only 2%, 3% and 8% of individuals followed up after 6, 12 and 18 months, respectively, having sero-converted or sero-reverted. A limitation of this longitudinal study was that study participants were followed up only once after 6, 12 or 18 months. This study design was chosen to reduce dropout rates of study participants due to repeated blood drawings. Follow up of all study participants at three time points would have strengthened conclusions on the stability of antibody titers and in combination with environmental studies might have shed further light onto contact patterns of individuals with M. ulcerans.
As part of our study we also analyzed serological responses of children below the age of 11 years to treponemal and non-treponemal antigens in order to assess exposure to T. pallidum subspecies pertenue, the causative agent of yaws, which is transmitted by direct contact with the fluid from the lesion of an infected person. In contrast to syphilis, caused by the closely related T. pallidum subspecies pallidum, yaws mainly affects children living in poor rural areas of tropical countries [36]. Traditionally, the recommended algorithm for the serological diagnosis of treponemal diseases includes a non-treponemal test for screening, and a treponemal test for confirmation. Non-treponemal tests detect antibodies to non-treponemal antigens such as cardiolipin and lecithin released from damaged host cells or lipoprotein-like material released from the treponemes. Due to the occurrence of false-positive reactions, treponemal test results based on T. pallidum antigen are required for a reconfirmation of non-treponemal tests [37].
While eradication campaigns in the 1950s and 1960s by mass treatment of affected communities led to a drastic reduction of worldwide cases, yaws has lately re-emerged in Africa, Asia and the western Pacific [38] and Ghana was recently reported to be among the three most endemic countries for yaws [39]. Official case notification rates were 32 and 383 per 100,000 population in 2010 for the Ashanti and Central regions, respectively [40], but underreporting is suspected.
In this study antibodies to non-treponemal and treponemal antigens were detected in four of 402 (1%) children below the age of 11, indicating active yaws transmission in the affected communities. Antibodies to the treponemal antigen only were found in 2.2% of the children. Since yaws and syphilis are serologically indistinguishable, the interpretation of test results in adolescents and adults would require careful clinical assessment. While a recently published study in the Northern Region of Ghana has not found evidence of active yaws despite of continued case reporting [41], our data demonstrate evidence of ongoing yaws transmission in communities of the Offin river valley. There is an urgent need for more comprehensive data on the prevalence of yaws in Ghana to better implement mass drug administration programs.
Comparative genome analyses, environmental studies, as well as serological and epidemiological studies of BU affected populations have in the last decades gradually broadened our knowledge of environmental reservoirs and probable infection routes of M. ulcerans. Future longitudinal sero-epidemiological and environmental studies over longer time periods combined with the monitoring of environmental contact patterns may be required to unravel mysteries of M. ulcerans transmission.
|
10.1371/journal.pcbi.1003140 | Distinct Allelic Patterns of Nanog Expression Impart Embryonic Stem Cell Population Heterogeneity | Nanog is a principal pluripotency regulator exhibiting a disperse distribution within stem cell populations in vivo and in vitro. Increasing evidence points to a functional role of Nanog heterogeneity on stem cell fate decisions. Allelic control of Nanog gene expression was reported recently in mouse embryonic stem cells. To better understand how this mode of regulation influences the observed heterogeneity of NANOG in stem cell populations, we assembled a multiscale stochastic population balance equation framework. In addition to allelic control, gene expression noise and random partitioning at cell division were considered. As a result of allelic Nanog expression, the distribution of Nanog exhibited three distinct states but when combined with transcriptional noise the profile became bimodal. Regardless of their allelic expression pattern, initially uniform populations of stem cells gave rise to the same Nanog heterogeneity within ten cell cycles. Depletion of NANOG content in cells switching off both gene alleles was slower than the accumulation of intracellular NANOG after cells turned on at least one of their Nanog gene copies pointing to Nanog state-dependent dynamics. Allelic transcription of Nanog also raises issues regarding the use of stem cell lines with reporter genes knocked in a single allelic locus. Indeed, significant divergence was observed in the reporter and native protein profiles depending on the difference in their half-lives and insertion of the reporter gene in one or both alleles. In stem cell populations with restricted Nanog expression, allelic regulation facilitates the maintenance of fractions of self-renewing cells with sufficient Nanog content to prevent aberrant loss of pluripotency. Our findings underline the role of allelic control of Nanog expression as a prime determinant of stem cell population heterogeneity and warrant further investigation in the contexts of stem cell specification and cell reprogramming.
| Nanog is a key factor influencing the decision of a stem cell to remain pluripotent or differentiate. Each embryonic stem cell (ESC) in a population exhibits fluctuating Nanog levels resulting in heterogeneity which affects cell fate specification. The allelic regulation of Nanog was demonstrated recently but its implications on population heterogeneity are unclear. We developed a multiscale population balance equation (PBE) model and compared our results with pertinent experimental studies. Under allelic control the profile of Nanog features three peaks or distinct states. Transcriptional noise causes the distribution to become bimodal as suggested previously. When stem cells carrying a reporter transgene in an allelically regulated locus were examined, we observed non-matching distributions of the endogenous and reporter proteins. This led us to investigate the performance of reporter systems depending on insertion of the transgene in one or both alleles and the protein degradation dynamics. Lastly, our model was employed to address how allelic regulation affects the maintenance of pluripotency in stem cells with a single Nanog allele deletion. A fraction of these cells remains pluripotent while deletion of a single allele does not simply reduce NANOG uniformly for all ESCs but modulates NANOG heterogeneity directly.
| Nanog is a principal pluripotency regulator of embryonic stem cells (ESCs) in the early blastocyst. Mouse ESC (mESC) self-renewal is supported by Nanog in the absence of leukemia inhibitory factor (LIF) [1], [2] while Nanog knockdowns experience changes in global gene expression and loss of pluripotency [3]–[5]. In human ESCs (hESCs), Nanog reduction or overexpression leads to differentiation or inhibition of lineage commitment, respectively [6], [7]. Growth factors such as basic FGF and activin A known to maintain the pluripotency of hESCs target Nanog [8]–[10] further illustrating its prominent role in the decision of stem cells to self-renew or differentiate.
Single ESCs in vivo and in vitro exhibit fluctuating levels of several markers [11], [12] including Nanog [13], [14], which appears to regulate the heterogeneity of stem cell populations through feedback mechanisms with other transcription factors [15]. A bimodal distribution of Nanog has been reported in mESCs and hESCs carrying a reporter gene encoding the green fluorescence protein (GFP) in the Nanog gene locus [16], [17]. These observations have prompted the development of mathematical models to gain further insights into the mechanisms underlying Nanog heterogeneity. Nanog dynamics depicted in gene regulatory networks (GRNs) featuring feedback loops with transcriptional partners (mainly Oct4 and Sox2), are elicited via excitability [18] or oscillatory patterns [19]. According to these models, stem cells shuttle between a pluripotent Nanoghigh state and a differentiation-permissive Nanoglow or Nanog− state [14], [18]. Cells from the latter state reestablish the bimodal distribution under non-differentiating conditions pointing to the robustness of Nanog expression heterogeneity.
However, recent findings of allelic regulation of Nanog expression in mESCs shine new light on the observed heterogeneity of stem cell populations with respect to their Nanog profile. In an elegant study, Miyanari et al. [20] showed that cells in early pre-implantation mouse embryos express Nanog monoallelically but transition to a biallelic pattern in the late blastocyst. Similarly, mESCs cultured under typical maintenance conditions with LIF and serum express Nanog from a single allele whereas most mESCs treated with GSK3 and MAP inhibitors (2i condition [21]) activate both Nanog alleles. Accordingly, four distinct subpopulations of mESCs are observed depending on whether Nanog is expressed from each of the two alleles, both or none. This suggests that allelic regulation is a previously unaccounted source of stem cell population heterogeneity with the Nanog distribution comprising three cell groups (i.e. with monoallelic, biallelic and no expression of the gene).
This seemingly contradicts previous studies reporting only two groups (Nanoglow and Nanoghigh) of cells and calls for the development of a new framework incorporating the recent findings on the allelic control of Nanog expression. In fact, the two-state Nanog expression view of mESCs was recently reassessed and an intermediate state (middle Nanog) was added but without accounting explicitly for the allelic expression of the gene [22]. Moreover, previous models yielded fluctuations in Oct4 as coincident with those in Nanog. Yet, Miyanari et al. [20] noted that Oct4 does not experience the same allelic regulation with Nanog illustrating further the need for reexamining current rationales proposed for the variability of Nanog in stem cell ensembles. The allelic control of Nanog expression also necessitates the reinterpretation of work utilizing mESC and hESC lines carrying a reporter gene (typically that of the green fluorescent protein or gfp) knocked in one of the two Nanog alleles (e.g. [16], [18]).
We assembled a multiscale stochastic population balance equation (PBE) model to investigate how the recently discovered allelic control of Nanog expression affects ESC population heterogeneity. The distribution of pluripotency regulators in stem cell ensembles is determined by multiple processes transpiring at different physical and temporal scales. In addition to its allelic regulation, Nanog interacts with several other known and still unidentified factors and signals at the molecular level in a stochastic fashion. However, events at the cellular level (e.g. mitosis) also affect the content of Nanog and its partners in self-renewing stem cells [23]. Here allelic regulation was incorporated in our model to describe the distribution of Nanog in self-renewing mESCs [20]. Most notably, the newly observed regulatory mechanism is shown to be sufficient to give rise to a multimodal distribution of Nanog in a stem cell ensemble even in the absence of transcriptional noise. Pitfalls stemming from the use of Nanog reporter cell lines are demonstrated indicating the value of the PBE framework as a tool aiding in the (re)interpretation of the data from such experiments. Finally, quantitative analysis was performed of the Nanog signature of Nanog+/− mutant ESCs, the capacity of which to maintain their pluripotent state has been debatable. Our data explicates how these stem cells maintain a group of self-renewing cells through allelic control despite their higher propensity for differentiation and lower average content of NANOG protein.
The model was developed in three stages. First, a linear system was constructed describing the temporal evolution of the mESC population achieving dynamic equilibrium. Cell proliferation kinetics and transition rates between different subpopulations of allelic Nanog expression were determined based on data from Miyanari and Torres-Padilla [20]. Then, a single-cell model was assembled for Nanog gene expression. Finally, a system of PBEs was casted linking the single-cell to the population dynamics.
Mouse ESC types were defined (Figure 1) depending on whether nanog is expressed from both alleles (type 1) simultaneously or from a single allele (types 2 and 3) while there are also cells with both alleles being inactive (type 4).
The percentage of each subpopulation was calculated based on data from mESCs cultured with leukemia inhibitory factor (LIF) [20]. Briefly, approximately 30% of the mESCs were NANOG− (type 4) as determined from immunocytochemisty data for a total population of 135 cells (suppl. Figure 4b in reference [20]). Of the remaining 70% of (NANOG+) mESCs, roughly 80% expressed the gene monoallelically and 20% from both alleles (Figure 2e in reference [20]). Hence, 56% (types 2 and 3) and 14% (type 1) of the mESCs expressed nanog from a single and both alleles, respectively. Because no bias was reported for nanog expression from a specific allele, one can assume that each of types 2 and 3 comprises 28% of the total mESC population.
Single-cell allele-specific RT-PCR results were also provided in the same report (suppl. Figure 3b in Miyanari et al. [20]). Out of 19 mESCs examined, four cells were biallelic, ten cells were monoallelic and the remaining were classified as type 4 cells corresponding to the following fractions: 21.1% of type 1, 52.6% of types 2 and 3 and 26.3% of type 4. This population composition was close to that derived from the immunocytochemistry and RNA FISH data. However, the mESC fraction values calculated based on immunocytochemistry/RNA FISH were preferred due to the significantly larger sample size compared to that in the allele-specific RT-PCR experiment.
The stochastic switching of mESCs from one allelic pattern of nanog expression to another can be modeled as a time homogeneous Markov chain with four states (Figure 1). Cells switching satisfies the Markov property that the future state of each cell depends only on its current state. The fractions of cells per state at equilibrium are the elements of the limiting (equilibrium) distribution of the chain .
The transition matrix can be calculated from the percentages of the mESC population shuttling between states (see the Materials and Methods section):(1)satisfying the condition: . The transition rates between states i and j provide information regarding the kinetics of the process and these can be calculated from the transition probabilities (see Materials and Methods) taking into account that the fractions of mESCs in each state and between states have been determined over a single cell cycle Td or about 10 hours (suppl. Figure 6 in reference [20]). This yields the transition rate matrix:(2)with .
In addition, the proliferation rate of cells in the ith state can be calculated based on the doubling time Td of the mESC population. All mESCs in the population have the same proliferation kinetics regardless of the allelic regulation of Nanog expression:(3)
The mESC population can be described by a row vector with four elements representing the number of mESCs of each type (i.e. F1(t), F2(t), F3(t), F4(t)). Taking an exponential growth for the mESC population, the vector satisfies the equation(4)
The matrix is the sum of the transition rate matrix and a diagonal matrix with the growth rates of mESCs belonging to the four types, i.e.(5)
Each subpopulation can also be described by a percentage Zi(t) so that Fi(t) = Zi(t)Ft(t) (Ft(t): total cell number). Then, Equation 4 can be re-casted (see Materials and Methods):(6)with a stationary distribution when .
After calculating the proliferation rate and kinetics of transitioning between subgroups with different allelic expression of nanog, a model was constructed for single-cell level gene expression. A set of differential equations can be written encompassing the active (“On”) and inactive (“Off”) states for each allele (Table 1). The production and degradation dynamics of Nanog protein per allele are represented by zeroth and first order kinetics, respectively. The same kinetics were assumed for both Nanog alleles.
For the jth Nanog allele (j = 1, 2) we write:(7)(8)
Noise is an integral part of the transcription of genes [24] and can be taken into account by representing gene expression dynamics with stochastic differential equations (SDEs). To that end, we also employed the SDEs below to describe Nanog expression:(9)(10)
The noise terms were taken as temporally uncorrelated, statistically independent unit Gaussian white noise. The term refers to the intensity of the noise linearly related to NANOG [19]. Solutions to the SDEs were obtained via the Euler-Maruyama method [25].
The half-life (t1/2) of the NANOG protein in mESCs was experimentally determined to be approximately 2 hours [26], [27], or(11)
The production rate of NANOG protein per allele is estimated at molecules/hr. The Nanoghigh population was considered as comprising cells producing Nanog biallelically and monoallelically at a ratio of 1∶4 [20]. Then, this value of yields approximately 3,700 NANOG molecules per nanog-expressing mESC (Nanoghigh) on average at steady state [18]. Flow cytometry analysis of Nanog expression in mESCs reveals that the means of the Nanoglow and Nanoghigh states are different by two orders of magnitude [16], [18]. Consequently, the rate of NANOG protein generation from the “off” state allele is set at 1% of the “on” rate i.e. molecules/hr. Hence, even mESCs with both alleles at the ‘off’ state exhibit a baseline of NANOG expression. It should be noted however, that varying the values of and by 20% did not alter the modeling results qualitatively (Figure S1). For case studies involving Nanog reporter systems (e.g. with GFP expression), Equations 7 and 8 were utilized to describe the expression of the reporter protein with the same values for and as above. The degradation rate however, was calculated based on the t1/2 of the reporter protein. A t1/2 of 20 hours is reported for GFP [28].
With the elements of determined for mESC proliferation and transition between patterns of allelic Nanog expression and the single-cell gene expression model in place, we proceeded to construct a PBE-based system to describe, analyze and predict the effects of allelic regulation on the NANOG heterogeneity of stem cell populations. This framework takes into account processes such as gene expression and division occurring at the single-cell and population levels and spanning multiple time scales. The framework below comprises four PBEs, i.e. one PBE for each of the four distinct mESC groups based on the allelic expression of Nanog.(12)
The integers i and j refer to the mESC type () with i≠j and .
The state vector variables N1 and N2 correspond to NANOG levels originating from each one of the two alleles and v represents the cell size (volume) indicative of the cell's division potential [29]. The growth rate of cell size is proportional to cell size as detailed previously [23] (see also Materials and Methods). The rates for Nanog expression (i.e. and ) have been derived above (Equations 7–10). The dividing rate and partitioning function have been reported previously for stem cells [23] and details are provided in the Materials and Methods section. In addition, the allelic switching rates correspond to the transition rates (Equation 2), i.e.:(13)
Numerical solutions of the PBE system were obtained via a stochastic kinetic Monte Carlo algorithm [23], [30] as described in Materials and Methods. This entails the calculation of the time between successive cell divisions and allelic switching (interval of quiescence) which is considered a Markov process.
According to the findings of Miyanari et al. [20], mESCs achieve an equilibrium state as a composite of four subpopulation. This is reflected in the non-trivial solution of Equation 6. Thus, we first examine if stem cells from each subpopulation can reconstitute the blended population at equilibrium in PBE model simulations. For this purpose, the starting population was set to 100% of single type mESCs and its temporal trajectory was tracked (Figure 2). Indeed, mESCs attained the same equilibrium state within 50–100 hr from initial ensembles of cells with a uniform nanog expression pattern. Thus, the allelic control of nanog allows mESCs to restore the population with constant fractions of different types at equilibrium.
We then set out to investigate the extent to which allelic gene regulation contributes to NANOG expression macro-heterogeneity utilizing the PBE representation of the mESC population. Simulation results yielded three distinct peaks in the distribution of NANOG (Figure 3A). The NANOGhigh region comprised mESCs with biallelic Nanog expression (peak ‘H’ in Figure 3A) and the NANOGlow area (peak ‘L’) contained cells with both alleles being inactive. These type ‘4’ cells may exhibit low levels of NANOG from leaky expression and from protein produced before entering this state (see below). In a flow cytometry assay, these cells may fall within the region of autofluorescence [18] or isotype control [23].
Notably, there is also a prominent third peak (‘M’) corresponding to mESCs with monoallelic nanog expression. The results seem to contrast the bimodal distributions of NANOG which have been reported for mESCs and hESCs carrying the gfp gene in their Nanog gene locus [16], [17]. We considered that noise associated with gene transcription and translation [31]–[33] may have an additional impact on NANOG variability. To test this, we employed SDEs (Equations 9–10) instead of deterministic ODEs (Equations 7–8) to depict the dynamics of Nanog. As shown in Figure 3B, a 20% noise relative to the NANOG protein level was sufficient for peaks ‘M’ and ‘H’ to merge yielding a bimodal profile. Potentially, other sources of noise during measurement (e.g. instrument noise) may also contribute to the dispersion in NANOG distribution effectively reducing the threshold of intrinsic noise in gene expression leading to the same two- (instead of three-) peak profile. In addition to stem cell lines expressing a reporter gene from the Nanog locus, stem cells stained with appropriate Nanog antibodies exhibit the same two-peak NANOG profile shown in Figure 3B after flow cytometry. Since our focus in this study was on how allelic regulation affects Nanog presentation in stem cell populations and fundamental experiments for quantification of intensity from other noise sources are still lacking, we considered the dynamics of Nanog as deterministic (Equations 7–8) in subsequent simulations.
The NANOG content of individual cells displaying different modes of allelic control was examined. For any snapshot of the mESC ensemble, four distinct subpopulations could be identified clustered approximately on the vertices of a square. Each vertex corresponded to mESCs with a specific pattern of allelic Nanog expression (Figure 3C). There were also cells transitioning between subgroups as indicated by their presence on the edges and inside the square area. Upon closer observation, a number of cells classified as type ‘4’ (green dots) were close to the vertex of cells expressing Nanog biallelically. Direct transition between types ‘1’ and ‘4’ was ruled out as it was not observed in experiments [20].
Another possible explanation is that cells having both of their Nanog alleles inactive which mapped close to the subgroup ‘1’ region most likely entered state ‘4’ only recently. Hence, their high Nanog content is mainly due to past elevated expression of nanog. This prompted us to investigate NANOG fluctuations in randomly selected single mESCs (Figure 3D) and the time required to build or deplete NANOG after allelic pattern switching. Division and allelic switching events transpired stochastically at timescales longer than the periods for production and degradation of the protein. When the NANOG levels of all daughter cells derived from a single mESC are plotted over 60 hours (Figure 3E), a state for the population can be clearly seen as in Figure 3A. A cell with high Nanog content reverts to a state of minimal intracellular Nanog 15–20 hours after turning off expression from both alleles. This explains the presence of type ‘4’ cells in the vicinity of the type ‘1’ cell vertex and along the ‘1’–‘2’ and ‘1’–‘3’ type edges (Figure 3B). In contrast, type ‘4’ cells switching to monoallelic Nanog expression build their Nanog content reaching higher levels faster (less than 5 hours). This explains the presence of only few type ‘2’ and ‘3’ cells close to the type ‘1’ vertex.
Altogether, the allelic regulation of Nanog expression leads to macro-heterogeneity of the population with the assortment of cells into three groups which become less distinct under the influence of noise. The population and its Nanog profile can be reconstituted from individual self-renewing mESCs despite differences in their Nanog expression patterns. Interestingly enough, the time for depletion of intracellular NANOG reserves by stem cells switching to the state of ‘no-active allele’ expression is significantly longer than for reaching a high NANOG content after cells enter states of monoallelic gene expression from state ‘4’.
The difference in the lag for adjustment of NANOG content after switching to a particular state of allelic Nanog expression is a function of parameters such as the cell Td, the average frequency of switching between patterns and the t1/2 of the protein. Whereas the first two parameters depend on the cell type, the latter is also largely specific to the protein of interest. This consideration is pertinent to mESC and hESC reporter lines, which are increasingly utilized in stem cell research. In these cells, a reporter gene such as gfp is knocked in one of the Nanog allelic loci [16], [17]. With this design, GFP and NANOG are assumed to have the same production rate [34] lending credence to the notion that GFP should track NANOG closely.
Nevertheless, two potential issues arise with such design. First, differences in the reporter and endogenous protein degradation kinetics may also effect divergence in the profiles of the two gene products. The GFP has a longer t1/2 (∼20 hours [28]) than that of NANOG (∼2 hours). Second, the existence of allelic regulation suggests that a reporter gene expressed from one allele via the promoter of a native gene may not be representative of the overall level of the target gene. These concerns prompted a more detailed investigation of the potential disparity in the expression of the reporter gene and endogenous NANOG subjected to allelic control.
With allelic regulation and stochastic partitioning during mitosis, our simulations clearly showed that unlike the NANOG profile the GFP distribution features two distinct modes even in the absence of transcriptional noise (Figure 4A). More importantly, the heterogeneity associated with GFP is more pronounced (coefficient of variation or CV = 1.0) than that of the actual NANOG distribution (CV = 0.74). Thus, caution should be exercised when examining data from experiments with stem cell lines featuring knocked-in reporter genes.
We further tested if the read-out of this reporter system was reliably reflecting the dynamic expression of the native gene under different assumptions. As expected, when the same t1/2 was assumed for both NANOG and the reporter protein (Figure 4B), their levels were perfectly correlated (Pearson product-moment correlation coefficient ρ = 1.00) if the reporter gene was inserted in both alleles. In the case of a single-allele knock-in (Figure 4C) however, a subpopulation of the stem cells (NANOGhigh/GFPlow) could not be correctly reported due to the effect of allelic regulation (ρ = 0.66).
The situation of NANOG and a reporter protein having the same t1/2 is unlikely when GFP (or several of its variants) is considered. This results in non-matching profiles for the two proteins (ρ = 0.89) even when both alleles carry the fluorescent marker gene (Figure 4D). Reporter systems used in practice carry the reporter gene in one of the two targeted alleles with different half-lives for the native and reporter gene products. Not surprisingly, cells with insertion of the reporter gene in one allele showed the lowest correlation (ρ = 0.60) between the expression of reporter and NANOG with different t1/2 (Figure 4E). Specifically, GFPhigh mESCs were also NANOGhigh but a portion of NANOGhigh mESCs fell within the GFPlow region. These cells may be misconstrued as autofluorescent or similar to isotype controls. Additionally, mESCs exhibit heterogeneity in reporter and NANOG levels but the reporter read-out does not vary linearly with the NANOG expression level as in Figure 4B.
Since allelic regulation is not universally applicable to stem cell genes (e.g. pou5f1 (Oct4)), we also analyzed a more general case for an endogenous gene X not subjected to this mode of expression control. When this gene is expressed at steady state, its level correlates qualitatively with the level of the corresponding reporter in a straightfoward manner. We therefore considered transient expression of gene X as in the case of pluripotency markers at the onset of differentiation or upon treatment with transcriptional inhibitors [35]. For this purpose, transcription of the native and reporter genes was turned off in the PBE model (see Materials and Methods) and the temporal evolution of the respective protein distributions was tracked (Figure 5A). Without allelic regulation (Figure 5B), the single-allele reporter system displayed a tighter correlation between the expression of the reporter and endogenous proteins (ρ = 0.83 vs. ρ = 0.60 in Figure 4E) and was maintained over 20 hours (ρ∼0.82; time equal to t1/2 of reporter). Hence, the reporter signal qualitatively still reflected the endogenous protein level. However, the relative decrease in reporter protein over time lagged the native protein reduction significantly (Figure 5C) although both eventually converged to distributions with lower mean values (Figure 5A).
Taken together, our data demonstrate that in stem cell lines expressing reporter genes from the Nanog gene locus, the reporter protein level is not reflective of the endogenous NANOG protein. Actually, stem cells carrying the reporter gene in one target allele are commonly utilized in research today. These cells exhibited the highest divergence in the profiles of the native protein and its surrogate reporter. Therefore, the effects of allelic regulation should be accounted for when interpreting relevant data. Lastly, differences in the reporter and target protein half-lives contribute to disparate profiles of transiently expressed genes regardless of whether the reporter gene is knocked in one or both alleles even when there is no allelic regulation.
We showed that normal cells having inactive both Nanog alleles (e.g. type ‘4’ cells in Figure 1C) eventually reconstitute a heterogeneous population featuring cells with high Nanog preventing commitment. Thus, we asked the question: How does the deletion of one copy of nanog affect the capacity of mESCs to maintain a pluripotent state given the allelic regulation of the gene? This segment of our work was motivated by conflicting findings in experiments utilizing Nanog mutant cells. Hatano et al. [5] observed that Nanog+/− mESCs readily differentiate in spite of being cultured with LIF. Others also reported that suppression of Nanog leads to reduced expression of other pluripotency markers [7] and induces differentiation [6] in hESCs. Still, Chambers et al. [16] in an elegant study reported that Nanog+/− and Nanog−/− mESCs continue to self-renew in the absence of differentiation stimuli and form colonies with similar morphology as pluripotent mESCs concluding that Nanog acts to safeguard pluripotency but is not an indispensable factor.
To that end, the PBE model was modified by turning off the expression of Nanog from one allele to account for the Nanog+/− genotype (see Materials and Methods). When comparing the distribution of NANOG in wild-type and mutant mESC populations, the latter cells still exhibited NANOG+ mESCs. However the fraction of NANOG+ mESCs dropped from approximately 73% for normal mESCs to almost 46% for Nanog+/− cells (Figures 6A–B). This was concomitant with an increase in the heterogeneity of the population (CV = 0.74 and 1.08 for Figures 6A and 6B, respectively). It should be noted that in flow cytometry assays the line separating the NANOG− and NANOG+ cells (500 molecules/cell here) between the first and second/third peaks is determined based on appropriate isotype controls. Shifting the line within this region did not alter the fractions of cells significantly. The average NANOG amount per cell was almost half in the Nanog+/− mESC population than in normal mESCs (Figure 6C) in line with western blot results by Hatano et al. [5]. Our findings show that deletion of one Nanog allele does not simply reduce NANOG uniformly for all mESCs but modulates NANOG heterogeneity directly.
Examination of the NANOG fluctuations in single cells further illustrated this effect (Figure 6D). Compared to wild-type mESCs, Nanog+/− cells had a lower chance of switching back to a NANOG+ state due to allele deletion. In fact, almost 60% of wild-type mESCs with both alleles in the ‘off’ state switched on at least one allele within five cell cycles and the steady-state mESC population was reconstituted within 100 hours (see Figure 1C). In contrast, the corresponding fraction of Nanog+/− mESCs was only 43%. Nonetheless, the higher fraction of NANOG− cells indicates that loss of one Nanog allele results in a commitment-permissive state. Thus, Nanog+/− cells remain pluripotent in the absence of differentiation signals but over half of the population will promptly differentiate upon induction with appropriate factors.
Nanog is a core pluripotency transcription factor influencing the decision of stem cells to self-renew or differentiate. The recent demonstration that Nanog is allelically regulated in mESCs calls for reexamination of findings about the role of Nanog on the maintenance of the pluripotent state and the propensity of stem cells for commitment to particular lineages. It also provides a new vista for the interpretation of data from engineered stem cell lines with reporter genes knocked in the Nanog gene locus. Allelic regulation of Nanog expression has not been demonstrated experimentally in human stem cells but we surmise that work in this direction is in progress, especially given that this mechanism is plausible when analyzing pertinent hESC data. With these considerations in mind, we developed a PBE model taking into account the allelic regulation of Nanog in conjunction with the asynchronous cell proliferation and gene expression dynamics. Besides recapitulating the experimental findings of Miyanari et al. [20], our results clearly demonstrate that any of the four mESC types under routine maintenance conditions (LIF and serum) gives rise to mESC populations with the same heterogeneity with respect to Nanog expression. This is particularly significant because Nanog coordinates multiple genetic programs during development and reprogramming and potentially regulates heterogeneity [15], which translates to variable proclivity for self-renewal or commitment among cells of the same population. Indeed, a subpopulation of self-renewing cells residing at a state with lower Nanog content is primed for specification upon induction with suitable factors. In its current form, the framework does not consider differentiation but work in this direction is underway [36].
The Nanog distribution in mESC populations at equilibrium features three peaks corresponding to types ‘1’ (biallelic), ‘2’+‘3’ (monoallelic) and ‘4’ (both alleles being inactive) seemingly contrasting previous reports of a bimodal NANOG (GFP) distribution in mESC and hESC lines with the gfp expressed from the Nanog locus [17], [18]. One may argue however that in a flow cytometry assay the lowest Nanog content (type ‘4’) peak ‘L’ would overlap most likely with the isotype (or autofluorescence) control and therefore the cells would be considered as NANOG− akin to the LN mESCs [18] and to hESCs [23]. Additionally, sorted LN cells reconstitute the original bimodally distributed population of LN/HN cells upon subculturing similar to our results with a starting population of type ‘4’ mESCs. We also showed that gene expression noise causes the peaks ‘M’ and ‘H’ (NANOG+ cells) to merge yielding a bimodal profile. The existence of three states based on Nanog expression for mESC populations has been recognized in a recent study with the introduction of a middle Nanog (MN) state [22]. Thus, discrepancies between the present and other studies regarding the NANOG profile of self-renewing stem cell populations appear to be largely reconciled.
Nevertheless, the underlying determinants of the NANOG distribution are significantly different. The existence of the LN population was explained earlier through the concept of excitability in a GRN of Nanog with Oct4 and Sox2 [18]. A transient low expression of Nanog (LN) ensues when the GRN featuring a negative feedback loop is perturbed by transcriptional noise. Others have also employed the same three-transcription factor GRN with modifications to study NANOG variability [19] concluding that oscillations or noise in Nanog expression leads to a similar two-peak profile. Gene expression noise is a major determinant of the distribution of NANOG in stem cells [23]. Our model provides alternative mechanisms driving the emergence of the NANOG heterogeneity observed in mESC cultures. The bimodal distribution of Nanog emerged in our analysis by considering allelic regulation, asynchronous cell proliferation, and stochastic partitioning of NANOG with or without transcriptional noise in a single-gene model. Allelic control of Nanog expression has been elegantly demonstrated in mouse embryos and mESCs [20] supporting our findings. Yet, practical methods for controlling noise in cellular processes are still lacking. This leaves open the question of whether (and if so how) allelic modulation of gene expression acts in concert with one or more excitable GRNs under transcriptional noise to promote diversity in isogenic stem cell populations.
Our framework also provides a rationale regarding the stability of the LN state. Sorted HN mESCs (GFP+) give rise to a population with a lower fraction of LN mESCs (GFP−; 7%) than the HN group (38%) of sorted LN mESCs cultured for the same period (48 hours) [18]. Supported by an excitable GRN model, this observation led to the conclusion that the LN state is unstable with frequent transitions to the HN, whereas the latter state is stable and conversions to the LN state are rare. Potential discrepancies between actual Nanog expression and GFP signal aside, we also observed that a number of type ‘4’ cells are classified as NANOGhigh cells especially if they have just exited the state in which both nanog alleles were active. Cells with both alleles recently inactivated, require longer time to deplete their NANOG reserve whereas those exiting this state build their protein content faster. Therefore, the experimentally observed dynamics of the HN and LN states are supported by our model mainly as a result of allelic regulation of Nanog.
Unlike other reports employing GRNs, the Nanog expression dynamics in this study were described by a single-gene model with “on” and “off” states. This approach was advantageous in two ways: First, adoption of a GRN model necessitates assumptions about the structure of the network. Structures of GRNs involved in stem cell fate decisions are not well-established. For example, Navarro et al. [37] recently reported that Nanog activity is autorepressive and independent of Oct4/Sox2 unlike GRNs utilized in previous studies. Second, GRN models typically involve several parameters which are currently impossible to determine through experiments. Although we utilized a single-gene expression model, the PBE framework is amenable to the incorporation of GRNs, especially as more information comes to light from research on the interactions of Nanog with other partners.
It should be noted that culture conditions affect the relative portions of stem cells in different Nanog states. Mouse ESCs maintained in medium containing serum and LIF achieve equilibrium with fractions reflected by [20]. We considered this as our model system since mESCs are commonly cultivated with LIF and serum. However, the same analysis can be carried out for other conditions. For example, growing mESCs in 2i leads to a significant enhancement in biallelic nanog expression thereby changing the relative portions of different subpopulations at equilibrium () [20]. An analysis of the Nanog distribution in mESCs under different culture conditions has been reported [22] without considering explicitly allelic gene regulation. The corresponding model is based on the calculation of a one-dimensional ‘potential energy’ function representing the ‘barrier’ for cells moving between intermediate states. Others [38] have also modeled the transition of stem cells between attractor states through a quasi-potential energy function in an epigenetic landscape introduced by Waddington [39].
The allelic control of Nanog expression calls for closer scrutiny of stem cell lines carrying reporter genes such as GFP and its variants. Use of such lines is warranted on the premise that the reporter signal can serve as a surrogate closely matching the expression of a protein from the same genetic locus. Our simulation results illustrate that the reporter signature varies drastically depending on whether its gene is inserted in one or both target alleles, even under the assumption of equal t1/2 for the reporter and native gene products. Thus, stem cell lines intended for monitoring genes subjected to allelic regulation should have the reporter gene inserted in both alleles. Obviously, this entails practical considerations as such construction is significantly more cumbersome than that of single-allele knock-ins. Reporter genes are inserted in the targeted locus usually by homologous recombination which is a notoriously inefficient process although certain modifications may enhance its efficiency [40]–[42]. Because allelic control of expression is not universal, single-allele residing reporter gene cell lines may be sufficient for monitoring genes not subjected to this mechanism.
Still, an important factor in monitoring gene expression via a reporter surrogate is the difference in the kinetics (typically exemplified by the t1/2) for net production of the native and reporter proteins translating to non-matching profiles. This disparity may be partially alleviated with the use of proper fast-degrading (destabilized) reporter variants [43], [44] but should not be overlooked as it is fundamental for proper interpretation of pertinent data. In fact, the PBE model described here can be used to back-calculate the actual expression profile of the protein of interest from reporter distributions. The process entails the estimation of parameter values for reporter production and switch on/off rates. The same values will apply to the native protein distribution due to the matching regulation by virtue of sharing the same chromatin site. If the gene is allelically regulated, transition rate parameters can be obtained, for example, from immunocytochemistry/RNA FISH or single-cell allele-specific RT-PCR data. Other PBE parameters can be determined as we described previously [23]. The cell doubling time (Td) can be measured in cell culture experiments and the t1/2 values of the reporter and the native protein also can be obtained through well-established methods [45]. With this information available, the PBE model can be run to generate the actual profile of the target protein. This approach is straightforward when the distributions of the reporter/protein are time-invariant. The same methodology can be applied to temporally fluctuating distributions but requires detailed knowledge of the mechanism(s) governing the evolution of reporter and protein production. Additional information may also be necessary, for example, in differentiation experiments where the expression of pluripotency and lineage-specific markers changes with time. A major challenge in these experiments is the identification of appropriate single-cell functions describing the dynamics of stem cell commitment. The timing of the measurements also becomes relevant since our results show that the decay in reporter protein with a longer t1/2 lags that of the target protein when both genes are not actively transcribed (Figure 6C). Analogous results can be obtained for a reporter and its target protein when the transcription of both is turned on under proper conditions.
The model also shines light on whether Nanog− stem cells in a self-renewing population may regain or lose irreversibly their pluripotent status. We demonstrated that starting with a group of wild-type mESCs having both nanog alleles ‘off’, over 60% of them switch on at least one within five cell cycles in non-differentiating conditions. In the same interval, Nanog+/− mESCs with their nanog allele initially inactive transition to a population where 43% of the cells are NANOG+. Thus, even NANOG− mESCs can self-renew and reestablish a NANOG+ population in agreement with previous studies [2], [5], [16], [18]. The framework in its present form helps to predict if a cell within an ensemble will continue to self-renew or commit to a particular fate if exposed to differentiation stimuli. Such prediction entails the knowledge of a Nanog content threshold for differentiation-preventive vs. -permissive stem cell self-renewal.
Nonetheless, further research is needed to address a distinct question, i.e. to which lineage a differentiation-primed NANOG− stem cell will convert. The lineage propensity of cells with low or no Nanog expression is debatable. According to Mitsui et al. [2], Nanog−/− mESCs primarily express markers of parietal and visceral endoderm, whereas others [5] showed that Nanog+/− cells express genes of the three embryonic germ layers. Their results suggest a Nanog content-dependent differentiation with extraembryonic endodermal fates favored in the absence of Nanog and mesodermal, endodermal and ectodermal progeny being generated from cells with Nanog content gradually decreasing by 0–50% compared to pluripotent state ESCs. Mouse ESCs at the LN state cultured in neuronal differentiation medium may still revert to the HN state albeit at a low fraction (16%) [18]. Since Nanog interacts with multiple partners in pluripotency and differentiation programs [4], long-term residence of stem cells in the NANOG− state may eventually lead to differentiation, even with small perturbations in their microenvironment. For instance, no changes are evident in transcriptional regulatory network partners of Nanog until at least three days after its depletion [15]. Longer-term expansion of Nanog−/− mESCs without loss of their pluripotency has also been reported with variable degrees of success [5], [16]. Thus, the kinetics of NANOG− stem cells undergoing differentiation vs. self-renewal and the balance with the NANOG+ cells remain to be elucidated.
The time span between the complete decline in Nanog content and loss of pluripotency is also an illustration of the multiscale nature of stem cell fate specification [15], [16]. We view that models for stem cell populations should consider together subcellular (e.g. regulation of pluripotent/differentiation marker expression, signal transduction), intercellular (e.g. paracrine signaling) and population-wide processes (e.g. cross-talk among subpopulations with distinct phenotypes). These phenomena are not only innate to the stem cell niche and major determinants of fate decisions but also transpire over markedly different time scales. Multiscale PBE approaches afford coping with the multiple temporal/spatial scales of stem cell processes. In the present study, rapidly fluctuating gene expression dynamics were combined with significantly slower events such as cell proliferation and allelic regulation. At the same time, there is flexibility in the implementation of models for deterministic or stochastic phenomena such as the transcription and allelic switching of nanog.
In conclusion, the stochastic PBE model developed in this study is aligned with the experimental findings on the allelic switching of Nanog expression and the heterogeneity of cells with single nanog allele deletion. Our results illustrate that allelic regulation is pivotal for the observed heterogeneity of ESCs with respect to Nanog content. The same mechanism may very likely influence the diverse presentation in stem cell populations of other markers (e.g. Oct4, Stella, Sox2, Rex1), which are intricately connected to the expression of Nanog. Lastly, the significant problems linked to the use of reporter cell lines for monitoring Nanog (or other genes) are portrayed. The PBE framework provides a platform for addressing these issues in practice and may serve as a tool complementing experiments to gain a deeper understanding of stem cell population heterogeneity in connection with fate specification. These outcomes will accelerate the development of efficient differentiation and reprogramming methods for the generation of therapeutically useful progeny.
The transition probabilities for a cell switching from state i to state j can be calculated considering (a) the limiting distribution and (b) information regarding the numbers of cells shuttling between these states. Such information is available per cell cycle (unit of time) in the report by Miyanari et al. [20] and as indicated in Figure 1. For instance, 12% of the total mESC population shuttles between states 1 and 2. The percentages of cells switching from i to j and from j to i states are assumed to be equal. Then, detailed balances can be written, i.e.(A1)where represents the fraction of the cell population transitioning from state i to j. For instance, yieldinggiven that 14% and 28% (elements and ) of the total population are in states 1 and 2, respectively. The other transition probabilities are calculated in the same fashion noting that since states 1 and 4 are not linked directly. Moreover, .
The transition rates for cells switching from state i to j are defined as [46]. Here, the data for calculation of the transition probabilities refer to a single cell cycle time Td (unit time of observation) and thus, the transition rates are approximated as . Moreover, holds true based on transition matrix properties.
The system of differential equations (Equation 4) describing the temporal evolution of the subgroups of mESCs exhibiting distinct allelic expression of Nanog in terms of cell numbers () can be re-written based on the corresponding percentages :(A2)
This results in the following expression:(A3)where is the (4x4) identity matrix and because(A4)then,(A5)
This yields Equation 6 in the main text. This can also be written as:(A6)At steady state there is a non-trivial solution since the rank of is 3 and the corresponding vector of the null space is: . The non-trivial solution is subjected to the constraint: . The results are consistent with the findings from mESC experiments yielding .
The master PBEs (Equation 12) contain the following functions:
A schematic of the Monte Carlo (MC) algorithm for obtaining numerical solutions of the PBE model has been described previously [23] and is shown in Figure S2. In addition, we detail below the selection of a specific event interrupting quiescence (i.e. cell division or switching between allelic Nanog expression patterns).
For this purpose, matrix was set up with Ft rows (i.e. equal to the total number of cells) and 4 columns for the four mESC states. The nth row corresponds to a cell from the ith subgroup and contains the pertinent transition rates () and proliferation rate (). Given a random number ran2 from a uniform distribution, we identify: (a) The kth cell which will disrupt quiescence, and (b) whether this cell will divide or switch to a different state:or (if falls between two successive cells in the matrix E)
The above inequalities allow for the identification of the kth mESC, which is at the ith state. If then this cell divides, otherwise it switches from ith to the lth state. Initially the algorithm is a constant volume MC as the cell population size increases up to a limit (typically 10,000 cells). Subsequently, the algorithm becomes a constant number MC [23] with the daughter cells replacing the mother cell and another randomly selected cell (see Figure S3). Simulation programs were written in FORTRAN. MATLAB (Mathworks, Natick, MA) was utilized mainly for post-processing of results.
The PBE model was modified to simulate the temporal evolution of an endogenous gene X and a reporter gene when both are not subjected to allelic regulation. The off-diagonal elements of matrix were set to zero because there is no allelic switch and all cells in the population belong to (sub)group ‘1’. Initially, all cells in the ensemble express the reporter and X genes. Subsequently, expression of X and the reporter was turned off by using the “off” state values for the parameters in the single-gene model.
We assumed without loss of generality that Nanog allele 1 was deleted in Nanog+/− mESCs by setting and maintaining the gene expression from allele 1 in the “off” state for the duration of the simulation as shown in Table S1. The expression dynamics and pertinent parameters for the functional allele (allele 2) remained the same as described in the model development paragraph.
|
10.1371/journal.pcbi.1002432 | Feedforward Inhibition and Synaptic Scaling – Two Sides of the Same Coin? | Feedforward inhibition and synaptic scaling are important adaptive processes that control the total input a neuron can receive from its afferents. While often studied in isolation, the two have been reported to co-occur in various brain regions. The functional implications of their interactions remain unclear, however. Based on a probabilistic modeling approach, we show here that fast feedforward inhibition and synaptic scaling interact synergistically during unsupervised learning. In technical terms, we model the input to a neural circuit using a normalized mixture model with Poisson noise. We demonstrate analytically and numerically that, in the presence of lateral inhibition introducing competition between different neurons, Hebbian plasticity and synaptic scaling approximate the optimal maximum likelihood solutions for this model. Our results suggest that, beyond its conventional use as a mechanism to remove undesired pattern variations, input normalization can make typical neural interaction and learning rules optimal on the stimulus subspace defined through feedforward inhibition. Furthermore, learning within this subspace is more efficient in practice, as it helps avoid locally optimal solutions. Our results suggest a close connection between feedforward inhibition and synaptic scaling which may have important functional implications for general cortical processing.
| The inputs a neuron receives from its presynaptic partners strongly fluctuate as a result of either varying sensory information or ongoing intrinsic activity. To represent this wide range of signals effectively, neurons use various mechanisms that regulate the total input they receive. On the one hand, feedforward inhibition adjusts the relative contribution of individual inputs inversely proportional to the total number of active afferents, implementing a form of input normalization. On the other hand, synaptic scaling uniformly rescales the efficacy of incoming synapses to stabilize the neuron's firing rate after learning-induced changes in drive. Given that these mechanisms often act on the same neurons, we ask here if there are any benefits in combining the two. We show that the interaction between the two has important computational consequences, beyond their traditional role in maintaining network homeostasis. When combined with lateral inhibition, synaptic scaling and fast feedforward inhibition allow the circuit to learn efficiently from noisy, ambiguous inputs. For inputs not normalized by feed-forward inhibition, learning is less efficient. Given that feed-forward inhibition and synaptic scaling have been reported in various systems, our results suggest that they could generally facilitate learning in neural circuits. More broadly, our work emphasizes the importance of studying the interaction between different plasticity mechanisms for understanding circuit function.
| As part of an ever-changing world, brain activity changes continuously. The fraction of neurons active in a region at each given moment fluctuates significantly driven by changes in the environment and intrinsic dynamics. Ideally, regions receiving this activity as input should be able to represent incoming signals reliably across the full possible range of stimulation conditions. Indeed, this type of regulation seems to be ubiquitous in the cortex. In the early visual system, contrast gain control begins in the retina [1] and is strengthened at subsequent stages of the visual system, such that the way an image is represented in V1 simple cells is largely contrast invariant [2], [3]. Similarly, in the olfactory system, neuronal representations remain sparse and odor-specific over thousand-fold changes in odor concentration [4]–[6].
To be able to achieve such invariance, neurons have evolved various mechanisms that adjust neuronal response properties as function of their total input. One instance of such normalization involves feedforward inhibition, in which afferent inputs induce both excitation and mono-synaptically delayed inhibition onto principal cells [7]–[12], shaping the temporal activity pattern of the postsynaptic neurons [8]–[10], and sparsifying population activity [5]. The degree of specificity of this inhibition can vary from stimulus specific to relatively unspecific [7], [12]. Here, we focus on fast but unselective feedforward inhibition, which has been reported in a range of circuits including hippocampus and sensory areas [11], [13]–[15]. This mechanism adjusts, virtually instantaneously, the sensitivity of pyramidal cells to the overall strength of the afferent stimulus. As a result, the influence of an individual afferent on the firing of the postsynaptic neuron is continuously normalized by the total number of active afferents. Functionally, it has been hypothesized that such input normalization is needed to expand the range of inputs that can be represented in a neuron population [11], however, its implications for learning in the circuit remain unclear.
Another mechanism with similar effects, but acting on a slower time scale, is synaptic scaling [16]–[18]. Specifically, it is believed that neurons detect sustained changes in their firing rates through calcium-dependent sensors and increase or decrease the density of glutamate receptors at synaptic sites to compensate for these changes in drive [19]. This results in an uniform rescaling of the strength of excitatory synapses as a function of average postsynaptic activity. Synaptic scaling often takes a multiplicatively form [17], which has the benefit of preserving the relative contribution of synapses and hence the information stored through Hebbian learning [20]. This type of weight normalization is believed to address a different kind of stability problem–the fact that synapses are plastic. As Hebbian learning alone would destabilize neural dynamics, due to a positive feedback loop, additional homeostatic mechanisms such as synaptic scaling are needed to ensure stable circuit function [18]–[20].
Fast feedforward inhibition and synaptic scaling have been reported for a range of circuits including hippocampal and neocortical pyramidal neurons [11], [19]. Given that both mechanisms effectively regulate the total incoming drive to neurons, it may be somewhat surprising that they co-occur in the same cell types. This suggests there may be some computational advantage in combining input normalization and synaptic scaling. However, based on the existing experimental evidence alone, it is unclear what possible benefits this interaction may have.
We show here that the role of input normalization and synaptic scaling goes beyond simply maintaining circuit homeostasis, and that they play important computational roles during synaptic learning. In the presence of neuronal competition by global lateral inhibition, the two enable efficient unsupervised learning from noisy or ambiguous inputs. Specifically, we consider an elementary circuit that incorporates synaptic scaling and fast feedforward inhibition. We analyze the learning dynamics in this circuit and show that, for certain input statistics, standard neural dynamics and Hebbian synaptic plasticity implement approximately optimal learning for this data–an observation that we further confirm in numerical experiments. The studied circuit learns an efficient representation of its inputs which can be used for further processing by downstream networks (e.g., for classification). Importantly, in the absence of feedforward inhibition, learning in the same circuit results in much poorer representations, as the system has a stronger tendency to converge to locally optimal solutions–a problem that neural and non-neural systems for unsupervised learning commonly face. This provides evidence for synaptic plasticity requiring normalized inputs for efficient learning. Given that feedforward inhibition and synaptic scaling seem to co-occur in various neural circuits, our results suggest that the interplay between the two mechanisms may generally facilitate learning in the cortex.
We construct a model of feedforward inhibition and synaptic scaling acting in a neural circuit in which excitatory synapses change by Hebbian learning. The analysis of their interaction proceeds in two steps. First, we study the dynamics of learning within the circuit, leaving details of the neural dynamics unspecified. This analysis reveals that the weights converge to final values that are fully determined by the input distribution and the neuronal transfer function. Second, when using a specific statistical model for the input distribution, we can identify biologically plausible neural dynamics that implement optimal learning for these stimuli. We show that a specific form of lateral inhibition implementing softmax competition between different neurons is sufficient for optimal learning in our setup, something which we then confirm by numerical simulations, using both artificially generated and natural data. Lastly, we show that learning performance is critically dependent on feedforward inhibition and, how the emerging representations can be used by higher processing layers, for instance, for efficient classification.
As a starting point, consider the elementary neural circuit shown in Fig. 1A. The network consists of neurons receiving excitatory inputs from input neurons through a set of excitatory weights , . We denote by the activity of input neuron and by the activity of the downstream processing neuron .
In the general case, the activity of neurons can be defined as a function of the activity of the input layer, , and of the weights :(1)This transfer function is not necessarily local, as it does not restrict the dependency to the afferent weights of neuron ; it allows us to also describe the interactions between neurons through lateral connections (marked by dotted lines in Fig. 1A). For the first part of the analysis, we assume the neural dynamics given by (1) to be arbitrary, though later we consider specific forms for the transfer function.
We model feedforward inhibition by explicitly normalizing the input vector to satisfy the constraint:(2)Such input normalization can remove undesired patterned variations (e.g. contrast, see Fig. 1D), potentially facilitating learning in the circuit. If we denote the un-normalized input by , the constraint can, for instance, be fulfilled by a simple division, , though alternative implementations are possible. This formulation abstracts away the details of the biological implementation, focusing on its functional implications [11]. Importantly, the simple form allows us to derive theoretical results about the role of this form of feedforward inhibition during learning. At the level of the neural circuit, however, input normalization relies on the presence of a set of fast spiking interneurons (in the hippocampus – predominantly basket cells [8]) innervated by the same afferent inputs, with unspecific projections onto the subsequent layer. The implications of this neural implementation are considered in more detail in the Discussion.
We model incoming synapses to be plastic and to change by Hebbian learning, with synaptic scaling implemented by an additional weight dependent term [20], [21]:(3)where is a small positive learning rate. This synaptic scaling model captures the important biological constraint that weight changes should rely only on information that is local to the synapse. It differs from global forms that use an explicit weight normalization in that the normalizing constant is not a separate parameter, but rather is implicitly determined by the circuit dynamics.
The circuit model above defines specific learning dynamics for the synaptic weights as function of the their initial values and of the incoming inputs . To investigate the evolution of the weights analytically, it is informative to first study the time course of the weight sums for an arbitrary neuron . Using the learning rule (Eq. 3) and the explicit input normalization constraint (Eq. 2), we obtain:(4)which shows that is a stationary point for the dynamics of . Furthermore, since neural activity and the learning rate are both positive, is a stable stationary point, i.e., increases when smaller than , and decreases when larger, independent of the input statistics. Consequently, synaptic plasticity automatically adjusts the sum of the incoming weights to each neuron to the total incoming drive (since ). Hence, the synaptic weights of a processing neuron adapt during learning to match the scale of its inputs. Rather than being a separate parameter, the norm of the weights is inherited from the properties of the input stimuli. We show below that this match of the normalizing constants for inputs and weights, respectively, is critical for achieving efficient learning in the neural circuit.
In contrast to the mean , which is independent of the inputs provided that the inputs are normalized, the stationary points for individual weights depend on the statistics of the input patterns. Such a dependency is, of course, needed if the circuit is to memorize properties of the input after learning. We can derive an analytical solution for learning in this system, something that has often proved difficult for other models. Specifically, if we consider the input vectors to be drawn independently and identically from a stationary but otherwise unspecified distribution , we can show (see Methods) that, at convergence, the weights associated with each neuron are uniquely determined by the statistics of input stimuli and the transfer function :(5)where the brackets denote the average of the expression under the input distribution. This approximation is very accurate for small learning rates and large numbers of inputs.
Although Eq. 5 gives a formal description for the outcome of learning in the neural circuit as a function of the neuron dynamics and the input statistics , it tells us little about the quality of the learning result. For this, we need to specify the input distribution . In particular, we use a generative model, which gives not only an explicit model for the input statistics , but also an expression for the theoretically optimal solution for inference and learning on such data, which we can use to evaluate the quality of learning in the neural circuit [22].
The specific generative model we chose is a mixture model, which is naturally associated with classification tasks [23]. Intuitively, a mixture model assumes each input stimulus to belong to one out of classes. Each class is described by a representative input and its typical variations. Mixture models have been well-investigated theoretically and are used to model a variety of data [23]. Moreover, although they may seem restrictive, mixtures are well-suited to model multi-modal data distributions also when the assumptions of the model are not satisfied exactly [23].
In generative model terminology, mixture distributions assume an input to be generated by one of model classes (see Fig. 1B). Each class is described by a representative pattern , which we will refer to as its generative field. The mixture distributions define the variations of the patterns within each class, where is the matrix of all generative fields. The prior probability specifies how many inputs are generated by the different classes. Here, we assume all classes to be equally likely, and, since inputs represent positive firing rates, we choose the Poisson distribution to model noise:(6)where is the number of input dimensions.
To capture the effects of feedforward inhibition, we assume the parameters to satisfy the constraint:(7)with parameter effectively determining the contrast of the inputs, see Fig. 1C. Note that this model only approximates the effect of feedforward inhibition, since individual stimuli are not normalized (the constraint in Eq. 2 is only true on average). However, the approximation gets increasingly accurate with increasing size of the stimuli, .
Having a model for the input distribution, we can derive the optimal solution for inference and learning on this data. In particular, we use the expectation maximization (EM) framework [24], [25] which enables us to learn the maximum likelihood solutions for the parameters from input stimuli. Intuitively, this optimal learning procedure alternates between what we call the E-step, estimating how likely the data are under the current model, and the M-step, when we change the model parameters. Iterating E- and M-steps is guaranteed to never decrease the data likelihood and, in practice, it increases the likelihood to (possibly local) likelihood maxima. If a global maximum likelihood solution is found, the parameters represent the best possible learning result (in the limit of many data points). Similarly, the posterior distribution with optimal represents the best possible inference given any specific input. For our model, we obtain the following update rules for optimal parameter learning:(8)(9)where the posterior probability required for the E-step takes the form of the well-know softmax function [26] with arguments .
With the concrete model of normalized input data, we can now ask how learning in our neural circuit is related to the theoretically optimal solutions for such data. First, recall that after learning in the neural circuit has converged, the synaptic weights are a solution of Eq. 5. Second, for the probabilistic model the (possibly local) optimum is obtained after the EM iterations have converged, which means that satisfies Eq. 9 with . Comparing the result of neural learning with the result of EM learning, we note that they have a very similar structure:(10)Indeed, synaptic weights can be easily mapped into the parameters of the generative model and if we choose the transfer function in the circuit to be equal to the posterior probability , the two expressions are the same. Hence, if we interpret neural activity as representing posterior probabilities under our model (compare [27]–[31]), any fixed point of EM optimization becomes an approximate fixed point of neural learning.
The transfer function makes learning in the neural circuit approximately optimal for normalized data, but what does this transfer function mean in neural terms? First, the optimal neural dynamics requires a specific form of lateral interactions, implementing the softmax function (Eq. 8, left-hand-side). Through these interactions, neurons compete for representing each input stimulus. Due to its importance for competitive learning, neural circuits giving rise to the softmax have extensively been investigated [26], [32]–[34]. Typically they involve unspecific feedback inhibition which suppresses neurons with weak inputs while those with strong inputs can maintain high activity rates. Most of the variants of the implementation should work for the purposes of our model (also compare [35]–[37]); hence we do not commit to one specific realization of this function.
The arguments of the softmax have a particularly simple form: they represent local summations of input activities weighted by synaptic strengths, . While the summation of inputs is biologically plausible, scaling by the logarithm of the weights may not be. It, for instance, implies that the contribution of an input to a neuron's activity may be negative or, unrealistically, change sign during learning. This problem can be addressed, however, while preserving the close correspondence between the circuit's fixed points and maximum likelihood solutions. To achieve this, we note that the only requirement for the input data is that the total input is preserved, . We therefore have some freedom when modeling how feedforward inhibition enforces this constraint. In particular, if the un-normalized input is , then feedforward inhibition could constrain the total inputs by:(11)which represents a slight alteration to the common choice . Practically, this form of normalization continues to scale the activity of an un-normalized input unit by the total activity , but it introduces an offset corresponding to having some spontaneous background activity in the input layer (which leads to a normalization constant ).
This model of feedforward inhibition guarantees that the weights will eventually converge to values larger or approximately equal to one. As a consequence, negative weight factors can be removed completely by linearizing the logarithm around one. We consider two forms of such a linearization: in the first, we use the linearization only for values of , in the second, we completely replace the logarithm by the linearized form (see inset of Fig. 1A):(12)where . For the linearization we exploited that for normalized inputs the softmax becomes invariant with respect to weight offsets (see Methods). The linear case recovers the conventional linear summation of synaptic inputs, while the logarithmic case is a closer approximation of the optimal dynamics (see Discussion).
The complete description of the final neural circuit is summarized in Table 1. It consists of essentially three elements: input normalization, Hebbian plasticity with synaptic scaling, and softmax competition (see also Fig. 1). Our analysis shows that these elementary models of neural interactions can be approximately optimal for learning on normalized inputs from mixture distributions. Notably, the neural circuit can process any type of un-normalized data as feedforward inhibition projects any stimulus to a subspace on which learning is optimal.
It is important to remark that no explicit knowledge about is required at the level of processing neurons, which would be difficult to justify neurally. Instead, synaptic scaling automatically adjusts the weights such that the constraint in Eq. 2 is satisfied. This, furthermore, means that synaptic plasticity can follow slow changes of the normalization constant , which could be used to further facilitate learning. Formally, manipulating during learning provides a simple implementation for simulated annealing, which is often used to prevent optimization from converging to locally optimal solutions [38], [39]. Alternatively, annealing can be achieved by changing the amount of spontaneous activity in the input layer (see Discussion for neural mechanisms implementing such changes).
Considering the details of the neural circuit and the generative model used here, some aspects of the analytical results presented may not seem very surprising. The similarity between the fixed points for the synaptic weights and the maximum likelihood solution is partly due to the fact that both models fulfill the same constraint, , at least approximately. However, this constraint has different origins in the two models: in the neural circuit it is a reflection of synaptic scaling, whereas in the generative model it appears due to the fact that the modeled data is normalized. Along the same lines, the fact that the softmax function emerges as the optimal transfer function for the circuit is somewhat expected, given that the softmax is closely associated with mixture models. However, the arguments of the softmax, , have a particularly compact form in our case, and they can be easily approximated through the integration of afferent inputs to the processing neurons. The compactness of the neural interactions is a direct consequence of the combination of Poisson mixture distributions, normalized inputs and synaptic scaling. Without any of these components, the interactions would be more complicated, or not optimal.
Although we have shown that learning in the neural circuit approximates optimal learning for our data model, several details remain to be investigated. First, it is unclear how close is learning in the neural circuit to the optimum in practice. Second, since real data rarely follows the assumptions of the model exactly, we would like to know how robust learning is in such cases. These questions can only be answered through numerical simulations using either simple artificial data for which the optimal solutions are known, or realistic inputs from a standard database.
Until now, we have evaluated the effectiveness of learning by measuring how well the final weights can describe the data (formally, the data likelihood under the generative model). Alternatively, we could ask how useful the emerging input representation is for performing higher level tasks in downstream circuits. The performance for such tasks can give a measure of learning quality that is more independent of specific assumptions about the input statistics. Moreover, such alternative performance measures become a necessity when comparing learning on normalized versus un-normalized data, as done in the following section. Since likelihoods are well-suited measures of learning performance only when computed using the same data, no such comparison is possible when trying to asses the benefits of normalization.
For the MNIST dataset, a natural task is classification, which has been extensively investigated in the literature, both in neural models and using purely functional approaches (e.g., [42]–[44]). Note, however, that the type of classification relevant for biological systems differs from the generic classification in several aspects. Perhaps most importantly, stimuli processed by neural circuits usually come without explicit labels. For instance, most visual stimuli we process are not accompanied by labels of the visual objects that caused them. However, during development we are provided (directly or indirectly) with the meaning of objects for some stimuli. In order to classify inputs accordingly, the model needs to have access to at least some stimuli for which the class membership (label) is known. These labels can then be used to associate the representations in the lower processing layer (obtained by unsupervised learning) with the corresponding class; for instance, all writing styles of a hand-written ‘2’ with digit class ‘2’. Having an overcomplete representation of the data becomes critical for the system to work in this setup. As we have seen in previous numerical experiments, learning with MNIST data yields representations of different classes of hand-written digits. Because of different writing styles, the variations of all patterns showing the same digit are too strong to allow for a representation of all digits with one class per digit. However, as already shown in Fig. 3D, with more neurons than classes, the emergent representation successfully captures all digit classes, with different neurons representing different writing styles (the more units, the more detailed the representation of different writing styles).
For classification, we extend the neural circuit to include an additional processing stage that makes use of the previously learned representation for assigning class labels. As done for the first processing layer, we formulate the classification process probabilistically, using a generative model assuming that a digit type generates different writing styles (Fig. 4A). This allows us to derive a probabilistic procedure for classifying a given input stimulus (see Methods). The focus here is assessing the utility of the first layer representation for higher level computations rather than the neural implementation of this later processing stage. Still, we can note that the dynamics of the second layer shares several features with the first layer model: the neural dynamics have a simple dependency on a weighted sum of incoming inputs (see Methods), and the inputs themselves are normalized (because of the softmax), suggesting this type of computation could be implemented in a neurally plausible circuit.
To illustrate classification based on the representations learned unsupervised, we first consider stimuli representing digits of types ‘0’ to ‘3’. For this data, the representations learned by unsupervised learning in the first processing layer (with units) is shown in Fig. 4B (bottom row). We label these representations using of the data used for training (i.e., we use the labels of of the training data). The probability distribution for the map between first layer representations and class labels is shown in Fig. 4B (computed using Eq. 34, see Methods), demonstrating a close to perfect assignment of representations to digit classes. For a quantitative analysis of this match, we can measure the classification performance of the system for a test dataset (i.e., for data not used for training; see Methods for details). For the four digit dataset, the classification performance as function of the number of neurons in the first processing layer is shown in Fig. 4C. For both the neural circuit and EM optimization classification performance increases with the number of units. As can be observed, the neural circuit with log-saturating synaptic efficacies shows virtually identical classification rates to EM learning. Likewise, the neural circuit with standard linear input summation shows a good classification performance, even slightly better for the complete case (four digit classes and four processing neurons). In an overcomplete setup, the rate of successful classifications is still high (e.g., around for the five times overcomplete setup), though a bit lower than for the log case and EM.
So far, we have used classification performance as an additional measure for the quality of learning in the circuit. However, the setup is interesting from a functional perspective as well, since it allows for relatively high rates of correct classification using a very limited amount of labeled data. Fig. 4E shows classification performance for different degrees of overcompleteness in the processing layer if normalized EM is applied to the full MNIST data (we use EM here as it can be efficiently scaled-up to the size of the full MNIST dataset; see Methods). As before, classification performance increases with an increasing number of units and with the number of labels used for classification (see Fig. 4E and Fig. 4F, respectively). Importantly, a small percentage of labels is already sufficient to obtain almost the same classification performance as when using all labels. For instance for processing units we obtained a performance of correctly classified stimuli using just of the MNIST labels. For rates above less than of labels were sufficient. Moreover, performance in our model is comparable to that of state-of-the-art methods, such as deep belief networks (DBN; [42]). Using all the labels, the performance of DBN reaches [42], but with a much more complex circuit (two processing layers and an associative memory), several learning mechanisms, and after the tuning of many free parameters. In contrast, learning in our model is very straightforward, with very few free parameters (), and requires just few labeled inputs. These properties seem particularly desirable in the biologically relevant setting.
Even if we assume that synaptic scaling is unavoidable to guarantee stability during Hebbian learning, it is still unclear why the system would need feedforward inhibition, or, more in formal terms, what are the benefits of learning using normalized data. This question can be addressed at two levels. First, at an abstract level, we can ask how different are the outcomes of optimal probabilistic learning when using unconstrained versus normalized data. Second, in neural terms, we can ask how learning changes when blocking feedforward inhibition in the neural circuit.
To answer the first question, we use our generative model approach to compare the optimal learning dynamics for data that is, or not, normalized (this difference will depend on the relative size of different stimuli; compare Fig. 5A and B). Formally, we construct an analog mixture model for un-normalized data, and derive optimal learning for this model. The analysis yields a similar set of update rules (see Methods, Eqs. 26 and 27), which we can use for unsupervised learning with similar (but un-normalized) data. Because the two learning procedures use different data, comparing them is nontrivial. While for data generated according to the assumed probabilistic model we can still use the percentage of trials converging to the optimum as a performance measure, comparison becomes very difficult for the digits data. Since the likelihoods are no longer comparable (because they are estimated from different data), we can only rely on the classification rates for estimating the quality of the learned representations in this case.
We compare the performance of the two learning procedures for the same two datasets described above. For the blocks dataset, learning performance is not significantly different in the two cases (not shown), probably because the task is too easy to be able to differentiate between the two learning procedures. The results for the digits are shown in Fig. 5C. The unconstrained learning procedure yields worse performance than the constrained case; the difference may seem small in absolute terms, but the classification rate for the unconstrained case is worse than the outcome of k-nearest-neighbour (k-NN) classification, which we may view as a lower bound for task difficulty. In itself, this result is not sufficient to prove that learning from normalized data is generally useful for unsupervised learning. Since we can only estimate learning performance indirectly, through the classification rates, it may be that data normalization improves classification in general, by removing task irrelevant variability, without having any specific benefit for learning per se. If this were the case, then we should observe a similar performance improvement for the normalized relative to the unnormalized data when using a standard classifier, such as k-NN. This is however not the case; on the contrary, for k-NN performance decreases to (from ) after data normalization, suggesting that the benefits of normalization are restricted to learning procedures that explicitly exploit this property, as does learning in our model.
For the neural circuit, the utility of the interaction between feedforward inhibition and synaptic scaling is further emphasized. When blocking feedforward inhibition (practically, this means using unnormalized stimuli as inputs to the circuit) the linear circuit converges to represent all classes very rarely, much less often than when feedforward inhibition is active in the circuit (Fig. 5D, compare grey and red bars). In principle, since the neural circuit approximatively implements optimal learning for normalized data, one could expect that performance should be similar to that obtained by constrained EM with un-normalized data, which is indistinguishable from that obtained when learning from normalized data. So why is there a the big difference in performance in the case of the neural circuit? The critical difference between EM and the network is that synaptic scaling only enforces the constraint of the weights through its (normalized) inputs. If the incoming stimuli are not normalized, the sum of the weights is not guaranteed to converge at all (Eq. 4 does not apply). This intuition is confirmed by the fact that when replacing synaptic scaling by an explicit weights normalization (see Methods) learning evolves similarly to the case when feedforward inhibition is active. These results suggest that feedforward inhibition is critical for correctly learning the structure of the data when the weights are constrained by biologically plausible synaptic scaling.
Our results reveal a close connection between feedforward inhibition and synaptic scaling, which could be important for cortical processing. We have shown that an elementary neural circuit with lateral inhibition, Hebbian plasticity and synaptic scaling can approximate optimal learning for normalized inputs. Furthermore, although our analysis demonstrates the approximate equivalence between learning in the neural circuit and the optimal theoretical solution only when inputs are generated by normalized mixture distributions with Poisson noise, numerical simulations using realistic data show that close to optimal learning is possible even when the inputs do not match these model assumptions exactly. Importantly, optimal learning is an outcome of a synergistic interaction between input and weight normalization, and learning is much less effective in absence of any of the two.
The mechanisms required for optimal learning in our model circuit have close correspondents in biology. First, the type of input normalization used in our model has been observed in both hippocampus and the cortex [11]. It involves a population of fast-spiking inhibitory neurons that deliver relatively homogeneous inhibition to the pyramidal cells. For a more detailed map of our model onto this circuit, we assume, in first instance, that the normalized version of the stimulus is explicitly represented in one layer, which then projects onto the processing layer. Alternatively, it is imaginable that the normalized stimuli could only be available in implicit form, without the need for an additional input layer; this would, however, require some corrections to the Hebbian learning rule, since the presynaptic term would depend on the input scale in this case. Second, learning in the circuit takes a simple local form, which has natural biological correspondents. In particular, for the linear approximation for synaptic currents, learning involves simple Hebbian plasticity and multiplicative synaptic scaling. The map to biology is somewhat more difficult for the model with logarithmic saturation of synaptic efficacies. This would translate in an unconventional type of weight-dependent Hebbian learning, and more complex additive synaptic scaling. Although there is some data on weight-dependent correlation learning [45] and additive synaptic scaling has been reported in some systems [46], the experimental evidence clearly favors the linear approximation for synaptic currents. The logarithmic version is nonetheless important, as the closest approximation to the optimal solution with bounded excitatory input currents. Moreover, it enables us to quantify the effect of the approximations in the linear model and hence to explain the difference in performance of the neural circuit relative to the theoretical optimal solution. Lastly, optimal learning requires a lateral interactions between the processing neurons, mathematically described by the softmax function. Due to its importance for competitive learning, different circuit models giving rise to softmax or softmax-like competition have been investigated previously [26], [32]–[34], [36], [37], typically involving lateral inhibitory networks with uniform connectivity onto the excitatory population. Experimentally, evidence for such lateral inhibition has recently been reported, for instance, in primary sensory cortex, where feedback inhibition relies on broadly tuned interneurons, that integrate information from pyramidal cells with diverse stimulus preference [47], confirming earlier anatomical observations (see [48] for an overview).
We have seen that the normalization constant plays an important role during learning, as it controls the sharpness of the posterior distribution which in turn influences the frequency to converge to locally vs. globally optimal solutions. Learning outcomes can be improved by annealing this parameter throughout learning. Biologically, several neuromodulators are known to affect the response properties of inhibitory neurons [49] in a way that would effectively change the normalization constant. Alternatively, the modulation of background noise can affect neuronal gain in cortical neurons [13], [15], which, in the model, has similar effects (since both change input contrast). It is tempting to speculate that the effectiveness of learning can be manipulated by systematic changes in background current or in the concentration of neuromodulators, such as acetylcholine, dopamine or noradrenaline [49], [50]. This would suggest that experimentally manipulating the concentration of these substances in the cortex should have predictable effects on learning efficiency, although these may be difficult to dissociate from other effects of such manipulations on arousal or attention [51].
Activity normalization is ubiquitous in the cortex. In particular, divisive normalization – when a neuron's response is rescaled as function of that of its neighbors – has been reported for a variety of sensory systems, from visual [52]–[54], to auditory [55], [56] or olfactory [57]. Correspondingly, a range of functions have been attributed to such normalization. It could optimize the representation of visual inputs in primary sensory areas [58], [59], facilitate the decoding of information from probabilistic population codes [60], explain attentional modulation of neural responses [61], or implement multi sensory cue integration [62]. While the form of normalization considered here is not equivalent to standard models of divisive normalization (which typically assume an L2 norm) and seems to have different neural substrates [63], several interesting parallels can be drawn with these models. In particular, we can view feedforward inhibition as a way to constrain the space of representations, similar to [59]. However, instead of asking how normalization affects the information that can be encoded in the population as a whole, we investigate how activity normalization constrains learning in neurons receiving it as inputs.
The simple, biologically plausible neural circuit proposed here achieves robust, close to optimal unsupervised learning through the interaction between feedforward inhibition and synaptic scaling. Moreover, the two are mirror processes, which need to work together for Hebbian learning to yield efficient representations of the inputs to the network. Since the type of neural mechanisms involved in our model can be found throughout the cortex, it is tempting to suggest that the interaction between feedforward inhibition and synaptic scaling could be a general strategy for efficient learning in the brain.
Learning in the neural circuit consists of iterative applications of Eq. 1 and Eq. 3 to normalized input data , which is drawn identically and independently from a stationary distribution . To facilitate numerical analysis, we assume that learning uses a finite dataset of stimuli, presented repeatedly to the network in random order. In the limit of large , this procedure becomes equal to drawing a new sample from each time.
For the learning dynamics Eqs. 1 to 3 we can show that the synaptic weights approximately satisfy Eq. 5 at convergence. The approximation holds for small learning rates and large numbers of inputs . Large learning rates would bias learning towards recent inputs. A small dataset would introduce a large sample bias such that averages across the dataset would be significantly different from expectation values w.r.t. the distribution in Eq. 5. For the derivation nested terms scaling with and applied times have to be considered, which requires a series of rather technical approximations. We, therefore, present the essential steps here and provide the details as supplemental information (Text S1).
For the derivation, we consider learning after convergence, i.e., after the changes of have reduced to changes introduced by random fluctuations due to online updates. For small these fluctuations are small. Let us denote by an iteration step after which only such small fluctuations take place. After iteration we can assume the weights to have evolved to satisfy for all (which follows from Eq. 4). For small the learning dynamics (1) to (3) is approximated by changing the weights according to followed by an explicit normalization to . More compactly, we can write:(13)where denotes the weights at the th iteration of learning, and .
We now consider another learning steps after iteration , i.e., we iterate through the inputs once again after learning has converged. By applying the learning rule (13) iteratively times, the weights are given by (see Text S1):(14)The right-hand-side can now be simplified using a sequence of approximations, all of which are based on assuming a small but finite learning rate and a large number of inputs . Below we present the main intermediate steps of the derivation and list the approximation used for each step:(15)(16)(17)where (note that is the mean of over iterations starting at iteration ).
For the first step (15) we rewrote the products in Eq. 14 and used a Taylor expansion (see Text S1):(18)
For the second step (16) we approximated the sum over in (15) by observing that the terms with large are negligible, and by approximating sums of over by the mean (see Text S1). For the last steps, Eq. 17, we used the geometric series and approximated for large (see Text S1). Furthermore, we used the fact that for small , (which can be seen, e.g., by applying l'Hôpital's rule). Finally, we back-inserted the definition of for .
By inserting the definition of into (17) and by applying the assumption that the are drawn from a stationary distribution , it follows that:(19)yielding the final expression:(20)For Eq. 19 we used the initial assumption that the weights have converged, i.e., that remains approximately unchanged after . If the same assumption is applied to Eq. 20, we obtain Eq. 5.
Note that although we have applied a number of different approximations during this derivation (compare [64] for proof sketches of some of them), each approximation is individually very accurate for small and large . Eq. 5 can thus be expected to be satisfied with high accuracy in this case; subsequent numerical simulations for a specific choice of the transfer function confirm such high accuracies.
Given a set of inputs drawn from an input distribution , optimal generative model parameters can be found by optimizing the likelihood: . A frequently used approach to find optimal parameters is expectation maximization (EM) [24], [25]. Instead of maximizing the likelihood directly, EM maximizes a lower-bound of the log-likelihood, the free-energy:(21)where and are the newly computed and previous parameters of the generative model, respectively, and where is an entropy term only depending on the previous parameters. To optimize the free-energy, EM alternates between two steps – the E-step and the M-step. First, in the E-step, the parameters are assumed fixed at and the posterior is computed for all data points . Second, in the M-step, the model parameters are updated using these posterior values. Note that for more general models, computations of expectation values w.r.t. the posteriors are considered part of the E-step. For mixture models such expectations are tractable operations, and we, therefore, often use E-step and computation of the posterior synonymously.
M-step solutions can be found by setting the derivative of the free-energy w.r.t. to zero. Applied to the concrete model of normalized input given by the mixture model (Eq. 6), we have to optimize the free-energy under the constrained of normalized weights: . We can satisfy the constraint by using Lagrange multipliers for the derivatives and obtain:Expanding the expression for the free energy and computing the partial derivatives gives (all drop out):Taking the sum over and applying the constraint , we can rewrite the above expression as:
Inserting the value of computed above and solving for yields:(22)For the normalized mixture model (Eq. 6 and Eq. 7), the posterior probability can be computed directly. By inserting the Poisson noise model and constant priors, , and by using the constraint on the weights, the posterior can be simplified as follows:(23)Note that the specific combination of normalization constraint and Poisson noise results in the final compact form of the posterior. The E-step consists of computing these posteriors for all inputs
To summarize, putting together 22 and 23, E- and M-step for our model of normalized data are given by:(24)(25)
To further simplify the computation of the posterior in Eq. 8, first note that due to normalized input, , the posterior computations remain unchanged for any offset value for the weights:(28)(29)(30)If we use an offset of we can approximate by applying a Taylor expansion around . If we use the linear approximation for values only, we obtain the function in Eq. 12. For data with as enforced by Eq. 11, the weights will converge to values greater or approximately equal to one, which makes to a very accurate approximation. If we use the linear approximation for all values of , we obtain the conventional linear summation in Eq. 12.
In order to use the representation of pattern classes in the first processing layer for classification, we consider the hierarchical generative model in Fig. 4A. The model assumes the patterns to be generated by the following process: First, choose a pattern type (e.g., for ten digit types), second, given choose a pattern class (e.g., different writing styles), and, third, given generate the actual pattern (with added noise). For the generation of pattern types we assume flat priors , i.e., we assume that each type is equally likely.
Under the assumption that the data is generated by the model, optimal inference is given by computing the posterior , where are the parameters of the model. By using the form of the graphical model in Fig. 4A, we obtain:(31)
The probabilities are given in Eq. 6 (right-hand-side). To estimate the probabilities let us first define the sets and let us assume these sets to be disjoint (no overlap). In this case we obtain:(32)(33)(34)Together with Eq. 31, the estimate for allows for a convenient way to approximate the posterior using input labels:(35)(36)
That is, we can compute the values using labeled inputs for each type . Having computed all , the approximate posterior given an unlabeled input is given by Eq. 36. Few labeled inputs can be sufficient to get good estimates for and thus for the posterior computation (compare Fig. 4B). Note that Eqs. 35 and 36 can only be regarded as approximations for optimal classification because of the assumptions made. However, they serve in providing good classification results (see Results), the can conveniently be computed after unsupervised learning, and, the can be interpreted as weights in a neural processing context.
After unsupervised learning and computation of using Eq. 35, an input is assigned to the digit type with highest posterior using Eq. 36. If the assigned type matches the true label of , the input is correctly classified. Note, in this context, that our approach would also allow for a quantification of the classifications' reliabilities by comparing the different values of .
Finally, note that the setting of few labeled inputs among many unlabeled ones is typical for semi-supervised learning. Algorithms for semi-supervised learning usually take labeled and unlabeled data into account simultaneously. As we focus on unsupervised learning and use the labels for a second stage of classification, we avoided to refer to our approach as semi-supervised.
For all simulations we initialize the weights with the mean pixel intensity averaged over all data points, with some additive uniform noise:(37)(38)(39)where is the uniform distribution in the range .
|
10.1371/journal.pgen.1006092 | Identification of Multiple Proteins Coupling Transcriptional Gene Silencing to Genome Stability in Arabidopsis thaliana | Eukaryotic genomes are regulated by epigenetic marks that act to modulate transcriptional control as well as to regulate DNA replication and repair. In Arabidopsis thaliana, mutation of the ATXR5 and ATXR6 histone methyltransferases causes reduction in histone H3 lysine 27 monomethylation, transcriptional upregulation of transposons, and a genome instability defect in which there is an accumulation of excess DNA corresponding to pericentromeric heterochromatin. We designed a forward genetic screen to identify suppressors of the atxr5/6 phenotype that uncovered loss-of-function mutations in two components of the TREX-2 complex (AtTHP1, AtSAC3B), a SUMO-interacting E3 ubiquitin ligase (AtSTUbL2) and a methyl-binding domain protein (AtMBD9). Additionally, using a reverse genetic approach, we show that a mutation in a plant homolog of the tumor suppressor gene BRCA1 enhances the atxr5/6 phenotype. Through characterization of these mutations, our results suggest models for the production atxr5 atxr6-induced extra DNA involving conflicts between the replicative and transcriptional processes in the cell, and suggest that the atxr5 atxr6 transcriptional defects may be the cause of the genome instability defects in the mutants. These findings highlight the critical intersection of transcriptional silencing and DNA replication in the maintenance of genome stability of heterochromatin.
| In eukaryotic genomes cellular processes such as transcription and replication need to be tightly controlled in order to promote genomic stability and prevent deleterious mutations. In Arabidopsis thaliana, two redundant histone methyltransferases, ATXR5 and ATXR6, are responsible for the deposition of a silencing epigenetic mark, histone H3 lysine 27 monomethylation. Loss of ATXR5/6 results in transcriptional activation of transposable elements (TEs), upregulation of DNA damage response genes and a genomic instability defect characterized as an excess of DNA corresponding to heterochromatin regions. Using a genetic screen, we sought to find suppressors of the atxr5/6 phenotype, and interestingly, we identified multiple genes implicated in general transcriptional activity. Through genomic characterization of the mutants our data suggest a model where transcriptional silencing of heterochromatin during S-phase is required for proper replication and maintenance of genome stability. These findings emphasize the important relationship between chromatin, transcriptional control and replication in the maintenance of genome stability in a eukaryotic system and identify new players involved in these processes.
| The genome represents a biological entity that is necessarily static yet retains a level of plasticity. Cells must faithfully replicate their genomes to avoid deleterious mutations, but also must be responsive to external stimuli. Eukaryotes have evolved multiple layers of epigenetic regulation that allow the genome to respond to environmental and developmental changes as well as provide a level of genome defense against parasitic genetic elements such as transposons. While epigenetic and replication fidelity pathways have traditionally been thought to be independent, multiple lines of evidence have recently implicated epigenetic mechanisms in the regulation of DNA replication [1–4].
In Arabidopsis thaliana, we previously identified two redundant histone methyltransferases, ATXR5 and ATXR6 (referred to in the aggregate as ATXR5/6), that are responsible for monomethylating lysine 27 of histone H3 (H3K27me1) [5]. Loss of these methyltransferases in the atxr5/6 double mutant leads to a severe loss of transcriptional silencing at repetitive transposable elements (TEs) [5,6]. The atxr5/6 mutants also display an unusual phenotype, wherein heterochromatin regions of the Arabidopsis genome exhibit an aberrant gain in DNA copy number (here referred to simply as over-replication). The over-replication phenotype appears to be mainly in cells which have undergone endoreduplication, a form of cell cycle without mitosis frequently observed in terminally differentiated cells [7]. The regions producing extra DNA are highly repetitive and carry epigenetic marks characteristic of silent chromatin such as DNA methylation and H3K27me1, and largely overlap with the pericentromeric regions transcriptionally derepressed in the atxr5/6 mutant.
We previously showed that mutations that strongly reduce DNA methylation in an atxr5/6 mutant background suppress the over-replication phenotype of atxr5/6 [6], suggesting that the heterochromatic nature of these regions is necessary to engender the gain in DNA copy number phenotype of the atxr5/6 mutant. In addition, while the DNA methylation mutants suppressed the over-replication phenotype of atxr5/6 mutants, they actually enhanced the transcriptional derepression phenotype [6]. Thus, the extra DNA phenotype and the transcriptional silencing phenotypes were decoupled in these mutants, showing that the extra DNA production in atxr5/6 is not required for the aberrant transcriptional activation of transposable elements.
In order to better understand the relationship between the atxr5/6 silencing and DNA copy number phenotypes, we carried out an extensive analysis of the atxr5/6 transcriptome, and compared this with the transcriptome of plants undergoing DNA damage. We also identified a number of modulators of the atxr5/6 phenotype by forward and reverse genetics approaches. Most notably, we isolated mutations in several genes, including those encoding members of the TREX-2 complex, a methyl-binding domain protein, and a SUMO dependent E3 ligase that suppressed the transcriptional defects in atxr5/6 mutants together with the genomic instability defects. Furthermore, we found that mutation of a gene involved in DNA repair and replication fork stability, BRCA1, enhanced both the atxr5/6 transcriptional and over-replication phenotypes. These results suggest a very close relationship between the loss of transcriptional silencing in atxr5/6 mutants and over-replication, consistent with a model wherein inappropriate transcription in atxr5/6 conflicts with the normal replication of heterochromatin to cause genomic instability.
We previously observed that the atxr5/6 mutants produce excess DNA corresponding to heterochromatin regions, also referred to as over-replication, that was most obvious in nuclei that had undergone endoreduplication [7], a process that is systemic in many tissue types of Arabidopsis and roughly correlates with tissue age [8]. We sought to test whether the transcriptional silencing defect of atxr5/6 mutants is similarly confined to specific tissues. To do this we analyzed two tissue types with different levels of endoreduplication, immature floral tissue that shows very low levels of endoreduplication, and cotyledons (embryonic leaves) that are very highly endoreduplicated. These tissues were analyzed by flow cytometry (Fig 1A) and whole-genome re-sequencing to confirm the state of over-replication in atxr5/6 (Fig 1B). Consistent with previous profiling of nuclei of different ploidy levels [7], we found an increase in DNA copy number in atxr5/6 mutants in cotyledons (Fig 1A) that was localized to regions of pericentromeric heterochromatin (Fig 1B). The excess DNA was absent in floral tissue of atxr5/6 mutants by both flow cytometry and sequencing analysis (Fig 1A and 1B).
We next performed RNA sequencing (RNA-seq) of flower and cotyledon tissue. We observed a reduction of the atxr5/6 transcriptional silencing defect in flowers (Fig 1C) that paralleled the lack of the extra DNA phenotype. While we were able to identify 487 TEs up-regulated in atxr5/6 cotyledon tissue relative to wild type (S1 Table), we found only 5 TEs up-regulated in floral tissue (S1 Table) and the TEs identified in cotyledon tissue showed greatly reduced transcription in atxr5/6 flowers (Fig 1C). Together these results suggest that the co-occurrence of the transcriptional defect and the over-replication defect in the atxr5/6 mutant may be connected and specific to endoreduplicated tissues.
We previously observed that, consistent with the over-replication phenotype in the atxr5/6 mutants that is likely to cause genomic instability, several genes involved in the homologous recombination (HR) DNA repair pathway were up-regulated in atxr5/6 [6]. To assess the extent of DNA damage pathway activation in the mutants, we generated RNA-seq data from wild type seedlings that had undergone gamma-irradiation, which is known to generate robust activation of DNA repair pathways. From this analysis, we identified 230 protein-coding genes that were activated 90 minutes post-irradiation (S2 Table). This identified gene set significantly overlapped (S1 Fig) with a previously published set of gamma-irradiation responsive genes identified by microarray analysis [9], though the RNA-seq method identified a larger set of genes than the microarray-based approach. Using this gene set we were able to conclude that the majority of the protein-coding genes up-regulated in atxr5/6 seedlings belong to genes that are upregulated upon irradiation (Fig 1D), implying that the atxr5/6 protein-coding gene expression changes mainly reflect a response to DNA damage. Furthermore, upon irradiation of atxr5/6 mutants, we observed a robust upregulation of the same DNA damage genes that were up-regulated upon irradiation in wild type plants (Fig 1E), indicating that DNA damage response signaling is intact in atxr5/6 mutants. Therefore, we concluded that the excessive DNA phenotype in atxr5/6 mutants is not due to a generalized failure to induce DNA damage pathways.
Using the seedling RNA-seq datasets, we also defined TEs up-regulated upon irradiation (Col 100Gy compared to Col 0Gy) as well as due to the atxr5/6 mutations (atxr5/6 0Gy compared to Col 0Gy) (Fig 1E). We identified fewer TEs up in the atxr5/6 seedling data (n = 69) than the cotyledon datasets (n = 487), likely due to the heterogeneous nature of tissues from whole seedlings that have fewer endoreduplicated nuclei than do cotyledons. Importantly, we failed to observe a large increase in transposon expression post-irradiation (Fig 1E). This was true at the 90 minute time point as well as at 24 hours post-irradiation (Fig 1E). These results indicate that the transposon silencing defect we observe in atxr5/6 mutants is most likely not simply a consequence of the DNA damage induced in those mutants.
Given our observation that atxr5/6 mutants show a generalized activation of DNA damage response pathways (especially HR genes) we sought to assess the effect of loss of DNA damage response gene in atxr5/6 mutants. To do this we generated atbrca1 atxr5/6 triple mutants. In mammals BRCA1 is a well characterized HR pathway protein important in maintaining genome stability with additional functions in cell-cycle check point and transcriptional regulation [10–13]. In plants AtBRCA1 has been shown to be necessary for efficient DNA repair [14] and is among the most highly up-regulated genes in an atxr5/6 mutant [6]. Interestingly, we found that the atbrca1 atxr5/6 mutants exhibited an enhancement of atxr5/6-induced extra DNA phenotype by flow cytometry and whole-genome sequencing of sorted 16C nuclei (Fig 2A and 2B).
We used RNA-seq of cotyledons to compare wild type, atbrca1, atxr5/6, and atbrca1 atxr5/6 plants. The atbrca1 atxr5/6 mutants showed a marked increase in expression of TEs identified as being overexpressed in atxr5/6 (Fig 2C). In addition, de novo identification of up-regulated TEs identified a greater number of reactivated TEs in atbrca1 atxr5/6 plants than atxr5/6 mutants (Fig 2D). Together these results suggest that wild type BRCA1 acts to restrict the atxr5/6 phenotype.
In order to learn more about the biological mechanisms underlying the apparent link between the over-replication and transcription phenotypes of atxr5/6, we established a forward genetic screen to identify suppressors of the atxr5/6 phenotype. We fused a GFP reporter to the promoter of the RAD51 DNA damage response gene that is highly over-expressed in atxr5/6 mutant. Transgenic lines carrying the RAD51 promoter-GFP fusion construct showed strong GFP fluorescence in cotyledons when in the presence of the atxr5/6 mutations but not when crossed to wild type plants (S2A Fig). Thus, GFP fluorescence correlated with endoreduplicated tissues showing upregulated expression of transposons and extra DNA in atxr5/6 mutants, and therefore appeared to be a suitable visual readout of the atxr5/6 phenotype. We carried out an EMS mutagenesis of atxr5/6 RAD51pro::GFP seed (referred to as RAD51pro::GFP from this point forward) and searched for suppressors of the atxr5/6 phenotype by screening for families segregating plants that had lost the cotyledon GFP expression (GFP-, S2B and S2C Fig).
Utilizing two M2 lines, EMS_2_37 and EMS_2_300, segregating for mutations causing loss of GFP signal (ems_2_37 and ems_2_300, S2C Fig), we performed flow cytometry on the GFP- plants as well as GFP+ plants from the same M2 family. For both the EMS_2_37 and EMS_2_300 lines, we observed clear suppression of the atxr5/6 extra DNA defect for GFP- plants (Fig 3A and 3B). We backcrossed the GFP- ems_2_37 and ems_2_300 mutants to the atxr5/6 line to confirm the function of the GFP protein in the F1 generation and to confirm the resegregation of GFP- plants in the resultant F2 generation (S2B Fig). We performed RNA-seq on cotyledons from the GFP- plants from this F2 generation as well as the GFP+ segregants from the same families. We found a striking reduction of TE expression in the GFP- plants as compared to both the GFP+ siblings and the starting RAD51pro::GFP line (Fig 3C), revealing that the suppression of the extra DNA in atxr5/6 by these mutants was accompanied by suppression of the atxr5/6 transcriptional silencing defect. This result is markedly different from the previously characterized suppression of atxr5/6-induced over-replication by DNA methylation mutants [6], where the loss of DNA methylation caused an increase in TE expression while suppressing the generation of extra DNA in atxr5/6 mutants.
We utilized EMS-induced mutations identified in RNA-seq datasets to map the ems_2_37 and ems_2_300 mutations (S3A and S3B Fig; S1 Text). The ems_2_37 mutation mapped to a splice site mutation at AtSAC3B (At3g06290, S3C–S3F Fig), a homolog of the yeast Sac3 protein [15]. The effect of the ems_2_37 mutation on AtSAC3B transcript splicing could be verified in the RNA-seq data, which showed clear intron retention relative to the control lines (S3C Fig). The ems_2_300 mutation was found to map to a nonsense mutation in AtTHP1 (At2g19560, S3F Fig), the homolog of yeast Thp1. AtSAC3B and AtTHP1 have been found to interact, analogously to their yeast homologs, in a complex termed TREX-2 [15]. The TREX-2 complex has been characterized in multiple systems as facilitating gene expression and RNA export from the nucleus via nuclear pore complexes (NPCs) [16–19]. Furthermore, the TREX-2 complex has also been found to act with transcribing RNA polymerase complexes to prevent the formation of deleterious transcriptional intermediates such as R-loops [20–22].
The molecular identities of ems_2_37 and ems_2_300 were confirmed by whole genome resequencing of the EMS lines to confirm the mapping of the EMS alleles (S3D and S3E Fig), as well as introgression of T-DNA-based insertional mutants for both AtSAC3B (atsac3b-3) and AtTHP1 (atthp1-1) into an atxr5/6 background to verify suppression of atxr5/6 extra DNA (S4A and S4B Fig) as well as atxr5/6-induced TE gene expression (Fig 3D). We also observed a reduction in expression of irradiation-induced genes in the trex-2 atxr5/6 triple mutants relative to atxr5/6, consistent with the suppression of the extra DNA and its relationship to the DNA damage response (Fig 3E). The RNA-seq analysis of atsac3b-3 and atthp1-1 single mutant cotyledon tissue revealed no gain in transcription for atxr5/6-induced TEs (Fig 3D) and de novo calling of up-regulated TEs in the single mutants identified only 2 and 4 TEs respectively. Finally, we verified the identities of the genes by performing complementation tests between the insertional mutants and EMS alleles (S4C Fig). We therefore renamed ems_2_37 and ems_2_300 as atsac3b-4 and atthp1-5 respectively. Subsequently we identified another EMS line, EMS_2_209, carrying a nonsense mutation in AtSAC3B (S3F Fig) by whole-genome re-sequencing and confirmed the identity of this mutation (renamed atsac3b-5) via non-complementation with the atsac3b-4 EMS line (S4C Fig).
We isolated another suppressor of the extra DNA in atxr5/6 mutants, ems_2_129, which we mapped via whole-genome re-sequencing to a nonsense mutation at mbd9 (At3g01460; Fig 4A–4D). MBD9 is a protein with a methyl CpG binding domain previously identified as a regulator of flowering time with potential roles in histone H4 acetylation [23,24]. We confirmed the identity of ems_2_129 via introgression of the mbd9-3 insertional allele into an atxr5/6 background, and by performing complementation analysis (Fig 4E). RNA-seq analysis of the mbd9-3 atxr5/6 triple mutant revealed that suppression of the atxr5/6 extra DNA was accompanied by suppression of transcription at atxr5/6-induced TEs and irradiation-induced genes (Fig 4F).
A fifth mutation, ems_2_325, like the TREX-2 and MBD9 mutants, showed suppression of both the atxr5/6 extra DNA and transposon over-expression phenotypes (Fig 5A and 5B). Mapping of the EMS-induced lesion by RNA-seq and subsequent whole-genome resequencing revealed a nonsense mutation in the coding region of At1g67180 (At-STUbL2), a relatively uncharacterized protein with a predicted N-terminal BRCT domain and C-terminal RING domain (S5A–S5C Fig). Complementation analysis was done using an insertional mutant (S5D–S5I Fig), the ems_2_325 allele was renamed stubl2-1 and the insertional FLAG_430E03 allele was renamed stubl2-2. At-STUbL2 was previously identified in a yeast two-hybrid screen for proteins that bind non-covalently to SUMO, and was shown to encode a SUMO-targeted ubiquitin E3 ligase capable of complementing the growth defects of Schizosaccharomyces pombe rfp1/rfp2 mutants [25]. In Arabidopsis, SUMO interacting proteins are highly enriched for those involved in chromatin regulation including histone and DNA methyltransferases [25]. Interestingly, At-STUbL2 was the only identified Arabidopsis STUbL protein containing the BRCT domain, which is a domain often found in proteins such as BRCA1 involved in DNA damage repair and cell cycle checkpoint control. Although the precise molecular function of At-STUbL2 is unknown, the At-STUbL2 RNA is co-expressed with ATXR6 as well as DNA repair and DNA replication genes, and the MET1 and CMT3 DNA methylation proteins that are known to function during DNA replication [26–28] (S3 Table).
Given our previous identification of DNA methylation mutants as suppressors of the atxr5/6 extra DNA phenotype, we questioned whether any of the newly identified suppressors of atxr5/6 from our forward genetic screen may have an effect on DNA methylation. To address this question, we performed whole-genome bisulfite sequencing on insertional mutants for each of the newly identified suppressors as well as wild type and atxr5/6 controls and performed differential methylated region (DMR) discovery for the mutants as well as on a previously published methyltransferase 1 (met1) dataset as a control [29]. MET1, a DNMT1 homolog, is a maintenance methyltransferase responsible for maintaining CG methylation as well as some of the non-CG cytosine methylation in the genome [29,30] and met1 mutants suppress the atxr5/6 extra-DNA phenotype while enhancing the TE silencing defect of atxr5/6 mutants [6].
Consistent with the lack of a strong TE silencing phenotype for any of the new suppressors of atxr5/6 (S6 Fig), we observed very limited alterations in DNA methylation in the mutants (Fig 6A). The one exception was the at-stubl2 mutant (Sample 5, Fig 6A), which showed a relatively large number of DMRs. We attribute this to the ecotype differences between the at-stubl2-2 Ws background and the control Col ecotype, since Arabidopsis ecotypes are known to contain differentially methylated regions [31]. In agreement with this interpretation, at-stubl2-2 mutants showed little to no alteration of DNA methylation patterns at the chromosomal level (Fig 6B). mbd9 mutants were previously reported to show global hypermethylation [24], however, while we observed a relatively higher number of CG-context hypermethylated DMRs for mbd9-3 mutants (Sample 2, Fig 6A) as compared to the other mutants, analysis of the overall genome levels of DNA methylation suggests this effect on DNA methylation is minor (Fig 6B).
Consistent with lack of strong methylation defects, the mutants isolated from our screen also did not exhibit dramatic morphological defects. This is also consistent with previously published data showing that a triple mutant of all three SAC3b related genes in Arabidopsis was reported to have no morphological defects [15] and that the mbd9 mutant displays only subtle flowering time and branching defects [23].
ATXR5/6 represents a novel link between epigenetic gene regulation and genomic instability because the atxr5/6 mutant exhibits both de-repression of transposons in pericentromeric heterochromatin, as well as an over-replication defect manifested as the production of excessive DNA from these pericentromeric regions [7]. An open question has been the causal relationship between these two phenomena. A priori we can pose three different models of this relationship: 1) atxr5/6 mutations cause replication defects that indirectly cause transcriptional defects, 2) atxr5/6 mutations cause transcriptional defects that indirectly cause replication defects, or 3) atxr5/6 mutations affect replication and transcription independently. Based on the work presented here, we favor model 2 to the exclusion of model 1, but we cannot rule out model 3. Model 1, the model wherein transcriptional defects in atxr5/6 are due to defects in DNA replication, seems very unlikely based on our previous work characterizing the impact of the loss of DNA methylation pathways on the atxr5/6 phenotype. This work indicated that the transcriptional silencing phenotype of atxr5/6 mutants was not dependent on the over-replication phenotype, since DNA methyltransferase mutants suppressed the over-replication phenotype but actually enhanced the transcriptional derepression phenotype of atxr5/6 [6]. In addition, in the current study we found that irradiation-induced DNA damage caused an upregulation of DNA damage response genes resembling that found in the atxr5/6 mutant, but did not cause a transposable element silencing defect. In contrast, the possibility that transcriptional defects in the atxr5/6 mutant are the cause of the genomic instability defects (model 2) are consistent with the data in the current study. First, by comparing immature flower tissue with cotyledon tissue we observed a co-occurrence of the transcriptional silencing defect and the extra DNA phenotype in cotyledons, both of which were absent in immature flower tissue. Furthermore, the brca1 mutation acted as an enhancer of both the transposon derepression phenotype and the extra DNA phenotype of atxr5/6. Lastly, we performed a screen for mutants that suppress the atxr5/6 phenotype, and found that every suppressor reduced both the transposon derepression phenotype and the extra DNA phenotype of atxr5/6. Thus the transcriptional silencing defects of atxr5/6 were inseparable from the abnormal pericentromeric DNA content in these mutants, suggesting that the transcriptional misregulation may be the cause of the genome instability phenotype.
If our preferred model to explain the phenotype of atxr5/6 is correct, there remain many open questions. For instance, it is not clear why ATXR5/6 appears to be exceptional among transcriptional silencers, such as DNA methylation proteins, with regards to a link to aberrant DNA replication. Because ATXR6 is expressed at the G1/S transition of the cell cycle, showing close co-expression with DNA replication licensing factors such as CDT1 and ORC2, we previously speculated that ATXR5/6 and H3K27 monomethylation may act to limit DNA replication initiation, and that extra DNA in the atxr5/6 mutant was due to inappropriate multiple firing of origins of replication (re-replication), creating “onion skins” [32] of excessive DNA near origins [7]. Although this is still a possibility, because of the tight linkage between transposon upregulation and extra DNA production in atxr5/6 as well as in our newly identified suppressors, we favor a model in which aberrant G1/S phase transcription in atxr5/6 leads to replication-transcription conflicts, which ultimately lead the production of excessive DNA in heterochromatin. As part of the normal cell cycle, there must be coordination of DNA replication with transcription, and in both prokaryotic and eukaryotic systems, failure to coordinate DNA transcription and replication results in genome instability [33–35]. This instability can be caused by direct collision between DNA and RNA polymerase complexes as well as by indirect conflicts between the complexes as is the case with R-loop formation by RNA polymerases which can act as a barrier to DNA replication fork progression [21,36]. In both the direct and indirect cases of replication-transcription conflict, the result can be replication fork stalling and collapse [36–38] which in turn leads to the formation of single-stranded DNA and recombinagenic structures that can lead to mutagenic outcomes for the genome [39]. It is thus possible that atxr5/6 mutations generate genomic instability via the release of transcriptional silencing at a critical point during the cell cycle such as S-phase, which then creates replication-transcription conflicts and hyper-recombinagenic structures that result in the amplification of repetitive DNA in pericentromeric regions. In this way, the timing of transcriptional derepression may differentiate atxr5/6 mutants from other mutants such as DNA methylation mutants that exhibit loss of TE silencing, but no effect on DNA replication.
A replication-transcription conflict model would be in line with studies of the human [40] and yeast [41,42] genomes where it has been proposed that replication stress can lead to the generation of copy number variants at repetitive DNA. In support of the notion of such a replication-specific silencing function, ATXR5/6 have been characterized as cell-cycle regulated proteins which act with the PCNA proteins normally found at replication forks [43], suggesting that these proteins function during S phase. Furthermore, ATXR5/6 have been implicated in genetic and epigenetic control of normal rDNA repeat behavior [44], and the ribosomal repeats are known sources of replication-transcription conflict in yeast [45]. The reason for the specificity of the genome instability defect for heterochromatin regions is not known, but it seems possible that the resolution of transcription and replication fork collisions may be more difficult to complete in heterochromatin regions due to the more inaccessible nature of heterochromatin. This could also help explain why DNA methylation mutants suppress the genome instability defect of atxr5/6 mutants, since severe reduction of DNA methylation would render these regions much less like heterochromatin and more like euchromatin, for instance through reduced levels of the H3.1 histone variant recently shown to be required for over-replication in atxr5/6 [46].
Our identification of the TREX-2 complex as being necessary for the genomic instability defect in atxr5/6 mutants also supports the hypothesis of replication-transcription conflict driving the atxr5/6 genome instability, because components of this complex were isolated in a yeast genetic screen for factors affecting the viability of a strain genetically predisposed to accumulate aberrant replication intermediates [47]. TREX-2 mutants were found to rescue the viability of a replication-deficient strain where replication forks were destabilized in a manner that is phenomenologically similar to our observations of the trex-2 atxr5/6 triple mutants. In yeast, genetic rescue by the TREX-2 mutants was proposed to act via the loss of topological strain created by the normal gene gating facilitated by TREX-2/THO. Interestingly, this defect depended on transcription but not R-loop formation. In addition, the TREX-2 complex is also required for transcriptional efficiency [17], and TREX-2 was shown to promote RNA Pol II transcription through its interaction with the Mediator complex [16]. This is consistent with our finding that TREX-2 mutants reduce the inappropriate transcription of heterochromatin seen in atxr5/6 mutants, and suggests that loss of TREX-2 may alleviate replication stress present in an atxr5/6 genome, which is otherwise undergoing heterochromatic transcription. Similarly, MBD9, which is identified here and which has been characterized as a transcriptional activator of the flowering gene FLC [24], likely acts to promote transcription, such that loss of MBD9 would alleviate transcription-induced replication blocks in a manner similar to TREX-2 mutants. Finally, although the function of At-STUbL2 is not known, since this mutant also reduces the transposon over-expression phenotype of atxr5/6, we propose that STUbL2 acts via similar mechanisms as the TREX-2 and MBD9 mutants, and may encode a transcriptional regulator.
An alternative model to explain the correlation between the transcriptional defects and over-replication defects in atxr5/6 is that R-loops generated by inappropriately transcribing transposons directly cause mutagenic events leading to excessive DNA production, even in the absence of replication-transcription conflicts [48]. R-loops are formed during the process of transcription where the RNA strand pairs with the complementary DNA strand, leaving the other DNA strand free, and exposing the cell to potentially mutagenic single stranded DNA[21]. If atxr5/6 mutants fail to properly resolve R-loops in heterochromatin, this could explain the heterochromatin specificity of the excessive DNA phenotype, and also explain why mutants that suppress the transcriptional upregulation of atxr5/6 also suppress the over-replication defect. Consistent with this model, BRCA1 is known to play a role in the prevention of DNA damage due to transcription associated R-loops [13], and brca1 mutants enhanced the excessive DNA damage phenotype of atxr5/6 mutants. Although TREX-2 is also known to be involved in resolving R-loop structures[49], the trex-2 mutants from our screen dramatically reduced the transposon de-repression defect of atxr5/6 mutants, and therefore would also dramatically reduce the abundance of R-loops. One observation that does not fit well with the R-loop model is that brca1 mutant enhanced both the magnitude of the transposon over-expression phenotype and the over-replication defect, and it is difficult to understand how failure to resolve R-loop-induced DNA damage would lead to an increase in transcription. Clearly, the mode of action of Arabidopsis BRCA1, TREX-2, and the other suppressors identified here, will be an important question for future studies.
Given the emerging importance of the interaction between epigenome and genome stability for models of disease such as cancer [2,50,51], it will be important to further test the replication-transcription conflict and other models of the atxr5/6 phenotypes since further understanding of this phenomena may inform other models and systems where loss of transcriptional control leads to genomic instability.
All strains used in this study, unless otherwise indicated, were in a Columbia (Col) ecotype background. Details regarding the strain information as well as the generation of the RAD51pro::GFP line can be found in the Supplemental Experimental Procedures (S1 Text). In addition, S4 Table details the genotypes of lines used in high-throughput sequencing experiments.
10-day old seedlings were irradiated on plates via exposure to a Cs-137 source following the general experimental design previously described [9].
EMS mutagenesis of ~2000 atxr5/6 seeds carrying the RAD51pro::GFP transgene was carried out as previously described [52].
All flow cytometry analysis and FACS was performed as previously described [6]. For cotyledon tissue, cotyledons from at least 20 plants were pooled, whereas for leaf or floral tissue 3 plants were typically pooled.
The DNA-seq libraries presented in Figs 1 and 2 were generated as previously described [6,7]. The DNA-seq libraries in S3, S5 Figs and Fig 4 were similarly prepared regarding DNA extraction and Covaris shearing, but the libraries were prepared using either the Illumina DNA TruSeq or Nugen Ultralow Ovation kits (see GEO accession GSE77735 for details).
All RNA-seq libraries were prepared using a standard Trizol (Life Technologies) RNA extraction followed by library generation with the Illumina RNA TruSeq kit. All RNA was derived from the cotyledon tissue of >20 plants unless otherwise indicated. For all libraries two biological replicates were performed unless otherwise indicated.
For whole genome bisulfite sequencing libraries, libraries were generated from 3-week-old adult leaf material using the NuGen Ovation Ultralow Methyl-Seq kit before being bisulfite converted with the Qiagen Epitect bisulfite kit using the FFPE protocol. All libraries were sequenced on an Illumina HiSeq instrument.
Base calls were performed using the standard Illumina pipeline and all reads were aligned to the TAIR10 genome (www.arabidopsis.org). For DNA-seq libraries reads were aligned using the Bowtie aligner [53], for RNA-seq Tophat2 [54]was used, and for whole-genome bisulfite data the BSmap aligner was used [55]. Protein-coding genes were defined as described in the TAIR10 annotation (www.arabidopsis.org) and transposable elements were defined using a previously described list [56] that had been updated to the TAIR10 assembly. All statistical analysis was performed in an R environment. Details of bioinformatics data analysis can be found in the Supplemental Experimental Procedures (S1 Text).
The sequencing data have been deposited in the Gene Expression Omnibus (GEO) database under accession number GSE77735.
|
10.1371/journal.pcbi.1006227 | General differential Hebbian learning: Capturing temporal relations between events in neural networks and the brain | Learning in biologically relevant neural-network models usually relies on Hebb learning rules. The typical implementations of these rules change the synaptic strength on the basis of the co-occurrence of the neural events taking place at a certain time in the pre- and post-synaptic neurons. Differential Hebbian learning (DHL) rules, instead, are able to update the synapse by taking into account the temporal relation, captured with derivatives, between the neural events happening in the recent past. The few DHL rules proposed so far can update the synaptic weights only in few ways: this is a limitation for the study of dynamical neurons and neural-network models. Moreover, empirical evidence on brain spike-timing-dependent plasticity (STDP) shows that different neurons express a surprisingly rich repertoire of different learning processes going far beyond existing DHL rules. This opens up a second problem of how capturing such processes with DHL rules. Here we propose a general DHL (G-DHL) rule generating the existing rules and many others. The rule has a high expressiveness as it combines in different ways the pre- and post-synaptic neuron signals and derivatives. The rule flexibility is shown by applying it to various signals of artificial neurons and by fitting several different STDP experimental data sets. To these purposes, we propose techniques to pre-process the neural signals and capture the temporal relations between the neural events of interest. We also propose a procedure to automatically identify the rule components and parameters that best fit different STDP data sets, and show how the identified components might be used to heuristically guide the search of the biophysical mechanisms underlying STDP. Overall, the results show that the G-DHL rule represents a useful means to study time-sensitive learning processes in both artificial neural networks and brain.
| Which learning rules can be used to capture the temporal relations between activation events involving pairs of neurons in artificial neural networks? Previous computational research proposed various differential Hebbian learning (DHL) rules that rely on the activation of neurons and time derivatives of their activations to capture specific temporal relations between neural events. However, empirical research of brain plasticity, in particular plasticity depending on sequences of pairs of spikes involving the pre- and the post-synaptic neurons, i.e., spike-timing-dependent plasticity (STDP), shows that the brain uses a surprisingly wide variety of different learning mechanisms that cannot be captured by the DHL rules proposed so far. Here we propose a general differential Hebbian learning (G-DHL) rule able to generate all existing DHL rules and many others. We show various examples of how the rule can be used to update the synapse in many different ways based on the temporal relation between neural events in pairs of artificial neurons. Moreover, we show the flexibility of the G-DHL rule by applying it to successfully fit several different STDP processes recorded in the brain. Overall, the G-DHL rule represents a new tool for conducting research on learning processes that depend on the timing of signal events.
| Most learning rules used in bio-inspired or bio-constrained neural-network models of brain derive from Hebb’s idea [1, 2] for which “cells that fire together, wire together” [3]. The core of the mathematical implementations of this idea is multiplication. This captures the correlation between the pre- and post-synaptic neuron activation independently of the timing of their firing.
Time is however very important for brain processing and its learning processes [4]. Differential Hebbian learning (DHL) rules [5, 6] are learning rules that change the synapse in different ways depending on the specific timing of the events involving the pre- and post-synaptic neurons. For example, the synapse might tend to increase if the pre-synaptic neuron activates before the post-synaptic neuron, and decrease if it activates after it. As suggested by their name, DHL rules use derivatives to detect the temporal relations between neural events. Here we will use the term event to refer to a relatively short portion of a signal that first monotonically increases and then monotonically decreases. Events might for example involve the activation of a firing-rate unit in an artificial neural network, or the membrane potential of a real neuron, or a neurotransmitter concentration change. DHL rules use the positive part of the first derivative of signals to detect the initial part of events, and its negative part to detect their final part. By suitably multiplying the positive/negative parts of the derivative of events related to different signals, DHL rules can modify the synapse in different ways depending on how their initial/final parts overlap in time.
To the best of our knowledge, current DHL rules are basically two: one proposed by Kosko [5] and one proposed by Porr, Wörgötter and colleagues [6, 7]. These rules modify the synapse in specific ways based on the temporal relation between the pre- and post-synaptic events. Formulating other ways to modify synapses based on event timing is the first open problem that we face here.
The development of dynamical neural-network models and learning mechanisms that, as DHL, are able to take time into consideration is very important. Indeed, the brain is an exquisitely dynamical machine processing the continuous flow of information from sensors and issuing a continuous flow of commands to actuators so its understanding needs such types of models [8–11]. In this respect, neuroscientific research on spike timing dependent plasticity (STDP; [12]) clearly shows how synaptic changes strongly depend on the temporal relation between the spikes of the pre- and post-synaptic neurons. Given the typical shape of spikes, an important class of STDP models, called phenomenological models [13], abstracts over the features of the spike signals and directly links the synaptic strengthening, Δw, to the time interval separating the pre-synaptic and post-synaptic spikes, Δt, on the basis of a function of the type Δw = f(Δt) [12, 14]. Such a function is usually designed by hand and reflects the synaptic changes observed in experimental data. [15]. The function f(Δt) generates a typical learning kernel that when plotted shows a curve where each Δt causes a certain Δw. Phenomenological models are simple but are applicable only to spike events. In comparison, DHL rules are more complex but have the advantage of computing the synaptic update as the step-by-step interaction (based on multiplication) between the pre-synaptic and post-synaptic events. Therefore they are applicable to any complex signal that might exhibit events with variable time courses.
When applied to the study of STDP, the property of DHL rules just mentioned also opens up the interesting possibility of using them to investigate the actual biophysical neural events following and caused by the spikes that actually lead to the synaptic change, as first done in [16]. The chain of processes changing the synapse is also captured by biophysical models (e.g., see [14, 17]). These models can capture those processes in much biological detail (mimicking specific neurons, neuromodulators, receptors, etc.) but at the cost of being tied to specific phenomena. Because the level of abstraction of DHL rules lies between that of phenomenological models and that of biophysical models, DHL represents an important additional research tool.
Experimental study of STDP [18, 19] shows that different types of neurons, for example excitatory/inhibitory neurons in different parts of the brain, implement a surprisingly rich repertoire of learning kernels. It is reasonable to assume that the brain employs such learning mechanisms to implement different computational functions. In this respect, an interesting fourth class of models appropriate for studying STDP, which might be called functional models, aims to derive, or to justify, specific STDP learning kernels based on normative computational principles [20–23].
Investigating the functions of different STDP kernels is not in the scope of this work. However, assuming that the variety of learning kernels discovered through STDP experiments supports different functions relevant to neural processing and that analogous functions might be needed in artificial neural networks, it is important to understand the computational mechanisms that might generate such a variety of learning kernels. In this respect, an important question is this: is there a DHL learning rule, or a set of them, that can generate the complete variety of learning kernels found in the brain? Some existing research shows how different STDP learning kernels can arise from the same biophysical mechanisms [17], or from the same DHL-based model [24]. However, these studies propose specific mechanisms to address a sub-set of STDP data sets rather than proposing a general way to systematically reproduce STDP learning kernels. Understanding the extent to which DHL can capture the known STDP phenomena, and how this can be done, is thus a second important open problem that we address here.
The rest of the paper addresses the two open problems indicated above in the following ways. As a first contribution of the paper, the Section ‘G-DHL and the systematisation of DHL’ considers the first open problem—how different DHL rules can be generated in a systematic fashion—by proposing a general framework to produce DHL rules. In particular, the section first reviews the DHL rules proposed so far in the literature; then it presents the G-DHL rule and shows how it is able to generate the existing DHL rules and many others; and finally it shows how one can filter the neural signals to generate events that correspond to the features of interest and can use memory traces to apply the G-DHL rule to events separated by time gaps.
As a second contribution of the paper, the Section ‘Using G-DHL to fit STDP data sets’ deals with the second open problem—understanding if and how G-DHL can be used to capture known STDP phenomena. To this end, the section first illustrates how the G-DHL synapse update caused by a pre- and post-synaptic spike pair can be computing analytically rather than numerically, and then it presents a collection of computational tools to automatically search the rule components and parameters to fit a given STDP data set.
Addressing the same second open problem, and as a third contribution of the paper, the Section ‘Using G-DHL to fit STDP data sets’ uses those computational tools to show how the G-DHL rule is able to reproduce several learning kernels from the STDP literature. To this end, the section first uses G-DHL to fit the classic STDP data set of Bi and Poo [25]; then it illustrates how the G-DHL components found by the fitting procedure can be heuristically useful to search for the biophysical mechanisms underlying a given STDP data set; and finally it shows how to apply the G-DHL rule to systematically capture different aspects of all the STDP data sets reviewed by Caporale and Dan [18] (such as their temporal span, long-term potentiation/depression, and variability around zero inter-spike intervals—e.g. sharp depression-potentiation passages, non-learning plateaus, Hebbian/anti-Hebbian learning).
The Section ‘Discussion’ closes the paper by analysing the main features of G-DHL and its possible development. All software used for this research is available for download from internet (https://github.com/GOAL-Robots/CNR_140618_GDHL).
As discussed in the introduction, different types of neurons exhibit surprisingly different STDP learning kernels. For this reason we tested the flexibility of G-DHL by using it to capture several different STDP learning kernels involving pairs of pre- and post-synaptic spikes. In the future G-DHL could be extended to capture STDP processes involving spike triplets or quadruplets ([41]; see [42] for a model) by considering three or more multiplication elements rather than only two as done here.
To apply G-DHL to spike pairs, we first outline the procedure used to derive the formulas to compute G-DHL analytically, rather than numerically as done so far. The procedure is illustrated in detail in Section 2.1 in S1 Supporting Information in the case in which one assumes that spikes and traces are described with some commonly used formulas. Sections 2.7 and 2.8 in S1 Supporting Information show a method that leverages these formulas to use G-DHL to fit STDP data sets; examples of this fitting are shown in the Section ‘Results’.
Before presenting the formulas, we discuss two important points. The closed-form formulas for synaptic updates by the G-DHL rule have two main advantages. First, they allow the mathematical study of the G-DHL rule (see Sections 2.2 and 2.6 in S1 Supporting Information). Second, the formulas allow a computationally fast application of G-DHL by computing the synaptic update through a single formula rather than as a sum of many step-by-step synaptic updates as done in its numerical application, an advantage exploited in the computationally intensive simulations of the Section ‘Results’.
A second observation concerns the relation between the G-DHL explicit formulas and phenomenological models discussed in the introduction. The G-DHL explicit formulas have the form Δw = f(Δt) typical of phenomenological models. This shortcut is possible because spikes have a fixed shape: this implies that Δt is the only information relevant for computing G-DHL. The resulting synaptic update is however the same as the one that would be obtained by numerically simulating the step-by-step interaction between the pre- and post-synaptic neural events mimicking more closely what happens in the real brain. Therefore, the possibility of computing Δw = f(Δt) formulas for DHL rules does not violate what we said in the introduction, namely that G-DHL captures the mechanisms causing the synaptic update at a deeper level with respect to phenomenological models.
The procedure for the automatic fit of STDP data sets was first employed to fit the classic STDP data set of Bi and Poo from rat hippocampal neurons [25]. Fig 8a summarises the results (for ease of reference, henceforth we will refer to synapse strengthening/weakening as ‘LTP—long term potentiation’ and ‘LTD—long term depression’). The model comparison technique selected two G-DHL components: an LTP component (σpp = 0.73) and an LTD component (ηps = −0.025). The parameters σ and η differ in scale as they refer to differential and mixed G-DHL components involving signal-derivative or derivative-derivative multiplications.
Fig 8b shows the target data and their fit obtained with the G-DHL components and parameters shown in Fig 8a. The G-DHL regression fits the data accurately (FVU = 0.2725). While the original paper performed the fit with the usual exponential function for both positive and negative Δt, the G-DHL regression captures the LTP with the σpp ‘sharp’ component (Fig 6), concentrated on small positive inter-spike intervals, and the LTD with the ηps = −0.025 ‘softer’ component (Fig 7), concentrated on negative intervals.
We now illustrate with an example the idea of using the components found by the G-DHL regression to heuristically search for biophysical mechanisms possibly underlying a target STDP data set. This example involves the Bi and Poo’s data set [25] analysed in the previous section. The idea relies on the observation that each multiplication factor of the G-DHL components identified by the regression procedure has a temporal profile that might correspond to the temporal profile of the pre-/post-synaptic neuron electrochemical processes causing the synaptic change.
The steps of the procedure used to search the biophysical mechanisms are as follows: (a) identify with an automatic procedure the G-DHL components and parameters fitting the target STDP data set; (b) define the temporal profile of the two pre-/post-synaptic factors of each found component, and the LTP/LTD effects caused by the component; (c) identify possible biophysical processes having a temporal profile similar to the one of the identified factors; (d) design experiments to verify if the hypothesised biophysical processes actually underlie the target STDP phenomenon in the brain. We now give an example of how to apply the steps ‘a’ and ‘b’, and some initial indications on the step ‘c’, in relation to the Bi and Poo’s data set [25]. The example aims to only furnish an illustration of the procedure, not to propose an in-depth analysis of this STDP data set.
Regarding step ‘a’, Fig 8 shows that the G-DHL regression identified two LTP and LTD components.
Regarding step ‘b’, Fig 9 shows the temporal profile of the factors of the two components. The first component is a ‘positive-derivative/positive-derivative’ component ([ u ˙ 1 ] + [ u ˙ 2 ] +; Fig 9a, left graph) with two factors (Fig 9b, left graph): (a) a relatively long pre-synaptic factor ([ u ˙ 1 ] +) lasting about 30 ms; (b) a shorter post-synaptic factor ([ u ˙ 2 ] +) lasting about 7 ms. These two factors, amplified by a positive coefficient (σpp = + 0.73), produce LTP concentrated on small positive inter-spike intervals (0 ms < Δt < 30 ms; Fig 9a, left graph).
The second component is a ‘positive-derivative/signal’ component ([ u ˙ 1 ] + u 2; Fig 9a, right graph) with other two factors (Fig 9b, right graph): (a) a relatively long pre-synaptic factor ([ u ˙ 1 ] +) lasting about 30 ms; (b) a longer post-synaptic factor (u2) lasting about 50 ms. The two factors, amplified by a negative coefficient (ηps = −0.025), produce LTD covering negative-positive inter-spike intervals (−30ms < Δt < 20ms; see Fig 9a, right graph).
When the two components are summed, LTP more than cancels out LTD for positive delays (0ms < Δt < 20ms). This causes the sharp passage from LTD to LTP around the critical Δt values close to zero, which characterise the target kernel (Fig 8).
Regarding step ‘c’ of the procedure, directed to identify possible biological correspondents of the component factors identified in step ‘b’, we now discuss some possible candidate mechanisms that might underlie the factors identified for the Bi and Poo’s data set. Note that these brief indications are only intended to show the possible application of the procedure, not to make any strong claim on the possible specific mechanisms underlying such STDP data set.
Pioneering studies on hippocampus have shown that a repeated stimulation of the perforant path fibres enhances the population response of downstream dentate granulate cells (long-term potentiation–LTP; [47–49]). LTP also takes place in other parts of brain such as the cortex [50], amygdala [51], and the midbrain reward circuit [52]. Other studies have shown the existence of long-term depression (LTD), complementary to LTP, in various parts of brain, for example hippocampus [53, 54] and motoneurons [55]. More recent research has shown that LTP and LTD, and their intensity, depend on the duration of the temporal gap separating the pre- and post-synaptic spikes (spike time-dependent plasticity—STDP; e.g. [56], see [18] for a review). The relation between the time-delay and the synaptic change depends on the types of neurons involved (e.g., glutamatergic vs. GABAergic neurons [57, 58]), the position of the synapse (e.g., [59]), and the experimental protocols used (e.g., [60]).
Early findings that blocking NMDA receptors (NMDARs) can prevent both LTP and LTD, while a partial blocking can turn an LTP effect into an LTD, has led to the proposal of several calcium-based models of synaptic plasticity (e.g., [61–64]). One view proposes that two independent mechanisms can account for the classic STDP learning kernel [19, 65]. This is in line with the two components, and their factors, found by our G-DHL based regression of Bi and Poo data set. The first component was an LTP ‘positive-derivative/positive-derivative’ component ([ u ˙ 1 ] + [ u ˙ 2 ] +) formed by two factors. The first factor was a pre-synaptic factor ([ u ˙ 1 ] +) lasting about 30 ms, compatible with a short-lived effect involving the pre-synaptic glutamatergic neuron spike and affecting the post-synaptic NMDARs [66]. The second factor was a post-synaptic factor ([ u ˙ 2 ] +) lasting about 7 ms, compatible with a back-propagating action potential (BAP; [67]). The second component was a ‘positive-derivative/signal’ LTD component ([ u ˙ 1 ] + u 2) formed by two factors: a relatively slow pre-synaptic element, ([ u ˙ 1 ] +), lasting about 30 ms, and a slow post-synaptic element, (u2), lasting about 50 ms. Different biological mechanisms might underlie these two factors. In this respect, there is evidence that post-synaptic NMDARs might not be necessary for spike-timing-dependent LTD [68], while this might be caused by metabotropic glutamate receptors (mGluR; [69]), voltage gated calcium channels (VGCC; [25, 69]), pre-synaptic NMDAR [70], or cannabinoid receptors [68, 69].
We tested the generality of G-DHL by fitting all STDP kernels reported in the review of Caporale and Dan [18]. The data sets addressed in this review encompass many different STDP experiments reported in the literature and proposes a taxonomy to group them into distinct, and possibly exhaustive, classes. The taxonomy is first based on the excitatory or inhibitory nature of the pre- and post-synaptic neurons, giving the classes: (a) excitatory-excitatory; (b) excitatory-inhibitory; (c) inhibitory-excitatory; (d) inhibitory-inhibitory. Some neurons in different parts of brain belong to the same class but exhibit different STDP learning kernels: in [18], these have been grouped in ‘subtypes’ (sub-classes) called ‘Type I’, ‘Type II’, etc.
For the G-DHL regressions we used the original data when the authors of the experiments could furnish them. When this was not possible, we used the data extracted from graphs in the publications. Figs 10 and 11 summarise the outcome of the G-DHL-based regressions for the different data sets. For each data set, the figures report this information: (a) left graph: original data and, when available, regression curve of the original paper; (b) right graph: regression curve based on G-DHL; (c) top-center small graph: function with which the review [18] proposed to represent the STDP class of the data set. In the following, we illustrate the salient features of these regressions. Section 3 in S1 Supporting Information presents more detailed data on all the regressions as those presented in Fig 8 for the data set of Bi and Poo.
Understanding the functioning and learning in dynamical neural networks is challenging but also very important for advancing our theories and models of the brain—an exquisitely dynamical machine. Differential Hebbian Learning (DHL) might become a fundamental means to do so. Existing DHL rules are few, basically two [5, 7], and are not able to model most spike-timing dependent plasticity (STDP) phenomena found so far in the brain. Building on previous pioneering research, this work addresses these limitations in multiple ways. First, it proposes a framework to understand, use, and further develop DHL rules. In particular, it proposes a general DHL (G-DHL) rule encompassing existing DHL rules and generating many others, and highlights key issues related to the pre-processing of neural signals before the application of DHL rules. Second, it proposes procedures and formulas for applying DHL to model STDP in the brain. Third, it shows how the proposed G-DHL rule can model many classes of STDP observed in the brain and reviewed in [18].
With respect to other approaches for modelling STDP, DHL represents a complementary tool in the toolbox of the modeller and neuroscientist. First, DHL differs from ‘phenomenological models’. Although simple and elegant, these models update the synapse based on mathematical functions directly mimicking the synaptic changes observed in empirical experiments in correspondence to different inter-spike intervals [14, 15]. Instead, DHL rules compute the synaptic update on the basis of the step-by-step interactions between levels of and changes in the neural variables of interest. DHL rules also differ from ‘biophysical models’. These models can reproduce many biological details but have high complexity and rely on phenomenon-specific mechanisms (e.g., [14, 17]). Instead, DHL rules reproduce fewer empirical details but at the same time, after the systematisation proposed here, they represent ‘universal mechanisms’ able to capture many STDP phenomena.
G-DHL relies on two main ideas. The first idea, elaborated starting from previous proposals [5] (see also [29]), is that the derivative of an ‘event’, intended as a monotonic increase followed by a monototic decrease of a signal, gives information on when the event starts and terminates. This information is used by G-DHL to update the connection weight depending on the time interval separating the pre- and post-synaptic neural events. The second idea is that the actual synaptic update can rely on different combinations of the possible interactions between the pre-/post-synaptic events and their derivatives, thus leading to a whole family of DHL rules.
Mathematically, this gives rise to a compound structure of the G-DHL rule which is formed by a linear combination of multiple components. In this respect, the capacity of G-DHL to capture different STDP phenomena is linked to the power of kernel methods used in machine learning [34, 35]. The linear form of the rule facilitates its application through manual tuning of its parameters, as shown here and in some previous neural-network models of animal behaviour using some components of the rule [80–82]. The linear form of the rule also facilitates the automatic estimation of its coefficients when used to capture STDP data sets, as also shown here.
G-DHL has a high expressiveness, as shown here by the fact that we could use it to accurately fit multiple STDP data sets. In particular, the G-DHL components form basis functions that are well suited to model key aspects of STDP, in particular its long-term potentiation/depression features, its time span, and its variability around the zero inter-spike interval (e.g., sharp depression-potentiation passages, non-learning plateau, Hebbian/anti-Hebbian learning). The regressions of the data sets targeted here employed seven out of eight components of the rule. The regressions are particularly reliable because the optimisation procedure used here is highly robust with respect to local minima, so they show the utility of most G-DHL components for modelling different STDP data sets. Future empirical experiments might search for STDP processes corresponding to the eighth non-used G-DHL component (encompassing a multiplication between the pre-synaptic stimulus and the post-synaptic derivative negative part): this corresponds to a relatively long LTD peaking at a negative inter-spike interval but also involving low-value positive intervals.
The results of our regression based on G-DHL of the classic STDP kernel, represented by the classic Bi and Poo data set [25], suggests the possible existence of two distinct mechanisms underlying LTP and LTD involved in such STDP learning kernel, so it is interesting to compare this result with different views in the literature. A specific hypothesis on calcium control of plasticity was formulated in [83] and was followed by significant experimental evidence. According to this hypothesis, post-synaptic calcium transients above a lower threshold cause LTD whereas calcium transients above a second higher threshold produce LTP. In a detail model [84], this phenomenon is captured with a single mechanism for which the synaptic change is caused by calcium concentrations at the post-synaptic neuron modulated by the temporal relation between the current at the pre-synaptic neuron (causing NMDAR opening) and the back-propagating action potential (BAP) at the post-synaptic neuron [67]: low levels of post-synaptic calcium cause the synapse depression whereas high levels cause its enhancement. Models of such type have been criticised on the basis of empirical evidence. According to [65], calcium models require a long-fading BAP-induced transients to account for LTD when the BAP occurs before the pre-synaptic action potential [12]. Moreover, calcium models also predict a pre-post form of LTD even when the BAP occurs beyond a given time from the pre-synaptic action potential. While this pre-post form of LTD has been registered in hippocampal slices [74], other data [25] indicate that it is not a general feature of STDP. In this respect, our findings agree with other proposals for which two independent mechanisms account for LTP and LTD in the classic STDP learning kernel [19, 65]. Future work might extend these preliminary results. In particular, it could aim to understand in detail how some of the mechanisms mentioned above implement change detectors and these lead to STDP, as predicted by the G-DHL core functioning mechanisms based on derivatives. Moreover, G-DHL could be used to heuristically guide the identification of the biophysical mechanisms underlying different STDP data sets beyond the classic kernel.
Future work might also investigate, both computationally and empirically, DHL rules different from G-DHL, namely: (a) DHL rules formed by three or more components (useful to model STDP involving more than two spikes [41]); (b) DHL rules using orders of derivatives higher than the first one used in G-DHL [32, 33]; (c) DHL rules generated by other types of filters, rather than [ u ˙ ] + and [ u ˙ ] - used in G-DHL, to detect the increasing and decreasing parts of events.
Another line of research might aim to investigate the possible computational and behavioural functions of the different G-DHL components. In this respect, the analysis presented here on the computational mechanisms underlying STDP might contribute to the current research on the possible functions of such plasticity [20–23]. Indeed, this research mainly focuses on the computational function of the classic STDP learning kernel [25], whereas the research presented here, by stressing how the brain uses different DHL rules, calls for the investigation of their different possible functions.
A different approach to understand the functions of different DHL rules and STDP kernels might use embodied neural models to understand their utility to support adaptive behaviour. The development of G-DHL was in fact inspired by the need to implement specific learning processes in neural-network models able to autonomously acquire adaptive behaviours [80–82]. Thus, it could for example be possible to establish a particular target computation or behaviour and then automatically search (e.g. with genetic algorithms or other optimisation techniques) the rule components and coefficients that are best suited for them. For example, previous work [85] used a learning rule based on Kosco’s DHL rule [5] to obtain interesting/surprising emergent behaviours in physical simulated agents. This approach might test other G-DHL components to produce different behaviours.
|
10.1371/journal.pgen.1000236 | Two Pathways Recruit Telomerase to Saccharomyces cerevisiae Telomeres | The catalytic subunit of yeast telomerase, Est2p, is a telomere associated throughout most of the cell cycle, while the Est1p subunit binds only in late S/G2 phase, the time of telomerase action. Est2p binding in G1/early S phase requires a specific interaction between telomerase RNA (TLC1) and Ku80p. Here, we show that in four telomerase-deficient strains (cdc13-2, est1Ä, tlc1-SD, and tlc1-BD), Est2p telomere binding was normal in G1/early S phase but reduced to about 40–50% of wild type levels in late S/G2 phase. Est1p telomere association was low in all four strains. Wild type levels of Est2p telomere binding in late S/G2 phase was Est1p-dependent and required that Est1p be both telomere-bound and associated with a stem-bulge region in TLC1 RNA. In three telomerase-deficient strains in which Est1p is not Est2p-associated (tlc1-SD, tlc1-BD, and est2Ä), Est1p was present at normal levels but its telomere binding was very low. When the G1/early S phase and the late S/G2 phase telomerase recruitment pathways were both disrupted, neither Est2p nor Est1p was telomere-associated. We conclude that reduced levels of Est2p and low Est1p telomere binding in late S/G2 phase correlated with an est phenotype, while a WT level of Est2p binding in G1 was not sufficient to maintain telomeres. In addition, even though Cdc13p and Est1p interact by two hybrid, biochemical and genetic criteria, this interaction did not occur unless Est1p was Est2p-associated, suggesting that Est1p comes to the telomere only as part of the holoenzyme. Finally, the G1 and late S/G2 phase pathways for telomerase recruitment are distinct and are likely the only ones that bring telomerase to telomeres in wild-type cells.
| Duplication of linear DNA is complicated by the fact that conventional DNA polymerases cannot copy their ends. From yeasts to humans, replication of DNA ends, called telomeres, is accomplished by a telomere-dedicated reverse transcriptase called telomerase that uses its RNA subunit as a template. We show that there are two genetically distinct pathways that recruit yeast telomerase, Est2p, to telomeres in a cell cycle–dependent manner. Only one of the two pathways, the pathway operating late in the cell cycle, is required for telomere maintenance. In telomerase-deficient strains, the amount of Est2p that is telomere-bound late in the cell cycle is about 50% of wild type levels. Maintenance of functional levels of Est2p late in the cell cycle requires that Est1p, another telomerase subunit, be telomere-bound. In addition, Est1p must be associated with Est2p via an interaction between it and telomerase RNA. Human telomerase is not active in most somatic cells, but is critical for stem cell longevity. Even a modest reduction in telomerase has a serious impact on human health. The sensitivity of yeast to reduced levels of telomere-associated telomerase may help us understand why human stem cells require high levels of telomerase for their survival.
| Telomerase is a specialized reverse transcriptase that lengthens the 3′ end of telomeric DNA. In Saccharomyces cerevisiae, the template for telomere elongation is a short stretch within the 1158 base TLC1 telomerase RNA. Est2p is the S. cerevisiae telomerase catalytic subunit, while Est1p and Est3p are two telomerase subunits whose roles in telomerase action are less well understood. Although Est2p and TLC1 RNA are sufficient for telomerase catalytic activity in vitro, all three EST proteins, as well as TLC1 RNA are required in vivo. Telomerase deficient strains such as tlc1Δ, est2Δ, est1Δ, and est3Δ are viable but slowly lose telomeric DNA (reviewed in [1]). After 50–100 generations, when telomeres are very short, chromosome loss increases in these strains, and most cells in the population die, a collection of behaviors known as the ever shorter telomere (est) phenotype [2].
The key events in S. cerevisiae telomere replication and processing occur in late S/G2 phase. Most of the ∼300 bp yeast telomere is replicated by semi-conservative DNA replication, which occurs very late in S phase. After semi-conservative replication, C-strand resection generates ∼50–100 base G-tails at both ends of DNA molecules. These G-tails are repaired by C-strand resynthesis prior to mitosis [3]–[5]. Telomerase lengthening of telomeres also occurs late in the cell cycle [6],[7].
Cdc13p is a single-strand TG1–3 sequence specific DNA binding protein [8],[9] that associates in vivo with the G-tails that constitute the very ends of yeast chromosomes [10],[11]. Although the Cdc13p complex has an essential role in protecting telomeres from degradation [12]–[14], there are also alleles of CDC13, such as cdc13-2, that have normal end protection activity but confer an est phenotype [9]. Cdc13p and Est1p interact by two-hybrid, co-immuno-precipitation [15], and genetic criteria [16]. Moreover, fusions between the DNA binding domain of Cdc13p and Est2p can maintain telomeres in the absence of Est1p [17]. Together these data suggest that Est1p acts by recruiting Est2p to the telomere in late S/G2 phase. This recruitment is thought to occur via a specific interaction between Est1p and Cdc13p that is lost in telomerase defective cdc13-2 cells. The association of Est3p with Est2p is Est1p-dependent [18].
In previous work, we used chromatin immuno-precipitation (ChIP) to test different aspects of this recruitment model [19]. Consistent with the model, Est1p binding to telomeres is limited to late S/G2 phase, and Cdc13p binding, which occurs throughout the cell cycle, increases enormously at this time, concomitant with the appearance of long G-tails. However, Est2p is telomere associated throughout most of the cell cycle, not just in late S/G2 phase as the model predicts. The high Est2p binding in late S/G2 phase is reduced by ∼50% in the telomerase deficient cdc13-2 strain, while Est2p binding earlier in the cell cycle is unaffected. The telomere association of Est2p at times when telomerase is not active is also inferred by fluorescent in situ hybridization, which shows co-localization of TLC1 telomerase RNA with telomeres in G1 and S phase cells [20]. Unexpectedly, Est1p binding was equivalent in wild type (WT) and cdc13-2 cells, although the signal to noise ratio for Est1p in these early ChIP experiments was low [19].
A more sensitive ChIP assay was used to determine the requirements for Est2p binding in G1 and early S phase [21]. Mutations in TLC1 (tlc1Δ48) or YKU80 (yku80-135i) that disrupt the ability of TLC1 RNA to interact with Ku80p both in vitro [22],[23] and in vivo [21] eliminate Est2p telomere binding in G1 and early S phase [21]. In addition, Est2p and Est1p telomere binding in late S/G2 phase is reduced to ∼50% of WT levels in these mutants. Although telomeres in tlc1Δ48 and yku80-135i cells are shorter than in WT cells, neither strain has an est phenotype [22],[23]. Thus, telomerase binding to yeast telomeres in G1 and early S phase is not required for telomere maintenance.
Here we extend the analysis of the requirements for Est2p telomere binding. Several additional est mutations, est1Δ, tlc1-SD, and tlc1-BD, had the same Est2p telomere binding profile as cdc13-2 cells. Thus, reduced levels of Est2p binding specifically in late S/G2 phase correlated with an est phenotype, while a WT level of Est2p binding in G1 was not sufficient to maintain telomeres. The specific interaction between TLC1 and Est1p that is lost in tlc1-SD and tlc1-BD cells [24], resulted in low, but detectable, Est1p telomere binding as did the cdc13-2 mutant. Est1p telomere binding was not detected in an est2Δ strain. Together, these data show that WT levels of Est2p telomere binding in late S/G2 phase require that Est1p bind telomeres, a binding that requires a specific interaction between Est1p and TLC1 RNA and that is reduced in cdc13-2 cells. No Est2p or Est1p was detected at telomeres when both the G1/early S and the late S/G2 phase pathways for Est2p recruitment were disrupted (as in tlc1Δ48 cdc13-2 or yku80-135i tlc1-SD double mutant cells). Thus, these pathways are likely the only ones that recruit telomerase to yeast telomeres in vivo.
Chromatin immuno-precipitation (ChIP) was used to determine the telomere association of proteins involved in telomere maintenance in WT and mutant cells. We used strains in which the protein being studied was multiply epitope tagged at its endogenous locus and was the only form of the protein in the cell. Functionality of epitope tagged proteins was determined by their effects on telomere length and other telomere phenotypes. In earlier experiments, we used a Myc-tagged Est2p that was not fully functional as telomeres were ∼50 bps shorter in its presence [19]. When this Myc-tagged Est2p was introduced into an est1Δ strain, cells senesced so rapidly that it was not possible to determine whether Est2p binds telomeres in est1Δ cells.
The function of epitope tagged Est2p was improved by inserting a flexible linker of eight glycine residues between the carboxyl terminus of Est2p and multiple Myc epitopes (hereafter called Est2-G8-Myc) [21],[25]. This allele supports WT telomere length, does not senesce when combined with deletion of YKU, and est1Δ cells carrying this Est2-G8-Myc can be grown for 50 to 100 cell divisions before they senesce. This Est2-G8-Myc allele was used to determine if Est2p telomere binding is Est1p-dependent.
Otherwise isogenic WT or est1Δ cells expressing Est2-G8-Myc were arrested in late G1 phase with alpha factor (0 minute time point) and then released into the cell cycle at 24°C. Samples were taken at 15 min time intervals and processed for FACS to determine position in the cell cycle, and by ChIP to determine Est2p association with telomeric DNA. Under these conditions, cells were in G1 phase at 0 and 15 min, in S phase at 30, 45, and 60 minutes, and in G2/mitosis at 75 and 90 min [19],[21] (and data not shown). None of the mutations or epitope tagged proteins had reproducible effects on cell cycle progression [19],[21] (and data not shown). ChIP samples were analyzed by quantitative multiplex PCR using primer pairs specific for the modified VII-L telomere (TEL), sub-telomeric VII-L DNA (ADH), or a sequence far from a telomere (ARO) (Figure 1A, B, left). Alternatively we examined association with the native VI-R telomere and ARO (Figure 1A, B right). For all synchronies, a representative gel from one of the three or more independent synchronies is shown. Fold enrichment is binding at telomere VII-L or VI-R relative to binding at ARO and normalized to input DNA [21]. Graphs are the compiled data from the three or more independent synchronies for a given strain; error bars are one standard deviation from the average for each time point.
As shown previously [19],[21], in WT cells Est2p had high telomere association in G1 and early S phase (0 through 30 min), a modest decline in mid-S phase, a second peak in late S/G2 phase (60 minutes), and then a decline in association as cells progressed to the end of the cell cycle (Figure 1A, white squares). Est2-G8-Myc association with the VII-L telomere in est1Δ cells was similar to what was seen in WT cells except that binding in late S/G2 phase (60–90 min) was significantly reduced (Figure 1A left, black circles; see figure legends for P values). A similar pattern of Est2-G8-Myc binding in the absence of Est1p was seen at telomere VI-R (Figure 1A right). This pattern of binding was very similar to that seen for Est2-G8-Myc binding in another telomerase defective strain, cdc13-2 (Figure 1B, black triangles). The results for cdc13-2 presented here with Est2-G8-Myc are similar to our published data carried out with the less functional Est2p-Myc and using different quantitation methods [19]. Reduced Est2p binding in est1Δ and cdc13-2 cells was not due to reduced abundance of Est2p (Figure 1C, lanes labeled est1Δ and 13-2). We conclude that Est1p is not essential for Est2p telomere binding but is required for WT levels of Est2p binding in late S/G2 phase.
In addition to changing our tagging strategy, a series of technical changes were made in the ChIP protocol that increased the signal to noise ratio in these experiments [21]. These changes were particularly important for Est1p since in our earlier work, the association of Est1p with telomeres in late S phase was enriched only five fold over background [19]. Under the conditions of these previous experiments, Est1p bound equally well to telomeres in WT and cdc13-2 cells. Since the result was surprising (see introduction), we redid this experiment using our more sensitive ChIP methods.
As shown previously [19], in WT cells, Est1p binding to the VII-L telomere peaked in late S phase (60 min) (Figure 2A, left, white boxes). With the improved ChIP protocols, peak Est1p binding was 20-fold above background. Contrary to our previous results, Est1p telomere binding to the VII-L telomere in cdc13-2 cells was much lower than WT, only ∼4 fold above background (Figure 2A, left, black triangles). Although this binding was low, it was significantly higher than the signal with the no-tag control strain at all time points (Figure 2A). Reduced but significant Est1p binding was also seen at telomeres VI-R (Figure 2A, right) and XV-L (data not shown) in cdc13-2 cells. Western analysis demonstrated that this reduced Est1p telomere binding was not due to an effect of the cdc13-2 mutation on Est1p levels (Figure 2B, compare WT lanes to lanes labeled 13-2). We attribute the difference in these results compared to our earlier studies to the increased sensitivity of the current ChIP assay.
By multiple criteria, Est1p interacts with Cdc13p, and this interaction is thought to recruit Est2p to telomeres (see introduction). If Est1p binding depends solely on its ability to interact with Cdc13p, Est1p might bind telomeres even in an est2Δ strain. However, using synchronous cells, Est1-Myc binding to telomeres VII-L and VI-R in est2Δ cells (Figure 2C, black triangles) was very low, similar to background levels (Figure 2C, no tag, white triangles). The very low Est1p at est2Δ telomeres supports the interpretation that the signal for Est1p telomere binding in cdc13-2 cells (Figure 2A) was real. The absence of Est1p binding was not due to reduced levels of Est1p in est2Δ cells (Figure 2B, compare WT to est2Δ lanes).
TLC1 RNA is immuno-precipitated with Est1p [26],[27]. RNA structure analysis identified a potential stem-bulge region contained within nucleotides 600 to 669 that is conserved among different yeasts [24]. Both the putative 9 bp stem and the 5 nucleotide bulge are essential for telomerase function in vivo. TLC1 alleles that delete either the bulge (tlc1-BD, bulge deletion allele) or that reduce base-pairing in the stem (tlc1-SD, stem disruption allele) yield an est phenotype. The stem-disruption compensatory mutation (tlc1-SC, stem compensatory), which restores the potential for stem formation in the tlc1-SD allele, maintains WT length telomeres. Several lines of evidence indicate that the stem-bulge region interacts with Est1p. Est1p (but not Est2p) over-expression suppresses the est phenotype of tlc1-BD cells. Moreover, neither TLC1-BD RNA nor TLC1-SD RNA immuno-precipitates with Est1p. The loss of interaction with TLC1 RNA is specific for Est1p as both mutant RNAs immuno-precipitate with Est2p [24].
We used ChIP to determine if the stem-bulge region is also essential for Est1p binding to telomeres (Figure 3). Est1p-Myc association with the VII-L telomere was determined in synchronized tlc1-SD cells (Figure 3A). Est1p association was low, 4-fold over background, although the difference between it and the no tag strain had high significance only at the 45 min timepoint (Figure 3A). This low association was not due to reduced Est1p levels (Figure 3C, lane labeled SD).
Because the WT and tlc1 mutant strains proceeded similarly through the cell cycle, the level of Est1p telomere binding in asynchronous cultures can be used to compare Est1p binding in different backgrounds. In asynchronous cells, Est1-Myc binding in WT cells was 11.6 fold (telomere VII-L) or 9.4 fold (telomere VI-R) (Figure 3B). The level of Est1-Myc binding in asynchronous cells expressing the telomerase defective tlc1-SD (3.2 fold, VII-L; 2.0 fold, VI-R; lanes labeled SD) or tlc1-BD (3.3 fold, VII-L; 2.4 fold, VI-R; lanes labeled BD) alleles was not significantly different from binding in tlc1Δ cells (2.5 fold, VII-L; 3.0 fold, VI-R) (Figure 3B). In all three tlc1 mutants, Est1p binding was low but significantly higher than in the no-tag control. In contrast, the telomerase proficient tlc1-SC allele supported high levels of Est1-Myc telomere binding, (9.9 fold, VII-L; 6.3 fold, VI-R; lanes labeled SC), a level that was not significantly different from WT. Low Est1-Myc binding was not due to difficulties detecting proteins at short telomeres as Cdc13p-Myc telomere binding was high in all backgrounds (Figure 3D). Est1p abundance was similar in WT and mutant strains (Figure 3C). We conclude that Est1p interaction with the conserved stem-bulge region of TLC1 RNA is required for normal levels of Est1p telomere binding.
By co-immunoprecipitation, Est2p interacts normally with both TLC1-SD and TLC1-BD RNAs [24]. Although the late S/G2 peak of Est2-G8-Myc binding did not occur in est1Δ cells (Figure 1A), this effect could be due to an off-telomere effect of Est1p on Est2p structure or function that can not occur in the absence of Est1p. Alternatively, Est1p telomere binding may be required for Est2p binding. Since Est1p was present at normal levels (Figure 3C) but had low telomere association in tlc1-SD and tlc1-BD cells (Figure 3A, B), it is possible to distinguish between these possibilities in these strains.
In synchronous tlc1-SD cells, Est2-G8-Myc binding to the VII-L telomere was very similar to the pattern of Est2-G8-Myc binding in est1Δ and cdc13-2 cells (compare Figure 4A, tlc1-SD, black diamonds, to Figure 1A, est1Δ, black circles and 1B, cdc13-2, black triangles). That is, Est2-G8-Myc binding was at WT levels in G1 and early S phase but was significantly reduced in late S/G2 phase. Est2-G8-Myc binding was also determined in asynchronous cells expressing the TLC1 alleles (Figure 4B). As shown previously [19], Est2p binding was absolutely dependent on TLC1: Est2-G8-Myc binding was at background levels in tlc1Δ cells (enrichment of 1.3 fold, VII-L; 0.8 fold, VI-R). The high Est2-G8-Myc binding in WT cells (17.3 fold, VII-L, 12.2 fold VI-R) was significantly reduced but still detectable in asynchronous tlc1-SD cells (VII-L 10.1 fold; VI-R, 7.6 fold), consistent with normal Est2-G8-Myc binding throughout most of the cell cycle and reduced binding in late S/G2 phase. Similar results were seen in asynchronous tlc1-BD cells (enrichment at VII-L, 8.0 fold; VI-R, 8.2 fold). The reduction in Est2-G8-Myc telomere binding in tlc1-SD and tlc1-BD cells was not due to a reproducible decrease in Est2p abundance (Figure 4C). These results demonstrate that Est1p is needed in cis (i.e., at the telomere) to support WT levels of Est2p telomere binding in late S/G2 phase.
Eliminating the specific interaction between Yku80p and TLC1 RNA with the tlc1Δ48 or yku80-135i mutations [21],[23] eliminates Est2p at the telomere in G1 and early S phase [21]. Mutations that strongly reduced the amount of telomere bound Est1p, such as cdc13-2, tlc1-SD and tlc1-BD, or an est1Δ allele that eliminates Est1p altogether, lack high levels of Est2p binding in late S/G2 phase [19]; (Figure 1A, B; 4A). To determine if these pathways are the only ones that recruit telomerase to yeast telomeres, Est2p telomere binding was examined in double mutants that eliminate both pathways. Est2-G8-Myc binding was determined in synchronous tlc1Δ48 cdc13-2 cells (Figure 5A). Although Est2p abundance was normal in this background (Figure 1C, lane DM, double mutant), Est2-G8-Myc binding to the VII-L telomere (Figure 5A, black triangles) was not significantly different from the no tag control (Figure 5A, white triangles). Est2-G8-Myc binding was also very low at both the VII-L and VI-R telomeres in asynchronous yku80-135i tlc1-SD, yku80-135i tlc1-BD and tlc1Δ48 est1Δ cells (Figure 5B). Est1-Myc binding to both the VII-L and VI-R telomeres was indistinguishable from the no-tag control strain in synchronous tlc1Δ48 cdc13-2 cells (Figure 6). We conclude that the TLC1-Ku mediated pathway that recruits Est2p to telomeres in G1 phase [21] and the pathway that requires specific interactions of Est1p with both Cdc13p (Figure 2A) and TLC1 RNA (Figure 3A, B) that maintains high levels of telomere bound Est2p in late S/G2 phase are the only pathways that recruit Est2p to DNA ends in otherwise WT cells.
A specific interaction between a 48 bp stem-loop region in TLC1 RNA and Yku80p brings Est2p to the telomere in G1 and early S phase [21]. This TLC1-Ku interaction is also required for WT levels of telomere-associated telomerase in late S/G2 phase as both Est2p (∼40–50% of WT) and Est1p (∼33% of WT) telomere binding are reduced in these backgrounds. Cells that lack the Ku-TLC1 RNA interaction (tlc1Δ48, yku80-135i, and ykuΔ) have short but stable telomeres and do not senesce.
In contrast, four est strains examined here (est1Δ, cdc13-2, tlc1-SD, and tlc1-BD) as well as est3Δ cells (data to be published elsewhere), had WT levels of telomere associated Est2p in G1 and early S phase but reduced (∼50% of WT) Est2p telomere binding in late S/G2 phase (Figure 1A, B; 4A). The only est strain (other than est2Δ) that did not have this pattern was tlc1Δ in which there was no Est2p telomere binding at all [19] (Figure 4B). Reduced Est2p telomere binding was not associated with a marked reduction in Est2p abundance except in the tlc1Δ strain [19] (Figure 1C, 4C).
Est2p binding in late S/G2 phase was Est1p dependent (Figure 1A). However, the presence of Est1p was not sufficient for WT Est2p telomere association as Est2p binding was equally reduced in late S/G2 phase in tlc1-SD and tlc1-BD cells (Figure 4A, B) where Est1p was present (Figure 3C), but is neither TLC1-associated [24] nor telomere-bound (Figure 3A, B). The fact that Est2p does not bind telomeres at all in a tlc1Δ strain [19] (Figure 4B) is consistent with TLC1 driving Est2p telomere association throughout the cell cycle via specific interactions with other proteins, Ku80p in G1/early S phase and Est1p in late S/G2 phase.
The G-tail binding Cdc13p is not a telomerase subunit [28]. Therefore, the late S/G2 phase reduction in Est2p telomere binding in cdc13-2 cells [19] (Figure 1B) is unlikely due to a change in telomerase structure. This strain also had low Est1p binding, ∼25% of WT levels (Figure 2A) yet Est1p and Est2p abundance was normal in cdc13-2 cells (Figure 1C, 2B). These data can be explained if the holoenzyme comes to the telomere (or is held at the telomere) in late S/G2 phase via a specific interaction between Est1p and Cdc13p that is impaired in cdc13-2 cells [17]. Low but detectable Est1p binding in cdc13-2 cells is consistent with the effects of this mutation on telomerase recruitment to a double strand break (DSB) that is generated next to a tract of telomeric DNA. In cdc13-2 cells, Est1p binding to the break was much lower than in WT cells for up to 2 hrs after DSB formation, but at 3 hrs, Est1p binding was >50% of the WT level [29]. Est1p telomere binding was also reduced when it was unable to interact with TLC1 RNA as in tlc1-SD and tlc1-BD cells (Figure 3A, B). The Cdc13p and TLC1 pathways for Est1p recruitment are not redundant as Est1p binding was low when either interaction was disrupted. However, the pathways must be somewhat independent as Est1p levels at the telomere in late S/G2 phase were ∼25% of WT in both cdc13-2 (Figure 2A) and tlc1-SD cells (Figure 3A), higher than the background level of Est1p telomere binding in est2Δ (Figure 2C) and tlc1Δ48 cdc13-2 cells (Figure 6). Likewise, in asynchronous cells, Est1p telomere binding was statistically indistinguishable in tlc1Δ, tlc1-SD, and tlc1-BD cells. However, in each strain, binding was higher in the mutant than in the no-tag control (Figure 3B). Thus, a small but significant amount of Est1p can associate with telomeres in the complete absence of TLC1 RNA. We speculate that this low level association is due to the Est1p-Cdc13p interaction.
Together with earlier findings, the data presented here support several conclusions. First, the G1 and the late S/G2 phase Est2p recruitment pathways must be the only ones that bring telomerase to yeast telomeres in otherwise WT cells since there was no telomere associated Est2p or Est1p in doubly mutant strains (tlc1Δ48 cdc13-2, tlc1Δ48 est1Δ, yku80-135i tlc1-SD, yku80-135i tlc1-BD; Figure 5, 6). Second, G1 bound Est2p was neither sufficient (e.g., est1Δ) nor necessary (e.g., tlc1Δ48) to maintain telomeres by telomerase. In fact, it is possible that the G1 recruitment pathway contributes to telomere length solely by protecting ends from degradation [30]. This view is supported by the finding that ≥60% of the Est2p that is telomere associated in G1 phase is located at least ∼100 bps from the chromosome end and thus is not in a position to lengthen telomeres [31]. Third, Est1p interacts poorly' with Cdc13p unless it is part of the holoenzyme. Three est mutants, tlc1-SD, tlc1-BD (Figure 3C) and est2Δ (Figure 2B) had WT levels of Est1p, but there was low (tlc1-SD, tlc1-BD; Figure 3A, B) or no (est2Δ; Figure 2C) Est1p at the telomere in these strains. Thus, in vivo, the interaction between Cdc13p and Est1p detected by biochemical and genetic methods [15],[16] either does not occur or is not stable at telomeres unless Est1p is part of the holoenzyme. Fourth, while reduced Est2p binding in late S/G2 phase correlated with an inability to maintain telomeres by telomerase, it was not sufficient to confer an est phenotype. The levels of telomere associated Est2p and Est1p in the non-senescing tlc1Δ48, yku80-135i, ykuΔ and tel1Δ strains [31],[32] are similar to what was seen here for four est mutants (Figure 1A, B; Figure 4A, B), yet telomeres in tlc1Δ48, yku80-135i, ykuΔ, and tel1Δ cells, while shorter than WT, are stable [22],[23],[33].
There are several mutually non-exclusive explanations for why similarly low levels of telomerase in late S/G2 phase support telomerase proficiency in some backgrounds (eg., tel1Δ, tlc1Δ48, yku80-135i, ykuΔ) and an est phenotype in others (cdc13-2, tlc1-BD, tlc1-SD). For example, there may be fairly subtle quantitative differences between Est2p and/or Est1p binding between the two mutant classes that are not detected by ChIP. Alternatively, there may be qualitative differences between the telomerase that is telomere associated in the two classes of mutants, such as post translational modification of telomerase subunits or the presence of Est3p. Another possibility is that telomere structure is different between EST and est strains in late S/G2 phase, and the telomeric structure found in est cells makes it harder for low amounts of telomerase to engage properly with chromosome ends. The idea that a specific telomere structure is required for telomerase activity is supported by the observation that forced association of Est1p and Est3p with Est2p in G1 phase cells is not sufficient to support telomerase-mediate telomere elongation [18].
From yeasts to humans, the amount of telomerase per cell is surprisingly low [34],[35]. Reducing this already low level by mutation in one of several telomerase components [36]–[39] or by altering a telomere structural protein [40] can cause fatal diseases, such as dyskeratosis congenita or idiopathic pulmonary fibrosis [41],[42]. Here we show that in some genetic backgrounds ∼50% of WT levels of Est2p at yeast telomeres in late S/G2 phase is not sufficient to maintain telomeric DNA and prevent cellular senescence. An understanding of why yeast cells are sensitive to reduced levels of telomerase in some genetic backgrounds but not in others may help clarify why even relatively modest reductions in telomerase levels in human stem cells affects their survival.
All experiments were carried out in YPH499 [43] background that was modified by insertion of URA3 immediately adjacent to the left telomere of chromosome VII [44] to generate YPH499-UT, and the BAR1 gene was deleted and replaced with kanMX6 [21]. Proteins were epitope tagged at their endogenous loci as described [11],[21],[25] in a manner that places TRP1 at the tagged locus. Briefly, Est1 was tagged at its carboxyl terminus with nine Myc epitopes [19], and Est2p was tagged at its carboxyl end with a Gly8 linker followed by 18 Myc epitopes [21],[25]. Cdc13p was tagged at its carboxyl terminus with 9 Myc epitopes [19]. Complete deletions of TLC1 (replaced by LEU2), EST1 (replaced by HIS3), and EST2 (replaced by HIS3) were generated using PCR-mediated transformation [45]. The TLC1 alleles tlc1-SD (stem disruption; 3 bp disrupted in predicted stem), tlc1-SC (stem disruption compensatory; potential for base-pairing restored to tlc1-SD), and tlc1-BD (bulge deletion; deletion of 5 bulged nucleotides) described in [24] were generously provided by Tom Cech and introduced by integration into the genome. The tlc1Δ48 and yku80-135i alleles are described in [23] and were generously supplied by Dan Gottschling. The cdc13-2 [16], est1Δ, tlc1Δ, est2Δ, tlc1Δ48, and yku80-135i mutations were generated as heterozygous diploids expressing Myc-tagged proteins. Doubly mutant strains were derived from heterozygous diploids at both loci that also expressed the desired Myc-tagged protein. In both cases, the heterozygous diploids were sporulated. Freshly dissected spores of the desired genotypes were identified by replica plating, grown up, and used immediately for ChIP analyses.
The ChIP experiments were carried out, analyzed by multiplex PCR, and quantified exactly as described [21]. Briefly, relative fold enrichment of a protein with telomeres was determined by (TELIP/AROIP)/(TELinput/AROinput) where input is the amount of the DNA sequence that was PCR amplified in the samples before precipitation and IP is the amount of the sequence in the anti-Myc immuno-precipitate. Cell synchrony experiments were carried out as in [21]. Briefly, 30°C grown, log phase cells (A660 = 0.3) cells were arrested in late G1 phase using alpha factor (Sigma), removed from alpha factor (0 minutes), and then allowed to proceed through a synchronous cell cycle at 24°C. Samples were removed at 15 min intervals and processed for ChIP and FACS (fluorescent activated cell sorting) analysis at each time point. For synchrony experiments, the data for each time point are presented as the mean of the three or more independent synchrony experiments plus or minus one standard deviation from the mean. Likewise, values for asynchronous cultures are the mean plus or minus one standard deviation for three independent cultures. Statistical significance was determined using a two-tailed Student's t test. For the purposes of this paper, P values ≤0.05 were considered significant.
|
10.1371/journal.pgen.1008057 | Cellular response to small molecules that selectively stall protein synthesis by the ribosome | Identifying small molecules that inhibit protein synthesis by selectively stalling the ribosome constitutes a new strategy for therapeutic development. Compounds that inhibit the translation of PCSK9, a major regulator of low-density lipoprotein cholesterol, have been identified that reduce LDL cholesterol in preclinical models and that affect the translation of only a few off-target proteins. Although some of these compounds hold potential for future therapeutic development, it is not known how they impact the physiology of cells or ribosome quality control pathways. Here we used a genome-wide CRISPRi screen to identify proteins and pathways that modulate cell growth in the presence of high doses of a selective PCSK9 translational inhibitor, PF-06378503 (PF8503). The two most potent genetic modifiers of cell fitness in the presence of PF8503, the ubiquitin binding protein ASCC2 and helicase ASCC3, bind to the ribosome and protect cells from toxic effects of high concentrations of the compound. Surprisingly, translation quality control proteins Pelota (PELO) and HBS1L sensitize cells to PF8503 treatment. In genetic interaction experiments, ASCC3 acts together with ASCC2, and functions downstream of HBS1L. Taken together, these results identify new connections between ribosome quality control pathways, and provide new insights into the selectivity of compounds that stall human translation that will aid the development of next-generation selective translation stalling compounds to treat disease.
| A fundamentally new approach to treat human diseases caused by “undruggable” proteins would be to use small molecules to selectively inhibit their synthesis by the ribosome. Here we compare two related compounds (PF846, PF8503) that selectively stall human translation to the effects of a more general translation inhibitor, homoharringtonin. We used genome-wide approaches to probe the effects of these compounds, including measurements of which messenger RNAs are being translated and the effects of knocking down gene expression on cell growth. These experiments revealed new and surprising genetic connections between ribosome quality control pathways. We then used biochemical and cell-based experiments to test the involvement of particular ribosome quality control proteins such as ASCC2, ASCC3 and HBS1L in the physiological response of cells to translation inhibitors. The genetic and biochemical insights presented here should aid the development of next-generation selective translation inhibitors to treat disease.
| Proteins are the main druggable therapeutic targets for the treatment of human diseases, ranging from metabolic disease, to cancer and dementia [1]. Most therapeutic strategies consist of inactivating key protein enzymatic or binding activities by using small molecules or antibodies. However, many proteins remain “undruggable” due to the difficulty in identifying small molecule or biological inhibitors of their function [2]. An alternative approach would be to prevent the synthesis of target proteins in the first place, by targeting the step of translation [3]. However, targeting translation in a highly specific way remains an unsolved problem. Many known translational inhibitors bind to the ribosome in close proximity to messenger RNA (mRNA) or transfer RNA (tRNA) binding sites, and thereby interfere with different steps of the translation process [4]. These mechanisms of action, even if context-specific, lead to substantial inhibition of translation which is extremely toxic, as exemplified by the many classes of antibiotics that target the bacterial ribosome [5,6]. In humans, the low specificity and high toxicity of homoharringtonine (HHT), the only translational inhibitor approved as a drug treatment, makes this compound useful as a last resort treatment for chronic myeloid leukemia (CML) [7,8], but it is not clear how its mechanism could be repurposed for specific drug targets.
As opposed to affecting translation for much of the proteome, a new class of compounds has been identified that selectively targets the translation of PCSK9, with few off-target effects [9–11]. These compounds bind in the ribosome exit tunnel, and seem to affect the trajectory of the growing nascent polypeptide chain in the exit tunnel as it is extended [12], thereby selectively stalling translation of a narrow spectrum of transcripts. Understanding the molecular basis for how these compounds selectively stall translation will be critical for the future design of transcript-specific translation inhibitors to target previously undruggable proteins [2]. However, designing new molecules as selective translation inhibitors to treat disease also necessitates a better understanding of how cells respond and adapt to compound-induced translational stalling.
Genomic screens serve as powerful and unbiased tools to identify genetic modifiers of the action of small molecule inhibitors. They can be used to discover proteins involved in an inhibitor’s mechanism of action, and to assess possible targets for combination therapy [13,14]. In particular, CRISPR interference screens (CRISPRi) allow robust and highly specific knockdown of gene transcription with minimal off-target effects [15]. These screens can be used to compare the effects of inhibitors on cell growth as a function of genetic background, and to identify components of genetically related pathways and how these are influenced by drug treatment [16]. Here, we used a genomic CRISPRi screen using a compound that selectively stalls the translation of human PCSK9, to test the effects of this class of compound on human cell fitness. We used compound PF-06378503 (PF8503) [9], a compound related to the PCSK9 selective translational inhibitor PF-06446846 (PF846) described previously [10], but which is slightly more toxic, to exert the selective pressure required for growth-based CRISPRi screens [17]. We show using ribosome profiling that PF8503 inhibits translation of an overlapping set of proteins compared to PF846, along with a distinct set of off-target proteins. In the CRISPRi screen, we identified proteins that suppress or enhance the toxicity of PF8503, a number of which are associated with translation and ribosome quality control pathways. We used targeted CRISPRi to validate the involvement of these proteins in cell fitness during PF8503 treatment, and to identify genetic interactions between ribosome quality control pathways. We also compared the genes identified in the PF8503 CRISPRi screen with those identified using the non-selective translation inhibitor HHT, in order to reveal the pathways likely to be general cellular responses to the stress induced by translation inhibition. Taken together, these results reveal pathways affected by selective stalling of translation, and suggest the cell dependencies most likely to be impacted by this class of inhibitors.
Growth based CRISPRi genetic screens require variation of cellular fitness of knockdowns to identify pathways and proteins involved in resistance or sensitivity to a stress condition. Therefore, to carry out a growth-based screen, a slightly toxic concentration of compound must be used [17]. Previous studies showed that cells from the hematopoietic lineage are more sensitive to the class of compounds that include PF846 and PF8503 (Fig 1A) [9]. In these experiments, which used rat bone marrow, PF846 and PF8503 had similar toxicity profiles [9]. The CRISPRi screen library was originally developed and validated in K562 cells [15], immortalized leukemia cells that can be induced to develop characteristics similar to early-stage erythrocytes, granulocytes and monocytes [18]. Given their close relationship to the hematopoietic lineage, we therefore used K562 cells to implement the CRISPRi screen. We first compared K562 cellular viability as a function of PF8503 and PF846 concentrations, and found PF8503 had a slightly higher negative impact on cellular metabolic activity and was slightly more toxic than PF846 (S1 Fig). We determined the best concentration of PF8503 to use in the CRISPRi screen to be 7.5 μM, which decreases cell viability 30–40% (S1 Fig).
Before using PF8503 in a CRISPRi screen, we first used ribosome profiling to compare proteins targeted by PF8503 and the previously described compound PF846 in liver-derived Huh-7 cells, which endogenously produce PCSK9 [10]. We first treated Huh-7 cells with increasing concentrations of PF8503 to identify its IC50 value with respect to inhibiting PCSK9 production, which was similar to that of PF846, or ~0.4 μM (Fig 1B). We then treated Huh-7 cells with 1.5 μM PF8503, which corresponds to ~70% of the maximal inhibition of PCSK9 production, or 0.5% DMSO (vehicle control) for 1 hour prior to isolating ribosome protected fragments for ribosome profiling library generation. The same experiment and pipeline was used in parallel with PF846 in order to compare our results to previously published data [10]. After using the bioinformatic pipeline described previously [9] (S2A Fig), we found that PF8503 affected the translation of 46 mRNAs, whereas PF846 affected translation of 24 mRNAs after 1 hr of treatment. As expected PCSK9 was affected by both compounds with a log2 fold change of -1.5 and -2 for PF8503 and PF846, respectively (Fig 1C and S1 Table).
The percentage of mRNAs affected by PF8503 (0.50%) is comparable to that for PF846 (0.27%), (S3A Fig). The protein targets of PF8503 overlap with those of PF846 (10 of 24 proteins, Fig 1C), but PF8503 and PF846 inhibit the translation of 36 and 14 distinct mRNAs, respectively (S3B and S3C Fig). Notably the potency of PF8503 stalling on many of the common targets is higher than that of PF846 (Fig 1C and S3 Fig), as assessed by differential readcounts 3’ of the stall sites (defined by DMax) (S2 Fig). Furthermore, some of the additional proteins stalled in PF8503-treated cells are slightly impacted in the PF846-treated cells, but to a lower extent that did not allow these to pass the statistical filters (S3C Fig and S1 Table). We found that the stall sites on a given mRNA occurred at identical or nearly the same codons of the transcript when comparing ribosomal footprints from the PF8503- and PF846-treated cells. As observed previously, stalling occurs near the N-terminus of the protein, although not exclusively (Fig 1D and S4 Fig).
We also compared our present PF846 results with the 21 PF846 targets previously identified in Huh-7 cells (S3D Fig and S2 Table) and observed that changes in the read count 3’ of the stall site defined by DMax were highly correlated between experiments (Pearson R = 0.82), showing the high reproducibility of our method of analysis (S2 Fig). However, we found that the main stall site position seen in the ribosome profiling read-counts varied slightly when comparing the present PF846 treatment and the previously published PF846 data [10], and when comparing PF8503 and PF846 stall sites on common targets (S4 Fig, S1 Table and S2 Table). This could be due to technical variability in the preparation of the ribosome profiling libraries, such as the efficiency of RNAse I digestion. Alternatively or in addition, this variability may reflect that compound-induced inhibition of translation is not due to a sharp stall at one codon position, but occurs due to a slow-down of translation over multiple adjacent codons [12]. Altogether, these results indicate that PF8503 is a selective inhibitor of translation and shares a similar mechanism of action to that of PF846 in stalling protein synthesis.
To identify pathways involved in the cellular response to high concentrations of PF8503, we used a genome-wide CRISPRi screen with an established whole-genome library of sgRNAs [17]. Human K562 cells constitutively expressing dCas9-KRAB-BFP were cultured with 7.5 μM PF8503 or 0.5% DMSO as a control. We used deep sequencing of the genomically-expressed sgRNAs in the cell population at the start and end of the experiment to determine their enrichment or depletion [17]. Briefly, sgRNA enrichments in the DMSO-treated control and PF8503-treated populations correspond to growth phenotypes termed Gamma and Tau, respectively, and quantify the impact of sgRNA expression on cell fitness independent of treatment. The difference in sgRNA enrichment between the control and PF8503-treated population corresponds to the impact of the compound on cell fitness independent of the impact of sgRNAs on cell growth, a parameter termed Rho (Fig 2A).
Genes from the PF8503 CRISPRi screen were filtered based on their compound-specific phenotypes (Rho > |0.1|) and Mann-Whitney adjusted p-value (<0.000001) leading to a total of 452 genes impacting cell fitness (Fig 2A). These genes were then assessed for pathway enrichment using the STRING database [19]. This analysis revealed a clear distinction between pathways protecting or sensitizing the cell to PF8503 toxicity (Fig 2B and S5 Fig). For example, proteins whose expression protects cells from PF8503 toxicity are highly enriched in mRNA synthesis and export pathways (transcription regulation, splicing, EJC-TREX complex) and cell cycle, apoptosis and DNA repair pathways (Fig 2B). By contrast, pathways sensitizing cells to PF8503 are concentrated in the mitochondrion, ribosome biogenesis and translation (Fig 2B and 2C). A third category of proteins are protective in the DMSO control (Gamma) but sensitizing in the PF8503-treated samples (Tau) (Fig 2D), including proteins involved in tRNA wobble nucleotide U34 modification [20], 60S ribosomal subunit maturation (ZNF622) and mRNA turnover (CNOT10, CNOT11).
Notably, proteins known to be involved in rescuing stalled ribosomes also affect cell viability in the presence of PF8503, suggesting that ribosome quality control pathways are triggered upon compound-induced stalling (Fig 2B and 2C). No-Go Decay proteins (NGD) Pelota (PELO) and HBS1L, involved in recognizing stalled ribosomes, enhance PF8503-induced cell toxicity, with PELO showing the highest sensitizing phenotype among these hits (Rho = 0.43). Other proteins that sensitize cells to PF8503 include Ribosome Quality Control (RQC) proteins NEMF and LTN1, involved in recycling the 60S subunit of the ribosome, and CNOT proteins involved in mRNA degradation (Fig 2). The sgRNAs most potent in decreasing cell fitness target two of the three subunits that constitute the activating-signal co-integrator complex (ASC-1), subunits ASCC2 and ASCC3. ASCC3 has been implicated in ribosome quality control, by aiding in resolving stalled ribosomes on poly-A sequences [21].
To further explore the connection between ribosome quality control pathways and PF8503-induced cell toxicity, we first validated the effects of the genes identified in the CRISPRi screen. We constructed CRISPRi K562_dCas9-KRAB cell lines expressing the most active sgRNAs from the CRISPRi screen and tested the effect of PF8503 on these cells using a similar treatment protocol for 7 days (S6A Fig). Knockdown efficiency was confirmed by RT-qPCR and Western blot analysis (S6B and S6C Fig). The individual knockdowns were highly correlated with the Rho phenotypes in the CRISPRi screen (Fig 3A, R2~0.97), allowing us to use the individual CRISPRi cell lines to evaluate the effects of these proteins on cell fitness.
The observation that knockdown of ASCC2 or ASCC3 is highly toxic in the presence of PF8503 suggests a major role of the ASC-1 complex in cell survival upon PF8503 induced translational stress. Induction of apoptosis has been observed in conditions of proteotoxic stress, i.e. downstream of the integrated stress response [22]. We therefore assessed the impact of decreased ASCC2 or ASCC3 expression in the presence of PF8503 by testing for activation of executioner caspases 3 and/or 7 upon long term compound treatment (Fig 3B). Whereas cells transfected with a non-targeting scrambled sgRNA and treated with PF8503 showed no significant induction of caspase 3/7 activation (apoptotic index ~0.1), knockdown of ASCC3 increased the apoptotic index to greater than 1 in the presence of PF8503 but not in DMSO-treated controls. ASCC2 knockdown also elicited a higher apoptotic index in the presence of PF8503 (Fig 3B). By contrast, in a survey of genes identified by the CRISPRi screen, no other knockdown cell lines showed an appreciable induction of apoptosis (S7 Fig).
Although ASCC2 and ASCC3 contribute to cell viability in the presence of PF8503, their role in rescuing stalled translation complexes of the type generated by PF8503 is not established [21]. Furthermore, these proteins in the ASC-1 complex have known roles in the nucleus as transcriptional activators and in the alkylated DNA damage response [23,24]. If these proteins have a direct role in the recognition or rescue of stalled ribosomes, they should interact in the cytoplasm. Immunoprecipitation of ASCC3 in the cytoplasmic fraction of HEK293T cells showed that ASCC2 and ASCC3 interact in the cytoplasm, independently of PF8503 treatment (Fig 3C), although the presence of ASCC2 in the non-bound fraction indicates that not all of ASCC2 is associated with ASCC3. None of three ribosomal proteins tested (RPL27 or RPS3, RPS19) was found in these immunoprecipitations (Fig 3C and S8 Fig). Interestingly, ASCC2 and ASCC3 can be found to co-fractionate with ribosomes in a sucrose cushion of K562 cellular extracts (Fig 3D). In high-salt sucrose cushions, ASCC2 remains bound to the ribosome, whereas ASCC3 levels are greatly reduced, suggesting that the interaction of ASCC3 with the ribosome is indirect and/or unstable. The third member of the ASC-1 complex, ASCC1, also fractionates with the 80S ribosome (Fig 3D), in both low- and high-salt conditions.
Since HBS1L is involved in NGD pathways, and ASCC2 and ASCC3 bind the ribosome and could also contribute to ribosome quality control, we wondered whether their knockdown would impact translation in the presence of PF8503. Using metabolic labelling with the methionine homologue L-AHA, we measured global translation in the presence and absence of PF8503. PF8503 did not decrease global translation in control cells with a scrambled sgRNA, whereas translation was significantly inhibited with non-specific translation inhibitor cycloheximide (Fig 4A). Furthermore, none of the knockdowns of HBS1L, ASCC2, or ASCC3 affected global translation compared to the control cell line, either in the presence or absence of PF8503 (Fig 4B). The ability of PF8503 to induce selective stalling was also unaffected by the knockdowns of HBS1L, ASCC2, or ASCC3. Using a reporter mRNA with codons 1–35 of PCSK9 encoded at the N-terminus of Renilla luciferase, the IC50 for PF8503 was found to be ~0.3 μM for control cells and all three knockdown cell lines (Control, 0.39 +/- 0.02 μM; ASCC2 KD, 0.31 +/- 0.04 μM; ASCC3 KD, 0.26 +/- 0.06 μM; HBS1L KD, 0.39 +/- 0.03 μM; Fig 4D). These results indicate that the effects of knocking down HBS1L, ASCC2, and ASCC3 on cell fitness in the presence of PF8503 is not due to generally lower rates of translation, or to a direct role for these proteins in PF8503-induced stalling.
In order to check whether the effects of ASCC2 and ASCC3 on cell fitness during PF8503 treatment are interdependent, we constructed a cell line in which both ASCC2 and ASCC3 were knocked down using dual sgRNAs, as described in previous Perturb-seq experiments [25] (Fig 5A and S9A Fig). We also generated cell lines in which both ASCC3 and NEMF or ASCC3 and HBS1L were knocked down, to check for genetic interactions with the RQC or NGD pathways, as hypothesized by a role of ASCC3 in a RQC-trigger (RQT) complex [21]. In these experiments, we noted that knockdown of ASCC3 in the context of using dual sgRNAs was not as efficient as the case with single sgRNAs (S9B and S9C Fig). We therefore generated a dual-sgRNA cell line with a scrambled sgRNA and ASCC3 sgRNA to serve as a control for the double-knockdown cell lines. We then used competitive growth assays as in the CRISPRi validation experiments to determine the effects of combined knockdowns in the presence of PF8503 (S9A Fig).
Interestingly, simultaneous knockdown of both ASCC2 and ASCC3 recovered wild-type fitness in the presence of PF8503 (Fig 5B). This negative epistasis provides strong genetic support that ASCC2 and ASCC3 are part of the same pathway in the context of cell response to PF8503-induced translational stress. On the other hand, the fitness of the NEMF-ASCC3 double knockdown cells is that expected based on simple addition of each protein’s individual phenotype in the presence of PF8503. This does not fully rule out a role of ASCC3 in activating RQC but suggests that these proteins can act in independent pathways. Interestingly, the double knockdown of HBS1L and ASCC3 also showed a negative epistasis, with a phenotype of the double knockdown (Rho~0.3) similar to the phenotype of the single knockdown for HBS1L (Rho~0.2) (Fig 5B). In separate experiments in which lentiviral vectors expressing sgRNAs for HBS1L and either ASCC2 or ASCC3 were introduced into cells sequentially, we also observed a strong negative epistasis (S10 Fig). Importantly, PF8503 retained its ability to stall PCSK9 reporters in the ASCC2-ASCC3 and HBS1L-ASCC3 double knock-down cells (S11 Fig). Taken together, these results suggest that ASCC3 and ASCC2 act in an overlapping pathway with the NGD pathway, which requires HBS1L.
In order to ascertain whether the genes identified by the PF8503 CRISPRi screen are specific to selective translational stalling, we compared the PF8503 screen to a CRISPRi screen carried out with the general translation inhibitor HHT, a non-specific translation inhibitor that stalls translation immediately after initiation and early in elongation [26,27][28]. The HHT CRISPRi screen used the same engineered K562_dCas9-KRAB cells, with HHT added at its LD50 value (100 nM). Overall, the HHT-dependent phenotypes for all genes did not correlate well with the PF8503-dependent phenotypes, using the stringent filters applied for the PF8503 CRISPRi screen (RhoPF8503 vs. RhoHHT, Pearson R = 0.11) (Fig 6A). By contrast, the growth phenotypes due to the inherent effects of knockdowns for both CRISPRi screens (GammaPF8503 vs. GammaHHT and TauPF8503 vs. TauHHT) were well correlated (Pearson R = 0.87 and 0.78, respectively, S12A and S12B Fig). In order to compare the PF8503 and HHT effects, we adjusted the p-value cutoff for compound-dependent phenotypes with RhoHHT > |0.1| to identify a similar number of genes in the HHT CRISPRi screen (496) when compared to the PF8503 screen (452) (S12 Fig). This allowed us to identify pathways shared between PF8503 and HHT treatments, as well as distinct pathways.
Notably, although only 126 genes were shared between the two screens using the less-stringent cutoff (Fig 6B), a KEGG pathway enrichment analysis performed using the STRING database [19] identified a significant enrichment of the same pathways in the HHT and PF8503 screens, including ribosome biogenesis, the spliceosome and RNA transport (Fig 6C). Proteins with functions related to mitosis, apoptosis, cell cycle, and response to DNA damage also significantly impacted cell fitness in the presence of either of the compounds (S3 Table and S4 Table). For some of the proteins with shared effects on cell fitness, we compared cell growth in the presence of PF8503, HHT, or PF846 (Fig 6D). Similar effects were observed with all 3 compounds suggesting that these proteins may play a more general role in stress response to translational inhibitors. For example, the two main proteins of the NGD pathway (PELO and HBS1L) and of the RQC pathway (NEMF and LNT1) strongly sensitize cells in the presence of either PF8503 or HHT, suggesting that NGD and RQC could be involved in rescuing stalled or paused ribosomes more generally. By contrast, HHT affected pathways associated with the proteasome and ubiquitin-mediated proteolysis, whereas these did not have a PF8503-dependent phenotype, possibly due to its lower toxicity (S3 Table and S4 Table). Interestingly, aminoacyl-tRNA biosynthesis pathways were only enriched in the PF8503 CRISPRi screen (Fig 6C). Most of these genes have a Rho phenotype close to 0 in the HHT screen. We also identified that knockdown of two factors involved in late stages of 60S ribosomal subunit assembly, EIF6 and ZNF622, had a positive effect on cell fitness in PF8503-treated cells (Fig 2C and Fig 3A). By contrast, in the presence of HHT, knockdown of EIF6 negatively affected cells, and knockdown ZNF622 did not have a statistically-significant effect (Fig 6A, S12C and S12E Fig).
PF8503 belongs to a new family of compounds able to selectively inhibit the translation of target proteins by the human ribosome. Originally discovered as an orally-available small molecule inhibitor of PCSK9 production [9,10,29], this class of compound could eventually serve as as a new paradigm for designing therapeutics for “undruggable” proteins [2]. These compounds have the unique ability to bind inside the ribosome exit tunnel, allowing them to interact with the protein nascent chain and selectivally stall translation [12]. However, the molecular basis for the strength of compound-induced stalling is still unknown. Furthermore, it has not been clear how the cell may respond to ribosomes stalled by these compounds. For example, the cell may clear selectively and non-selectively stalled ribosomes using quality control pathways that remain to be determined. Depending on the strength of the stall, quality control mechanisms could rescue or arrest translation on some transcripts more efficiently than others, participating in the selectivity of these compound and impacting cellular fitness during treatment [30]. Here we used ribosome profiling to map the selectivity landscape of PF8503, allowing comparisons to the related compound PF846 [10], and CRISPRi screens to uncover pathways impacting the toxicity of this class of compound. Taken together, these results identify new connections between ribosome quality control pathways, and should inform future designs of new molecules of this class.
We originally determined the selectivity of compound PF846 (Fig 1A) using human liver-derived Huh-7 cells [10]. Although highly selective, PF846 had some toxic effects in certain cell types and a rat model [10], which would preclude its use for chronic conditions. While optimizing this class of compound, PF8503 (Fig 1A) was shown to have similar potency to PF846 in inhibiting PCSK9 production, as well as a similar cellular toxicity using rat bone marrow as a model for the most sensitive cells to these compounds [9]. Here, we confirmed that PF8503 and PF846 inhibit PCSK9 production by Huh-7 cells with nearly identical IC50 values (Fig 1B), with PF8503 being slightly more toxic when treating human K562 cells (S1 Fig). Although compounds with similar potency to PF8503 and PF846, but with dramatically reduced cellular toxicity, have been identified [9], it is still important to understand why compounds like PF8503 and PF846 decrease cell viability at high doses, as these pathways may arise in the development of new compounds in the future.
Interestingly, although PF8503 and PF846 have similar potency inhibiting PCSK9, they induce ribosome stalling on a surprisingly different array of off-target mRNAs, as determined by ribosome profiling. Of the 46 mRNAs subject to PF8503-induced stalling, only 12 overlap with the mRNAs affected by PF846 (Fig 1C and S3 Fig). For example, translation of CNPY4, TM4SF4, and DHFRL1 mRNAs is potently inhibited by PF8503, but these are unaffected or barely affected by PF846 (S3C Fig). PF846 potently inhibits translation of FAM13B and HSD17B11, whereas these mRNAs are unaffected by PF8503 (S3B Fig). Other mRNAs targeted by one of the compounds may also be stalled by the other, but only at much lower levels that do not lead to a statistically significant reduction in translation past the stall site (S3 Table and S4 Table, S3 Fig). These results suggest that future efforts to tune the structure of these compounds could lead to selective stalling of new targets beyond PCSK9.
To identify genetic interactions with PF8503 toxicity, we used PF8503 to carry out a whole-genome CRISPRi screen in human K562 cells. By comparing the genetic interactions with PF8503 and the more general translation inhibitor HHT, we were able to identify shared pathways that affect cell viability in the presence of these compounds, as well as ones distinct to each compound class. Although it is possible that common pathways that interact genetically with treatment with PF8503 or HHT (S3 Table and S4 Table) may reflect a general stress response to translation inhibition, PF8503 does not have an observable effect on global translation (Fig 4A and 4B). This suggests that even the low level of stalled translation induced by PF8503 is sufficient to induce a cellular response that can be detected using CRISPRi. Common pathways that decrease overall translation due to, for example, a decrease in ribosome biogenesis may lower the overall burden on cells to monitor translation defects, and enable cells to mount a more robust response to PF8503-stalled ribosomes or HHT treatment. In support of this model, we observe sgRNA enrichment (i.e. protein reduction) for a number of factors in ribosome biogenesis and translation (Fig 2 and Fig 6C). An additional explanation for the shared pathways is that HHT, although an inhibitor of late steps of translation initiation [26,27], generates a low level of stalled translation elongation complexes [28] that behave similarly to those generated by PF8503.
Intriguingly, we found that genes involved in tRNA maturation and in particular modification of tRNA nucleotide U34 at the wobble position affect cell viability only in the presence of PF8503 (Fig 2C and 2D), not HHT (S4 Table). Modifications of U34 in tRNAs have been shown to increase the rate of translation [31]. In the present experiments, knockdown of tRNA U34 modification enzymes would be predicted to slow down overall translation, providing a protective effect against PF8503-induced stalling similar to that expected from a decrease in ribosome biogenesis. The fact the tRNA synthesis genetically interacts only with PF8503 and not HHT may reflect that fact that PF8503 stalling could be more dependent on the speed of translation, whereas HHT inhibits translation primarily before elongation begins, and would not depend as much on U34 modified tRNAs. This interpretation is also consistent with the known binding sites of both drugs. Whereas HHT is bound to the ribosomal A site in the 60S subunit and inhibits the first step of elongation [26,27], PF8503 binds inside the ribosome exit tunnel [12], and likely depends on the speed of translation to stall ribosome nascent chain complexes [10]. Notably, some forms of cancer are dependent on high levels of the enzymes that modify U34 in tRNA [32], suggesting that PF8503-class compounds could be developed in the future to target this type of cancer cell addiction.
Interestingly, we found that knocking down NGD proteins PELO and HBS1L, as well as RQC proteins NEMF and LTN1, had a positive impact on cell fitness upon either PF8503 or HHT treatment (Fig 6A and 6D), suggesting that NGD and RQC are involved in resolving at least some PF8503- or HHT-stalled ribosomes. Notably, whereas the NGD and RQC pathways in yeast have been associated with mRNA turnover, treatment of cells with PF846 has been shown not to lead to decreased levels of PF846-targeted mRNAs [10,11]. The fact that ribosome quality control-related pathways sensitize cells to PF8503 at first seems surprising, since it implies that rescuing translation in response to PF8503-induced translational stalling has a negative impact on cell fitness. However, an alternative explanation is that PF8503-stalled complexes saturate the NGD or RQC pathways in the context of normal cells, preventing these pathways from carrying out their normal quality control functions. Consistent with this saturation hypothesis, the PF846 class of compounds are relatively more toxic in the hematopoietic lineage [9,10], which is more dependent on a functional NGD pathway [33]. Furthermore, ribosome profiling experiments likely miss weak translation stalling events that may occur across the human transcriptome, or those that are robust substrates for ribosome quality control pathways and are therefore not detected in the ribosome profiling analysis. It is notable that some derivatives of the PCSK9-stalling compounds can retain potency without the associated toxicity in rat bone marrow cells [9], suggesting that the NGD- and RQC-dependent cellular responses can be managed for new therapeutic targets.
Surprisingly, the genes that protect cells the most from the toxic effects of PF8503-dependent stalling encode components ASCC2 and ASCC3 of the transcriptional activating signal cointegrator complex ASC-1 [24]. Recent experiments have also identified ASCC2 and ASCC3 as integral to the cellular response to DNA alkylation damage [23]. ASCC3 is an RNA helicase and has been shown to be the human ortholog of yeast Slh1 (SLH1 gene) [21,34]. ASCC2 harbors a CUE (coupling of ubiquitin conjugation to ER degradation) domain that binds K63-linked polyubiquitin chains as part of the DNA alkylation damage response [23]. In yeast, there is emerging evidence of a role for Slh1 in RQC [21,34]. Furthermore, K63 polyubiquitination has been found to be essential for certain stalled ribosome quality control pathways [35]. These studies led to the proposal that Slh1 and Cue3, the presumed orthologue of ASCC3 and ASCC2, are components of a new complex called the RQC-trigger complex (RQT), triggered by ribosomal stalling but not necessary for RQC [21]. Notably, we do not observe a phenotype for the ubiquitin E3 ligase ZNF598 (S3 Table), which has been found to ubiquitinate the ribosome in certain contexts of ribosome stalling, i.e. on challenging sequences such as CGA codons and poly(A) stretches in yeast [21,34] and humans [30], and in the presence of colliding ribosomes [36],[30]. These observations prompted us to investigate further the function of ASCC2 and ASCC3 in cellular response to PF8503-induced selective stalling.
Although ASCC2 and ASCC3 are primarily known for their roles in the transcription activator complex ASC-1 [24] and in DNA alkylation damage repair [23], there is substantial evidence that ASCC2 and ASCC3 also act in the cytoplasm. First, the role of the ASC-1 complex in alkylated DNA damage repair is not phenocopied in the CRISPRi screen in the presence of PF8503. For example, loss of splicing factor BRR2 (SNRNP200 gene) increases sensitivity to MMS DNA damage [23]. PRP8 knockdown (PRPF8 gene) also decreases ASCC3 foci upon DNA damage [23]. In the PF8503 CRISPRi screen, BRR2 knockdown is mildly protective (Rho = 0.13) and knockdown of PRP8 has only a mildly negative effect (Rho = -0.05). Knockdown of RNF113A and UBC13 (UBE2N gene) in this pathway [23] also had mildly negative effects (Rho = -0.13 and -0.05, respectively). Finally, knockdown of the key dealkylase connected to ASCC helicase complex, ALKBH3 [23], had no effect (Rho = 0.03). Similarly, other proteins involved in alkylated DNA damage repair had no discernable phenotype, which suggests ASCC3 is unlikely linked to alkylation damage in the presence of PF8503 (MGMT, -0.02; ALKBH2, 0.00; ALKBH5, 0.00, source [37], under “DNA Repair -> DNA Damage Reversal” tab. Second, in addition to the results described above in relation to the yeast RQT system, there is substantial evidence for the cytoplasmic localization of the ASC-1 subunits. In the Human Protein Atlas [38], ASCC2 and ASCC3 are robustly cytoplasmic as well as nuclear. ASCC1 is less abundant in the cytoplasm. Furthermore, ASCC3 has been identified to be in close proximity to the translational repressor 4EHP (eIF4E2), based on proximity ligation experiments [39].
We were able to confirm a role for ASCC2 and ASCC3 in response to PF8503-induced stalling using a combination of CRISPRi knockdowns, cell-based assays and cell fractionation (Fig 3 and S6 Fig). We observed decreased cell fitness and increased induction of apoptosis in PF8503-treated cells when either ASCC2 or ASCC3 were knocked down. Although this could be attributed to nuclear signaling events based on the reported functions for the ASC-1 complex [23,24], we found all three components of ASC-1 associated with cytoplasmic ribosomes (Fig 3D), consistent with previous results connecting K63 polyubiquitination [35] and ASCC3 to translation [21], and suggesting that ASCC1 might be the human counterpart of the third protein of the RQT complex, yKR023W (Rqt4) [21]. Additionally, the fact that ASCC2 and ASCC3 co-IP in the cytoplasm (Fig 3C) is consistent with a role for ASC-1 in translation, independent of its role in DNA alkylation repair and transcription activation in the nucleus. Also supporting a role for ASC-1 in translation quality control, both ASCC2 and ASCC3 knockdowns negatively affect cell fitness of HHT-treated cells (Fig 6A and 6D, S12 Fig).
We also observe that EIF6 and ZNF622, two proteins known to be involved in late stages of ribosomal 60S subunit assembly [40], play a role in PF8503-induced stalling (Fig 2 and S6D Fig). Knockdown of either factor has a positive effect on cell fitness. Intriguingly, both EIF6 and ZNF622 can be affinity-purified with ASCC2 in a K63-linked polyubiquitin chain-dependent manner [23]. These affinity-purification experiments were conducted in the context of defining the DNA alkylation damage response. However, based on their role in PF8503-induced stalling, it is possible that EIF6 and ZNF622 function in translation quality control pathways, acting downstream of ASC-1.
We were surprised not to identify ZNF598, the ubiquitin E3 ligase that ubiquinates ribosomal proteins in a ribosome stalling-dependent manner [36,41], in the CRISPRi screen with either PF8503 or HHT. Its absence could be due to the fact that the CRISPRi screen reveals genetic and not necessarily all biochemical connections. For example, residual ZNF598 levels in the CRISPRi knockdown cells may be sufficient for ribosome quality control in the present experimental conditions. Furthermore, there may be other ubiquitin ligases that serve an overlapping role with ZNF598. Finally, ZNF598 is thought to mark ribosomes after ribosome collisions [36,41]. In the case of PF8503 and HHT, there may be no or few ribosomal collisions since most stalling events occur very near the N-terminus of the protein, i.e. the 5’ end of the open reading frame. Future experiments will be required to reconstitute the initial ubiquitination events responsible for ribosome quality control observed here.
Importantly, we were able to use double-knockdown cell lines in genetic interaction experiments to identify connections between different translation quality control pathways in human cells. We found that the HBS1L phenotype in the presence of PF8503-induced stalling is dominant over the ASCC3 phenotype (Fig 5B), suggesting that HBS1L and ASCC3 function in a common translation quality control pathway, with ASCC3 involved in steps after translation stall recognition. This result implies that the NGD pathway and ASC-1 intersect to resolve at least some stalled ribosome-nascent chain complexes. The model that ASC-1 acts downstream of HBS1L, after the stalling event, is further supported by the fact that that ASCC2 and ASCC3 knockdowns do not impact general translation or the IC50 for PF8503 (Fig 4B and 4C). Interestingly, we found all three ASC-1 subunits bound to the 80S ribosome independent of active translation (Fig 3D), implying that ASC-1 surveillance of translation may be widespread. By contrast to HBS1L and ASCC3, NEMF which is part of the RQC pathway seems to act independently of ASCC3 (Fig 5B), suggesting that PF8503-induced stalling may not be a “classic” RQC substrate like those identified previously, such as poly-lysine stretches [21,30]. Alternatively, the RQC pathway components may not be as limiting in the presence of PF8503, precluding our ability to detect a genetic interaction. Interestingly, we find that a double-knockdown of ASCC2 and ASCC3 rescues the negative effect of single knockdowns of ASCC2 or ASCC3 upon PF8503 treatment (Fig 5B). We suggest that the negative effect of individually knocking down ASCC2 or ASCC3 could result in a partially activated translation quality control pathway that leads to a build-up of stalled ribosomes that cannot be resolved, thereby leading to severe cell stress.
The genetic interactions we observe between components of the NGD and RQT, but not RQC, pathways in the presence of specific translational stalling by PF8503-related compounds (Fig 7) opens new avenues for exploring the mechanisms of translation quality control pathways in humans. Although the exact role of ASC-1 complex on the ribosome is still unclear, the fact that ASCC2 and ASCC3 strongly impact cell fitness in the presence of PF8503 suggest that ASC-1 may play an essential role in either stall recognition, ribosomal degradation or recycling, or stress signaling to the nucleus. Furthermore, it is possible that ribosome assembly factors EIF6 and ZNF622 also play a role in translation quality control pathways (Fig 7), an idea that can now be explored in depth. Combined with the new insights into changes in selectivity when comparing PF8503 and PF846, these results provide a foundation for the future development of compounds that selectively stall the translation of diverse protein targets involved in human disease.
The human chronic myeloid leukemia (CML) cell line K562, as well as a CRISPRi derivative of this cell line constitutively expressing a catalytically inactive Cas9 fused to a KRAB effector domain (dCas9-KRAB)[17] was kindly provided by the Innovative Genomics Institute, UCSF. K562 cell lines and human hepatocellular carcinoma Huh7 cells (ATCC) were cultured in RPMI 1640 medium (Life Technologies, CARLSBAD, CA, USA) supplemented with 0.2 mM L-glutamine (Glutamax, Life Technologies), 10% FBS (F4135, Sigma-Aldrich, St. Louis, MO, USA), and antibiotics (Penicillin/Streptomycin, 0.1 mg/mL, Gibco) unless otherwise stated. HEK293T cells (UC Berkeley Cell culture facility) were maintained in DMEM (Gibco) supplemented with 10% FBS (Seradigm) and antibiotics unless otherwise stated. Cells tested negative for mycoplasma infection before use.
The compounds PF-06446846 (PF846) and PF-06378503 (PF8503) were synthesized as described in [9] and provided by Pfizer. To determine the IC50 value for inhibition of PCSK9 production, 100 μL of PF846 and PF8503 dilutions in DMSO (0.5% final concentration in media) were added to an overnight culture of 3000 Huh7 cells/mL. PCSK9 production was estimated after overnight incubation by the determination of PSCK9 concentration using a solid phase sandwich ELISA (PCSK9 Quantikine ELISA Kit, RD systems).
96-well plates were seeded with 100 μL of K562 cells at 105 cells/mL and treated with different concentrations of translational inhibitor for 72 hr. For longer experiments, cells were diluted in a new plate and treated again with translational inhibitor in order to remain under the maximum cell density (<106 cells/mL). Cell viability was determined by ATP level measurements (CellTiterGlo2.0, Promega) as an indicator of cell titer and metabolic activity. To determine the level of caspase induction, caspase3/7 levels were measured from 50 μL of cell suspension and ATP levels were measured from the remaining volume (Caspase3/7Glo, Promega). The apoptosis index for each well was calculated as the ratio of Caspase and ATP luminescence signals.
Overnight cultures of 7x105 Huh7 cells in 10 cm dishes were treated with 1.5 μM PF846, 1.5 μM PF8503, or 0.5% DMSO control for 1 hr, in biological triplicate, rinsed with phosphate-buffered saline (PBS) containing cycloheximide (100 μg/mL) and triturated in lysis buffer (20 mM Tris-Cl pH 7.4, 150 mM NaCl, 5 mM MgCl2, 1 mM DTT, 100 μg/mL cycloheximide, 1% Triton X-100 and 25 U/mL DNAse I, Promega). The lysates were aliquoted, flash frozen and stored at -80°C until used for ribosome footprint library preparation.
Ribosomes and libraries were prepared as previously described [42]. In short, 300 μL of thawed lysate were digested with RNAse I (Invitrogen, AM2294) for 45 min at room temperature, monosomes were collected using a 1 M sucrose cushion and ultracentrifugation (603,000 g, 2 hr, 4°C), and total RNA was purified using miRNAEasy kit (Qiagen, 217004). RNA fragments of 26 to 34 nucleotides (corresponding to ribosome footprints) were size-selected using a denaturing urea gel. The isolated RNA was dephosphorylated using T4 polynucleotide kinase (New England Biolabs, M0201S), and Illumina-compatible polyadenylated linkers were ligated to the footprint RNAs using truncated T4 RNA ligase 2 (200 U, New England Biolabs, M0242S). First-strand cDNA was synthesized using Protoscript II (200 U, New England Biolabs, M0368L) and circularized using Circligase I (100 U, Epicentre, CL4111K). Ribosomal RNA was depleted using custom oligonucleotides [10] attached to myOne streptavidin beads (Invitrogen, 65001). The cDNA libraries were indexed and amplified using 8 to 14 cycles of semi-quantitative PCR with high fidelity Phusion polymerase (New England Biolabs, cat. no. M0530S). Ribosome footprint libraries were sequenced at the QB3 Vincent J. Coates Genomics Sequencing Laboratory, UC Berkeley on an Illumina Hiseq 2500 sequencer.
De-multiplexed reads were stripped of 3′ adapters with the FASTX toolkit (http://hannonlab.cshl.edu/fastx_toolkit/index.html) and aligned to human ribosomal RNA sequences using Bowtie [43] to remove ribosomal RNA-mapped reads. A reference transcriptome for Bowtie was built using the UCSC/Gencodev24 known coding Canonical Transcripts of the Grch38 human reference genome [44,45]. To avoid ambiguously mapped reads, only the alignments that mapped to one position of the reference transcriptome were selected for further analysis. To map the reads along the transcripts, an mRNA P site-offset was determined depending on the size of the fragment as 14 nt for fragment of 26 nucleotides or less, 15 nt for fragments of 27 to 29 nucleotides, and 16 nt for fragments of 30 nucleotides or more [46]. Codon maps were then generated for the CDS of each transcript. The data was then processed using custom python scripts and R (Supplemental document 1) as previously described with some modifications [10] (S2 Fig). In short, a DMax value was calculated for each transcript as the maximum difference between cumulative normalized reads in PF8503- or PF846-treated and untreated (DMSO control) samples. For each sample, a Z-score transformation of the DMax values was calculated using the scale function in R and the DMax positions with a Z-score greater than or equal to 2 were considered putative stalled transcripts. A differential expression analysis was then performed in R using DEseq for transcripts that had more than 30 counts mapping to CDS reads located 3’ of the DMax position, or 10 codons after the start codon for transcripts lacking a clear DMax value. Reads 10 codons before the end of the transcript were also omitted. Transcripts showing a log2 fold change and FDR adjusted p-value < 0.05 were considered significant.
To generate ribosomal footprint density plots, the number of ribosomal footprints aligning to each codon position was divided by the total number of reads aligning to the protein-coding regions, then multiplied by 100 to yield reads percentage. All read density plots represent average values for 3 biological replicates. For mRNAs affected by treatment with PF846 and/or PF8503, putative pause sites were found using R and defined as the codon position at which the average read density is at least 10 times higher than the median of the positions on the transcript with more than 0 reads. Footprint densities were then analyzed manually to identify the main stall sites.
Illumina sequencing data and processed ribosome footprints have been deposited at the NCBI Gene Omnibus Database under accession number GSE121981.
The “Top5” and “Supp5” plasmid sub-pools of the whole-genome human CRISPRi sgRNA library hCRISPRi-v2 [17] were first pooled to obtain the full 10 sgRNA/gene library and then amplified. Amplification was performed by electroporation into Endura electrocompetent cells (Lucigen, 60242) using the manufacturer’s instructions and amplified in 1 L LB broth with 100 μg/mL ampicillin. The transformation efficiency assured coverage of at least 400X the size of the library. Even coverage of the libraries was confirmed by PCR amplification and deep sequencing.
PF8503 and HHT screens were conducted separately. For each screen, the sgRNA libraries were transduced into K562 cells expressing dCas9-KRAB-BFP as previously described [15]. In short, lentiviral vectors were produced in HEK293T cells by transfection of the hCRISPRi-v2 plasmid library along with packaging (pCMVdeltaR8.91) and envelope plasmids (pMD2.G) using TransIT-293 Transfection Reagent (Mirius, MIR 2704). Virus-containing media was harvested after 72 hours, supplemented with polybrene (8 μg/mL), and applied to 250x106 K562_dCas9-KRAB cells. Cells were spun in 6-well plates (2 hr, 200 g, 33°C) to enhance infection efficiency. Cells were then suspended in fresh culture medium. 60% (for PF8503) or 75% (for HHT) of the cells were infected determined using flow cytometry detection of BFP expression at 2 days post-infection. Cells were then treated with puromycin (0.75 ng/mL, Gibco), for 2 days, at which point the population of BFP-expressing cells reached 80–90% (for both screens). Cells were split into two replicates and the screen was started after one day of recovery in puromycin-free media (T0).
The K562_dCas9-KRAB cells expressing the genome-wide sgRNA library were cultivated in 3 L spinner flasks. For the PF8503 screen, cells were treated 3 times with 7.5 μM PF8503 or vehicle (DMSO), on days 0, 4, and 7. For the HHT screen, cells were treated with 100 nM HHT on day 8. In each screen, the total amount of cells was kept above 250x106, and cell density between 0.25x106 cells/mL and 1x106 cells/mL for coverage of at least 1000X the sgRNA library throughout the screen. Cell size, number, and fraction of sgRNA-containing cells were monitored using a Scepter cell counter and by flow cytometry during the screen. Cells were harvested for library sequencing on the initial cell population (T0) and after 11 days (for PF8503 and corresponding vehicle) or 15 days (for HHT and vehicle).
The CRISPRi screen libraries were prepared from the cells as previously described [15]. Briefly, genomic DNA was extracted from about 300x106 harvested cells using Nucleospin XL Blood kit (Machery-Nagel, 740950) and digested overnight at 4°C with SbfI-HF restriction enzyme (New England Biolabs, R3642S). Fragments of about 500 bp, corresponding to the sgRNA cassette, were size selected on a large scale 0.8% Agarose gel and purified using a Qiagen gel extraction kit (28706X 4). The libraries were then indexed and amplified by 23 cycles semi-quantitative PCR using HF-Phusion polymerase (New England Biolabs, M0530S). The libraries were multiplexed and sequenced at the QB3 Vincent J. Coates Genomics Sequencing Laboratory, UC Berkeley, or the Center for Advanced Technology, UCSF, on an HiSeq4000 Illumina sequencer using custom primers as previously described [15]. Computational analysis of the CRISPRi screens was carried out as previously described [15], using the pipeline available on github (https://github.com/mhorlbeck/ScreenProcessing), with the hCRISPRi-v2 library tables/alignment indices. Briefly, the quantified genomically-integrated sgRNAs were used to calculate the effect of each sgRNA on growth without drug (Gamma; T0 vs vehicle), with drug (Tau; T0 vs drug), and the effect of drug treatment only (Rho; vehicle vs drug). Gene-level scores in each condition were obtained by averaging the top 3 sgRNAs for each gene (as ranked by absolute value) and by Mann-Whitney p-value of all 10 sgRNAs/gene compared to non-targeting control sgRNAs.
For each gene candidate to be tested for validation of the CRISPRi results, two sgRNA protospacers giving strong phenotypes in the CRISPRi screen were selected (S5 Table) and cloned into a pSLQ1371-GFP or -BFP vector using restriction sites BstXI and BlpI as previously described [17]. Lentiviral vectors were produced in HEK293T cells by transfection of the pSLQ1371-GFP or -BFP sgRNA lentiviral expression vectors along with packaging (pCMVdeltaR8.91) and envelope plasmids (pMD2.G) using TransIT-293 Transfection Reagent (Mirius, MIR 2704). K562_dCas-KRAB cells were then infected with the resulting lentiviruses in 6-well plates, and puromycin selected (1 μg/mL) two days after infection. The MOI was determined by flow cytometry using GFP or BFP fluorescence before puromycin selection. Knockdown of PELO levels was too toxic to obtain stable cell lines. We therefore used HBS1L knock-down cell lines to explore disruption of NGD.
The double knockdown cell lines were obtained in a similar manner, either by carrying out two sequential lentiviral infections with lentiviruses each expressing one sgRNA. Alternatively, we prepare double knock-down cell lines by a single infection of a lentivirus containing two sgRNA expression cassettes. These plasmids were obtained by ligating a synthetic DNA sequence containing the human U6 (hU6) promoter followed by the sgRNA sequence targeting ASCC3 and a different constant region cr2 [25] downstream of a murine U6 (mU6) promoter and sgRNA expression cassette, in the pSLQ1371-GFP or -BFP vector.
Cell pellets were lysed using an NP40 lysis buffer (50 mM Hepes pH = 7.4, 150 mM KCl, 2 mM EDTA, 0.5% NP40, 0.5 mM DTT) with complete protease inhibitor (Sigma-Aldrich, 4693116001). Proteins were separated on 4–12% Bis-Tris gels (NuPage, Invitrogen,NP0322BOX) and transferred to a nitrocellulose membrane (Ultracruz, sc-3718). Membranes were blocked with 5% skim milk and incubated with primary antibodies overnight. The antibodies used in these experiments are given in S6 Table. The secondary antibodies used were conjugated with horseradish peroxide and detected with ECL substrate (PerkinElmer, NEL103E001EA). Western blot films were developed (Optimax, PROTEC), and imaged using Image Studio Lite Software (LI-COR Biosciences, Lincoln, NE, USA).
For the quantification of gene expression of knockdown cell lines, as compared to negative control cell lines with a scrambled sgRNA, total RNA was isolated from cells using an RNA Miniprep kit (Zymo Research, R2060) and analyzed using a SYBR Green 1-step reverse transcriptase real time PCR kit (RNA-to-Ct, Applied Biosystems, 4389986) on Biorad Quantstudio 3 according to the manufacturer's instructions. Target-specific primers (S7 Table) were designed using Primer-BLAST [47] (https://www.ncbi.nlm.nih.gov/tools/primer-blast/) or the Primer3 web tool ([48], http://primer3.ut.ee/). The cycle threshold (Ct) values were determined using ThermoFisher Connect (https://www.thermofisher.com/us/en/home/cloud.html). Levels of each target RNA were determined relative to the housekeeping gene PPIA mRNA levels as internal control for each cell line and the percent of inhibition was calculated using the ΔΔCt method.
For competitive growth assays, in order to simulate the conditions of the screen, cell lines expressing the scrambled sgRNA were mixed with each target cell line expressing an sgRNA of interest to be tested at a ratio of ~1:1. Each cell line also produced either the GFP or the BFP reporter individually, which allowed for the determination of the cell proportion of each cell population over time using flow cytometry. Cell populations were treated with PF8503 (7.5 μM), PF846 (7.5 μM), or HHT (20 nM), near the LD50 for each compound. Growth of the population was also determined by cell counting in order to calculate the enrichment phenotypes as was done in the CRISPRi screen.
ASCC3 immunoprecipitation was carried out using HEK293T cells. 50–60% confluent HEK293T cells in a 10 cm dish were treated overnight with 0.05% DMSO or 7.5 μM PF8503. Cells were washed 3 times with HBSS buffer (Gibco, 14025092) and detached by scraping in 1 mL HBSS. Cell pellets were collected after centrifugation and resuspended in 200 μL IP lysis buffer (25 mM Tris-HCl pH 7.4, 150 mM KCl, 0.5 mM DTT, 1X Protease inhibitor, Sigma-Aldrich 4693116001, 4 mM EDTA, 0.5% NP40, 100 U/mL RNAsin, Promega N2511), incubated 10 min on ice, and centrifuged at 4°C for 10 min (14000xg) to remove cell debris. Total cell lysates were incubated with 1 μg of ASCC3 antibody (Bethyl laboratories, A304-015A) for 1 hr at 4°C on a rotator. Then 40 μL of cell lysate was stored at -20°C as the input fraction of Western blot gels. The remaining 160 μL was mixed with 50 μL protein G beads that had been washed and resuspended in 50 μL of NT2 buffer (25 mM Tris-HCl pH 7.4, 150 mM NaCl, 1 mM MgCl2, 1X protease inhibitor, Sigma-Aldrich 4693116001, 0.05% NP40, and 100U/mL RNAsin, Promega N2511) and the samples were incubated overnight at 4°C on a rotator. Protein G beads were washed 4 times with NT2 buffer. Proteins bound to the beads were eluted by denaturation in 1X NuPage buffer at 95°C for 10 min and protein content of each fraction were analyzed by Western blot, along with the loading control.
Actively-growing K562 cells were washed 3 times with warm PBS and suspended in Methionine free RPMI (Gibco, A1451701), supplemented with 10% dialyzed FBS at equal cell concentrations (~0.5x106 cells/mL). Cells were distributed in 6-well plates and incubated for 2 hr in a cell culture incubator at 37°C. Then, 7.5 μM PF8503 or 0.5% DMSO was added to the medium immediately before addition of L-AzidoHomoAlanine (L-AHA) at 25 μM final concentration. Cells were incubated 30 min, washed 3 times with PBS (centrifugation at 400g, 5 min, 4°C) and cell pellets were then flash frozen in liquid nitrogen. Cell pellets were incubated 30 min on ice in 200 μL lysis buffer (1%SDS, 50 mM Tris HCl, pH 8, protease inhibitor 1X, Sigma-Aldrich, 4693116001, and 150 U/mL Benzonase nuclease, Sigma, E1014) and vortexed 5 min. Cell lysates were collected after centrifugation of the cell debris at 18000 g for 10 min at 4°C, then incubated 2 hr with 50 μM of IRDye 800CW DBCO (LI-COR, 929–50000), to label L-AHA with fluorescent dye using a copper-independent Click reaction [49]. Excess dye was removed using a desalting column (Zeba Spin Desalting Columns, 40K MWCO, 0.5 mL, 87767). Total protein levels were determined using Bradford protein assay (Biorad, 5000006) for each sample. Total proteins were separated by electrophoresis on 12% or 4–12% Bis-Tris gel (NuPage, Invitrogen, MES-SDS buffer, NP0002) and the gels were washed with PBS. Newly synthesized proteins were labelled with IRDye800 and were imaged in the gel on a LI-COR Infrared imager (Odyssey, 800 nm wavelength channel). After imaging of labelled proteins, total proteins were stained with Coomassie (SafeStain Blue, Invitrogen, LC6060). Images were analyzed using LI-COR Image studio LiTe software (LI-COR Biosciences, Lincoln, NE, USA).
All mRNA reporters were produced from plasmids linearized using PmeI (New England Biolabs, R0560S), in which the mRNA was encoded downstream of a T7 RNA polymerase promoter. Reporters used for in-cell assays were transcribed, capped and tailed using HiScribe T7 ARCA mRNA Kit (with tailing, New England Biolabs, E2060S) following the manufacturer’s instructions. Reporters used for in vitro translation assays and sucrose cushions were transcribed using HiScribe T7 Quick yield (New England Biolabs, E2050S). All mRNA reporters were purified by LiCl precipitation (2.5 M final concentration), washed with 70% ethanol, and quantified by spectrometry (NanoDrop ND-1000 UV-Vis Spectrophotometer).
K562_dCas-KRAB cell lines expressing different sgRNAs were grown in 3 L spinner flasks and collected when the density was ~0.5x106 cells/mL. Cells were pelleted at ~100 g for 4 min at 4°C, washed in isotonic buffer (20 mM MOPS/KOH pH 7.2, 130 mM NaCl, 5 mM KCl, 7.5 mM magnesium acetate, MgOAc, 5 mM sucrose), and suspended in 1–1.5 volumes of cold hypotonic buffer (10 mM MOPS/KOH pH7.2, 10 mM KCl, 1.5 mM MgOAc, 2 mM DTT). Cells were incubated in hypotonic buffer for 10 min on ice and shredded 15X through a 26G needle. Cell extracts were cleared by a 4 min centrifugation at 10,000 g, 4°C and cell extract aliquots were flash frozen and stored at -80°C until used.
In vitro translation reactions (IVT) contained 50% thawed cell extract in energy and salt mix buffer at the final concentrations: 20 mM MOPS/KOH pH 7.5, 0.2 mM ATP, 0.05 mM GTP, 15 mM creatine phosphate, 0.1 mg/mL creatine phosphokinase, 1 mM DTT, 1.5 mM MgOAc, and 200 mM KCl, 1x amino acid mix (88x amino acid mix was prepared with 16.7x MEM essential amino acids, 33.3x MEM nonessential amino acids, and 6.7 mM glutamine, Life Technologies, 11140050, 11140050, and 25030024), 40 ng/μL reporter mRNA, 0.05% DMSO or various concentrations of the specific translation inhibitor. Total reaction volumes of 10 μL to 100 μL were incubated in a thermocycler at 30°C for 15 to 45 min. Nanoluciferase signals were determined in each well by mixing 8 μL of the IVT reaction with 50 μL of luciferase substrate (NanoGlo, Promega, N1110), and measured using a Veritas microplate luminometer (Turner Biosystems).
To test the effect of salt concentration on protein adsorption to ribosomes in sucrose cushion experiments, potassium acetate (KOAc) was added after the IVT reaction to a final concentration of 400 mM (high salt) or 200 mM (low salt) prior to layering on the sucrose cushion in MLA-130 tubes (Beckman Coulter, 343778). 100 μL of the adjusted IVT were layered on top of a 600 μL sucrose cushion (1 M sucrose, 1 mM DTT, 20 U/mL Superasin, Invitrogen, cat. no. AM2694, 4 mM HEPES 7.4, 30 mM KOAc, 1 mM MgCl2) and ultracentrifuged at 603,000 g for 50 min at 4°C (Optima MAX Ultracentrifuge, Beckman Coulter). To identify proteins that co-isolate with the ribosome, the ribosome pellet was resuspended in 1X Nupage sample buffer (NuPage, NP0008) and protein content of this fraction was determined by Western blot.
The mRNA reporters used in cell reporter assay contained the 5’-untranslated region (5’-UTR) of β-globin (HBB) followed by either the stalling sequence of PCSK9 (codons for amino acids 1–35), the p2A sequence [50] and Renilla luciferase (PCSK9(1–35)-Rluc), or by just the Firefly luciferase sequence (Fluc). Briefly, 125 μL of K562 cells were seeded in 96-well plates at ~0.4x106 cells/mL and incubated overnight (37°C, 5% CO2). The test reporter (PCSK9(1–35)-Rluc) and the control reporter (Fluc) were co-transfected (0.05 μg per mRNA per well) in K562 cell lines using JetMessenger mRNA transfection kit (Polyplus, 150–01) according to the manufacturer's instructions. Serial dilutions of 200x PF8503 were prepared in DMSO and added to the cells immediately after transfection (0.5% DMSO final per well). Rluc and Fluc luminescence signals were measured 7–8 hr after transfection with Dual-Glo (Promega, E2920) on a Veritas microplate luminometer (Turner Biosystems). The normalized signal for each well (Rluc/Fluc) was calculated and normalized to the signal of non treated (DMSO only) wells. IC50 values were calculated using GraphPad Prism version 7.00 for Mac (GraphPad Software, La Jolla California USA, www.graphpad.com).
|
10.1371/journal.pgen.1003941 | Activin Signaling Targeted by Insulin/dFOXO Regulates Aging and Muscle Proteostasis in Drosophila | Reduced insulin/IGF signaling increases lifespan in many animals. To understand how insulin/IGF mediates lifespan in Drosophila, we performed chromatin immunoprecipitation-sequencing analysis with the insulin/IGF regulated transcription factor dFOXO in long-lived insulin/IGF signaling genotypes. Dawdle, an Activin ligand, is bound and repressed by dFOXO when reduced insulin/IGF extends lifespan. Reduced Activin signaling improves performance and protein homeostasis in muscles of aged flies. Activin signaling through the Smad binding element inhibits the transcription of Autophagy-specific gene 8a (Atg8a) within muscle, a factor controlling the rate of autophagy. Expression of Atg8a within muscle is sufficient to increase lifespan. These data reveal how insulin signaling can regulate aging through control of Activin signaling that in turn controls autophagy, representing a potentially conserved molecular basis for longevity assurance. While reduced Activin within muscle autonomously retards functional aging of this tissue, these effects in muscle also reduce secretion of insulin-like peptides at a distance from the brain. Reduced insulin secretion from the brain may subsequently reinforce longevity assurance through decreased systemic insulin/IGF signaling.
| It is widely known that reduced insulin/IGF signaling slows aging in many contexts. This process requires the forkhead transcription factor (FOXO). FOXO modulates the expression of many genes, and the list of those associated with slow aging is impressive. But there are few data indicating the mechanisms or genes through which FOXO actually slows aging. Here, we identify a novel FOXO target, dawdle, the Activin-like ligand in fruit flies. We show that down-regulation of Activin signaling in muscle, but not in adipose tissue, leads to extended lifespan. In part it does so when it alleviates the negative transcriptional repression of its Smox transcription factor (a Smad transcription factor) upon a keystone autophagy gene, Atg8a. This double signaling cascade autonomously improves muscle performance (measured at cellular and functional levels) and nonautonomously extends lifespan as it reduces the secretion of insulin peptides from the brain. The work develops the emerging model for interacting autonomous-nonautonomous roles of insulin/IGF signaling as a systems integrative mechanism of aging control.
| Reduced insulin/IGF-1 signaling increases the lifespan of nematodes, flies and rodents [1], [2]. In Caenorhabditis elegans, mutants in insulin-like receptor daf-2 live twice as long as wild type [3], [4]. Mutation of insulin receptor InR and insulin receptor substrate (chico) increase adult lifespan in the fruit fly Drosophila melanogaster [5], [6]. It is reported that mice with mutation at the IGF-1 receptor (Igf1r) extend lifespan [7], as do mutants of the insulin receptor substrate (Irs2) [8] and of the insulin receptor within adipose tissues [9].
Genetic evidence places the forkhead transcription factor FOXO as the downstream effector of insulin/IGF-1 signaling [3], [10], [11], [12]. Activated insulin/IGF-1 signaling enhances the phosphorylation of FOXO, which is sequestered in the cytoplasm. Conversely, reduced insulin results in FOXO nuclear translocation, which thus promotes or represses the transcription of FOXO target genes [11] (Figure S1A). In C. elegans lifespan extension of daf-2 and age-1(PI3 kinase) mutants requires daf-16, a FOXO homolog in worms [3]. Recent work likewise shows that FOXO is required for insulin-mediated lifespan extension in Drosophila [13], [14]. FOXO also appears to function in human aging where independent studies found polymorphisms of FoxO3A to associate with exceptional longevity [15], [16]. Insulin signaling through its control of FOXO is a potentially conserved system to regulate aging but despite this emerging consensus, the proximal targets of insulin/FOXO signaling that orchestrate these mechanisms of longevity assurance are essentially unknown.
One aspect that is clear is that insulin/FOXO signaling operates both nonautonomously and autonomously to control Drosophila aging. Systemically reducing insulin signaling by mutations of the insulin receptor (InR) and insulin receptor substrate (chico) slows the decline in cardiac performance of aging flies while a similar outcome is produced by overexpressing FOXO and PTEN just within cardiomyocytes [17]. Likewise, overexpressing FOXO in muscle maintains muscle protein homeostasis and delays muscle function decline with age while dFOXO expressed in muscles extends lifespan [18], as does expression of dFOXO only from fat body [19], [20]. dFOXO expressed from fat body reduces secretion of systemic insulin-like peptides (DILP2 and DILP5), which are produced predominantly in the brain. dFOXO of Drosophila fat body modulates lifespan by inducing fat body dilp6 transcription, which in turn suppresses neuronal DILP secretion [21]. These findings suggest that insulin/FOXO signaling within some organs controls both the systemic level of circulating DILPs while systemic DILPs regulate somatic maintenance of insulin sensitive tissues. Identifying the FOXO target genes and somatic maintenance pathways in such tissues will elucidate how reduced insulin/IGF-1 assures longevity.
Genome-wide studies with C. elegans have been used to probe how daf-16 controls lifespan in response to insulin signaling. Microarray analyses have identified many mRNA that are affected directly or indirectly by daf-16, and reducing some of these genes by RNA interference (RNAi) increases longevity [22], [23]. Chromatin immunoprecipitation (ChIP) and DNA adenine methyltransferase identification (DamID) have been used to identify the direct targets of DAF-16 to clarify which pathways are proximally responsible for the impact of daf-16 upon aging [24], [25]. Oh et al. [24] thus described 103 genes to be direct targets of daf-16 in the long-lived daf-2(e1370) mutant, and three out of 33 tested target genes were found to increase lifespan when gene expressions were reduced by RNAi: lin-2, egl-10 and sca-1.
In Drosophila, work to date has identified binding targets of dFOXO primarily from wild type adults. Alic et al identified 1423 dFOXO binding sites in wildtype adult female Drosophila using a ChIP-on-chip approach [26]. Among 1755 unique genes that are less than 1 kb away from these dFOXO binding sites, about 365 genes are transcriptionally regulated by dFOXO. These targets could potentially modulate many insulin related phenotypes including growth, reproduction, and metabolism, as well as aging.
Here we aim to understand how reduced insulin/IGF-1 signaling extends Drosophila lifespan by identifying genes transcriptionally regulated by dFOXO in long-lived insulin signaling mutants. We conducted ChIP analysis with a dFOXO antibody followed by Illumina high-throughput sequencing from chico heterozygous mutants, which are long-lived and normal sized, and from adult flies with ablated insulin producing cells (IPCs), which are also long-lived [27]. dFOXO was seen to bind at promoters of 273 genes common to these genotypes, thus providing a candidate set of potential factors in control of aging. Pathways enriched within this set include those for G-proteins, Wnt and Transforming growth factor-beta (TGF-β). We subsequently focused on TGF-β signaling because dFOXO binds in the promoter and represses dawdle (daw), an Activin-like ligand in TGF-β superfamily. In genetic trials, reducing daw or its downstream transcription factor Smox increases lifespan, preserves muscle function and reduces poly-ubiquitinated protein accumulation. The muscle specific benefits of activated dFOXO are mediated through the control of autophagy by Smox, which we find to bind and transcriptionally repress Atg8a/LC3, a reported longevity assurance gene of Drosophila [28]. Expressing Atg8a in muscle was also sufficient to increase lifespan. Furthermore, reducing daw in muscle decreased DILP2 peptide secretion from the brain while peripheral insulin/IGF signaling was correspondingly reduced. Our results suggest that insulin/IGF signaling controls lifespan in part through dFOXO-mediated repression of muscle Activin signaling and its downstream functions including muscle autophagy, muscle proteostasis and subsequent remote control of systemic insulin/IGF signaling.
To understand how dFOXO extends Drosophila lifespan we sequenced promoters derived from chromatin-immunoprecipitation with antibody against dFOXO in two genotypes of long-lived flies with reduced insulin signaling. Heterozygotes of chico1 live 36% longer than co-segregating wildtype sibs [13], [29]. Unlike many mutants of the insulin-signaling pathway, chico+/− adults have normal development time, body size and fecundity. Aging is likewise retarded by partially ablating adult IPCs by inducing apoptosis with a cell specific inducible driver (Dilp2-GeneSwitch-gal4>UAS-reaper) [27]. We conducted ChIP-Seq analysis from 15-day old female adults from both genetic manipulations. This revealed dFOXO to bind at 1331 and 763 promoter regions (Figure S1B), from chico and IPCs ablated flies respectively, corresponding to 2042 and 1012 candidate genes (Figure 1A and Table S7).
We identified 273 genes common to both longevity-assurance genotypes (Figure 1A). Biological functions defined by Gene Ontology (david.abcc.ncifcrf.gov) in this overlapping set include development, growth and neuron differentiation (Figure 1B). Pathway analysis (david.abcc.ncifcrf.gov) revealed enrichment in Wnt and TGF-β signaling (Figure S1C). Corresponding to previous work, we also found significant binding of dFOXO at puckered (puc) in both longevity assurance genotypes (Table S1). In the JNK signaling pathway, puc is a negative regulator of JUN kinase basket (bsk), and mutation of puc extends Drosophila lifespan [30].
To determine how candidate dFOXO targets affect longevity we selected 23 genes for further analyses (Figures 1C–1E, Table S1) based on their placement in recognized signaling pathways or because they showed a strong dFOXO binding. dFOXO binding at the promoter of these candidates was verified by ChIP followed with gene specific qPCR. In this analysis dFOXO was significantly enriched at all candidate targets in both insulin mutants when compared to wildtype (Figures S1E–S1F).
To measure the impact of insulin/IGF-1 on candidate transcription we quantified mRNA in adults of wildtype (WT), chico null mutant (chico −/−) and chico; foxo double mutant (chico −/−; foxo −/−) (Figures 1C–1E). Transcripts of 12 genes were up-regulated in chico −/− relative to wildtype but not in chico −/−; foxo −/−, indicating that dFOXO induces these genes. The expression of seven genes was repressed in chico −/− relative to wildtype but not in chico −/−; foxo −/−, suggesting that dFOXO directly represses these genes. Four genes were not differentially expressed despite their enriched dFOXO binding in the insulin mutants; activated dFOXO may be required but not sufficient to control the expression of these genes. Thus, dFOXO can function as both a transcriptional activator and repressor [26], but this factor may also become poised at genes upon reduced insulin signaling and not affect transcriptional changes until the required co-factors are induced by other signals.
To determine if candidate dFOXO targets contribute to aging regulation we measured lifespan and age-specific mortality when each was reduced by RNAi or over-expressed from transgenes. Cohorts of control and mis-expression genotypes were coisogenic; misexpression was induced only in adults via GeneSwitch (GS)-Gal4 driving either UAS-RNAi or UAS-transgene. The effect of RNAi on lifespan was assessed for all 23 candidates. Knockdown of three genes (daw, Glyp and Tsp42Ef) extended lifespan by consistently reducing age-specific mortality (Figures 1F–1H, Figure S2 and Table S1), while knockdown of 14 genes shortened lifespan (Table S1). Among the candidates whose transcriptions were positively regulated by dFOXO, seven transgenic lines were available to test the effect of overexpression on lifespan; two cases had no effect on lifespan while five cases reduced survival (Table S2).
Among the observed longevity assurance genes, daw-RNAi induced by two independent ubiquitous GeneSwitch drivers respectively extended lifespan 12% to 35% (mean lifespan) by consistently reducing mortality rate (Table S1, S3). Dawdle is one of two Drosophila Activin-like ligands [31], [32], belonging to the Transforming Growth Factor-β (TGF-β) protein superfamily. To date, daw is reported to function in axon guidance [31], [32], cell proliferation and larval brain development [33]. Our results indicate that daw acts as a downstream target of dFOXO to modulate lifespan, suggesting that the Activin branch of TGF- β signaling may participate in control of aging.
Drosophila has two TGF-β ligand subfamilies: bone morphogenetic proteins (BMP) (ligands: Dpp, Gbb and Scw) and Activin (ligands: Daw and Act-β). These ligands signal through subfamily-specific Type I receptors (Tkv and Sax for BMP, Babo for Activin) and shared Type II receptors (Punt, Wit). BMP-like ligands and Activin-like ligands activate distinct downstream signaling cascades leading respectively to phosphorylation of the Smad transcription factors Mad and Smox (Figure 2A) [34].
Since daw-RNAi increases lifespan, we determined whether other elements of either TGF-β pathway could likewise control aging (Figures 2C–2F and Figure S3). RNAi for Smox, the Activin associated Smad transcription factor, extended lifespan 10% (Figure 2F). RNAi for Activin receptor babo and the Activin-like ligand Act-β did not affect survival. Repressing the BMP branch of TGF-β signaling via RNAi for dpp, gbb, Mad and Tkv consistently reduced survival (Table S3). Ubiquitously overexpressing genes in either Activin or BMP subfamily shortened lifespan (data not shown), as did RNAi for co-Smad (Med), the shared Type-II receptor (Punt and Wit) and two other TGF- β ligands (Myo and Mav) (Table S3).
The TGF-β signaling pathways of Drosophila are homologous to C. elegans TGF-β/dauer and Sma/Mab. Recent reports clarify that the TGF-β/dauer pathway can regulate somatic aging, while the Sma/Mab pathway appears to modulate reproductive aging [35], [36]. We performed a phylogenetic analysis on TGF-β ligands of C. elegans, Drosophila and mouse (Figure 2B). Similar to previous published phylogenetic analysis [37], [38], we found that the TGF-β/dauer ligand of C. elegans, DAF-7, is closely related to Activin-like ligands in Drosophila (Daw and Activin-β) and mouse (Activin-A, B, C and E), while the Sma/Mab ligand in C. elegans, DBL-1, is similar to BMP-like ligands in Drosophila and mouse. Together these results suggest that Activin may be a conserved longevity pathway.
To understand how Activin regulates Drosophila aging we determined which tissues produced this control. daw mRNA is highly expressed in muscle and fat body, a tissue with both liver and adipose-like activities (Figure 3A). Smox protein is more widely distributed (Figure 3B). To assess the role of Activin in muscle and fat body we knocked down daw, Smox and babo with tissue-specific drivers. Lifespan was extended by inactivating each of these genes in muscle, but not in fat body (Figures 3C–3H, Figure S4 and Table S4). Since daw is a dFOXO downstream target that is down-regulated in chico mutants (Figure 1D), we determined whether reduced insulin/IGF-1 signaling modulates Activin within muscle. chico mutants showed reduced daw mRNA sampled from thorax (containing mostly flight muscle), and this effect was reversed in chico; foxo double mutants (Figure 3I). Furthermore, Smox protein was less phosophorylated in chico mutants (Figure 3J). Insulin/IGF-1 signaling, through dFOXO, thus appears to modulate muscle Activin signaling, which in turn is sufficient to regulate longevity.
Muscle performance in many animals declines in parallel to the accumulation of misfolded protein aggregates [39]. Insulin/IGF-1 signaling in Drosophila may affect this process since over-expressing dFOXO in Drosophila muscle slows the aggregate accumulation and promotes macroautophagy [18]. Here we determine whether dFOXO mediates its effects on muscle proteostasis and function through its control of Activin.
Experimentally reducing Activin prevents the decline of muscle function with age. Flight activity typically declines in aging flies [18], as it does in our wildtype control (Figures 4A–4C). RNAi against the Activin factors daw, Smox and babo each retarded this decline (Figures 4A–4C). Likewise, the ability to climb at advanced ages was preserved in daw RNAi flies relative to wildtype (Figure S6D). Progression of these composite movement traits was associated with changes in protein aggregates within muscle. Aggregates visualized with Poly-Ubiquitin FK2 antibody increase with age in wildtype muscle, but this change was significantly delayed by muscle specific RNAi against daw, Smox or babo (Figure 4D, F).
Macroautophagy modulates protein aggregate accumulation [40]. We used two markers of lysosome/autophagy activity, lysotracker and cherry-tagged-Atg8a (homolog of LC3), to determine if Activin regulates muscle proteostasis through macroautophagy. The intensity of lysosome markers decreased with age in wildtype flight muscle, but was maintained in aged muscle expressing RNAi for daw, Smox, or babo (Figure 4E, quantified in Figure 4G). We likewise observed more autophagosomes in flight muscle with inactivated TGF- β/Activin signaling (via RNAi for daw, Smox, or babo) (Figure 5A). In contrast, constitutively activated Activin signaling (via overexpressing babo-Act) reduced the number of autophagosomes (Figure 5A, 5B). Since Activin signaling is transcriptionally regulated by dFOXO via daw, these results may explain reported associations between reduced insulin signaling and elevated autophagy [18]. Reduced insulin signaling represses Activin, which in turn releases repression of autophagy and thereby reduces accumulation of protein aggregates.
Drosophila encode 18 autophagy genes [41]. Many of these are less expressed in aged flies (Figure 5C) [18]. Since reduced Smox mRNA produces elevated autophagy in aged muscle, we studied the phosphorylation of this transcription factor in old flies. Smox phosphorylation is increased in aging muscle (Figure 5D), suggesting that Activin may be a negative regulator of Atg gene expression. Indeed, Atg6 and Atg8a mRNA were increased when daw and Smox were reduced in muscle (Figure 5E, 5F), while mRNA of Atg5, Atg6 and Atg8a were reduced by over-expressing constitutively active form of the babo receptor (Figure 5G).
Drosophila Smox protein is homologous to vertebrate Smad2 and Smad3 transcription factors. Human Smad3 protein recognizes the consensus sequence GTCTAGAC [42], although a single copy of the Smad box (GTCT) is also reported to support Smad3 binding at the MH1 domain [43]. We searched the promoter regions of Atg8a and identified at least two adjacent Smad boxes located within Atg8a (Figure 6A). ChIP-PCR with affinity-purified Smox antibody showed that Smox binds to the promoter region of Atg8a (Figure 6B), but not Atg1 and Atg6 (Figure 6C). In contrast to Smox, dFOXO does not bind to the promoter of Atg8a (Figure S6E). This is unlike mammalian FoxO3 which induces autophagy by directly binding to the promoters of LC3b, Gabarapl1, and Atg12l in C2C12 myotubes [44]. Consistent with our model for negative regulation on Activin signaling by activated dFOXO, chico −/− inhibits Smox binding at Atg8a (Figure 6D).
An electrophoretic mobility shift assay (EMSA) confirms that Smox binds directly within Atg8a promoter. We expressed and purified a recombinant protein of the Smox-MH1 DNA binding domain (amino acids 1–140) and measured its interaction with biotin-labeled Atg8a oligonucleotide probes containing Smad box (5′-AGAC AGAC-3′). Smox-MH1 strongly bound to the Atg8a probe, and this interaction was blocked by addition of unlabeled wildtype cold probes (Figure 6E). To define the required sequences of this Smad box (AGAC) we competed labeled wildtype probe with mutated cold probes. Unlabeled cold probes with mutations in both Smad boxes (Mut2 and Mut3) did not compete with the wildtype binding, but cold probe mutants for only a single only Smad box (Mut1 and Mut4) retained some competitive ability (Figure 6E). Furthermore, in vitro expressed Smox-MH1 also binds to the oligonucleotide probe for the vertebrate Smad binding element (5′-GTCT AGAC-3′) (Figure 6F). Together these data identify an invertebrate Smad binding element (AGAC AGAC) in the promoter region of the autophagy gene Atg8a. This Smad binding element contains a direct repeat of two Smad boxes (AGAC). Upon activation, Smox, the Drosophila homologue of Smad2/3, binds to the Smad box located within the promoter of Atg8a. Activin signaling represses autophagy via direct transcriptional regulation on the key autophagy gene Atg8a.
To test whether increasing Atg8a expression within muscle is sufficient to promote lifespan, we over-expressing Atg8a using a muscle-specific driver (MHC-Gal4). Lifespan was modestly but significantly increased, suggesting that Atg8a gene is a specific instance of a longevity assurance genes that that functions through muscle downstream of Activin signaling (Figure 7A, Table S4). To further examine whether Atg8a is required for Activin-mediated lifespan extension, we silenced both Atg8a and daw using muscle-specific RNAi (Figure 7B, Figure S7, Table S5). Lifespan extension when daw RNAi was expressed in muscle was rescued when Atg8a was simultaneously reduced by RNAi in this tissue, indicating Activin regulates longevity through muscle Atg8a.
Previous studies from C. elegans recognize cross-talk between insulin/IGF-1 and TGF-β pathways [36], [45]. DAF-16 is nuclear localized in many mutants of the TGF-β/dauer pathway [36]. To determine if muscle Activin signaling affects Drosophila lifespan through systemic insulin/IGF-1signaling we measured circulating insulin-like peptides in adults with tissue specific daw-RNAi. Knockdown of daw in muscle reduced the level of hemolymph DILP2 (Figure 7C), while dilp2 mRNA remained constant (Figure 7D), suggesting the daw specifically modulates DILP2 secretion. In contrast, knockdown of daw in fat body increased the level of circulating DILP2 (Figure S6F). These contrasting tissue associated changes correspond to the observed effects upon lifespan when daw is reduced in each tissue (Figure 3C–3H). Notably, dilp2 mRNA is reduced in other tissue-limited genetic manipulations that extend lifespan [19], [46], and knockout of the dilp2 locus is sufficient to extend lifespan [47]. We now see that reducing muscle Activin signaling via daw RNAi also remotely controls DILP2 secretion from the brain. This is sufficient to decrease systemic insulin signaling because insulin/FOXO sensitive 4ebp mRNA is elevated in peripheral tissues (e.g. fat body) (Figure 7E).
Unlike the consistent response of 4eBP in longevity mutants with reduced insulin, some but not all insulin/IGF pathway mutants have reduced fecundity. Here we found normal fecundity in females with reduced muscle Activin signaling (Figure 7F), suggesting the effects of muscle Activin on lifespan are not mediated through trade-off between longevity and reproduction.
Insulin/IGF-1 signaling modulates longevity in many animals. Genetic analysis in C. elegans and Drosophila shows that insulin/IGF-1 signaling requires the DAF-16/FOXO transcription factor to extend lifespan, while in humans several polymorphisms of FoxO3A are associated with exceptional longevity [15], [16]. Although many downstream effectors of FOXO have been identified through genome-wide studies [22], [24], [25], [26], the targets of FOXO responsible for longevity assurance upon reduced insulin signaling are largely unknown [24]. Here we found 273 genes targeted by Drosophila FOXO using ChIP-Seq with two long-lived insulin mutant genotypes. We focused on daw, an Activin ligand, which is transcriptionally repressed by FOXO upon reduced insulin/IGF signaling. Inactivation of daw and of its downstream signaling partners babo and Smox extend lifespan. These results are reminiscent of observations from C. elegans where reduced TGF-β/dauer signaling extends longevity [36]. Notably, the lifespan extension of TGF-β/dauer mutants (e.g. daf-7 (e1372) mutants) can be suppressed by daf-16 mutants, suggesting that TGF-β signaling intersects with the insulin/IGF-1 pathway for longevity in C. elegans [36]. In our phylogenetic analysis, DAF-7, Daw and mammalian Activin-like proteins share common ancestry. Activin signaling, in response to insulin/IGF-1, may thus represent a taxonomically conserved longevity assurance pathway.
Longevity benefits of reduced Activin (TGF-β/dauer) in C. elegans were resolved only when the matricide or ‘bagging’ (due to progeny hatching within the mother) was prevented by treating daf-7(e1372) mutants with 5-fluorodeoxyuridine (FUdR) to block progeny development [36]. This approach made it possible to distinguish the role of Activin in somatic aging from the previously recognized influence of BMP (Sma/Mab signaling) upon reproductive aging in C. elegans [35], [48]. Activin, of course, is a somatically expressed regulatory hormone of mammalian menstrual cycles that induces follicle-stimulating hormone (FSH) in the pituitary gland. In young females, FSH is suppressed within a cycle when maturing follicles secrete the related TGF-β hormone Inhibin [49]. In mammalian reproductive aging, the effect of Activin in the pituitary becomes unopposed as the stock of primary follicles declines, thus inducing elevated production of FSH. We now find that reduced Activin but not BMP signaling favors somatic persistence in Drosophila. These parallels between reproductive and somatic aging among invertebrate models and humans suggest that unopposed Activin signaling is pro-aging while favoring reproduction.
Reduced insulin/IGF signaling extends lifespan through interacting autonomous and non-autonomous actions. Reducing IIS in some distal tissues has been shown to slow aging because this reduces insulin secretion from a few neurons: reducing IIS by increasing dFOXO in fat body or muscle extends Drosophila fly lifespan while decreasing IPC production of systemically secreted DILP2 [18], [19]. Here we identify Activin as a direct, downstream target of insulin/dFOXO signaling within muscles that has the capacity to non-autonomously regulate lifespan. Knockdown of Activin in muscle but not in fat body is sufficient to prolong lifespan. RNAi for muscle Activin signaling led to decreased circulating DILP2 and increased peripheral insulin signaling. Muscle is thus proposed to produce a signaling factor, a myokine, which impacts organism-wide aging and metabolism [18], [50], [51] (Figure S8).
Aging muscle may produce different myokine-like signals in response to their physiological state. Aged muscles degenerate in many ways including changes in composition, mitochondria, regenerative potential and within-cell protein homeostasis [52]. Protein homeostasis is normally maintained, at least in part, by autophagy [40], [53]. Loss of macroautophagy and chaperone-mediated autophagy with age will accelerate the accumulation of damaged proteins [54]. Expression of Atg8a in Drosophila CNS is reported to extend lifespan by 56% [28], while recent studies find elevated autophagy in long-lived mutants including those of the insulin/IGF-1 signaling pathway [18], [55], [56]. Our results now show that insulin/IGF signaling can regulate autophagy through its control of Activin via dFOXO. Poly-ubiquitinated proteins accumulate in aging Drosophila while lysosome activity and macroautophagy decline. Muscle performance with age (flight, climbing) was preserved by inactivating Activin within this tissue. This genetic treatment also reduced the accumulation of protein aggregates. These effects are mediated by blocking the transcription factor Smox, which otherwise represses Atg8a. Smox directly regulates Atg8a through its conserved Smad binding motif (AGAC AGAC). These results, however, contrast with an observation where TGF-β1 promotes autophagy in mouse mesangial cells [57].
Insulin/IGF-1 signaling is a widely conserved longevity assurance pathway. Our data indicate that reduced insulin/IGF-1 retards aging at least in part through its FOXO-mediated control of Activin. Furthermore, affecting Activin only in muscle is sufficient to slow its functional decline as well as to extend lifespan. Autophagy within aging muscle controls these outcomes, and we now find that Activin directly regulates autophagy through Smox-mediated repression of Atg8a. If extrapolated to mammals, pharmaceutical manipulations of Activin may reduce age-dependent muscle pathology associated with impaired autophagy, and potentially increase healthy and total lifespan through beneficial signaling derived from such preserved tissue.
Flies were reared and maintained at 25°C, 40% relative humidity and 12-hour light/dark. Adults were maintained upon agar-based diet with cornmeal (0.8%), sugar (10%), and yeast (8% unless otherwise noted). RU486 (mifepristone, Sigma, St. Louis, MO, USA) used to activate GeneSwitch-Gal4 was dissolved in ethanol to a concentration of 200 µM and added to the food.
Fly stocks included: MHC-Gal4 [18], Tub-GS-Gal4 [58]; da-GS-Gal4 [59]; S106-GS-Gal4 [58]; dilp2-GS-Gal4 (provided by H. Jasper); dilp2-GS-Gal4,UAS-dicer2 (provided by S. Helfand). RNAi lines for dFOXO target genes and TGF-β pathway are from Bloomington stock center (TRiP line) and Vienna Drosophila RNAi Center (see the supplemental Table for the stock number). UAS-lines are: UAS-hairy [60], UAS-puc [61], UAS-RhoGAP18B [62], UAS-vri [63], UAS-wit [64], UAS-cv-2 [65], UAS-babo-Act (also known as UAS-babo*1A2) [66], UAS-Atg8-Cherry [67]. MHC-Gal4 is a constitutive muscle-specific driver, with expression restricted to the thoracic region and legs (data not shown).
Chico mutants were made in our lab as describe previously [13], [29]: y1; cn1; ry506 (wildtype), y1; cn1 chico1/cn1; ry506 (chico−/+), y1; cn1 chico1/cn1 chico1; ry506 (chico−/−), y1; cn1 chico1/cn1 chico1; foxo21 ry506 (chico−/−, foxo−/−). Adult on-set IPC ablation flies were made by crossing Dilp2-GS-Gal4 to UAS-rpr and inducing the cell death in IPC cells by feeding flies with RU486 for 15 days.
Two insulin mutants were used in ChIP-Seq experiments, chico −/+ and IPC ablation. Chromatin immunoprecipitation (ChIP) was performed according to previously published methods with modification [68], [69], [70]. About 200–250 adult females (∼200 mg) at the age of 15-day-old were pooled for each ChIP sample. Two biological replicates were prepared for each genotype. Flies were homogenized and cross-linked in 1× PBS containing 1% formaldehyde. The fly lysate were sonicated using a Branson 450 sonicator to break down the chromatin into a pool of DNA fragment with average size of 500 bp. Immunoprecipitation was performed using Dynal protean A beads (Invitrogen, Grand Island, NY, USA) and affinity purified anti-dFOXO antibody made in our laboratory. Following the wash with LiCl and TE buffer, the DNA-protein complex was eluted from the Dynal beads and reverse cross-linked. After Proteinase K digestion, dFOXO-bound DNA fragments were purified and diluted in Tris-HCl buffer. About 20 ng of ChIP DNA (dFOXO-bound DNA) and input DNA (DNA sample before the immunoprecipitation) were used in library preparation following the methods described in [71]. The libraries were then size-selected (150 bp-350 bp) and purified by agarose gel, and subjected to the Illumina Genome Analyzer IIx Sequencer (Illumina, San Diego, CA, USA).
To map the dFOXO binding sites, we pooled the raw reads (about 20 million reads per sample) from two replicates into one data file and aligned it to Drosophila reference genome using Bowtie short read aligner [72]. About 70% of raw reads have at least one alignment. The enrichment of dFOXO binding between ChIP DNA and input DNA was determined using peak calling package PeakSeq [73]. Enriched regions with FDR of 0.01 were selected. Target genes, which were detected 5 kb away from the center of the binding sites, were also obtained. The ChIP-Seq raw data are archived at NCBI GEO with Accession # GSE44686.
For ChIP-PCR analysis, the binding enrichment was calculated as the fold change of ChIP DNA versus input DNA. The binding to the coding region of Actin gene (Act5C) and sry genomic region were used as negative controls.
The DAVID functional classification tool was used for pathway and molecular function analysis on the dFOXO target genes [74]. Genomic sequence near the dFOXO binding region (∼200 bp) was downloaded from the Flybase (http://flybase.org/) and de novo motif analysis was performed using MEME Suite [75].
Total RNA was extracted from 10 whole flies or from tissue of 15 flies in Trizol reagent (Invitrogen, Grand Island, NY, USA). DNase-treated total RNA was quantified with a NanoDrop ND-1000. About 50–100 ng of total RNA was used for quantification with SuperScript One-Step RT-PCR reagent (Invitrogen, Grand Island, NY, USA) and measured on an ABI prism 7300 Sequence Detection System (Applied Biosystems, Carlsbad, CA, USA). Three biological replicates were used for each experimental treatment. mRNA abundance of each gene was normalized relative to ribosomal protein L32 (RpL32, also known as rp49) by the method of comparative CT. Primer sequences are shown in Table S8.
Full length TGF-β ligands from worm, fly and mouse were aligned in ClustalW. From the alignments, a phylogenetic tree was constructed using MEGA 5.0 [76], according to the neighbor-joining method with a bootstrap test calculated with 2000 replicates and a poisson correction model. Mouse Glial cell line-derived neurotrophic factor (GDNF) was used as the out-group.
Two to three-day-old female adult flies were collected with light CO2 anesthesia and pooled in 1 L demography cages at a density of 100 to 125 flies per cages. Three independent cages were initiated per genotype. Food vials with media containing vehicle only or RU486 were changed every two days, at which time dead flies were removed and recorded. Survival analysis was conducted with JMP statistical software with data from replicate cages combined. Survival distributions were compared by the Log-Rank test. Cox proportional hazard survival analysis was used to assess how reduced daw and Atg8a interacted to affect mortality.
Flying and climbing assays were scored as described in [18]. In the flying assay, flies were released at the top of a 250 ml cylinder (about 30 cm long). The number of flies that didn't fall straight to the bottom of the cylinder was recorded. A total of 40 females were scored for each genotype.
In the climbing assay (also known as negative geotaxis assay), flies were first tapped down to the bottom of a standard (empty) food vial, and the percentage of flies that climbed up 8 cm within 20 seconds was recorded. A total of 80 females (10 flies per vial) were scored for each genotype.
Antibodies for immunostaining included: anti-polyubiquitin FK2 (1∶200) (Assay Designs/Enzo Life Sciences, Farmingdale, NY, USA), and anti-rabbit IgG-DyLight 488 (1∶300) (Jackson ImmunoResearch, West Grove, PA, USA). F-actin was visualized by Alexa Fluor 488-conjugated Phalloidin (Invitrogen, Grand Island, NY, USA). Lysosome was monitored by LysoTracker Red DND-99 at the concentration of 100 nM (Invitrogen, Grand Island, NY, USA). DNA was stained with Hoechst 33342 (1 µg/ml) (Invitrogen, Grand Island, NY, USA). Samples were processed as described in [18], and imaged with a Leica SP2 laser scanning confocal microscope. To quantify the area of protein aggregates and the number of lysotracker or Atg8a-positive dots, grayscale images were converted to binary images (halftone or black & white) with a grayscale cutoff of 20 pixels using ImageJ software [77]. The number/area of positive immunostaining was measured with the “Analyze Particles” function.
Smox polyclonal antibody was generated against the peptide sequence (DSIVDYPLDNHTHQ) corresponding to amino acids 143–156 (Covance, Dedham, MA, USA) and affinity purified (Thermo/Pierce, Waltham, MA, USA) (specificity documented in Figure S5). Phospho-Smad2 antibody was from Cell Signaling Technology (#3108) (Danvers, MA, USA). Thorax tissue from ten female adults was homogenized in RIPA buffer (Thermo/Pierce, Waltham, MA, USA) with protease inhibitor cocktail (Sigma, St. Louis, MO, USA). Supernatant was incubated with SDS loading buffer (Invitrogen, Grand Island, NY, USA) at 70°C for 10 min. About 30 µg of denatured protein was separated on 10% SDS-polyacrylamide precast gels (Invitrogen Grand Island, NY, USA) and transferred to nitrocellulose membranes. Following incubation with primary and secondary antibodies, the blots were visualized with Pierce ECL Western Blotting Substrate (Thermo Fisher Scientific, Waltham, MA, USA). Band intensity was quantified with Image Lab software (Bio-Rad, Hercules, CA, USA).
cDNA for Smox-MH1 DNA binding domain (1–420 nt) was cloned into pFN29K-His6HaloTag protein expression vector (Promega, Madison, WI, USA). After expression in E. coli, recombinant proteins were purified using HaloTag purification kit (Promega, Madison, WI, USA). Empty vector was used as a negative control.
Biotin-labeled DNA probes were generated using 3′-end biotin labeling (Fisher/Thermo, Waltham, MA, USA). The binding reactions were carried out in a 10 µl of assay mixture containing 10 mM Tris·HCl (pH 7.5), 150 mM KCl, 5 mM MgCl2, 10 ng/µL poly(dI-dC), ∼50 ng labeled probe and 20 µg purified recombinant protein. After incubation at room temperature for 20 min, the mixtures were electrophoresed on 0.8% agarose gels in 0.5× Tris/borate/EDTA buffer. Biotin-labeled DNAs were transferred to a positive-charged nylon membrane (Invitrogen, Grand Island, NY, USA) and detected using GelShift Chemiluminescent EMSA (Active Motif, Carlsbad, CA, USA).
10-day-old mated female flies were maintained on standard food (2% yeast) for five days at three females per vial and 8–10 vials per group. Flies were passed daily to new vials over five days and eggs were counted daily.
We followed our recently reported EIA assay to measure hemolymph DILP2 [21]. Briefly, about 0.5 µL of hemolymph was collected by decapitation of 15 female flies. Hemolymph was then incubated overnight in a 96-well EIA/RIA plate (Corning Incorporated, Corning, NY, USA) at room temperature. Anti-DILP2 antibody (gift from P. Leopold) was used at 1∶2500 dilution. After the incubation with a HRP-conjugated secondary antibody (1∶2500), hemolymph samples were treated with TMB solution (3,3′,5,5′-teramethylbenzidine; American Qualex antibodies, San Clemente, CA);absorbance was recorded at 450 nm upon a plate reader.
Data are presented as mean ± SEM from three independent biological replicates, unless otherwise noted. Statistical significances were evaluated by t-test and one-way ANOVA analyses using GraphPad Prism Software.
|
10.1371/journal.pcbi.1005291 | A General Shear-Dependent Model for Thrombus Formation | Modeling the transport, activation, and adhesion of platelets is crucial in predicting thrombus formation and growth following a thrombotic event in normal or pathological conditions. We propose a shear-dependent platelet adhesive model based on the Morse potential that is calibrated by existing in vivo and in vitro experimental data and can be used over a wide range of flow shear rates (100 < γ ˙ < 28 , 000 s - 1). We introduce an Eulerian-Lagrangian model where hemodynamics is solved on a fixed Eulerian grid, while platelets are tracked using a Lagrangian framework. A force coupling method is introduced for bidirectional coupling of platelet motion with blood flow. Further, we couple the calibrated platelet aggregation model with a tissue-factor/contact pathway coagulation cascade, representing the relevant biology of thrombin generation and the subsequent fibrin deposition. The range of shear rates covered by the proposed model encompass venous and arterial thrombosis, ranging from low-shear-rate conditions in abdominal aortic aneurysms and thoracic aortic dissections to thrombosis in stenotic arteries following plaque rupture, where local shear rates are extremely high.
| Hemostasis (thrombus formation) is the normal physiological response that prevents significant blood loss after vascular injury. The resulting clots can form under different flow conditions in the veins as well as the arteries. The excessive and undesirable formation of clots (i.e., thrombosis) in our circulatory system may lead to significant morbidity and mortality. Some of these pathologies are deep vein thrombosis and pulmonary embolism and atherothrombosis (thrombosis triggered by plaque rupture) in coronary arteries, to name a few. The process of clot formation and growth at a site on a blood vessel wall involves a number of simultaneous processes including: multiple chemical reactions in the coagulation cascade, species transport and platelet adhesion all of which are strongly influenced by the hydrodynamic forces. Numerical models for blood clotting normally focus on one of the processes under a specific flow condition. Here, we propose a general numerical model that encompass a wide range of hemodynamic conditions in the veins and arteries, with individual platelets and their adhesive dynamics included explicitly in the models. Further, we include the biochemistry of coagulation cascade, which is essential to modeling thrombus formation, and couple that to our platelet aggregation model. The simulation results—tested against three different experiments—demonstrate that the proposed model is effective in capturing the in vivo and in vitro experimental observations.
| Platelets are fundamental to both hemostasis and thrombosis in many vascular diseases, including abdominal aortic aneurysm (AAA), thoracic aortic aneurysm and dissection (TAAD), and carotid atherosclerosis [1–3]. Normal platelets do not interact with the healthy artery wall. In cases of endothelial injury or exposure of extracellular matrix to blood flow, however, platelets can quickly activate and cover the injured area to stop bleeding. The initial adhesion of platelets on the thrombogenic area can be attributed to a variety of platelet membrane receptor-ligand interactions, such as glycoprotein Ib(GPIb)-V-IX with immobilized von Willebrand Factor (vWF), GPIIb-IIIa (αIIbβ3) with vWF, GPVI with collagen, α2β1 with collagen, αIIbβ3 with fibrinogen, and so on, depending on the nature of the lesion [4] and the local shear rate of blood flow [5–7]. At low shear rates (γ ˙ < 1000 s - 1), platelets adhere to the thrombogenic area through different pathways, relying on the exposed extracellular matrix (ECM) proteins [4, 5, 8]. On the other hand, as shear rate increases, interactions between immobilized vWF and GPIb become exclusive in initializing platelet aggregation while other interactions are broken down due to high bond failure rates [9–11]. The reason that vWF-GPIb interactions persist at such high shear rates (≈ 25,000 s−1 shown in in vitro experiments [11]) is that the vWF proteins, which are normally in a coiled state, tend to extend several fold in high-shear environments. The conformational change of vWF exposes the repeating functional A-1 domains in multimeric vWF, leading to enhanced adhesive interactions between GPIb and vWF [12–15]. Recently, experiments showed that the effect of vWF multimer extension was more pronounced in elongational flows, like in stenotic arteries, than in pure shear flows in a straight vessel [14].
The exposure of the subendothelial matrix triggers coagulation, which involves a network of tightly regulated enzymatic reactions leading to the production of the enzyme thrombin. Thrombin activates platelets and creates fibrin monomers that polymerize into a fibrous gel that stabilizes the clot. Coagulation is believed to be initiated when tissue factor (TF) molecules embedded in the vessel wall are exposed by injury and bind plasma enzyme factor VIIa [16]. Platelet activation can be induced by direct contact of platelets with collagens exposed in the subendothelium, by the action of thrombin, or by exposure to a threshold level of adenosine diphosphate (ADP) and thromboxane-A2 (TxA2). A finite quantity of ADP and TxA2 is released by a platelet during a time interval following the platelet’s activation. Numerous models are proposed for the systems biology of coagulation cascade among which the Kuharsky and Fogelson [16] is considered the most comprehensive one as it takes into account plasma-phase, subendothelial-bound and platelet-bound enzymes and zymogens. An extended version of this model was introduced by Leiderman and Kuharsky [17] to incorporate the spatial variations, represented by a system of partial and ordinary differential equations for the reactive transport of the chemical species. In this work, to reduce the computational cost, we use a slightly reduced-order model of coagulation proposed by Anand et al. [18], which has the advantage of including both TF and contact pathways in plasma.
The above-mentioned platelet-wall interactions and coagulation occur in the presence of blood flow. Hemodynamics plays a key role in transporting the platelets to the thrombogenic area via advection and diffusion. Begent and Born [19] performed an in vivo study on the effect of blood flow rates (or equivalently shear rates) on thrombus formation in a venous flow. They discovered that thrombus growth in venules with diameters of 40 − 60μm reached a maximum at a blood flow velocity around 400μm/s due to the balance between the number of platelets transported to the injured sites and the shear stress on the surface of the growing thrombus. Transport of platelets and other proteins involved in thrombus formation (fibrinogen and plasminogen, among others) becomes particularly important in the pathological conditions of AAA and TAAD. For example, platelets and reactants flow into an AAA and initiate intraluminal thrombus at specific locations in the aneurysm bulge [20, 21]. Such intraluminal thrombus can affect the mechanical properties of the local vessel wall, leading to increased risk of aneurysm rupture [22]. In TAAD, however, clinical evidence suggests that a completely thrombosed false lumen within the dissection results in an improved prognosis whereas a partially thrombosed false lumen may render the wall more vulnerable to further dissection or rupture [23]. Whether a fully thrombosed TAAD is formed or not could be attributed to the hemodynamics in the false lumen.
Numerical models have been developed to study platelet activation, adhesion, and aggregation in both physiological and pathological conditions [17, 24–30]. Pivkin et al. [25] developed a platelet model based on the force coupling method (FCM) to simulate platelet aggregation in a circular vessel. This model reproduced the experimental results in [19] and explored the effect of flow pulsatility on thrombus formation. Xu et al. [26, 27] developed a 2D multiscale model to simulate thrombus formation at different stages. Kamada et al. [24] used spring models for a variety of ligand-receptor interactions between platelets to investigate effects of ligand-receptor deficiencies on thrombus formation at different shear rates. Mountrakis et al. [29] used a 2D immersed boundary model and simulated platelets and red blood cells (RBCs) in blood vessels with saccular-shaped aneurysms. Biasetti et al. [31] solved advection-diffusion-reaction for multiple biomolecules in the coagulation cascade in fusiform-shaped AAAs to predict the location of intraluminal thrombus formation. In a very recent work, Tosenberger et al. [30] investigate the interaction of blood flow, platelet aggregation and plasma coagulation using a hybrid dissipative particle dynamics-continuum model in a 2D channel. The flow of plasma with the suspending platelets are solved using dissipative particle dynamics, while the regulatory network of plasma coagulation is described by a system of partial differential equations. Although considerable work has been conducted to simulate the advective and diffusive motions of platelets and other blood components in arterial flows, most studies focused on simplified arterial geometries. Transport and aggregation of platelets in dissections and stenoses have not yet been well studied due to the complex geometries and varying mechanisms of platelet adhesion under different hemodynamic conditions.
Our main goal in this paper is to develop a phenomenological model for platelet-wall and platelet-platelet adhesion, whose strength depends on the local shear rate, to represent different adhesion mechanisms. We model platelets as rigid spherical particles using the Lagrangian description within the context of FCM [32], as adopted in [25], whereas the hemodynamics and chemical transport are obtained from the solution of the Navier-Stokes (NS) equations and advection-diffusion-reaction (ADR) equations on a fixed Eulerian grid, respectively. We present the calibration of parameters in Eq (10) based on carefully chosen experimental data from the literature, where the platelet aggregation process is mainly separated from the complex biochemistry of the coagulation cascade. More specifically, we use the in vivo experimental data of Begent and Born for venous thrombus formation in mice [19] to calibrate our model for low-shear-rate regimes, where platelet aggregation is induced by the release of ADP in vivo causing the formation of white thrombi. In the high-shear regime, we use the results reported by Westein et al. for stenotic microchannels [14], where the shear rates can reach as high as 8,000 s−1. Here, platelet aggregation is caused by perfusing whole blood over surfaces coated by vWF/fibrinogen. Further, we use the experimental results in [33] for the purpose of testing our platelet aggregation model in a stenotic channel coated with collagen where shear rates are as high as 15,000 s−1. In the second part of the paper, we use a detailed model for blood coagulation coupled with our platelet aggregation model to address thrombus formation in arteriole-sized vessels similar to the in vitro experiment of Shen et al. [34] Our simulations agree well with the wide range of experimental data considered, thus suggesting the effectiveness of the proposed approach in modeling thrombus formation in blood vessels having complex geometries and under a broad range of flow conditions.
Platelet motion within a flow field and adhesion to a damaged surface are solved together by coupling a spectral/hp element method (SEM) [35] with a FCM [32]. Specifically, SEM is used to solve the flow field and the reactive transport of chemical species on a fixed Eulerian grid, whereas FCM is implemented to describe the two-way interactions between the blood flow and Lagrangian particles (i.e., platelets).
Simulations with fully resolved RBC and platelet suspensions in blood are challenging due to the computational cost of modeling millions of particles. In order to reduce the computational cost, we take blood as a continuous medium, and the effect of RBCs on platelet margination is taken into account by assuming that blood flow at the inlet of the simulated vessels is fully developed and platelets are already marginated toward the vessel wall. We prescribe the distribution of the platelets at the inlets based on the reported experimental distributions of Yeh et al. [36]. The reported distributions are obtained for platelet-sized latex beads suspended in whole blood flowing in tubes with ≈ 200 μm diameter at 40% hematocrit, where the average wall shear rate is ≈ 500 s−1.
In FCM, the translational velocity of each platelet particle is estimated by the local average of the fluid velocity weighted by a Gaussian kernel function. In our simulations, we assume platelets to be spheres with radius of 1.5 μm and number density of 300,000 mm−3, while blood is assumed to be an incompressible Newtonian fluid. Applying the FCM method detailed in [32], the governing equations for the incompressible flow are
ρ ( ∂ u ∂ t + u · ∇ u ) = - ∇ p + μ ∇ 2 u + f ( x , t ) , (1) ∇ · u = 0 , (2) f ( x , t ) = ∑ n = 1 N F n Δ ( x - Y n ( t ) ) , (3)
where u, p, and μ are the flow velocity, pressure and blood viscosity, respectively, and Fn in Eq (3) is the force due to particle n (discussed later). The effect of the platelets on the flow field is incorporated into the body force term f (x, t) in the Navier-Stokes Eq (1). The contribution of each platelet whose center of mass is located at Yn to the flow at position x is smoothed by a Gaussian distribution kernel Δ(X ≡ (x − Yn)), where Δ(X) is
Δ ( X ) = ( 2 π σ 2 ) - 3 / 2 exp ( - X · X / 2 σ 2 ) , (4)
with σ the standard deviation of the kernel, which is related to particle radius a through σ = a / π. The governing equations are written in weak form and the domain is discretized using spectral elements that allow high order Jacobi polynomials. Time integration is performed using a semi-implicit splitting scheme [35].
The velocity of each platelet Vn is calculated by interpolating the local flow velocity at the location of a platelet using the same Gaussian kernel of Eq (4) (different standard deviations may be used for force and velocity interpolations)
V n = d Y n d t = ∫ u Δ ( x - Y n ( t ) ) d x , (5)
where the position vectors for all the platelets are updated at each time step using a second-order Euler forward scheme. As mentioned above, a near-wall excess distribution of platelets is imposed for platelets entering the domain to take into account the effect of margination. It is known that the lateral platelet diffusion is enhanced through its collisions with RBCs, which is on the order of Dp = 10−7 cm2/s. This value varies with local shear rates and hematocrit. In one of our sensitivity studies, we augment platelet transport through the following equation for the displacement
d Y n = V n d t + ( 2 D p d t ) 1 / 2 R (6)
where R is a Gaussian random variable with mean 0 and variance 1. The effective diffusion coefficient is taken to be a function of the local shear rate based on the equation proposed by Wootton et al. [37], D p = 7 ( 10 - 9 ) γ ˙ ( c m 2 / s ), where the enhanced diffusion is considered in the lateral direction only.
The net force acting on each platelet Fn is written by
F n = - 4 3 π r 3 ( ρ p l a t - ρ f l u i d ) d V n d t + F i n t e r , (7)
where the first term is the inertial force resulting from the density difference between the platelets and blood flow. The second term accounts for the interaction forces between platelets with each other and the wall, which represent overall effects of different ligand-receptor interactions.
We propose a phenomenological model based on Morse potential UMorse to model the attractive/repulsive interactions between platelets, namely
U M o r s e = D e [ 1 - e - β d ( r / d - 1 ) ] 2 , (8)
where De is the energy depth contributing to the strength of the interaction force and β controls the width of the energy well; r is the distance between the platelets centeroids and d = 3 μm is the equilibrium distance between two platelets and is selected to be the diameter of platelet.
As shown in Fig 1, the Morse potential is similar to a Lennard-Jones potential; it consists of both attractive (at r > d) and repulsive parts (at r < d). The Morse potential possesses a softer repulsive-core, however, which is much more stable when simulating platelet aggregation. The magnitude of the interaction forces resulting from this Morse potential can be obtained by taking the variation of the potential with respect to interparticle distance r, which gives
F i n t e r = - ∂ U M o r s e ∂ r = 2 D e β [ e - 2 β d ( r / d - 1 ) - e - β d ( r / d - 1 ) ] . (9)
The maximum attractive forces between the two platelets can be calculated by ∂Finter/∂r = 0, which gives Fmax = βDe/2 occurring at (r/d − 1) = log(2)/βd. In our simulations, βd is selected to be 2.5 and thus the maximum attractive force is obtained at r ≈ 1.27d. The undetermined parameter De, which mainly controls the magnitude of the platelet interaction forces, is determined from experimentally measured thrombus formation and growth under different hemodynamic conditions. Toward this end, platelets are assumed to exist in three different states, namely passive, triggered, or activated. In passive or triggered states, platelets are non-adhesive, hence only repulsive forces are applied between them to prevent cellular overlap as shown by the blue line segment in Fig 1. If a passive platelet interacts with an activated platelet, it becomes triggered and will switch to an activated state after an activation delay time τact. When two activated platelets interact with each other, a repulsive force results when r < d and an attractive force when r > d as shown by the red line segment in Fig 1. For calibrating our platelet aggregation model, we consider an interaction distance of 2d between platelets within which resting platelets can get activated. It should be noted that in the in vitro experiments for platelet aggregation, platelets can bind directly to the collagen or vWF-coated surfaces without activation. This may be followed by irreversible platelet activation and the release of ADP, whereas thrombin production is excluded from these experiments.
Next, we present a phenomenological model that correlates the adhesion force to the local shear rate. The correlation has to be able to cover different flow conditions (e.g., clotting in venules vs. arteries) and adhesive mechanisms (e.g., adhesion at low vs. high shear rates). For that purpose we propose a shear-dependent correlation for De following a hyperbolic tangent formula
D e ( λ 2 ) = D e h [ tanh ( λ 2 - λ 2 l 1000 ) + D e l D e h + 1 ] , (10)
where λ 2 = 2 D : D is the square root of the second invariant of the fluid strain-rate tensor D, D e l and D e h determine the adhesive forces at low and high shear rates, respectively, and λ 2 l is the shear rate threshold value where transition from low to high shear regime takes place.
The constants in Eq (10) are calibrated using in vivo and in vitro experiments, which results in the function plotted in Fig 2. The model is tested for clotting in venules at low shear rates and microfluidic devices with a constriction resembling atherosclerosis plaques, which can induce high shear rates of the order of ≈ 20,000 s−1 [11]. We determined the constants to be D e h = 500 D e l, where D e l ≈ 2 . 1 × 10 - 17 N m and λ 2 l = 5 , 500 s - 1. The calibrated values are both inspired by the available data from the recent study by Mehrabadi et al. [11], which presents a predictive model for high-shear thrombus growth, and by the observations from our numerical simulations mimicking the in vivo and in vitro experiments. Additional details will be discussed in section Results.
In this section, we briefly describe the mathematical model for the combined TF and contact pathway of blood coagulation originally proposed by Anand et al. [18]. More details of the model as well as the reaction rate constants and parameters are given in the SI Text. The model takes into account plasma-phase enzymes and zymogens, and coagulation inhibitors, where the advection-diffusion-reaction (ADR) equations for plasma-phase enzymes, zymogens or complexes lead to a system of 20 partial differential equations (PDEs) in the following form
∂ c i ∂ t + u · ∇ c i = D i ∇ 2 c i + S i i = 1 to 20, (11)
where ci and Di are the concentration and diffusion coefficient for each reactant, respectively, and Si represents the rate of production or destruction of that reactant. A zero-flux boundary condition is imposed for most reactants in the ADR equations except for a few reactants (factors IX/IXa and X/Xa) to initiate the coagulation, which is in the form: −Di∂ci/∂n = Bi, where n is the unit normal on the boundary and Bi is the related surface reaction. For zymogens, the upstream concentrations at the inlet and initial concentrations are set to their normal plasma values, whereas all enzyme and complex concentrations are initially set to a very small nonzero value.
Passive platelets can directly bind to the collagen on the subendothelium and become activated. They also become activated by exposure to sufficiently high concentrations of thrombin, TxA2 and ADP. We define an activation function ω(x, t) = [IIa]/[IIa]thr + [ADP]/[ADP]thr + [TxA2]/[TxA2]thr, where the subscript “thr” corresponds to the threshold concentration that activates the platelets. Platelets transporting through the regions with values of ω > 1 will become activated. In this study, we assume a negligible activating effect for TxA2, and the threshold values of [IIa]thr = 1 nM [38] and [ADP]thr = 1,000 nM [39].
Experimental measurements show that platelets release a finite quantity of ADP to the blood stream within 5 seconds following activation [40]. Here, we assign a normal distribution for the release function R ( t ) = exp [ - ( t - μ ) 2 / σ r 2 ] / 2 π σ r 2 with the mean release time of μ = 3 s and variance of σ r 2 = 2 s 2. We use the same FCM Gaussian kernel function to evaluate the spatial distribution of ADP release from each platelet
S r e l ( x , t ) = ∑ n = 1 N A ′ R ( t - t a c t n ) Δ ( x - Y n ( t ) ) , (12)
where A′ = 3 × 10−8 nM is the ADP content for each platelet [39], and t a c t n is the time at which platelet n becomes activated.
Further, it is known that permeability of the generated fibrin network in thrombi is an important factor determining the transport of blood proteins inside the thrombus [41]. To couple the porosity of fibrin network to the local flow field we introduce a Brinkman term in the form of − (μ/k) u to the right hand side of the NS Eq (1), where μ is the blood viscosity and k is permeability inside the fibrin network, and is considered to be locally varying with the concentration of fibrin. The experimental measurements of Kim et al. showed an inverse power law permeability with respect to the fibrin volume fraction [41]. Assuming that the same correlation exists with respect to the fibrin concentration [Ia], we write k as
k = 8 ( 10 ) - 12 [ Ia ] [ Ia ] t h r - 1 . 8 ( m 2 ) , (13)
where [Ia]thr = 5,000 nM is the threshold concentration at the core of the clot causing the lowest clot permeability k = 8(10)−12 m2.
To initiate and drive the coagulation, a spatially varied concentration level of subendothelium-bound TF-VIIa complex is prescribed at the site of injury. This results in a few reactions at the wall (represented by flux conditions) that form enzymes IXa and Xa that drive the TF pathway. The concentration of [TF-VIIa]0 was initially set at 0.25 for venous flows, which is in the range of concentration levels in the numerical study of Kuharsky and Fogelson [16] (estimated to be initially O ( 1 ) n M at shear rate 500 s−1). The concentration of TF-VIIa complex decreases with increasing concentration of fibrin as there will be less binding sites on the vessel subendothelium for the complex. The exact correlation for the variation of [TF-VIIa] is not known, and thus, we assume a cubic function in the form of [TF-VIIa] = [TF-VIIa]0(1 − ([Ia]/[Ia]thr)3).
As introduced above, platelet adhesion and aggregation in blood flow at low shear rates (< 1,000 s−1) may stimulate multiple ligand-receptor interactions, depending on the exposed ECM proteins (but is not strongly dependent on GPIb-vWF binding). We assume that the overall effect of interactions between receptors and ligands is incorporated into the adhesive model of Eqs (9) and (10), with D e l the undetermined parameter.
First, we consider venous thrombus formation and growth similar to the in vivo experiment of Begent and Born [19]. The geometry consists of a straight tube of 50μm diameter and 300μm length as shown in Fig 3a. A parabolic velocity profile is imposed at the inlet with variable average velocities in the range of 100 − 1,000 μm/s, which result in a maximum Reynolds number Re ≈ 0.02, whereas a zero-Neumann velocity boundary condition is imposed at the outlet. To mimic the site of injury and initiate platelet aggregation, we place fixed activated particles (green particles in Fig 3a) uniformly at the bottom of the channel 150 − 180 μm from the inlet. These fixed infinitesimal particles only interact with moving platelets in the blood flow without interfering with the flow field. Fresh platelets (red particles) are inserted at the inlet proportional to the local flow rate with a density of 300,000mm−3, and are removed from the system once they exit the channel.
The snapshots of the developed thrombi are given in Fig 3b–3d for several flow rates, where red particles represent activated platelets that can adhere to the site of injury and blue particles are resting platelets. We also plot λ2 contours on the circular cross-sections located at the middle of clots in Fig 3b–3d. The contours clearly show the elevated shear rates on the thrombus surface upon increasing blood velocity, which lead to disaggregation at higher blood velocities.
The number of platelets in the aggregate at the injured area is recorded for a period of 10 seconds, from which we can calculate the aggregate growth rate. A representative thrombus growth rate is plotted in Fig 4a on a semi-log axes, which shows an initial transient followed by a steady exponential growth of the form ∼exp(αgt), similar to in vivo observations of Begent and Born. After fitting the numerical data, we are able to extract the exponential growth rate αg for different blood flow velocities, which were then normalized by the maximum growth rate and plotted in Fig 4b. Note that, at a lower blood velocity 100 μm/s, aggregation occurs slowly due to the smaller number of platelets transported to the injured region. As blood velocity increases to 400 μm/s, more platelets are delivered to the injured region, contributing to faster growth rate. If blood velocity is increased further to 800 μm/s, the higher shear stresses on the surface of the platelet aggregate limit further aggregation, and thus reduces the growth rate. Similar non-monotonic trends can be observed in the experimental data of Begent and Born, which are extracted from their article and plotted in Fig 4b for comparison. Similarly, Tosenberger et al. [30] observed non-monotone dependence of clot growth rate followed by the clot detachment upon increasing the shear rate. Our numerical values for exponential growth rates are close to the results in Pivkin et al. and [25] Kamada et al. [24], although the magnitude of the exponential growth rates from experiment is several fold higher than from the simulation. There could be a few reasons for this discrepancy, including the mismatch in the size of the injury site and the difference in normal platelet concentration between in vivo experiments and our simulations. We looked at this problem more closely by separately increasing the size of injury to 60μm or increasing the platelet density in our simulations to 500,000mm−3. These additional results are shown in Fig 4(c) along with the original results of Fig 4b. We observe similar trends in all three curves. The effect of increasing the size of injury marginally affects the exponential growth rates, whereas the increase in platelet density increases the exponential growth rates more notably. Another process that could potentially affect the growth rates is the shear-induced platelet drift toward the wall or the thrombus. Although in our numerical scheme particles are inserted close to the walls as suggested by the experimental observations, the particles may be subjected to further margination as they pass through the narrower vessel at the site of injury. We tested this hypothesis by using an empirical equation for enhanced diffusion of platelets at higher shear rates proposed by Wootton et al. [37] (see Eq (6)). The results are plotted in Fig 4(C), which indeed show an increase in the growth rates by 50%.
By adjusting the interaction forces between the platelet particles, we were able to reproduce the dependence of the growth rate on blood velocity reported in [19]. The resulting maximum attractive force applied in the simulation is found to be Fadh,max ≈ 10pN corresponding to D e l ≈ 2 . 1 × 10 - 17 N m.
In atherosclerotic arteries, the presence of plaques generates fluid mechanical conditions that promote high-shear platelet aggregation and thrombus formation [14, 15]. Nesbitt et al. [15] observed that platelet aggregation was predominately in the post-stenosis region and proposed that the aggregation of platelets was resulted from platelet tethering. Westein et al. [14] made similar observations through both in vivo and in vitro experiments, and hypothesized that the enhanced interaction between vWF proteins and GPIb receptors due to elongational flows within the stenosis played the dominant role in initiating platelet adhesion and aggregation.
In order to estimate platelet interaction forces that cause platelet aggregation at elevated shear rates, we first use the data of Westein et al. from a microfluidic device with different degrees of stenosis. A schematic of the simulation domain is shown in Fig 5, where the channel height is 50μm and its depth is 35μm. We consider four different occlusion levels of 20, 40, 60, and 80%. The mean flow velocity at the channel inlet is set as u ¯ i n l e t = 12 m m / s, equivalent to Re ≈ 0.5 and an inlet wall shear rate γ ˙ w = 1 , 000 s - 1 consistent with the microchannel experiment [14]. Fixed and activated particles (green particles in Fig 5) representing vWF, are placed on discontinuous strips on the lower side of the channel wall.
We performed numerical simulations for different occlusion levels to calibrate the platelet-wall and platelet-platelet adhesive forces, which suggested that an approximately two orders of magnitude higher adhesive force is required for platelet aggregation at such elevated shear rates. We plot snapshots of platelets aggregated in the channel at different occlusion levels in Fig 6. In the first column (Fig 6a–6c) we present results where the adhesive forces are increased uniformly (through D e h = 500 D e l), irrespective of the local shear rate magnitude. Based on these snapshots and their related curves for the density of aggregated platelets in Fig 6d, we observe that platelets aggregate inside the stenosis for all geometries and flow conditions, even at 20% occlusion where no aggregation was reported in the experiment of Westein et al. This nonphysical trend necessitates the use of a shear-dependent model for adhesive forces similar to Eq (10). Next, we present snapshots of platelet aggregation simulated using Eq (10) in the second column (Fig 6e–6g) along with their aggregate density curves in Fig 6h. Here, we observe a significant improvement in the results with no aggregation for 20% occlusion, a delayed aggregation for 40% occlusion, and a significant increase in the density of aggregated platelets for 60% stenosis. Altogether, these trends successfully capture the behavior observed in the experiment of Westein et al. [14]
One important trend in the in vivo experimental results of Westein et al. [15] is the enhanced platelet aggregation at the outlet of stenosis compared to its inlet. This effect may be attributed to several factors, including elongation of vWF multimers and enhanced diffusion of agonists at the outlet of the stenosis. To model these effects using the current numerical approach, we introduce a new parameter τact that delays the activation of platelets once stimulated by other activated platelets. This new parameter can be adjusted to control the distribution of aggregated platelets in the stenotic region. Either no or very short delay times will lead to aggregation at the inlet toward the middle parts of stenosis, whereas platelets with properly adjusted activation delays do not become adhesive until they pass the apex of the stenosis. We plot the results of platelet aggregation in an 80% stenosis in Fig 7 with both numerical and experimental platelet density profiles inside the stenosis. We assume that activation delay time is a random number with a uniform distribution and set as τact = 6 ± 3ms for each platelet. Fig 7b reveals that this model produces results similar to the experiment given a shear-dependent adhesive force and activation delay time.
Having the adhesion model calibrated for different flow conditions, we include the coagulation process in blood flowing over a site of injury that expresses tissue factor (TF), which is the primary stimulus for initiation of coagulation. We perform simulations in a circular tube of 50 μm diameter and 350 μm length representing a venule. A cylindrical patch with seeded platelets is placed in the middle of the tube to represent the site of injury, where thrombosis is allowed to initiate (see Fig 9a). We consider the lower range of flow rates that are normally seen in venous blood flows corresponding to a shear rate of 64 s−1. The time course of aggregate density is plotted in Fig 9b; it shows similar exponential growth after an initial transient time (≈ 5s), and in agreement with the experiments of Begent and Born [19].
Our initial numerical observations based on the kinetic rates taken from Anand et al. [18] showed negligible to no thrombin production. This signifies the effect of blood flow on the transport of coagulation reactants away from the site of injury before they can initiate the cascade. Only when we increase each reaction constant by approximately 10 fold, could we observe the production of thrombin mostly downstream of the injury site (see Fig 10a). As shown in the snapshots of Fig 10a, platelets can adhere directly to the exposed collagen and initially form aggregates independent from the coagulation process. As the aggregation grows both radially and axially, blood flow becomes stagnant at the site of aggregation, which in turn, reduces the advective transport of coagulation reactants away from the injury. This can be further seen in Fig 10b and 10c, where thrombin and fibrin concentration profiles are plotted at three different axial locations. The profiles show an almost independent thrombin burst and the subsequent fibrin generation at the center and downstream of the injury, whereas a delayed thrombin burst occurs at t ≈ 12 s proximal to the injury, where platelet aggregation is more pronounced. Further, the concentration profiles of ADP in Fig 10c show an increase as more platelets aggregate and release their granule including ADP. We observe significant oscillation in the concentration profile proximal to the injury as platelets activity and aggregate is higher in that region.
Platelet adhesion occurs via receptor-ligand bindings, but many different receptors and ligands are active under different shear conditions. Specifically, three shear rate regimes have been described: low shear “venous flow” (< 200 s−1), primarily governed by fibrinogen and the GPIIb-IIIa; intermediate shear “arterial flow” (500 − 4,000 s−1), primarily governed by GPIb, GPIIb-IIIa; and high shear “pathologic flow” (> 4,000 s−1) commonly found in diseased, constricted, or stenosed arteries, primarily governed by vWF and GPIb [42, 45]. The binding kinetics are thus diverse and for some integrins not very well characterized, thus inclusion of these details in numerical models will increase their uncertainty as well as the associated computational cost. In this study, our primary objective was to establish a phenomenological shear-dependent model for platelet adhesive dynamics based on the available experimental data for low [19], intermediate [14], and high shear flow [33] conditions. The various quantities reported in these experiments, such as thrombus shape and growth rate as well as platelet aggregate densities, enable us to tune our model for a wide range of shear rates.
We chose a Morse potential to generate the attractive/repulsive forces with a shear-dependent parameter i.e., the strength of the potential D e ∼ f ( γ ˙ ), that is calibrated through Eq (10) for different flow conditions. The repulsive forces rise exponentially for inter-platelet distances less than r < d to prevent cellular overlap. As mentioned in section Materials and Methods, we set the interaction range of the Morse potential βd = 2.5 so that the potential strength De is the only parameter left to be tuned. The adjusted interaction range implies that particles will not induce forces for distances r ⪆ 3d as shown in Fig 1. Further increase in βd is not physiologically correct as the potential and adhesive forces become long-range. Although the present adhesive potential is not capable of directly addressing the kinetics of bond formation/dissociation, it can capture different binding phenomena implicitly due to the effect of local flow conditions and shear rates. The transport velocity of a platelet moving close to the vessel wall is proportional to γ ˙ w meaning that at low shear rates the change in the inter-particle distance r within a time interval Δt is small. Therefore, adhesive forces are stronger representing slow, but strong bonds formed by GPIIb-IIIa. At higher and intermediate shear rates, the energy landscape still remains unchanged. However, faster platelets move a larger distance away from each other leading to weaker adhesive forces, which may represent fast, but weak bonds formed by GPIb-vWF. The maximum value of the bond forces in our model based on the calibrated parameters is ≈ 10 pN, which is in the range of bond forces measured for GPIb-vWF (catch-slip bonds with maximum lifetime at 20 pN [46]), and GPIIb-IIIa-fibrinogen (slip bonds with maximum lifetime at 10 − 20 pN [47]) for which the longest bond lifetimes were observed. Further, two activated platelets in our model can only form one bond with each other, whereas each one in the pair can form multiple bonds with the other platelets in its neighborhood, which may result in the distribution of hydrodynamic drag among several bonds. Under pathologic flow conditions where the shear rates are extremely high, the inter-platelet distance r is most likely to be ≈ 3d, where the same adhesive energy landscape will not be able to slow down or arrest the platelets. Hence, the landscape has to be scaled up with increasing shear rate, which explains the use of D e h in the hyperbolic tangent Eq (10).
Experimental results of Westein et al. [14] allowed model calibration at medium to high shear rates where the maximum wall shear rate at the apex reaches 8,000 s−1. One important finding in their work is the marked increase (between two to three fold) of platelet aggregation post-stenosis. Regardless of the molecular mechanisms that can cause such enhanced aggregation at the following edge of a stenosis, we are able to produce similar trends by introducing a platelet activation delay time parameter, τact. Although there is a physical intrinsic delay in the activation of platelets [28], this parameter is introduced for modeling purposes only; it, too, can be considered as a function of the local blood velocity. Microfluidic experimental results of Li et al. [33] show a different trend, however, where platelet aggregation initiates at the apex with the highest wall shear rate and then spreads to the inlet and outlet of stenosis. We tested our shear-dependent model against their results, and can achieve similar trends and threshold shear rates at which occlusion occurs.
Numerical modeling of thrombus formation and growth is a challenging problem due to multiscale and multiphysics nature of clotting process, which involves fluid mechanics, cell mechanics, and biochemistry. Diverse studies have addressed this problem on different scales such as cellular, meso and continuum levels (e.g., refer to [48–52]) whereas attempts have been made to bridge these different scales to model the process at the initial phase of platelet activation and aggregation (e.g., [53–55]). These studies may be broadly put in three distinct modeling strategies: cellular/sub-cellular modeling of platelet transport and aggregation in whole blood; continuum-based modeling of blood flow treating platelets as Lagrangian particles; and continuum-based modeling of thrombus formation and growth using empirical correlations for platelet deposition rates.
Cellular and multiscale modeling of platelets were used in several studies [28, 30, 48, 51, 53, 54, 56], where the hydrodynamics of blood is resolved and used to model transport of platelets and coagulation enzymes. The kinetic reactions of the coagulation cascade leading to the generation of thrombin and fibrin can be resolved by solving the related advection-diffusion-reaction (ADR) equations. Such detailed models are normally very expensive due to the presence of individual cells and the large set of differential equations related to the biochemistry of coagulation. As a result, they are typically used for mesoscale simulations, and are conducted to explain the relevant microscopic mechanisms and experimental microfluidic observations.
It is possible, however, to reduce the cost of simulations by treating blood and red blood cells as incompressible Newtonian fluid (or non-Newtonian in small arterioles and capillaries), thus leading to continuum fields for blood velocity and pressure and the transport of enzymes, which can be resolved using an Eulerian approach while individual platelets are treated as Lagrangian particles (e.g., refer to [24, 57]). This numerical approach has the advantage of tracking thousands of platelets forming aggregates at the site of injury and effectively capturing the shape and extent of thrombus. Our proposed model based on FCM falls in this category. FCM provides a flexible platform for two-way coupling of platelets (treated as rigid spherical particles) with the background flow. As a result, the thrombus shape modeled by FCM is affected by the local hydrodynamics and fluid stresses. Further, it is possible to introduce porosity to the formed thrombus by adjusting the radius of influence of each particle on the fluid. The major drawback for this kind of approach, however, is the limitation on long-time simulation of large-scale particulate systems for several minutes, which is the physiological time scale of most clotting processes (e.g., thrombosis following the atherosclerosis plaque rupture or aortic dissection).
Several continuum models treat platelets as concentration fields similar to chemical species that follow specific ADR transport equations [17, 38]. These models could also become expensive depending on the number of species considered, and their outputs are generally more prone to uncertainty due to a large set of input parameters. In a recent work, Mehrabadi et al. [11] developed a continuum-based model of thrombus formation using empirical correlations for thrombus growth rate as a function of local shear rate using whole blood experiments over a wide range of experimental shear rates. The model has the advantage of predicting thrombus occlusion time with no significant computational cost using a well-trained model by data extracted from different experiments. However, several contributing factors are neglected, including mechanisms of thrombus formation in a low-shear regime, thrombus mechanics, and embolization. These issues can potentially be addressed by introducing platelets as FCM particles, thus forming a hybrid scheme in which the mechanistic behavior of thrombus formation can be resolved while the continuum model accumulates platelets in the thrombus based on empirical correlations until occlusion has been reached.
Including transport equations for different species involved in the coagulation cascade is crucial for accurate predictions of final thrombus shapes, and is straightforward in the current Eulerian-Lagrangian framework. Our numerical simulations of coupled coagulation and platelet aggregation at lower venous flow rates suggest that initiation of coagulation of flowing blood displays a threshold response to shear rate and to the size of the site of injury. This is mainly due to the competition between coagulation reactions at the site of injury and the advection of species from the injury. Similar threshold response was also observed in the in vitro experiments of Shen et al. for the whole blood flowing on a surface patch coated with TF [34]. Further, our results show that at lower shear rates platelet aggregation and coagulation can occur independently from each other on two isolated spots at the site of injury leading to the enhanced appearance of fibrin monomers and fibrin deposition. Clinically, stasis and low blood flow are considered risk factors for deep vein thrombosis. As shear rate increases in blood flow through arterioles, advective effects become more dominant, which could eliminate thrombin production on the subendothelium. Therefore, the role of heterogeneous coagulation reactions on the surface of adhered platelets would become more crucial to the progression of thrombosis, and must be included in future numerical models.
One of our goals is to improve our understanding of the effects of hemodynamics on the initiation and development of intramural thrombus within a false lumen caused by an aortic dissection. Besides their greater complexities in geometry and flow conditions compared to the microscopic systems considered in this study, the size of aortic dissections are rather large. Therefore, simulations may require hundreds of thousands of FCM particles to represent platelets. Even the computational cost for such lower-fidelity simulations in large domains could become restrictive, and may require additional modeling strategies that will be addressed in future work.
We developed an Eulerian-Lagrangian model to predict thrombus shape and growth, where motions of Lagrangian platelets are coupled with the background blood flow using a force coupling method. Further, platelet adhesion to the site of injury and to each other is modeled by a shear-dependent Morse potential, which is calibrated with experimental data for different shear conditions. Our simulation results show good agreement with experiments for a wide range of shear rates, thus suggesting that the proposed method is suitable for modeling venous thrombosis and embolization as well as thrombosis in arteries.
|
10.1371/journal.ppat.1004856 | An Ultrasensitive Mechanism Regulates Influenza Virus-Induced Inflammation | Influenza viruses present major challenges to public health, evident by the 2009 influenza pandemic. Highly pathogenic influenza virus infections generally coincide with early, high levels of inflammatory cytokines that some studies have suggested may be regulated in a strain-dependent manner. However, a comprehensive characterization of the complex dynamics of the inflammatory response induced by virulent influenza strains is lacking. Here, we applied gene co-expression and nonlinear regression analysis to time-course, microarray data developed from influenza-infected mouse lung to create mathematical models of the host inflammatory response. We found that the dynamics of inflammation-associated gene expression are regulated by an ultrasensitive-like mechanism in which low levels of virus induce minimal gene expression but expression is strongly induced once a threshold virus titer is exceeded. Cytokine assays confirmed that the production of several key inflammatory cytokines, such as interleukin 6 and monocyte chemotactic protein 1, exhibit ultrasensitive behavior. A systematic exploration of the pathways regulating the inflammatory-associated gene response suggests that the molecular origins of this ultrasensitive response mechanism lie within the branch of the Toll-like receptor pathway that regulates STAT1 phosphorylation. This study provides the first evidence of an ultrasensitive mechanism regulating influenza virus-induced inflammation in whole lungs and provides insight into how different virus strains can induce distinct temporal inflammation response profiles. The approach developed here should facilitate the construction of gene regulatory models of other infectious diseases.
| Vaccines suffice for protecting public health against seasonal influenza viruses, but when unexpected strains appear against which the vaccine does not confer protection, alternative treatments are necessary. In this work, we used gene expression and virus growth data from influenza-infected mice to determine how moderate and deadly influenza viruses may invoke unique inflammatory responses and the role these responses play in disease pathology. We found that the relationship between virus growth and the inflammatory response for all viruses tested can be characterized by ultrasensitive response in which the inflammatory response is gated until a threshold concentration of virus is exceeded in the lung after which strong inflammatory gene expression and cytokine production occurs. This finding challenges the notion that deadly influenza viruses invoke unique cytokine and inflammatory responses and provides additional evidence that pathology is regulated by virus load, albeit in a highly nonlinear fashion. These findings suggests immunomodulatory treatments could focus on altering inflammatory response dynamics to improve disease pathology.
| Invading pathogens induce acute inflammation when molecular signatures are detected by pattern recognition receptors (PRRs; e.g., RIG-I like receptors [RLRs] and Toll-like receptors [TLRs]) expressed on tissue-resident immune cells and non-immune cell types. PRR ligation triggers innate immune responses and leads to the induction of inflammatory and antiviral gene expression, which together function to limit pathogen growth, activate the adaptive immune response, and ultimately resolve the infection [1,2]. Precise regulation of PRR-mediated signaling is necessary to both avoid inadvertent tissue damage in response to non-pathogenic stimuli, and to prevent immunopathology resulting from excessive expression of inflammatory molecules. In essence, the ideal inflammatory response must exhibit a balance between appropriate activation against a genuine threat and self-limiting behavior once that threat has been controlled. Despite its importance in maintaining normal tissue homeostasis and limiting pathogen-associated diseases, the mechanisms underlying the regulation of this balance are poorly understood.
Influenza A viruses are recognized by both TLRs and RIG-I-like receptors (RLRs) [3–7], and some strains are potent inducers of inflammatory and antiviral gene expression. Generally, lung tissues infected with pathogenic isolates exhibit high virus titers and robust inflammatory gene expression, as has been documented in in vivo studies with the 1918 Spanish influenza virus [8,9], highly pathogenic H5N1 avian influenza viruses [10–12], and the 2009 H1N1 pandemic influenza virus [13,14]. In contrast, seasonal influenza viruses typically replicate less efficiently, elicit more restrained inflammatory responses, and are usually not associated with lethal infections. Recent evidence has implicated the level of virus replication in infected lung tissues as the primary phenotypic variable driving inflammation and lethal outcomes [15,16]. Other data indicate that influenza viruses that exhibit significant differences in pathogenicity stimulate qualitatively similar host responses that differ primarily at the level of magnitude and kinetics [17]. However, these studies have not revealed the mechanisms that account for the different profile dynamics observed in infections by high and low pathogenic viruses. Such information would aid in clarifying not only how some influenza viruses induce lethal disease, but also the general mechanisms that regulate inflammatory balance.
To characterize the dynamics of influenza virus-induced inflammation, we developed a novel approach to infer gene regulatory models from dynamic gene expression data. Referred to as systems inference microarray analysis, our method builds on current approaches that use co-expression analysis to isolate modules of functional signatures in gene expression data and then extends these methods by fitting the gene expression modules to mathematical equations (models) by using segmented regression analysis. Models can be created to look for strain-dependent responses and, unlike traditional differential expression analysis, to predict gene expression under new experimental conditions. By using this method, we set out to determine how influenza viruses that exhibit variable pathogenicity profiles influence the dynamics of the inflammatory response.
To characterize the dynamics of the host immune response to specific virus isolates, we infected mice with 105 PFU of three virus isolates with distinct pathogenicity profiles and harvested lung tissues at 14 time points after infection (from 0 to 7 days post-infection; n = 3 per virus per time point; see Fig 1) for several parallel analyses. These viruses included a low pathogenicity seasonal H1N1 influenza virus (A/Kawasaki/UTK4/2009 [H1N1]; referred to as ‘H1N1’), a mildly pathogenic virus from the 2009 pandemic season (A/California/04/2009 [H1N1]; referred to as ‘pH1N1’), and a highly pathogenic H5N1 avian influenza virus (A/Vietnam/1203/2004 [H5N1]; referred to as ‘H5N1’). An initial inoculation of 105 PFU was used as previous studies indicated that a high virus dose was needed to invoke different pathologies in H1N1 and pH1N1-infected mice [13]. As expected, lung virus titers (virus titers determined by plaque assay and reported in plaque forming units [PFU] per gram lung; Fig 1) indicated a clear hierarchy of mild, moderate, and severe virus-induced disease. Specifically, the H5N1 virus produced the highest lung titers and between days 5 and 7 post-infection, this virus also caused mortality in the animals whose lungs were to be collected on day 7 post-infection (i.e., ‘severe’ disease). In contrast, all animals infected with the H1N1 or pH1N1 viruses survived the duration of the time course study; however, pH1N1-infected animals were visibly sicker and exhibited higher lung titers relative to those infected with H1N1 at all time points observed after the first 30 h post-infection (i.e., ‘moderate’ and ‘mild’ disease, respectively). Histopathological analysis of tissue samples collected on days 1, 2, and 5 post-infection (Fig 2) also showed that H5N1-infected tissue exhibited the earliest, most severe signs of inflammation and inflammatory immune cell infiltrates followed by pH1N1-infected tissue, whereas H1N1-infected tissue showed mild evidence of inflammation and was most similar to tissue from the control mice (mock-infected mice).
We next used co-expression analysis to integrate inflammation-associated gene expression differences between influenza-infected and control lungs into a systems level context. We first asked whether the expression of inflammation-associated genes clustered into modules of co-expressed genes. Tissues from the same animals that were used to determine virus growth were used to evaluate changes in global lung transcriptional profiles. A total of 168 microarrays were developed (three per time point for H1N1-infected, pH1N1-infected, H5N1-infected and control mice). One microarray was removed after reviewing replicate quality. After filtering transcripts for minimally confident variation (we required at least one time-matched, infected condition compared with mock-infected absolute fold change ≥ 2 and a false discovery rate [FDR]-adjusted P-value < 0.01), the log2 of the normalized intensity of the retained transcripts (16,063) for all 167 samples were then clustered by using the Weighted Gene Co-expression Network Analysis (WGCNA) algorithm [18]. In all, 45 distinct co-expression modules were identified (referred to as N1, N2, etc.; S1 Table provides the module assignments for all transcripts). To identify the biological role of the host response modules, we performed functional enrichment analysis on each gene module by using DAVID [19] and ToppCluster [20]. Because each module was comprised of positively and negatively correlated transcripts, we used the module eigengene (i.e., the first principle component of the gene expression matrix) to divide each module into two submodules containing transcripts that were positively or negatively correlated with the parent module’s eigengene, denoted as kME+ and kME- (referred to as module membership), respectively. This procedure allowed us to look for biological processes with similar but opposing dynamic responses to the virus infections. Functional enrichment analyses were then applied to each submodule by using two bioformatics platforms to ensure robust results.
We identified two submodules (N1 kME+ and N2 kME-, referred to as simply N1 and N2 in the remainder of the text) that were enriched for inflammatory response and inflammation-associated pathway signatures by using both bioinformatics platforms, and these two modules became the focus of our study (Table 1 summarizes the functional enrichment results for the immune and inflammatory related annotations. The complete enrichment results from ToppCluster and DAVID are available in S2 Table and S1 and S2 Files). The N1 module was uniquely enriched for cytokine activity and type I interferon (IFN) regulating TLR and RLR pathways [21], as well as transcriptional signatures associated with IFN-regulated activity (i.e., the transcription factor binding sites [TFBS] of Irf1, Irf7, Irf2, ISRE, and NF-κB). Additionally, N1 was the only module that exhibited significant enrichment with a compendium of established IFN-stimulated genes (‘ISGs’; Table 1; see Methods. The list of ISGs is available in S3 File). A more recent study identified 147 IFN stimulated genes in immortalized, human airway epithelial (Calu3) cells [22]. Of these, 90 mouse homologs were annotated on the microarrays and 70 of the homolog probes were assigned to the N1 module (Fisher’s exact test; P-value < 10–16; odds ratio = 36.4), further associating N1 interferon-stimulated gene activity.
In contrast, the N2 module was only weakly enriched for some cytokine activity related annotations and not enriched for any of the binding sequences of transcription factors that are members of canonical inflammatory pathways (such as interferons, interferon regulatory factor proteins, or NFκB). Instead, it was primarily associated with several annotations related to leukocyte and lymphocyte activity (see summary of ToppCluster enrichment results in Table 1; see S1 and S2 Files). Further analysis with CTen [23], a platform for associating clustered gene expression data with specific cell types, found N2 to be highly enriched for genes expressed in macrophages in various cellular states (e.g., bone marrow-derived macrophages exposed to lipopolysaccharide [LPS]) (Fig 3A; additional details available in S3 Table). The remaining immune-associated submodules (the kME+ N22, N25, N31, and N35 submodules; described in Table 1) were enriched for several key immune processes such as antigen presentation, and T cell and natural killer (NK) cell activity, but their further assessment would be beyond our focus and the scope of this study. Thus, bioinformatics analyses robustly associated the N1 module with inflammation, cytokine production, and type I IFN pathway activity—likely activated in resident lung cells—whereas the N2 module is associated with migration and activation of macrophages in the lung.
A closer examination of the expression dynamics of each of the inflammation response-associated modules revealed patterns of expression that were consistent with the biological roles predicted by our bioinformatics analyses. We used the scaled difference of the module eigengene to characterize the expression of all genes within each module. We subtracted the mean of the eigengene of the control samples from the eigengene of time-matched, virus-infected samples and then divided by the largest observed average difference across all conditions. The resulting scaled difference eigengene (SDE) represents the fraction of the maximum log fold change in gene expression observed across all experimental conditions (Fig 3B and 3C. S1 Fig briefly illustrates the eigengene scaling in greater detail. S4 Table provides a heatmap of the gene expression of all probes in each module).
Within the inflammation-associated modules (N1 and N2), the H5N1 virus induced the earliest gene expression changes and the highest peak expression levels, corroborating previous observations that H5N1 viruses are strong inducers of inflammatory and IFN response signaling in vivo [10,24,25]. Consistent with the prediction that N1 is involved in detecting virus in infected tissues, the N1 module eigengene was the most highly correlated with virus titer (Pearson pairwise correlation, ρ = 0.70).
Module gene expression dynamics further suggested that the N2 module gene is associated with lymphocyte infiltration. Exudate macrophages [26] and neutrophil [27] have been identified as factors of severe disease during influenza infection. To associate gene expression dynamics with changes of immune cell counts, a new population of mice were infected with the three influenza viruses, five mice per infection group were sacrificed on days 1, 2, 3 and 7 and the changes in the number of macrophages and neutrophils was assessed (see Methods). Unlike the previous study by Brandes, et al. [27], strong neutrophil infiltration was not specific to fatal infections but, instead, infiltration of both cell types had a clear hierarchical relationship with the severity of the infection (Fig 3D and 3E). The N2 module eigengene exhibited a lesser correlation to virus titer (ρ = 0.55), but was tightly correlated to macrophage influx into the lung (ρ = 0.90; P-value < 0.01 Student’s t-test). Of the 45 module identified in the studied, N2 had the highest correlation to both macrophage and neutrophil influx (the correlation of macrophage and neutrophil influx and all 45 module eigengenes are provided in S5 Table; ρ = 0.80; P-value < 0.01). The N1 module on the other hand was weakly but significantly correlated with immune cell infiltration (ρ = 0.67 [P-value = 0.03] for neutrophils; ρ = 0.60 [P-value = 0.05] for macrophages), but its eigengene was not the most highly correlated (5 other module eigengenes had a greater absolute correlation). These results further associate N2 with immune cell-specifically macrophage- infiltration, while the sum of the bioinformatics, virus titer correlations and immune cell infiltration evidences associated N1 with inflammation and type I IFN pathway activity.
A further advantage of a network approach is that the functional relevance of genes might be inferred from their positions within the co-expression network [28]. We used the module membership (the correlation between the gene’s expression and the module eigengene, kME) to isolate potential regulators of the N1 module. Among the top intramodular hub genes (i.e., genes with the highest module memberships, see S4 Table), we found Mnda, Herc6 and Cd274 and several interferon regulated, virus replication inhibitory genes such as Oas2 and Oas3 [29]. Herc6 is involved in ubiquitination [30]. Mnda is significantly up-regulated in monocytes exposed to interferon α [31] while Cd274 is a transmembrane protein expressed on antigen presenting cells and modulates activation of T cells, B cells and myeloid cells. We also observed that interferon stimulated genes tended to have higher intramodular hub rankings, suggesting a regulatory role for interferon (Wilcoxon rank sum test, P-value < 10–12). We then considered the module membership rankings of transcription factors known to regulate interferon. Of the established interferon regulatory factors that are members of the N1 module (e.g., Irf1, Irf2, Irf7, Irf9, Stat1, and Stat2. Nfkb1 and Nfkb2 were not assigned to N1), Irf1 and Irf7 had the highest module memberships (kME = 0.94 and 0.93; ranking = 118 and 225, respectively). Irf7 expression was also several orders of magnitude greater than Irf1 (S4 Table). Together, these findings corroborate our bioinformatics analyses by suggesting that N1 is regulated by interferon, and N1 expression likely results in enhanced cytopatchic effects and regulation of the lymphocyte immune response. Network analysis further suggests that Irf7 may play a regulatory role upstream of interferon transcription.
Previous studies have suggested that highly pathogenic influenza virus infections induce an irregular or disproportionate inflammatory response relative to seasonal influenza viruses, and that these differences occur early in the host response [24]. For this reason, we sought to further explore the possibility of isolate-specific or isolate-independent response patterns of the cytokine-associated N1 module. We wanted to infer mathematical relationships that could describe when inflammatory-associated gene expression occurs and what magnitude of expression is expected. By using the eigengene as a representation of the scaled gene expression dynamics, we attempted to infer simple mathematical models that can be related to common signaling mechanisms.
Surprisingly, when we plotted the N1 SDE for each isolate against the corresponding virus titer, we observed a consistent profile for all three viruses; regardless of intrinsic virulence, the fold change in N1 gene expression remained initially low and rapidly increased only after a virus titer of approximately ~108 PFU/g (of lung) was reached (Fig 4A). Following activation, N1 gene expression increased as a function of virus concentration at the same apparent rate for all infection conditions, and more complicated dynamics were observed only during the later phase of the infection when virus clearance was observed (i.e., when the virus titers began to decrease). These observations suggest that IFN-regulated (N1) gene expression was induced by an ultrasensitive response mechanism controlled at the level of the virus titer rather than the virus’s intrinsic virulence.
Ultrasensitive responses characterize the dynamics of several signaling pathways that regulate essential and often toxic biological processes such as the cell cycle [32] and apoptosis [33]. As shown in Fig 4B, ultrasensitive responses are typified by an attenuated response to low levels of stimulation but a strong response occurs once a threshold level of stimulus is reached. Cooperativity [33] and positive feedback [34] are two mechanisms that produce ultrasensitive responses. To formalize the hypothesis that the inflammatory gene response follows an ultrasensitive response profile, we selected a segmented linear model (SLM, defined in Fig 4C) to be a simplified representation of the ordinary differential equations normally used to model ultrasensitive responses, and we fit the N1 SDE to an SLM that was strictly a function of the virus titer. The optimal fit showed a threshold of 107.78±0.14 PFU/g is required for N1 module activation to occur, after which the SDE’s rate of activation (a2) was 0.5±0.07 log10(PFU/g)-1 with an intercept (b2) of -3.8 (unitless) (see the Methods and S2 Fig for additional details). Below this threshold, the model predicted minimal gene expression (a1 = 0.17; b1 = 0.03). The SLM goodness of fit on the training data was an adjusted R2 = 0.72 while an adjusted R2 = 0.41 was observed when the data was fit to a linear model. A Davie’s test confirmed that a segmented model was a significantly better fit than a linear correlation model (P-value < 2.2e-16). While the H1N1-infected lung tissue did not exceed an average peak virus titer of 107.4 PFU/g (peak titer occurs at 48 hpi in Fig 4A), we observed increased transcriptional activity in H1N1-infected mouse lung tissues after this time point, suggesting either that the actual peak virus titer occurred between 48 hpi and the subsequent time point (60 hpi), or that the model-predicted threshold was slightly over-approximated.
We next sought to validate the threshold model by attempting to predict cytokine-associated gene expression in influenza virus-infected lung when only the virus titers are known. For this, we selected the H5N1 virus, which has previously been associated with an excessive cytokine response [10]. We infected mice with 103 PFU of the H5N1 virus (a 100-fold reduced dose compared with that used in the experiments to fit the model), determined lung virus titers at the same time points used for the initial experiment (S3 Fig), and then evaluated the segmented linear model’s ability to predict cytokine-associated gene expression. First, we confirmed that the original transcripts assigned to the N1 module were again co-expressed, and thus we used the same transcripts originally assigned to the N1 module to determine the eigengene (see S4 Fig for an analysis of the conservation of the N1 module between the two experiments). In this experiment, as expected, the peak average virus titer (109.3±0.21 PFU/g) for the 103 PFU dose was delayed compared with that for the 105 PFU dose (S3 Fig, compare to Fig 1). Moreover, the SDE exhibited an 18-h delay in activation and a 42-h delay in peak expression compared with the 105 PFU N1 eigengene (Fig 4D). Importantly, based on the virus titers alone, the fitted segmented linear model accurately predicted N1-like SDE behavior (R2 = 0.71 for all time points and R2 = 0.87 for time points up to peak expression; Fig 4E and S5 Fig), and this could be further demonstrated at the individual gene level for specific N1-associated transcripts (e.g., Herc6, Stat1 and Irf7; Fig 4F). These observations provide strong evidence that activation of inflammatory-associated gene expression is dictated by a specific virus concentration in infected tissue, and further suggest the novel possibility that the pulmonary innate inflammatory response has a nonlinear, ultrasensitive-like activation profile that promotes tolerance to low concentrations of virus.
Although transcriptional activation of IFN-stimulated and inflammatory gene expression is a reasonable measure of the effects of inflammation response stimulation, we reasoned that a bona fide ultrasensitive mechanism that regulates this response should be reflected in other aspects of the associated signaling pathway(s). Indeed, of the 17 cytokines associated with the N1 module, 15—including key inflammatory proteins, such as interleukin 6 (IL-6) and monocyte chemoattractant protein-1 (MCP-1)—were significantly correlated with the N1 module eigengene (Pearson’s ρ≥0.5; FDR-adjusted P-value < 0.01; see S6 Table). In addition, when the protein expression levels of these 15 cytokines were plotted against the corresponding titer data, we observed dynamics similar to that of the inflammatory-associated N1 gene module. Initially, protein expression was low but strongly increased only after the virus titers exceeded the threshold of ~108 PFU/gram determined in the gene regulation model (Fig 5A). In contrast, most of the protein levels of the other measured cytokines with transcripts that were not assigned to the N1 module did not show any obvious relationship to the proposed threshold response (S6 Fig). The only major exceptions were LIF, RANTES, and IL18 (S6 Fig). LIF’s gene transcript was not annotated on the arrays whereas IL18’s transcript was not identified as differentially expressed and therefore was not included in the clustering study. The RANTES transcript was included in the clustering study and assigned by the WGCNA algorithm to N2 although the transcript’s correlation to the N1 eigengene suggests it could also have been assigned to the N1 module (Pearson correlation = 0.84 and 0.89 to the N1 and N2 eigengene, respectively). Some proteins appeared to conform to the threshold model in pH1N1-infected, but not H1N1- or H5N1-infected, mice. These may be cytokines that have strain-dependent responses and do not conform to the model. Overall, changes in N1-associated cytokine protein levels in influenza virus-infected mouse lungs were consistent with the proposed virus-titer regulated threshold-mechanism underlying the IFN-mediated response.
We then searched for evidence of threshold-like behavior in the upstream signaling events leading to the activation of IFN-mediated, inflammation-associated gene expression: namely IFN-α/β protein expression, and IRF3 and STAT1 transcription factor phosphorylation (Figs 5B, S7 and S8. Images of representative blots are available in S9 Fig). Significant increases in the concentration of IFN-α and phosphorylated STAT1 (pSTAT1) were detectable in infections with all three virus isolates and occurred at time points after the threshold level of virus was exceeded (Fig 5B). For the H1N1 data, significant levels of pSTAT1 were observed at 36 hpi which—as noted previously—corresponds to the time immediately after the virus titers in H1N1-infected mice reached their peak (see previous discussion). On the other hand, significant increases in phosphorylated IRF3 (pIRF3) were observed only in pH1N1 and H5N1 infections, whereas significant increases in IFN-β were observed only in the H5N1 infection. Changes in the levels of IFN-β and pIRF3 occurred after significant increases in pSTAT1 and IFN-α occurred. As such, the change in the percentage of pSTAT1 was more closely correlated to the N1 eigengene than was that of pIRF3 (correlation = 0.77±0.06 and 0.67±0.09 respectively), and significant increases in pSTAT1 corresponded to time points at which the mean virus titer exceeded the threshold level identified in the gene expression analysis (~107.78 PFU/g). The greater correlation of IFNα and STAT1 activation to the inflammation-associated, N1 module’s gene expression dynamics and the enrichment of the N1 module for the IRF7 binding sequence (Table 1) suggest that the primary driver of the threshold-regulated, inflammatory gene response originates along the IRF7 → IFN-α → STAT1 axis (Fig 5C).
Our data reveal that the activation of the IFN-associated inflammatory and antiviral response program in influenza-infected mouse lung is characterized by an ultrasensitive response driven by the virus load. The power of the threshold model is illustrated by its ability to accurately predict gene expression in infected mice, and the data further suggest that the molecular basis of threshold behavior originates upstream of IFN-α production. Threshold-like and ultrasensitive mechanisms are hypothesized to be necessary for effective management of critical cellular machinery in noisy environments, and are recognized players in the activation of the cell cycle [34], mitogen-activated protein kinase signaling [35], and apoptosis [33,36]. However, while a role for threshold behaviors have been postulated to be essential for filtering noise or errant signaling in complex biomolecular environments [35], our study is the first to directly link threshold-like behavior to the virus-induced innate immune response.
The ultrasensitive response observed in this study provides additional insight into the mechanisms that drive severe pathologies during influenza infection. Several works have suggested that viral load is a key determinant of pathology [12,15] while other works suggest that highly pathogenic influenza viruses induce early, strong inflammatory responses that are independent of the viral load [9,24,37]. Recently, it was observed that fatal influenza infections in mice coincide with a strong influx of neutrophils in what the author’s describe as a viral load-independent, “feedforward” inflammatory circuit [27]. The ultrasensitive response suggested by our study consolidates these hypotheses by suggesting that viral load drives cytokine production (and in turn immune cell infiltration) in a nonlinear manner which is capable of producing states of high and low innate immune responses. Characterization of key aspects of the inflammatory response, such as the onset and peak inflammatory gene expression, require a high temporal resolution of the virus growth and host response dynamics; an experimental design that was unique to our study. The ultrasensitive response model does not negate the significance of neutrophil infiltration [27] in determining fatal infections but suggests that viral load drives the high and low innate immune states. The observed threshold may represent the transition to immunopathology; as indicated by the histopathology results (Fig 2). Moreover, the influenza virus’ NS1 protein is crucial for inhibiting the interferon-mediated antiviral response [38]. The NS1 protein of three viruses used in this study have the SUMO1 acceptor site that indicates interferon antagonism capability [39–41]. Additional studies with NS1-mutated viruses and other pathogens may better reveal strain-dependencies for the observed thresholding behavior.
The ultrasensitive response further suggests that the innate immune response has a limited capacity to respond to influenza virus infection and supports the development of immunomodulatory therapies. Interestingly, after the threshold was exceeded, the rate of activation for inflammatory and interferon-associated gene expression (N1) was conserved for a moderately pathogenic and deadly viruses (Fig 4A). The conserved rate of activation implies that the immune response detects the virus concentration but not the virus growth rate; suggesting the innate immune response is naturally limited in its ability to respond to high growth influenza viruses. Additionally, studies in knockout mice indicate that type I IFN-associated pathways are essential for protection during primary infection [42] and that earlier initiation of these pathways coincides with increased survival in mice infected with highly pathogenic isolates [43]. In combination with these studies, the findings here suggest a novel means of protecting high risk groups by treating them with compounds that target the molecular mechanisms responsible for the threshold behavior. Lowering the threshold required to induce the cytokine response may be a means of providing protection from severe influenza infection. Since these compounds would target host proteins, such treatments would be effective against various influenza virus strains. Data from the viruses studied here suggest that post-threshold, inflammatory gene expression primarily reflects interferon-regulated tissue damage, but time-course data from additional highly pathogenic viruses are needed to assess the degree to which interferon activity is associated with virus growth suppression.
The A/California/04/09 H1N1 virus (pH1N1) was received from the Centers for Disease Control and Prevention. The A/Kawasaki/UTK-4/09 H1N1 virus (H1N1) served as a reference for a seasonal influenza, whereas a fatal human isolate, A/Vietnam/1203/04 H5N1 virus (H5N1), served a highly pathogenic virus.
All mouse experience were performed in accordance to the University of Tokyo's Regulations for Animal Care and Use. These regulations were approved by the Animal Experiment Committee of the Institute of Medical Science, the University of Tokyo (approval number: PA10-13). The committee acknowledged and deemed acceptable the legal and ethical responsibilities for the animals, as detailed in the Fundamental Guidelines for Proper Conduct of Animal Experiment and Related Activities in Academic Research Institutions under the jurisdiction of the Ministry of Education, Culture, Sports, Science and Technology, 2006.
All experiments with H5N1 viruses were performed in biosafety level 3 containment laboratories at the University of Tokyo, which are approved by the Ministry of Agriculture, Forestry, and Fisheries, Japan.
Five-week old C57BL/6J, female mice were obtained from Japan SLC. For all experiments, mice were anesthetized with isoflurane and intranasally inoculated with either 103 or 105 PFU of virus. Initially, 42 mice were inoculated with 105 PFU of the H1N1, pH1N1, or H5N1 virus or mock-infected with PBS (a total of 168 mice). At 14 time points, 3 mice per group were humanely sacrificed and their lungs harvested. The lungs were sectioned and used to assess virus titers (right-upper lobe), cytokine levels (right lower), and the initial gene expression that was used to train the segmented linear model (left-lower). Separately, 42 mice were infected with 103 PFU of the H5N1, sacrificed at the same 14 time points, and lungs sections obtained as described previously to provide the model validation data. The same inoculation method was used for all mice in this study. The numbers of mice used for flow cytometry, western blot, and interferon protein assay experiments are specific in the corresponding sections.
For cytokine and chemokine measurements, mouse lungs were treated with the Bio-Plex Cell Lysis Kit (Bio-Rad laboratories, Hercules, CA) according to the manufacturer’s instructions. Concentrations of other cytokines were determined with the Bio-Plex Mouse Cytokine 23-Plex and 9-Plex panels (Bio-Rad laboratories) Array analysis was performed by using the Bio-Plex Protein Array system (Bio-Rad laboratories).
Virus titers were determined by plaque assay using MDCK cells.
Mouse lung tissues were placed in RNALater (Ambion, CA), an RNA stabilization reagent and stored at -80°C. All tissues were thawed together and homogenized for 2 minutes at 30 Hz using a TissueLyser (Qiagen, Hilden, Germany) as per the manufacturer’s instructions. From the homogenized lung tissues, total RNA was extracted with the RNeasy Mini Kit (Qiagen, Hilden, Germany) in accordance with the manufacturer’s instructions. Cy3-labeled cRNA preparations were hybridized onto Agilent-014868 Whole Mouse Genome 4x44K microarrays for 17 h at 65°C. Feature Extraction Software version 7 (Agilent Technologies) was used for image analysis and data extraction, and Takara Bio provided whole array quality control metrics.
Differential expression was assessed by using a linear regression model. By using the limma package [44] 26 version 3.14.1 from BioConductor, probe intensities were background corrected by using the “norm-exponential” method, normalized between arrays (using the quantile method), and averaged over unique probes IDs. Replicate quality was assessed using hierarchical clustering, resulting in the removal of a single array (the array corresponded to a sample collected at 3 hpi in H5N1-infected mice.) Probe intensities were fit to a linear model that compared data from infected samples to time-matched data collected from uninfected mice. Probes were annotated by matching to the probe names in the mgug4122a version 2.1 mouse annotation database available from BioConductor. All arrays in this study have been deposited on the GEO Expression Omnibus (GSE63786).
Unsigned co-expression networks were constructed by using the blockwiseModules program from the WGCNA package version 1.23.1 [18] in R. The analysis was performed with several different parameterizations to ensure robust clustering. For the results reported in this text, we removed all probes that did not have a confident fold-change greater than 2 (FDR < 0.01) for at least one infected-tissue to control-tissue, time-matched comparison. We then clustered the log2 of the normalized intensities for all 167 microarrays (corresponding the three samples for each time-point for each infected or control population with the exception of data from H5N1-infected mice at 3 hpi which had two samples). Based on the scale-free topology characteristics curve, a power of n = 7 with no reassignment after clustering (reassignThreshold = 0) and a maximum cluster size of 6000 probes was used. We generally observed that allowing gene reassignment between the modules led to poorer clustering based on the distribution the gene’s module memberships (the correlation between a gene and the eigengene of the module to which it had been assigned. S10 Fig illustrates the distribution of the gene kMEs for modules N1 and N2). We then repeated the clustering using different powers (ranging from 7 to 11), allowing different cluster sizes, different subsets of the expression data (e.g., clustering data from each infection separately or together), or relaxing the differential expression condition. In all clusterings performed, the two modules discussed in the text were identifiable. Fisher exact tests between each clustering run were used to determine whether the initial modules were significantly conserved under different parameterizations.
We also considered if the N1 module would be isolated when using signed versus unsigned network construction. We constructed a signed co-expression network and found that 92% of the kME+ N1 genes are again clustered and confirmed that the gene expression dynamics were maintained (see S11 Fig).
ToppCluster [20] and DAVID [19] were used for gene ontology and pathway enrichment, and ToppCluster was also used for transcription factor binding site enrichment analyses. DAVID uses clusters of related annotations constructed from several annotation databases (e.g., pathway and gene ontology annotations) to determine the function of a set of genes and scores the enrichment by averaging the unadjusted P-values (determined by Fisher’s Exact test) of the annotations within the cluster. ToppCluster uses hypergeometric tests to determine the enrichment between a set of genes and gene lists contained in 18 categories (databases) detailed in the ToppGene Suite [45]. The databases include cis-regulatory motif data [46,47], referred to as transcription factor binding sites (TFBS) in the text. Both tools were used with their default settings and the gene universe was considered to be all annotated mouse genes. For each module, we considered the enrichment of all genes assigned to the module and the kME+ and kME- subsets. Generally, the enrichment analysis of the whole module gene set reiterated the enrichment results of the kME+ and kME- subsets albeit with slightly lower but still significant enrichment. Since both tools returned similar GO and pathway enrichment results, we summarized the functional and pathway enrichment results in Table 1 using the results from DAVID.
The enrichment of interferon stimulated genes was determined by using a list of interferon stimulated genes from the Interferon Stimulated Gene Database [48] that was downloaded on May 9, 2012 (see S3 File). For each module, all module genes and the kME+ and kME- subsets were tested for enrichment using Fisher’s exact test in R. The p values were adjusted to control the false discovery rate.
CTen [23] was used to determine enriched cell signatures in select co-expression modules. The enrichment score reported is the—log10 of the false discovery rate.
Model fitting and validation was performed in R using the ‘segmented’ package [49].
Five mice per time point per infection were infected with 105 PFU of the described virus. Five uninfected (naïve) mice served as negative controls. Whole lungs were collected from mice, and incubated with Collagenase D (Roche Diagnostics; final concentration: 2 μg/mL) and DNase I (Worthington; final concentration: 40 U/mL) for 30 minutes at 37°C. Single-cell suspensions were obtained from lungs by grinding tissues through a nylon filter (BD Biosciences). Red blood cells (RBCs) in a sample were lysed with RBC lysis buffer (Sigma). Samples were resuspended with PBS containing 2 mM EDTA and 0.5% bovine serum albumin (BSA), and cell number was determined by using a disposable cell counter (OneCell). To block nonspecific binding of antibodies mediated by Fc receptors, cells were incubated with purified anti-mouse CD16/32 (Fc Block, BD Biosciences). Cells were stained with appropriate combinations of fluorescent antibodies to analyze the population of each immune cell subset. The anti-F4/80 (BM8; eBioscience) antibodies were used. All samples were also incubated with 7-aminoactinomycin D (Via-Probe, BD Biosciences) for dead cell exclusion. Data from labeled cells were acquired on a FACSAria II (BD Biosciences) and analyzed with FlowJo software version 9.3.1 (Tree Star).
Three mice per time point per infection group were infected with 105 PFU of the described virus. The primary antibodies of mouse anti-STAT1 (phospho Tyr701) mAb (ab29045, abcam), rabbit anti-IRF3 (phospho Ser396) mAb (4947, Cell Signaling), and mouse anti–β-actin (A2228; Sigma-Aldrich) were used; the secondary antibodies were HRP-conjugated anti-mouse IgG antibody (GE Healthcare) and HRP-conjugated anti-rabbit IgG antibody (GE Healthcare). Mouse lungs were collected and homogenized with RIPA buffer (Thermo Scientific, Rockford, IL, USA) containing proteinase inhibitor (Roche, Mannheim, Germany) and phosphatase inhibitor cocktails (Sigma-Aldrich, Saint Louis, Missouri, USA). The lysates were then briefly sonicated and centrifuged. Each sample was electrophoresed on sodium dodecylsulfate polyacrylamide gels (Bio-Rad Laboratories, Hercules, CA, USA) and transferred to a PVDF membrane (Millipore, Billerica, MA, USA). The membranes were blocked with Blocking One (Nacalai Tesque, Kyoto, Japan) for 30 min at room temperature, and then were incubated with the primary antibodies overnight at 4° C, followed by the secondary antibodies. They were then washed 3 times with PBS plus Tween 20 (PBS-T) for 5 min and incubated with secondary HRP-conjugated antibodies (as described above) for 30 min at room temperature, followed by three washes with PBST. Specific proteins were detected by using SuperSignal West Femto Maximum Sensitivity Substrate (Thermo Scientific, Rockford, IL, USA). Photography and quantification of band intensity were conducted with the VersaDoc Imaging System (Bio-Rad Laboratories, Hercules, CA, USA). The quantity of target bands from each sample was standardized by their respective β-actin.
Three mice per time point per infection group were infected with 105 PFU of the described virus. Half the lung of each mouse was dissolved in 1 mL of RIPA buffer. We measured the Interferon-alpha and Interferon-beta by using ELISA kits (#12100, #42400, PBL Assay Science, NJ, USA) according to the manufacturer’s instructions. Plates were read at an absorbance of 450 nm using a Versa Max plate reader (MolecularDevices, Menlo Park, CA).
Additional gene set overlap tests were performed in R with all of the genes annotated on the array as the reference (background) set. Statistical tests to compare means within the western blot, flow cytometry, immune cell count and protein assay data sets were performed in R using the ‘multcomp’ package [50].
|
10.1371/journal.pgen.1004455 | The EJC Binding and Dissociating Activity of PYM Is Regulated in Drosophila | In eukaryotes, RNA processing events in the nucleus influence the fate of transcripts in the cytoplasm. The multi-protein exon junction complex (EJC) associates with mRNAs concomitant with splicing in the nucleus and plays important roles in export, translation, surveillance and localization of mRNAs in the cytoplasm. In mammalian cells, the ribosome associated protein PYM (HsPYM) binds the Y14-Mago heterodimer moiety of the EJC core, and disassembles EJCs, presumably during the pioneer round of translation. However, the significance of the association of the EJC with mRNAs in a physiological context has not been tested and the function of PYM in vivo remains unknown. Here we address PYM function in Drosophila, where the EJC core proteins are genetically required for oskar mRNA localization during oogenesis. We provide evidence that the EJC binds oskar mRNA in vivo. Using an in vivo transgenic approach, we show that elevated amounts of the Drosophila PYM (DmPYM) N-terminus during oogenesis cause dissociation of EJCs from oskar RNA, resulting in its mislocalization and consequent female sterility. We find that, in contrast to HsPYM, DmPYM does not interact with the small ribosomal subunit and dismantles EJCs in a translation-independent manner upon over-expression. Biochemical analysis shows that formation of the PYM-Y14-Mago ternary complex is modulated by the PYM C-terminus revealing that DmPYM function is regulated in vivo. Furthermore, we find that whereas under normal conditions DmPYM is dispensable, its loss of function is lethal to flies with reduced y14 or mago gene dosage. Our analysis demonstrates that the amount of DmPYM relative to the EJC proteins is critical for viability and fertility. This, together with the fact that the EJC-disassembly activity of DmPYM is regulated, implicates PYM as an effector of EJC homeostasis in vivo.
| The multi-protein exon junction complex (EJC) is deposited at exon-exon junctions on mRNAs upon splicing. EJCs, with Y14, Mago, eIF4AIII and Barentsz proteins at their core, are landmarks of the nuclear history of RNAs and play important roles in their post-transcriptional regulation. In mammalian cells, the Y14-Mago interacting protein PYM associates with ribosomes and disassembles EJCs in the cytoplasm. However, the physiological function of PYM and its regulation in vivo remains unknown. We have analysed PYM function during Drosophila oogenesis, where the EJC is essential for oskar mRNA localization in the oocyte, a prerequisite for embryonic patterning and germline formation. We find that Drosophila PYM interacts with Y14-Mago but, in contrast to mammalian PYM, does not bind ribosomes. We demonstrate that EJCs associated with oskar mRNA in vivo are disassembled by PYM over-expression in a translation-independent manner, causing oskar mislocalization. Our in vivo analysis shows that the Drosophila PYM C-terminal domain modulates PYM-Y14-Mago interaction, revealing that PYM is regulated in Drosophila. Furthermore, PYM is essential for viability of flies lacking one functional copy of y14 or mago, supporting a role of PYM in EJC homeostasis. Our results highlight a distinct mode of regulation of the EJC-dissociating protein PYM in Drosophila.
| In eukaryotes, post-transcriptional regulation of gene expression plays important roles in development and differentiation. These include RNA processing events in the nucleus, such as splicing, which also affects 3′ end processing of the RNA, mRNA export, localization, translational enhancement and decay [1]–[5]. The multi-protein exon junction complex (EJC), which is recruited to RNAs upon splicing, has been linked to most of these steps in RNA maturation. The EJC assembles 20–24 nucleotides (nt) upstream of splice junctions and is organized around a core complex of four proteins: the DEAD box RNA helicase eIF4AIII, which is deposited by the spliceosomal protein CWC22 [6]–[8] and binds the mRNA independently of its sequence, the Y14 (Tsunagi)-MAGOH (Mago nashi, Mago) heterodimer, which stabilizes the complex, and MLN51 (Barentsz, Btz), which associates with the EJC upon RNA export [9].
In Drosophila, asymmetric localization of several key mRNAs during oogenesis is essential for embryonic patterning [10]. While in transport, these mRNAs are translationally repressed, and protein is produced only upon mRNA localization and at a particular developmental stage. The localization of oskar mRNA to the posterior pole of the oocyte requires splicing and the EJC core proteins [4], [11]–[15], indicating that nuclear events determine mRNA targeting within the cytoplasm. However, in vivo association of an assembled EJC with oskar has not been shown and the basis for the requirement of the complex in RNA transport remains unclear.
Partner of Y14-Mago (PYM) was identified through its association with the Y14-Mago heterodimer in Drosophila S2 cells [16]. The crystal structure of the PYM-Y14-Mago trimeric complex revealed that the PYM N-terminal residues are necessary for its interaction with Y14-Mago [17]; in mammals, this interaction can provoke disassembly of the EJCs from spliced mRNAs [18]. Furthermore, in HeLa cells, the PYM C-terminus, which bears similarity to eIF2A, associates with the 40S ribosomal subunit in the cytoplasm [19]. These observations led to the proposal that cytosolic ‘free’ PYM binds ribosomes and dislodges EJCs from mRNAs during the pioneer round of translation, thus restricting EJC disassembly to translating mRNAs. However, the function of PYM and its relationship to the EJC has not been characterized in vivo.
In this study, we characterize the function of PYM during Drosophila oogenesis. We show that Drosophila PYM (DmPYM) binds Y14-Mago but that, in contrast to its mammalian ortholog, it does not appear to interact with ribosomes. While DmPYM is not required for viability, it is essential in flies lacking one functional copy of y14 or mago. We demonstrate that over-expression of the N-terminus of DmPYM in the ovary is sufficient to dissociate EJCs from mRNAs in the cytoplasm independently of translation and causes female sterility due to oskar mislocalization in the oocyte. Finally, we show that assembly of the PYM-Y14-Mago ternary complex is modulated by the PYM C-terminal domain, indicating that PYM activity is controlled by a distinct mechanism in Drosophila.
Drosophila pym (wibg, CG30176), situated within intron 1 of the bgcn gene (Figure 1A) [20], is expressed in the ovary, and the protein maternally deposited in the embryo (Figure 1C, lane 1 and Figure S1A, lane 1) [21]. Immunostaining of ovaries revealed that DmPYM is present in the germarium, nurse cell and follicle cell cytoplasm, and within the oocyte is uniformly distributed in the cytoplasm (Figure S1B). A cytoplasmic distribution of PYM has also been reported in Drosophila S2, HeLa and plant cells [17], [19], [22].
To assess the role of PYM in vivo, we made use of a Drosophila line bearing a P element insertion in the gene (P{lacW}wibgSH1616) that constitutes a molecular null allele of pym (Figure 1B; Figure S1E). The flies were viable, although the females displayed defective ovarian development, due to impaired bgcn function. Transgenic expression of a bgcn cDNA tagged with GFP (bgcnGFP) restored normal oogenesis, although no PYM protein nor pym RNA was detected (Figure 1C, lane 2; Figure S1A, lane 2; Figure S1E). We used such flies, to which we refer as “pym null”, for our subsequent analyses (Table S1). Severe knockdown of PYM in ovaries and early embryos (Figure S1A, lanes 3 and 4) by expression of shRNAs targeting pym in the female germline [23] also did not appear to affect oogenesis or embryonic development.
In HeLa cells, PYM (HsPYM) enhances translation of intron-containing reporter mRNAs and is required to stimulate translation of intronless herpesvirus gM transcript by ORF57 protein during the lytic cycle [19], [24]. In addition, HsPYM knockdown resulted in increased association of EJC with spliced reporter mRNAs [18], implying a role of PYM in EJC removal. As oskar mRNA transport to the oocyte posterior pole requires the EJC core proteins, and tight control of oskar translation is critical for normal embryonic development [11]–[15], [25], we examined the distribution of oskar mRNA and protein in pym null egg-chambers. As shown in Figure 1D, oskar mRNA was transported into the oocyte during the early stages of oogenesis and accumulated at the posterior pole in of the oocyte at stages 8–9 and Oskar protein was first detected at the posterior pole at stage 9, when the mRNA is localized. Localization of gurken and bicoid mRNAs, as well as expression of Gurken protein, also appeared normal in pym null egg-chambers (Figure S1C, D).
In Drosophila S2 cells and HeLa cells, PYM interacts with the Y14-Mago heterodimer [17]–[19] which, together with eIF4AIII and Btz, constitute the EJC core [26], [27]; in Arabidopsis thaliana, PYM (AtPYM) interacts both with the heterodimer and with Y14 and Mago monomers [22]. Furthermore, in mammalian cells, PYM over-expression has been shown to destabilize EJCs [18]. To probe whether DmPYM might have a role in EJC regulation in vivo, we performed genetic interaction tests between pym and EJC components in the fly (Table S1). y14, mago or eIF4AIII heterozygous mutant flies, like pym null flies, are viable and fertile. Remarkably however, we failed to generate pym null flies harbouring only one functional copy of y14 or mago. In contrast, pym null, eIF4AIII heterozygous flies were viable. The lethality we observed was specific, as transgenic expression of FLAG-tagged Y14 at close to endogenous levels rescued the lethality of the pym null, y14 heterozygous flies. These results show that DmPYM function is essential when Y14 and Mago are present at reduced levels, indicating an important relationship between PYM and these EJC components in vivo.
To determine the endogenous binding partners of DmPYM during oogenesis, we performed co-immunoprecipitations (coIPs) from cytoplasmic extracts of wild-type ovaries. As shown in Figure 2A, the EJC core proteins Y14 and Mago co-precipitated with PYM when using an anti-PYM antibody, but not an unrelated antibody, demonstrating specificity of the interaction. In contrast, eIF4AIII and Btz did not co-precipitate. Addition of RNase during the coIP did not affect Y14 and Mago recovery, indicating that DmPYM binds to Y14 and Mago by direct protein-protein interaction. This is consistent with a previous study showing that a 35 residue N-terminal domain of Drosophila PYM interacts with Y14 and Mago at their heterodimerization interface [17]. Indeed, the amino acid residues of PYM necessary for this interaction are conserved across metazoa (Figure S2A).
In HeLa cells, over-expressed PYM interacts with the 48S pre-initiation complex, and components of the eIF4F complex, ribosomal proteins, and translation factors such as CBP80 and PABP coIP with HsPYM [19]. Furthermore, sucrose density gradient analysis revealed that the HsPYM C-terminus, which shows a high degree of homology to HseIF2A, is necessary for co-sedimentation with ribosomal fractions [18], [19]. Thus it was proposed that PYM physically links the EJC to the translation machinery, enhancing translation of spliced mRNAs [18], [19]. To test whether the endogenous Drosophila PYM associates with ribosomal subunits, we performed sucrose cushion centrifugation to pellet the ribosomes from ovarian cytoplasmic extracts and examined the distribution of the endogenous PYM by western blot analysis. As shown in Figure 2B, ribosomal proteins were enriched in the pellet, whereas actin, a predominantly cytoplasmic protein, fractionated in the post-ribosomal supernatant, validating the assay. Cap-binding proteins and PABP were also detected in the pellet, indicating the presence of translationally competent mRNPs in this fraction. In contrast, the quasi-totality of DmPYM was recovered in the post-ribosomal supernatant (Figure 2B), suggesting a lack of interaction with ribosomes. In addition, coIP assays using an anti-PYM antibody failed to reveal a significant interaction of DmPYM with ribosomal subunits or components of the translation initiation complex, as compared with Y14 and Mago (Figure 2A). This suggests that, unlike HsPYM, DmPYM does not associate with the translation initiation machinery. In addition, the absence of a significant association with CBP20 and eIF4E excludes a major role of PYM in cap-dependent translation regulation in Drosophila.
To analyze the function of the DmPYM during Drosophila oogenesis, we divided the protein into N-terminal (N), middle (M), and C-terminal (C) domains, and generated a set of eGFP-tagged PYM deletion transgenes (Figure S3A). Upon expression in the female germline, the bulk of the GFP signal in the PYM-GFP egg-chambers was distributed uniformly throughout the cytoplasm of the nurse cells (Figure S3B), similar to endogenous PYM (Figure S1B); however, in the case of N-, M- and C-PYM, some GFP signal was also detected in the nurse cell nuclei.
In spite of their similar distribution, the different PYM-GFP proteins had dramatically different effects on embryonic development. Females expressing FL-PYM or ΔN-PYM in the germline were fertile (Table S1). In contrast, those expressing ΔC- and N-PYM had reduced fertility: most of the progeny embryos failed to hatch due to abdominal patterning defects, and those that did hatch developed into sterile adults. Such a “grandchildless” phenotype is suggestive of reduced Oskar protein function. Indeed, immunoblot analysis of ovaries of PYM transgenic females confirmed that Oskar protein levels were substantially reduced in ΔC- and N-PYM-GFP expressing ovaries, compared with FL- or ΔN-PYM-GFP expressing, or wild-type ovaries (Figure S4A). This suggested that expression of the PYM N-terminus interferes with Oskar expression.
Expression of the posterior determinant oskar is spatio-temporally controlled such that Oskar protein is produced and accumulates stably only upon localization of the mRNA at the posterior pole of the oocyte during mid-oogenesis [28]–[30]. The low levels of Oskar protein in ΔC- and N-PYM expressing ovaries could therefore be due to a failure in oskar mRNA localization or translation at the posterior pole. To distinguish between these possibilities, we examined the distribution of the oskar mRNP component Staufen and of Oskar protein by immunostaining. As shown in Figure 3C, D, Staufen failed to enrich at the posterior pole of ΔC- and N-PYM oocytes indicating a failure in oskar mRNA localization (see also Table S1), and Oskar protein was not detected in these oocytes during oogenesis, consistent with the western blot analysis (Figure S4A). These results show that over-expression of the DmPYM N-terminal domain is sufficient to disrupt posterior localization of oskar and thus explains the absence of Oskar protein and the consequent female sterile phenotype of ΔC- and N-PYM expressing females.
We also noted that oskar localization was somewhat impaired in FL-PYM expressing egg-chambers: although Staufen accumulated at the oocyte posterior pole, the protein was also detected around the cortex (Figure 3, compare panel A with B, E and F), suggesting that, while less potent than ΔC- and N-PYM, FL-PYM also has some capacity to interfere with oskar transport.
To test if the amount of PYM relative to oskar mRNA and EJCs might be important for transport, we expressed FLAG-tagged FL-, ΔN- and ΔC-PYM transgenes in the germline of oskA87/+ females, which produce only half the normal dose of oskar mRNA [31] (Table S1). Both FL-PYM and ΔC-PYM transgenes caused oskar mislocalization, and no Oskar protein was detected in the oskA87/+ oocytes (Figure 3G, H). Western blot analysis revealed a substantial reduction in Oskar protein levels in FL- and ΔC-PYM expressing ovaries, as compared with ΔN-PYM ovaries or the wild-type control (Figure S4B, C). Both FL- and ΔC-PYM expressing females produced embryos with a strong posterior group phenotype (data not shown); only ∼5% of embryos produced by FL-PYM females hatched, and these developed into sterile adults.
In contrast to oskar, gurken and bicoid mRNAs were correctly localized to the antero-dorsal corner and anterior cortex of the PYM over-expressing oocytes (Figure S4D). The mislocalization of oskar mRNA upon PYM over-expression was independent of the tag and its position in the protein, as expression of PYM transgenes tagged at their C-terminus with eGFP produced a similar effect (data not shown). All subsequent analyses of PYM function in the flies were performed using egg-chambers expressing epitope-tagged PYM constructs in the oskA87/+ genetic background.
The presence of endogenous PYM does not appear to affect oskar localization in wild-type or oskA87/+ egg-chambers. To assess if the observed effects of PYM over-expression on oskar localization might be due to aberrant interaction of the over-expressed PYM proteins with EJC components, we performed anti-FLAG coIPs from cytoplasmic extracts from ovaries expressing the different FLAG-PYM transgenes. As observed for endogenous PYM (Figure 2A), FL- and ΔC-PYM proteins co-precipitated Y14 and Mago, but not eIF4AIII, whereas ΔN-PYM did not interact with any of the EJC core proteins tested (Figure 3J, K). These data, consistent with previous observation in HeLa cells [18], [19], show that the N-terminal portion of DmPYM exclusively mediates its interaction with the Y14-Mago heterodimer.
To test for possible interaction of over-expressed PYM proteins with the ribosomal subunits, we performed coIPs from cytoplasmic extract of the GFP-tagged PYM expressing ovaries. As shown in Figure S5A, similar to the endogenous protein, the over-expressed PYM proteins did not show significant interaction with either of the ribosomal subunits. Consistent with this, sucrose cushion centrifugation assays revealed that the tagged PYM proteins predominantly fractionated in the post-ribosomal supernatant (Figure S5B). Importantly, the GFP-tagged FL-PYM detected in the ribosomal pellet fraction was not associated with the small ribosomal subunit (Figure S5B, lane 10), as treatment of the extract with EDTA, which led to redistribution of the small ribosomal subunit to the supernatant fraction, did not affect FL-PYM distribution (Figure S5C, lane 10). We conclude that epitope-tagging does not affect the interactions of PYM or its distribution in vivo.
Taken together, our data show that increased levels of DmPYM relative to oskar mRNPs disrupt localization of the mRNA. Furthermore, this property of DmPYM is mediated by its N-terminal domain and is down-regulated by the C-terminal domain.
Although the in vivo association of the EJC with oskar mRNA has never been demonstrated, the EJC core proteins are considered to be essential components of oskar mRNPs [4], [11], [14]. We therefore hypothesized that oskar mislocalization upon PYM over-expression was due to the loss of EJC association with the mRNA, and assessed the overall integrity of EJCs in the ovary. Over-expression of either FLAG-FL-PYM or FL-PYM-GFP in oskA87/+ ovaries caused a substantial reduction in coIP of Y14 and Mago with eIF4AIII (Figure 4A, compare lanes 9 and 10 with lanes 11 and 12), indicating disassembly of the EJC core complex. The interaction of eIF4AIII with Mago was also impaired in egg-chambers over-expressing ΔC-PYM, but not ΔN-PYM (Figure 4C, lower panel, lanes 7 and 8). Hence, over-expression of the DmPYM N-terminal domain causes disassembly of the Drosophila EJC core. The fact that in the absence of PYM over-expression intact EJCs are recovered (Figure 4A, lanes 9 and 10) confirms that endogenous levels of DmPYM are not deleterious to EJC integrity.
To test whether elevated levels of FL- or ΔC-PYM cause the EJC to dissociate from oskar mRNA, we first performed in vitro splicing assays coupled with RNase H cleavage using osk(E1E2(iftz)) RNA and assessed the degree of protection of the EJC binding site, in the presence or absence of recombinant PYM proteins [32]. Briefly, after the completion of splicing and concomitant EJC deposition, reactions were incubated with GST-tagged FL-, ΔN- or ΔC-PYM, followed by RNase H cleavage induced by an oligonucleotide (Ol-25) complementary to the EJC deposition site. As shown in Figure 4B, a significant decrease of oskar mRNA correlated with an increase of the corresponding RNase H cleavage product was observed in the presence of GST-FL- and ΔC-PYM (lanes 6, 7 and 10, 11, respectively; * shows the position of the cleavage product), compared to GST-ΔN-PYM or the GST control, indicating a loss of EJC binding. This result shows that elevated amounts of DmPYM can cause mature EJCs to dissociate from oskar mRNA in vitro, raising the possibility that oskar-associated EJCs are destabilized by ectopic PYM in vivo.
To test directly whether high PYM levels cause the EJC to dissociate from oskar mRNA in vivo, we co-expressed GFP-Mago and FLAG-tagged PYM constructs in the germline of oskA87/+ flies and performed RNA-coimmunoprecipitation (RIP) from cytoplasmic extracts of ovaries. The EJC was immunoprecipitated using GFP-Trap beads and the precipitated RNA extracted for semi-quantitative RT-PCR analysis. As shown in Figure 4C, oskar mRNA was enriched in immunoprecipitates of ΔN-PYM extract, demonstrating that EJCs are assembled on oskar mRNA in vivo and that over-expression of the C-terminal and middle region of DmPYM does not alter its integrity. In contrast, although similar amounts of GFP-Mago were recovered, considerably less oskar mRNA was immunoprecipitated from FL-PYM or ΔC-PYM ovaries (Figure 4C). These results show that over-expression of the DmPYM N-terminal domain causes EJC dissociation from oskar mRNA in the ovary. We further investigated whether the association of the EJC with other localized mRNAs was also affected by the expression of PYM transgenes. Interestingly, a profile similar to oskar was observed with bicoid, nanos and gurken mRNAs: these mRNAs were also co-precipitated with a considerably lower efficiency from FL-PYM and ΔC-PYM, compared with ΔN-PYM extracts, demonstrating that the N-terminus of PYM is sufficient to dissociate the EJC from the spliced mRNAs in vivo. Since FL-PYM or ΔC-PYM over-expression leads to oskar mislocalization in the oocyte (Figure 3G, H), but not that of bicoid or gurken (Figure S4D), we conclude that although the EJC associates with bicoid and gurken mRNAs in the oocyte, its function is dispensable for their localization. This is consistent with normal localization of bicoid and gurken in y14 mutant oocytes [11].
Previous studies have shown that HsPYM interacts with the translation pre-initiation complex and disassembles mature cytoplasmic EJCs from spliced RNAs [18], [19]. Having found no evidence of association of DmPYM with ribosomal subunits (Figure 2 and Figure S5), yet shown that DmPYM has the capacity to dismantle EJCs from mRNAs in the egg-chamber (Figure 4), we assessed whether this effect is translation-dependent. We monitored the distribution and translational output of oskΔi(2,3)-boxB transgenic mRNA [32] in PYM-GFP expressing egg-chambers. Although oskΔi(2,3)-boxB mRNA is spliced and localized (Figure 5A), it is not translated due to the presence of 5× boxB stem-loops between the two oskar start codons (Figure 5E, lane 2). As shown in Figure 5B–D, upon expression of FL- or ΔC-PYM, but not ΔN-PYM, oskΔi(2,3)-boxB mRNA was mislocalized in the oocyte, as revealed by anti-Staufen immunostaining. This demonstrates that increased levels of DmPYM cause EJC dissociation from oskar mRNA in the absence of translation. Furthermore, these data provide evidence that the EJC is deposited on the oskΔi(2,3)-boxB mRNA in the oocyte and that integrity of the complex is essential for oskar transport.
To investigate the contribution of the different DmPYM domains to EJC binding and disassembly, we performed coIPs from extracts of Drosophila S2 cells co-expressing HA-tagged eIF4AIII and either GFP (control) or GFP-tagged PYM proteins. To monitor both the PYM-Mago interaction and EJC integrity, PYM-GFP and HA-eIF4AIII were immunoprecipitated separately using GFP-Trap and HA-beads, respectively, and western blots of the bound fractions were probed with anti-Mago antibodies. Neither of the ectopically expressed proteins bound detectably to the agarose beads (Figure 6A, lanes 8–14, mock), demonstrating specificity of the assay.
Consistent with our previous experiments (Figures 3J, K and 4A, C), Mago co-precipitated exclusively with the FL-, ΔC- and N-PYM-GFP fusion proteins, which contain the N-terminal Y14-Mago binding domain, but not with ΔN-, M and C-PYM, or the GFP control (Figure 6A, lanes 15–21). In addition, we noted that a greater amount of Mago was recovered in ΔC-PYM than in FL-PYM immunoprecipitates (Figure 6A, lanes 16 and 18), suggesting a regulatory function of the C-terminus in DmPYM binding to Mago-Y14. Surprisingly, we observed only low co-precipitation of Mago with N-PYM, compared with ΔC-PYM, which contains both the N-terminal and middle domain (Figure 6A, lanes 18 and 19). This implies a role of the PYM middle domain, which itself does not bind Mago, in stabilizing the N-PYM-Y14-Mago interaction.
We next investigated the effect of the different PYM-GFP fusion proteins on EJC integrity, monitoring the ability of HA-eIF4AIII to co-IP Mago (Figure 6A, lanes 22–28). Remarkably, in spite of their differential ability to co-precipitate Mago, all three fusion proteins, FL-, N- and ΔC-PYM-GFP, displayed a similar capacity to disassemble the EJC (Figure 6A, lanes 23, 25 and 26). In contrast, neither ΔN-, M-, nor C-PYM-GFP, which failed to bind Mago, affected integrity of the EJC (Figure 6A, lanes 24, 27 and 28). These results show that the ability of the different DmPYM truncations to provoke EJC disassembly correlates with their ability to bind Mago, but not with its co-precipitation efficiency.
Although both N-PYM and ΔC-PYM affected EJC stability in S2 cells (Figure 6A, lanes 25 and 26) and were equally potent in causing oskar mislocalization (Figure 3C, D), the two proteins differed considerably in their ability to co-precipitate Mago (Figure 6A, lanes 18 and 19). This seemingly low binding of N-PYM to Y14-Mago might reflect a short half-life of the complex. To test this hypothesis, we prepared cytoplasmic extracts from S2 cells expressing the FL-, ΔC-, N-PYM-GFP or GFP proteins, added the protein cross-linking agent DSP and performed IPs using GFP-Trap beads. The overall immunoprecipitation efficiency was reduced in the presence of DSP. However, substantially greater amounts of Mago and Y14 co-precipitated with N-PYM upon cross-linking, consistent with stabilization of the trimeric complex (Figure 6B, lane 22).
Quantification of the western blots revealed that, although the binding of both FL- and ΔC-PYM to Y14-Mago increased upon DSP treatment (1.7 and 3.6 fold, respectively), ΔC-PYM bound Y14-Mago more effectively than FL-PYM under both native and cross-linking conditions (Figure 6C). In stark contrast, the efficiency with which N-PYM co-precipitated Mago increased 92-fold (from 0.92%, to 84.88%) upon DSP cross-linking, such that it approximated that of ΔC-PYM (98.9%, Figure 6C). Hence, while the interaction of N-PYM with Y14-Mago is labile, the binding capacity of the N-terminal domain alone to the EJC is nearly equal to that of ΔC-PYM, and is far greater than that of FL-PYM. This explains the potent effect of N-PYM over-expression on EJC integrity and thus, on oskar RNA localization.
Previous studies carried out in cultured cells have led to the model that PYM - a Y14-Mago binding protein, by virtue of its association with the small ribosomal subunit, dissociates EJCs from spliced mRNAs during the first round of translation (ref. 9 and references therein). To date, however, the physiological role of PYM has remained unclear. Here we have addressed the function of PYM in an animal context.
Although a direct association of the EJC core components with oskar mRNA in vivo has been presumed [4], [14], [32], our RNA-coimmunoprecipitation experiments on Drosophila ovarian extracts provide the first evidence of a “physical” association of EJC core components with oskar mRNA. The effect of PYM on oskar localization can therefore be seen as a direct consequence of EJC dissociation from oskar RNA. This further underscores the importance of EJC association in oskar mRNA localization. PYM over-expression does not affect bicoid or gurken mRNA localization in the oocyte, consistent with previous genetic studies indicating no role of the EJC in this process. However, the fact that upon PYM over-expression EJCs are removed not only from oskar, but also from other templates such as bicoid, gurken and nanos mRNAs, indicates that the activity of DmPYM in vivo is not restricted to specific mRNPs, but that the protein acts more globally on EJC-containing mRNP complexes.
Our in vivo analysis shows that the amount of PYM relative to the EJC core proteins, Y14 and Mago, is crucial for Drosophila development: PYM function is essential when the gene dosage of either of its interacting partners Y14 and Mago is reduced. Interestingly, under steady-state conditions, flies that exclusively lack pym function are viable and are easily maintained as a stock in the laboratory. The viability of pym null flies suggests either that in vivo the binding activity of endogenous DmPYM to Y14-Mago has no physiological impact or that it may be negatively regulated. The latter hypothesis is supported by our finding that, in wild-type flies, the EJC-dependent localization of oskar mRNA is only abolished by expression of PYM constructs lacking the C-terminal domain. Over-expression of full-length DmPYM has a mild effect that is increased in “sensitized” oskA87/+ flies, in which oskar RNA dosage is reduced. Such striking differences between the PYM truncations with respect to their capability to mislocalize oskar RNA both suggests that the full-length PYM - in contrast to its truncations - is regulated, and points to the C-terminal domain as a “key player” in such a regulatory pathway.
Studies carried out in human cell cultures showed that the C-terminal eIF2A-like domain of HsPYM mediates its association with components of the 48S translation pre-initiation complex [19]. Thus an attractive hypothesis would be that, in Drosophila, PYM is regulated through the interaction of its C-terminal 54 residues with the small ribosomal subunit. However, none of our IP or ribosome pelleting experiments performed on Drosophila ovarian lysates support such an association of endogenous or ectopically expressed DmPYM protein with components of the small ribosomal subunit. This is most likely due to divergence of the amino acid residues in DmeIF2α (Figure S2B). Furthermore we show that over-expression of DmPYM in oocytes not only impairs localization of endogenous oskar, but also of a non-translatable oskar RNA reporter. Thus a ribosomal interaction with DmPYM in EJC regulation seems unlikely.
Although the interaction of DmPYM with Y14-Mago is essential for EJC disassembly, the stability of the ternary complex is not important for the dissociation. The middle and the C-terminal domains of DmPYM influence its interaction with Y14-Mago, albeit in opposing manners: binding is stabilized by the former and antagonized by the latter (Figure 6). Whether the middle domain somehow stabilizes the PYM N-terminus:Y14-Mago interaction or whether it promotes proper presentation of the N-terminus for Y14-Mago binding remains to be addressed. In support of an antagonistic effect of the PYM C-terminus on Y14-Mago binding, the interaction of FL-PYM with Y14 and Mago does not approximate saturation even in the presence of the protein cross-linker DSP. This clearly indicates that, in vivo, DmPYM must be present in an equilibrium between active and dormant states. The C-terminal domain might be modified post-translationally or might serve as a binding platform for co-factors that enhance or inhibit the EJC-dismantling activity of DmPYM. Further analysis, including an unbiased proteomics approach to identify novel interacting partners of DmPYM, should provide insights into the regulation of DmPYM and its interaction with the EJC. The fact that PYM function is essential in flies lacking one functional copy of y14 or mago suggests that PYM activity is regulated by a pathway ensuring EJC homeostasis in the fly.
The full length pym (FL-PYM) and eIF4AIII coding regions were PCR amplified from ovarian cDNAs using specific primers (Table S2) and cloned into pENTR/SD/D-TOPO plasmid (Invitrogen) to generate the entry clones. The PYM deletion (ΔN, ΔC, N, M and C-PYM) entry clones were generated by PCR using FL-PYM entry clone as template. The entry clones were used for recombination with the destination vector (pPFMW for N-terminal FLAG; pPWG for C-terminal eGFP; pPFMW for N-terminal FLAG-Myc) from the Drosophila Gateway vector collection (gift of Terence Murphy, Carnegie Institution for Science). The eIF4AIII entry plasmid was used for recombination with pAHW vector to generate an N-terminal HA-tagged protein. The primers are listed in Table S2.
All fly stocks were maintained at 25°C. The following fly stocks used in this study were; w1118 (wild-type), w−;P{lacW}wibgSH1616/CyO (BL#-29502), w−;Df(2R)BSC600/SM6a (BL#-25433), w-,pCOGGal4::VP16;;oskA87, nosGal4::VP16/TM3Sb [31], w-;;nosGal4::VP16, w-;[bmP-BamHA]/Cyo;bgcnGFP/TM3 (gift of D. McKearin), w−,GFP::Mago;;Sb/TM3Ser [33], y1scv1;P{TRiP.HMS01488}attP2 (BL#-35746), y1scv1;P{TRiP.GL00515}attP40 (BL#-36096), y1scv1;P{TRiP.GL00596}attP40 (BL#-36636). mago3 [13], tsuΔ18 [34] and eIF4AIII19 [14] alleles were used to test genetic interaction with pym.
Flies lacking pym function were generated by crossing the recessive-lethal pym allele P{lacW}wibgSH1616 to w−; Df(2R)BSC600/SM6a, which contains a chromosomal deletion encompassing the bgcn locus. The viable pym null adults had defective oogenesis, which was rescued by expression of a bgcnGFP transgene (gift of D. McKearin). The stock eventually lost the Df(2R)BSC600 chromosome and P{lacW}wibgSH1616/P{lacW}wibgSH1616; bgcnGFP flies were used for the analyses shown in the manuscript. We refer to both genotypes (P{lacW}wibgSH1616/Df(2R)BSC600;bgcnGFP and P{lacW}wibgSH1616/P{lacW}wibgSH1616; bgcnGFP) as “pym null”, as they behaved identically in our assays (see Table S1).
For the generation of transgenic flies, pUASp-based destination plasmids containing the PYM fragments were injected together with helper plasmid as described [4].
The PYM-GFP and FLAG-GFP constructs were expressed in the germline using nosGal4::VP16 driver. The oskΔi(2,3)-boxB transgene was expressed in oskA87/Df(3R)pXT103 background.
For each biological replicate, 6 adult females (wild-type and pym null) were homogenized in 350 µl TRIzol LS (Invitrogen) and RNA was extracted according to the manufacturer's protocol. The RNA samples were treated with TURBO DNase I (Ambion) for 20 min and purified using RNeasy kit (Qiagen). Reverse transcription was performed using Superscript III First-Strand synthesis kit (Invitrogen) following the manufacturer's protocol. The cDNA samples were used for qRT-PCR in One-Step ABI PCR cycler using the primers listed in Table S2.
Cells were grown in Express Five SFM Medium (Life technologies) at 25°C in the presence of penicillin-streptomycin (100 U/ml; Life technologies), puromycin (1.8 µg/ml; Sigma) and L-glutamine (2 mM; Life technologies). An S2 cell line stably transfected with pMT-Gal4-puro (gift of S. DeRenzis) was used in this study. For transient transfections, plasmids were transfected into cells in 75 cm2 cell culture flasks using Effectene reagent (Qiagen) following manufacturer's instructions. After 36 h, CuSO4 was added to 0.75 mM and the cultures incubated for 6 h for Gal4 induction. Cells were harvested using a cell scraper, washed with PBS and placed on ice for further processing.
All procedures were carried out at 4°C. Cytoplasmic extracts, from fly ovaries or S2 cells, were prepared using the NE-PER kit (Thermo Scientific) following manufacturer's instructions in presence of Halt protease inhibitor cocktail (Thermo Scientific). For protein cross-linking, extracts were supplemented with 1 mM DSP (Dithiobis[succinimidyl propionate], Sigma) and incubated at 4°C for 1 h with mixing. Cross-linking was stopped by addition of 50 mM Tris-HCl pH 7.8 followed by incubation for 15 min. Lysates were pre-cleared with Protein A/G beads (Roche) for 30 min. For immunoprecipitation, extracts were incubated either with rat anti-PYM (1∶250), rabbit anti-eIF4AIII (1∶150), mouse anti-FLAG (1∶100, Sigma #F3165) antibodies or with pre-blocked GFP-Trap (Chromotek) and anti-HA (Sigma) agarose beads for overnight with mixing. The beads were pre-blocked using Western Blocking reagent (Roche). For RNase treatment, 1 µl of RNase cocktail (Ambion) was added to the lysate together with the antibody. When using antibodies, the immuno-complexes were isolated by incubating the lysate with pre-blocked Protein A or G beads for 3 h. The beads were washed four times, 10 min each, with wash buffer (25 mM HEPES-KOH pH 7.5, 300 mM KCl, 4 mM MgCl2, 1 mM DTT, 125 mM Sucrose, 0.2% NP-40, 1× Halt protease inhibitor) and once with PBS. The beads were boiled in presence of 2× Laemmli sample buffer and the bound proteins analyzed by SDS-PAGE followed by western blotting.
The primary antibodies used for western blot staining were: Rat anti-PYM (1∶5,000; gift of E. Izaurralde), Rat anti-Y14 [11] (1∶2,500), Rabbit anti-Mago [11] (1∶2,000), Rabbit anti-eIF4AIII (1∶4,000; gift of I. Palacios), Rat anti-Btz (1∶1,000; A. Ephrussi (unpublished)), Rabbit anti-RpS6 (1∶2,000; gift of M. Hentze), Rabbit anti-RpL32 (1∶2,000; gift of M. Hentze), Rabbit anti-eIF4E (1∶2,000; gift of A. Nakamura), Rabbit anti-CBP20 (1∶2,000; gift of M. Hentze), Rabbit anti-eIF4G (1∶2,000; gift of A. Nakamura), Rabbit anti-Kinesin heavy chain (KHC, 1∶25,000; Cytoskeleton), Rabbit anti-actin (1∶2,500; Sigma), Rabbit anti-Oskar (1∶2,000), Rabbit anti-GFP (1∶2,000, Torrey Pines). Goat anti-Rabbit (1∶2,500) and anti-rat (1∶2,500) conjugated with HRP (GE Healthcare) were used as secondary antibodies.
For quantification of the coIP assays using S2 cell extracts under native and DSP cross-linked conditions, signals obtained from western analysis were subjected to densitometry measurements using ImageJ (http://imagej.nih.gov/ij/) and processed with Excel (Microsoft Inc.). The relative enrichments (RE) for PYM-GFP proteins, Mago and Y14 were defined as relation of measured signals in precipitates and inputs:Since DSP treatment led to lower enrichments for GFP or PYM-GFP fusions (Figure 6B), all values plotted for Mago and Y14 from the individual immunoprecipitates were defined as the co-IP efficiency (CoIPe) relative to corresponding enrichments, estimated for GFP and GFP-PYM fusion proteins. For example, the CoIPe of Mago in ΔC-PYM-GFP IP was defined as:
The ovaries from fly lines co-expressing GFP-Mago and FLAG-PYM constructs were used for preparation of cytoplasmic extract using the NE-PER kit (Thermo Scientific). The extract was pre-cleared using Protein A beads for 1 h at 4°C and incubated with GFP-Trap agarose (ChromoTek) for 3 h. The beads were washed four times, 15 min each, with wash buffer containing 600 mM KCl and once with PBS. 20% of the input and beads were aliquotted for SDS-PAGE analysis and the rest processed for RNA extraction using TRI Reagent (Ambion). The recovered RNA was treated with 2 U TURBO DNase I (Ambion), and used for cDNA synthesis using SuperScript III First-Strand synthesis kit (Invitrogen) according to the manufacturer's protocol. Primers specific for oskar, bicoid, nanos and gurken were used for the PCR amplification (see Table S2). The data from 28 cycles are shown.
Dissected ovaries were lysed in hypotonic buffer (5 mM Tris-HCl pH 7.5, 1.5 mM KCl, 2.5 mM MgCl2, 0.5% TX-100, 0.5% DOC, 1× Halt protease inhibitor), incubated on ice for 15 min with or without 20 mM EDTA and centrifuged at 13,000 g for 10 min. The supernatant was adjusted to the sucrose cushion buffer (10 mM Tris-HCl pH 7.5, 150 mM KCl, 5 mM MgCl2), layered on 1 M sucrose (200 µl extract on 700 µl sucrose solution) and centrifuged at 200,000 g for 2 h. The supernatant was concentrated using Amicon Ultra-10K filters and the pellet suspended in sucrose cushion buffer containing 1 mM DTT. The fractions were analysed on SDS-PAGE followed by western blotting.
The ovaries were dissected in PBS and processed for FISH and immunostaining as described previously [4]. For immunostaining, the primary antibodies were: Rat anti-Staufen (1∶2,500), Rabbit anti-Oskar (1∶3,000), mouse monoclonal anti-Gurken 1D12 (1∶200, Drosophila Studies Hybridoma Bank), Rat anti-PYM (1∶7,500). The DIG-labeled antisense probes to oskar, bicoid and gurken, used for FISH analysis has been described previously [32].
The pENTRY/SD/D-TOPO vector containing FL-PYM was used for recombination with the Gateway destination vector pDEST15 (Invitrogen) to generate GST-PYM plasmid. The ΔN- and ΔC-PYM constructs were generated by PCR using primers pairs O-388/O-390 and O-389/O-391, respectively, and cloned into EcoRI and NotI site of pGEX-4T1 (gift of E. Loeser). All plasmids were fully sequenced.
The proteins were expressed in E. coli (BL21 DE3 Rosetta 2) and purified using glutathione beads under standard conditions, dialyzed against the following buffer: 1.5× PBS, 1 mM MgOAc, 10% Glycerol, 2 mM DTT in Spectra/Por Membrane 1 (cut-off: 6–8,000), flash-frozen in aliquots and stored at −80°C till further use.
Preparation of DNA templates, in vitro transcription and preparation of embryonic nuclear extracts have been described previously [32]. Soluble nuclear extract was dialyzed against Buffer D (15 mM HEPES-KOH pH 7.9, 20% glycerol, 120 mM KCl or KOAc, 0.2 mM EDTA pH 8.0, 1 mM DTT) in Spectra/Por Membrane 1 (cut-off: 6–8,000), aliquoted, quick-frozen in liquid nitrogen and stored at −80°C.
In vitro splicing reactions were carried out in 25 µl, containing 10 µl of the embryo nuclear extract, 32P-labeled pre-mRNA substrate in a buffer (26 mM HEPES-KOH pH 7.9, 40 mM KCl, 80 mM KOAc, 4 mM MgOAc, 5 mM Creatine-Phosphate, 4 mM ATP, 2.5% PVA) for 180 min at 20°C. Purified proteins (0.5 or 1 µM of either GST, GST-FL PYM, GST-ΔN PYM or GST-ΔC PYM) were added to the 25 µl splicing reactions and further incubated for 30 min at 20°C. RNase H assays were carried out as previously described [32] using oligonucleotide Ol-25.
|
10.1371/journal.pntd.0004312 | 3D Architecture of the Trypanosoma brucei Flagella Connector, a Mobile Transmembrane Junction | Cellular junctions are crucial for the formation of multicellular organisms, where they anchor cells to each other and/or supportive tissue and enable cell-to-cell communication. Some unicellular organisms, such as the parasitic protist Trypanosoma brucei, also have complex cellular junctions. The flagella connector (FC) is a three-layered transmembrane junction that moves with the growing tip of a new flagellum and attaches it to the side of the old flagellum. The FC moves via an unknown molecular mechanism, independent of new flagellum growth. Here we describe the detailed 3D architecture of the FC suggesting explanations for how it functions and its mechanism of motility.
We have used a combination of electron tomography and cryo-electron tomography to reveal the 3D architecture of the FC. Cryo-electron tomography revealed layers of repetitive filamentous electron densities between the two flagella in the interstitial zone. Though the FC does not change in length and width during the growth of the new flagellum, the interstitial zone thickness decreases as the FC matures. This investigation also shows interactions between the FC layers and the axonemes of the new and old flagellum, sufficiently strong to displace the axoneme in the old flagellum. We describe a novel filament, the flagella connector fibre, found between the FC and the axoneme in the old flagellum.
The FC is similar to other cellular junctions in that filamentous proteins bridge the extracellular space and are anchored to underlying cytoskeletal structures; however, it is built between different portions of the same cell and is unique because of its intrinsic motility. The detailed description of its structure will be an important tool to use in attributing structure / function relationships as its molecular components are discovered in the future. The FC is involved in the inheritance of cell shape, which is important for the life cycle of this human parasite.
| Trypanosoma brucei is an uni-cellular parasite transmitted to humans and cattle by the bloodsucking tsetse fly. Once swimming in the mammalian bloodstream, it causes the devastating African sleeping sickness in humans and nagana in cattle. During its complex life cycle, it undergoes many cell shape changes, which are important for efficient parasite transmission. Here, we have studied a cell structure intrinsically involved in shape acquisition during division of the T. brucei life cycle form that multiplies in the fly midgut. Using electron tomography we show the 3D architecture of a motile cellular junction that slides with the tip of the growing new flagellum along the side of the old flagellum. This enables the new flagellum to zip in to the cell body structure alongside the old flagellum after which the cleavage furrow is established between these two flagella, producing two daughter cells of similar cell shape. We present here a detailed architectural overview of this junction; we show that it matures with time and pushes the old flagellum’s axoneme sideways as it passes. This structural map enables insight into the function of this extraordinary mobile cellular junction.
| Cellular junctions are crucial for the formation of tissues, pathogen/host cell interactions and communication between cells, e.g., the plasmodesmata in plants and the gap junctions in animals. However, junctions can also exist within a single cell, such as the top connectors between sterocilia and the kinocilium on outer hair cells in the ear [1].
Trypanosoma brucei are unicellular protozoa able to form multiple kinds of cellular junctions. These parasites cause the devastating African sleeping sickness that is transmitted to humans and cattle by the bite of an infected tsetse fly (Glossina spp). The ability to adapt to a changing environment is essential to their complex life cycle [2]. One such adaptation is the asymmetric intercellular junctions between the T. brucei flagellum and the microvilli in the tsetse fly salivary gland epithelia [3]. When the parasites are attached like this, the cells divide asymmetrically to generate daughter cells of a different shape. Similar cellular junctions between the flagellum and the host species tissue are also found in T. congolense, T. vivax and Leishmania mexicana [4–6], providing not only a physical tethering to the substrate but also a signaling opportunity such the one described between the parasitophorous vacuole and the amastigote L. mexicana flagellum [5].
Procyclic T. brucei, the form that infects the fly mid-gut, possess a single flagellum that originates in the flagellar pocket and exits the cell body near the posterior end of the cell [7,8]. The extracellular part of the flagellum contains an extra-axonemal structure called the paraflagellar rod (PFR; [9–12]), and is attached to the plasma membrane through a region called the flagellar attachment zone (FAZ; [2,3,13]. Inside the FAZ, a specific complex junctional component, the recently discovered ‘staple’ is found [12]. These are extracellular plate-like structures with fibrous connections into both the flagellum and cell body. These are, in contrast to the first example of T. brucei cellular junctions, intracellular connections, connecting one part of the cell to another.
A third cellular junction in T. brucei is the flagella connector (FC); a specialization that is unique to procyclic cells in division that are assembling a second flagellum that will be inherited by a daughter cell [5,14–16]. The FC is a mobile trans-membrane junction that links the tip of the new flagellum to the side of the old flagellum (Fig 1A; [14,17]). Once the new flagellum tip, and the FC, has reached a point roughly 50% along the length of the old flagellum, it stops migrating. From then on the tip of the new flagellum is immobile on the surface of the old flagellum, and continued flagellar growth is temporally accompanied by independently separating basal bodies and kinetoplasts [18]. The physical connection between old and new flagellum probably ensures that the elongating new flagellum copies the left-handed helical path of the old flagellum [19], facilitates flagellar attachment zone formation and thus imposes a similar cell shape on the ensuing daughter cells after division.
The basic outline of the FC structure has been described using conventional thin-section electron microscopy of chemically fixed material [14,17]. This work showed that the FC consists of a tri-laminar structure composed of three distinct electron dense layers found in the new flagellum, the interstitial space, and the old flagellum. Each layer is subdivided into three plates. Interconnecting these layers with the axonemal microtubule doublets are thin intra-flagellar filaments (Fig 1B; [7,8,17]).
However, many aspects of the FC structure and behaviour have remained elusive. For example it is not known how it moves along the old flagellum, although we do know this motion is separate from the extension of the new flagellar axoneme [9–12,18], indicating the presence of some sort of molecular motor. We have now performed (cryo) electron microscopy and (cryo) electron tomography with the hope of further clarifying the function and mechanism of this junction and its motility. The combination of techniques used has resulted in our developing a comprehensive 3D architecture, presented here, that provides insight into the physical properties of the FC.
To investigate the FC ultra-structure, we performed transmission electron microscopy on both formaldehyde fixed cells and cells cryoimmobilised by high pressure freezing. In thin cross-sections of chemically fixed flagella, both the new and old flagella had irregular outlines (Fig 1C). The tip of the new flagellum lay quite distant from the plasma membrane of the cell, and some of its doublet microtubules were missing[15]. The FC displayed partitions in the electron dense material (previously named “plates” [17]) throughout the interstitial zone. In high pressure frozen cells, on the other hand (Fig 1D), interstitial zone material was visible, but no clear partitions into plates were visible. In this preparation, both flagella were round in cross-section. New and old flagella are both in close proximity to the cellular surface. The new flagellum in Fig 1D had a central pair that was parallel with the underlying sub-pellicular microtubule array, but the old flagellum’s central pair axis was rotated in comparison to the cellular microtubules (Fig 1D and S1 Fig).
The FC was then examined using cryo-electron microscopy of sections cut from high pressure frozen cells embedded in vitreous ice. Because this sample preparation does not involve dehydration of the cells, nor coating of proteins with heavy metals, it displays cell structure in a close-to-native state [12,20]. This image shows the two flagella, both close to the cell membrane and between them we find the FC (Fig 1E). Distinct, regular filamentous densities project from both flagella membranes in the FC and a region of darker electron density is found in the middle of the interstitial zone. This filamentous arrangement is interrupted in three areas by smoother electron densities across the FC (Fig 1F–1F’; blue boxes).
We conclude that studying the FC ultra-structure using various sample preparations and imaging techniques yields new information about its ultrastructure. Therefore, we progressed to study its 3D structure using electron tomography of the conventional chemically fixed, high pressure frozen and vitrified samples.
A tomographic reconstruction of chemically fixed FCs was performed. A thin slice of one reconstruction (Fig 2A; S1 Movie) shows the tri-laminate structure and filamentous connections from it to both axonemes (arrows). A 3D model of the FC was produced by drawing around the structural features of interest in the tomogram. The top view (Fig 2B; S2 Movie) displays the flagellar membranes and shows how the tip of the new flagellum is apposed to the old flagellum. In the en face view (Fig 2C), membranous components were subtracted to visualise the microtubule doublets and the 3D morphology of the tri-laminar complex and its associated filamentous network (doublet microtubules are colour coded with a gradient from doublet 1 in pale yellow to doublet 9 in dark red).
Previously, each layer within the tri-laminar structure has been described as partitioned into three electron dense plates [14–18, 21]. However, in the tomographic reconstruction, the 3D structure of complete layers shows that they are subdivided into a range of 2–4 plates, of which 3 plates is the most common (Fig 2D). The 3D reconstruction also shows a correlation between the angles of partitions and the two axes underlying axonemes (Fig 2E).
The FC tri-laminar structure was less noticeable in the high pressure frozen material (Fig 2F; S3 Movie). A new flagellum plate was not detectable, probably because of the electron dense cap coating the entire inside of the new flagellum tip (Fig 2G–2H; S4 Movie). The interstitial layer was thinner than previously seen in the chemically fixed sample. Also here, the plates within the tri-laminar structure were not detectable (Fig 2I), neither was the filamentous network between the FC and the axonemes. Thus, the high pressure frozen FC reveals a more compact FC morphology with an electron dense cap instead of the new flagellum plate.
The time of progression of a cell through the cell cycle is directly correlated to the length of the new flagellum. We therefore measured the length of the new flagellum in those high pressure frozen cells in which the FC had been reconstructed.
When the new flagella were short (below 2 μm), the FC was in the process of being formed inside the flagella pocket (Fig 3A; S5 Movie). At this stage, the FC was seen as a thin electron density in the interstitial zone extending from the wall of the old flagellum. We correlated the new flagellum length with the FC interstitial zone thickness (distance between the flagella) and found that this parameter decreased considerably as the FC matures (Fig 3C), being 28 ± 6 nm in short flagella (n = 4) and 13 ± 2 nm in longer flagella (Fig 3B; n = 4).
The structural maturation of the old flagellum FC layer during the cell cycle was, however, less clear (Fig 3D). The old flagellum layer in two cells with long new flagella had the same thickness (~20 nm) as found in cells earlier in the cell cycle; however two flagella connectors had old flagellum layers almost twice as thick (~35 nm). There was no difference in FC length (273 ± 61 nm; n = 9; Fig 3E) or depth (116 ± 30 nm; n = 8; Fig 3F) over the cell cycle.
We conclude that the thickness of the interstitial layer of the FC changes as the cell cycle progresses.
To image the FC protein architecture in a more native state, we made cryo-electron tomograms of vitreous sections. One electron tomogram of such a section, contained the most distal ~70 nm of the FC between the old and new flagellum, as well as the cell body (Fig 4A; S6 Movie). To examine the 3D architecture of the region we modelled the FC, membranes and microtubules (Fig 4B). The generated 3D model shows the complete FC structure, including membranes, as a ~100 nm wide connection between the old and new flagella (Fig 4C).
In views sliced through the tomogram, it is apparent that the distance between the old and new flagella is the greatest (~50 nm) closest to the distal tip of the new flagellum (Fig 4D); that this distance shrinks to ~20 nm as one moves more proximal in the new flagellum (Fig 4E–4F). The extracellular density we interpret as the FC lies close to the old flagellum throughout this volume, perhaps suggesting that it originates from that organelle. A very electron dense structure is also seen in close vicinity to the FC, which relevance we do not know, but a similar electron density was seen in a similar position in another cryo-electron microscopy image (Fig 4A–4B; S2 Fig). The FC layer in the new flagellum was not distinguishable, nor the axonemal microtubules in the new flagellum in this sample preparation. A faint density was visible in the old flagellum at the location of the FC. The dimensions measured in cryo-sectioned specimens have to be carefully interpreted due to compression, a characteristic artefact of vitreous sectioning [14,17,22]. The compression factor is estimated here at 50% based on the ovoid shape exhibited by the microtubules and the old flagellar membrane, and should be considered when reading the measurements here and in Table 1.
The electron density that forms the interstitial FC component had a clear periodicity when seen in cross-section, even though there were tomographic slices where this periodicity was not as strong (possibly correlating to the areas indicated with blue boxes in Fig 1F’). When the distance between the flagella was larger (Fig 4D), three stacks of periodical electron densities were present. As the distance grew closer, this decreased to two (Fig 4E) and then one line (Fig 4F) of periodical densities.
When the tomogram was rotated, these periodical densities showed as parallel lines (Fig 4G) 7 nm apart, and with denser areas spaced 11 nm along them (arrows; Fig 4G). We conclude that cryo electron tomography of frozen hydrated sections of T. brucei cells has revealed the structural periodicity of the interstitial zone of the FC.
In the old T. brucei flagellum, defined numbers can be assigned to each doublet microtubule, as the central pair does not rotate [18,23] and a fixed, external structure, the paraflagellar rod (PFR), exists. The attachment of the axoneme to the PFR occurs at microtubule doublets 4–7 [19]. As previously described [14,17], the FC complex faces microtubule doublets number 7–9 in the old flagellum, but here we also see a close proximity between the FC and microtubule doublet 1. Previous studies were unable to distinguish which microtubule doublets the FC was facing in the new flagellum, or indeed if this side of the junction is more flexible, interacting with a range of microtubule doublets. Despite the apparent disorganisation inside the growing new flagellum tip[15], we have seen that the FC was consistently aligned with doublets 3–5 in the new flagellum (Fig 5A; n = 3). Thus, the FC follows specific microtubule doublets within both the new and old flagellum (Fig 5B).
To examine if the location of the axoneme in the old flagellum was perturbed by the passage of the FC, several images of flagellar cross-sections from multiple cells were aligned with the microtubule quartet in the sub-pellicular MT array to the bottom left of the flagellum (e.g. Fig 5C). An ellipsoid was placed where the axoneme was found in each image, which was invariably in the top left corner of the flagellum (Fig 5D). However, close to the FC, the position of the axoneme shifted ~100 nm away to a more central location (red ellipsoids). This shift in axonemal position around the FC can also be seen in longitudinal sections of the region (e.g. Fig 2F), and tells us that the FC presence rearranges the internal space of the old flagellum.
For several reasons, discussed later, it is possible that the filamentous network seen in chemically fixed cells between the FC and the axonemes represents a physical link between these three structures. We measured the distances the filamentous network would span, from the membrane of the flagellum to the closest doublet microtubule at the level of the FC. The membrane-axoneme distance was 61 ± 1 nm and 55 ± 9 nm (n = 2) respectively in the new and old flagella in chemically fixed cells. In high pressure frozen cells, the axoneme was found closer to the membrane in new flagella (33 ± 11 nm; n = 7), than in old flagella (65 ± 16 nm; n = 7; Fig 5C). Surprisingly, we also found a fibrous structure in this space between the FC and the axoneme in the old flagellum.
In both chemically fixed electron tomograms of the FC, a novel, electron dense fibre was observed inside the old flagellum near the old FC layer and the flagellar membrane (Fig 6A). This fibre appears as a filament, ~20 nm by ~45 nm in cross-section and longer than the field of view in a single tomogram (Fig 6B). We have named this component the FC fibre. When tilted to show the flagellum in cross-section, the position of the FC fibre is off-centre towards the cellular side of the axoneme (Fig 6C). This is also shown in the 3D model, where the FC is consistently found between MT doublets 7 and 8 (Fig 6D). In tomograms of both FCs, the proximal extremity of the FC fibre initiates ~400 nm prior to the FC (Fig 6E). The complete length of the FC fibre (~870 nm) could only be measured in one tomogram, and its distal end extends ~200 nm further than the distal extremity of the old flagellum plate. The FC fibre had connections to the old flagellum FC plate (Fig 6F), but only a few connections to the doublet microtubules (Fig 6G).
In this paper, we have used a combination of fixation and imaging procedures to reveal the three dimensional ultrastructure of the FC of the procyclic form of T. brucei, a mobile cellular junction [14,17,18]. A combined detailed analysis of these data and previous publications on this structure shows that the FC behaves like a motorized double-sided vehicle that travels along microtubule doublets 7–9 in the old flagellum and in contact with microtubule doublets 3–5 in the new flagellum (Fig 7). Inside the new flagellum, the distance to the axoneme is 33 nm, a distance a kinesin molecule could easily span [24,25]. On the other side of the junction, the protein or protein complexes involved in linking the FC and the old axoneme must span the greater ~65 nm. This connection must also be very strong as it does not only move the FC, but also displaces the axoneme as it passes by.
We revealed structural variations of the FC and correlated them with the cell’s cycle stage judged by the length of the new flagellum. When the new flagellum is at, or close to, its stop point (at which the new flagellum tip stops translocating along the old flagellum[18]), the interstitial layer of the FC was reduced to about half the original thickness. The thickening of the old-flagellum layer of the FC in some of these flagella introduces the possibility of rearrangement of the external components. The maturation of the FC structure over the cell cycle is a novel finding and might hint towards the mechanism for removal of this structure after it has fulfilled its purpose.
A novel component of the FC was described—the FC-fibre. Because of this fibre’s morphology, length and location that all correlate well with those described of intraflagellar transport trains (IFT particles delivers flagellum building material to the flagellum tip using molecular motor proteins walking on the axonemal microtubules) [26–28], we suggest that it might represent a row of IFT particles (further discussion in S3 Fig and S1 Text).
This study also has methodological interest, as we studied the same structure with an unprecedented combination of electron microscopy methods. We showed that the high pressure frozen and freeze substituted FC appears more similar to the cryo-sectioned and cryo-visualised FC, with smooth flagellar membranes and unfragmented extracellular material. The thicknesses of the various FC plates vary depending on sample preparation (Table 1). Even though frozen hydrated sections of cells revealed the periodic organisation of the FC at the molecular level, cutting artefacts such as compression [22] added to the difficulty of visualizing the old and new flagellar plates especially when the FC plates are oriented perpendicular to the cutting direction. Therefore, all fixation methods are valuable for specific purposes, and quantitative results achieved using only one method should be interpreted with caution.
The FC structure provided by the 3D architecture presented here establishes a new level of insight into a junctional apparatus that possesses the capacity for lateral mobility. This insight into the FC substructure and morphogenesis is a necessary platform for future studies of molecular components and their assignment within the highly organized structure. In addition the structural definition will be critical for studies designed to reveal where the molecular motor is located and how it operates.
Logarithmically growing procyclic 427 cells in SDM-79 medium were fixed by a) adding 2.5% glutaraldehyde to the culture or b) high pressure frozen using Leica EM Pact II (Leica Microsystems, Vienna) as in [15,29].
In brief, chemically fixed cells were postfixed (2.5% glutaraldehyde, 2% formaldehyde in 100 mM phosphate buffer pH 7–7.4 for 2 hours; then 1% osmium tetroxide in 100 mM phosphate buffer for 1–2 hours), en-bloc stained (2% magnesium uranyl acetate in water for 2 h) and dehydrated with increasing concentrations of ethanol, immersed in propylene oxide and infiltrated by increasing concentrations of epon.
High pressure frozen samples were freeze substituted (2% uranyl acetate from a 20% methanolic stock solution, in dehydrated acetone for 1 h). Infiltration with increasing concentration of HM20 (3:1, 2:1, 1:1, 1:3, 0:1 acetone:HM20 for several hours each) was performed at -50°C, where polymerization using UV light was initiated. Polymerization was finished with 48 h UV illumination at room temperature.
Thin sections (75 nm) were cut using an UltraCut microtome (Leica Microsystems,Vienna), and post stained with 3 min lead citrate only (chemically fixed samples), or 8 minutes 2% uranyl acetate followed by 3 minutes Reynold’s lead citrate (high pressure frozen samples) [30].
Sections 250–300 nm thick was cut, post-section stained and 15 nm colloidal gold particles (BBInternational, Cardiff, UK) was applied to both surfaces of the grid. Serial sections incorporating the entire FC were imaged using a Ultrascan 785 4k x 4k camera binned to 2k x 2k (Gatan, Pleasanton, CA, USA) every degree, ±60°, at 23000 x magnification on a F30 Tecnai microscope (FEI Company, Eindhoven, The Netherlands), then rotated 90° and a second axis was acquired. Pixel size was ~1 nm. Tomograms were reconstructed using the IMOD software [31], and 3D models were made by outlining objects of interest in the tomograms.
The lengths of new flagella were measured by taking lower magnification images of the thick serial sections containing the cell in which the FC had been imaged. Serial sections were aligned and 3D models of the new flagellum were made.
Cells were prepared by harvesting with centrifugation and resuspended in 20% dextran and 0.2% sucrose in medium. Within 3–4 minutes of resuspension, cells were high pressure frozen and then treated as in [12]. In brief, 80–100 nm thick frozen hydrated sections were cut and for tomography imaged every 1.5° and tilted ±60°on F20 Tecnai microscope (pixel size 0.76 nm; FEI Company, Eindhoven, The Netherlands). Fourier transform image was made from a subarea of a single 0.76 nm slice in IMOD.
|
10.1371/journal.pgen.1005211 | Early Lineage Priming by Trisomy of Erg Leads to Myeloproliferation in a Down Syndrome Model | Down syndrome (DS), with trisomy of chromosome 21 (HSA21), is the commonest human aneuploidy. Pre-leukemic myeloproliferative changes in DS foetal livers precede the acquisition of GATA1 mutations, transient myeloproliferative disorder (DS-TMD) and acute megakaryocytic leukemia (DS-AMKL). Trisomy of the Erg gene is required for myeloproliferation in the Ts(1716)65Dn DS mouse model. We demonstrate here that genetic changes specifically attributable to trisomy of Erg lead to lineage priming of primitive and early multipotential progenitor cells in Ts(1716)65Dn mice, excess megakaryocyte-erythroid progenitors, and malignant myeloproliferation. Gene expression changes dependent on trisomy of Erg in Ts(1716)65Dn multilineage progenitor cells were correlated with those associated with trisomy of HSA21 in human DS hematopoietic stem and primitive progenitor cells. These data suggest a role for ERG as a regulator of hematopoietic lineage potential, and that trisomy of ERG in the context of DS foetal liver hemopoiesis drives the pre-leukemic changes that predispose to subsequent DS-TMD and DS-AMKL.
| An excess number of genes in trisomy on human chromosome 21 leads to the development of specific diseases in human Down syndrome. An excess copy of the gene, ERG, an ETS family transcription factor, has been implicated in abnormal blood system development in Down syndrome. In this study we show how trisomy of Erg in a murine Down syndrome model perturbs hematopoietic progenitor cells in a manner similar to that observed in human Down syndrome by inducing gene expression changes and lineage priming in early multi-potential progenitors. We show that the gene expression signature specifically attributable to trisomy of Erg in the murine model is strongly correlated with gene expression changes in human Down syndrome hematopoietic cells. The data suggest that Erg is an important regulator of megakaryocyte-erythroid lineage specification in multipotential hematopoietic cells and that trisomy of Erg in the context of DS prediposes to a transient myeloproliferative disorder and acute megakaryocyte leukaemia in a multi-step model of leukemogenesis.
| Down syndrome (DS) is the commonest human aneuploidy [1]. DS infants with trisomy of human chromosome 21 (HSA21) are uniquely predisposed to a transient myeloproliferative disorder (DS-TMD) and acute megakaryocytic leukemia (DS-AMKL) [2]. DS-TMD, usually characterised by the presence of peripheral immature myeloblasts/megakaryoblasts and the variable involvement of other organs, is restricted to the neonatal period, spontaneously regresses and is the result of genetic co-operation between trisomy of HSA21 gene(s) with an acquired somatic mutation in GATA1 in virtually all cases [3]. However, up to 30% of children will subsequently develop DS-AMKL, a malignancy clonally related to the preceding DS-TMD. Candidate gene analysis and genome-wide exome sequencing have identified somatic mutations and deletions implicated in the progression of DS-TMD to DS-AMKL, in genes including JAK1, JAK2, JAK3, FLT3, TP53, TRIB1, MPL, EZH2, APC, PARK-2, PACRG, EXT1, DLEC1 and SMC3, and further suggested that GATA-1 mutations alone in the context of HSA21 trisomy were sufficient for development of DS-TMD [4–12].
Preceding acquisition of GATA1 mutations, human DS foetal livers exhibit perturbed hematopoiesis. Increased numbers and clonogenicity of hematopoietic stem (HSC) and progenitor cells, increased frequency of bi-potential megakaryocyte-erythroid progenitors, and reduced numbers of granulocyte-macrophage-committed progenitor cells have been described [13–15]. This perturbation must be attributed to a specific trisomic gene or genes on HSA21 that drive the pre-leukemic DS phenotype from which DS-AMKL and DS-TMD subsequently arise. Murine DS models with germline transmissible segmental trisomies of human or murine orthologues of HSA21 genes have allowed genetic analyses of the contributions of genes within the DS critical interval to specific DS phenotypes [16–19]. A well studied model is the Ts(1716)65Dn mouse, which is trisomic for orthologs of ~104 human chromosome 21 genes [17]. Ts(1716)65Dn mice display progressive myeloproliferation chracterised by thrombocytosis, megakaryocyte hyperplasia, dysplastic megakaryocytic morphology and myelofibrosis. Similarly, blasts with erythro-megakaryocytic features and myelofibrosis are commonly observed in organs affected by DS-TMD/AMKL, while DS foetal livers show increased numbers of bipotential megakaryocyte-erythroid progenitors with increased clonogenicity and megakaryocyte/erythroid potential as well as megakaryocytosis [13–15].
We previously implicated the ETS family transcription factor ERG as a critical HSA21 gene in DS hematopoietic disease by demonstrating that specific reversion of Erg gene dosage to functional disomy, while the other ~103 orthologs remained trisomic, abrogated the myelo-megakaryocytic proliferation in Ts(1716)65Dn mice [20]. Erg has previously been shown to be essential for normal hematopoietic stem cell function [21–23]. Moreover, Erg deregulation can cause erythro-megakaryocytic leukemia in mice [24,25], and is implicated in acute myeloid and lymphoid malignancy in humans. In t(16;21) AML that carry the ERG/TLS-FUS fusion, complex karyotype AML with amplification of 21q, normal karyotype adult AML, MLL-rearranged paediatric AML and in T-ALL, high levels of ERG expression correlate with poor prognosis [26–29].
The detailed mechanisms by which Erg contributes in trisomy to myeloproliferation in Ts(1716)65Dn mice, and whether molecular changes specifically driven by three copies of Erg in this model reflect those associated with human DS, remained to be elucidated. To address these questions, we detailed hematopoietic progenitor perturbations associated with malignant myeloproliferation in 4 month-old Ts(1716)65Dn mice, the youngest age at which myeloid progenitor abnormalities have been observed [30]. We then performed transcriptome analysis of hematopoieitic stem cell (HSC) and myeloid progenitor cell-enriched populations from the bone marrow of Ts(1716)65Dn mice to define the relevant biological and genetic changes by which trisomy predisposes to development of myeloproliferation in this DS model [20]. Transcriptome changes in multipotential progenitor cells that were attributed specifically to trisomy of Erg in the Ts(1716)65Dn mouse were then compared to expression changes due to trisomy of chromosome 21 (HSA21) in human DS CD34+CD38- hematopoietic cells [31] to explore the role of ERG gene dosage in human disease.
The stem-cell enriched lineage-negative cKit+Sca1+ (LSK) population has been previously shown to be expanded in trisomic Ts(1716)65Dn DS mice [20,30]. This population, which functionally resembles the expanded stem cell-enriched compartment in human DS foetal livers [15], was corrected to wild-type levels in Ts(1716)65Dn mice when Erg was specifically reduced from trisomy to functional disomy [20]. Moreover, it has been observed that, in addition to changes in stem cell numbers, bipotential erythroid-megakaryocyte progenitor populations are also perturbed in human DS [15]. We therefore sought to define the effects of trisomy on specific murine hematopoietic progenitor cells in the Ts(1716)65Dn DS model. Ts(1716)65Dn trisomic mice were crossed to mice carrying the non-functional ErgMld2 allele as previously described [20]. The four resulting genotypes: mice trisomic for ~104 orthologs of human chromosome 21 genes including Erg (Ts65Dn/Erg+/+/+), mice disomic for functional Erg and trisomic for the remaining ~103 genes in the trisomic segment (Ts65Dn/Erg+/+/Mld2), euploid mice (Erg+/+) and disomic mice with one functional Erg allele (Erg+/Mld2), were analysed at 4 months of age for abnormalities in common myeloid progenitors (CMP), granulo-monocytic progenitors (GMP), and megakaryocyte-erythroid progenitors (MEP) as previously defined [32]. Consistent with previous data [30], increased numbers of GMPs and a deficit of MEPs were evident in Ts65Dn/Erg+/+/+ mice (Fig 1A). These abnormalities were corrected in Ts65Dn/Erg+/+/Mld2 mice (Fig 1A), suggesting that trisomy of Erg is specifically associated with perturbations of myeloid progenitors of several hematopoietic lineages in the Ts(1716)65Dn DS model.
Myeloid progenitors in Ts65Dn/Erg+/+/+ mice were then examined in greater detail, with particular emphasis on the recently described series of progressively maturing BEMP, CD150+CD9hi and CD150+FcγR+ bipotential erythroid-megakaryocyte progenitors [33,34], as well as pre-granulocyte macrophage (PreGM), granulocyte-macrophage (GMP), pre-colony forming unit erythroid (PreCFU-E) and colony forming unit erythroid (CFU-E) populations [33,35], defined in S1 Table (Fig 1B and 1C). Trisomic Ts65Dn/Erg+/+/+ mice contained increased numbers of GMP and CD150+CD9hi bipotential progenitor cells, and these were normalised in number when Erg was reduced to functional disomy in Ts65Dn/Erg+/+/Mld2 mice. Conversely, the numbers of GMPs and bipotential megakaryocyte-erythroid progenitor populations (BEMP, CD150+CD9hi, CD150+FcγR+) were reduced in Erg+/Mld2 mice haploinsufficient for functional Erg. Trisomic Ts65Dn/Erg+/+/+ mice were noted to have significantly fewer CFU-E, an anomaly that was not evident in Ts65Dn/Erg+/+/Mld2 mice and Erg+/Mld2 mice had an expanded population of these late erythroid progenitors, supporting previous data that suggested Erg normally restrains terminal erythroid differentiation [22,24,25]. Consistent with the immunophenotypic analyses, in clonogenic assays, Ts65Dn/Erg+/+/+ bone marrow demonstrated stimuli-specific increases in the numbers of granulocyte, macrophage and megakaryocyte colony-forming units (CFU). This myeloproliferation was moderated in Ts65Dn/Erg+/+/Mld2 mice (Table 1). Together, these data support a role for Erg in regulation of multiple hematopoietic lineages. The bias toward megakaryocyte and away from erythroid progenitor formation that was specifically attributable to trisomy of Erg in Ts(1716)65Dn mice resembles progenitor abnormalities observed in pre-leukemic human DS foetal livers prior to acquisition of GATA1 mutations, although human DS foetal livers exhibit reduced numbers of GMP, in contrast to the mouse model [15].
We undertook gene expression profiling of prospectively isolated LSK, CMP, GMP and MEP populations from Ts65Dn/Erg+/+/+, Ts65Dn/Erg+/+/Mld2, Erg+/+ and Erg+/Mld2 mice. Gene expression changes within each cell population that were attributable to changes in functional Erg gene dosage were determined by specific pair-wise comparisons using linear modeling and empirical Bayes moderated t-statistics [36]. Changes due to full trisomy of all ~104 syntenic genes were inferred from comparison of Ts65Dn/Erg+/+/+ versus Erg+/+ mice, trisomy-induced changes specifically attributable to trisomy of Erg were evident from comparison of Ts65Dn/Erg+/+/+ versus Ts65Dn/Erg+/+/Mld2 (effects specific to Erg trisomy), changes due to trisomy with two copies of functional Erg (effects due to non-Erg gene trisomy) emerged from comparison of Ts65Dn/Erg+/+/Mld2 versus Erg+/+, and changes due to Erg haplo-insufficiency from comparison of Erg+/Mld2 versus Erg+/+.
We initially explored the expression in LSK, CMP, GMP and MEP cells of the ~104 HSA21 orthologs present in trisomy in Ts65Dn/Erg+/+/+ mice compared with their expression in Erg+/+ cells. Increased expression of these genes in trisomc cells varied and was also dependent on the specific hematopoietic cell type examined (Fig 2A). At a false discovery rate of 5% across the entire transcriptome, Son, Usp16, Cyrzl1, Gart, Cct8 were upregulated in LSK cells, Sfrs15, Ifngr2, Ifnar2, Gart, Atp5j in CMPs, and Gart, Chaf1b, Hlcs, Ttc3 in GMPs of Ts(1716)65Dn mice (Fig 2A and S2 Table).
The greatest number of gene expression changes due to full trisomy (Ts65Dn/Erg+/+/+ versus Erg+/+) or specifically to Erg trisomy (Ts65Dn/Erg+/+/+ versus Ts65Dn/Erg+/+/Mld2) occurred in the LSK compartment, while gene expression changes attributable to trisomy of non-Erg genes in the interval (Ts65Dn/Erg+/+/Mld2 versus Erg+/+) were fewer and occurred principally in MEP and GMP populations (Table 2). In addition, in LSK cells there was a strong degree of overlap in differentially expressed genes due to full trisomy and specifically due to trisomy of Erg, both in up- and down-regulated genes (Fig 2B, 2C and 2D).
Given the correlation of gene expression changes induced by full trisomy with those associated specifically with trisomy of Erg, a Genuine Association Analysis (GENAS) was undertaken. GENAS is a relatively new statistical technique that estimates the biological correlation between two differential expression profiles, correcting for any technical correlation and for the statistical uncertainty with which the fold changes are estimated [37,38]. Unlike the initial analysis in Fig 2B, GENAS does not depend on a significance cut-off and instead calculates an overall correlation using all expressed genes. In confirmation of the initial analyses, there was a very high degree of biological correlation in LSK cells between gene expression changes caused by trisomy of the full HSA21 syntenic segment and those associated specifically with trisomy of Erg. The strongest correlation between expression changes resulting from full trisomy and those due to trisomy of non-Erg genes occurred within MEP cells (Fig 3). Notably, the degree of biological correlation between gene expression changes attributable specifically to Erg trisomy and those due to trisomy of non-Erg genes was weak in all cell types analysed, suggesting gene expression changes induced specifically by trisomy of Erg were distinct from the combined effect of other genes within the Ts(1716)65Dn trisomic segment.
Together, these findings establish that in hematopoietic progenitor cells, Erg is the dominant influence within the Ts(1716)65Dn trisomic interval, three copies of which specifically drives gene expression changes that are evident primarily within multipotential hematopoietic progenitor populations rather than more lineage-restricted progenitors.
We next sought to understand how gene expression changes induced specifically by trisomy of Erg could lead to perturbations of hematopoietic progenitors and the myeloproliferative phenotype observed in Ts(1716)65Dn mice. To do this, we undertook an expression signature analysis of the expression changes in the Ts65Dn/Erg+/+/+ versus Ts65Dn/Erg+/+/Mld2 cells using curated gene sets from the Molecular Signatures database (MSigDB version 2.5) and pathway gene sets defined by the BioCarta, KEGG and Reactome databases. A ROAST test was conducted for each signature gene set. ROAST is a gene set test suitable for small samples and linear models that accounts for inter-gene correlation [38,39]. ROAST evaluates whether the overall expression signature defined by a gene set is up- or down-regulated within a specific comparison. This analysis found that expression signatures associated with progenitor cells of specific hematopoietic lineages, including granulocyte-monocyte progenitors, megakaryocytes, platelets and platelet processes, were significantly upregulated in the Ts65Dn/Erg+/+/+ cells (ROAST P-value < 0.05, S3 Table). By contrast, signatures associated with myeloproliferation were not significantly enriched in Ts65Dn/Erg+/+/+ versus Ts65Dn/Erg+/+/Mld2 LSK cells.
Given these findings, and as we had observed perturbations in specific hematopoietic progenitor populations in Ts65Dn/Erg+/+/+ bone marrow (Fig 1 and Table 1), we examined whether the gene expression changes specific to trisomy of Erg in Ts(1716)65Dn mice could be interpreted in terms of cell lineage priming. We generated gene expression signatures that define specific cell populations: LSK, GMP and their precursors (Pre-GM Flt3+, Pre-GM Flt3-), megakaryocytes and their progenitors (BEMP, CD150+ CD9hi, CD150+ FcγR+), and CFU-E and pre-CFU-E. An expression signature was defined for each cell population by compiling the significantly up-regulated genes in that cell type compared to the average of all the other populations, together with the magnitude of the up-regulation for each gene as measured by the moderated t-statistic (S4 Table). ROAST tests were then conducted, with genes weighted by their magnitude of change, to determine whether these signatures were associated with the gene expression changes due to full trisomy (Ts65Dn/Erg+/+/+ versus Erg+/+), the specific effects of Erg trisomy (Ts65Dn/Erg+/+/+ versus Ts65Dn/Erg+/+/Mld2), the effects of trisomy of non-Erg genes (Ts65Dn/Erg+/+/Mld2 versus Erg+/+) and Erg haploinsufficiency (Ts65Dn/Erg+/+ versus Erg+/Mld2) in LSK, CMP, GMP and MEP populations.
Signatures of more committed progenitors, particularly multipotential PreGM Flt3- cells, granulo-monocytic progenitors (PreGM Flt3+ and GMP), bipotential megakaryocyte-erythroid progenitors (CD150+CD9hi and CD150+FcγR+) and megakaryocytes, were enriched in the gene expression changes due to trisomy of Erg in LSK cells, and this was also evident in CMPs by this comparison (Fig 4). In LSK cells, this was accompanied by downregulation of the normal LSK gene signature indicative of a more differentiated profile (Fig 4). In contrast, erythroid progenitor signatures (PreCFU-E and CFU-E) were enriched in the gene expression changes caused by haploinsufficiency of functional Erg in LSK and CMP cells (Fig 4). Strikingly, these genetic changes attributable specifically to trisomy of Erg in Ts(1716)65Dn LSKs and CMPs were concordant with the increased numbers of GMP and megakaryocyte-committed cells and fewer erythroid progenitors that were evident in Ts65Dn/Erg+/+/+ but not Ts65Dn/Erg+/Mld2+ mice, as well as the relative enrichment of erythroid progenitors in mice with Erg haploinsufficiency (Fig 1).
Finally, we investigated the expression of genes implicated by somatic mutations and deletions in the progression of DS-TMD to DS-AMKL. Of these genes, expression of the murine orthologues for JAK1, JAK2, JAK3, FLT3, TP53, TRIB1, MPL, EZH2, APC, EXT1 and SMC3 was detected in the microarray data from LSK, CMP, GMP or MEP cells. While modest differences were evident in expression of some of these genes in comparisons between cells from Ts65Dn/Erg+/+/+ and Erg+/+ mice, differential expression of these genes did not reach statistical significance (S1 Fig and S2 Table).
Thus, the data support a model in which gene expression changes attributable specifically to trisomy of Erg in Ts(1716)65Dn mice occur primarily in multipotential hematopoietic cells with priming for specific hematopoietic progenitor lineages leading to myeloid progenitor perturbation, myeloproliferative changes and megakaryocytosis.
Transcriptome profiling of human DS Lin-CD45+CD34+CD38- bone marrow cells enriched for HSCs and multipoential hematopoietic progenitors that had been previously combined with transcriptomes of DS neurospheres derived from foetal cortical precursors, demonstrated differences in gene expression when compared to diploid controls [31]. We sought to establish if DS Lin-CD45+CD34+CD38- bone marrow cells specifically demonstrated gene expression changes associated with trisomy of HSA21 when compared to diploid controls, and if these were related to differential gene expression due to Erg trisomy in LSKs in the Ts(1716)65Dn DS model. From the published data [31], we derived gene expression signatures for human DS HSC and multipotential progenitor cells that were induced by trisomy of HSA21 (S5 Table). We then used ROAST to compare these human DS gene signatures to gene expression changes induced by full trisomy in Ts(1716)65Dn LSK cells, as well as those specific to trisomy of Erg and those due to trisomy of non-Erg genes (Fig 5). The human DS gene signatures were strongly and significantly correlated with gene expression changes due specifically to Erg trisomy in Ts(1716)65Dn LSK cells, but not with those attributable to trisomy of non-Erg genes (Fig 5A). The complementary analysis was then undertaken. ROAST tests showed that gene expression changes specifically induced by trisomy of Erg in Ts(1716)65Dn LSK cells significantly correlated with gene expression changes due to trisomy of HSA21 in human DS Lin-CD34+CD38- cells (Fig 5B). This genetic data supports the role of ERG as a key gene on HSA21 which in trisomy, drives the hematopoietic phenotype in human DS.
We show here, in comparisons between Ts65Dn/Erg+/+/+ mice, in which a ~104-gene interval syntenic to HSA21 is present in trisomy, and their Ts65Dn/Erg+/+/Mld2 counterparts, in which the Erg gene within this interval is specifically reduced to functional disomy, that trisomy of Erg is specifically required for the characteristic perturbations in specific hematopoietic progenitor cell populations in the Ts(1716)65Dn DS model. Unlike transgenic and retroviral models of induced Erg expression, the use of the non-functional ErgMld2 allele in this model allows accurate characterisation of effects of functional Erg gene dosage while expression is endogenously regulated. The hematopoietic changes in Ts(1716)65Dn mice reflected the pre-leukemic changes observed in human DS foetal livers prior to acquisition of GATA-1 or other somatic mutations. While these analyses do not exclude contribution from other genes within the trisomic interval, they clearly establish that three copies of Erg is specifically required to drive transcriptome changes in HSCs and early multipotential progenitors in Ts(1716)65Dn mice. Moreover, the data provide a mechanism by which deregulation of Erg leads to perturbation of myeloid progenitor populations, myeloproliferation and megakaryocytosis in trisomy. Our findings show that perturbation of functional Erg gene dose by trisomy or by haploinsufficiency, is associated with lineage specific gene expression changes and "lineage priming" of multipotential hematopoietic progenitor cells. Notably, the effects specific to trisomy of Erg on gene expression and progenitor cell alterations in Ts(1716)65Dn mice were strongly reflected in Erg+/Mld2 mice, but in the opposite direction. Finally, we demonstrate that the gene signature specifically associated with trisomy of Erg in Ts(1716)65Dn LSK cells is directly related to gene expression changes in human DS Lin-CD34+CD38- cells enriched for HSCs and early multipotential progenitors. These data support a key role for ERG as a critical gene in trisomy of HSA21 that drives hematopoietic changes in human DS.
Previous analyses established that Erg is an important regulator of HSC self-renewal after establishment of definitive hematopoiesis in the embryo [23] and in emergency hematopoiesis in adult HSCs [21,22]. The data presented here imply additional roles for Erg for lineage priming in early multipotential hematopoietic cells which affects subsequent myeloid lineage development (Fig 6). This includes granulocyte-monocyte progenitors and bipotential erythroid-megakaryocyte progenitors, with effects on megakaryocyte and erythroid lineage specification being of particular relevance. In these latter roles, haploinsufficiency of functional Erg resulted in fewer bipotential erythroid-megakaryocyte progenitors and a propensity toward CFU-E formation, consistent with previous transplantation studies which demonstrated a bias toward erythroid lineage formation from Erg+/Mld2 HSCs [22]. Trisomy specifically related to Erg in the Ts(1716)65Dn model led to expansion of bipotential megakaryocyte-erythroid progenitors, fewer committed erythroid progenitors and megakaryocytosis [20]. This finding is consistent with changes in human DS foetal livers [15] and other independent murine models of Erg overexpression [24,40]. Indeed, aberrant megakaryocyte-erythroid differentiation and erythroid maturation block may be a potential unifying mechanism for Erg in predisposing to acute bipotential megakaryocytic-erythroid leukemia [25]. The selective expansion of CD150+CD9hi cells within bipotential megakaryocyte-erythroid progenitor compartment in Ts(1716)65Dn mice is similar to that observed in murine models of thrombopoietin driven myeloproliferation [33,34], providing additional evidence that this progenitor population correlates with the degree of megakaryocytosis in disease models. It was also noted that Ts(1716)65Dn mice exhibited increased numbers of GMP, while human DS foetal livers exhibit reduced numbers of these progenitors [15]. This discrepancy may by attributable to the comparatively low level of endogenous Erg expression in murine models relative to human hematopoietic cells [40]. Indeed, in a transgenic model of ERG overexpression, reduction in GMP number was observed in murine foetal livers [40], suggesting the level of gene expression and the hematopoietic stage is important in explaining this apparently discordant phenotype.
It is notable that the expansion of the LSK compartment in Ts65Dn/Erg+/+/+ mice, which is specifically dependent on trisomy of Erg [20], reflects expansion of HSPCs in DS foetal livers [15], and was associated with downregulation of the LSK gene signature and up-regulation of progenitor cell-specific gene expression patterns, making trisomic LSKs less "stem cell-like". This suggests that an excess of Erg favors progenitor specification at the expense of HSC maintenance. This finding is in keeping with the increased multi-potential pre-progenitor cell frequency in Ts65Dn/Erg+/+/+ bone marrow [20] and the increased propensity to form "blast-my" colonies upstream from CFU-GEMM in DS foetal liver [15]. These observations also argue against increased Erg gene dose inducing a program of self-renewal as a leukemogenic mechanism in the development of DS-AMKL and DS-TMD. Indeed, competitive transplantation assays using the Ts1Rhr trisomic murine model of DS, which carries a trisomic segment that includes the Erg gene, demonstrated fewer competitive repopulating units compared to disomic controls [41]. Nevertheless, the observation that the multipotential PreGMFlt3- progenitor signature was also upregulated in Ts65Dn/Erg+/+/+ CMPs relative to Ts65Dn/Erg+/+/Mld2 suggests that in trisomy, Erg allows maintenance of multipotential progenitor potential while promoting specific lineage commitment toward megakaryocyte and granulocyte-monocyte differentiation. This would be in keeping with the increase in numbers of total colonies, particularly megakaryocyte-containing colonies, observed by in vitro culture experiments.
It remains to be determined whether the consequences of genetic perturbations due to trisomy and effects specific to ERG trisomy in the context of DS as our data may suggest, could be amenable to therapeutic targeting in the prevention or treatment of DS-TMD or DS-AMKL in addition to currently available strategies.
Derivation and genotyping of the ErgMld2 mutant allele has been described [21]. Ts(1716)65Dn mice (The Jackson Laboratory) were maintained as previously described [20]. All mice were derived from the first-generation progeny of matings between Erg+/Mld2 and Ts(1716)65Dn mice and genotyped as previously described [42]. This study was performed in accordance with the Australian Code for the Care and Use of Animals for Scientific Purposes, published by the Australian National Health and Medical Research Council. Procedures were approved by the Walter and Eliza Hall Institute of Medical Research Animal Ethics Committee (Approval number 2012.003).
Single-cell suspensions from bone marrow were prepared in balanced salt solution (0.15 M NaCl, 4 mM KCl, 2 mM CaCl2, 1 mM MgSO4, 1 mM KH2PO4, 0.8 mM K2HPO4, and 15 mM N-2-hydroxyethylpiperazine-N'-2-ethanesulfonic acid supplemented with 2% [vol/vol] bovine calf serum). Clonal analysis of bone marrow cells (2.5x104) was performed in 1 mL semisolid agar cultures of 0.3% agar in Dulbecco/s modified Eagles medium containing 20% newborn calf serum and stem cell factor; SCF (100 ng/mL), erythropoietin; EPO (2 U/mL), and interleukin-3; IL-3 (10 ng/mL), granulocyte colony stimulating factor; G-CSF (103 U/mL), granulocyte-macrophage colony stimulating factor; GM-CSF (103 U/mL) and/or macrophage colony stimulating factor; M-CSF (103 U/mL). Cultures were incubated at 37°C for 7 days in a fully humidified atmosphere of 10% CO2 in air, then fixed, dried onto glass slides, and stained for acetylcholinesterase, Luxol fast blue, and hematoxylin, and the number and type of colonies were determined.
Staining was performed using rat anti-mouse biotinylated or fluorochrome-conjugated antibodies specific for Ter119 (Ly-76), Gr1 (Ly6G and Ly6C), Mac1 (CD11b), B220 (CD45R), CD4, CD8, CD41, CD34, CD16/32, Sca1 (Ly6A/E), cKit (CD117) and CD150 (Biolegend) and IL-7 receptor α (IL7Rα), CD48, CD105, and CD9 (eBioscience). Secondary staining used streptavidin PE-Texas-Red (BD Pharmingen). Cells were analyzed using a LSR Fortessa flow cytometer (Becton Dickinson), or cells were sorted using a FACSAria II (Becton Dickinson) flow cytometer after antibody staining with lineage depletion.
Student's unpaired two-tailed t-tests were used using GraphPad Prism v. 5.0a for Mac Os X (GraphPad Software), unless otherwise specified.
Bone marrow LSK, CMP, GMP and MEP populations were isolated by FACS from 2–4 mice including males and females from each genotype (Ts65Dn/Erg+/+/+, Ts65Dn/Erg+/+/Mld2, Erg+/+ and Erg+/Mld2) at ~ 4 months of age. Total RNA was isolated from 100,000–500,000 cells pooled from genotype and cell population matched samples using the RNeasy Micro kit (Qiagen). RNA quality was assessed with the Agilent Bioanalyzer 2100 (Agilent Technologies) by using the Agilent RNA 6000 Nanokit (Agilent Technologies) according to the manufacturer's protocol. Up to 200ng of RNA was labelled with the Total Prep RNA amplification kit (Ambion), and complementary RNA (1.5 μg) was hybridized to 48 arrays using six Illumina Mouse WG-6 V2.0 Expression BeadChips (Illumina, Inc., San Diego, CA) according to Illumina standard protocols. The resultant microarray probe level data were analyzed by using the limma software package Version 3.21.1 [38]. Raw intensities were normalized by using the neqc function, which performs normexp background correction followed by quantile normalization using control probes [43]. Probes were filtered if not detected in any sample (detection P value < 0.05). Pairwise comparisons were made by using linear modeling and empirical Bayes moderated t statistics [36]. Empirical array quality weights were estimated and incorporated into the linear models [44]. Allowance was made for possible correlations between RNA samples drawn from the same pool of mice [45]. The false discovery rate (FDR) was controlled by using the Benjamini-Hochberg algorithm. Probes with FDR of less than 5% were considered to be differentially expressed. The microarray data have been deposited to the Array Express database (http://www.ebi.ac.uk/arrayexpress) with accession number E-MTAB-2574.
Validation of the microarray data was undertaken via RNA-seq. Total RNA was extracted using the RNeasy Plus minikit (Qiagen) from LSK cells sorted independently from three trisomic Ts65Dn/Erg+/+/+ (two male and one female) and Erg+/+ controls (two male and one female) at ~ 4 months of age. Sequencing was performed on an Illumina Hi-Seq 2500, producing at least 15.9 million 100bp paired-end reads per sample. Reads were mapped to the mm10 mouse genome (Genome Reference Consortium GRCm38) using the Subread aligner [46]. Read counts were summarised at the gene level by featureCounts [47] using NCBI RefSeq gene annotation. Differential expression analysis utilised the edgeR [48] and limma software packages. Genes were filtered as not expressed if they failed to achieve at least 0.5 counts per million reads in at least 2 of the 6 samples. All Entrez gene IDs without an official symbol were removed from further analysis, as were Y chromosome genes, Xist, and immunoglobulin genes, leaving 14,551 genes for downstream analysis. Library sizes were normalised using the TMM method [49]. The voom function of the limma package was used to convert read counts to log2 counts per million with associated precision weights [50]. Differential expression was assessed using empirical Bayes moderated t-statistics [36]. There was a strong correlation between the gene sets identified as differentially expressed by microarray (S2 Table) with RNA-seq data in Ts65Dn/Erg+/+/+ versus Erg+/+ LSK cells (ROAST P value = 0.00795) and of the differentially expressed genes identified from the microarray comparison, over fifty genes were also identified as differentially expressed using non-pooled independent samples by RNA-Seq with a P value of less than 0.05.
Affymetrix microarray CEL files containing gene expression profiles of human DS CD34+CD38- HSCs and multipotential progenitors [31] were obtained from the Paterson Institute for Cancer Research at the University of Manchester and analysed with the affy (version 1.34) and limma software packages. Raw intensities were background corrected, normalised and summarized using the Robust Multiarray Average algorithm. Pairwise comparisons were made by using linear modeling and empirical Bayes moderated t statistics.
Gene sets from the Molecular Signatures Database (Broad Institute, Version 2.5) were mapped from human to mouse orthologs (http://bioinf.wehi.edu.au/software/MSigDB/). Genuine Association analysis [37] used the genas function of the limma package. Rotational gene set tests (ROAST) were performed with the roast function of the limma package, using the “mean” set statistic, array quality weights, Holm modification for multiple testing and 10000 rotations [39]. Moderated t-statistics were used to weight genes in the gene set tests. Barcode plots were made using the barcodeplot function of the limma package.
Heatmaps were plotted using the Heatmap.2 function from the gplots software package, using Pearson correlation for hierarchical clustering for rows and columns. Gene signatures were represented in a heatmap as Z score equivalents. The Z scores were derived from the standard normal distribution and correspond to the continuity-corrected single-tailed P values obtained from the ROAST tests, with a positive Z score for an up-regulated gene set and negative Z score for a down-regulated gene set.
|
10.1371/journal.pbio.1001319 | Competing Sound Sources Reveal Spatial Effects in Cortical Processing | Why is spatial tuning in auditory cortex weak, even though location is important to object recognition in natural settings? This question continues to vex neuroscientists focused on linking physiological results to auditory perception. Here we show that the spatial locations of simultaneous, competing sound sources dramatically influence how well neural spike trains recorded from the zebra finch field L (an analog of mammalian primary auditory cortex) encode source identity. We find that the location of a birdsong played in quiet has little effect on the fidelity of the neural encoding of the song. However, when the song is presented along with a masker, spatial effects are pronounced. For each spatial configuration, a subset of neurons encodes song identity more robustly than others. As a result, competing sources from different locations dominate responses of different neural subpopulations, helping to separate neural responses into independent representations. These results help elucidate how cortical processing exploits spatial information to provide a substrate for selective spatial auditory attention.
| When a listener is presented with many sound sources at once, it is easier to understand a particular source when it comes from a different spatial location than the other competing sources. However, past studies of auditory cortex generally find that in response to a single sound source, there is not a precise representation of spatial location in the cortex, which makes this effect of spatial location hard to understand. Here, we presented zebra finches with two simultaneous sounds (a birdsong target and a noise masking sound) from distinct spatial locations and recorded neural responses in field L, which is analogous to primary auditory cortex in mammals. When the target sound was presented by itself, the location of the source had little effect on the ability to identify the target song based on neural activity in field L. However, when the target was presented with a masker sound, the location of both sources strongly affected neural discrimination performance. Moreover, different subpopulations of neurons preferentially encoded either target or masker, providing a potential substrate for spatial selective attention. Thus, even though location is not well coded in cortical neurons, spatial information strongly modulates cortical responses.
| Past studies of spatial effects in auditory cortex have focused on how spatial location is encoded. These studies typically find that single-unit spatial tuning in cortex is weak [1]–[4], not topographically organized [5],[6], and not encoded independently of other perceptually important features [7]. There is good evidence for a specialized “where” pathway in auditory cortex, in which spatial information plays a larger role than in other cortical areas [8]. However, although we know of no single study that directly compares spatial tuning in cortex to that of lower stages of the auditory pathway, spatial tuning of cortical neurons is generally broad compared to both behavioral sensitivity [4] and spatial encoding in the midbrain [9],[10]. One hint for how to resolve these apparent discrepancies is that in an awake animal performing a spatial task, spatial information in cortical responses is enhanced [10]. Together, these results suggest that although spatial information is available, it is not the primary feature represented in the cortical auditory regions. Instead, spatial information may modulate neural responses in a way that depends on task demands, thus enabling analysis of sound sources in realistic auditory scenes [11],[12].
It may be that spatial effects are not best revealed by looking at how well source location is encoded by neural responses, but rather by examining how source location affects other aspects of information in cortical spike trains. In everyday perception, source location matters most in auditory scenes in which sounds compete with each other. Although listeners can localize a sound source in quiet, this ability is degraded in more typical, real-world settings containing reverberant energy or competing sources [13]. In contrast, in exactly those kinds of realistic situations where there are competing sources, spatial separation helps listeners segregate sounds and enables them to focus selective attention, a critical skill for understanding a source of interest [14],[15]. In this sense, behavioral results support the idea that the locations of competing sources strongly influence auditory perception, regardless of whether the listener can effectively localize in such a setting.
Motivated by these observations, we hypothesized that the effects of spatial location on cortical processing would best be revealed by a study that uses competing sound sources. Rather than focusing on how accurately spatial location of a source was encoded, we explored how competing source locations influenced the ability to encode the identity of a target communication signal (in this case, birdsong). We found that, consistent with our hypothesis, source location of a target song presented in isolation had little effect on how well neurons in avian field L (the analog of mammalian primary auditory cortex [16]) encoded song identity; however, in the presence of a competing noise masker, both target and masker locations strongly influenced encoding of song identity. Moreover, depending on the location of target and masker, different neurons were “best” at encoding identity. Such a coding scheme may provide a substrate for spatial auditory attention, as top-down modulatory control signals could selectively suppress responses of neurons favoring a masker in order to reduce competition and allow more precise analysis of a target from a desired location.
We recorded neural responses from male zebra finches in the auditory forebrain (field L, based on stereotactic coordinates [17]–[19]) to stimuli from four azimuthal locations in the frontal hemifield. Target stimuli were two conspecific songs, presented either in quiet (“clean”; Figure 1A) or in the presence of a spectrally similar noise masker coming from the same or a different location as the target song (Figure 1B). We assessed neural performance using a single-trial spike-distance-based [20] nearest-neighbor classification scheme [21], calculating a percent correct score that indicates how well neural responses coded stimulus identity. Chance performance was 50%. Consistent with prior studies [17],[22]–[24], rate coding alone was insufficient to allow reliable stimulus discrimination; mean performance when no masker was present was only 54%, averaged across recording sites.
In each experimental session, there were four loudspeaker locations, leading to 16 target-masker spatial configurations. If the recording electrode was in the left hemisphere, loudspeaker locations were on the left side (−90°, ipsilateral to the electrode), in front (0°), halfway between front and right (+45°), and on the right (+90°, contralateral). These locations were flipped about the midline when recording in the right hemisphere. Henceforth, coordinates are referenced to the recording electrode, so that ipsilateral azimuths have negative signs and contralateral azimuths have positive signs. Discrimination performance was calculated for all 16 configurations and three signal-to-noise ratios (SNR; −6 dB, 0 dB, +6 dB). To assess the extent to which the head created an acoustical obstruction (“head shadow”) to the ear opposite the sound source, we measured sound level at both ears from all four locations using a masker token as the probe stimulus. The differences between left and right ears were 1.5, 0.1, −0.8, and −1.3 dB for −90, 0, +45, and +90°, respectively.
For the example site in Figure 1A and B, clean performance was near ceiling at all tested locations. Masked performance was much lower and varied substantially as the target was moved from the ipsilateral side (−90°) to the contralateral side (+90°), holding the masker at −90°. Across recording sites, the masked performance varied much more than clean performance did as a function of location. To quantify this, we computed the spatial sensitivity (defined as the difference between the best and worst performance for a given experimental condition; see Materials and Methods) for each site for both clean and masked targets. Spatial sensitivity was 3-fold higher with a masker present than without (p<.001; Figure 1D). The driven spike rate in response to clean songs did not vary significantly with location (r = .16, p = .068; Figure 1E). This distinction is important: while target azimuth was at best weakly coded by the rate response of the neurons, information about song identity encoded in spike trains varied greatly with target and masker locations.
The way in which classification performance varied with spatial configuration varied from site to site. Indeed, some sites responded best when the target was in a particular hemisphere (Figure 2A, site 1), some for a particular target-masker location configuration (sites 2 and 3), and some in idiosyncratic configurations that fit no simple description (site 4).
To explore how such a population of neurons might encode song identity, we considered two population-coding schemes. The first was based on a previous study, which assumed that behavioral performance was determined by the best thresholds across a population of neurons, an approach termed the “lower envelope” principle [25]. Here we define the corresponding neural “upper envelope” as the best classification performance across the entire neuronal population. The performance of individual sites and the upper envelope are shown in Figure 2B as a function of target and masker location for an SNR of −6 dB. While no one site performs well for all spatial configurations, almost all configurations yield at least some sites that encode target identity well.
At higher SNRs, the upper envelope is at ceiling (Figure 3A). To better reveal the effects of spatial configuration, we calculated the mean performance across sites for each spatial configuration. Despite the complex dependence of performance on spatial configuration for many of the sites, the mean performance varies smoothly with spatial configuration for each SNR. Specifically, mean performance is best when the target is contralateral and the masker ipsilateral to the neural recording site and worst in the reverse configuration (Figure 3B). Figure 3C shows the mean performances across sites in which the target is farther than the masker from the recording site, in the contralateral direction. Representing the data this way assumes a simple population model in which the neurons in one hemisphere are favored over the other (i.e., the responses from the hemisphere contralateral to the target are enhanced and the ipsilateral responses are suppressed). Using this model (which includes only the values in the lower right half of the grids in Figure 3B, including the diagonal), the effect of spatial separation (as well as SNR) is highly significant (p<.001 for both); moreover, linear regression fits at each SNR show that performance improves with increasing spatial separation of target and masker. Such performance increases are parallel with results from behavioral studies in humans [26] and birds [27] that report spatial unmasking.
Maskers degrade responses to target songs. A simple way to evaluate the masker interference is to compare the response elicited by the target in quiet to that of responses to the target plus masker. Differences between the two responses can be categorized into orthogonal categories of spike additions, where the presence of the masker causes extra spikes (usually in the gaps between syllables), and spike subtractions, where spikes that are elicited by the target alone are reduced by the presence of the masker (usually during syllables; see Figure 4A–C). Both types of interference have been studied before [18]; here we extended that analysis. We modeled spike trains that had only subtractions or only additions (Figure 4D; see Materials and Methods), and then calculated performance for these modeled spike trains just as we did for the measured ones.
We first validated our modeling approach by comparing predictions for modeled spike trains containing both additions and subtractions (i.e., the full effect of the masker) to measured data (see Figure 4, “modeled” rasters and performances). The example model rasters look similar to the measured masked spike trains, and target song identification performance closely matched performance using the masked spike trains. These results validate our methods for modeling additions and subtractions.
Following validation, we modeled spike trains that included only spike additions or only spike subtractions to separate their relative effects on performance. When modeling spike additions only (i.e., when no subtractions were modeled), target identification was better than for the measured response. On the other hand, performance for subtractions-only spike trains was only slightly better than the measured responses for two of the three configurations. For the target-contralateral, masker-ipsilateral configuration (right column of Figure 4), performance was essentially equal for the subtractions-only and masked spike trains. These results suggest that additions did not impair discrimination performance when the target was contralateral to the recorded site. However, including additions had some impact on the other two configurations. Overall, this analysis shows that the masker degraded performance more by preventing spikes that a clean target would have elicited than by causing additional spikes.
The times at which spikes are likely to be added by the masker tend to occur when the clean response rates are low. This can be quantified by correlating the clean stimulus response rate (Figure 4A) with the rate of subtractions (blue depths in Figure 4C) as a function of time. This correlation is significant and negative, confirming that subtractions reduce spikes the most when the likelihood of a spike in response to the clean stimulus is great (r = −.75, p<.001). In contrast, the correlation between the time-dependent spike additions (red peaks in Figure 4C) and the clean rate is weak (r = .08, p<.001). Taken together, these results suggest that the effect of removing spikes from the peaks interferes with target identification more than adding spikes. This holds true even in spatial configurations where the number of spikes added is greater than the number of spikes removed.
Here we show that, in quiet, sound source location has only a modest impact on coding of song identity in field L, an analog of auditory cortex [16]. In general, spatial tuning in brainstem is sharper than in cortex, demonstrating that cortical auditory neurons do not directly inherit the already encoded spatial information present in lower centers of the auditory processing stream [1],[28]. However, our results show that the spatial configuration of competing sources strongly affects the coding of those sources' content.
Spatial effects in cortical neurons are far greater when there are competing sounds than when there is only a single source. This observation suggests that spatial information acts to modulate competition between sources, even in an anesthetized preparation. The fact that these effects arise under anesthesia is important because it shows that they are preattentive. Competition between spatially separated sources helps segregate neural responses, so that information about competing sound sources from different locations is concentrated in distinct subpopulations of cortical neurons. Specifically, most neurons preferentially encode information about contralateral sources; however, some neurons show more specific preferences.
Thus, even though location is not directly coded in cortical neurons, spatial information strongly modulates cortical responses. This idea fits with recent results showing that the effects of source location on neurons in cortex are enhanced when an awake animal is engaged in a task requiring a localization response [10]. The degree to which spatial information affects cortical responses changes with intention: depending on the importance of spatial information to the task being undertaken, spatial coding may be either enhanced or weakened. It is possible that inhibition driven by activity in prefrontal cortex (in mammals) or its analog (in avian species as studied here [29]) causes the sharpened tuning observed during spatial tasks [10]; if so, such connections may also be engaged during selective attention tasks to down-regulate responses of neurons preferentially encoding a masking stimulus that is to be ignored or to up-regulate responses of the distinct population of neurons preferentially encoding the target.
In an anesthetized preparation like that tested here, the enhancement of spatial effects due to the presence of a competing source cannot be coming from top-down modulation from executive centers of the brain. Instead, these effects must be the result of neural circuitry that is “hard wired.” It may be that weak spatial tuning, which is not strong enough to cause observable changes in neural responses with changes in the location of a single sound source played alone, causes large effects when there are multiple sources from different positions. The preattentive spatial competition we find provides a substrate to realize selective spatial auditory attention. Once responses to competing sounds are partially segregated through this kind of preattentive, spatially sensitive process, attentional signals, including inhibitory feedback from executive control areas, can enhance the spatial selectivity already present.
In humans, many spatial effects are explained by the fact that the head causes a significant acoustic shadow at many audible frequencies [26],[30]. When competing sound sources come from different azimuthal locations, the SNR at the ear closer to the target will be greater than if the sources were co-located. This kind of “better-ear” effect has nothing to do with neural processing but is a simple consequence of physics. For the human, such effects can be very significant for speech perception, because the head shadow can be 15 dB or more for frequencies important for speech. Thus, although not interesting from a neural processing perspective, these acoustic effects are important for perception.
Here, in the zebra finch, better-ear effects are small. The zebra finch head is diminutive; its width corresponds to only one quarter of the wavelength of the highest frequency present in our bandlimited stimuli (8 kHz). Given that appreciable acoustic interactions only arise when the wavelength of the sound is comparable to or smaller than the size of the physical object in the environment, the stimuli we presented did not contain frequencies high enough to cause large interaural level differences. This bears out in our measurements, which show an amplitude difference between the ears of approximately 1.5 dB when the stimulus is at ±90°. The better-ear effect is thus limited to 3 dB.
Performance in the target contralateral, masker ipsilateral configuration was 16.8% better than performance in the target ipsilateral, masker contralateral configuration, on average. In contrast, the performance benefit of lowering the masker noise level by 6 dB is only 8.8%. Moving the masker and target in space, then, has nearly double the effect on identification performance as a 6 dB increase in SNR. Given that the maximum effect of acoustic head shadow is only 3 dB, the better-ear acoustic effects cannot explain the spatial effects obtained. Moreover, although a better-ear effect may contribute to the processing of natural broadband signals that contain frequencies high enough to interact acoustically with the zebra finch head, it is unlikely to play a major role in the effects observed here, where we used low-pass filtered stimuli.
Interference from a masker on the response encoding a target can be broken down into two forms: spike additions (primarily in the gaps between syllables) and spike subtractions (primarily during syllables) [18]. Here, we quantified the effects of spatial configuration on spike additions and subtractions, and then evaluated modeled spike trains to determine the relative impacts of these effects on neural discrimination performance. In general, additions were more likely than subtractions when the target was ipsilateral to the recording site and masker was contralateral (see Figure 4C), while subtractions were the more prevalent form of interference in the reverse configuration. Because additions and subtractions were calculated by comparing the responses at each spatial configuration to responses to the corresponding target-only stimulus, they represent only the effect of the masker on the response, independent of the minor changes that occur due to absolute target location.
By modeling spike trains with only additions or only subtractions, we were able to gauge their effects on performance. Spike subtractions degraded performance at all configurations (in Figure 4C, blue bars are lower than white bars). In contrast, the spike trains with only additive interference coded target song identity nearly as well as the responses in quiet (red bars are nearly the same as white bars). Additions have a modest impact when subtractions are also present; additions-only performance was better than the fully masked responses in some configurations (compare blue and black bars). Subtractions, on the other hand, interfere with encoding of song identity more seriously and consistently across all spatial configurations.
Although this analysis does not reveal the mechanisms by which a masker interferes with coding of a target, it does give some insight into the complex interactions that take place when two competing sounds are present in an environment. For instance, one might expect, a priori, that the presence of an ongoing masker would cause activity to increase overall, so that the stereotypical target response in quiet is hidden amidst added spikes elicited by the masker. Yet, instead, the detrimental effects of the masker come about primarily from suppression of responses to key features in the target; moreover, the influence of the masker on the target response depends on spatial configuration. This pattern of spatial-configuration-dependent suppression of spikes suggests that competing sources, each preferentially encoded by a distinct neural subpopulation, mutually suppress each other, giving rise to enhanced spatial modulation of responses compared to when a single, unchallenged sound source is presented in isolation.
For a given site, song identity coding tended to vary with both the target and masker locations and generally was best when the target was contralateral from the recording electrode and the masker was ipsilateral to it. For a single site to show spatial release from masking, performance for that site should increase monotonically with increasing spatial separation between the sources. Thus, neither any single recording site nor mean performance averaged over all sites (shown in Figure 3B) exhibits spatial release from masking. Similar results have been seen in the midbrain, in inferior colliculus [28], where, as here, single units showed preferences for encoding responses to different sources, depending on the spatial configuration. However, the activity of thousands of forebrain neurons, not just a single unit, combines to govern perception and behavior. As shown in Figure 3, across the population of neurons in forebrain, there are typically neurons contralateral to the target source that encode target identity well. By looking at the mean performance of neurons at recording sites for which the target sound is more contralateral than the masker (or at the best neuron in that population), performance is predicted to improve with spatial separation (see Figure 3C). Thus, the ensemble of responses, even from an anesthetized bird, can explain behavioral spatial unmasking if one assumes a mechanism as simple as attending to neurons in the hemisphere that favors encoding of the target and ignoring those from the opposite hemisphere.
In behavioral experiments, performance improves with increasing separation between target and masker sounds both for speech and for non-speech sounds [26],[31],[32]. As noted above, better-ear acoustics contribute to spatial release from masking for many sounds important to human behavior, such as speech. Indeed, when a target sound is easily distinguished from a masker (such as when a communication signal is played in steady-state noise), better-ear acoustics can fully account for spatial release from masking in human studies. Interestingly, avian studies do not show the same pattern. The amount of spatial release from masking is essentially identical when behaving birds identify target birdsongs embedded in either a chorus of songs that sound qualitatively like the target songs or a steady-state masker (with the same long-term spectral content as the chorus, but has different short-term structure) [27]. This different pattern suggests that humans can segregate a target from a dissimilar masker even when the two sources are near each other in space, rendering spatial cues redundant [33]. In contrast, birds may be less sophisticated in segregating competing sources, relying more heavily on spatial attributes even when target and masker have distinct spectro-temporal content. Regardless, the current results demonstrate how spatial separation of target and masker can support spatial release from masking in those situations where it is observed behaviorally, no matter what species.
All experimental procedures involving animals were done in accordance with the protocol approved by the Boston University Institutional Animal Care and Use Committee. All subjects were male zebra finches (Taeniopygia guttata).
Prior to the day of recording, a preparatory surgery was performed. In this surgery, the location of field L was marked as the point 1.2 mm anterior and 1.5 mm lateral of the midsagittal sinus and a headpin was fixed to the skull. On the day of recording, the bird was first placed in a soft cloth restraining jacket in a quiet, dark room. Injections of urethane anesthetic (20%) were administered every half hour in decreasing amounts (starting with 35 µL) until the bird was unresponsive to its head being patted and its foot being squeezed. Once anesthetized, the bird was placed in a stereotactic frame with its head secured by the previously implanted pin. A craniotomy was performed in which an approximately 2 mm square of skull was removed centered about the spot previously marked as field L. Tungsten microelectrodes (FHC, Bowdoin, ME) ranging in impedance from 2 to 4 MΩ were advanced into the brain using a micron-precision stepper motor. Extracellular potentials were amplified at the headstage, bandpassed between 500 and 10,000 Hz, and recorded with a low-noise soundcard at a sampling rate of 44.1 kHz.
Stimuli were constructed from combinations of two different target zebra finch songs and masking noise (see Figure 1A and B for spectrograms), all filtered between 500 and 8,000 Hz. The songs were chosen to have similar durations (∼2 s); they were songs never before heard by the subjects. To generate the masking noise, several songs were concatenated, the discrete Fourier transform computed, the phase randomized uniformly between 0 and 2π (preserving symmetry), and the inverse Fourier transform computed. The result was noise with a magnitude spectrum identical to the average of the spectra of those songs, but with no temporal structure. Ten independent, random tokens of noise were created so that any residual temporal structure was averaged out across repeated presentations. Independent noise tokens were used on each trial instead of using a single, frozen token because individual noise tokens with the same statistics can have drastically different masking effects [34]. Additionally, the use of independent masker tokens better simulates what happens in natural settings, where, over time, a bird repeatedly hears highly stereotyped songs from its familiar colony mates, but hears them in a different background of masking sources each time.
Stimuli were presented using four single-driver loudspeakers in a sound-treated booth (IAC, Winchester, UK) at a sampling rate of 44.1 kHz. Target songs were normalized so that their root-mean-square amplitudes were 72 dB SPL (c-weighted). The loudspeakers were at four locations in the azimuthal plane: ipsilateral to the implanted hemisphere (−90°), in front of the bird (0°), contralateral to the implant (+90°), and at the angle halfway between the front and contralateral angles (+45°). The speaker locations were referenced relative to the recording electrode, with the side ipsilateral to the implant assigned the negative sign.
Each recording session consisted of 10 blocks. In each block, each of the two target songs was played in isolation from all four locations. Additionally, for each target song, 16 target-masker spatial configurations were tested, each at three SNRs. This resulted in 2×(4+4×4×3) = 104 stimuli per block in which targets were present. We also played the masker alone from each location in each block, resulting in a total of 108 stimuli per block. Each of the 10 blocks used a different, independent token of masking noise semi; the order of the stimuli within each block was randomized. Overall, there were 1,080 two-second stimuli presented with 1.5 s between the end of one trial and the beginning of the next, resulting in a recording session that lasted 63 min for each neural site.
Extraction of action potentials (spikes) was performed off-line. First, neural traces were thresholded. The recording to 1 ms on either side of each local maximum was windowed out and considered a potential spike. These waveforms were sorted into user-defined template spike waveforms using a correlation-like coefficient:where xS is a spike waveform and xT a template waveform, and the sums are taken over time. Spikes were sorted into classes based on the template that yielded the highest r or were thrown out if they were not above a minimum r to any of the templates. This sorting was verified using principal components analysis clustering. Using this method, single units as well as multiunit clusters (which could not be separated into single units) were extracted. Of the sites that met the minimum performance criterion (see below), 17 were single units and 16 were multiunit clusters. Multiunit activity should produce weaker spatial effects than well-isolated single units. By including both isolated and multiunit recordings in our analysis, our approach is likely to underestimate (if anything) the effects of spatial configuration on neurons in the forebrain. Recordings were made in both hemispheres, but a relative coordinate system was used so that negative azimuths always correspond to the hemisphere ipsilateral the recording site and positive azimuths to the contralateral side.
Discrimination performance was calculated using a nearest-neighbor template-matching scheme and a spike distance metric. Methods used were similar to those used in past studies [17]–[19],[22],[24],[35],[36]. To compute pairwise distances between recorded spike trains, each spike train was convolved with a decaying exponential kernel whose time constant determined the effective integration time of the spike comparisons; then the sum of the squared difference was calculated. These distances were used to compare a test spike train against two templates, one from each target song. Each spike train was classified as being elicited by the song whose template was closest to the measured spike train. This process was repeated many times for all spatial configurations for kernel time constants of 1, 4, 16, 63, 251, and 1,000 ms. All but one of the recording sites had an optimal time constant of 16 or 63 ms, in the same range as time constants found in similar past studies (the outlier had an optimal time constant of 251 ms) [17],[22]. In this way, a percent correct score was calculated as a function of time constant, representing how well the spike trains from each spatial configuration matched the target spike trains from the template configuration. The time constant that yielded the best clean target discrimination for each site was used.
Spatial sensitivity was computed as the difference between the maximum and minimum discrimination performance for a given stimulus type. For clean songs, these extrema were determined across the four target locations. For masked stimuli, they were determined across all 16 location configurations, at the SNR that had the highest variance. So that sites with poor performance did not appear spuriously insensitive to spatial configuration, only sites that had an unmasked discrimination performance of 90% or more were included in the analysis.
To analyze the effects of spike additions and deletions separately, modeled spike trains were generated. Spike trains were binned into time bins of 2.5 ms. To generate a new spike train, the mean and standard deviation of the number of spikes in each time bin were computed across the 10 responses to each stimulus. Then, the number of spikes in each bin was chosen randomly from a Gaussian distribution with the same mean and standard deviation, with negative spike counts fixed to zero. Time bins in which the masked rate was higher than the clean rate were labeled “additions.” Thus, to simulate a spike train that had only additions, the higher of the masked and clean rates (and the corresponding SD) was chosen at each time bin. Similarly, to simulate a subtractions-only spike train, the minimum of the masked and clean rates was chosen for each bin. The discrimination performance of these simulated spike trains was then calculated in the same manner as real recordings, described above. Refractory period violations (interspike intervals of less than 1 ms) had a negligible effect on analysis.
The significance of the difference between the spatial sensitivity for clean and masked responses was computed using a paired Student's t test. All correlation r values were computed using Pearson's product-moment coefficient, with p values calculated using a Student's t distribution. The significance of spatial separation and SNR for the data shown in Figure 3C were computed using a two-way repeated measures ANOVA. After using a one-way ANOVA to confirm a significant effect of interference type (p<.001), Tukey's HSD test was used to compute post hoc comparisons between all performance values at each of the three spatial configurations in Figure 4G. All statistics were done using Matlab's built-in functions. A p value of .05 or less was considered significant.
|
10.1371/journal.ppat.1004369 | Symbionts Commonly Provide Broad Spectrum Resistance to Viruses in Insects: A Comparative Analysis of Wolbachia Strains | In the last decade, bacterial symbionts have been shown to play an important role in protecting hosts against pathogens. Wolbachia, a widespread symbiont in arthropods, can protect Drosophila and mosquito species against viral infections. We have investigated antiviral protection in 19 Wolbachia strains originating from 16 Drosophila species after transfer into the same genotype of Drosophila simulans. We found that approximately half of the strains protected against two RNA viruses. Given that 40% of terrestrial arthropod species are estimated to harbour Wolbachia, as many as a fifth of all arthropods species may benefit from Wolbachia-mediated protection. The level of protection against two distantly related RNA viruses – DCV and FHV – was strongly genetically correlated, which suggests that there is a single mechanism of protection with broad specificity. Furthermore, Wolbachia is making flies resistant to viruses, as increases in survival can be largely explained by reductions in viral titer. Variation in the level of antiviral protection provided by different Wolbachia strains is strongly genetically correlated to the density of the bacteria strains in host tissues. We found no support for two previously proposed mechanisms of Wolbachia-mediated protection — activation of the immune system and upregulation of the methyltransferase Dnmt2. The large variation in Wolbachia's antiviral properties highlights the need to carefully select Wolbachia strains introduced into mosquito populations to prevent the transmission of arboviruses.
| In recent years it has been discovered that many organisms are infected with bacterial symbionts that protect them against pathogens. Wolbachia is a bacterial symbiont that is found in many species of insects, and several strains are known to protect the insects against viral infection. We took 19 strains of Wolbachia from different species of Drosophila fruit flies, transferred them into Drosophila simulans, and then infected these flies with two different viruses. We found that about half of the strains slowed the death of flies after viral infection. Given that 40% of terrestrial arthropods may be infected with Wolbachia, this suggests that many species may benefit from this protection. These increases in survival were tightly linked to reductions in the levels of the virus in the insect, suggesting that Wolbachia is reducing the viruses' ability to replicate. Despite the two viruses we used being very different, the level of protection that a Wolbachia strain provided against the two viruses tended to be very similar, suggesting that a single general mechanism underlies the antiviral effects. The extent to which a Wolbachia strain provides protection against viral infection depends largely on the bacterial density— the more Wolbachia, the greater the protection.
| Heritable symbionts are major players in arthropod evolution owing to their high incidence and the diversity of effects they have on their host's phenotype [1]–[3]. Primary (obligate) symbionts are mutualists that play some essential role — typically synthesizing nutrients missing from the insect's diet — and they often form stable associations with their hosts that can last for many millions of years [4]–[6]. Secondary (facultative) symbionts have more diverse effects, which range from parasitism to mutualism [3], [7]. The parasites mostly manipulate their host's reproduction to enhance their transmission to the next generation, for example by distorting the sex ratio towards females (the sex that transmits the bacteria). The mutualists can supply nutrients, or protect against environmental stresses [8] or natural enemies [9]–[11]. Furthermore, symbionts can combine several strategies at once, with some ‘Jekyll and Hyde’ strains simultaneously exhibiting mutualistic and parasitic phenotypes [12].
As secondary symbionts occasionally jump between different host species [13]–[16], they can result in rapid evolutionary change in their hosts. This process may be quite different to selection acting on the host genome, as when a host acquires a novel symbiont it can instantly acquire a complex adaptation encoded by many genes. Striking examples of rapid evolution resulting from the spread of symbionts include a Rickettsia bacterium infecting whiteflies which rapidly spread through US populations by causing sex ratio distortion as well as increased fecundity and survival [17], and a Spiroplasma bacterium that spread through populations of Drosophila neotestacea, protecting the hosts against a parasitic nematode [10].
Wolbachia is a maternally-transmitted alphaproteobacterium that is estimated to infect around 40% of terrestrial arthropods [18] and can act as both a parasite and a mutualist. Until recently it was viewed primarily as a parasite that manipulates host reproduction, most commonly by inducing cytoplasmic incompatibility (CI) [19]–[23]. CI allows Wolbachia to invade populations by causing embryonic mortality when uninfected females mate with infected males, thus conferring a selective advantage to infected females [24]. Recently it was discovered that Wolbachia can also protect Drosophila melanogaster against several RNA viruses [25], [26], and subsequently similar antiviral effects have been reported in other Drosophila species [27], [28], as well as in mosquitoes [29]–[32]. In most cases, Wolbachia has been shown to be associated with a decrease in viral titer [26], [27], [29]. However, Wolbachia increased the survival of the flies but had no effect on viral titer in D. melanogaster infected with Flock House virus (FHV) [26] as well as in one case in D. simulans infected with Drosophila C virus (DCV) [27], suggesting that Wolbachia might also allow its host to tolerate viral infections without affecting the pathogen load. Wolbachia has also been associated with protection against filarial nematodes, Plasmodium parasites and pathogenic bacteria in mosquitoes [29], [33]–[36]. However, it is not known whether the mechanisms of protection acting against these parasites are the same that are involved in protection against viruses.
Antiviral protection by Wolbachia could potentially be used to control vector-borne diseases such as dengue fever [37], [38]. When artificially introduced into Aedes aegypti, the main vector of dengue virus, Wolbachia was shown to limit the replication of dengue virus as well as chikungunya, yellow fever and West Nile viruses [29], [39], [40]. Furthermore, when Wolbachia infected mosquitoes were released into the wild, the bacterium spread through the mosquito populations due to the induction of CI [41], [42].
In both Drosophila and mosquitoes, different Wolbachia strains are associated with different levels of antiviral protection [27], [32], [39], [43], [44], even among very closely related strains [45]. The causes of this variation are not entirely clear, as relatively few Wolbachia strains have been characterized for their level of protection, not all studies control the host genetic background, and none have controlled for the confounding effects of the bacterial phylogeny. Nonetheless, several studies found that the Wolbachia strains with the highest density within the host provide the strongest protection against viruses, and tissue tropism may also play a role [27], [30], [44]–[47]. Overall, little is known about how commonly Wolbachia protects insects against viral infection, how this trait is distributed across the Wolbachia phylogeny, and therefore to what extent it has contributed to the evolutionary success of Wolbachia.
The mechanisms by which Wolbachia protects hosts against viruses remain to be elucidated. The protection could be caused by direct interactions between Wolbachia and viruses, competition for shared resources, or indirectly through the regulation of host gene expression [26]. In particular, it was speculated that Wolbachia infection may up-regulate the host immune system. While this was shown to occur after transinfection of Wolbachia from Drosophila into Ae. aegypti and Anopheles gambiae [33], [48]–[50], such an effect was not observed in D. melanogaster naturally infected with Wolbachia [50]–[55] or transinfected with a non-native strain [44]. Similarly, the small interfering RNA (siRNA) pathway, which provides broad-spectrum antiviral defense in insects, is not required for Wolbachia to confer antiviral protection in flies [56]. Recent results suggest that, in Ae. aegypti, Wolbachia has an indirect effect on viral replication through the manipulation of host microRNAs [57], [58]. In this species Wolbachia suppresses the expression of AaDnmt2, a methyltransferase gene, by up-regulating the microRNA aae-miR-2940 [57]. Overexpression of AaDnmt2 decreases Wolbachia density and increases the titer of Dengue virus, suggesting a causal link between Wolbachia and viral replication. However, Dnmt2, the homolog of AaDnmt2 found in D. melanogaster, was shown to have an antiviral effect against Drosophila C virus [59], contradicting the effect observed in case of dengue infection.
To overcome the lack of experimental tools available for Wolbachia — the bacterium cannot be cultured outside of insect cells and cannot be genetically manipulated or cloned — we have taken a comparative approach, looking for genetic correlations between levels of antiviral protection and potential causes such as changes in gene expression. To allow us to do this, we compared the level of protection of 19 Wolbachia strains from a diverse range of Drosophila species that we transferred into a common D. simulans genetic background. We used Drosophila C virus and Flock House virus, which are both RNA viruses with positive-sense single-stranded genomes. Drosophila C virus belongs to the family Dicitroviridae [60] and is naturally found in D. melanogaster and D. simulans [61]–[63]. FHV belongs to the family Nodaviridae and was initially isolated from a beetle [64].
Using this comparative approach, we show that Wolbachia strains vary considerably in the extent to which they increase the survival of flies after viral infection. There is little specificity, with strong genetic correlations between protection against FHV and DCV, despite these viruses being distantly related. The increases in survival can largely be explained by Wolbachia reducing viral titer. The variation in antiviral protection is largely explained by differences in the density of Wolbachia in host tissues. However, there is no evidence that either activation of the humoral immune response or up-regulation of the methyltransferase gene Dnmt2 play any role in antiviral protection.
We assembled a panel of 19 Wolbachia strains that naturally infect 16 different species of Drosophila (Table 1). We reconstructed the phylogeny of these strains using sequences from eight multilocus sequence typing (MLST) genes and a Bayesian method that accounts for recombination between strains [65]. The phylogeny reveals that 18 of the strains clustered in what is commonly regarded as the supergroup A [66], with wMa being the only strain from the supergroup B (Figure 1A). Many of the strains are very closely related.
To allow us to compare the level of antiviral protection that the different Wolbachia strains provide to their hosts, we transferred them into the same inbred D. simulans genetic background. Eleven of the strains were transferred as part of this study, and the remaining eight have been reported before (Table 1).
Flies from the 19 D. simulans lines carrying the different Wolbachia strains, together with a Wolbachia-free control, were stabbed with a needle that had been dipped in DCV, FHV or Ringer's solution (2528, 2527 and 2492 flies were stabbed for each treatment respectively). We then followed their survival for 25 days (Figure 2A–C). In the mock-infected flies (Ringer), there was significant heterogeneity among the 20 fly lines (Cox's mixed-effect model, χ2 = 47, d.f. = 19, P = 0.0004; Figure S1), possibly reflecting either intrinsic effects of Wolbachia on survival or some other difference between the lines, such as remaining differences in the host genetic background. The overall survival of the mock-infected flies was low, likely due to this being a weak inbred stock (Figure 2C).
There was a substantial variation among Wolbachia strains in the degree to which they protect their hosts against viral infection (Figure 2A & B). Twelve of the Wolbachia-infected lines showed significantly reduced mortality relative to the Wolbachia-free flies within either the DCV or FHV treatments (Figure 1B & 1C). To account for the slight variation in the survival of the mock-infected controls, we tested whether survival of the 20 fly lines was affected by a statistical interaction between the Wolbachia strain and infection (whether the flies were infected with a virus or mock-infected). There was a highly significant interaction for both DCV and FHV (Cox's mixed-effect models; DCV: χ2 = 127.4, d.f. = 19, P<10−15; FHV: χ2 = 107.6, d.f. = 19, P<10−14). Using this more conservative approach of testing for an interaction of Wolbachia and infection treatment, nine of the 19 Wolbachia strains provided a significant level of antiviral protection, with six protecting against both viruses, one protecting against just FHV and two protecting against just DCV (Table S1). This protective phenotype is widespread across the Wolbachia phylogeny and is not restricted to particular clades (Figure 1A–C).
To examine the extent to which the effects of Wolbachia are specific to different viruses, we estimated the genetic correlation between protection to DCV and FHV (the proportion of the genetic variance shared by the two traits). The genetic correlation between protection to DCV and FHV was high (Model 1: rg = 0.81, 95% CI: 0.55,0.97; Figure 3A), indicating that most of the genetic variance in antiviral protection affects both DCV and FHV. There was no evidence of a genetic correlation between the survival of virus-infected and mock-infected flies (Model 1, DCV-Ringer: rg = 0.31, 95% CI = −0.37,0.87; Model 1, FHV-Ringer: rg = 0.61, 95% CI = −0.14,0.99).
This genetic correlation between viruses could arise either because there is a causal relationship between DCV and FHV protection, or as a consequence of common ancestry (phylogenetic non-independence). We used a phylogenetic mixed model to partition the variance in the two traits into a component that can be explained by correlations across the Wolbachia phylogeny and a strain specific component that is independent of phylogeny. If there is a causal link between the traits, then the strength of their association will be the same for the phylogenetic and strain components. There was no significant difference in the strength of the genetic correlation for the phylogenetic and strain-specific components, consistent with a causal link between the traits. As we have limited power to separate these effects in a single model, we also fitted a model with just the phylogenetic effect. This model again produced a very similar correlation between viruses (rg = 0.95, 95% CI: 0.84,0.99), and the convergence of parameter estimates was improved.
To investigate the effect of Wolbachia on viral titers, flies were stabbed with DCV or FHV, and relative viral RNA levels measured at two days post infection (dpi) by reverse transcription-quantitative PCR (RT-qPCR). For both DCV and FHV, viral titer was affected by the Wolbachia-infection status of flies (ANOVA on ln(viral titer); DCV: F19,165 = 23.4, P<10−16; FHV: F19,166 = 12.1, P<10−16; Figure 4A & 4B). Wolbachia strains tended to have similar effects on the two viruses, with a strong positive genetic correlation across the Wolbachia strains between titers of DCV and FHV (Model 2: rg = 0.75, 95% CI = 0.48,0.94; Figure 3B).
Wolbachia could increase the survival of Drosophila after infection either by reducing viral titers (increasing resistance), or by allowing flies to better cope with infection damage (increasing tolerance). To test whether Wolbachia provides resistance to infection, we compared the survival of flies and viral titers across the 19 Wolbachia strains. Relative to the Wolbachia-free control, seven of the strains were individually associated with significantly reduced DCV titers, and five with reduced FHV titers (Figure 4A & 4B). Interestingly, for the strains wHa in DCV- and wSan in FHV-infected flies, there was a significant increase in viral RNA levels (Figure 4A & 4B) and this result was replicable by repeating the experiment on these lines (J. Martinez, personal observation). Overall, we found that the titer of both DCV and FHV was negatively genetically correlated to the survival of DCV and FHV infected flies respectively (Model 3, DCV: rg = −0.77, 95% CI = −0.95,−0.49; Model 4, FHV: rg = −0.88, 95% CI = −1.00,−0.64; Figure 3C & 3D). Therefore, resistance could be the primary explanation for Wolbachia-induced protection.
To understand how Wolbachia affects the dynamics of infection, we followed DCV and FHV titers for five days in Wolbachia-free flies and flies infected with the protective strain wAu that conferred strong protection against both viruses, the non-protective strain wSh and wAna, which protected against DCV but not FHV. In all treatments, including in the presence of the strongly protective strain wAu, DCV and FHV were able to replicate within the flies (Figure S2). For example, at 2 dpi, DCV titres had increased ∼18.000-fold in wAu-infected flies compared to ∼290.000-fold in Wolbachia-free flies. At 2 dpi, the timepoint chosen in the previous experiment, viruses were still in their growth phase, with a plateau of DCV titres being reached at around 3 dpi (Figure S2). Resistance conferred by wAu seemed to occur earlier against FHV (1 dpi at start growth phase) than against DCV (2 dpi at end growth phase) and viral titers were reduced from those points on, including at the end of the growth phase.
There was significant variation in the relative density of the different Wolbachia strains (ANOVA on ln(relative density): F18,150 = 115.6, P<10−16; Figure 1D) in 3–6 day old virus-free flies (the same age as the flies that were infected to estimate survival and viral titer). The highest density strain, wMelCS, had 114 fold higher density than the lowest density strain, wBor.
Flies infected with high density Wolbachia strains tended to live longer after viral infection (Figure 3E & F) and had lower viral titers. This is reflected in strong genetic correlations between bacterial density and survival after both DCV and FHV infection (Model 1; DCV: rg = 0.76, 95% CI: 0.49,0.93; FHV: rg = 0.57, 95% CI: 0.20,0.86). This appears to be a specific effect of Wolbachia on survival after viral infection as there is no support for a correlation between Wolbachia density and the survival of the mock-infected flies in this assay (Model 1; Ringer: rg = 0.18, 95% CI: −0.41,0.77). Similar to the survival analysis, bacterial density is negatively genetically correlated with DCV titers (Model 2: rg = −0.53, 95% CI: −0.82,−0.13). However, there is no support for a correlation between FHV titer and Wolbachia density (Model 2: rg = −0.29, 95% CI: −0.65,0.18).
To examine whether there is likely to be a causal relationship between Wolbachia density and survival after viral infection, we partitioned the variance in these traits into components that are dependent and independent of the Wolbachia phylogeny. The regression coefficient of survival against bacterial density was not significantly different for the phylogenetic and strain components for the two viruses. We are therefore unable to find evidence to suggest that this correlation is an artefact of phylogenetic relatedness, although we would caution that this analysis has very limited statistical power (non-significant phylogenetic component). Given this limited power to partition the variance across the two components, we also fitted a model with phylogenetic component only, and again the correlations between protection and Wolbachia density were similar to the model without the phylogeny (DCV: rg = 0.71, 95% CI: 0.32,0.92; FHV: rg = 0.63, 95% CI: 0.24,0.90).
We finally investigated if antiviral protection could be explained by an effect of Wolbachia on host gene expression. We first tested the hypothesis that Wolbachia prime the immune system of flies, by measuring the expression of Drosomycin and Diptericin as reporters of Toll and IMD pathway activation respectively. We did not detect a significant genetic correlation between expression of either immune genes and DCV protection (Model 5; Drosomycin: rg = 0.26, 95% CI: −0.37,0.83; Diptericin: rg = −0.05, 95% CI: −0.79,0.66; Figure S3A & S3C) or FHV protection (Model 6; Drosomycin: rg = 0.25, 95% CI: −0.29,0.72; Diptericin: rg = 0.28, 95% CI: −0.37,0.95; Figure S3B & S3D). Finally, the expression of a putative candidate for protection, the methyltransferase gene Dnmt2, was not significantly affected by the Wolbachia-infection status (ANOVA on ln(expression level); DCV: F16,122 = 1.56, 0.09; FHV: F16,122 = 0.85, P = 0.63; Figure S3E & S3F) and did not show any correlation with level of protection in DCV-infected (Model 5; Dnmt2: rg = 0.12, 95% CI: −0.59,0.90) or FHV-infected flies (Model 6; Dnmt2: rg = −0.35, 95% CI: −0.98,0.49).
Protective symbionts can be an important component of an organism's defenses against infection, in some cases even being the primary mode of defense [67]. Despite their importance being increasingly recognized, studying these symbionts remains challenging. Many cannot survive outside of host cells, be genetically manipulated or cloned. Our approach to circumvent these problems has been to assemble a large panel of different Wolbachia strains in a common host genetic background. This allows us to detect genetic correlations between traits, and infer whether these traits are causally linked.
Our results suggest that symbionts may play a role in protecting a substantial proportion of insect species against viral infection. Wolbachia is probably the most widespread symbiont in arthropods [68] and its wide distribution is partly attributable to its ability to manipulate host reproduction as well as its tendency to be horizontally-acquired between different host species over evolutionary time scales [2]. Recently, it has been shown to confer protection against natural enemies, in particular against RNA viruses [25], [26], [29]. By assessing the level of protection among several Wolbachia strains, we showed that, far from being an exception, Wolbachia-mediated protection is a common phenomenon, which could potentially have contributed to its evolutionary success.
Among the tested strains, about half were able to confer some level of protection in D. simulans. Assuming that Wolbachia is found in 40% of arthropod species [18], our results suggest that 20% of arthropods may benefit from such a protection. This extrapolation relies on strains retaining their ability to protect their original host, but the host species could also influence the expression of the protective phenotype. It was previously shown that protective strains native to D. melanogaster also protect mosquito hosts after artificial transfer. In contrast, the strain wInn, which did not confer protection in this study, was previously found to protect against FHV in its original host D. innubila [28]. Host genotype effects on the Wolbachia density have previously been found [69], [70]. Given the correlation between protection and density, the expression of protection is likely to be under the control of both the Wolbachia strain and the host genotype.
Wolbachia has previously been shown to protect insects against a remarkably taxonomically diverse array of RNA viruses [25], [26], [29], [31], [40]; and this could either reflect a broad-spectrum antiviral mechanism or Wolbachia may have independently evolved different ways of targeting different viruses. Previously it has been observed that strains that protect strongly against one virus tend to protect against other viruses, suggesting the former explanation is true [27], [44]–[45]. We found that an estimated 81% (rg = 0.81) of the genetic variation among strains in DCV and FHV protection is common to the two viruses. Furthermore, this pattern appears to be independent of the bacterial phylogeny, indicating that the same genes underlie the level of protection to the two viruses tested. This supports the hypothesis that Wolbachia has a single broad-spectrum mechanism of antiviral protection.
The increased survival of Wolbachia-infected flies after viral infection could result from the symbiont increasing either resistance or tolerance to infection [71], [72]. Resistance occurs where increases in survival are caused by reductions in viral titers, while tolerance describes the situation where hosts are better able to survive a given viral load [71]. Both of these effects have been ascribed to the antiviral properties of Wolbachia in the past [26], [27], [29], [31], [40], [44], [45]. Our analysis allows us to test the effect of resistance by estimating the proportion of the variation in survival that can be explained by differences in viral titer. The genetic correlation between titer and survival was very high for both viruses, so in this instance it seems likely that the between-strain variation in survival is mainly due to resistance to virus infection. Our data cannot exclude a role for tolerance, as Wolbachia may be altering the amount of harm that a given viral titre causes. However, were this to be the case, then it is likely to be a common underlying link, such as both traits relying on Wolbachia density or the same mechanism.
In some studies it was shown that Wolbachia infection can lead to higher viral titers or virus-induced mortality [73], [74]. Interestingly, in our experiment two Wolbachia strains were associated with an increase in viral titer, although not with increased mortality. This is a tantalizing result, which would suggest care should be taken when introducing Wolbachia into disease vector populations. However, we would caution that this result needs to be investigated in more detail – we measured many traits across many strains, so rare outliers could be an artefact of confounding factors like remaining differences in the genetic background of the strains.
The density of Wolbachia plays a key role in determining the level of antiviral protection it provides to its host. This has been previously demonstrated experimentally by manipulating Wolbachia density using antibiotics, and by comparisons of high and low density strains [27], [44], [45], [46]. Our results strengthen this conclusion, as we show that the relationship of density and survival is strong and highly significant across a large panel of strains. Furthermore, this association does not appear to be a consequence of phylogenetic relatedness, suggesting that higher Wolbachia density is causing higher levels of resistance to viruses.
Do any factors other than Wolbachia density cause between-strain variation in the level of resistance to viruses? Our analysis provides only weak support for other factors being important, as while the genetic correlation (the proportion of genetic variance shared) between Wolbachia density and survival ranges from 0.57–0.76, the upper confidence intervals for all estimates are greater than 0.86. Therefore, while our data suggests other factors are important, the evidence is not strong. Any of these Wolbachia strains may have the intrinsic ability to provide resistance to viruses – they simply need to be present at a sufficiently high density. If true, it is tempting to speculate what this might imply about the underlying mechanism of resistance. It seems more compatible with a mechanism whereby the presence of Wolbachia per se makes cells or the host less hospitable to viruses, such as through competition for resources [75] or remodeling of the cellular environment. In contrast, if Wolbachia was expressing specific antiviral factors, then these might be easily gained or lost through evolution, breaking the genetic correlation of resistance and Wolbachia density. It is likely that various mechanisms can lead to variation in bacterial density and thus affect within-host density. For example, it was found recently that a ∼21 kb region encoding eight genes is amplified three to seven times in different wMelPop isolates relative to wMelCS [45], [76], and is associated with much higher density and stronger protection against viruses [45]. However, copy number variation of this region does not explain differences in density or protection between wMelCS and wMel-like strains [45]. It is therefore tempting to speculate that genomic differences between Wolbachia strains that confer differential protection to viruses will only reveal different ways of varying the bacterial density rather than the actual antiviral mechanisms.
One way of rendering the host less hospitable for viruses is through the regulation of host genes. It was argued in the past that Wolbachia infection may lead to the activation of immune pathways that in turn could limit the multiplication of other parasites. Previous studies in mosquitoes showed that even if Wolbachia can prime the host immune system and increase antiviral resistance, such an effect is absent in D. melanogaster [44], [50], [51], [55], and flies deficient in both the Toll and IMD pathways still display Wolbachia-mediated resistance [54]. In agreement with previous studies, our results support the conclusion that Toll and IMD pathways are not required for antiviral protection since both Drosomycin and Diptericin expression level (reporters of Toll and IMD pathways respectively) were uncorrelated to the survival of virus-infected flies. However, other immune pathways and restriction factors could still be involved.
In the mosquito Ae. aegypti, the methyltransferase AaDnmt2, whose homolog in Drosophila methylates transfer RNAs and other nucleic acids, has been proposed as a potential candidate to explain the antiviral effect of Wolbachia [57]. Wolbachia was shown to decrease the expression of AaDnmt2 through the induction of the expression of aae-miR-2940 microRNA. Conversely, the overexpression of AaDnmt2 led to a decrease in Wolbachia density and an increase in the titer of dengue virus. However, it was recently shown that the Drosophila homolog Dnmt2 has an antiviral effect against DCV and Nora virus, the opposite to the pattern seen in mosquito cells infected with dengue virus [59]. We found that Wolbachia has no consistent effect on Dnmt2 expression in D. simulans, and variation in Dnmt2 expression does not explain any of the variation in survival after infection. This suggests changes to Dnmt2 expression are not a general explanation of the antiviral effects of Wolbachia. It is possible that a different mechanism of resistance applies to mosquitoes and dengue virus. However, we would argue that the critical experiment to reach this conclusion would be to show that the antiviral effects of Wolbachia on dengue virus require AaDnmt2 or aae-miR-2940, and this experiment has yet to be performed.
Together with previous studies, our results show that antiviral protection is very common among Wolbachia strains. As such, it has to be taken into account if we are to draw a complete picture of Wolbachia ecology and evolution. For example, protection may favor the rapid sweeps of Wolbachia observed in natural populations [77], [78] and explain why strains such as wMel and wAu, that induce weak or no CI, can be maintained in natural populations [79], [80]. Owing to the high incidence of Wolbachia and the broad spectrum of viruses affected by the protection, it is likely that Wolbachia-mediated protection has substantially contributed to the evolution of arthropods. By protecting against infection, symbiont-based immunity may in turn influence the evolution of the host immune system. Although the mechanisms remain to be elucidated, protection is tightly linked to the bacterial density. Therefore, variation in the selective pressure exerted by viruses could partly explain why Wolbachia strains vary so much in density, and why some are found in somatic tissues whereas other are restricted to the germ cells [81]. From an applied perspective, our study extends the panel of Wolbachia strains that could be introduced into mosquito populations to limit the spread of arboviruses. However, for successful introduction, the choice of a strain should not only be based on the level of protection but also consider costs on host fitness and strength of CI that will affect the invasive potential of Wolbachia.
The origin of the 19 Wolbachia strains used in this study and their original host line are listed in Error! Reference source not found. To control for host genetic effects, all Wolbachia strains were transferred into the D. simulans STCP genetic background. This line was previously obtained through six generations of brother-sister crossing [23], [82], [83]. Eight of the strains were transferred into the STCP background in previous studies (Table 1; [23], [82], [83]). Of these, the three strains naturally-infecting D. simulans were generated by six generations of backcrossing Wolbachia-infected females to STCP males, and the remaining five were transferred by microinjection (Table 1). We microinjected eleven more strains into the STCP line (Table 1). Microinjections were performed as previously described using a microcapillary needle to transfer cytoplasm of infected embryos into uninfected STCP embryos [82]. All microinjected lines were maintained in the lab for at least 10 generations before the beginning of the experiments.
Two generations before the beginning of the experiments, the Wolbachia infection status of the STCP lines was checked by PCR using the diagnostic primers wsp81F and wsp691R [84], and the PCR products were sequenced. For strains microinjected in this study, vertical transmission was also assessed with PCR by testing 48 offspring per strain originating from Wolbachia-infected mothers (data not shown). Three fly stocks, transinfected with the strains wBai, wBic and wBor, showed imperfect vertical transmission (54%, 91% and 62% respectively). For those three strains, the presence of Wolbachia was checked by PCR one generation before each experiment and only offspring from infected mothers were used in the experiments. Additionally, in qPCR assays (see below), flies of those three strains that were used in the experiments were first isolated individually, their Wolbachia infection status was confirmed and only Wolbachia-infected individuals were kept and pooled in groups of 4–6 flies.
For all the experiments, flies were maintained on a cornmeal diet (agar: 1%, dextrose: 8.75%, maize: 8.75%, yeast: 2%, nipagin: 3%) at a constant temperature of 25°C with a 12-hour light/dark cycle at 70% relative humidity.
The phylogeny of the 19 Wolbachia strains was inferred from the partial sequences of the eight genes 16S rRNA, aspC, atpD, ftsZ, sucB, groEL, coxA and fbpA previously used in Multilocus Sequence Typing studies [66], [85]. The gene sequences were either obtained from GenBank, or were sequenced using the protocol described in Paraskevopoulos et al. (2006). Accession numbers and the origins of the sequences are described in Table S2. Each gene was individually aligned using Mauve v2.3.1 [86] and the phylogeny was inferred using ClonalFrame v1.2 to take into account recombination between strains [65]. To check for convergence, 9 independent runs were done with 100,000 MCMC iterations after 100,000 burn-in iterations with parameter recording every 100 iterations. For the first 8 runs, a uniformly chosen coalescent tree was used as the initial tree, and for the 9th run, a UPGMA tree was used. The UPGMA starting tree was compared to the eight other trees and showed a good convergence with seven of them based on the tree comparison tool implemented in ClonalFrame v1.2 [65]. Parameter estimates for the UPGMA starting tree also showed a good convergence based on the Gelman and Rubin test [87] in ClonalFrame v1.2. A consensus tree with branch support values was built from the posterior sample of the UPGMA starting tree at 50% majority-rule in MEGA 5.2 [88] and visualized in R using the ape package [89]. The consensus tree was visually compared with a tree inferred from the concatenated sequence of the eight genes using PhyML v3.1 with 500 bootstrap replicates [90] to assess the effect of recombination on the phylogenetic signal. All clades inferred from ClonalFrame were retrieved in the maximum likelihood tree. Therefore, the ClonalFrame tree with the UPGMA starting tree was used in further analyses.
Viruses were produced and titrated as in [26], with minor changes. DCV was produced and titrated in Schneider's Line 2 cells (SL-2), while FHV was titrated in Schneider Drosophila line 2 cells (DL2). For each infection assay, one viral aliquot was defrosted on the day of infection and diluted in Ringer's solution [91] to reach a viral concentration of 5×108 TCID50/mL for DCV and 3.38×108 TCID50/mL for FHV.
For each fly line, 3–6 day-old female flies were collected. After being anaesthetized with CO2, flies were either infected with DCV, FHV, or mock-infected with Ringer's solution [91]. The inoculum was administered by stabbing flies into the left pleural suture on the thorax with a 0.15 mm diameter anodized steel needle (Austerlitz Insect Pins) bent ∼0.25 mm from the end (∼half of the dorsal width of the thorax), dipped into viral or Ringer's solution as in [92]. Twenty stabbed flies were placed in a vial of fly cornmeal medium and dead flies were recorded every day for 25 days after infection. Flies were transferred into fresh vials of food every 3 days.
The survival assay was replicated on six consecutive days. On each day two vials of flies from each Wolbachia strain were assigned to two of the three treatments. The same was done for the Wolbachia-free flies, except that the number of replicates was doubled to increase statistical power. The stabbing order of the fly lines as well as the sequence of treatments were randomized each day. Mortality that occurred on the day following infection was attributed to stabbing injuries and was discarded from the analyses.
The Wolbachia density, DCV and FHV titers as well as the expression of the three host genes Drosomycin, Diptericin and Dnmt2 were measured by qPCR on a BioRad iQ5 thermocycler using primers, probes and cycle conditions listed in Table S3. Wolbachia density was measured on pools of 10 virus-free 3–6 day-old female flies (n = 10 pools) from which DNA was extracted using the Gentra Puregene kit (Qiagen). For viral titers and host gene expression, 3–6 day-old flies were first infected with DCV or FHV, as described above, and, 2 days after infection, 10 flies were pooled (n = 10 pools per virus), homogenized in TRIzol Reagent (Ambion) and frozen at −80°C. Total RNA was extracted using the Direct-zol-96 RNA kit (Zymo Research) by following the manufacturer's instructions, including a 15 min DNase I digestion step.
For host gene expression, total RNA was reverse-transcribed using the GoScript Reverse Transcription System (Promega) with random primers. Host gene expression and the Wolbachia density were measured relative to the endogenous control gene actin 5C (Table S3) using the SensiFAST SYBR & Fluorescein kit (Bioline).
The copy-number of viral genomic RNA was measured relative to the control gene EF1α100E (Table S3) in a one-step RT-qPCR reaction using the QuantiTect Virus kit (Qiagen). For each virus, both viral and fly cDNAs were amplified in a duplex reaction using virus and fly primers in association with dual-labeled (hydrolysis) fluorescent probes (Sigma) (Table S3). For each sample, two RT-qPCR reactions were carried out and the mean of these two technical replicates was used as the relative viral titer in the statistical analysis.
The efficiency of the PCR amplication was checked using a dilution series for each set of primers. The relative Wolbachia density and viral titers were calculated as follows: , where Ct is the cycle threshold and .
Because qPCR efficiencies tended to be different between the control gene actin 5C and both the immune and the methyltransferase genes, we use the Pfaffl method to take into account those differences [93]. As dilution series analysis shows the qPCR efficiency for actin 5C to be 100%, the relative efficiency E for the gene of interest can be estimated from the experimental data as . Following [93], Ct values for the gene of interest were corrected for differences in qPCR efficiency as . Levels of gene expression were then estimated as follow: , where . We also normalized the results across 96 well plates (sets of samples were kept in plates for both RNA extraction and qPCR). Thus, expression level for a given sample j was normalized by the mean, , and standard deviation of the corresponding plate i for each gene of interest as follow: . The strains wBai, wBic and wBor were not included in the analysis of host gene expression.
In addition to the single timepoint analysis of viral titers, variation of titers was measured in another infection experiment for Wolbachia-free flies and for the Wolbachia strains wAu, wSh and wAna over a 5 day period. Flies were infected with DCV or FHV and maintained in the same conditions as for the other infection experiments. Live flies were frozen everyday from the day of infection until 5 dpi. For each day and each strain, the RNA was extracted from two pools of ten flies and viral titers measured as explained above except that the RT-qPCR was run on a StepOnePlus thermocycler (Applied Biosystems).
Survival data were analyzed with a Cox's proportional hazards mixed-effect model using the coxme package in R [89]. The Cox's model estimates hazard ratios, which is the probability of a Wolbachia-infected fly dying at a given time-point divided by the probability of a Wolbachia-free fly dying. The infection treatment (DCV, FHV or mock-infected), the Wolbachia infection status (the 19 strains and no Wolbachia) were treated as fixed effects, and the replicate vial as a random effect. The overall significance of multilevel factors or their interactions was tested using likelihood ratio tests to compare models with or without these terms. Flies that were alive at the end of the experiment were treated as censored data. Variation in Wolbachia density and viral titers was analysed using linear models on ln-transformed data to reach the assumptions of normality and homoscedasticity. Differences in viral titer with the Wolbachia-free control were assessed using Dunnett's tests in order to correct for multiple comparisons.
Genetic correlations between traits were estimated by fitting a series of multi-response mixed models using a Bayesian approach in the R package MCMCglmm [94] as follows:where ytwi is the response of the ith biological replicate for Wolbachia strain w, for which we have measured trait t. βt is the intercept term for trait t with a level for each trait, and it can be interpreted as the mean trait value across the Wolbachia strains. The Wolbachia strain random effect, us:tw, is the deviation from the expected value for trait t in strain w. Random effects were assumed to be from multivariate normal distributions with zero mean vectors (illustrated for a model with three traits):where σ2s:t1 is the genetic variance for trait t1, and σs:t1,t2 is the genetic covariance between trait t1 and t2. etwi is a residual capturing the between-vial variation for each trait (within-strain effects, environmental effects and experimental error). Residuals were assumed to be normally distributed and a separate variance was estimated for each trait with the following variance-covariance structure (illustrated for a model with three traits):where I is an identity matrix indicating that strain effects within traits are independent of each other since traits were measured on different biological replicates.
The traits included in these models included the survival of DCV-, FHV- or mock-infected flies (estimated as a negative ln hazard ratio for each vial of flies), the Wolbachia density, viral titer and gene expression (all estimated as ln ). We fitted six different models with different trait combinations. Model 1 included four traits: survival after DCV-, FHV- and mock-infection, and Wolbachia density. Model 2 included three traits: DCV titer, FHV titer and Wolbachia density. Models 3 and 4 included two traits for each virus respectively: survival after viral infection and viral titer. Models 5 and 6 included four traits for each virus respectively: survival as well as Drosomycin, Diptericin and Dnmt2 expression levels after viral infection.
Genetic correlations between two traits can arise either because the traits are causally related or because of phylogenetic non-independence. To explore these explanations we also fitted a phylogenetic mixed model, which included an additional random effect, up:tw, which is the deviation from the expected value for trait t in strain w due to the phylogeny, i.e. the component of the between-strain variation that is explained by the phylogeny [95], [96]:In this model the strain random effect us:tw is the variation that is not accounted for by the phylogeny under a Brownian model of evolution [95]. The intercept βt can be interpreted as the trait value in the Wolbachia strain at the root of the phylogeny. For the phylogenetic effect, up:tw, the following variance-covariance structure was assumed (illustrated for a model with three traits):where A is a matrix with elements ajk standing for the proportion of time that strain j and k have had shared ancestry since the root of the phylogeny. σ2p:t1 is the variance of the phylogenetic effect for trait t1, and σp:t1,t2 is the covariance of phylogenetic effects between trait t1 and t2. Under a Brownian model of evolution, the phylogenetic covariance between two Wolbachia strains is inversely proportional to the time since they diverged from their common ancestor. The phylogenetic effects themselves were poorly estimated and are therefore not reported. We also fitted these phylogenetic models without the strain effect as this improved model convergence and statistical power to test for genetic correlations.
Independent normal priors with zero mean and large variance (1010) were used for the fixed effects and are virtually non-informative in this context. We used several prior probability distributions for the Vp and Vs covariance matrices to ensure our results were robust to the prior selected. Results presented were obtained using parameter expanded priors, but we also fitted models with inverse-Wishart and flat priors that gave equivalent results. We also repeated the analyses after removing outliers so that the distribution of the residuals was normal. In all cases our conclusions were unaffected by these changes. Finally, models were run after removing strains wBai, wBic and wBor, since those strains showed unstable Wolbachia infection status. Estimates of genetic correlations as well as their statistical significance were very similar to models that include these 3 strains, and are therefore not reported. The models were run for 13,000,000 iterations with a burn-in of 3,000,000. We checked for convergence by visually examining the trace of the posterior sample and ensuring the autocorrelation between successive samples in the MCMC chain was <0.1. Credible intervals (CI) were estimated from the posterior distribution of parameter estimates as the 95% highest posterior density intervals.
|
10.1371/journal.ppat.1007102 | Manganese acquisition is essential for virulence of Enterococcus faecalis | Manganese (Mn) is an essential micronutrient that is not readily available to pathogens during infection due to an active host defense mechanism known as nutritional immunity. To overcome this nutrient restriction, bacteria utilize high-affinity transporters that allow them to compete with host metal-binding proteins. Despite the established role of Mn in bacterial pathogenesis, little is known about the relevance of Mn in the pathophysiology of E. faecalis. Here, we identified and characterized the major Mn acquisition systems of E. faecalis. We discovered that the ABC-type permease EfaCBA and two Nramp-type transporters, named MntH1 and MntH2, work collectively to promote cell growth under Mn-restricted conditions. The simultaneous inactivation of EfaCBA, MntH1 and MntH2 (ΔefaΔmntH1ΔmntH2 strain) led to drastic reductions (>95%) in cellular Mn content, severe growth defects in body fluids (serum and urine) ex vivo, significant loss of virulence in Galleria mellonella, and virtually complete loss of virulence in rabbit endocarditis and murine catheter-associated urinary tract infection (CAUTI) models. Despite the functional redundancy of EfaCBA, MntH1 and MntH2 under in vitro or ex vivo conditions and in the invertebrate model, dual inactivation of efaCBA and mntH2 (ΔefaΔmntH2 strain) was sufficient to prompt maximal sensitivity to calprotectin, a Mn- and Zn-chelating host antimicrobial protein, and for the loss of virulence in mammalian models. Interestingly, EfaCBA appears to play a prominent role during systemic infection, whereas MntH2 was more important during CAUTI. The different roles of EfaCBA and MntH2 in these sites could be attributed, at least in part, to the differential expression of efaA and mntH2 in cells isolated from hearts or from bladders. Collectively, this study demonstrates that Mn acquisition is essential for the pathogenesis of E. faecalis and validates Mn uptake systems as promising targets for the development of new antimicrobials.
| Enterococcus faecalis is a leading cause of hospital-acquired infections that are often difficult to treat due to their exceptional multidrug resistance. Manganese (Mn) is an essential micronutrient for bacterial pathogens during infection. To prevent infection, the host limits Mn bioavailability to invading bacteria in an active process known as nutritional immunity. To overcome this limitation, bacteria produce high-affinity Mn uptake systems to scavenge Mn from host tissues. Here, we identified the main Mn transporters of E. faecalis and show that, by working collectively, they are essential for growth of this opportunistic pathogen in Mn-restricted environments. Notably, the inability to acquire Mn during infection rendered E. faecalis virtually avirulent in different animal models, thereby revealing the essentiality of Mn acquisition to enterococcal pathogenesis. The results reported here highlight that bacterial Mn transport systems are promising targets for the development of novel antimicrobial therapies, which are expected to be particularly powerful to combat enterococcal infections.
| While normal residents of the gastrointestinal (GI) tract of animals and humans, enterococci are also the third most frequent cause of hospital-acquired infections and a major threat to public health due to the alarming rise of multidrug-resistant isolates [1]. Enterococcal infections in humans are mainly caused by Enterococcus faecalis and Enterococcus faecium, with the great majority of infections (~70%) caused by E. faecalis. Collectively, enterococci rank as the third leading etiological agent in infective endocarditis (IE) [2], second in complicated urinary tract infections (UTI) [3], and one of the leading causes of device-associated infections and bacteremia [1]. Despite the recent introduction of new antibiotics active against both E. faecalis and E. faecium (i.e. daptomycin, linezolid, tigecycline), the indiscriminate use of antibiotics and the rise in elderly and severely ill populations susceptible to infection continues to contribute to a worldwide increase in enterococcal infections [4, 5]. The pathogenic potential of E. faecalis, and more generally, of all enterococci, has been largely attributed to their harsh and extremely durable nature, which includes intrinsic tolerance to commonly-used antibiotics (such as cephalosporins), chlorine, alcohol-based detergents, and an ability to survive extreme fluctuations in temperature, pH, oxygen tension, humidity and nutrient availability [1].
We recently showed that virulence-related phenotypes of a strain unable to synthesize the nutritional alarmone (p)ppGpp, known as (p)ppGpp0 strain, were directly linked to manganese (Mn) homeostasis, as important phenotypes of the (p)ppGpp0 strain could be reverted by addition of Mn to laboratory growth medium or serum [6]. Mn is an essential micronutrient for bacterial pathogens and hosts alike [7]. In lactic acid bacteria such as E. faecalis, Mn is the co-factor of enzymes of central metabolic pathways, such as energy generation and DNA biosynthesis, and has been shown to play an important role in oxidative stress responses [6, 8]. Similarly, vertebrates require Mn for a variety of cellular pathways such as lipid, protein, and carbohydrate metabolism [9, 10] and, depending on the tissue, Mn levels range from 0.3 to 2.9 μg per gram of tissue [11, 12]. While the lowest range concentration found in tissues is more than enough to promote bacterial growth, the mammalian host restricts the availability of essential metals such as Mn and iron (Fe) to invading pathogens by producing small molecules and proteins that tightly bind to metals, an active process termed nutritional immunity [13]. For example, Fe-binding proteins such as transferrin (TF) and lactoferrin (LF) are utilized by the host to chelate Fe in serum (TF) and mucose (LF), thereby restricting its bioavailability to invading pathogens [14]. The bioavailability of Mn and zinc (Zn) in the host is restricted, at least in part, by calprotectin, a Mn/Zn-sequestering protein of the S100 family that accounts for more than 40% of the total protein content of neutrophils [15]. In addition, macrophages express an Nramp (Natural Resistance-Associated Macrophage Protein)-type Fe/Mn transporter that starves phagocytosed bacteria to promote clearance in the phagolysosome [7]. Notably, mice defective of calprotectin or the Nramp-type macrophage transporter are significantly more susceptible to bacterial infections [16, 17].
To overcome host-imposed Mn restriction, bacteria produce their own high-affinity Mn transporters. Presently, three classes of Mn transporters are known in bacteria: i) ABC-type permeases, ii) Nramp-type H+/Mn transporters, and iii) the less common P-type transporters [7]. Not surprisingly, Mn uptake systems have been identified as major virulence factors for several gram-negative and gram-positive pathogens [7, 8, 18]. For instance, in Salmonella enterica serovar Typhimurium, deletion of the ABC transporter sitABCD alone or in combination with the Nramp transporter mntH led to virulence attenuation in a murine systemic infection model [19]. Similarly, in Staphylococcus aureus, combined inactivation of the ABC-type transporter mntABC and the Nramp-type transporter mntH reduced staphylococcal virulence in murine skin abscess [20] and systemic infection models [17]. In streptococci, which are more closely related to the enterococci, deletion of the ABC-type Mn transporter attenuated virulence of Streptococcus mutans, S. parasanguinis and S. sanguinis in rat or rabbit models of IE as well as of Streptococcus pyogenes in a murine model of skin infection [21–25]. Remarkably, deletion of the ABC-type Mn transporter psaBCA rendered Streptococcus pneumoniae completely avirulent in at least three different animal models [26].
The significance of Mn homeostasis and specifically Mn acquisition systems to E. faecalis pathogenesis has not been explored in detail. In silico analysis indicates that the core genome of E. faecalis encodes three putative Mn transporters: the ABC-type transporter EfaCBA and two Nramp-type transporters designated MntH1 and MntH2 [27, 28]. Previous global transcriptional (microarray) analysis revealed that transcription of these genes, particularly efaCBA and mntH2, is strongly induced in blood or in urine ex vivo as well as in a murine peritonitis model [29–31]. In addition, transcription of efaCBA, mntH2 and to a much lesser extent mntH1 is induced in Mn-depleted laboratory medium while repressed in Mn-rich medium [6, 32–34]. Notably, an earlier study identified EfaA (the lipoprotein component of the EfaCBA complex) as a prominent antigen in enterococcal IE that can be used as an immunodiagnostic tool to discriminate E. faecalis from other IE-causing bacteria [35, 36]. Finally, virulence of a strain lacking efaA was slightly attenuated in the murine peritonitis model [36].
To determine the significance of Mn homeostasis in the pathophysiology of E. faecalis, we generated a panel of single (Δefa, ΔmntH1, ΔmntH2), double (ΔefaΔmntH1, ΔefaΔmntH2, ΔmntH1ΔmntH2) and triple (ΔefaΔmntH1ΔmntH2) deletion mutants in strain OG1RF and tested their ability to grow under metal-restricted conditions and to cause infection in three different animal models. We found that production of only one of the three Mn transporters is sufficient to support growth of E. faecalis in Mn-depleted media, a finding that is supported by the drastic reduction in cellular Mn levels in the triple ΔefaΔmntH1ΔmntH2 mutant but not in single or double mutant strains. Loss of efaCBA alone impaired growth in serum or in the presence of calprotectin in vitro. In most cases, inactivation of mntH2 exacerbated the phenotypes of the Δefa mutant, suggesting that EfaCBA and MntH2 are the primary high-affinity Mn transporters of E. faecalis. While only ΔefaΔmntH1ΔmntH2 showed virulence attenuation in the Galleria mellonella invertebrate model, both ΔefaΔmntH2 double and ΔefaΔmntH1ΔmntH2 triple mutant strains were virtually avirulent in the two mammalian models tested (i.e., rabbit IE and murine catheter-associated UTI). Collectively, this investigation highlights the essentiality of Mn acquisition to E. faecalis pathogenesis, suggesting that pathways associated with Mn homeostasis are promising targets for the development of new antimicrobials.
The annotations of EF2074-2076 (OG1RF11677-11679), EF1901 (OG1RF11567) and EF1057 (OG1RF10838) as efaCBA, mntH1 and mntH2, respectively, have been inferred based on their similarities to previously characterized Mn transport systems and their Mn-dependent transcriptional profiles [32–34]. To uncover the role of these systems in metal uptake, single (Δefa, ΔmntH1, ΔmntH2), double (ΔefaΔmntH1, ΔefaΔmntH2, ΔmntH1ΔmntH2) and triple (ΔefaΔmntH1ΔmntH2) deletion strains were generated using a markerless system [37]. Brain Heart Infusion (BHI) broth has been previously reported to contain relatively low levels of Mn when compared to other laboratory media, such as the chemically-defined FMC medium [6, 25]. Using inductively coupled plasma-optical emission spectrometry (ICP-OES), we confirmed this to also be the case for BHI prepared in our laboratory—Mn, Fe and Zn concentrations were ~ 1.1 μM (Mn), ~ 3.9 μM (Fe) and ~ 11.3 μM (Zn) (Fig 1A). With those values in mind, all mutant strains were isolated on BHI plates supplemented with 150 μM MnSO4. Upon genetic confirmation that all gene deletions occurred as planned, we tested the capacity of all mutant strains to grow on unsupplemented BHI. While single and double mutant strains grew well on plain BHI plates, the triple ΔefaΔmntH1ΔmntH2 mutant could only grow on Mn-supplemented BHI (Fig 1B). To confirm this initial observation, growth of the parent OG1RF and mutant strains was monitored in the chemically-defined FMC medium depleted of Mn (Mn < 90 nM) [6]. Growth of all mutant strains (singles, doubles and triple mutant strains) in complete FMC (Mn-replete) was indistinguishable from growth of the parent strain (Fig 1C and 1E). Single Δefa, ΔmntH1, and ΔmntH2, as well as double ΔefaΔmntH1 and ΔmntH1ΔmntH2 behaved similarly to the parent strain when Mn was depleted from the medium (Fig 1D and 1F). In contrast, the ΔefaΔmntH2 double mutant strain displayed an extended lag phase and slightly reduced final growth yields in Mn-depleted FMC (Fig 1F). Most notably, both growth rate and final growth yields were severely impaired in the triple mutant strain under these conditions (Fig 1F). Ectopic expression of any one of the three Mn transporters from a plasmid (pTG-efa, pTG-mntH1 or pTG-mntH2) partially rescued the growth defect of the triple mutant on unsupplemented BHI agar and restored growth in Mn-depleted FMC (Fig 1G and 1H). Collectively, these results strongly support that EfaCBA, MntH1 and MntH2 are bona fide, functionally redundant, Mn transporters.
Next, we used ICP-OES to quantify cellular Mn content in the different Mn transport mutants. In agreement with the results shown in Fig 1, single deletions of efaCBA, mntH1 or mntH2 did not affect cellular Mn content (Fig 2A). However, combined deletion of mntH2 with either efaCBA or mntH1 (ΔefaΔmntH2 and ΔmntH1ΔmntH2) caused a significant reduction (~ 40%) in intracellular Mn pools (Fig 2B). Most notably, Mn pools were below detection limit in the triple mutant strain, thereby providing unequivocal evidence of the cooperative nature of EfaCBA, MntH1 and MntH2 in Mn acquisition. Taking into account that a balanced Mn/Fe ratio is essential for cellular homeostasis, and that bacterial systems homologous to EfaCBA have been shown to also transport Fe [7, 38], we used ICP-OES to also determine the cellular Fe content in our panel of mutant strains. While Fe pools were not affected in single and double mutants of Nramp-type transporters (ΔmntH1, ΔmntH2 and ΔmntH1ΔmntH2), loss of EfaCBA, either alone or in combination with MntH1 and MntH2, led to ~ 40% reduction in cellular Fe content (Fig 2), indicating that, similar to streptococci, EfaCBA is a dual Mn/Fe transporter [7, 25]. To rule out that loss of mntH1, mntH2 or both can lead to further reductions in intracellular Fe in the Δefa background strain, we repeated the ICP-OES analysis of Δefa, ΔefaΔmntH1, ΔefaΔmntH2, and ΔefaΔmntH1ΔmntH2 strains in the same experiment. Inactivation of mntH2 (or mntH1) in the Δefa background did not result in further reductions in Fe content (S1 Fig, panel A).
Considering that EfaCBA participates in Fe uptake, we evaluated whether Fe supplementation (150 μM FeSO4) could restore growth of the ΔefaΔmntH1ΔmntH2 triple mutant strain on BHI plates. We found that, different than Mn (Fig 1), Fe supplementation does not restore growth of the triple mutant in BHI agar (S1 Fig, panel B). We also monitored growth of the different mutant strains in FMC depleted of Fe (FMC ↓ Fe), or both Fe and Mn (FMC ↓ Mn ↓ Fe) (S1 Fig, panels C-F). While the single Δefa strain displayed a moderate growth defect during Fe limitation, the double and triple mutant strains were indistinguishable from the parent strain. When Fe and Mn were simultaneously depleted from the medium, the ΔefaΔmntH2 and ΔefaΔmntH1ΔmntH2 mutants displayed an exacerbated growth defect when compared to growth in Mn-depleted medium (FMC ↓ Mn) (Fig 1, panel F), supporting the notion that Fe and Mn can act as interchangeable co-factors in E. faecalis [6]. Collectively, these results indicate that the growth defect of the ΔefaΔmntH1ΔmntH2 strain can be primarily associated to its inability to acquire Mn.
Metal sequestration is an ancient infection defense mechanism that spans various kingdoms of life [39]. Insects, such as G. mellonella, produce Fe-chelating proteins homologous to mammalian TF and ferritin and could theoretically use similar strategies to chelate Mn [40–42]. To probe the significance of Mn transport in E. faecalis virulence, we first used the G. mellonella larvae model of systemic infection. All single and double mutants killed G. mellonella at rates that were comparable to the parent OG1RF strain (S2 Fig), with similar averages of larvae survival (~ 27%) 72 hours post-infection (Fig 3). However, virulence of the triple ΔefaΔmntH1ΔmntH2 mutant was dramatically impaired with ~ 87% larvae survival after 72 hours (Fig 3, S2 Fig).
The Mn- and Zn-binding calprotectin is a major protein used by vertebrates to restrict the availability of these two transition metals to microbes during infection [15]. Normally found in circulating blood and tissues at relatively low levels, calprotectin rapidly accumulates to concentrations of up to 1mg ml-1 in response to inflammation and infection [15], thereby depleting Zn and Mn from the infected site. Of note, previous studies indicated that the antimicrobial activity of calprotectin against E. faecalis is, at least in part, due to Mn sequestration [15]. Here, we assessed the ability of the Mn transport mutants to grow in the presence of a sub-inhibitory concentration of purified calprotectin. At the selected concentration, addition of calprotectin delayed growth of the parent OG1RF strain at the initial time-points, but did not impact final growth yields after 24 hours of incubation when compared to untreated cells (S3 Fig). Of note, growth of the ΔefaΔmntH1ΔmntH2 triple mutant strain was impaired even in the absence of calprotectin, most likely due to the low Mn levels present in this medium (S3 Fig). Loss of mntH1 (ΔmntH1) did not affect enterococcal tolerance to calprotectin when compared to the parent strain. However, inactivation of mntH2, either by itself (ΔmntH2) or combined with mntH1 (ΔmntH1ΔmntH2), moderately impaired growth of E. faecalis in the presence of calprotectin (Fig 4 and S3 Fig). Interestingly, growth rates and growth yields of any one of the efaCBA mutants was severely impaired by calprotectin. As expected, double inactivation of efaCBA and mntH2 (ΔefaΔmntH2) rendered E. faecalis even more susceptible to the inhibitory effects of calprotectin. These results revealed that EfaCBA, closely followed by MntH2, is the primary Mn transporter used by E. faecalis to overcome the inhibitory effects of calprotectin.
A member of the S100 family, calprotectin forms a S100A8/S100A9 heterodimer with two distinct metal-binding sites: a high-affinity Zn-binding site and a dual high-affinity Zn/Mn-binding site. To confirm that the increased sensitivity of Δefa strains to calprotectin was due to Mn sequestration and not to Zn sequestration, we tested the sensitivity of wild-type and selected mutant strains against a calprotectin variant (ΔMn Tail calprotectin) that retains the ability to chelate Zn but is unable to chelate Mn due to amino acid substitutions of key histidine residues (His103 and His105) in the C-terminal tail of calprotectin [15]. The inability to chelate Mn by the calprotectin ΔMn Tail variant restored final growth yields of all mutant strains tested to wild-type levels (Fig 4, and S4 Fig, panel A), strongly suggesting that EfaCBA, and to a lesser extent MntH2, promote tolerance to calprotectin in a Mn-dependent manner. In line with this finding, the antimicrobial activity of native calprotectin could be fully overcome by Mn supplementation (10 μM MnSO4) (S4 Fig). Because calprotectin has been shown to also scavenge Fe in vitro [43], we also tested whether Fe supplementation (10 μM FeSO4) could rescue calprotectin tolerance in E. faecalis. Addition of Fe rendered the wild-type and ΔmntH2 strains even more sensitive to calprotectin, while partially restoring growth of the Δefa single mutant (S4 Fig). More importantly, Fe did not rescue calprotectin tolerance of the double ΔefaΔmntH2 strain. Collectively these results indicate that EfaCBA and MntH2 mediate calprotectin tolerance by primarily facilitating Mn uptake.
Next, we monitored growth and survival of the Mn transport mutants in pooled human serum at 37°C—serum was used instead of whole blood to avoid free metal contamination released by lysing erythrocytes. While inactivation of mntH1 and mntH2, alone or in combination, did not affect the ability of E. faecalis to grow and survive in serum, loss of efaCBA led to a significant decrease in survival after 48 hours incubation when compared to the parent strain (~1.3 log Δefa, ~1.5 log ΔefaΔmntH1, ~2.3 log ΔefaΔmntH2, ~1.8 log, ΔefaΔmntH1ΔmntH2) (Fig 5A and 5B). To determine if the survival defect of Δefa strains was due to an inability to scavenge Mn or Fe from the environment, we supplemented serum with 1 mM MnSO4 or 1 mM FeSO4. Of note, the metal stocks chosen for supplementation were 99% pure to minimize the presence of other transition metals. The addition of either Mn or Fe restored serum growth and survival of Δefa strains to wild-type OG1RF levels (Fig 5C). Moreover, Mn or Fe supplementation increased final growth yields of all strains, including the parent strain, confirming that both metals are growth-limiting factors in the human serum. Collectively, these results indicate that the EfaCBA system plays a primary role in promoting E. faecalis serum survival, likely because of its dual capacity to function as a Mn and Fe transporter.
Patients with enterococcal IE are known to generate specific antibodies against EfaA, the substrate-binding lipoprotein component of the EfaCBA system [32, 35]. While an efaA single mutant was previously shown to be slightly attenuated in a mouse peritonitis model [36], the contribution of EfaCBA, or the other Mn transport systems MntH1 and MntH2, in enterococcal IE has not been explored. Here, we used a catheterized rabbit IE model [44] to determine the ability of selected Mn transport mutants to colonize a previously-formed sterile heart vegetation and, then, systemically spread to different organs by using spleen homogenates as a readout. In the first experimental set, the parent OG1RF strain was co-inoculated systemically (via ear vein injection) with single Δefa and triple ΔefaΔmntH1ΔmntH2 strains. Forty-eight hours post-infection, an average of 7.5 (+/- 0.7 SD) and 4.1 (+/- 0.9 SD) log10 CFU E. faecalis were recovered from hearts and spleens, respectively, confirming that E. faecalis can efficiently colonize the injured heart endothelium and spread systemically in this model. PCR was used to screen E. faecalis colonies (50 to 100 per animal and organ) to distinguish between the three strains based on the different amplification products obtained using efaCBA and mntH2 flanking primers. Of the total colonies screened from heart vegetations, ~ 55% corresponded to the parent OG1RF and ~ 45% to the Δefa single mutant strain (Fig 6A). While the average recovery rates of OG1RF and Δefa were not statistically significant (p>0.05), there was a great variation from animal to animal, with the parent OG1RF corresponding to the large majority (>90%) of colonies recovered in two animals and Δefa predominating at ~ 80% in the remaining three animals (Fig 6A). Similar trends were observed in the corresponding spleens of the individual animals (~ 49% OG1RF and ~ 51% Δefa, p>0.05) (Fig 6B), indicating that systemic dissemination is likely a direct consequence of bacterial seeding from the heart vegetation. Most importantly, none of the screened colonies in this experiment corresponded to the triple mutant strain in either heart vegetations or spleens (Fig 6A and 6B).
Based on the strong growth defect of the double ΔefaΔmntH2 in the presence of calprotectin and in serum (Figs 4 and 5), we next tested the ability of the parent OG1RF, ΔmntH2 and ΔefaΔmntH2 strains to colonize the heart vegetations in a triple co-infection experiment. In this second set of experiments, the average numbers of OG1RF and ΔmntH2 recovered from heart vegetations and spleens were not significantly different (~ 40% OG1RF, ~ 60% ΔmntH2, p>0.05) (Fig 6C and 6D). Interestingly, the double ΔefaΔmntH2 mutant phenocopied the triple mutant strain, as none of the colonies screened from each animal corresponded to the ΔefaΔmntH2 strain. Of note, none of the mutant strains tested displayed a growth/fitness defect in vitro when co-cultured in BHI supplemented with 150 μM MnSO4 (S5 Fig). Collectively, these results strongly support that EfaCBA and MntH2 are the primary Mn transporters of E. faecalis during infection and further underscore the functional redundancy shared by these two Mn transporters in vivo.
Previous studies have shown that transcription of efaCBA and mntH2 is induced when E. faecalis is grown in urine, indicating that Mn is also a growth-limiting nutrient in the bladder environment [30, 45]. We used ICP-OES to determine the Mn content in pooled human urine obtained from healthy donors and found this batch to be low in Mn (~ 0.7 μM), which is within the normal clinical range described elsewhere [10]. Copper (Cu) (~ 0.2 μM) and Fe (~ 0.7 μM) levels were also low in pooled urine, whereas Zn levels (~ 9.2 μM) were comparatively much higher. Nevertheless, it should be noted that these values represent total metal concentrations present in urine not taking into account their bioavailability. Next, we tested the ability of the mutant strains to grow in human urine supplemented with bovine serum albumin (BSA) at 37°C—BSA was used to supplement urine to mimic protein infiltration during catheter-associated urinary tract infections (CAUTI), a condition shown to promote growth of E. faecalis in the bladder environment [46, 47]. We found that both the double ΔefaΔmntH2 and triple ΔefaΔmntH1ΔmntH2 mutant strains displayed a significant growth defect after 24 hours of incubation (Fig 7A). While the parent, single and double mutant strains entered stationary phase and maintained the same growth yields for up to 48 hours, the ΔefaΔmntH1ΔmntH2 triple mutant also displayed a survival defect after 48 hours (~ 1 log10 reduction in CFU ml-1, p < 0.05) when compared to the other strains (S6 Fig). Importantly, MnSO4 supplementation (1 mM) fully restored growth of the triple mutant strain in urine, while addition of FeSO4 (1 mM) only partially rescued the defective phenotype (Fig 7B).
To determine whether Mn acquisition is also relevant to enterococcal UTI, we tested the parent OG1RF and each individual mutant strain in a murine CAUTI model [48]. Twenty-four hours post-infection, the OG1RF, Δefa and ΔmntH1 strains were recovered in similar numbers from the bladders of infected animals, whereas the ΔmntH2 strain had a ~ 1.5 log10 CFU reduction (p<0.05) in total bacteria recovered (Fig 8A). Inactivation of mntH1 did not exacerbate the phenotype of the mntH2 single gene inactivation (ΔmntH1ΔmntH2) but simultaneous inactivation of efaCBA and mntH1 (ΔefaΔmntH1) resulted in ~ 1 log10 CFU reduction in bacteria recovered from the bladder (p<0.05). In agreement with the mounting evidence supporting that EfaCBA and MntH2 are the primary systems for Mn acquisition, combined deletion of efaCBA and mntH2 (ΔefaΔmntH2) or all three Mn transporters (ΔefaΔmntH1ΔmntH2) further reduced bacterial loads recovered from the bladder to below or near the detection limit (Fig 8A). The exact same trends found in the bladder were observed for bacteria recovered from biofilms formed on the surface of the implanted catheters, i.e. significantly lower bacterial burden for strains lacking mntH2 and very little to no bacterial cells recovered from strains lacking both efaCBA and mntH2 (Fig 8B). Interestingly, the bacterial counts of strains with an attenuated phenotype were lower on catheters than in the bladders, suggesting that Mn uptake is especially critical for biofilm formation on urinary catheters. Moreover, while the parent OG1RF was able to ascend to the kidneys in 55% of the mice and disseminate to spleen and heart in ~ 30% of the animals infected, the Mn transport mutants were rarely recovered from kidneys and were almost never isolated from more distant organs such as spleen and heart (S7 Fig). The overall impaired systemic dissemination of Mn transport mutants is in line with the importance of Mn acquisition systems for enterococcal survival in serum (Fig 5) and for infection in the rabbit IE model (Fig 6).
Biofilm formation of E. faecalis on urinary catheters is a critical step in UTI and is primarily mediated by the enterococcal surface adhesin EbpA [47, 49]. Notably, the metal ion-dependent adhesion site (MIDAS) of EbpA is required for biofilm formation in urine and for virulence in the experimental CAUTI model [47, 50]. While the identity of the metal bound to the tip of EbpA remains elusive, we wondered if the colonization defect of strains lacking one or more Mn transporters could relate to impaired biofilm formation and/or defects in EbpA production. First, we tested the ability of the parent and mutant strains to form biofilms on the surfaces of tissue culture plate wells and plastic catheters that were pre-coated with fibrinogen to promote biofilm formation in an EbpA-dependent manner [47]. After 24 hours of incubation in urine supplemented with BSA, the single Δefa, double ΔefaΔmntH2 and triple ΔefaΔmntH1ΔmntH2 mutant strains formed significantly less biofilm as measured by either crystal violet staining of tissue culture plates (Fig 9A) or catheter immunostaining (Fig 9B). In agreement with these findings, the double ΔefaΔmntH2 and triple ΔefaΔmntH1ΔmntH2 mutant strains displayed a ~ 1 log10 CFU reduction in catheter-associated bacteria when compared to the parent strain (Fig 9C). While compatible with the slight reduction in biofilm formation (Fig 9A and 9B), the ~ 0.5 log10 CFU reduction observed with the Δefa single mutant on the catheter surface was not statistically significant (Fig 9C). Finally, we used ELISA to quantify the surface expression levels of EbpA in the different strains grown under the same conditions used in the biofilm assay, i.e. urine supplemented with BSA. Despite defects in biofilm formation of Δefa, ΔefaΔmntH2 and ΔefaΔmntH1ΔmntH2 strains on fibrinogen-coated surfaces, there were no apparent differences in EbpA levels between parent and mutant strains (Fig 9D), suggesting that the defective biofilm phenotype is not EbpA-dependent.
The results described above suggest that EfaCBA is the chief Mn transporter in bloodstream infections while MntH2 appears to play a more prominent role in CAUTI. Given the functional redundancy of these Mn transporters, we wondered if efaCBA and mntH2 were differentially expressed when E. faecalis is in the bloodstream or in urine, thereby providing an explanation for the different phenotypic behaviors of Δefa and ΔmntH2 strains in different sites of infection. To address this possibility, we used quantitative reverse transcription-PCR (qPCR) to determine efaA, mntH1 and mntH2 transcription levels in vivo. When compared to cells grown in trypsinized beef heart medium, we found that efaA transcription was strongly induced in cells isolated from rabbit heart valves (~ 15-fold after two days and ~ 100-fold after three-four days) (Fig 10A). On the other hand, transcription of mntH1 was moderately repressed after infection (7-fold and 3-fold repression after two and three-four days, respectively), whereas mntH2 transcription remained largely unaltered over time (Fig 10A). Strikingly, mntH2 was the only gene up-regulated (~ 10-fold induction) in cells recovered from the bladders of mice 24 hours post-infection, while efaA and mntH1 transcription was not significantly altered (Fig 10B). Since Mn plays a crucial role in promoting oxidative stress tolerance in lactic acid bacteria [6], in part by acting as the co-factor of superoxide dismutase (sodA), we also determined the transcription levels of sodA (EF0463, OG1RF10348) as a readout for oxidative stress in IE and CAUTI. Surprisingly, sodA transcription was downregulated (~ 15-fold) at the early stage of endocarditis infection (two days) and did not differ from the inoculum condition at the later stage of infection (three-four days) (Fig 10A). Similarly unexpected, cells recovered from the bladders of infected mice displayed a ~ 25-fold reduction in sodA one day post-infection (Fig 10B).
In this study we confirm previous in silico predictions that the E. faecalis core genome encodes three bona fide Mn transporters: one ABC-type (EfaCBA) and two Nramp-type transporters (MntH1 and MntH2) [27]. Early studies showed that the virulence of an E. faecalis efaA mutant was slightly delayed in a mouse peritonitis model despite the fact that the strain failed to display noticeable phenotypes in vitro [33, 36]. Here, we provide an explanation for such findings by showing that only the simultaneous inactivation of two or all three transporters can drastically impair Mn homeostasis under laboratory as well as in vivo conditions. Specifically, the ΔefaΔmntH1ΔmntH2 triple mutant strain was unable to grow (or grew very poorly) in Mn-restricted environments—a condition commonly encountered in human tissues—and was unable to robustly infect vertebrate and invertebrate animal hosts. To our knowledge, this is the first description of an E. faecalis mutant strain being virtually avirulent in multiple animal infection models. Moreover, this is also the first demonstration that the ability to acquire Mn using high affinity transporters within the urinary tract environment was shown to be an important aspect for the development of bacterial UTI.
While EfaCBA, MntH1 and MntH2 appear to be functionally redundant, simultaneous inactivation of efaCBA and mntH2 often phenocopied the triple mutant strain. Specifically, the double ΔefaΔmntH2 mutant strain was highly sensitive to calprotectin, displayed a biofilm formation defect in human urine and was virtually avirulent in the IE and CAUTI models. None of these phenotypes were further exacerbated in the triple mutant strain that also lacked mntH1. Considering that mntH1 transcription is not as strongly affected by in vitro Mn fluctuations as efaCBA and mntH2 [6, 34], it is tempting to speculate that EfaCBA and MntH2 are the primary high-affinity Mn transporters of E. faecalis responsible for maintaining Mn homeostasis in Mn-restricted environments. In contrast, MntH1 may serve as a housekeeping transporter, possibly with lower affinity for Mn than EfaCBA and MntH2, under Mn-replete conditions. The specificity and affinity (high/low) of each transport system for Mn, as well as other important transition metals such as Cu, Fe and Zn, will be determined in future studies.
Lactic acid bacteria like enterococci and streptococci are notoriously Mn-centric organisms, having a much higher nutritional demand for Mn than other bacterial groups [38]. This metabolic particularity might explain why the striking loss of virulence observed here contrasts with the moderate virulence attenuation of Mn transport mutants in gram-negative and other gram-positive pathogens [7, 8, 17]. Indeed, the complete loss of virulence of ΔefaΔmntH2 and ΔefaΔmntH1ΔmntH2 strains shown here is only comparable to the virulence defects of streptococcal strains lacking EfaCBA orthologs [25, 51, 52]. Interestingly, the observation that E. faecalis produces three highly conserved Mn transporters instead of two, such as most streptococci [53], may be an indication that this opportunistic pathogen is better equipped to acquire Mn from the host environment than some of the closely-related pathogenic species such as S. pyogenes and S. pneumoniae. Alternatively, it is possible that E. faecalis may simply have a higher cellular demand for Mn than streptococci, having to rely on multiple Mn transporters to meet such high demand. This would resemble Lactobacillus plantarum, a non-pathogenic organism with five annotated Mn transporters (one P-type, one ABC-type, and three Nramp-type transporters) that has one of the highest cellular requirements for Mn among gram-positive and gram-negative bacteria [38, 54].
While Mn is the co-factor of several growth-promoting bacterial enzymes, Mn is thought to mediate bacterial virulence mainly by protecting cells from host-derived reactive oxygen species (ROS) [7, 8, 18, 44, 53, 55]. Previous studies have shown that Mn contributes to elimination of damaging ROS via three distinct mechanisms: (i) by direct non-enzymatic scavenging of superoxide radicals, (ii) by serving as the co-factor of the Mn-dependent superoxide dismutase (SOD) enzyme, and (iii) by replacing Fe as an enzymatic co-factor thereby protecting Fe-binding proteins from Fenton chemistry damage [8, 38, 56]. Not surprisingly, Mn transport mutants of different bacterial species display enhanced sensitivity to oxidative stresses in vitro and reduced macrophage survival [18, 19, 53, 57]. We were previously able to show that Mn availability is critical for enterococcal tolerance to hydrogen peroxide [6]. Here, we demonstrate that similar to S. aureus and Neisseria gonorrhoeae [20, 58], inability to acquire Mn with high affinity increases the sensitivity of the E. faecalis ΔefaΔmntH2 strain to paraquat, a superoxide generator (S8 Fig). Addition of Mn to the medium fully restored paraquat tolerance of the mutant strain to wild-type levels, indicating that EfaCBA and MntH2 contribute to superoxide tolerance only in Mn restricted environments. Global transcriptional analyses of E. faecalis grown in whole blood or urine ex vivo or isolated from a murine peritonitis model showed that transcription of several genes associated with ROS detoxification, including sodA, is induced when compared to cells grown in laboratory medium, indicating that bacterial cells have to cope with ROS stress during invasive infections [29–31]. Unexpectedly, we found that transcription of sodA was downregulated in the early stages of IE and CAUTI infection (Fig 10). While further studies are needed to determine if E. faecalis does not encounter oxidative stress during these infections, it is possible that the downshift in sodA is a response mechanism to lower the cellular Mn requirement when this nutrient is already restricted. In support of this possibility, sodA transcription has been previously shown to be repressed in E. faecalis during Mn limitation in vitro and in cells recovered from a rabbit subdermal abscess model 8 hours post-infection [6, 59]. Nevertheless, the discrepant transcriptional induction of sodA in different studies remains to be resolved. One possibility is that, during infection, oxidative stress fluctuates according to the dynamic and temporal immune responses mounted by the host.
By obtaining the transcriptional profile of efaA, mntH1 and mntH2 during IE and CAUTI, we found that the transcriptional responses of each system to different environmental cues in vivo can greatly differ. This was particularly noticeable for efaA and mntH2, since efaA was strongly induced in cells recovered from heart valves whereas mntH2 was induced in CAUTI (Fig 10). In contrast, transcription of mntH1 was repressed in heart valves while remaining largely unaltered in CAUTI. Considering that efaCBA, mntH1 and mntH2 have been previously shown to be regulated by EfaR [33], a metalloregulator from the DtxR family, the different transcriptional profiles of these genes were somewhat unexpected. However, the same study proposed that, similar to Bacillus subtilis and S. enterica, there could be other factors regulating transcription of Mn transport genes in E. faecalis [8, 33, 60, 61]. In fact, in silico analysis of the E. faecalis OG1RF genome identified two Fur-binding consensus sequences upstream efaC, the first gene in the efaCBA operon, indicating that this dual Fe/Mn transporter may also be regulated in response to Fe availability, oxidative stress or both. Future studies are warranted to identify the environmental cues present in blood and urine that trigger the different transcriptional responses of efaCBA and mntH2, as well as the potential cis and trans-acting elements regulating these responses.
In addition to differences in expression and, possibly, metal-binding affinity, the distinct behaviors of efaCBA and mntH2 mutants in the context of IE and CAUTI may have additional explanations. For instance, EfaA homologs of S. sanguinis (SsaB), S. parasanguinis (FimA) and S. pneumoniae (PsaA) have been proposed to act as adhesins to a variety of relevant surfaces [53]. While the role of Mn ABC transport permeases as moonlighting proteins that participate in cell adhesion is a subject of debate [53], the possibility that EfaA can also function as a surface adhesin should not be completely excluded at this point. Alternatively, considering that the availability of free Mn and Fe is greatly restricted in the bloodstream, it is possible that the dual ability of EfaCBA to import both metals, as suggested by cellular metal quantifications of efa mutants in this and other studies [62], is physiologically relevant during bloodstream infections but not during UTI. In support of this possibility, both Fe and Mn are growth-limiting factors for E. faecalis in serum (Fig 5) [6], while only Mn could fully rescue the growth defect of the triple Mn transport mutant in urine (Fig 7). Importantly, the biological significance of Fe to bacterial infection is highlighted by the high incidence of opportunistic infections in individuals with elevated levels of circulating free Fe due to hemochromatosis [14]. Similarly, elevated tissue levels of Mn were recently show to promote staphylococcal abscess formation in the murine heart [55], collectively supporting the notion that Fe and Mn availability facilitates bloodstream infections.
At this time, the dominant role of MntH2 in CAUTI is less clear. Recent work in Streptococcus agalactiae showed that the Nramp-type MntH is induced at low pH and facilitates macrophage and acid stress survival [63]. It is thus possible that the transcriptional induction of mntH2 in the urinary tract is attributed to a low pH response. Alternatively, we hypothesized that MntH2 might contribute to Cu detoxification during CAUTI. Specifically, Cu and ceruloplasmin—the major mammalian Cu-binding protein—were recently found to accumulate during bacterial UTI and uropathogenic Escherichia coli (UPEC) has been shown to up-regulate Cu efflux systems during clinical UTI to avoid Cu toxicity [64, 65]. Thus, we tested the ability of parent and ΔmntH2 strains to grow in human urine supplemented with Cu. However, both strains were completely resistant to physiological concentrations of Cu (up to 0.5 μm) and showed identical levels of Cu sensitivity at supra-physiological concentrations (S9 Fig). While in need of further study, it appears that the prominent role of MntH2 in CAUTI could be simply attributed to its higher expression levels in the bladder environment when compared to EfaCBA.
In summary, here we show that maintenance of Mn homeostasis is an essential trait enabling E. faecalis to fully exert its full pathogenic potential. Specifically, simultaneous inactivation of efaCBA and mntH2 renders E. faecalis avirulent in two mammalian infection models while only a triple ΔefaΔmntH1ΔmntH2 mutant strain was attenuated in the G. mellonella invertebrate model. These results expand the knowledge that bacterial Mn transport systems are promising targets for the development of novel antimicrobial therapies, which should be particularly effective to combat E. faecalis infections [66]. Work will soon be underway to uncover the significance of dietary Mn in the pathophysiology of E. faecalis, to elucidate the organ-dependent differential transcriptional regulation of each Mn transporter, and to understand the contribution of Mn-dependent responses to oxidative stress survival in the context of infection.
Bacterial strains and plasmids used in this study are listed in Table 1. All E. faecalis strains were routinely grown overnight at 37°C in BHI supplemented with 150 μM MnSO4. When required, 10 μg ml-1 erythromycin was added to the growth medium for stable maintenance of plasmids in the complemented strains. The primary antibodies used in the study were rabbit anti-Streptococcus group D antigen (anti-E. faecalis lipoteichoic acid) [67] and mouse anti-EbpAFull [47]. Horseradish peroxidase (HRP)-conjugated goat anti-mouse and goat anti-rabbit antisera from KPL and IRDye 680LT goat anti-rabbit from LI-COR Biosciences were used as secondary antibodies.
Deletion of efaCBA, mntH1, and mntH2 from the E. faecalis OG1RF strain was carried out using the pCJK47 markerless genetic exchange system [37]. Briefly, ~ 1 kb PCR products flanking the efaCBA, mntH1, or mntH2 coding sequences were amplified with the primers listed in S1 Table. The amplicons included the first and last residues of the coding DNA to avoid unanticipated polar effects. Cloning of amplicons into the pCJK47 vector, electroporation and conjugation into E. faecalis strains and final isolation of single mutant strains were carried out as previously described [37]. Double mutants were obtained by conjugating the different pCJK constructs accordingly into Δefa, ΔmntH1 or ΔmntH2 single mutants. A triple mutant was obtained by conjugating the pCJK-mntH2 plasmid into the ΔefaΔmntH1 double mutant. All gene deletions were confirmed by PCR sequencing of the insertion site and flanking sequences in single, double and triple mutants. BHI agar was supplemented with 150 μM MnSO4 to enable isolation of the triple mutant strain.
The shuttle vector pTG001, a modified version of the nisin-inducible pMSP3535 plasmid [69] with an optimized RBS and additional restriction cloning sites (a gift from Dr. Anthony Gaca, Harvard Medical School), was used to complement the ΔefaΔmntH1ΔmntH2 triple mutant strain. Briefly, the coding sequence of efaCBA, mntH1, or mntH2 was amplified from OG1RF (efaCBA, mntH1, mntH2) using the primers listed in S1 Table, digested with the appropriate restriction enzymes and ligated into pTG001 digested with compatible restriction enzymes to yield plasmids pTG-efa, pTG-mntH1 and pTG-mntH2. Upon propagation in E. coli DH10B, pTG001 (empty plasmid) and the complementation vectors were electroporated into OG1RF wild-type or ΔefaΔmntH1ΔmntH2 strains using a standard protocol [70] modified such that electroporated cells were immediately recovered in BHI supplemented with 0.4 M sorbitol. Presence of plasmids were confirmed via PCR amplification of the DNA insert region using plasmid-specific primers.
In vitro growth of strains in metal-depleted medium was performed using the chemically-defined FMC medium [71] depleted for Mn, Fe, or Mn and Fe as previously described [6]. When indicated, the Mn and Fe concentrations were depleted from ~ 110 μM (Mn) and ~ 75 μM (Fe) in metal-replete FMC to < 90 nM in Mn- and/or Fe-depleted FMC. Overnight cultures were diluted 1:40 in complete FMC (metal-replete) and grown aerobically at 37°C to an OD600 of 0.25 (early exponential phase). Then, cultures were diluted 1:100 into fresh FMC depleted for Mn, Fe, or Mn and Fe, and cell growth was monitored using the Bioscreen growth reader monitor (Oy Growth Curves AB Ltd.). To assess colony formation on BHI plates, overnight cultures were washed once in sterile PBS containing 0.1 mM EDTA to chelate extracellular divalent cations, followed by a second PBS wash to remove EDTA. Washed cells were diluted 1:100 in PBS, and 5 μl aliquots were spotted on plain BHI agar plates or BHI plates containing 150 μM MnSO4, or 150 μM FeSO4. Plates were incubated overnight at 37°C before growth was recorded. For growth assessment of E. faecalis strains in the presence of 35 mM paraquat, overnight cultures were first diluted 1:40 in BHI and grown to an OD600 of ~ 0.25 prior to inoculation (1:100) into fresh BHI with or without 150 μM MnSO4 and 35 mM paraquat. Growth was monitored using the Bioscreen growth reader monitor over a 24 hours period. Growth in the presence of purified calprotectin and the His (103–105)—Asn calprotectin variant (a gift from Dr. Walter Chazin and Dr. Eric Skaar, University of Vanderbilt) was adapted from previous reports [15, 72, 73]. Briefly, overnight cultures were diluted 1:50 in fresh BHI broth and incubated 1 hour at 37°C. Then, cultures were diluted 1:100 into calprotectin medium, consisting of 38% BHI and 62% CP buffer [20 mM Tris pH 7.5, 100 mM NaCl, 3 mM CaCl2, 5 mM β-mercaptoethanol]. Native calprotectin and its variant unable to bind Mn [15] were added at a sub-inhibitory concentration (120 μg ml-1), and growth at 37°C was recorded at 30 min intervals for up to 24 hours. When indicated, 10 μM MnSO4, ZnSO4, or FeSO4, were added to the medium to test for growth rescue.
Total metal concentration in BHI, human urine and within bacterial cells was determined as previously described by inductively coupled plasma—optical emission spectrometry (ICP-OES) at the University of Florida Institute of Food and Agricultural Sciences Analytical Services Laboratories [6]. Briefly, for quantification of metals in liquids, 18 ml BHI or urine were digested with 2 ml trace-metal grade 35% HNO3 prior to analysis. For quantification of the cellular metal content of bacterial cells, overnight cultures were washed twice with BHI and subsequently used to inoculate fresh BHI (1:20) and incubated statically at 37°C. After reaching an OD600 ~ 0.5, cells were harvested by centrifugation, washed twice in ice-cold PBS supplemented with 0.5 mM EDTA to chelate extracellular divalent cations, and aliquots were collected for metal analysis. Bacterial pellets were resuspended in 1 ml 35% HNO3 and digested at 90°C for 1 hour in a high-density polyethylene scintillation vial (Fisher Scientific). Digested bacteria were diluted 1:10 in reagent-grade H2O prior to ICP-OES metal analysis. Metal composition was quantified using a 5300DV ICP Atomic Emission Spectrometer (Perkin Elmer), and concentrations were determined by comparison with a standard curve. Metal concentrations were normalized to total protein content determined by the bicinchoninic acid (BCA) assay.
Larvae of G. mellonella was used to assess virulence of OG1RF and its derivatives as described elsewhere [74]. Briefly, overnight cultures were washed to reduce Mn carryover. Groups of 15 larvae (200–300 mg in weight) were injected with 5 μl of bacterial inoculum containing ~ 5x105 CFU. Larvae injected with heat-inactivated E. faecalis OG1RF (30 min at 100°C) were used as negative control. After injection, larvae were kept at 37°C and G. mellonella survival was recorded at selected intervals for up to 72 hours. Experiments were performed independently at least six times with similar results.
Survival of strains in pooled human serum and pooled human urine (Lee Biosolutions) was monitored as previously described [6]. Briefly, overnight cultures grown in BHI+Mn were washed in sterile PBS as indicated above to remove excess Mn and inoculated 1:200 in human serum or urine. Cultures were incubated aerobically at 37°C and, at selected time points, aliquots were serially diluted and plated on Mn-supplemented BHI plates for colony-forming unit (CFU) determination. For determination of growth in human urine, overnight cultures were normalized to an OD600 of 1.0 in BHI+Mn. Cells were washed 3 times with 1X PBS followed by 1:1000 dilution into undiluted pooled female urine supplemented with 20 mg ml-1 of BSA and incubated at 37°C. Bacterial growth was monitored by quantifying CFU as described above. When indicated, serum and urine were supplemented with a final concentration of 1 mM FeSO4 (99%, Sigma), 1 mM MnSO4 (99%, Sigma) or with increasing concentrations (0.5 μM, 0.5 mM or 1 mM) of CuSO4 (99%, Acros Organics). Urine was pooled from 3 healthy female donors, clarified by centrifugation, filter-sterilized, and adjusted to pH 6.5 prior to use. All urine samples were collected after obtaining written consent as per the study approval from the Washington University School of Medicine Internal Review Board (approval ID #201207143).
Male and female specific pathogen-free New Zealand White rabbits (2–4 kg; RSI Biotechnology) were utilized in an endocarditis model to assess virulence as previously described [25]. Prior to surgery, the rabbits were sedated and anesthetized with a cocktail of ketamine, xylazine and buprenorphine. A 19-gauge catheter was inserted into the aortic valve by way of the right carotid artery to induce minor damage. Each catheter was trimmed and sutured in place and the incision was closed with staples. Later the same day, duplicate inoculum preparation began by inoculating all strains into BHI for overnight growth in 6% O2. For the ΔefaΔmntH1ΔmntH2 strain only, BHI medium was supplemented with 100 μM Mn. The next morning, cultures were diluted 1:10 in fresh BHI with Mn supplementation and incubated with lids tight at 37°C until an OD600 of ~0.8 was obtained. The cells were then washed twice with 7 ml Chelex-treated (Bio-Rad) PBS containing 0.1mM EDTA. As determined in previous pilot experiments, a portion of the duplicate resuspended cells was removed and diluted to achieve equal strain representation and a total inoculum of ~107 CFU/ml. From each sample, 1 ml was removed, sonicated and plated for enumeration, while 0.5 ml were injected into the peripheral ear vein of each of three rabbits per inoculum. A total of six animals were used for each set of experiments. However, one animal per group had to be euthanized before the set time point and was not included in the final analysis. Forty-eight hours after inoculation, rabbits were euthanized via intravenous injection of Euthasol (Med-Pharmex Inc.). Catheter placement was verified upon necropsy. Harvested cardiac vegetations and spleens were placed into PBS, homogenized and plated on Mn-supplemented BHI agar plates. Resulting colonies were picked and patched onto a new BHI plate supplemented with Mn as appropriate. PCR screens were used to determine the frequency with which each strain was recovered from heart vegetations and spleens. Thus, 50 to 100 colonies per organ and animal were randomly chosen and screened by colony PCR (primers listed in table S1) for their efaCBA and mntH2 status (either wild-type or deleted gene). To verify that the ΔefaΔmntH2 and the ΔefaΔmntH1ΔmntH2 strains are not growth inhibited in vitro when co-cultured with the parent OG1RF and their respective single mutant strains (Δefa or ΔmntH2), overnight cultures were diluted in PBS to an initial inoculum of ~ 2 x 103 CFU of each strain. Then, we co-cultured OG1RF with Δefa and ΔefaΔmntH1ΔmntH2, or with ΔmntH2 and ΔefaΔmntH2 strains in BHI supplemented with 150 μM MnSO4. After 9 hours of incubation, PCR screens were used to determine the frequency with which each individual strain was recovered from the medium as described above.
The mice used in this study were 6-week-old female wild-type C57BL/6Ncr mice purchased from Charles River Laboratories. Mice were subjected to transurethral implantation and inoculated as previously described [48]. Mice were anesthetized by inhalation of isoflurane and implanted with a 5-mm length platinum-cured silicone catheter. When indicated, mice were infected immediately following catheter implantation with 50 μl of ~2 × 107 CFU of bacteria in PBS introduced into the bladder lumen by transurethral inoculation as previously described [48]. To harvest the catheters and organs, mice were euthanized 24 hours post-infection by cervical dislocation after anesthesia inhalation, and bladder, kidneys, spleen and heart were aseptically harvested. Subsequently, the silicone implant was retrieved from the bladder.
Silicone catheters (1 cm, Nalgene 50 silicone tubing, Brand Products) or 96-well polystyrene plates (Grenier CellSTAR) were coated overnight at 4°C with 100 μg ml-1 human fibrinogen free of plasminogen and von Willebrand Factor (Enzyme Research Laboratory). The next day, E. faecalis overnight cultures were diluted to an optical density (OD600) of 0.2 in BHI broth. The diluted cultures were centrifuged, washed three times with 1x PBS, and diluted 1:100 in urine supplemented with 20 mg ml-1 BSA. Bacterial cells were allowed to attach to the fibrinogen-coated silicone catheters or 96-well polystyrene plates for 24 hours at 37°C under static conditions. After bacterial incubation, catheters and microplates were washed with PBS to remove unbound bacteria. Half of the catheters were vortexed for 30 sec, sonicated for 5 min, and vortexed again for 30 sec to retrieve bacteria in the biofilm for CFU quantification. The other half of catheters were used to visualize biofilms. Briefly, catheters were fixed with formalin for 20 min and then washed three times with PBS. Catheters were blocked at 4°C with 5% dry skin milk PBS, followed by three washes with PBS-T. After the washes, plates were incubated for an hour at room temperature with rabbit anti-Streptococcus group D antigen antisera (1:500) in dilution buffer. Plates were washed with PBS-T, incubated with the Odyssey secondary antibody (goat anti-rabbit IRDye 680LT, diluted 1:10,000) for 45 min at room temperature and washed three times with PBS-T. As a final step, plates were scanned for infrared signal using the Odyssey Imaging System (LI-COR Biosciences). Fibrinogen coated-catheters were used as a control for any auto-fluorescence. For assessment of biofilm formation on fibrinogen-coated 96-well polystyrene plates, microplates were stained with 0.5% crystal violet for 10 min at room temperature. Excess dye was removed by rinsing with sterile water and then plates were allowed to dry at room temperature. Biofilms were resuspended with 200 μl of 33% acetic acid and the absorbance at 595 nm was measured on a microplate reader (Molecular Devices). Experiments were performed independently in triplicate per condition and per experiment.
Surface expression of EbpA by E. faecalis OG1RF parent and Mn transporter mutants was determined by ELISA as previously described [75] Bacterial strains were grown for 18 hours in urine supplemented with 20 mg ml-1 of BSA. Then, cells were washed three times with PBS, normalized to an OD600 of 0.5, resuspended in 50 mM carbonate buffer (pH 9.6) containing 0.1% sodium azide and used to coat (100 μl aliquots) Immulon 4HBX microtiter plates overnight at 4°C. The next day, plates were washed three times with PBS-T (PBS containing 0.05% Tween 20) to remove unbound bacteria and blocked for 2 hours with 1.5% BSA–0.1% sodium azide–PBS (BB) followed by three washes in PBS-T. EbpA surface expression was detected using mouse anti-EbpAFull antisera, which was diluted 1:100 in PBS dilution buffer (PBS with 0.05% Tween 20, 0.1% BSA, and 0.5% methyl α-d-mannopyronoside) before serial dilutions were performed. A 100-μl volume was added to the plate, and the reaction mixture was incubated for additional 2 hours. Subsequently, plates were washed three times with PBS-T, incubated for 1 hour with HRP-conjugated goat anti-rabbit antisera (1:2,000), and washed again three times with PBS-T. Detection was performed using a TMB substrate reagent set (BD). The reaction mixtures were incubated for 5 min to allow color to develop, then the reactions were stopped by the addition of 1.0 M sulfuric acid. The absorbance was determined at 450 nm. Titers were defined as the last dilution with an A450 of at least 0.2. As an additional control, rabbit anti-Streptococcus group D antiserum was used to verify that whole cells of all strains were bound to the microtiter plates at similar levels. EbpA expression titers were normalized against the bacterial titers at the same dilution.
The aortic valve homogenate RNAs used for these experiments were collected and analyzed in another study from rabbits infected with E. faecalis OG1RF in an experimental model of IE [76]. Gene expression was studied in RNA samples collected from animals infected for two days (early infection, n = 5 rabbits) or three to four days (advanced infection, n = 8 rabbits). For transcript quantifications during CAUTI, OG1RF cells were recovered from bladders of three infected mice, pooled and immediately placed in RNAlater solution (in triplicate, total n = 9 animals). RNA extraction, reverse transcription and real-time PCR were carried out following standard protocols [6, 77].
Data were analyzed using GraphPad Prism 6.0 software unless otherwise stated. Differences in cellular metal uptake, final growth in Mn-depleted FMC or human urine, the percentage of strains recovered from the rabbit IE experimental model, and the log10-transformed CFU values recovered from in vitro catheter biofilms were determined via ordinary one-way ANOVA followed by post-test comparisons. Log10-transformed CFU values from serum survival and calprotectin growth experiments were analyzed via a two-way ANOVA followed by comparison post-tests. While differences in the G. mellonella killing rate of mutant strains was assessed with the Mantel-Cox log-rank test, final 72 hour survival of larvae were similarly compared via ordinary one-way ANOVA with Dunnett’s multiple comparison post-test. Of note, two outliers (one for a ΔmntH1 replicate, another for a ΔefaΔmntH1ΔmntH2 replicate) were identified using the ROUT method (Q = 1%) and removed from the final analysis. For CAUTI experiments, biofilm assays and EbpA expression, data from multiple experiments were pooled and Two-tailed Mann-Whitney U tests were performed. To determine statistical significance in fold-change transcription of selected genes, fold-change values for each gene were plotted as geometric mean with the corresponding 95% confidence interval (error bars). A line at y = 1 denotes equal transcripts in inoculum control vs in vivo condition. 95% confidence intervals that did not cross y = 1 were significantly different (* p<0.05) from the inoculum control.
Urine samples for growth curves were collected after obtaining informed written consent for all subjects enrolled in the study as per the study approval from the Washington University School of Medicine Internal Review Board (approval ID #201207143). Populations generally classified as vulnerable, including children under the age of 18, were not enrolled in the study. Only subjects 18 years of age or older at the time of consent were eligible for the study. No group or persons were excluded from the study due to race, ethnicity or gender. Case subjects were enrolled solely based on the eligibility criteria.
Animal procedures for rabbit IE were approved by the Institutional Animal Care and Use Committee of the Virginia Commonwealth University as part of protocol number AM10030 as well as the University of Minnesota Institutional Animal Care and Use Committee as part of protocol number 0910A73332. The Washington University Animal Studies Committee approved all mouse infections and procedures as part of protocol number 20150226. All animal care was consistent with the Guide for the Care and Use of Laboratory Animals from the National Research Council and the USDA Animal Care Resource Guide.
|
10.1371/journal.ppat.1002898 | Roles of ATM and ATR-Mediated DNA Damage Responses during Lytic BK Polyomavirus Infection | BK polyomavirus (BKPyV) is an emerging pathogen whose reactivation causes severe disease in transplant patients. Unfortunately, there is no specific anti-BKPyV treatment available, and host cell components that affect the infection outcome are not well characterized. In this report, we examined the relationship between BKPyV productive infection and the activation of the cellular DNA damage response (DDR) in natural host cells. Our results showed that both the ataxia-telangiectasia mutated (ATM)- and ATM and Rad-3-related (ATR)-mediated DDR were activated during BKPyV infection, accompanied by the accumulation of polyploid cells. We assessed the involvement of ATM and ATR during infection using small interfering RNA (siRNA) knockdowns. ATM knockdown did not significantly affect viral gene expression, but reduced BKPyV DNA replication and infectious progeny production. ATR knockdown had a slightly more dramatic effect on viral T antigen (TAg) and its modified forms, DNA replication, and progeny production. ATM and ATR double knockdown had an additive effect on DNA replication and resulted in a severe reduction in viral titer. While ATM mainly led to the activation of pChk2 and ATR was primarily responsible for the activation of pChk1, knockdown of all three major phosphatidylinositol 3-kinase-like kinases (ATM, ATR, and DNA-PKcs) did not abolish the activation of γH2AX during BKPyV infection. Finally, in the absence of ATM or ATR, BKPyV infection caused severe DNA damage and aberrant TAg staining patterns. These results indicate that induction of the DDR by BKPyV is critical for productive infection, and that one of the functions of the DDR is to minimize the DNA damage which is generated during BKPyV infection.
| BK polyomavirus (BKPyV) is a human pathogen that establishes a persistent sub-clinical infection in healthy humans. When patients are immunosuppressed, particularly in kidney and bone marrow transplantation, the virus can reactivate and result in severe disease. BKPyV-related disease has risen due to the use of newer immunosuppressive regimens and an increase in the number of transplants performed each year. We are interested in understanding the interactions between BKPyV and host cell components or pathways, with the aim of developing more BKPyV-specific antiviral treatment options. In this study we characterized the relationship between BKPyV infection and the cellular DNA damage response (DDR), a signaling cascade that is initiated by cells to repair damaged DNA molecules. Our study indicated that BKPyV activates and hijacks the DDR to facilitate its infection and that various components of the DDR may play distinct roles during this process. These data suggest that the DDR may provide a potential host target to control BKPyV reactivation in transplant recipients.
| BK polyomavirus (BKPyV) was first isolated in 1971 from a renal transplant patient [1] and has gained much interest in the past two decades due to its disease prevalence in immunocompromised patients [2]. Infection with BKPyV is ubiquitous in healthy individuals but does not lead to any known clinical disease. Under immunosuppressed conditions, especially in renal transplant and bone marrow transplant recipients, the virus can reactivate from a persistent state to lytic infection, which results in severe disease including polyomavirus-associated nephropathy (PVAN) and hemorrhagic cystitis (HC), respectively [2]. Unfortunately, there is currently no FDA-approved, specific anti-BKPyV drug available for treating these diseases. The common approach to control BKPyV reactivation is palliative care for HC patients, or combining immunosuppression reduction with drugs that inhibit viral DNA replication for PVAN, although there are often conflicting outcomes with these treatment options [3].
Much of the knowledge about the polyomavirus lytic life cycle comes from research performed on Simian Virus 40 (SV40). The viral genome, a circular double-stranded DNA molecule, is delivered into the nucleus. Following nuclear entry, early proteins including the large T antigen (TAg) are expressed. TAg sets up the host environment for viral DNA replication by inducing cells into S phase and, at the same time, inhibiting the p53-dependent apoptotic pathway [4]. Initiation of viral DNA replication requires the concerted efforts of TAg, replication protein A (RPA), DNA polymerase alpha-primase (Pol-prim), and topoisomerase I [5]. Newly replicated viral DNA is encapsidated by the capsid proteins VP1, VP2, and VP3, and this is followed by viral egress and cell lysis, thus completing the life cycle. Although SV40 is well-studied, there are differences between it and BKPyV. There is still much that is unknown with regard to the interaction between host nuclear components and viral factors during BKPyV replication. Identification and characterization of these interactions is extremely important, since it may reveal novel anti-viral therapeutic targets.
The DNA damage response (DDR) is emerging as a cellular process that is targeted by a number of DNA and RNA viruses [6], [7]. The DDR involves signaling cascades that are initiated when cells experience various types of DNA damage, and it coordinates many cellular processes including cell cycle arrest, chromatin modification, and DNA repair to allow cells to repair the damaged DNA [8], [9]. This response is largely orchestrated by two major phosphatidylinositol 3-kinase-like kinases (PI3KKs): ataxia telangiectasia mutated (ATM), and ATM and Rad3-related (ATR) kinase. ATM mainly responds to double-stranded breaks (DSBs) resulting from conditions such as ionizing irradiation (IR). The Mre11-Rad50-NBSI (MRN) protein complex serves as a sensor for DSBs, and is crucial for the activation of ATM upon DSB damage [10], [11]. ATR, on the other hand, is activated by single-stranded DNA lesions and is important for resolving replication stress from conditions such as stalled replication forks and ultraviolet (UV) light [8]. Both kinases, when activated, can phosphorylate numerous downstream targets that are involved in DNA repair and cell cycle arrest, including Chk1, Chk2, and a histone variant H2AX [8], [12]. The phosphorylated H2AX (serine 139, referred to as γH2AX) is considered a hallmark of the DDR, and it is crucial for recruiting and maintaining downstream mediator and repair proteins to sites of damage [13], [14].
There is accumulating evidence suggesting that the DDR is closely linked to polyomavirus infections. Mouse polyomavirus (MPyV) infection increases phosphorylated ATM (pATM), and using either an ATM inhibitor or ATM-deficient cells, it has been demonstrated that an ATM-mediated DDR is required for MPyV replication [15]. ATM phosphorylates SV40 TAg, and knockdown of ATM decreases the level of TAg-pS120 and viral DNA synthesis [16]. The ATM-mediated DDR has also been reported to be required for SV40 infection, and SV40 infection is found to lead to the proteasome-dependent degradation of MRN complex [17], [18]. Another human polyomavirus, JC polyomavirus (JCPyV), also activates the ATM DDR. It is suggested that DDR activation induces G2 arrest to promote JCPyV replication [19]. The contribution of the ATR-mediated DDR to polyomavirus infection is less clear. Although knockdown of ATR during JCPyV infection results in a decrease in TAg levels, ATR knockdown does not seem to affect SV40 DNA replication [18]. This lack of effect on viral replication has been attributed to incomplete knockdown of ATR and kinase redundancy [18]. Using a cell line that expresses a dominant negative form of ATR, another group showed that ATR is required for the activation of ATR-Δp53 signaling pathway during SV40 infection [20]. The activation of this pathway is believed to lead to intra-S checkpoint activation and promote TAg- Pol-prim complex formation [20]. Finally, a recent microarray analysis has shown that several genes that are involved in DNA damage repair are also upregulated with BKPyV infection, indicating that the DDR may be important during BKPyV infection [21].
In this study we examined the roles of both the ATM and ATR-mediated DDR during lytic BKPyV infection in a primary human renal proximal tubule epithelial (RPTE) cell culture model [22]. Our results showed that BKPyV activates both branches of the DDR. Using small interfering RNA (siRNA) knockdowns, we demonstrated that ATM and ATR each contribute to DDR activation caused by BKPyV infection. Our data also clearly suggested the importance of ATR during lytic BKPyV infection, and more importantly, that both proteins function additively to ensure efficient viral DNA replication and synergistically to affect production of viral progeny. In the absence of either kinase, severe DNA damage accumulated during BKPyV infection, indicating that part of the roles that ATM and ATR play is to repair DNA damage caused by BKPyV infection.
To investigate the relationship between BKPyV and the DDR, we began our studies with the ATM-mediated branch of DDR, because it has been implicated in many polyomavirus infections [15]–[18]. To examine whether BKPyV infection activates an ATM-mediated DDR, we harvested total protein lysates from mock- and BKPyV-infected RPTE cells over a 3-day time course. The proteins were immunoblotted for markers that are indicative of ATM-mediated DDR activation (Figure 1A). The gradual increase in the viral early protein TAg allowed us to monitor the progression of infection. In our Western blots for TAg, we routinely observe two bands using a monoclonal anti-TAg antibody, which are labeled as TAg and TAg* (filled arrowheads). We think these are two forms of TAg because TAg is known to undergo multiple post-translational modifications [23]–[25]. While the total level of ATM remained constant throughout the course of infection, the level of ATM-pS1981 increased dramatically by 2 days post infection (dpi). Concomitant with the increase of ATM-pS1981, there was an induction of NBSI-pS343, γH2AX, Chk2-pT68, RPA32-pS4/8, and p53-pS15, all of which are downstream targets of active ATM-pS1981 [8], [12]. The induction of both ATM-pS1981 and γH2AX was similar to that seen when RPTE cells were treated with IR (data not shown), suggesting that our detection of DDR markers is specific. There was an increase in total p53 levels following BKPyV infection, consistent with previous reports of TAg stabilizing p53 [26]–[28]. In contrast to SV40 [17], BKPyV did not cause degradation of MRN components. In addition, we detected that BKPyV infection resulted in unique γH2AX and Mre11 staining patterns (Figure S1). There was a marked increase in both bright foci and pan-nuclear staining of γH2AX, as well as bright nuclear foci of Mre11. Some, but not all of these foci co-localized with small TAg foci. Together, these results demonstrated that the ATM signaling pathway is activated upon BKPyV infection in RPTE cells.
We next determined whether BKPyV infection led to cell cycle dysregulation (Figure 1B), which is commonly seen with DDR activation. Flow cytometry analysis of cells stained with propidium iodide (PI) showed that starting from 2 dpi, in cells that were infected with BKPyV, there was a gradual increase of a cell population with >G2 DNA content (Figure 1B and 1C). The accumulation of these polyploid cells persisted until 6 dpi (data not shown), whereas in mock-infected cells, most cells were in the G0/G1 phase throughout the time course. These data indicated that multiple rounds of DNA replication can occur within a single cell cycle in BKPyV-infected cells, and are consistent with the activation of the DDR.
To address the functional roles of the ATM-mediated DDR in productive BKPyV infection, we knocked down ATM using siRNA and assessed the effects of the knockdown on BKPyV infection. RPTE cells were first transfected with siRNAs that targeted ATM, followed by infection with BKPyV when the knockdown of ATM was achieved at 3 days post transfection (dpt) (Figure 2). By Western blotting, ∼90% of the total ATM was knocked down, which was also confirmed by immunoblotting against ATM-pS1981 (Figure 2A and data not shown). The knockdown lasted throughout the course of infection and was not affected by BKPyV replication (Figure 2A). Moreover, knockdown of ATM did not affect the morphology or viability of RPTE cells compared to cells that did not receive siRNA or cells that were transfected with non-targeting siRNA control (data not shown). We first compared the level of TAg in ATM knockdown cells to control cells using immunoblotting. The expression of TAg was similar among all the cells over a 3-day time course, suggesting that ATM was not required for viral early gene expression (Figure 2A). We then examined whether ATM knockdown resulted in a defect in viral DNA replication or infectious progeny production. Low molecular weight DNA was extracted from the samples and real-time PCR was employed to measure the viral DNA load (Figure 2B). The knockdown led to an ∼60% decrease of viral DNA compared to control cells by 3 dpi (Figure 2B). Consistently, we also observed an approximately 50% decrease of viral infectious progeny in ATM knockdown cells (Figure 2C). These data suggested that ATM contributes to optimal BKPyV DNA replication and viral growth.
To ask whether ATM was responsible for the activation of the DDR we observed during BKPyV infection, total proteins from mock- and BKPyV-infected ATM knockdown cells were immunoblotted for Chk2-pT68 and γH2AX and compared to control cells (Figure 2D). ATM knockdown partially abolished the induction of Chk2-pT68 caused by BKPyV infection (Figure 2D), consistent with ATM being the main kinase that phosphorylates Chk2 during the DDR [8], . γH2AX, however, was still induced in BKPyV-infected, ATM knockdown cells (Figure 2D). These data suggested that ATM may not be the sole contributor to the BKPyV-induced DDR.
The fact that ATM knockdown did not completely eliminate the BKPyV-activated DDR suggested that there are other DDR factors involved. Both ATM and ATR share a number of substrates and therefore it is possible that the DDR activation we observed during BKPyV infection was, in part, a result of ATR activation. To determine whether the ATR-mediated DDR is also activated during BKPyV infection, we immunoblotted proteins from mock- and BKPyV-infected RPTE cells for ATR, Chk1-pS317, and Chk1-pS296 (Figure 3A). ATR levels remained relatively unchanged during infection; however, both Chk1-pS317 and Chk1-pS296 increased starting at 2 dpi. Chk1-pS317 phosphorylation is a direct result of ATR activation upon DNA damage, which is followed by Chk1 autophosphorylation on S296 [29]. The strong induction of both Chk1-pS317 and Chk1-pS296 clearly suggested that an ATR-mediated DDR is also activated by BKPyV infection.
Next, we used siRNAs to either singly knock down ATR or doubly knock down both ATM and ATR, and examined the effects of these knockdowns on the DDR activation during BKPyV infection. Quantitative Western blotting showed that the knockdowns were ∼90% complete, and no significant cellular viability changes were noticed in knockdown cells (Figure 3B and data not shown). Having established that, we went on to determine whether ATR single knockdown or ATM+ATR double knockdown could eliminate the BKPyV-induced DDR in RPTE cells. Total proteins from mock- and BKPyV-infected control or knockdown cells were immunoblotted for Chk1-pS317, Chk1-pS296, Chk2-pT68, and γH2AX (Figure 3B). ATM knockdown reduced Chk2-pT68 in infected cells, whereas ATR single or ATM+ATR double knockdown almost completely abolished the induction of both Chk1-pS317 and Chk1-pS296. ATR knockdown, however, did not significantly alter the induction of Chk2-pT68. These data suggested that ATM was mainly responsible for the activation of pChk2 upon BKPyV infection, while ATR was more important in inducing pChk1. γH2AX induction was still evident even in ATM+ATR double knockdown cells (Figure 3B), suggesting that ATM and ATR may not be the only kinases contributing to the phosphorylation of this DDR marker.
We then investigated the role that ATR plays alone and in conjunction with ATM during productive BKPyV infection. We first measured viral gene expression in these knockdown cells. At 2 dpi, we did not detect a significant difference in the level of TAg between any knockdown cells and control cells (Figure 4A and S2), but a third TAg band appeared in the ATR and double knockdown cells (Figure 4A, open arrowhead). This band was no longer present at 3 dpi. In contrast to the ATM knockdown alone, at 3 dpi the ATR knockdown displayed a more marked difference in TAg and TAg*. There was about a 50% decrease of TAg in ATR knockdown cells, and an ∼80% decrease of TAg*. The ATM and ATR double knockdown sample had a pattern similar to the ATR knockdown alone (Figure 4B and 4C). None of the knockdowns, however, had a major effect on late gene VP1 expression (Figure 4D and S2). To ensure that we did not saturate the cells with large amounts of BKPyV, we repeated these experiments with a low multiplicity of infection (MOI) (0.01 IU/cell vs. 0.5 IU/cell). Quantitative Western blotting showed that there was no significant difference in either TAg or VP1 levels among all the knockdown and control cells (Figure S3A and S3B) at 2 dpi, suggesting that the lack of difference in TAg and VP1 levels at 2 dpi is not due to compensation from a high MOI infection.
TAg is required for the initiation of viral DNA synthesis and it provides a conducive host replication environment by inactivating the retinoblastoma susceptibility family proteins and p53 [4]. We therefore determined whether a change in TAg in ATR knockdown cells was accompanied by a defect in viral DNA replication. Using quantitative PCR, we found that ATR or ATM knockdown caused a similar decrease in viral DNA synthesis at 2 dpi (Figure 4E). The double knockdown cells, however, exhibited a more severe defect in viral DNA levels at both high and low MOIs (Figure 4E and Figure S3C), suggesting that ATM and ATR function additively to ensure full viral DNA replication. Finally, we measured infectious viral progeny production under these knockdown conditions (Figure 4F). ATR single knockdown led to a slightly larger decrease in viral titer compared to ATM single knockdown. The double knockdown cells displayed a dramatic decrease (∼90%) compared to the control cells.
To confirm the importance of the ATR/Chk1 pathway in BKPyV infection, we treated RPTE cells with the Chk1 inhibitor UCN-01 at 1 dpi (at which time the viral genome is delivered into the nucleus) [30]. The presence of UCN-01 slightly reduced TAg and VP1 levels (Figure 5A). Treatment with UCN-01 also resulted in a partial decrease in viral DNA replication and a severe reduction in viral titer (Figure 5B and 5C), consistent with ATR/Chk1 pathway being required for BKPyV productive infection. We have also tested the effect of the ATM inhibitor KU55933; however, this drug seemed to have some off-target effects in RPTE cells and therefore we could not draw conclusions from these experiments (data not shown).
It has previously been shown that another PI3KK that is involved in nonhomologous end-joining DNA repair, DNA-dependent protein kinase (DNA-PKcs), is able to phosphorylate H2AX independent of ATM and ATR [31], [32]. Additionally, for certain viruses such as adeno-associated virus (AAV), the induction of the DDR in infected cells is mostly mediated through DNA-PKcs [33]. We therefore tested the effect of DNA-PKcs single knockdown and ATM/ATR/DNA-PKcs triple knockdown on BKPyV infection and the induction of γH2AX (Figure 6). The DNA-PKcs-targeting siRNAs efficiently knocked down DNA-PKcs, but they also reduced ATM (Figure 6A). The single knockdown slightly increased TAg, TAg*, and VP1, with a concomitant ∼2 fold increase in viral DNA and viral titer (Figure 6C and 6E). The impact of the ATM/ATR/DNA-PKcs knockdown is similar to that of the ATM/ATR double knockdown. There is an ∼60% drop in viral DNA and a more striking difference in viral titer (Figure 6D and 6F). Surprisingly, with all three kinases absent, we could still detect a strong induction of γH2AX in BKPyV-infected cells (Figure 6B). These data suggested that additional cellular kinases may be involved in the activation of DDR during BKPyV infection.
To further dissect the roles that ATM and ATR play during BKPyV infection, we next examined the localization of TAg in knockdown and control cells (Figure 7). In all the mock-infected cells, the nuclei appeared normal judged by DAPI staining (data not shown). With BKPyV infection, however, we observed some abnormal DAPI staining in ATM/ATR single or double knockdown cells (Figure 7A). In these cells, we observed that the nuclei sometimes appeared smaller or fragmented (open arrowheads, Figure 7A, and Figure 7B), which are very similar to micronuclei that arise when chromosomes are broken or damaged [34]. At the same time, we also observed aberrant TAg staining patterns in the knockdown cells (Figure 7A). While TAg appeared mostly nuclear in control cells, in ATM and ATR knockdown cells we often observed cells that had diffuse TAg staining, sometimes throughout the whole cytoplasm (Figure 7A, filled arrowheads, and Figure 7C). In addition, in cells that received ATR siRNA, some of the cells showed fragmented TAg staining patterns (Figure 7A, arrows, and Figure 7C), similar to the fragmented DAPI staining.
To confirm that the observed aberrant DAPI and TAg staining is indeed associated with DNA damage, we performed metaphase chromosome analysis in control and knockdown cells, with or without BKPyV infections (Figure 8). In all the mock-infected cells (including all the knockdown cells), or BKPyV-infected cells with no knockdowns, most chromosome spreads were normal. In ATM or ATR knockdown cells, there were a very high percentage of metaphases showing a shattered phenotype, with ATR single knockdown showing the most severe defect (Figure 8A and 8C). These numbers might even be underestimated considering that we could not distinguish between uninfected and infected cells in the metaphase spreads. Shattered chromosomes were never observed in uninfected cells. Among the metaphases that were not shattered, the average number of chromosome breaks per cell was also higher in BKPyV-infected, ATM/ATR knockdown cells (Figure 8B). These data suggested that ATM, and to a greater extent, ATR, contribute to repairing DNA damage that is triggered by BKPyV infection.
In this report we characterized the relationship between the ATM- and ATR-mediated DDR and lytic BKPyV infection. Our results indicated that both branches of the DDR were activated by BKPyV; moreover, ATM and ATR functioned in parallel and contributed to the activation of individual DDR pathways (Figure 9A). In particular, ATM was mainly responsible for activating pChk2, whereas ATR was more important in activating pChk1 and also might be important for regulating TAg and its modified forms. Our results showed that both ATM and ATR were required to achieve maximal viral DNA replication and infectious progeny production. This is the first such report for any polyomavirus family member to our knowledge. Knocking down each individual kinase resulted in a partial defect in viral replication and the double knockdown had a more dramatic effect. When comparing ATM and ATR double knockdown cells to control cells, the degree of decrease in viral titer was consistently greater than the decrease in viral DNA, suggesting that there may be some post-DNA replication regulation that involves both ATM and ATR function. For example, ATR may be involved in alleviating the DNA replication stress generated during the resolution of the two daughter circular BKPyV chromosomes, and therefore ATR knockdown caused a more marked defect in viral progeny production than in DNA replication. Our results also showed that in the absence of ATM or ATR, severe chromosome damage accumulated upon BKPyV infection. These data point out that another possible function of ATM and ATR during BKPyV infection is to help repair the DNA damage caused by BKPyV (Figure 9B).
The exact role of ATM during polyomavirus early gene expression remains unclear, but it may be different depending on the specific virus and cell type. For example, one report showed that ATM knockdown in African green monkey CV1 cells reduced the level of SV40 TAg-pS120, but it did not have much effect on the total level of TAg [16]. In another report using BSC40 monkey kidney cells, the ATM inhibitor KU55933 reduced total SV40 TAg levels [17]. Our ATM knockdown data suggest that ATM is not required for TAg expression during BKPyV infection. Consistent with this, BKPyV TAg expression was similar in wild-type and ATM-knockout mouse embryonic fibroblast cell lines (data not shown). Unfortunately we cannot study the entire BKPyV life cycle in mouse cells due to a block to DNA replication [35].
Our data demonstrate the importance of ATM and ATR during BKPyV DNA replication, DDR activation, and progeny production. It is possible that ATM, ATR and downstream effectors directly participate in replication-related events. For example, ATM has been reported to phosphorylate SV40 TAg in vivo and therefore is required for optimal SV40 replication [16]. In our ATR single knockdown or ATM+ATR double knockdown cells, we observed a decrease in a modified form of TAg (TAg*) at 3 dpi, and the appearance of a third band of TAg at 2 dpi (Figure 4A). Preliminary experiments indicated that phosphatase treatment did not affect the level of TAg or TAg* (data not shown), but this does not definitively rule out the possibility that TAg* may represent a phosphorylated form of TAg. Additionally, proteins involved in homologous recombination or the Fanconi anemia pathway (required for repair of stalled replication forks) have been found to be necessary for SV40 replication [18]. It is possible that DDR proteins that are downstream of ATM and ATR directly contribute to BKPyV DNA replication. Alternatively, instead of being directly involved in viral replication, ATM and ATR may participate indirectly in BKPyV infection by affecting cell cycle status. For example, ATM and ATR activation leads to G2 arrest in JCPyV-infected human neuroblastoma cells and oligodendrocytes [19]. It is hypothesized that G2 arrest contributes to JCPyV replication by maintaining the cellular replication machinery and preventing mitosis [19]. Our cell cycle analyses also showed that BKPyV infection of RPTE cells results in the accumulation of polyploid cells, consistent with the idea that mitosis is inhibited to allow for maximal DNA replication.
A number of DDR proteins including γH2AX and Mre11 were re-localized into small TAg-containing foci during BKPyV infection. Similar co-localization between DDR proteins and TAg foci is seen during SV40 infection, although the MRN proteins are eventually degraded during SV40 infection [17], [18]. It has been hypothesized that these foci represent sites of viral DNA replication because they also contain proteins that are required for viral replication including RPA and Pol-prim, but not Pol-prim-associated host replication factor [17]. In addition to the small, bright foci of γH2AX, there was also an increase in pan-nuclear staining of γH2AX in BKPyV-infected cells. Although the functional significance of this increase is not clear, it has previously also been reported in other DNA virus infections such as adenovirus and AAV [33], [36], [37]. It is thought that this pan-nuclear staining increase may represent modification of histones on cellular chromatin and that the modification can be stimulated by viral replication.
Intriguingly, the induction of γH2AX by BKPyV still occurred in ATM, ATR, and DNA-PKcs triple knockdown cells, suggesting that there might be other cellular kinase(s) responsible for phosphorylating this molecule. Alternatively, it is possible that BKPyV infection alters a cellular phosphatase activity that, together with residual PI3KK activity due to incomplete knockdown, leads to an increase in the steady state level of γH2AX. One phosphatase candidate is PP2A, which has been demonstrated to dephosphorylate γH2AX in an ATM-, ATR-, and DNA-PKcs-independent manner [38]. Polyomavirus small T antigen is well known for its interaction with PP2A and its ability to inhibit PP2A enzymatic activity [39]. It will be interesting to determine the mechanism of γH2AX induction and its functional significance during BKPyV infection in the future.
What triggers the activation of both the ATM- and ATR-mediated DDR during BKPyV infection requires more careful examination. It is possible that a viral protein(s) alone is able to achieve the induction. For example, expression of SV40 TAg without a viral replication origin in normal human BJ/tert fibroblasts induces both an ATM- and ATR-mediated DDR, and this induction is dependent on the interaction of TAg with Bub1, a mitotic spindle checkpoint kinase [40]. Moreover, polyomavirus TAg alone is able to induce cellular DNA damage as judged by comet assays and cytogenetic analyses [18], [41]–[44]. For JCPyV, the ability of TAg to associate with cellular DNA is important for TAg induction of G2 cell cycle arrest [19]. It is also possible that either incoming or replicating viral genomes serve as the trigger for the DDR. For example, both wild-type and UV-inactivated AAV2, but not recombinant AAV2 vectors, are capable of inducing a DDR, suggesting that it is the viral DNA sequence, but not the viral capsid, that is responsible for the activation of DDR [37].
In conclusion, our results demonstrate the unique activation of various DDR components upon BKPyV infection and the essential roles of both ATM and ATR for viral replication and growth. The study of BKPyV infection and the DDR not only reveals novel knowledge about the cellular pathways with which the virus needs to interact in order to complete the lytic life cycle, but may also have important clinical implications for BKPyV reactivation and its related disease. For example, BKPyV reactivation is a severe problem in bone marrow transplant patients, who might have experienced radiation as part of the preparative regimen. Thus, research focusing on BKPyV and DDR may shed light on the analysis of the causality of BKPyV reactivation in these patients.
RPTE cells (Lonza) were maintained in renal epithelial cell growth medium (REGM) as previously described [45]. All cells were grown at 37°C with 5% CO2 in a humidified incubator.
BKPyV (Dunlop) was grown, purified, and titered using an infectious unit (IU) assay as previously described [46]. For infections, RPTE cells were pre-chilled for 15 min at 4°C. The cells were then exposed to purified BKPyV diluted in REGM at the indicated MOIs and incubated for 1 h at 4°C. Infection was initiated by replacing the viral inoculum with pre-warmed REGM and transferring the cells to 37°C. Total cell proteins and viral lysates were harvested as previously described [45].
UCN-01 (Sigma) was reconstituted according to the manufacturer's recommendations. The drug was added at 1 dpi at 100 nM and was left on for the time of the experiment. A cell metabolism WST-1 assay (Roche) was used to ensure that the drug treatment did not cause significant cytotoxic effects (data not shown).
Total cell proteins were harvested, quantified, and immunoblotted as previously described [46]. For quantitative blots using the Odyssey Infrared Imaging System, the membrane was processed according to the manufacturer's instructions (LI-COR). The membrane was scanned using the Odyssey Infrared Imaging system, and the relevant bands were quantified using the Odyssey software. See Table S1 for a list of antibodies and concentrations used in this study.
Mock or BKPyV-infected cells were trypsinized, resuspended in PBS, and fixed with 100% cold ethanol. DNA was labeled with 50 µg/ml PI+100 µg/ml RNAse in PBS at room temperature for 30 min. Samples were analyzed with a BD FACSCalibur flow cytometer and the cell cycle data were modeled using ModFit LT software.
At the indicated times post infection, RPTE cells were fixed and immunostained as previously described [47] with antibodies listed in Table S1. For standard fluorescence microscopy, samples were examined using an Olympus BX41 microscope with a Plan 40×/0.65 objective and processed using the Olympus DP manager software. For laser-scanning confocal microscopy, all images were obtained using a Zeiss LSM 510 confocal microscope with a 63×/1.2 objective and 1 µm optical section. Images were analyzed and processed using LSM image browser (Zeiss).
siRNA ON-TARGET plus SMART pools were purchased from Thermo Scientific Dharmacon: Non-targeting (D-001810-10-05); ATM (L-003201-00-0005); ATR (L-003202-00-0005); and DNA-PK (L-005030-00-0005). siRNAs were resuspended in 1× siRNA buffer (Dharmacon) to 20 µM stocks. RPTE cells were reverse transfected with indicated siRNAs using Lipofectamine RNAiMAX transfection reagent (Invitrogen) according to manufacturer's instructions. siRNAs were diluted in REGM without serum or antibiotics and mixed with Lipofectamine RNAiMAX (2–4 µl per well of a 12 well plate, or 33 µl per T75 flask). The complexes were allowed to form at room temperature for 15 min, followed by the addition of RPTE cells (60,000 cells per well, or 660,000 cells per flask). The optimal final concentrations of siRNA were determined empirically. ATM and DNA-PK siRNAs were used at 10 nM, and ATR siRNA was used at 20 nM. For double or triple knockdowns, non-targeting siRNAs were added to ensure that the total concentrations of siRNAs in all the samples were the same. Transfection complexes were washed out 1 dpt and replaced with REGM containing serum and antibiotics [45]. The cells were infected with BKPyV at 3 dpt as described above. For some batches of RPTE cells, siRNA transfection resulted in uneven cell death among different wells. Under these circumstances, cells were trypsinized, counted, and re-plated prior to infection to ensure that same number of cells were present in all samples for infection.
To quantify the viral DNA load in cells, low molecular weight DNA was isolated using a modified Hirt protocol [47], real-time PCR reactions were performed, and data were analyzed as previously described [47].
Cells were harvested for chromosome preparations with a modified protocol [48]. Briefly, cells were treated with colcemid (50 ng/ml) for 1 h followed by an 18 min incubation in 0.8% sodium citrate at 37°C and multiple changes of Carnoy's fixative (3∶1 methanol∶acetic acid). Cells were dropped onto slides and slides were baked overnight at 55°C before staining with Giemsa (Sigma). Metaphase chromosomes were observed using an Olympus BX41 microscope with a Plan 100×/1.25 oil objective or a Nikon OPTIPHOT microscope with a Plan 100×/1.40 oil objective.
|
10.1371/journal.pbio.2005388 | Centromeric signaling proteins boost G1 cyclin degradation and modulate cell size in budding yeast | Cell size scales with ploidy in a great range of eukaryotes, but the underlying mechanisms remain unknown. Using various orthogonal single-cell approaches, we show that cell size increases linearly with centromere (CEN) copy number in budding yeast. This effect is due to a G1 delay mediated by increased degradation of Cln3, the most upstream G1 cyclin acting at Start, and specific centromeric signaling proteins, namely Mad3 and Bub3. Mad3 binds both Cln3 and Cdc4, the adaptor component of the Skp1/Cul1/F-box (SCF) complex that targets Cln3 for degradation, these interactions being essential for the CEN-dosage dependent effects on cell size. Our results reveal a pathway that modulates cell size as a function of CEN number, and we speculate that, in cooperation with other CEN-independent mechanisms, it could assist the cell to attain efficient mass/ploidy ratios.
| It has been known for a long time that cells increase their size with the amount of DNA across the whole evolutionary scale. Moreover, within a given species and differentiation status, cells increase their size in an exquisitely proportional manner to the number of chromosomes. This correlation has been observed in fungi, plants, and animals, and it is assumed to have important implications in development and physiology. However, the molecular mechanisms that set cell size as a function of ploidy are still a mystery. Here, we uncover a pathway that links the number of chromosomes to the molecular network that triggers cell cycle entry in budding yeast. More specifically, we have found that an excess number of centromeres (CENs) increases degradation of Cln3, a G1 cyclin that plays a key role at Start, thus delaying entry into the cell cycle and, as cell growth is not affected, increasing cell size. Our results point to the existence of dedicated mechanisms operating during entry into the cell cycle to adjust cell size to ploidy.
| Most cells under unaltered conditions of growth are able to maintain their size within a strict range, and a current view sustains that the cell cycle and cell growth machineries should be interconnected by specific molecular mechanisms ensuring cell size homeostasis [1–5]. Budding yeast cells control their size mainly at Start [6,7], when a G1 cyclin, Cln3, acts as the most upstream activator [8]. Cyclin Cln3 forms a complex with Cdc28, the cell cycle Cdk in budding yeast, which phosphorylates the transcriptional inhibitor Whi5 and induces a transcriptional wave in circa 200 genes to trigger cell cycle entry [9]. Cln3 modulates cell volume at Start in a precise, dose-dependent manner [10–12], which suggests that mechanisms regulating its levels or activity likely play important roles in cell size determination. In this regard, Cln3 is present at low and nearly constant amounts throughout G1 [8, but see 13,14], and its nuclear levels are restrained by retention at the ER [15,16] and ubiquitin-mediated degradation by the proteasome [17,18].
It has long been known that cell size increases linearly with ploidy in fungi [19–21], plants [22], and animals [23,24], a function that is maintained across the enormous DNA content variation among eukaryotes [25] and has been used to infer ploidy in the fossil record [26]. Although ploidy has direct implications in cell growth and development, the underlying mechanisms that set cell size as a function of ploidy remain elusive [27]. Here, we describe a pathway linking the centromere (CEN) to the Start network in budding yeast. Briefly, we have found that an excess number of CENs increases degradation of Cln3 in the nucleus by a mechanism that involves physical and functional interactions between Cdc4, the specific F-box protein that targets Cln3 to SCF for ubiquitination, and Mad3, a centromeric signaling protein.
In control experiments in which the size of yeast cells was carefully measured, we had previously observed that the presence of an empty yeast centromeric plasmid (YCp) produced a slightly larger volume at budding. Interestingly, this effect was exacerbated by increasing the number of empty centromeric vectors with different auxotrophic markers, suggesting that G1 length could be modulated by a genetic determinant present in these extrachromosomal DNA molecules. After ruling out possible effects due to plasmid-borne auxotrophic markers (S1A Fig), we analyzed newborn daughter cells during cell cycle entry in time-lapse experiments and found that, while initial volume was very similar, YCp caused a strong delay in G1 and a larger cell size at budding (S1B and S1C Fig). To assess the effects of YCp copy number at the single-cell level, we inserted a TEF1p-driven transcription unit expressing green fluorescent protein (GFP) in YCp vectors and mCherry in chromosome 5 and used different approaches to increase the number of centromeric sequences in the cell, some of them in a conditional manner (Fig 1A). We first analyzed cells in the simplest scenario, i.e., containing three GFP-expressing YCp vectors. Budding volume of control cells displayed a large variability [21,28] but steadily increased with the GFP/mCherry ratios (Fig 1B, see S1 Data for a detailed statistical analysis). Intriguingly, the observed trend was compatible with the doubling in budding volume displayed by diploid cells. A yeast episomal plasmid (YEp), which is present at much higher copy numbers, did not significantly alter budding volume (Fig 1C), thus pointing to the autonomous-replicating sequence (ARS) or the CEN as the YCp-specific genetic determinants modulating cell size at budding. To discern between these possibilities, we used a yeast CEN placed immediately downstream from the inducible GAL1 promoter as a conditional CEN that, by growing cells under conditions that activate (galactose) or repress (glucose) transcription from the GAL1 promoter, can be switched off or on, respectively [29]. We introduced this conditional CEN (Fig 1A) into three different YCp vectors and observed that, under permissive conditions, cell volume at budding increased with a much steeper slope compared to unmodified YCp (Fig 1D). To rule out possible topological effects due to the circular conformation of YCp vectors, we used a linear yeast artificial chromosome (YAC) containing a conditional CEN to obtain a wide range of copy numbers per cell. As shown in Fig 1E, budding volume correlated with YAC copy number in a similar manner to that obtained with YCp vectors. Moreover, as this effect was also observed with a circular YAC derivative (S2 Fig), we were able to rule out possible additional effects of telomeric sequences. Finally, introducing conditional CENs into chromosomes 4 and 7 caused a significant increase in the budding volume of newborn daughter cells obtained by differential gradient centrifugation when allowed to enter the cell cycle under permissive conditions (Fig 1F). As previously described [30,31], high copies of centromeric vectors caused a short mitotic delay (S3A Fig) that depended on the spindle-assembly checkpoint (SAC) [32,33]. However, this delay was much shorter than that observed during cell-cycle entry (S1C Fig), suggesting that elevated CEN copies have a greater impact in G1. Accordingly, additional CEN sequences caused a small but significant increase in the proportion of cells in G1 phase in asynchronous cultures (S3B Fig). Overall, these data indicate that CEN number modulates G1 length in daughter cells and regulates their size at budding.
Budding yeast cells mainly determine their size at Start [4]. Thus, we reasoned that signals originating from the CEN could target specific components of the Start network. YCp vectors clearly increased budding volume in cells deficient in Whi5 (Fig 2A and 2B), thus ruling out this transcriptional repressor of the G1/S regulon [34,35]. By contrast, cells lacking Cln3, the most upstream G1 cyclin [8,10,36] acting at Start, did not increase their size further, indicating that Cln3 is essential in the mechanisms that allow centromeric signals to modulate cell size. Overexpression of wild-type Cln3, which causes a strong nuclear accumulation of this G1 cyclin [15,37], also suppressed the YCp-mediated effects on budding size. However, a Cln3–1 hyperstable mutant that also reaches high levels but lacks the C-terminal nuclear-localization signal (NLS) that is essential for nuclear import of Cln3 [38] was as sensitive as wild type to the presence of YCp (Fig 2A and 2B). Supporting the notion that centromeric-dependent effects take place in the nucleus, a different hyperstable Cln3ΔPEST mutant protein that retains the C-terminal NLS and strongly accumulates in the nucleus [15,37] fully suppressed YCp-mediated effects in cell size. Together, our data point to the idea that centromeric-dependent signals target, directly or indirectly, the yeast G1 cyclin in the nucleus.
A high-throughput two-hybrid analysis in budding yeast [39] had revealed an interaction between the Cln3 cyclin and Mad3, a component of the kinetochore-signaling network involved in the SAC [40,41]. Thus, we tested whether centromeric signaling proteins could have a role in modulating budding size as a function of YCp copy number (Fig 3A). The budding size of cells lacking either Mad3 or Bub3 was absolutely refractory to increasing copies of YCp while, contrarily, kinase Bub1 did not have any effect. These results suggest that Mad3/Bub3 inhibit Cln3 function in a Bub1-independent manner, thus defining a mechanism different to that executing the SAC.
Newborn daughter haploid and diploid cells lacking Mad3 displayed a smaller volume at budding compared to wild type (Fig 3B). However, size reduction was only moderate compared to the difference between haploid and diploid wild type, and mad3 cells reduced their size normally from diploid to haploid status. These data suggest that either Mad3 is not required per se in the sensing mechanism or cells must have additional or backup mechanisms to adjust cell size to ploidy (see below). Although Mad3 could not be overexpressed to much higher levels compared to the endogenous copy (S4A and S4B Fig), budding size displayed a clear increase under these mild overexpression conditions (Fig 3B). Considered together, these data reinforce the notion of an inhibitory role for Mad3 in cell cycle entry and cell size determination at budding.
Mad proteins use different but complementary mechanisms to modulate degradation of Cdc20 targets by the anaphase-promoting complex (APC/Cdc20), including mitotic cyclins [32,33], which suggests that the Mad3-dependent effects of YCp vectors on budding volume could be mediated by degradation of Cln3. Supporting this idea, Skp1 is a highly expressed centromeric protein that is also present in SCF, the E3 ubiquitin ligase required to degrade Cln3 [18]. We found that the presence of YCp vectors strongly increased the degradation rate of Cln3 in promoter shut-off experiments, and more importantly, this effect required Mad3 (Fig 4A, 4B, S5A and S5B Fig). To support these findings further, we used a partially hyperstable and hypoactive mutant (Cln3–11A) fused to mCitrine that has no gross effects on cell cycle progression [42] but allows detection of this cyclin in G1 cells by fluorescence microscopy to monitor Cln3 degradation specifically in the nucleus. Although cells expressing mCitrine–Cln3–11A displayed an increased volume at budding when compared to wild-type cells, the presence of YCp vectors caused a similar relative increment in their budding size (S6A Fig), which validated its use. Notably, by measuring mCitrine–Cln3–11A levels in G1 cells after cycloheximide addition, we found that the presence of YCp vectors also increased the degradation rate of this G1 cyclin in the nucleus in a Mad3-dependent manner (Fig 4C and S6B Fig). Accordingly, mCitrine–Cln3–11A steady-state levels were strongly decreased by YCp in the nucleus of G1 cells within the same volume range (Fig 4D). Since Cln3 is rate-limiting for triggering Start and setting the critical size at budding, these results would explain why the presence of YCp vectors causes a larger cell size. Next we analyzed the interaction between Mad3 and Cln3 by affinity purification and found that they yielded relative coprecipitation efficiencies similar to Cln3 and Cdc4ΔFbox (Fig 4E), the adaptor protein that recruits Cln3 to SCF in the nucleus [18]. Interestingly, we were able to detect an interaction between Cdc4 and Mad3 (Fig 4F), which suggests that Mad3 is present with Cdc4 in SCF complexes. Mad3 contains a GLEBS domain that is known to interact with Bub3 and, as a likely consequence, with Skp1 [43], and we found that Mad3 lacking the GLEBS domain does not efficiently interact with either Cdc4 or Cln3 (Fig 4E and 4F). Finally, modulation of budding size as a function of YCp copy number was strongly dampened by deleterious SCF mutations or deletion of the Mad3 GLEBS domain (Fig 4G and S7 Fig), supporting the essential role of a SCF–Cdc4/Mad3 complex in boosting Cln3 degradation to modulate cell size at budding as a function of CEN copy number.
In summary, we have uncovered a pathway that links centromeric signaling proteins to G1 cyclin stability and, hence, cell size determination in budding yeast (Fig 5). SCF–Cdc4 is estimated to be at low levels in the nucleus of yeast cells, and we envisage that Mad3, which is present at much higher levels, could act as a co-adaptor to increase the affinity of Cdc4 for Cln3. Strikingly, Cdc4 and Cdc20 display a high degree of similarity (34.2%) and contain WD40 segments that are used to interact with client proteins. However, the interaction of Mad3 would have different outcomes: (1) prevent Cdc20 from binding its targets in metaphase and (2) acting as an adaptor bridging Cln3 to Cdc4 in G1.
Mad3 is present at rather constant levels throughout the cell cycle [41], and Spc105, a scaffold protein involved in Mad3 activation at kinetochores by SAC [32], is already present in CENs in G1 [44]. Thus, by mechanisms different from those operating the SAC, Mad3 could be specifically activated in G1 at the kinetochore and sustain degradation of Cln3 at levels proportional to the number of CENs during G1 progression. Alternatively, we would like to speculate that the pathway uncovered here could belong to a Mad3-dependent checkpoint triggered by the excess of a kinetochore component that, being synthesized as a function of cell mass, would act as ploidy-mass reporter. Since Mad3-deficient or overexpressing cells do not display strong alterations in cell size, Mad3 would have a role as an effector of the checkpoint, not as sensor. Furthermore, the uncovered mechanism could be used to ensure that CENs congregate at the spindle-pole body (SPB) [45] before cell-cycle entry in budding yeast. While structural determinants of centromeric DNA are strikingly different in yeast and mammalian cells, kinetochore structural and signaling proteins are very well conserved. For this reason, we envisage that the mechanism operating in budding yeast could also exist across the evolutionary scale.
Previously proposed mechanisms to adjust cell size to ploidy [3] have not received sufficient experimental support. Although Whi5 is expressed at levels that depend on ploidy [42], diploid cells lacking one WHI5 copy are larger than haploid wild-type cells. On the other hand, introduction of additional Cln3-targeted promoters delays cell-cycle entry and increases cell size at budding [46]. However, it remains unclear whether titration of Cln3 by genome duplication is sufficient to produce a diploid cell size. We propose that, most likely with the contribution of these mechanisms, CEN-dependent degradation of Cln3 may play a pivotal role in scaling size with ploidy, a universal property of cells.
Cells were grown in SC medium with 2% glucose at 30 °C unless stated otherwise. Late G1-arrested cells were obtained by treating exponential cultures at OD600 = 0.5 with 5 μg/ml α factor for 105 min at 30 °C. Conditional CENGALp CENs were inhibited by addition of 2% galactose to culture medium. Cycloheximide was added at 25 μg/ml to inhibit protein synthesis. Cln3–3HA half-life was analyzed in tet-promoter shut-off experiments by adding tetracycline to 1 μg/ml [47]. MAD3 overexpression was attained by inducing a GAL1p-MAD3 construct with 1 mM estradiol in cells expressing the Gal4–hER–VP16 transactivator [48]. Yeast parental strains and methods used for chromosomal gene transplacement and PCR-based directed mutagenesis have been described [37]. Centromeric plasmids and yeast artificial chromosomes were obtained by multiple-fragment recombination [49] in yeast cells. Conditional CENGALp CENs (GAL10p–CEN4) were inserted in chromosomes 4 and 7 by CRISPR/Cas9-driven recombination [50]. The ΔGLEBS mutant of Mad3 lacked the C-terminal 155 amino acids. Cln3–1 [11] and Cln3ΔPEST [15] are both hyperstable mutant proteins, but only Cln3ΔPEST retains the C-terminal NLS [38]. The Cln3–11A mutant protein is a hypoactive and hyperstable cyclin that contains 11 amino acid substitutions (R108A, T420A, S449A, T455A, S462A, S464A, S468A, T478A, S514A, T517A, T520A) [42]. The Cdc4ΔFbox protein has been already described [18].
Yeast cells were analyzed by time-lapse microscopy in 35-mm glass-bottom culture dishes (GWST-3522, WillCo) essentially as described [28] using a fully-motorized Leica AF7000 microscope. Time-lapse images were analyzed with the aid of BudJ, an ImageJ (Wayne Rasband, NIH) plugin that can be obtained from www.ibmb.csic.es\home\maldea to obtain cell dimensions and fluorescence levels in cellular and nuclear compartments [28]. Briefly, cell boundaries are detected as pixels markedly darker compared to both the surrounding background and the cell interior. Once outliers have been removed, an ellipse is fitted to the obtained boundary pixel array, and major and minor axes are used to calculate the cell volume assuming a prolate as shape. The same cell is followed through consecutive time-lapse images by using the center of the ellipse as seed point to obtain radial profiles in the following image.
GST-tagged proteins were affinity purified with glutathione beads (GE Healthcare) from cell extracts as described [37]. Immunoblot analysis [51] was performed with antibodies against HA (12CA5, Roche), FLAG (M2, Sigma), myc (9E10, Sigma), and GST (polyclonal, Millipore).
Small daughter cells were isolated from Ficoll gradients as described [52]. DNA content distributions were obtained by Fluorescence Activated Cell Sorting [51].
Pairwise comparisons were performed with non-parametric tests. Specifically, median cell volumes at budding were compared with a Mann–Whitney U test. On the other hand, correlation of cell volume at budding with GFP/mCherry ratios was analyzed with a Spearman rank test. For pairwise analysis, data were subject to bootstrap resampling (N = 100), and the resulting median slopes were compared by a Mann–Whitney U test. For both median and regression analysis, the resulting p values are shown in the corresponding figure panels.
|
10.1371/journal.ppat.1005030 | BRCA1 Regulates IFI16 Mediated Nuclear Innate Sensing of Herpes Viral DNA and Subsequent Induction of the Innate Inflammasome and Interferon-β Responses | The innate immune system pattern recognition receptors (PRR) are the first line of host defenses recognizing the various pathogen- or danger-associated molecular patterns and eliciting defenses by regulating the production of pro-inflammatory cytokines such as IL-1β, IL-18 or interferon β (IFN-β). NOD-like receptors (NLRs) and AIM2-like receptors (ALRs) are cytoplasmic inflammasome sensors of foreign molecules, including DNA. IFI16, a sequence-independent nuclear innate sensor ALR, recognizes episomal dsDNA genomes of herpes viruses such as KSHV, EBV, and HSV-1 in the infected cell nuclei, forms an inflammasome complex with ASC and procaspase1, and relocates into the cytoplasm leading into Caspase-1 and IL-1β generation. IFI16 also induces IFN-β during HSV-1 infection via the cytoplasmic STING-TBK1-IRF3 pathway. Thus far, whether IFI16 recognizes foreign DNA directly or utilizes other host protein(s) is unknown. Here, we demonstrate that BRCA1, a DNA damage repair sensor and transcription regulator, is in complex with IFI16 in the host cell nucleus, and their association increases in the presence of nuclear viral genomes during de novo KSHV, EBV and HSV-1 infection, and in latent KSHV or EBV infection, but not by DNA damage responses (DDR) induced by bleomycin and vaccinia virus cytoplasmic dsDNA. BRCA1 is a constituent of the triggered IFI16-inflammasome and is translocated into the cytoplasm after genome recognition along with the IFI16-inflammasome. The absence of BRCA1 abrogated IFI16-viral genome association, inflammasome assembly, IFI16 cytoplasmic localization, and Caspase-1 and IL-1β production. The absence of BRCA1 also abolished the cytoplasmic IFI16-STING interaction, downstream IRF3 phosphorylation, nuclear translocation of pIRF3 and IFN-β production during de novo KSHV and HSV-1 infection. These findings highlight that BRCA1 plays a hitherto unidentified innate immunomodulatory role by facilitating nuclear foreign DNA sensing by IFI16, subsequent assembly and cytoplasmic distribution of IFI16-inflammasomes leading into IL-1β formation and the induction of IFN-β via cytoplasmic signaling through IFI16-STING, TBK1 and IRF3.
| Invasion of a host cell by pathogens, including viruses, is sensed by pattern-recognition receptors resulting in the elicitation of the host innate defenses such as the formation of multi-protein inflammasome complexes, inflammatory IL-1β and IL-18 cytokine production and interferon-β production via the cytoplasmic STING molecule. We have shown that nuclear episomal viral DNA genomes of herpes viruses (KSHV, EBV and HSV-1) are sensed by the nuclear resident IFI16 protein, resulting in the formation of the IFI16-ASC-procaspase-1 inflammasome complex. Here, we show that BRCA1 promotes viral DNA sensing by IFI16 in the nucleus and is a constituent of the triggered IFI16-ASC-procaspase-1 inflammasome. IFI16 and BRCA1 are in complex in the nucleus and their association increases in the presence of KSHV, EBV or HSV-1 genomes, but not by the DNA damage response or vaccinia virus cytoplasmic dsDNA. The absence of BRCA1 results in abrogated IFI16-genome association, IFI16 cytoplasmic translocation, IL-1β production, IFI16 interaction with STING, IRF3 phosphorylation, pIRF3 nuclear translocation, and IFN-β induction. Taken together, these results demonstrate a crucial and novel role of BRCA1 in the innate sensing of viral DNA and subsequent induction of the inflammasome and interferon-β responses.
| Sensing of microbial nucleic acids by pattern-recognition receptors (PRRs) is a crucial step for an effective innate immune response [1]. The best established function of PRRs like NLRPs (NOD-like receptors with PYRIN (PYD) domain) and ALRs (absent in melanoma 2 [AIM2]-like receptors) is their ability to sense pathogens and other danger signals. This leads into the formation of a multiprotein inflammasome complex consisting of a sensor protein, adaptor protein ASC (apoptosis-associated speck-like protein containing CARD) and procaspase-1 resulting in active Caspase-1 generation which cleaves the proforms of interleukin-1β (IL-1β), IL-18, and IL-33 cytokines.
Our studies have demonstrated that IFI16 (interferon inducible protein 16), a resident nuclear ALR protein in a variety of cells, functions as a sensor and detects nuclear replicating herpesvirus genomes such as Kaposi's sarcoma-associated herpesvirus (KSHV), Epstein-Barr virus (EBV), and herpes simplex virus type-1 (HSV-1) leading to IFI16-inflammasome formation [2, 3, 4, 5]. De novo KSHV infection of primary human microvascular dermal endothelial (HMVEC-d) cells and HSV-1 infection of primary human foreskin fibroblast (HFF) cells induces IFI16-ASC-procaspase-1 inflammasome formation in the nucleus and its redistribution to the cytoplasm [2, 5]. KSHV latency in endothelial and B cells also constitutively activates the IFI16-inflammasome and cytoplasmic relocalization, and IFI16 colocalizes with the KSHV and HSV-1 genomes in the nuclei of infected cells [2, 3]. EBV latency in B and epithelial cells also constitutively activates the IFI16 inflammasome and cytoplasmic relocalization, and IFI16 colocalizes with the EBV genomes in the nucleus [4].
IFI16 has also been shown to interact with STING (stimulator of interferon genes) leading to phosphorylation and nuclear translocation of IRF3 via the IFI16-STING-TBK signaling axis, resulting in IFN-β production during HSV-1 infection [6, 7]. The role of IFI16 as a silencing factor for the HSV-1 genome has also been reported [8]. We have also recently demonstrated that, independent of its innate immune response, IFI16 inhibited HSV-1 replication by repressing viral gene expression via its binding to the transcription start sites of viral genes, reducing the association of transcription factors to these sites and by promoting global histone modifications on the viral genome [9].
The linear epigenetically naïve virion-associated herpesviral dsDNA genome circularizes after entry into the nucleus, associates with histones and nucleosome proteins leading to epigenetic control [10, 11, 12]. The host cell DNA damage response (DDR) is a signal cascade event that includes the phosphorylation of repair mediators (H2AX, BRCA1, 53BP1, and Mdc1) and effectors of the checkpoint responses (CHK1 and CHK2). The DDR also recognizes exogenous genomes of nuclear DNA viruses which manipulate the DDR for their advantage [13]. Entry of KSHV DNA into the nuclei of endothelial cells during de novo infection induces an immediate DDR response of ATM kinase, H2AX and BRCA1 (breast cancer tumor suppressor protein) activation [14].
IFI16, believed to be a part of the large BRCA1-associated genome surveillance (BASC) DDR complex, interacts with BRCA1 and is implicated in BRCA1-mediated apoptosis and inflammation signaling [15]. IFI16 contains two DNA binding HIN domains, a transcriptional regulatory domain, an APIN/PAAD domain associated with the IFN response and an ASC binding PYD domain that also binds BRCA1 [16]. In a non-nuclear artificial system, IFI16 has been shown to bind to superhelical plasmid DNA and cruciform DNA [17]. A study suggested that IFI16 recognizes DNA in a non-sequence specific manner by electrostatic attraction between its positively charged HIN domain residues and the sugar-phosphate backbone of dsDNA [1]. However, questions such as whether pathogen DNA is recognized by IFI16 directly or in association with other host proteins and how IFI16 differentiates host vs. pathogen DNA remain unknown.
Here, we demonstrate that IFI16 is associated with BRCA1 in the nuclei of uninfected cells and that interaction increased during de novo KSHV and HSV-1 infection as well as during latent KSHV and EBV infection. Our studies show that BRCA1 is an essential component of the IFI16-inflammasome complex. In the absence of BRCA1, KSHV and HSV-1 genome recognition by IFI16, inflammasome complex formation and cytoplasmic localization, cleavage of caspase-1 and IL-1β, cytoplasmic association of IFI16 with STING, phosphorylation of IRF3, nuclear localization of pIRF3, and IFN-β production were abrogated. Collectively, these studies demonstrate that BRCA1 is a positive regulator of the foreign DNA sensing capability of IFI16 and a stabilizer of the IFI16 inflammasome complex and interferon responses.
IFI16 forms a functional IFI16-ASC-procaspase-1 inflammasome early during de novo KSHV infection of primary HMVEC-d cells as well as in latently infected cells, and translocates to the cytoplasm [2, 3]. We have also observed the phosphorylation of ATM, H2AX, CHK2 and BRCA1, the key early DDR components, as soon as KSHV DNA enters the infected cell nuclei at 30 min post-infection (p.i.) [14]. Since IFI16 is suggested to be a part of the DDR [18, 19], we hypothesized that IFI16 forms complexes with different proteins to mediate different functions, and that one or more of these IFI16 complexes recognizes the KSHV DNA to induce inflammasomes.
To test this hypothesis, we utilized de novo KSHV infection that is well-characterized in our earlier studies [2, 3, 14] in which infected cells are identified by a variety of methods such as i) entry of KSHV into the cytoplasm measured by immunofluorescence assay (IFA) for viral envelope and capsid proteins or 5-bromo-2’-deoxyuridine (BrdU) or 5-ethynyl-2’-deoxyuridine (EdU) labeled KSHV genome, and ii) entry of viral DNA into the nucleus measured by IFA for BrdU or EdU and nuclear expression of viral latency associated LANA-1 protein by IFA [2, 3, 14]. In addition, relocalization of nuclear IFI16 into the cytoplasm is also considered as an indicator of infection [2, 3]. When HMVEC-d cells were infected with KSHV containing BrdU genome (30 DNA copies/cell) and immunostained with anti-BrdU antibodies (Table 1), several BrdU labeled viral particles (Fig 1A, green spots) were detected both in the cytoplasm as well as in the infected cell nucleus at 15 min (0.25h) p.i. (Fig 1A, white and yellow arrows, respectively). Similarly, when HFF cells were infected with KSHV containing EdU labeled viral genome (30 DNA copies/cell), we observed a gradual increase in nuclear entry of viral DNA from the cytoplasm during the observed 0.25 h, 0.5 h and 4 h p.i. (Fig 1B, arrows). When these cells were stained for IFI16, we observed the colocalization of IFI16 with the viral genomes in the nucleus (Fig 1C, panels 3 and 4). In addition, IFI16 was observed only in the cytoplasm of cells with nuclear-EdU KSHV genome (Fig 1C, white arrows) and not in the uninfected HFF cells. These observations also support our assertion that relocalization of nuclear IFI16 into the cytoplasm is an indication of KSHV infection. We utilized similar concentrations of labeled and unlabeled virus in all our experiments.
To determine whether IFI16 interacts with members of the DDR under physiological conditions, extracts from uninfected HMVEC-d cells or cells infected with KSHV for 4 and 24 h were immunoprecipitated (IP-ed) with anti-IFI16 antibody and western blotted for BRCA1, IFI16, CHK2 and H2AX. Compared to uninfected cells, we observed increased association of IFI16 with BRCA1 (~2.5 and ~4-fold at 4 and 24 h p.i., respectively) in the infected cells (Fig 1D, panel 1, lanes 1–4). In contrast, the interaction of IFI16 with CHK2 and H2AX proteins was not detected (Fig 1D, panels 3 and 4, lanes 1–4). To confirm these observations, the same extracts were IP-ed with anti-BRCA1 antibody and Western blotted for IFI16, BRCA1, CHK2 and H2AX. In these reactions we also observed an increased association of BRCA1 with IFI16 in infected cells (~ 3.1 and ~ 4.8 fold at 4 and 24 h p.i., respectively) compared with the uninfected cells (Fig 1E, panel 2, lanes 1–4). In contrast, we observed that BRCA1 interacted with CHK2 and H2AX similarly in both uninfected and infected cells without additional infection induced association (Fig 1E, panels 3 and 4, lanes 1–4). There were no overall changes in the total BRCA1, IFI16 and CHK2 levels except for the increase in H2AX levels as shown before [14] upon KSHV infection (Fig 1F, lanes 1–4). IP using control IgG showed no interaction with BRCA1, IFI16, CHK2 and H2AX (Fig 1Ga) with similar input levels of those proteins except increased H2AX in infected cell extracts (Fig 1Gb) as observed before [14] demonstrating the specificity of the reactions. These observations suggested that multiple complexes involving IFI16 and DDR proteins may be present under physiological conditions. However, IFI16 interacts with BRCA1 and not with CHK2 and H2AX proteins, and infection increases the BRCA1-IFI16 association.
The similar total protein level of cellular BRCA1 and IFI16 in both uninfected and virus infected cells as observed in the WB of input in Fig 1F and Fig 1Gb as well as in IP reactions shown in Fig 1D, panel 2 and Fig 1E, panel 1 imply that there are no overall changes in the expression of these proteins as a result of KSHV infection. However, our co-IP and reverse co-IP results (Fig 1D, panel 1 and Fig 1E, panel 2) show that compared to uninfected cells, increased co-IP of BRCA1 and IFI16 occur in the infected cells demonstrating that the increased association between IFI16 and BRCA1 is an infection induced phenomenon.
To further assess that the observed increased interactions between IFI16 and BRCA1 are specifically due to viral infection in HMVEC-d cells, we studied the interactions between IFI16 and BRCA1 as well as interactions of H2AX and CHK2 with IFI16 and BRCA1 during bleomycin (10 mU/ml for 4 h) induced DDR and during KSHV infection. Under these conditions, increased levels of nuclear IFA spots representing phosphorylated H2AX (γH2AX), a well-established hallmark of DDR, were observed (Fig 2A, white arrows).
We performed a proximity ligation assay (PLA) to study the interactions between IFI16 and DDR proteins. PLA can detect an endogenous individual protein or interactions of two proteins and their localization. It relies on the principle that if two epitopes/proteins are within the proximity of 40 nm or below, the PLA oligo probes linked to two secondary antibodies bound to primary antibody-antigen complexes will be amplified to give a PLA signal visualized as a fluorescent dot (Fig 2B and 2C). PLA of untreated and bleomycin treated cells was performed using only one primary antibody in combination with anti-rabbit MINUS probe and anti-mouse PLUS probe. The lack of detection of any PLA signals under all experimental conditions validated the specificity of the antibodies and observed reactions (S1 Fig).
We performed PLA to assess the respective protein-protein interactions using the following combinations of primary antibodies: IFI16+BRCA1, BRCA1+CHK2, BRCA1+γH2AX, IFI16+ASC, IFI16+CHK2, and IFI16+γH2AX (Fig 2D, S1 Table). ASC, a non-DDR protein, was included as a negative control. Bleomycin induced DDR led to increased nuclear interactions between BRCA1-CHK2 and BRCA1-γH2AX (Fig 2D, 2nd row, 2nd and 3rd blocks, white arrows) and no such interaction was observed between IFI16 and ASC (Fig 2D, 2nd row, 4th block). Interestingly, we observed similar levels of interactions between IFI16 and BRCA1 in both untreated and bleomycin treated cells (Fig 2D, 1st and 2nd rows, 1st block, white arrows). Furthermore, no interactions were observed between IFI16 and CHK2 or between IFI16 and γH2AX under any conditions (Fig 2D, 1st and 2nd rows, 5th and 6th blocks). In contrast, during KSHV infection, we observed significant increased nuclear and cytoplasmic interactions between IFI16 and BRCA1 (Fig 2D bar graph) as well as increased nuclear and cytoplasmic interactions between IFI16 and ASC (Fig 2D, 3rd row, 1st and 4th blocks). In addition, we also observed only a limited association of BRCA1 with CHK2 and γH2AX (Fig 2D, 3rd row, 2nd and 3rd blocks) and no overall interaction of IFI16 with CHK2 and γH2AX during infection (Fig 2D, 3rd row, 5th and 6th blocks; S1 Table).
These observations suggested that: a) IFI16 interacts only with BRCA1 and not with H2AX and CHK2 under physiological conditions, b) the increase in IFI16-BRCA1 interaction observed in the infected cells was due to KSHV which is distinct from the complexes formed by BRCA1, H2AX and CHK2 proteins during DDR (S1 Table), and c) the BRCA1 and IFI16 complex may have roles during KSHV infection.
In response to KSHV infection, IFI16 interacts with the adaptor ASC, which recruits procaspase-1 to form an active inflammasome complex that gets relocated to the cytoplasm [2]. Since we detected increased association between BRCA1 and IFI16 in KSHV infected cells and observed increased IFI16-BRCA1 in the cytoplasm of infected cells (Fig 2D, 3rd row, 1st block), we next determined whether BRCA1 relocalizes to the cytoplasm in a way similar to that of IFI16 during de novo KSHV infection [2]. Although BRCA1 was observed in the nucleus and in the cytoplasm of both uninfected and infected cells, BRCA1 levels in the cytoplasm of infected cells increased during the observed periods (0.5 h, 4 h and 24 h) of infection similar to that of IFI16 (Fig 3A, panels 1 and 2, lanes 1 to 8). Forward and reverse co-IP experiments with the same uninfected and KSHV infected HMVEC-d nuclear lysates showed increased nuclear association between BRCA1 and IFI16 during the observed 0.5 h, 4 h and 24 h p.i. (Fig 3B). No such association was observed with IgG control (Fig 3B).
To determine whether BRCA1 was in complex with IFI16 in the cytoplasm of infected cells, cytoplasmic fractions were IP-ed with anti-BRCA1 antibodies and Western blotted for IFI16 (Fig 3C, panel 1). IFI16 was detected in the BRCA1 immunoprecipitates only in the cytoplasmic fractions of infected cells. In addition, we observed that the IFI16-BRCA1 interactions increased significantly with the duration of infection (Fig 3C, panel 1, lanes 3 to 5 vs lane 2). This observation was further confirmed by performing reverse-IP with anti-IFI16 antibodies and western blotting for BRCA1 (Fig 3C, panel 2, lanes 3 to 5 vs. lane 2).
Next, we determined whether the cytoplasmic BRCA1 was part of the active IFI16 inflammasome complex that gets relocalized to the cytoplasm during KSHV infection. Cytoplasmic fractions were IP-ed with anti-ASC, anti-IFI16, anti-BRCA1 and anti-procaspase-1 antibodies and Western blotted for IFI16, BRCA1 and procaspase-1 (Fig 3C, panels 3 to 9). Infection induced IFI16-inflammasomes were shown by the increased detection of IFI16 and cleaved Caspase-1 in the cytoplasm. The purity of the cytoplasmic preparation was confirmed by the absence of TBP and the presence of tubulin (Fig 3C, bottom input panels). Cytoplasmic relocalization of the active IFI16 inflammasome complex during KSHV infection was observed by the progressively increased interactions between ASC, IFI16 and procaspase-1 (Fig 3C, panels 3, 5, 7 and 9, lanes 3 to 5 vs lane 2). We also observed the presence of increasing amounts of BRCA1 in ASC and procaspase-1 immunoprecipitates during the course of infection (Fig 3C, panels 4 and 8, lanes 3 to 5 vs. lane 2). These observations for the first time identified the previously unknown association of BRCA1 with ASC and procaspase-1. Immunoprecipitation of procaspase-1 by anti-BRCA1 antibodies (Fig 3C, panel 6, lanes 3 to 5 vs. lane 2) suggested that BRCA1 relocated to the cytoplasm during KSHV infection as a component of the active IFI16 inflammasome complex.
Fig 3A showing KSHV infection induced cytoplasmic distribution of a fraction of nuclear resident BRCA1 and IFI16 coupled with the observed increase in co-IP of BRCA1 and IFI16 in the nuclear as well as the cytoplasmic fractions during virus infection (Fig 3B and 3C) suggest that BRCA1-IFI16 association in the nucleus increase perhaps to sense the viral genome which is followed by the BRCA1 aided cytoplasmic translocation of IFI16.
To further confirm the association of IFI16, ASC and procaspase-1 with BRCA1, we performed PLA in uninfected HMVEC-d cells and cells infected with KSHV for 24 h using anti-IFI16, anti-BRCA1, anti-ASC and anti-procaspase-1 antibodies. Specificity of the assay was confirmed by using secondary antibodies linked to probes alone (Fig 4B) and also by using only one primary antibody plus both secondary antibodies linked to probes (S2 Fig, A to E). The lack of detection of any signal following PLA in all cases confirmed the specificity of all the antibodies tested (Fig 4B and S2 Fig, A to E). When we used rabbit and mouse anti-IFI16 antibodies against different IFI16 epitopes in PLA for detecting cellular localization of IFI16 during infection, we detected IFI16 in the nucleus of uninfected cells (Fig 4A, UI, red arrows). In contrast, as we have demonstrated previously [3], at 24 h p.i., IFI16 was detected in both the nucleus and the cytoplasm of infected cells (Fig 4A, KSHV 24 h, red and yellow arrows).
To assess the associations with inflammasome related proteins, next, we performed PLA in uninfected and KSHV infected HMVEC-d cells at 24 h p.i. using various combinations of anti-IFI16 and anti-BRCA1 antibodies. We observed that, compared to uninfected cells, in addition to nuclear IFI16-BRCA1 interactions (Fig 4C, UI versus KSHV 24 h, red arrows), a substantial amount of BRCA1 and IFI16 were detected in the cytoplasm of infected cells (Fig 4C, yellow arrows and bar graph). PLA performed to determine the interactions of BRCA1, IFI16 and ASC in infected and uninfected HMVEC-d cells (Fig 4D to 4G) showed increased interactions between IFI16 and ASC (Fig 4D, red and yellow arrows), BRCA1 and ASC (Fig 4E, red and yellow arrows), BRCA1 and procaspase-1 (Fig 4F, red and yellow arrows) and IFI16 and procaspase-1 (Fig 4G, red and yellow arrows) in the cytoplasm of infected cells at 24 h p.i. compared to the uninfected cells as indicated by the detection of a significantly higher number of PLA spots (respective bar graphs). However, comparison between the number of BRCA1 and IFI16 PLA spots as well as IFI16 and ASC PLA spots detected in the infected cell nucleus and cytoplasm were not significant (Fig 4C and 4D, respective bar graphs) which could be due to complex formation as a result of continuous and dynamic viral genome sensing events and/or formation of different molecular complexes with distinct functionality. In contrast, PLA analysis showed that there were significant increases in the association of IFI16 with Caspase-1 as well as BRCA1 with ASC and Caspase-1 in the cytoplasm of infected cells compared to the nucleus of infected cells (Fig 4E–4G, respective bar graphs) which probably represent the active inflammasome complex in the cytoplasm.
These observations confirmed that BRCA1 interacts with components of the IFI16 inflammasome complex and relocalizes to the cytoplasm during de novo KSHV infection.
Although BRCA1 lacks a PYD domain to interact directly with the PYD domain of ASC, the similarities of the BRCA1 association with ASC to that of IFI16 with ASC prompted us to investigate whether the BRCA1-ASC interaction requires the presence of IFI16. Therefore, we analyzed the effect of IFI16 knockdown on the ASC-BRCA1 association early during KSHV infection (4 h) in HMVEC-d cells. Compared to control Si-RNA, >90% of IFI16 was knocked down with Si-IFI16 RNA (Fig 4H, panel 3, lanes 2 and 4 vs. lanes 1 and 3). Specificity of IFI16 knockdown was shown by the absence of any overall changes in the expression of BRCA1 and ASC protein levels (Fig 4H).
Cell lysates from control or IFI16-Si-RNA transfected HMVEC-d cells either uninfected or infected with KSHV were used for IP with BRCA1 antibody followed by Western blot with ASC antibodies. While the uninfected cells showed very little or no association of ASC with BRCA1 (Fig 4H, panel 1, lane 1–2), ASC was found in BRCA1 immunoprecipitates from control Si-RNA treated KSHV infected cells (Fig 4H, panel 1, lane 3). In contrast, little or no ASC was co-IP-ed with BRCA1 in Si-IFI16 treated virus infected cells (Fig 4H, panel 1, lane 4). There were no overall changes in the IP-ed total BRCA1 (Fig 4H, panel 2) or in total ASC and β-actin levels (Fig 4H, panels 5 and 6), which demonstrated the specificity of the IFI16 knockdown with few off-target effects. The presence of cleaved Caspase-1 in control Si-RNA treated KSHV infected cells but not in uninfected cells and little or none in Si-IFI16 treated virus infected cells (Fig 4H, panel 4) confirmed the activation of the IFI16 inflammasome during KSHV infection [2] and specificity of the experiment.
Next, we analyzed the BRCA1 cellular distribution by PLA in IFI16 knockdown HMVEC-d cells infected with KSHV (4 h p.i.) (S3A and S3B Fig). As shown before, IFI16 PLA spots were detected in the nucleus and in the cytoplasm of infected cells which were significantly absent in the Si-IFI16 cells and the efficiency of IFI16 knockdown is shown by the significant >95% reduction of IFI16 PLA spots (S3A Fig and bar graph). We observed that compared to uninfected Si-control cells with some amount of BRCA1 in the cytoplasm, KSHV infection resulted in increased BRCA1 cytoplasmic distribution which was significantly less in the IFI16 knockdown virus infected cells with a level comparable to that of the uninfected cells (S3B Fig and bar graph).
These results suggested that KSHV induced association of BRCA1 (with no PYD domain) with inflammasome adaptor ASC (with PYD domain) relies on the presence of IFI16 with PYD domain, and demonstrated the dependence on IFI16 for the observed BRCA1–ASC association and increased cytoplasmic BRCA1 translocation during KSHV infection.
To determine whether BRCA1 is a general inflammasome component or is specific only for the IFI16 inflammasome, protein lysates from HMVEC-d cells infected with KSHV for 4 h (early) and 24 h (late) were IP-ed with anti-BRCA1 antibodies and Western blotted for NLRP3 and AIM2 proteins. As shown in Fig 4I, panels 1 and 2, NLRP3 and AIM2 were not associated with BRCA1 in the uninfected or infected cells. No significant changes in the total protein levels of BRCA1, NLRP3 and AIM2 were observed (Fig 4I, input panels). These results not only demonstrated the specific interaction of BRCA1 with IFI16 but also ruled out its association with other cytoplasmic inflammasome sensors under physiological conditions as well as during KSHV infection.
Primary Effusion Lymphoma (PEL) derived BCBL-1 cells are KSHV latently infected B-cells with >80 copies of nuclear KSHV genomes and we have shown previously that IFI16 colocalizes with nuclear KSHV genome in latently infected cells resulting in the constitutive activation of IFI16 inflammasomes [3]. To examine the IFI16-BRCA1 interaction and redistribution in cells latently infected with KSHV, we used a combination of IP reactions with anti-IFI16,-BRCA1 or-ASC antibodies and Western blot reactions with cytoplasmic fractions from latent KSHV positive BCBL-1 and TIVE-LTC (endothelial) cells (Fig 5A and 5B) and also de novo KSHV infected (24 h) human foreskin fibroblast (HFF) cells (S4A Fig). We observed the association of BRCA1 with IFI16, ASC and Caspase-1 in the cytoplasm of infected cells (Fig 5A and 5B and S4A Fig) but not in the cytoplasmic fractions of control KSHV negative BJAB and TIVE (endothelial) (Fig 5C and 5D) or uninfected HFF cells.
These observations were further validated by PLA. Compared to TIVE cells (Fig 5E to 5I, left panels), we observed increased interactions between IFI16-BRCA1, IFI16-ASC, BRCA1-ASC, ASC-Caspase1 and BRCA1-Caspase1, and their distribution in the cytoplasm of KSHV latent TIVE-LTC cells (Fig 5E to 5I, right panels, yellow arrows, graphs in rightmost column). Analysis of average PLA spots showed a significant increase in the association of ASC with Caspase-1 (Fig 5H, graphs in rightmost column) as well as BRCA1 with ASC and Caspase-1 (Fig 5G and 5I, graphs in rightmost column) in the cytoplasm compared to nucleus of virus infected cells as observed before.
Similarly, using PLA, we also observed increased interactions and cytoplasmic distribution of IFI16-BRCA1 and IFI16-ASC complexes in HFF cells infected with KSHV for 24 h (S4B and S4C Fig, lower panels, yellow arrows).
Together with the biochemical studies, these PLA studies strongly demonstrated that KSHV latent infection also induced the association of nuclear sensor IFI16 with BRCA1 to interact with ASC and Caspase-1 and their redistribution to the cytoplasm of infected cells.
Similar to the biochemical studies with B cells and TIVE-LTC cells latently infected with KSHV described above, PLA studies also validated the increased interactions between IFI16-BRCA1, IFI16-ASC, and BRCA1-ASC and their distribution in the cytoplasm of KSHV positive BCBL-1 cells (Fig 6A to 6C, right panels, white arrows) compared to KSHV negative BJAB cells (Fig 6A to 6C, left panels). These results demonstrated that BRCA1 is part of the IFI16 inflammasome complex in B cells with latent nuclear KSHV genome.
Our previous studies demonstrated the constitutive activation of IFI16 inflammasomes in EBV infected cells as well as the colocalization of nuclear IFI16 with the nuclear EBV genomes [4]. In contrast, in cells infected with vaccinia virus replicating its dsDNA in the cytoplasm of infected cells, we observed the activation of the cytoplasmic AIM2-ASC inflammasome and not the IFI16-ASC inflammasome [2]. To determine whether BRCA1 and IFI16 interactions also occur in EBV infected cells, we performed PLA in cells infected with EBV and vaccinia virus. As expected, we observed a significantly increased interaction between IFI16 and ASC, and their distribution in the cytoplasm in EBV latency III positive lymphoblastoid cells (LCL) compared to control EBV negative B-lymphoma Akata and Ramos cells (Fig 6D, panels 1 and 2 versus panel 3, yellow arrows). Furthermore, we also observed increased interactions between IFI16 and BRCA1 and their cytoplasmic distribution in EBV positive cells compared to Ramos and Akata cells (Fig 6E, panels 3 vs panels 1and 2, yellow arrows). De novo KSHV or EBV infection (4 h p.i.) in PBMC also showed IFI16-BRCA1, BRCA1-ASC and BRCA1-Caspase-complex formation and their cytoplasmic distribution as observed by PLA (Fig 6F and 6G).
In contrast to the nuclear replicating herpesvirus infected cells, compared to uninfected cells, we observed very high association of ASC and AIM2 in the cytoplasm of vaccinia infected cells (Fig 6H versus Fig 6I, panel 1) and very little or no interaction between IFI16 and ASC (Fig 6H versus Fig 6I, panel 2). More importantly, we did not observe any significant increase in the IFI16-BRCA1 interactions in vaccinia infected cells compared with uninfected cells (Fig 6H versus Fig 6I, panel 3). These results demonstrated the specificity of BRCA1-IFI16 interactions observed in KSHV and EBV infected cells and suggest that the presence of nuclear viral DNA is necessary for increased BRCA1-IFI16 interactions, association with ASC and their cytoplasmic redistribution.
We next determined whether BRCA1, IFI16 and ASC are in a single macromolecular complex in cells latently infected with KSHV. For this, we performed double sequential PLAs with an initial reaction for IFI16 and BRCA1 (Fig 7A–7D, red spots) followed by a reaction for IFI16 and ASC (Fig 7A–7D, green spots). Very low numbers of green and red PLA spots, mostly distinct, were detected predominantly in the nucleus of KSHV negative BJAB and EBV negative Ramos cells (Fig 7A and 7C). In contrast, significantly higher red and green PLA spots with a higher number of colocalizing yellow spots were observed in KSHV positive BCBL-1 and EBV positive LCL cells (Fig 7B and 7D, white arrow). Interestingly, some white PLA colocalization spots were also detected in the nuclei of BCBL-1 cells (Fig 7B); however, yellow spots were mainly distributed in the cytoplasmic region. These results showed that IFI16, BRCA1 and ASC were in close proximity, suggesting the formation of an individual macromolecular complex, thus corroborating the results presented above. In addition, these results clearly demonstrated that the presence of nuclear viral genomes is critical for the formation of the innate BRCA1-IFI16-ASC inflammasome complexes.
The association of BRCA1-IFI16 with ASC and Caspase-1 prompted us to determine the functional role of BRCA1 in the regulation of IFI16-inflammasome formation leading to cleavage of procaspase-1 into active Caspase-1 followed by cleavage of pro-IL-1β into functional IL-1β.
To assess the role of BRCA1, we performed BRCA1 knockdown using a pool of Si-RNAs against BRCA1 (Si-BRCA1). The uninfected and KSHV infected Si-Control and Si-BRCA1 HMVEC-d and HFF cells (at 0.5 h, 4 h and 24 h p.i.) were used for biochemical studies to determine the effects of BRCA1 knockdown on the cleavage of procaspase-1 and/or pro-IL-1β. Specificity of BRCA1 knockdown was shown by the absence of any overall changes in the expression of IFI16 and ASC protein levels (Fig 8). These results demonstrated that although BRCA1-IFI16 are functionally related to each other, the expression level of these proteins are independent of their functionality.
Compared to Si-Control, >90% of BRCA1 knockdown was observed in Si-BRCA1 cells (Fig 8A and 8B, panel 1, lanes 1–5 vs lanes 6–8). Cleaved Caspase-1 (p20) and cleaved, mature IL-1β (p17) was detected in KSHV infected Si-Control HMVEC-d cells (Fig 8A, panels 2 and 3, lanes 3–5). Similarly, the presence of mature IL-1β was observed only in KSHV infected Si-Control HFF cells (Fig 8B, panel 2, lanes 3–5). However, no detectable cleavage of Caspase-1 and/or IL-1β was observed in Si-BRCA1 treated KSHV infected cells at any time of infection (Fig 8A, panels 2 and 3, lanes 6–8 and Fig 8B, panel 2, lanes 6–8). Additionally, we did not observe any marked changes in the expression levels of IFI16, ASC or β-actin as a result of BRCA1 knockdown (Fig 8A, panels 4, 5 and 8 and Fig 8B, panels 3,4 and 7), indicating the specificity of knockdown and the absence of off-target effects due to BRCA1 knockdown.
To further determine the role of BRCA1 in KSHV infection-induced inflammasome activation, we utilized a human mammary epithelial cell line with wild type BRCA1 (184B5 BRCA1+) and with a BRCA1 null mutant (HCC1937, BRCA1–) [20] (Fig 8C panel 1). Similar to BRCA1 knockdown in HMVEC-d and HFF cells, significant levels of active Caspase-1 (p20) and mature IL-1β (p17) were observed in BRCA1+ cells both at early (0.5 h and 4 h) as well as late (24 h) times p.i. but not in BRCA1– cells (Fig 8C, panels 2 and 3, lanes 2–4 vs. lanes 6–8) or in uninfected cells (Fig 8C, panels 2 and 3, lane 1 vs lane 5). To further confirm the active involvement of BRCA1 in KSHV induced inflammasome activation, we transduced BRCA1– HCC cells with lentiviruses expressing BRCA1 (Fig 8D, panel 1, lanes 5–8) or control lentiviruses (Fig 8D, panel 1, lanes 1–4). As expected, cleaved Caspase-1 and mature IL-1β were observed in BRCA1 expressing but not in BRCA1 negative HCC 1937 cells during either the early or late period post-KSHV infection (Fig 8D, panel 2 and 3, lanes 6–8 vs lanes 1–5). Together, these results demonstrated the active participation of BRCA1 in the formation of functionally active Caspase-1 and mature IL-1β during KSHV infection.
We next determined the potential molecular mechanism responsible for the observed dramatic reduction in inflammasome activation due to BRCA1 knockdown or absence. Since IFI16 inflammasome activation needs simultaneous interaction of adaptor ASC with both IFI16 and procaspase-1, we investigated the physical associations among IFI16 inflammasome components (IFI16, ASC, Caspase-1) in the BRCA1 knockdown cells, in the absence of functional BRCA1 or in lenti-BRCA1 add- back to BRCA1 negative cells. Cell lysates from control or BRCA1-Si-RNA-transfected HMVEC-d or HFF cells either uninfected or infected with KSHV were used for IP with anti-ASC antibody followed by WB with anti-IFI16 or Caspase-1 antibodies. While the uninfected cells showed very little or no association of ASC with IFI16 and Caspase-1, these associations were prominently observed in ASC immunoprecipitates of Si-Control treated KSHV infected cells both at early (0.5 and 4 h) as well as late (24 h) times p.i. (Fig 8A, panels 6–7 and Fig 8B, panels 5–6, lanes 3–5 vs lanes 1–2). In contrast, little or no IFI16 or Caspase-1 was co-IP-ed with ASC in Si-BRCA1 treated virus infected cells (Fig 8A, panels 6–7 and Fig 7B, panels 5–6, lanes 6–8).
Furthermore, we observed the presence of IFI16 and Caspase-1 in ASC immunoprecipitates predominantly in KSHV infected BRCA1 + 184B5 cells but not in BRCA1 −HCC1937 cells (Fig 8C, panel 6–7, lanes 2–4 vs lane 1 and 5 and lanes 6–8) as well as in lenti-BRCA1 expressed but not in control lenti-vector transduced HCC1937 cells upon early and late KSHV infection over uninfected cells (Fig 8D, panels 5–6, lane 6–8 vs lanes 1–5). However, there were no marked changes in the expression of IFI16, ASC and β-actin in any of these conditions.
We further analyzed the effect of BRCA1 knockdown on IFI16-ASC interactions by PLA during de novo KSHV infection in HMVEC-d and HFF cells. Although a considerable amount of IFI16 was detected in the nucleus of both uninfected and KSHV infected Si-Control and si-BRCA1 HMVEC-d cells (Fig 9A and 9B), cytoplasmic redistribution of IFI16 molecules was predominantly observed in Si-Control cells compared with that in Si-BRCA1 cells (Fig 9A vs. Fig 9B, yellow arrows). Next, we assessed the IFI16-ASC interactions in uninfected and KSHV infected Si-Control and si-BRCA1 HMVEC-d and HFF cells. Compared to uninfected HMVEC-d and HFF cells, we observed more IFI16-ASC PLA spots, indicating increased interactions in Si-Control cells at both early and late times post-KSHV infection (Figs 9C and 8E vs Fig 9D and 9F). In addition, considerable numbers of the observed PLA spots were redistributed to the cytoplasm after 4 h and 24 h of infection (Fig 9C, white arrows and Fig 9E, yellow arrows). In contrast, in Si-BRCA1 cells, very few IFI16 and ASC PLA spots were seen at either early or late times post-KSHV infection, implying very few interactions (Fig 9D and 9F). These findings suggested that BRCA1 is essential for IFI16-ASC complex formation and its cytoplasmic translocation during de novo KSHV infection.
Collectively, these results clearly demonstrated the essential role of BRCA1 in regulating KSHV induced IFI16 cytoplasmic redistribution and assembly of the IFI16-ASC-Caspase-1 inflammasome complex and IFI16 inflammasome activation.
IFI16 acts as a nuclear DNA sensor for KSHV, EBV and HSV-1 genomes and induces inflammasome activation [2, 4, 5]. It also recruits the endoplasmic resident protein STING to interact with TBK1 and IRF3, leading to phosphorylation and nuclear translocation of IRF3 in herpes virus infected cells [6,7]. Based on our findings that (i) BRCA1 interacts with IFI16 and ASC, and undergoes cytoplasmic translocation as part of the IFI16 inflammasome complex during de novo and latent KSHV and EBV infections, (ii) the presence of nuclear herpes viral DNA is necessary for BRCA1-IFI16 and IFI16-ASC interactions and, (iii) the absence of BRCA1 impaired IFI16 cytoplasmic translocation, IFI16-ASC-Caspase-1 associations and IFI16 inflammasome activation, we hypothesized that although IFI16 can directly bind to DNA in artificial systems [1], in the dynamic nuclear environment during herpes virus infections, BRCA1 is necessary for recognizing nuclear viral DNA by IFI16. This phenomenon results in cytoplasmic translocation of IFI16, activation of IFI16 inflammasomes and induction of host innate responses such as type-I interferon production.
To test this, we determined whether the BRCA1-IFI16 complex recognizes the KSHV genome by infecting HMVEC-d and HFF cells for 24 h with KSHV containing 5-ethynyl-2’-deoxyuridine (EdU) labeled viral genomes (Fig 10A and 10B, panels 1, 3 and 4, red spots within DAPI stained nuclei) and performed PLA for IFI16-BRCA1. In the combined PLA-IFA images, many of the IFI16-BRCA1 (green) complexes colocalized with KSHV genomes (red) in the nuclei (Fig 10A and 10B, yellow spots (BRCA1-IFI16+EdU) in panel 2 and white spots (BRCA1-IFI16+EdU+DAPI) in panel 3 and the yellow arrows of enlarged images in panel 4), which demonstrated possible viral genome sensing events mediated by this complex. A considerable number of BRCA1-IFI16 complex green spots were observed in the cytoplasm in addition to those in infected cell nuclei (Fig 10A and 10B, panels 2–4, red arrows) implying a virus dependent translocation of IFI16-BRCA1 complexes as demonstrated in our results described earlier.
To further demonstrate that the presence of nuclear KSHV genome drives the formation of BRCA1, IFI16 and ASC single macromolecular complexes, a sequential PLA with IFI16-BRCA1 and then IFI16-ASC was performed in HMVEC-d cells infected with EdU KSHV for 24 h. Red EdU spots within the nucleus demonstrated the infection and presence of viral genome (Fig 10C, panel 1) which were also represented as blue spots for the convenience of analysis (Fig 10C, panel 3 and 4). Considerable numbers of BRCA1-IFI16 red spots and IFI16-ASC green spots colocalized (yellow spots) in the nucleus as well as in the cytoplasm (white arrows). These results demonstrated the simultaneous presence of BRCA1-IFI16-ASC in a single complex. Most interesting was the detection of numerous white spots in the nucleus (Fig 10C, orange arrows) representing colocalization of BRCA1-IFI16 and IFI16-ASC complexes with the KSHV genome. These results demonstrated that the BRCA1-IFI16-ASC complexes recognize the presence of the foreign KSHV genome in the host cell nuclei.
Although some of the IFI16-ASC complexes (green spots) were also observed in very close proximity to the BRCA1-IFI16 complexes (red spots), distinct clear colocalization yellow spots were not visualized in all cases and some spots were distant from each other (Fig 10C). These might be due to the dynamic nature of one or more precomplexes followed by formation of mature inflammasome complexes that are >40 nm apart or formation of a complex with separate entities and distinct functionality.
To investigate the functional implications of BRCA1 in KSHV genome recognition by the BRCA1-IFI16 complex, we verified the consequence of BRCA1 knockdown affecting KSHV genome sensing by IFI16 in HMVEC-d cells at 24 h post-EdU KSHV infection. In control Si-RNA treated cells, we observed the colocalization of nuclear IFI16 molecules (green PLA spots) with EdU KSHV genome (red) (Fig 10D, panels 4 and 5, and Fig 11A, yellow spots, white arrows). As seen before, IFI16 was also detected in the cytoplasm without any colocalization with the viral genome probe (Fig 10D, panels 2, 3 and 4, green spots, red arrows and Fig 11A). These results demonstrated the cytoplasmic redistribution of IFI16 during KSHV infection and the absence of free viral genome in the cytoplasm. Multiple IFI16-DNA colocalization spots indicated interactions with several viral genomes in the presence of functional BRCA1. In contrast, IFI16 was confined to the nucleus of the BRCA1 Si-RNA treated cells with notably decreased colocalization with the EdU KSHV genome at 24 h p.i. (Fig 10E, panels 4 and 5 and Fig 11B and 11F). Consistent with this result, in contrast to Si-Control cells infected with KSHV, very little colocalization of IFI16 PLA spots and EdU KSHV genome was observed in BRCA1 knockdown HMVEC-d cells even at the early time of KSHV infection (Fig 11D and 11E vs Fig 11C). Taken together, these results convincingly demonstrated that in the dynamic nuclear environment, IFI16 relies on BRCA1 to increase its affinity to foreign KSHV DNA leading into stable inflammasome complex formation and translocation into the cytoplasm of infected cells.
To elaborate on the functional significance of BRCA1, we measured the bioactive form of IL-1β secreted as a consequence of inflammasome activation by KSHV infection. Compared to uninfected HMVEC-d cells, a significant but gradual increase in IL-1β secretion (~4.8–19 pg/ml) was observed from early to late times p.i. in control Si-RNA treated cells (Fig 12A). In contrast, treatment of BRCA1 Si-RNA resulted in dramatic reduction in IL-1β secretion (~1.6–3.2 pg/ml) (Fig 12A) which corroborated the active role of BRCA1.
Because a fraction of IFI16-BRCA1 complexes did not colocalize with IFI16-ASC in our sequential PLA experiments (Figs 7B, 7D and 10C) and there was significant restriction in the IFI16 nuclear to cytoplasmic translocation by BRCA1 knockdown, we set out to determine the role of BRCA1 in IFI16 mediated type I interferon production. When we first investigated the association of IFI16-STING responsible for the initial signal cascade for IFN-β via PLA, compared to uninfected HMVEC-d cells, control Si-RNA treatment followed by KSHV infection resulted in increased association of IFI16-STING at 4 h and 8 h p.i. which decreased at 24 h p.i. In contrast, BRCA1 Si-RNA treatment resulted in a nearly complete abrogation of IFI16-STING PLA spots (Fig 12B, green PLA spots).
Next, we determined the downstream signal molecule pIRF3 levels by WB in BRCA1 knockdown or BRCA1 reintroduction to BRCA1 negative cells following KSHV infection. A gradual increase in the pIRF3 level was observed from 0.5 h p.i. which decreased somewhat at 24 h p.i. in Si-Control treated KSHV infected HMVEC-d cells (Fig 12C, upper panel, lanes 3–5 vs lanes 1–2). In contrast, there was a substantial decrease in pIRF3 levels in Si-BRCA1 treated cells at early as well as late times p.i. (Fig 12C, upper panel, lanes 6–8 vs lanes 3–5) with no apparent change in total IRF3 (Fig 12C, bottom panel).
Similarly, compared to uninfected and lenti-vector control transduced cells, lenti-BRCA1 expression in BRCA1 negative HCC1937 cells resulted in increased phosphorylation of IRF3 from 0.5 h p.i. which decreased somewhat at 24 h p.i. (Fig 12D, upper panel, lanes 6–8 vs lanes 2–4 and lanes 1 and 5) with no significant changes in total IRF3 levels (Fig 12D, bottom panel). Consistent with these findings, IFA studies also showed a significant increase in pIRF3 levels and translocation to the nucleus in control Si-RNA treated KSHV infected HMVEC-d cells up to 8 h p.i. and a decrease at 24 h p.i. (Fig 12E, 1st and 2nd rows). However, with BRCA1 knockdown, IRF3 phosphorylation and nuclear translocation was significantly reduced (Fig 12E, 3rd and 4th rows). When IFN-β release was measured by ELISA, compared to uninfected cells (~26 pg/ml), we observed a gradual increase of IFN-β secretion (~42 pg/ml, ~74 pg/ml, ~125 pg/ml, ~67 pg/ml, ~61 pg/ml at 0.5, 4, 8, 18 and 24 h p.i., respectively) in Si-Control treated KSHV infected cells. In contrast, in BRCA1 depleted KSHV infected cells, we observed a significant reduction in IFN-β secretion (Fig 12F).
Together, these results clearly highlighted the essential role of BRCA1 in the regulation of KSHV genome recognition by IFI16 and correlated with the subsequent IFI16 mediated host innate responses of induction of inflammasomes and IFN-β.
Since BRCA1 in complex with IFI16 enhanced the nuclear KSHV genome affinity of IFI16 we further postulated a similar crucial role for BRCA1 in facilitating nuclear HSV-1 genome recognition by IFI16 with concomitant innate response activation. Uninfected HFF cells or HFF cells infected with EdU genome labeled HSV-1 (KOS) with 1 PFU/cell (~25 DNA copies/cell) for 30 min were processed for PLA. Compared to uninfected HFF cells, where BRCA1 was associated with IFI16 only in the nucleus (Fig 13B, green PLA spots) as IFI16 was predominantly present in the nucleus (Fig 13A, green PLA spots), some of the IFI16 and IFI16-BRCA1 PLA spots colocalized with Edu-HSV-1 genome in the nucleus within 30 min of infection (white arrows, Fig 13D and 13E; left and also rightmost panels). Similar to KSHV infection, increased levels of IFI16-BRCA1 spots were observed in the cytoplasm of HSV-1 infected cells (Fig 13E, green spots; yellow arrows). BRCA1 was not associated with ASC in uninfected cells (Fig 13C) and in contrast, BRCA1-ASC association spots were observed in the cytoplasm of HSV-1 infected cells (Fig 13F, green spots; yellow arrows) which suggested that BRCA1 is a constituent of the HSV-1 induced IFI16 inflammasome. More importantly, in BRCA1 knockdown cells, IFI16 was detected only in the nucleus with little colocalization with EdU HSV-1 genome compared to Si-Control cells (Fig 13H vs. Fig 13G).
To further demonstrate the role of BRCA1 in IFI16 mediated viral genome recognition, we used a biochemical method of DNA mediated chromatin pull down (Fig 14, schematic diagram) [21]. HFF cells pretreated with control Si-RNA or BRCA1 Si-RNA were infected with EdU labeled or unlabeled HSV-1 for 1 h. After protein-DNA cross-linking, biotin-TEG azide was selectively linked to the reactive alkyne group of EdU containing viral DNA via Click reaction. This was followed by DNA shearing, and the small biotin tagged chromatin fragments of HSV-1 genome were captured on streptavidin beads (Fig 14, schematic diagram steps I to V). HSV-1 genome associated IFI16 was analyzed by the elution of streptavidin captured genome bound proteins followed by Western blotting (Fig 14A, schematic diagram step VI). In control Si-RNA treated cells; IFI16 was detected in the eluted fraction from the EdU viral genome pulled down, which suggested the association of IFI16 with HSV-1 genome (Fig 14B, upper 1st panel, lane 3). In contrast, in BRCA1 Si-RNA infected cells, significantly less IFI16 was detected from EdU genome pulled down fractions (Fig 14B, upper 1st panel, lane 4) although similar levels of genome associated histone H3 and TBP were observed in these conditions (Fig 14B, upper 2nd and 3rd panels, lanes 3 and 4). A similar experiment with unlabeled HSV-1 did not detect any protein (Fig 14B, upper panels, lanes1 and 2) pulled down from equal chromatin fractions (Fig 14B, bottom input panels) which demonstrated the specificity of the EdU genome pull down method.
DNA purified from input and streptavidin captured materials showed similar levels of EdU viral genome pulled down in both control Si-RNA and BRCA1 Si-RNA treated chromatin fractions (Fig 14C, schematic diagram step VII). We also observed that DNA was recovered only from labeled virus infected cells but not from unlabeled virus infected cells (Fig 14C, upper panel, lanes 3 and 4 vs lanes 1 and 2). These studies further demonstrated the specificity of the method.
Overall, these data clearly suggested the involvement of BRCA1 in virus genome recognition by IFI16.
We next determined the effect on IFI16 mediated innate immune induction of inflammasomes and IFN-β as a consequence of HSV-1 genome sensing in the presence or absence of BRCA1. Immunoblot analysis from cell lysates of BRCA1 positive (184B5) and BRCA1 negative (HCC1937) cells either uninfected or infected with HSV-1 (1 pfu/cell) demonstrated clear cleavage of caspase-1 and IL-1β at 1, 2 and 4 h p.i. which were nearly absent at 8 h p.i. (Fig 15A, panels 3 and 4, lanes 2–5) which is consistent with IFI16 degradation at later time points p.i. with HSV-1 [5]. In contrast, little or no cleaved caspase-1 or IL-1β was detected in uninfected or HSV-1 infected HCC cell lysates (Fig 15A, panels 4 and 5, lanes 2 and 5 and lanes 7–10). A similar initial increase of pIRF3 at 1, 2 and 4 h p.i. but not at 8 h p.i. was observed in BRCA1 positive HSV-1 infected cells whereas pIRF3 levels were comparatively low or absent in BRCA1 negative cells (Fig 15A, panel 6) with no apparent change in total IRF3 levels (Fig 15A, panel 7). Though HSV-1 infection induced IFI16 at 1 h p.i. (Fig 15A, panel 3), as demonstrated before [5, 7], it was reduced by 2 h p.i. and nearly absent after 4 h p.i. The observed reduction of IFI16, caspase-1, IL-1β and p-IRF3 post HSV-1 infection in BRCA1 positive cells could be due to the increased expression of viral E3 ubiquitin ligase ICP0 (Fig 15A, panel 1) which results in the degradation of IFI16 [5, 7]. In contrast, in the absence of BRCA1, IFI16 levels were not decreased even after 8 h p.i. and actually increased (Fig 15A, panel 3) which could be due to the absence of BRCA1-IFI16 complex resulting in little or no genome recognition by IFI16 and thus unresponsive to ICP0 or some other factors that sequestered IFI16 away from ICP0 leading into delayed degradation kinetics.
HSV-1 infection in BRCA1 positive cells resulted in an initial increase in secreted IFN-β and IL-1β at 2 h and 4 h p.i. followed by a decrease at 8 h p.i. (Fig 15B and 15C) which were significantly reduced in BRCA1 negative cells (Fig 15B and 15C). Interestingly, similar to our earlier observation in IFI16 negative cells [9], compared to BRCA1 positive cells, >4 fold increase in HSV-1 yield was observed in BRCA1 negative cells upon HSV-1 infection (Fig 15D). This could be either due to less IFN-β production as a result of the absence of an IFI16-STING-TBK1-IRF3 signal cascade or loss of inflammasome activation in these cells and/or decreased transcriptional repressor activity of IFI16 as a result of decreased viral genome affinity of IFI16.
Furthermore, in the co-IP experiments, BRCA1 and Caspase-1 were detected in ASC immunoprecipitates from Si-Control RNA treated HSV-1 infected (1 pfu/cell) HFF cells. In contrast, little or no BRCA1 or Caspase-1 were co-IP-ed with ASC in IFI16 knockdown (Si-IFI16) cells (Fig 15E, 1st and 2nd panels) and the presence of cleaved Caspase-1 was observed only in control-Si-RNA treated HSV-1 infected cells (Fig 15E, 5th panel). Whole cell lysates from Si-Control and Si-IFI16 HFF cells (Fig 15E, 3rd panel) showed BRCA1, ASC, Caspase-1 and HSV-1-ICP0 levels (Fig 15E, panels 4, 5, 6 and 8).
When we analyzed the BRCA1 cellular distribution in IFI16 knockdown HFF cells infected with HSV-1 for 2 h (Fig 15F and 15G), as observed before, IFI16 PLA spots were detected in the nucleus and in the cytoplasm of infected cells which were significantly absent in the Si-IFI16 cells which demonstrated the efficiency of IFI16 knockdown (Fig 15F). Compared to uninfected Si-control cells with some amount of BRCA1 in the cytoplasm, HSV-1 infection resulted in significantly increased BRCA1 cytoplasmic distribution which was significantly reduced in the IFI16 knockdown virus infected cells with a level comparable to that of the uninfected cells (Fig 15G and bar graph).
These results suggested the dependence on IFI16 for the observed BRCA1-ASC and ASC-Caspase-1 association during inflammasome complex formation and increased cytoplasmic BRCA1 translocation during HSV-1 infection. These results also clearly demonstrated that similar to KSHV, BRCA1 is also essential for IFI16 to sense the HSV-1 genome, activation of inflammasomes, cytoplasmic translocation, and in the induction of the IFN-β response.
The innate immune system of mammalian cells utilizes distinct sensors and methods to detect DNA in endosomes, cytoplasm, or the nucleus which include recognizing differences in physicochemical structure such as CpG motifs or AT-rich regions of microbial DNA, abnormal occurrence of microbial DNAs or accumulation of endogenous or microbial DNA and elicitation of DDR against incoming viral DNAs [22]. Our studies presented here reveal that formation of distinct complexes, such as IFI16-BRCA1, is another way to detect herpes viral DNA entering the nucleus. These comprehensive studies demonstrate that BRCA1 plays a hitherto unidentified role as a cofactor to IFI16 in the nuclear innate sensing of foreign DNA and subsequent assembly and cytoplasmic distribution of stable IFI16-inflammasomes leading into IL-1β formation as well as the induction of IFN-β via cytoplasmic signaling through IFI16, STING, TBK1 and IRF3 (Fig 16).
Similar to the IFI16-BRCA1 interaction, previous studies suggest that some of the cytoplasmic inflammasome sensor molecules require cofactors to recognize their ligands or to be stabilized in their activated states. For example, NLRP1 requires NOD2 protein to interact with muramyl dipeptide ligand and to activate the NLRP1 inflammasome [23]. NLRC4 inflammasome formation requires NAIP (NLR family, apoptosis inhibitory protein). Mouse NAIP5 and NAIP6 detect flagellin, NAIP2 senses the type III secretion component Prg J and human orthologue NAIP senses a type III secretion needle protein to activate the NLRC4 inflammasome [24]. Similarly, formation of the NLRP3 inflammasomes in response to non-crystalline activators is also promoted by guanylate-binding protein 5 [25].
The presence of foreign microbial or host DNA in the cytoplasm is recognized as abnormal by AIM2 resulting in the AIM2-inflammasome [26]. Studies of the crystal structures of HIN domains of AIM2 and IFI16 in complex with dsDNA demonstrated that IFI16 and AIM2 recognize DNA in a non-sequence specific manner through electrostatic attraction between the positively charged HIN domain residues and the sugar phosphate backbone of dsDNA. In these studies with overexpressed proteins out of context of the nuclear environment, the PYD and HIN domains of AIM2 are shown to be in an autoinhibited intramolecular complex state which is liberated upon DNA binding thus facilitating DNA-mediated assembly of an inflammasome along the DNA staircase [1]. Whether AIM2 requires additional factor(s) for its function in the cytoplasm is not known.
In the nucleus, the nuances of sequence independent DNA sensing by IFI16 distinguishing between foreign and host DNA are not clear, since unlike other danger signals inducing inflammasomes and the interferon response, DNA is not unique to pathogens. Our studies demonstrate that the BRCA1 interaction with IFI16 predisposes IFI16 to recognition of viral genome as shown by the less IFI16 association in viral EdU genome pull down analysis and poor colocalization of IFI16 with EdU viral genome in the BRCA1 knockdown cells. Hence, the interaction of BRCA1 with the pyrin domain of IFI16 [15,19] could be liberating the autoinhibited complex state thus facilitating the HIN domains of IFI16 to sense the viral genome to trigger a rapid inflammasome and interferon-β response.
A recent study suggested that the PYD domain is essential for driving IFI16 to cooperatively assemble into filaments on DNA when IFI16 encounters long dsDNAs (greater than 20 bp in size), which can elevate the dsDNA binding efficiency of IFI16 [27]. In the nucleus, the major exposed self-dsDNAs are thought to be the dsDNA linkers between nucleosomes or the transcription bubbles, which are about 20 bp in size [28, 29]. It is believed that the exposed self-dsDNA in such short lengths would bind nuclear IFI16 with limited affinity and would not be capable of providing a sufficient platform for IFI16 to oligomerize into filaments or inflammasomes, thereby preventing spurious activation of immune pathways [27]. In contrast, IFI16, already in an open conformation aided by the interaction of BRCA1 with the PYD domain, probably rapidly engages the non-chromatinized dsDNA viral genomes of KSHV, EBV and HSV-1 soon after their nuclear entry more efficiently than self-DNAs likely due to a higher binding affinity leading to the rapid assembly of IFI16 inflammasomes and induction of the IFN-β response. Non-availability of BRCA1 may render a limited IFI16 response/affinity for such exposed viral DNA, which is insufficient to induce inflammasome activation and IFN-β production.
The immunomodulatory role of IFI16 is shown by its ability to recognize and respond to the presence of nuclear herpesviral and non-integrated HIV genomes, leading to inflammasome induction and/or to participate in STING-mediated IFN-β expression during HSV-1 and vaccinia virus infection [30]. Although IFI16 was demonstrated to interact with cytoplasmic STING in a DNA-dependent manner resulting in the recruitment of TBK1 and IRF3 plus phosphorylation of IRF3 [1,6], the key unresolved question concerns the mechanisms by which nuclear DNA sensing by nuclear resident IFI16 can trigger cytosolic STING dependent IFN-β induction. This would either require delivery of IFI16 as a nuclear IFI16-DNA complex to the cytoplasm for direct association with STING or for a cGAS (cGAMP-Synthase)-mediated cGAMP (cyclic GMP-AMP) production to bind with STING [31]. Absence of viral genome in the cytoplasm of productively infected cells together with the near absence of IFI16-STING in the BRCA1 knockdown cells infected with KSHV or HSV-1 clearly suggest the formation of a distinct IFI16 molecular complex with one or more nuclear protein(s), such as BRCA1, drives IFI16’s cytoplasmic trafficking, leading to the interaction of IFI16 with the STING signalosome.
Although initially identified as a cytoplasmic DNA sensor [31], in a recent study, Orzalli et al., [32] reported the detection of cGAS in both the nucleus and cytoplasm of HFF and immortalized oral keratinocyte cells. IFI16 was detected in HEK293 cells and not in 293T cells, and cGAS was absent in HEK293 and 293T cells. Knockdown of both cGAS and IFI16 reduced the IFN-β responses in HSV-1 infected HFF cells. IFI16 appeared to be stabilized by cGAS in the uninfected HFF cells as in the absence of cGAS, IFI16 half-life was moderately reduced [32]. The endogenous IFI16 co-precipitated with cGAS from the whole cell extracts of normal as well as HSV-1 infected HFF cells (3 h p.i.). However, whether this interaction occurs in the nucleus or in the cytoplasm and the kinetics of this interaction after nuclear viral genome recognition is not known. From the observations of the localization of IFI16 at the site of Wt-HSV-1 (KOS) viral genome and replication and the absence of cGAS at this site, Orzalli et al., [32] have concluded that IFI16 is the primary sensor of viral DNA in the nucleus and cGAS is not involved in genome recognition. They also reported that cGAS is involved in the early induction of the IFN-β response against transfected DNA, probably by recognizing the DNA in the cytoplasm and is also required for the IFN-β response at 8 h post-HSV-1 infection. Inflammasome response against the transfected DNA by AIM2 was not examined. In addition, how the message of nuclear viral genome sensing by IFI16 in the presence of nuclear cGAS is transmitted to the cytoplasm for STING-IRF3 mediated IFN-β production is not clear. Our studies suggest that IFI16 requires BRCA1 to increase its affinity to nuclear viral DNA leading into the sustained innate responses including the translocation of IFI16 into the cytoplasm of infected cells and the activation of STING. Unlike cGAS, which has transient association with IFI16 [32], our studies show that BRCA1 is associated with IFI16 in the nucleus under physiological conditions, and is involved in the sensing of herpes viral (KSHV, EBV and HSV-1) genomes by IFI16 and in the consequent innate inflammasome and IFN-β responses. It is likely that the interaction of IFI16 and STING is within a complex in the cytoplasm that may be stabilized by cGAS. Whether cGAS interacts with IFI16-BRCA1 or not and the kinetics, as well as the role of cGAS in IFI16-BRCA1-STING mediated innate IFN-β response requires additional studies which are beyond the scope of the present study.
A recent study utilized a PMA stimulated human monocytic THP-1 cell line for infection with KSHV and reported that IL-1β and IFN-β were induced via an IFI16 independent pathway [30]. However, the authors did not demonstrate that the TPA stimulated THP-1 cell actually supported KSHV and HSV-1 infection, and both viruses are known to infect unstimulated THP-1 cells. Moreover, due to increased phagocytosis in PMA induced THP-1 cells, herpesviruses undergo abortive infection, resulting in the accumulation of virus particles in lysosomes and the release of viral DNA to the cytoplasm [33]. This DNA is likely recognized by AIM2 and cGAS, to stimulate the IL-1β and interferon responses. In contrast, during in vitro infection of permissive cells, viral DNA from the capsid enters the IFI16 rich nucleus resulting in the consequences presented by our studies.
Besides the role of IFI16 in inflammasome and type 1 IFN induction, IFI16 is also shown to be a transcriptional modulator; however, the detailed mechanisms are poorly defined [16]. Recent studies by us and others demonstrated that in HSV-1 infected cells, IFI16 promoted the addition of repressive heterochromatin H3K9me3 markers and reduced the active euchromatin H3K4me3 markers on the viral gene promoters resulting in the reduced binding of transcription factors Oct1 and TBP, plus RNA pol II [8, 9]. Studies suggest that IFI16 may target exogenous DNA not associated with nucleosomes as virion associated dsDNA of SV40 genome containing nucleosomes and the adenoviral genome with core protein VII were resistant to the restriction effect of IFI16 [7]. However, from our studies showing that IFI16 is associated with latent chromatinized KSHV and EBV genomes, continuation of viral latent gene expression in the presence of IFI16 and IFI16-inflammasomes [3, 4], IFI16’s ability to mediate global differences in HSV-1 genome chromatin modifications [8, 9], and our ongoing studies showing the binding of IFI16 to the KSHV gene promoters in latently infected cells, we theorize that IFI16 has very complex roles in gene regulation and there may be other factors involved in IFI16’s ability to discriminate foreign vs. host DNA. Similar to the IFI16-BRCA1 complex involved in pathogen DNA recognition, inflammasome induction and IFN-β production shown here, it is possible that IFI16 could be forming distinct complexes with different proteins and each complex could be mediating distinct functions, such as transcription and other responses, which may differ between various virus nuclear lytic and latent infections and as per host cell types. Determining these possibilities and deciphering whether BRCA1 plays a role in IFI16’s ability to influence viral promoters and in the nuclear life cycles of DNA viruses requires further extensive studies which are beyond the scope of the present study.
Lack of association of BRCA1 with ASC in IFI16 knockdown cells demonstrated that BRCA1 likely doesn’t interact directly with ASC but relies on IFI16 for ASC association and for subsequent stable inflammasome formation. The multiple binding sites of ASC-PYD with distinct positively and negatively charged surfaces are proposed to transition between folded and unfolded states to regulate PYD function [34, 35]. The interaction of BRCA1 with the PYD of IFI16 could facilitate the spatial distribution of the binding sites between IFI16 and ASC to drive their stable association and inflammasome assembly. BRCA1 knockdown significantly reduced cytoplasmic IFI16 suggesting that BRCA1 facilitates cytoplasmic trafficking of nuclear IFI16 during KSHV and HSV-1 infection, which could be due to BRCA1 dependent post-translational modification (acetylation or phosphorylation) of IFI16. Different IFI16 motifs have been demonstrated to regulate IFI16 subcellular localization [36], and the role of BRCA1 in post-translational modification and cytoplasmic distribution of IFI16 are under investigation.
BRCA1 has been reported to be distributed in the nucleus and cytoplasm with distinct functionality such as DDR, chromatin remodeling, checkpoint control and apoptosis [37]. However, less cytoplasmic BRCA1 in IFI16 knockdown than control Si-RNA treated virus infected cells observed by us probably demonstrates the need for IFI16 in complex with BRCA1 for viral genome sensing followed by translocation to the cytoplasm. Although a study suggests the role of BRCA1-IFI16 interaction in a p53-mediated apoptosis pathway [15], IFI16 or BRCA1 knockdown did not affect endothelial cell viability. As shown here, formation of the IFI16-inflammasome was not detected in cells induced for DDR by bleomycin. Furthermore, the presence of DDR proteins CHK2 and H2AX in BRCA1 immunoprecipitates, but not in IFI16 immunoprecipitates despite the increased BRCA1-IFI16 interaction during KSHV infection, suggests that the BRCA1-IFI16 complex is probably not related to host DDR or apoptosis responses during DNA virus infection and may have evolved to mediate innate sensing and other functions.
The protective immune response triggered by the detection of microbial effectors termed “effector triggered immunity (ETI)” enables the host to distinguish pathogens from non-pathogen [38, 39, 40] and this is probably of relevance in non-professional immune cells too, such as epithelial, endothelial and other cells that come in contact with pathogens at different in vivo sites, including at the portal of pathogen entry. Our studies demonstrating that IFI16-BRCA1 functions as an innate sensor of KSHV, EBV and HSV-1 genomes in non-immune professional cells, such as endothelial, epithelial and fibroblast cells, and in B-lymphoma cell lines as well as the IFI16 and ASC interactions in tissue sections from PEL and KS patients [3], clearly suggested that these cells have evolved to respond to danger signals. Additional studies are required for a clear understanding of the molecular mechanism of BRCA1 and IFI16 in the fundamental relation between innate DNA sensing, viral and host genome regulation, chromatin remodeling, inflammation and immunity.
HMVEC-d and HFF cells (Clonetics, Walkersville, MD), TIVE, TIVE-LTC, BCBL-1, LCL, BJAB, Akata, Ramos, and BRCA1 positive and negative breast carcinoma (ATCC CRL 8799 and 2336) cells were grown as described before [2,4,41].
KSHV DNA was metabolically labeled during lytic replication by adding the thymidine analogue 5-Bromo-2’-deoxyuridine (BrdU) (Life Technologies) (1:100 v/v from supplied stock) or 5-ethynyl-2’-deoxyuridine (EdU) (Sigma) (10 μM) in DMSO to the culture medium of BCBL-1 cells on day 1 and day 3 of phorbol ester induction, and labeled virus from the day 5 culture supernatant was purified and genome copy numbers determined [2]. HSV-1 (KOS) genome was labeled by adding EdU to the Vero cell medium at 8 h, day 1 and 2 post-infection and labeled virus in the day 4 supernatant was purified and titer determined [5].
Deidentified peripheral blood mononuclear cells (PBMCs) were obtained from the University of Pennsylvania CFAR Immunology Core, and 1 × 107 PBMCs were infected [4]. Briefly, PBMCs were infected with KSHV or EBV in 1 ml of RPMI 1640 medium with 10% FBS and 5 ng/ml of Polybrene (Sigma, Marlborough, MA). After 4 h at 37°C (time point 0), infected and uninfected cells were centrifuged for 5 min at 1,200 × g; the pellet was washed twice with fresh RPMI medium, resuspended in fresh RPMI medium with 10% FBS, and cultured in six-well plates at 37°C. These cells were collected at different times p.i., and washed twice with 1× phosphate-buffered saline (PBS) before being spotted on slides.
Mouse monoclonal antibodies against human BRCA1 and rabbit polyclonal antibodies against human BRCA1 were from GeneTex, Irvine, CA and Millipore, Billerica, MA, respectively. Rabbit anti-human H2AX, γH2AX and CHK2 antibodies were from Cell Signaling Technology, Beverly, MA. Mouse monoclonal antibodies against human ASC were from MBL International, Woburn, MA. Goat polyclonal antibodies against human ASC/TMS1 were from Ray Biotech, Norcross, GA. Mouse monoclonal anti-IL-1β antibodies were from R&D Systems, Inc., Minneapolis, MN. Rabbit, mouse and goat antibodies against human IFI16 were from Sigma, St Louis, MO and Santa Cruz Biotechnology, Inc., respectively. Rabbit polyclonal anti-caspase-1 antibodies were from Bio Vision, Milpitas, CA. Rabbit anti-AIM2 and mouse anti-NLRP3 antibodies were from Abcam Inc., Cambridge, MA. Rabbit monoclonal antibodies against STING, p-IRF3 and Histone H3 are from Cell Signaling Technology. Mouse monoclonal anti-IRF3 antibody was from Abcam Inc. Alexa 594 or 488 anti-rabbit and anti-mouse secondary antibodies were from Molecular Probes, Invitrogen, Carlsbad, CA. Anti-rabbit and anti-mouse antibodies linked to horse-radish peroxidase were from KPL Inc., Gaithersburg, MD. Mouse monoclonal antibodies against β-actin and β-tubulin were from Sigma. Mouse monoclonal anti-TATA binding protein (TBP) antibody was from Abcam Inc. Protein A-Sepharose 6 MB and Protein G-Sepharose CL-4B Fast Flow beads were from Amersham Pharmacia Biotech, Piscataway, NJ.
Cells were harvested and used for preparation of nuclear and cytoplasmic extracts using a nuclear extract kit (Active Motif Corp., Carlsbad, CA) as per the manufacturer's instructions. After measurement of protein concentrations with BCA protein assay reagent (Pierce, Rockford, IL), nuclear and cytoplasmic extracts were subjected to Western blot (WB) analysis with different antibodies. The purity of the nuclear extracts and cytoplasmic extracts was assessed by immunoblotting with anti-TBP and anti-β-tubulin antibodies, respectively.
Cells were lysed in radioimmunoprecipitation assay (RIPA) lysis buffer (15 mM NaCl, 1 mM MgCl2, 1 mM MnCl2, 2 mM phenylmethylsulfonyl fluoride and protease inhibitor mixture (Sigma)), sonicated, and centrifuged at 10,000 rpm at 4°C for 10 min. Protein concentrations were estimated by BCA protein assay reagent (Pierce). Equal concentrations of proteins were separated by SDS-PAGE, transferred to nitrocellulose and probed with the indicated specific primary antibodies followed by incubation with species-specific HRP-conjugated secondary antibody and chemiluminiscence based detection (Pierce) of immunoreactive protein bands according to the manufacturer’s protocol. Blots were scanned and quantitated using FluorChemFC2 software and an AlphaImager system (Alpha Innotech Corporation, San Leonardo, CA).
Cells were lysed in lysis buffer (25 mM Tris-HCl, pH: 7.5, 150 mM NaCl, 1% NP40, 2 mM EDTA, 10% Glycerol, and protease inhibitor mixture). Two hundred μg of clarified and precleared cell lysate proteins were incubated overnight with immunoprecipitating antibody at 4°C, the resulting immune complexes were captured by protein A- or G-sepharose and analyzed by western blotting, using specific detection antibodies.
Transfection of primary HMVEC-d and HFF cells with siRNA was performed using a Neon Transfection System (Invitrogen) according to the manufacturer’s instructions. Briefly, subconfluent cells were harvested and washed once with 1x phosphate-buffered saline (PBS) and resuspended at a density of 1 x107 cells/ml in resuspension buffer R (provided by the company). Ten microliters of this cell suspension was mixed with 100 pmol of Si-RNA and then microporated at room temperature using a single pulse of 1,350 V for 30 ms for HMVEC-d and 1,700 V for 20 ms for HFF. After microporation, cells were distributed into complete medium and placed at 37°C in a humidified 5% CO2 atmosphere. At 48 h post-transfection, cells were analyzed for knockdown efficiency by Western blotting. All Si-RNA oligonucleotides (siGenome SMARTpool) for BRCA1, IFI16 and non-targeting Si-RNA pool no. 2 were purchased from Thermo Scientific (catalog no. M-003461-02-0010, M-020004-01-0010, and D-001206-14-20).
HMVEC-d cells were seeded on glass 8-well chamber slides (Nalgene Nunc International, Naperville, IL) and uninfected or KSHV infected (30 DNA copies/cell or latently infected) cells were fixed for 15 min with 4% paraformaldehyde, and permeabilized with 0.2% Triton X-100 for 5 min. Cells were then washed and blocked with Image-iT FX signal enhancer (Life Technologies) for 20 min. The cells were reacted with primary antibodies against the specific proteins, followed by fluorescent dye-conjugated secondary antibodies. To detect EdU labeled viral genome, cells were fixed, permeabilized and blocked with Image-iT FX signal enhancer (Life Technologies) for 20 min. A CLICK reaction was performed for 30 min at RT using Click-iT EdU reaction additive (Life Technologies), EdU reaction buffer, copper sulphate and Alexa Fluor 594 azide. Cells were observed by Nikon Eclipse 80i microscope, and analyzed with Metamorph digital imaging software. All experiments were performed three independent times and three different fields with a minimum of 20 cells were analyzed. All images were acquired at 40 X magnification.
BRCA1 (Human) Gateway V5-tagged lentiviral expression vector (pLX304) was from DNASU Plasmid Repository (Arizona State University Biodesign Institute, Tempe, AZ). BRCA1 lentivirus was produced using a four plasmid transfection system as previously described [42]. Briefly, 293T cells were transfected with BRCA1 expressing vector and packaging plasmids and the media was changed 16 h after transfection. Supernatants containing packaged lentivirus were collected at 48 h, passed through a 0.45 μm filter and used to transduce BRCA1 negative cells (HCC 1937) in the presence of polybrene (5 μg/ml, Pierce, Rockford, IL). 48 h post-transduction with control lenti and BRCA1 lenti expression vectors, cell lysates were processed for IP and Western blotting.
Secretion of IL-1β was detected using Ray Bio Human IL-1β ELISA Kit (Ray Biotech, Norcross, GA) according to the manufacturer’s instructions. Supernatant from uninfected or virus infected cells (~3X105) was collected at different times p.i., incubated in the assay wells overnight at 4°C, washed, incubated 1 h with the biotinylated antibody, and with HRP conjugated streptavidine solution at room temperature for 45 min. Wells were washed, incubated with TMB substrate in the dark for 30 min at room temperature followed by addition of stop solution, and the absorbance at 450 nm was read using a Synergy2 Biotek Plate Reader (Biotek, Winooski, VT).
IFN-β secretion levels were measured using the Verikine Human IFN Beta ELISA kit (PBL, Interferon Source, Piskataway, NJ) according to the manufacturer’s instructions. Supernatants from uninfected or virus infected cells (~3X105) were collected at different times post-infection, diluted 1:1 with sample dilution buffer and incubated in the assay wells at room temperature for 1 h [9].
A DUOLink PLA kit (Sigma) was used to detect protein–protein interactions. HMVEC-d and HFF cells were cultured and infected with KSHV (30 DNA copies/cell) in 8 chamber microscope slides, fixed in 4% PFA for 15 min at room temperature, permeabilized with 0.2% Triton X-100 and blocked with Duolink blocking buffer for 30 min at 37°C. TIVE and TIVE-LTC cells were fixed, permeabilized and blocked with Duolink blocking buffer as described above. Equal numbers of BJAB, BCBL-1, Akata, Ramos and LCL cells were washed extensively with PBS by centrifugation at 200xg at 4°C and spotted on 10-well glass slides and were fixed/permeabilized with prechilled acetone and blocked with Duolink blocking buffer. Cells were incubated with primary antibodies diluted in Duolink antibody diluents for 1 h, washed and then further incubated for another 1 hr at 37°C with species specific PLA probes (PLUS and MINUS probes) under hybridization conditions and in the presence of 2 additional oligonucleotides to facilitate hybridization of PLA probes only if they were in close proximity (<40 nm). A ligation mixture and ligase were then added to join the two hybridized oligonucleotides to form a closed circle. Several cycles of rolling-circle amplification using the ligated circle as a template were performed by adding an amplification solution to generate a concatemeric product extending from the oligonucleotide arm of the PLA probe. Lastly, a detection solution consisting of fluorescently labeled oligonucleotides was added, and the labeled oligonucleotides were hybridized to the concatemeric products. The signal was detected as a distinct fluorescent dot in the Texas red or FITC green channel and analyzed by fluorescence microscopy. Specificity of PLA was determined by using negative controls consisting of samples treated as described but with only secondary antibodies. An additional series of negative control PLA was conducted using each single species primary antibody followed by incubation with both species secondary antibodies. PLA was quantified using DUOLink software.
For in situ double sequential PLA of IFI16-BRCA1 and IFI16-ASC complexes, two independent PLA reactions were performed sequentially. Briefly, the PLA reaction for IFI16 and BRCA1 was performed first using rabbit anti-IFI16 and mouse anti-BRCA1 antibodies and detected by DUOLink red detection agent. Cells were then washed, blocked with PLA blocking buffer and subjected to a second PLA reaction using mouse anti-IFI16 and goat anti-ASC antibodies and detected with DUOLink green detection agent.
For EdU-labeled genome (chromatin) pull down, we used methods described by Kliszczak et al. [21] with minor modifications. Briefly, HFF cells (8 X 106) pretreated with control Si-RNA or BRCA1 Si-RNA for 72 h were infected with unlabeled or EdU labeled HSV-1 (10 pfu/cell) for 1 h and then cross-linked with 1% formaldehyde for 10 minutes at 4°C. Unreacted formaldehyde was quenched with the addition of 0.125 M glycine for 10 minutes at 4°C. Cells were harvested and permeabilized with 0.1% (v/v) Triton X-100 in PBS for 10 minutes on ice and washed with PBS. Biotin was attached to EdU genome via Click reaction with sequential addition of the following reagents, 10 mM (+)-sodium-L-ascorbate, 0.1 mM biotin-TEG azide and 2 mM copper (II) sulfate followed by 30 minutes incubation in the dark at room temperature and by addition of 10 volumes of 1% (w/v) BSA, 0.5% (v/v) Tween 20 in PBS for 10 minutes. After three PBS washes, soluble proteins were extracted in 500 μl CL lysis buffer (50 mM HEPES, pH 7.8, 150 mM NaCl, 0.5% (v/v) NP-40, 0.25% (v/v) Triton X-100, 10% (v/v) glycerol) with protease inhibitors by end-over-end mixing at 4°C for 10 minutes and then slow speed centrifugation (1800 rpm / 300 x g). The pellet was washed with 500 μl wash buffer (10 mM Tris-HCL pH 8.0, 200 mM NaCl, 0.5 mM DTT) for 10 minutes at 4°C by end-over-end mixing. Following low speed centrifugation, the pellet was then resuspended in 500 μl RIPA buffer (10 mM Tris-HCl, pH 8.0, 140 mM NaCl, 1% (v/v) Triton X-100, 0.1% (v/v) Na-Deoxycholate, 0.1% (w/v) SDS) with protease inhibitor cocktail and processed for shearing of the chromatin via sonication on ice at an amplitude of 40, 10 seconds on, 10 seconds off for 10 minutes. The extract was clarified by centrifugation at 15,000 x g for 10 minutes at 4°C. Protein content was quantitated and 30 μg of the supernatants were used for input western blotting. 1 mg of the extract was used for pull down with 50 μl of streptavidin magnetic beads which were washed with wash buffer, equilibrated with RIPA buffer and blocked overnight at 4°C with 0.5 mg/ml BSA and 0.4 mg/ml presheared salmon sperm DNA to minimize non-specific binding. Next day, beads were washed with wash buffer and incubated with chromatin extracts for 8 h at 4°C. Beads with bound complexes were then washed with wash buffer and subjected to reverse protein-DNA cross-linking and elution of proteins by incubation with 1X Laemmli sample buffer for 10 minutes at 95°C before SDS-PAGE. For DNA purification, complexes from beads were eluted in elution buffer (1% SDS; 0.1M NaHCO3). The cross-linking was reversed by treatment with 0.1 mg/ml RNase A and 0.3 M NaCl at 37°C for 30 minutes followed by incubation at 65°C for 2 h with 0.1 mg/ml Protenase K and DNA was column purified via Qiagen DNA extraction kit according to manufacturer’s instructions.
Data are expressed with means ± SD of at least three independent experiments (n≥3) using a Student’s T-test. In all tests, p<0.05 was considered statistically significant. Experiments in which p is <0.05 are marked with a single asterisk and p<0.01 are marked with double and p<0.001 with triple asterisks.
|
10.1371/journal.pcbi.1003273 | Molecular Mechanical Differences between Isoforms of Contractile Actin in the Presence of Isoforms of Smooth Muscle Tropomyosin | The proteins involved in smooth muscle's molecular contractile mechanism – the anti-parallel motion of actin and myosin filaments driven by myosin heads interacting with actin – are found as different isoforms. While their expression levels are altered in disease states, their relevance to the mechanical interaction of myosin with actin is not sufficiently understood. Here, we analyzed in vitro actin filament propulsion by smooth muscle myosin for -actin (A), -actin-tropomyosin- (A-Tm), -actin-tropomyosin- (A-Tm), -actin (A), -actin-tropomyosin- (A-Tm), and -actin-tropomoysin- (A-Tm). Actin sliding analysis with our specifically developed video analysis software followed by statistical assessment (Bootstrapped Principal Component Analysis) indicated that the in vitro motility of A, A, and A-Tm is not distinguishable. Compared to these three ‘baseline conditions’, statistically significant differences () were: A-Tm – actin sliding velocity increased 1.12-fold, A-Tm – motile fraction decreased to 0.96-fold, stop time elevated 1.6-fold, A-Tm – run time elevated 1.7-fold. We constructed a mathematical model, simulated actin sliding data, and adjusted the kinetic parameters so as to mimic the experimentally observed differences: A-Tm – myosin binding to actin, the main, and the secondary myosin power stroke are accelerated, A-Tm – mechanical coupling between myosins is stronger, A-Tm – the secondary power stroke is decelerated and mechanical coupling between myosins is weaker. In summary, our results explain the different regulatory effects that specific combinations of actin and smooth muscle tropomyosin have on smooth muscle actin-myosin interaction kinetics.
| Dependent on the required physiological function, smooth muscle executes relatively fast contraction-relaxation cycles or maintains long-term contraction. The proteins driving contraction – amongst them actin, tropomyosin, and the contraction-driving myosin motor – can show small changes in the way they are constructed, they can be expressed as different “isoforms”. The isoforms are supposedly tailored to support the specific contraction patterns, but for tropomyosin and actin it is unclear exactly how the isoforms' differences affect the interaction of actin and myosin that generates the muscle contraction. We measured actin movement outside the cellular environment, focusing on the effects of different isoform combinations of only actin, myosin, and tropomyosin. We found that the actin isoforms cause differences in the mechanical interaction only when tropomyosin is present, not without it. Also, all different actin-tropomyosin combinations affected the mechanical interactions in a different way. In our experiments we could not directly observe the mechanical interactions of actin, tropomyosin, and myosin, so we reconstructed them in a mathematical model. With this model, we could determine in detail how the different actin-tropomyosin combinations caused the differences that we observed in our experiments.
| Differential expression of smooth muscle contractile proteins has been associated with organismal development [1], contractile phenotypes [2]–[4], and pathologies, e.g. preterm labour, hypertrophic bladder, or airway hyper-responsiveness [5]–[7]. While the role of the smooth muscle myosin isoforms has been extensively investigated [7]–[9], the functional implications of the differential expression of specific actin and actin-regulatory protein isoforms remain elusive [4].
In smooth muscle, actin isoforms are expressed from four different genes, yielding “vascular muscle” - and “enteric muscle” -actin, as well as non-muscle (cytoplasmic) - and -actin. The muscle isoforms are associated with the contractile apparatus, the non-muscle isoforms with cytoskeletal structures [5]. Muscle -actin is generally associated with tonic, -actin with phasic smooth muscles [5], [10], [11]. An anti-proportional relationship between the absolute levels of - and -actin has been established [2]. Disease-related expression differences in - vs. -actin have been found [6]. Functional differences between - and -isoforms were searched for in molecular mechanics experiments, but, to our knowledge, no differences were detected [12]–[15]. Insight from tissue level mechanics seems lacking, too [4].
Smooth muscle tropomyosin affects the weak to strong binding of ATP-activated myosin to actin: tropomyosin can be in an ON state supporting myosin strong binding, or an OFF state hindering myosin strong binding [10], [16]. When regulated by caldesmon-calmodulin, dependent on the caldesmon-calmodulin activation state, smooth muscle tropomyosin is stabilized in the open or the closed state, increasing or decreasing the rate of myosin cycling compared to the rate without any tropomyosin being present [10], [17]. Tropomyosin forms chains along actin filaments by a head-to-tail overlap of consecutive tropomyosin molecules. This overlap leads to an increased cooperativity in the switching between the ON and the OFF state. Compared to striated muscle tropomyosin isoforms, a stronger cooperativity between tropomyosin displacement due to stronger end-to-end binding between tropomyosin molecules is observed, as well as a greater bias for the ON conformation [16], [18], [19]. Similar to striated muscle tropomyosin, smooth muscle tropomyosin facilitates cooperative binding of myosin to actin: above a critical ratio of myosin heads per actin monomer, myosin heads cooperatively displace tropomyosin into the ON state so that further myosin binding is facilitated; below a critical density or activation by phosphorylation, tropomyosin remains mostly in the OFF state [20], [21] and inhibits myosin cycling [10], [19].
Tropomyosin is expressed from the same two genes in non-muscle, striated muscle, and smooth muscle cells. In smooth muscle, alternative splicing yields two smooth muscle specific isoforms (tropomyosin- and tropomyosin-), one from each gene [22]. In vivo, tropomyosin- and tropomyosin- mostly occur as heterodimers, making functional differentiation between the isoforms difficult [10], [22]. In disease states, however, expression differences between both isoforms can be observed [6], raising the question of functional differences between these two isoforms, especially in interaction with other differentially expressed contractile protein isoforms. Crystallized N-terminal fragments of tropomyosin- and tropomyosin- displayed differences in the heterodimerization properties of tropomyosin- vs. tropomyosin- and a greater head-to-tail overlap of tropomyosin- than that of tropomyosin- [23]. These structural results were interpreted as indication of negligible differences in tropomyosin's interface for actin binding and more important differences in the surfaces available for mediation of actin-myosin interactions as well as the binding of other proteins [23]. However, actin affinity (in terms of binding constants) of smooth muscle tropomyosin- was found to be times greater than that of tropomyosin- [24], [25].
In this study, we use an in vitro motility assay to investigate differences in the propulsion of “vascular” -actin vs. “enteric” -actin by smooth muscle myosin in the presence of smooth muscle tropomyosin-, tropomyosin-, or in the absence of tropomyosin, see Fig. 1 A and Tab. 1. We develop and simulate a mathematical model to establish the differences in actin-myosin interaction kinetics that underlie the experimentally observed differences.
Using our specifically developed analysis software, we extracted the following features of actin sliding: mean sliding velocity (), the motile fraction (), the average run time (), and the average stop time () (Fig. 1 B, C). These features were extracted for the different experimental conditions (Tab. 1) and resolved by actin filament length () (Fig. 2). For A-Tm a consistent increase is apparent (Fig. 2 A). , , and do not immediately suggest consistent differences, (Fig. 2 B–D). In spite of high filament counts (Fig. 2 D, inset), the width of the confidence intervals compared to potential differences makes a direct, conclusive inference difficult, especially for and at .
The resolved features represent a simultaneous measurement of values, whose interdependence cannot be judged a priori. We applied a Principal Component Analysis (PCA) to reduce the dimensionality of our data and remove correlations between values, which would otherwise inflate statistical significance. Transformation into the three Principal Components (PCs) explaining most of the variance indicates that consistent differences between the experimental conditions exist (Fig. 3 A, B). Our statistical analysis detected no differences between A, A, and A-Tm, which will therefore be referred to as baseline conditions that show no effect; A-Tm, A-Tm, and A-Tm are all different from the baseline conditions, as well as from each other (Fig. 3 C, D). To support the conclusions from our statistical analysis, we executed a hierarchical cluster analysis. Based on the relatively large reduction of linkage when going from four to five clusters, a number of four clusters was chosen (Fig. 3 E). In the PC space, the four clusters appear similar to the above separation into one baseline and three regulated conditions (Fig. 3 F, G). Indeed, the four clusters form a clear representation of the A, A, A-Tm baseline conditions, and the three distinctly regulated conditions A-Tm, A-Tm, and A-Tm, (Fig. 3 H). Thus, two independent methods of statistical assessment indicate that only A-Tm, A-Tm, and A-Tm are significantly regulated, while for each of them the regulation affects actin sliding in the in vitro motility assay in a distinctly different manner (Fig. 3 I).
Next, we wanted to attribute the differences that had been detected using PCA to molecular mechanical features. Thus, we evaluated the motility features' fold changes relative to A, averaged over . For A-Tm, is statistically significantly increased to 1.12 times the baseline value (Fig. 4 A). For A-Tm, is decreased to 0.96-fold, is increased by a factor of 1.6 relative to the baseline value, though both changes show up only as strong trends (Fig. 4 B, C). For A-Tm, is elevated 1.3-fold, which also shows up as a strong trend only (Fig. 4 D). When and are analyzed together, the joint fold changes for A-Tm become statistically significant (Fig. 4 E). When only short actin is considered, is statistically significantly elevated to 1.7 times the baseline value (Fig. 4 F). Note that each condition's differences are found in different features, which is coherent with the PCA finding that the regulated conditions are each affected by tropomyosin in a distinct manner.
To theoretically understand the regulatory effect that tropomyosin has on actin-myosin interactions, we constructed a mathematical model of the kinetics of a myosin-coated surface interacting with actin filaments of different length . Stochastic simulations of our model produce time courses (Fig. 1 C). Averaging these time courses gives , all other features of actin sliding can be extracted in exactly the same way as from experimental data. Our model is an extension of our earlier model of the group action of myosins propelling actin filaments in the in vitro motility assay [26]. Briefly, the model assumes that myosin moves actin by two mechanical steps, the main power stroke and a secondary mechanical step preceding myosin detachment [27], [28]. When several myosins are simultaneously bound to the same actin filament, they are mechanically coupled via the filament. Thus, the individual myosins' steps cause a change in the mechanical configuration of the overall system of bound myosins and the actin filament. Consequently, mechanical work might have to be exerted on or might be released from the actin-myosin system during the execution of an individual myosin's mechanical step. This mechanical work affects the strain-dependent rates of both mechanical transitions, the main power stroke and the secondary pre-detachment step. The overall number of myosin binding sites that are accessible on a given actin filament () is assumed to be proportional to . Using the helix repeat of actin (0.0355 μm) as an approximate binding site distance [27], [29], the ranges were adjusted to correspond to the ranges used in the different analysis steps. For details regarding our mathematical model, see Text S1.
A set of model parameters was determined to mimic the baseline condition (Fig. 5). These baseline parameters were altered so as to mimic the changes in resolved features that were observed experimentally for the A-Tm, A-Tm, and A-Tm conditions (Fig. 5). The scalar fold changes in motility features were determined in the same way as from the experimental data (Fig. 6). The resolved motility features as well as the fold changes capture the experimentally observed differences between the baseline conditions and the conditions that exhibited statistically significant effects.
The changes in model parameters that were necessary to mimic the experimentally observed differences point towards the aspects of actin-myosin interaction kinetics that are changed in the different conditions (Fig. 7). For A-Tm, all kinetic rates (, , ) are increased 1.15-fold. For A-Tm, the impact of mechanical coupling between myosins on the rate of the mechanical transitions () is increased by a factor of 1.2. For A-Tm, is reduced to of the baseline value, and is reduced to of the baseline value.
We investigated in vitro the relevance of actin and smooth muscle tropomyosin isoforms to the mechanical action of smooth muscle myosins on actin. In accordance with prior studies [4], [12]–[15], no differences between actin isoforms could be detected. However, the sequence differences between actin isoforms are confined to regions of interaction with regulatory proteins [30], suggesting potential mechano-chemical differences in the presence of such regulatory proteins. In vitro studies in solution (i.e. not on a motility surface) showed a different binding affinity between actin and smooth muscle tropomyosin [24], [25]. Here, we establish that, in the presence of both tropomyosin- and tropomyosin-, the molecular mechanics differ between - vs. -actin. Thus, the sequence differences between actin isoforms not only affect actin-tropomyosin interactions, but also actin-myosin mechano-chemistry. Importantly, we found that -actin is significantly regulated only by tropomyosin-, while -actin is regulated by both tropomyosin- and tropomyosin-.
More specifically, the regulation by tropomyosin has distinct effects on in vitro molecular mechanics in three regulated actin-tropomyosin combinations (experimentally determined), suggesting three different modes by which tropomyosin regulation affects actin-myosin mechano-chemistry (determined by model parameter adjustment): (1) A-Tm – is increased 1.2-fold. This is caused by a 1.15-fold increase in the myosin attachment rate to actin, the unstrained myosin main power stroke rate, and the unstrained rate of detachment of unloaded myosin from actin. (2) A-Tm – is reduced to 0.96-fold and is increased 1.6-fold. This is caused by an increase in the impact that myosin-to-myosin mechanical coupling has on rates of mechanical steps of myosin by a factor of 1.2. (3) A-Tm – is increased 1.7-fold for short actin. This is caused by a decrease in the unstrained rate of detachment of myosin from actin to 0.75 times the baseline value and a decrease to 0.8-fold in the impact that myosin-to-myosin mechanical coupling has on rates of mechanical steps of myosin.
Note that no quantitative adjustment, e.g. minimization of sum of squared errors, was used to determine the model parameter changes stated above. In consequence, the numeric parameter changes stated above should be understood as qualitative indicators of the general nature of changes in actin-myosin interaction kinetics.
The changes in kinetic parameters determined for A-Tm using our model-based assessment are in line with what is known for this condition from ATPase assays with skeletal muscle myosin and actin. Sobieszek determined that gizzard smooth muscle tropomyosin increases the ATPase , while the affinity of myosin for the actin-tropomyosin complex was not affected at myosin∶actin ratios of less than one myosin head per 4 to 6 actin monomers – which is the relevant regime for our experiment [31]. These observations were attributed to increases in the rates of the kinetic steps after myosin binding to the actin-tropomyosin complex, which is concurrent with the general increase in the unstrained kinetic rates we observed for A-Tm. Williams et al. found results that are similar to Sobieszek's and were measured at low myosin concentrations and low ionic strengths corresponding to those used in our motility assays [18].
Sufficient evidence exists to state that smooth muscle tropomyosin does regulate smooth muscle myosin interactions with actin, and thus, the resulting molecular mechanics [10], [20], [21], [32]. Regarding the functional relevance of the smooth muscle tropomyosin isoforms, however, several not mutually exclusive mechanisms by which the isoforms could affect molecular mechanics have been put forward [22]:
With regards to smooth muscle contraction, smooth muscle myosin is the most central interaction partner of actin. We investigated its mechanical action on actin in the background of different actin and tropomyosin isoforms' interaction. Because we found that tropomyosin isoforms are indeed relevant to the regulation of actin-myosin interactions, all three mechanisms are possible for actin-tropomyosin-myosin interactions. However, the observed difference between the tropomyosin isoforms depends on the actin isoform. This suggests direct interactions between the actin filament and tropomyosin, highlighting the second mechanism.
Our mathematical model does not include tropomyosin-mediated myosin binding cooperativity. Binding cooperativity is often assessed by changing the myosin-actin ratio or the myosin activation level [10], [19]–[21]. Within the scope of this study, one detectable effect of binding cooperativity differences would be a shift in the actin length at which bifurcations between non-motile and motile behavior occur [26]. These bifurcation lengths depend on the number of myosins effectively bound to actin and would be affected by cooperativity-mediated changes in the effective rate of myosin binding to actin. We found no significant shifts in these lengths between the conditions, and therefore no indication of differences in binding cooperativity.
Like any automation of a manual analysis procedure, our video analysis software makes the analysis of large data sets feasible and prevents differences occurring between different days or operators. A specific advancement is the automated machine learning-based approach to quality control of the filament traces. Further, a result management framework was devised, which allows keyword-based queries into annotated data sets and the application of custom analysis functions. Utility functions allow the creation of customized MatLab scripts to interact with results. This supports customized analyses of existent data sets also by computational scientists without their own motility assays, as well as the “high throughput” necessary for determining statistical distributions and resolved curves of motility features. The MatLab scripts with instructions are released as open source (In Vitro Motility Assay Automated Analysis – ivma3, http://code.google.com/p/ivma3/). FIESTA is another openly accessible analysis software that can be used for in vitro motility assays [35]. It reaches nanometer precision and allows interactive assessment of filament motility in a graphical user interface. Differently, our software provides less precise image analysis and tracking at the benefit of fast processing of a high number of experiments and the possibility to execute specific analyses on large data sets in an automated fashion.
The statistical assessment uses bootstrapping to maintain the high filament count that is necessary for a high resolution while still giving account of the variation present in the experiment. To explore the results and counteract inflation of statistical significance resulting from resolved analysis, PCA was used on the bootstrapped data sets. We could not find existent examples of this combination of PCA and bootstrapping – other studies estimate the variation of PCA itself [36], [37], or assess the variation of bootstrap scores (loadings) [38], [39].
More detailed assessment of in vitro motility and the observed specificity of regulation require more specific theoretical explanations of the molecular mechano-chemistry underlying these observations. Our relatively simple stochastic model generated data sets that were analyzed in the same way as actual experimental data, indicating how the different actin and tropomyosin isoform combinations affect actin-myosin interaction kinetics. While providing a perspective beyond mere presentation of our experimental findings, the simplicity of our model as well as the procedure by which model parameters were adjusted to mimic the experimental observations call for future work. From an experimental perspective, molecular mechanical assays using expression and site-directed mutagenesis of actin and tropomyosin seem promising.
We developed an automated video analysis software which executes the following steps. Raw video data are preprocessed (image enhancement and frame merging to a time resolution of s) and turned into binary images. Filament objects and their properties are extracted from individual frames using connected components methods. Filaments are tracked throughout consecutive frames based on their centroid position and area. Frame-to-frame velocities () are calculated from centroid displacements between two consecutive frames. Filament length () and travelled path lengths are determined based on a transformation of image objects into rectangles of same area and perimeter, the longer edge representing lengths. A filament's mean trace velocity () is determined by dividing the total distance that the filament's tip has travelled by the time the filament was present for (). Filament traces with filament crossing events or signs of irregular motion were removed by a machine-learning algorithm, which was trained on subset of our data that we scored by hand. The automated video analysis was assessed using computer-generated mock motility videos, the automated quality control was evaluated against hand-scored data sets. For details see Text S1.
Statistical significance was assumed for . Statistical comparisons were executed by bootstrapping of the compared statistic; statistical significance was assumed where no overlap exists between the confidence intervals of the compared conditions. For details see Text S1.
|
10.1371/journal.ppat.1003580 | Polydnaviral Ankyrin Proteins Aid Parasitic Wasp Survival by Coordinate and Selective Inhibition of Hematopoietic and Immune NF-kappa B Signaling in Insect Hosts | Polydnaviruses are mutualists of their parasitoid wasps and express genes in immune cells of their Lepidopteran hosts. Polydnaviral genomes carry multiple copies of viral ankyrins or vankyrins. Vankyrin proteins are homologous to IκB proteins, but lack sequences for regulated degradation. We tested if Ichnoviral Vankyrins differentially impede Toll-NF-κB-dependent hematopoietic and immune signaling in a heterologous in vivo Drosophila, system. We first show that hematopoiesis and the cellular encapsulation response against parasitoid wasps are tightly-linked via NF-κB signaling. The niche, which neighbors the larval hematopoietic progenitors, responds to parasite infection. Drosophila NF-κB proteins are expressed in the niche, and non cell-autonomously influence fate choice in basal and parasite-activated hematopoiesis. These effects are blocked by the Vankyrin I2-vank-3, but not by P-vank-1, as is the expression of a NF-κB target transgene. I2-vank-3 and P-vank-1 differentially obstruct cellular and humoral inflammation. Additionally, their maternal expression weakens ventral embryonic patterning. We propose that selective perturbation of NF-κB-IκB interactions in natural hosts of parasitic wasps negatively impacts the outcome of hematopoietic and immune signaling and this immune deficit contributes to parasite survival and species success in nature.
| Parasitoid wasps are insects whose development takes place within the body of other insects. To survive, wasp larvae must overcome the immune defenses of their hosts. How parasitic wasps overcome host immunity is not fully understood even though we know that different strategies using venoms, virus-like particles, or viruses are involved. A unique class of viruses, called polydnaviruses is found in two families of wasps that comprise more than 20,000 wasp species. The genomes of polydnaviruses encode proteins with ankyrin repeats. Ankyrin repeats are also found in Cactus, the inhibitor protein of NF-κB signaling in Drosophila. Viral ankyrin proteins, or Vankyrins, however, lack the amino acid sequences necessary for turnover found in Cactus and mammalian IκB family members. We show that Vankyrins produced by polydnaviruses of a parasitic wasp that attacks caterpillars of many common agricultural pests can block NF-κB signaling in fruit fly larvae. This inhibition supports parasite success. Our work highlights the crucial role of NF-κB signaling across insect taxa in insect-insect and insect-virus interactions. Studies of polydnaviral ankyrin proteins in Drosophila reveal that immune-suppressive viruses may block both cellular and humoral immunity in insects to win the biological ‘arms race’.
| Parasitic wasps develop within their insect hosts as they devour host bodies. Wasp oviposition in Drosophila larvae simultaneously activates humoral and cellular immune reactions. In a systemic acute inflammatory reaction, humoral antimicrobial secretions and cytokines from the fat body synergize with hematopoietic proliferation and differentiation, to encapsulate the wasp egg and protect host larvae [1], [2], [3]. Immune response against wasp eggs alters hematopoietic development in the larval lymph gland and in the hemolymph [4], [5], [6]. Genetic and molecular analysis of wasp-infected Drosophila hosts has revealed the fundamental role for the Toll-NF-κB pathway in both humoral and cellular immunity [2], [3], [7]. Toll signaling is also essential for basal hematopoiesis in the lymph gland [8], although the precise functions of the Toll effector proteins, the NF-κB family transcription factors Dorsal (dl) and Dorsal-related immunity factor (Dif), in either basal or activated hematopoiesis are not understood.
The nuclear translocation and functions of Dorsal and Dif are inhibited by their interactions with Cactus, the cognate IκB inhibitor [9], [10]. The direct physical interaction with NF-κB proteins depends on several ankyrin repeats in the IκB protein sequences [11], [12]. A Toll-dependent degradation signal is interpreted by the N-terminal regulatory domain of Cactus [13]. Indeed, chronic inflammatory defects wrought by excessive Toll activation are ameliorated by a mutant Cactus without the N-terminal domain responsible for signal-dependent degradation [14]. Interestingly, ankyrin repeat sequence motifs, homologous to Cactus, are found in the genomes of all sequenced polydnaviruses [15], [16], [17]. However, whether closely-related members of this large family of insect viral proteins support parasite development by redundantly or differentially blocking NF-κB signaling in host hematopoiesis and immunity, and mechanisms underlying such differences, is not known.
Double-stranded DNA-carrying mutualistic and pathogenic polydnaviruses (PDVs) fall into the evolutionarily distinct bracovirus (BV) and ichnovirus (IV) genera that are associated with an estimated 20,000 species of parasitic wasp families Braconidae and Ichneumonidae, respectively [18]. Each wasp species has co-evolved with a unique, vertically-transmitted PDV [19], [20] that they introduce into their Lepidopteran larval hosts upon oviposition. The polydnaviral particles express their gene products in infected tissues to ensure wasp success [18].
The viral ankyrin (vankyrin) gene family is common to both BV and IV genomes; each genome carries several members [21]. The Campoletis sonorensis IV (CsIV) genome contains seven copies of vankyrin genes, with 47% to 83% amino acid sequence identity. Because Vankyrins lack the N- and C-terminal regulatory domains of cellular IκBs, it was suggested that these proteins effectively inhibit NF-κB signaling in parasitized insects [16], [22].
In this study, we first defined new functions of Dif and Dorsal in basal and activated hematopoiesis. We then tested the inhibitory functions of two of the seven vankyrin genes of the Campoletis sonorensis Ichnovirus, CsIV-P-vank-1 (P1) and CsIV-I2-vank-3 (I3). These Vankyrins (1) are the most similar to each other, with 83% amino acid sequence identity; (2) are derived from different DNA segments (P and I) of the multipartite CsIV genome; (3) share only four of the six 33-amino acid ankyrin repeats of Cactus; (4) possess a putative functional zinc-binding motif in their N-termini not present in other PDV ankyrins; (5) and are also among the most expressed in Lepidoptera immune tissues after parasitization. We reasoned that a Vankyrin-based immune-suppressive strategy between BVs and IVs reflects the broad functional conservation of NF-κB-dependent immune responses in insects and an intrinsic ability of Vankyrins to dominantly interfere with Toll-NF-κB signaling in a context-independent manner.
In a novel application of Drosophila to examine insect-insect and insect-virus interactions, we tested if P1 and I3 might differentially block hematopoietic and immune signaling in Drosophila models of acute (Leptopilina spp. wasp egg encapsulation) and chronic (ectopic NF-κB signaling) inflammation [2]. We also examined their ability to temper the maternal NF-κB pathway essential for embryonic dorsal/ventral (d/v) axis formation. We report specific and dose-dependent inhibition of NF-κB signaling in hematopoiesis, innate immunity, and embryonic patterning by P1 and I3. These results suggest that NF-κB signaling is pervasive across taxa and offer rational means for its selective inhibition by viral-ankyrin proteins.
Sumoylation-deficient animals exhibit ectopic NF-κB signaling which correlates with persistent high levels of an active ligand for the Toll receptor (Spatzle), and low levels of Cactus protein in immune cells. These changes contribute to continuous hematopoietic overproliferation and chronic inflammation [2], [23]. Loss of the Dif and dl loci (or of dl alone) suppresses chronic inflammation and aberrant hematopoiesis of Ubc9 mutants [14]. Wasp infection of Drosophila larvae activates NF-κB functions [2], [3] and also alters hematopoietic development in the lymph gland [5], [7], [24]. These results suggested that Dorsal and Dif likely control normal hematopoiesis in the lymph gland and their activity may be subject to immune suppression by parasitic wasps. The lymph gland is a lobed organ. Anterior lobes of the third instar larva are most developed and harbor a quiescent medullary zone (MZ) with relatively undifferentiated cells, maintained by the niche or posterior signaling center, while the peripheral cortical zone (CZ) contains cells at various developmental stages, including mature immune cells [25]. The niche is specified by the homeobox transcription factor Antennapedia (Antp) [26]. The phagocytic plasmatocytes make up the majority of mature cells, while crystal cells carry pro-phenol-oxidase crystals for melanization. Specialized large adhesive cells, lamellocytes, differentiate when Toll-NF-κB signaling is hyperactive or upon wasp infection [1]. In antibody staining experiments, we found Dorsal expression throughout the lymph gland lobes with somewhat higher signal in the CZ (Fig. 1A–B′). In contrast, Dif expression is high in the Antennapedia (Antp)-expressing niche cells. Dif signal is lower and variable in the MZ/CZ regions (Fig. 1C–D″).
We utilized a D4-lacZ reporter, which contains four Dorsal/Dif binding sites [27]. In uninfected larvae, D4-lacZ expression (Fig. 2A–A′) co-localizes with Antp>GFP expression. This basal D4-lacZ expression suggests high Dorsal/Dif transcriptional activity in the niche even in uninfected animals. Upon L. boulardi infection, D4-lacZ expression is four times higher in infected compared to uninfected larvae (compare Fig. 2B′ to Fig. 2A′). Additionally, numerous cells of the anterior lobes are also positive for anti-β-galactosidase staining (Fig. 2B–B′). The basal D4-lacZ expression in the niche is not observed in glands from dl8/Df119 animals (compare Fig. 2C to Fig. 2D), although, surprisingly, these mutant glands express the D4-lacZ reporter in many cortical cells (Fig. 2C–D′). Thus, it appears that (a) consistent with NF-κB function in anti-wasp response, the D4-lacZ reporter is sensitive to and is differentially activated (in distinct cell populations) by parasitization; (b) transcriptional activity of Dorsal in the niche is essential for D4-lacZ expression; and (c) Dorsal possibly represses transcription of gene targets in the lobe cortex.
To test their individual effects on the niche and on hematopoiesis, we modulated NF-κB levels in the niche. (1) RNAi knockdown with either Antp>GFP, DifRNAi or Antp>GFP, dlRNAi did not yield a significant difference in the number of GFP-positive cells (Fig. 3A–C, F), although, unexpectedly, the intensity of the Antp>GFP signal in cells with RNAi was significantly reduced (Fig. 3A–C, G). Conversely, ectopic expression of Dif or Dorsal in the niche increased Antp>GFP expression (Fig. 3D–E, G). (2) Overexpression of either wild type protein (Antp>dl or Antp>Dif) also showed supernumerary lamellocytes in the lymph gland lobes (Fig. S1G, J–K). (3) We found no significant difference in the number of crystal cells in Antp>GFP, DifRNAi or Antp>GFP, dlRNAi glands, slight increase in Antp>GFP, Dif but no change in Antp>GFP, dl (Fig. S1A–F). However, Df(2L)119/Df(2L)J4 mutants lacking both NF-κB proteins show significantly more crystal cells in the lymph gland and in the sessile compartment, compared to heterozygous controls (Fig. 3H–L), indicating inhibitory and redundant NF-κB roles in crystal cell development. Anti-Antp staining of lobes from Df(2L)119/Df(2L)J4 glands revealed that the niche is specified in the absence of Dif and Dorsal and the Antp protein expression levels appears comparable to those in heterozygous controls (data not shown). Together, these results suggest that Dorsal and Dif can modulate Antp-Gal4 transgene expression in the niche and are required non cell-autonomously for the proportional development of crystal cell and lamellocytic lineages.
We next made flp-out clones marked with GFP alone (control) or additionally expressing the fusion protein GFP-Dorsal to test for non cell-autonomous requirement. Consistent with observations above, glands with GFP-Dorsal clones showed a reduction in crystal cells (Fig. S2A–C) and an increase in lamellocytes (Fig. S2D–F), relative to glands with control clones. In both cases, mature cells appeared GFP-negative and outside the clone boundary (Fig. S2B, E). This observation supports a non cell-autonomous requirement for Dorsal/NF-κB in the development of these lineages.
Viral protein, P1 or I3, was co-expressed with GFP using the Antp-Gal4 driver. Both P1 and I3 are not only cytoplasmic (Fig. 4A, B, C; yellow), but also nuclear in cells of the niche (Fig. 4B, C; purple). While P1 is more uniformly distributed (Fig. 4B′), I3 distribution is speckled (Fig. 4C′).
Antp>P1 expression had no detectable effect on crystal cell numbers (Fig. 4E–E1, I), on the number of cells in the lymph gland niche (Fig. 4H), or on the intensity of Antp>GFP signal (Fig. 4G). However, like Antp>DifRNAi or Antp>dlRNAi, Antp>I3 reduces the intensity of Antp>GFP (Fig. 4G), and similar to Dif− dl− mutants, Antp>I3 increases the number of circulating/sessile crystal cells in the posterior larval segments (Fig. 4F1, I). Additionally, its expression reduces the niche cell count (Fig. 4H).
Consistent with the interpretation that I3 may be able to block Dorsal/Dif function, we found, using the D4-lacZ reporter, that Antp>GFP, I3 (but not P1) reduces basal levels of β-galactosidase expression in the niche (Fig. 5A–C″). Moreover, wasp infection-induced boost in D4-lacZ expression in the lobes is also significantly inhibited by I3 with reduced β-galactosidase signal intensity (Fig. 5D–F).
To determine if Vankyrins might co-localize with Dorsal and whether their subcellular localization in relation to Dorsal changes upon wasp-infection, we examined their presence in GFP-Dorsal-expressing blood cells. The GFP-Dorsal fusion protein, when expressed alone (Cg>GFP-Dorsal) is punctuate in the cytoplasm of uninfected larval blood cells and some fusion protein translocates to the nucleus upon wasp infection (Fig. 6A–B″) [2]. I3 distribution in uninfected cells (Srp>GFP-Dorsal; Srp>I3) is also cytoplasmic and punctuate (Fig. 6C–C″ - arrowhead). P1 distribution (Srp>GFP-Dorsal; Srp>P1) in these blood cells is also largely cytoplasmic and punctuate even though more evenly distributed than I3 (Fig. 6E–E″ - arrowhead). In control cells, both I3 and P1 signals show little to no overlap with the GFP-Dorsal signal (Fig. 6C″, E″). Upon wasp infection, both GFP-Dorsal and Vankyrin signals in blood cells are higher compared to cells from uninfected animals and their distribution is variable. While the I3 signal is intensely nuclear (Fig. 6D - arrow), there is strong and clear co-localization of some I3 with GFP-Dorsal in vesicular pattern in the cytoplasm (Fig. 6D″ - yellow). Interestingly, GFP-Dorsal levels remain relatively low in the nuclei of most blood cells from wasp-challenged I3-expressing animals (Fig. 6D′). In contrast, some of the P1 signal co-localizes with nuclear GFP-Dorsal (Fig. 6F″ - white). These results suggest that, unlike their behavior in cultured Lepidopteran cells (where lipopolysaccharide or laminarin exposure alters localization from nucleus to cytoplasm [28]), in Drosophila larval blood cells, both I3 and P1 proteins respond to wasp infection by relocalizing from the cytoplasm to the nucleus. These results suggest that Vankyrins may block signaling by interaction with NF-κB proteins in different subcellular compartments.
The circulating blood cells constitute a separate hematopoietic compartment and are derived from embryonic hemocytes [29]. These cells contribute to the cellular immune encapsulation response [30]. To examine if either P1 or I3 affect larval circulating blood cell count (also referred to as circulating hemocyte concentration, or CHC [7]) or Cg>GFP expression, we monitored the CHC and GFP expression in Cg>GFP, Cg>GFP, P1, and Cg>GFP, I3 larvae and found no significant difference (Fig. S3A). In addition, there was no detectable effect of P1 or I3 on zygotic development or viability (data not shown).
We next examined if Vankyrins block immune signaling in models of acute and chronic inflammation. Exposure of fly larvae to L. victoriae elicits strong encapsulation [2]. Wasp infection induces limited blood cell division and differentiation and cells of the lymph gland and circulation are mobilized to encapsulate the wasp egg [6] [30]. This systemic immune reaction is resolved within hours after infection and is akin to acute inflammation in its requirement for NF-κB pathway components [2]. Cg>GFP larvae, with 0, 1 or 2 copies of P1 or I3 were infected with L. victoriae and levels of encapsulation were compared (Fig. 7). All four Vankyrin-expressing lines showed a significant decrease in their ability to encapsulate wasp embryos compared to the controls (Fig. 7A–B′). Lines expressing two copies of either P1 or I3 were completely immune-compromised and unable to encapsulate L. victoriae eggs (Fig. 7A′, B′).
Continuous expression of Toll10b protein leads to the growth of chronic inflammatory hematopoietic tumors; this overgrowth is supported by extra rounds of mitosis triggered by Toll10b expression (Fig. 7C–F, Fig. S3B). Co-expression of either P1 or I3 inhibited mitosis down to wild type levels (Fig. S3B), shrinking growth and abundance of these inflammatory hematopoietic tumors: both the size and number of microtumors per animal induced by Cg>GFP, Toll10b showed significant reduction (Fig. 7C–F, Fig. S3C–E′). Furthermore, melanization of some of the largest microtumors induced in Cg>GFP, Toll10b animals was reduced by Vankyrin expression (Fig. S3C–E′). These results suggest that Vankyrins block tumor growth by interfering with Toll10b-dependent pro-mitotic signals.
Two BV ankyrins (Ank-H4 and Ank-N5) from the wasp Microplitis demolitor reduce the ability of Dif to bind to κB consensus sequence in the Drosomycin promoter [17]. Biochemical studies show strong binding of Ank-H4 and Ank-N5 to homodimers of Dif and Dorsal [31]. We therefore investigated the in vivo effects of Vankyrins on the expression of Drosomycin, a direct target of Dif and Dorsal, and three ProPO genes, involved in melanization.
Drosomycin is highly induced in larvae poked manually with a glass needle, compared to unchallenged controls (t = 34, df = 5, p<0.001). In the presence of either Vankyrin, Drosomycin expression is reduced compared to challenged larvae without Vankyrins (Fig. 8A). A similar trend is observed after wasp infection, although the induction was significantly more variable (data not shown).
We also examined the effect of Toll10b on the transcription of pro-phenol oxidase-encoding genes ProPO59, ProPO54 and ProPO45 [32], [33]. ProPO59 and ProPO54 were upregulated in Cg>GFP, Toll10b animals compared to Cg>GFP (Fig. 8A), while ProPO45 was not (data not shown). Either P1 (Cg>GFP, Toll10b, P1) or I3 (Cg>GFP, Toll10b, I3) reduced ProPO59 expression compared to the control Cg>GFP levels (Fig. 8B). I3 expression (Cg>GFP, Toll10b, I3) reduced the expression of ProPO54 relative to Cg>GFP, Toll10b larvae, P1 expression (Cg>GFP, Toll10b, P1) did not have this effect (Fig. 8B′).
The Toll pathway specifies dorsal-ventral fates during early embryogenesis. Embryos lacking maternal dl become dorsalized. To examine the effects of vankyrins on d/v patterning, we took advantage of the temperature-dependent haploinsufficiency of dl [34]. At 29°C, only 47% of the embryos derived from heterozygous dl1/+ females hatch, while the remaining, unable to develop normally, show slight dorsalization (Fig. 9A). We introduced 1 or 2 copies of each Vankyrin using the maternal driver Mat-Gal4. With only one copy of either P1 or I3 in dl1/+ females, the percent hatch did not differ significantly from the control (Fig. 9A). But with two copies of P1, only 8% of the embryos hatched, and, with two copies of I3, 18% of the embryos hatched. In both cases, the degree of dorsalization of unhatched embryos is more severe such that the ventral denticle belts, markers of ventral fate, are visibly reduced (Fig. 9A, B). Although, Mat-Gal4>P1 or Mat-Gal4>I3 in wild type background do not have strong effects; P1 expression has a mild and general effect on embryonic development. These results suggest that immune-suppressive Vankyrins can also block Toll signaling in the embryo and that multiple copies of vankyrin genes in PDV genomes ensure physiological and specific inhibition.
Despite common features, mutualistic polydnaviruses of Braconid and Ichneumonid wasps derive from different ancestral viruses [19], [20]. Yet, genomes from both families encode several copies of vankyrin genes. While high sequence similarity and multiple gene copies in the PDV genome may suggest similar localization or redundant biological effects, we observed surprising differences in localization of I3 and P1, and qualitative and quantitative differences of their effects on NF-κB signaling in hematopoiesis, immunity, and development.
The role of NF-κB proteins in larval hematopoiesis has been an open question for over a decade. Here we have uncovered a novel role for NF-κB signaling in the niche, where it controls the proper proportion of mature hematopoietic lineages. This discovery is validated by inhibitory effects of I2-vank-3. Dif and Dorsal are expressed in the niche. Excessive Dorsal or Dif in the niche is sufficient to trigger constitutive differentiation and release of lamellocytes. While single RNAi knockdowns have weak effects, genetic removal of both Dif and dl encourages supernumerary crystal cells and also reduces wasp-induced lamellocytes (our unpublished results). Thus, it seems that moderate levels of NF-κB activity (not essential either for Antp expression or for the specification of a particular lineage) are required for gene expression in the niche to guarantee the correct ratio of specific lineages. Elevated or diminished NF-κB activity perturb this balance: high levels parallel conditions of wasp infection (more lamellocytes; fewer crystal cells; this study [4]) whereas low levels have the opposite effect (this study). These data synthesize our view of how wasp infection shifts the balance of NF-κB functions, favoring lineage development for egg encapsulation. The identities of NF-κB target genes and their effects on lineage commitment remain to be discovered.
This interpretation that basal versus activated hematopoieses support distinct lineage programs suggests that cells of the niche, very likely, sense the systemic environment of the larval hemocoel, and respond to the host's immune status by switching states. It is therefore reasonable to conclude that the mechanism governing this switch includes NF-κB signaling itself, whose activation state depends directly on the infection status (this study; [3], [7]). How is infection sensed by the niche? Recent work suggests that the Toll ligand, Spätzle (spz) and Spätzle-processing enzyme (SPE cleaves and activates inactive Spz) are involved. (1) Spz protein and SPE transcripts are expressed at high levels in circulating blood cells and uniformly in cells of the lymph gland lobes [2], [35], [36]. (2) Spz and SPE expression is activated by wasp infection in both compartments [2]. (3) Mis-expression of either transgene in blood cells, or even just in the fat body, induces lamellocyte differentiation and systemic inflammation [2]. (4) The niche also senses the animal's nutritional status [37]. Thus, it appears that the niche is functionally flexible and responds to hemolymph factors by reprogramming hematopoiesis.
The observation that niche-specific expression of I3 alone reduces niche cell count, reduces Antp>GFP expression, and modulates both basal and activated hematopoiesis suggests that NF-κB signaling has complex and specific transcriptional effects that directly and indirectly control multiple parameters of niche function in response to organismal physiology. A lack of inhibition by P1 suggests that I3 may be a better inhibitor of Dif, which is expressed more strongly in the niche relative to other parts of the lobe. It is not surprising, then, that immune-suppressive viruses aiding parasite survival are poised to paralyze NF-κB signaling in hematopoiesis regardless of its activation status.
It is of interest that P1 and I3 share 83% amino acid identity and yet have strikingly different biological effects: (1) In basal and activated lymph gland hematopoiesis, I3 promotes crystal cell development, P1 does not; (2) their expression in circulating cells (via Cg>Gal driver) does not alter blood cell counts, but Antp>I3 encourages crystal cell development in sessile and circulating cell compartments; (3) both I3 and P1 reduce wasp egg encapsulation; (4) both Vankyrins block Cg>Toll10b-induced mitosis and resulting tumorogenesis, antifungal peptide gene expression and proPO gene expression (although I3's inhibitory effect on ProPO54 expression is stronger than that of P1's); and (5) in embryogenesis, P1's effects on dorsalization are stronger than those of I3's. We explain these differences on NF-κB signaling by postulating pre-existing differences in concentration and localization of a particular Vankyrin and NF-κB protein(s) (fly embryos do not express Dif) and their complexes in different cell types.
Differences in sub-cellular localization between P1 and I3 offer additional clues. Both P1 and I3 are cytoplasmic and punctate in blood cells from uninfected larvae, but on infection, both proteins assume a nuclear bias, co-localizing with GFP-Dorsal. While I3 and GFP-Dorsal appear vesicular and perinuclear, P1 and GFP-Dorsal co-localize within the nucleus. Recent experiments in the fly embryo [38] show that endocytosis is central to Toll signaling. It is, thus, possible that I3 interaction in blood cells with Dorsal blocks its endocytosis and/or nuclear uptake. P1 additionally may inhibit transcriptional activation of GFP-Dorsal. We noted, in these staining experiments, significant changes not only in their sub-cellular localization in resting versus immune-active cells, but also in their amounts. We interpret that Vankyrins themselves are subject to translational and/or post-translational regulation, a conclusion that is supported by measurements in transgenic cell culture studies [28]. Specificities in translational and/or post-translational regulation in different cell types may result in different biological outcomes.
Finally, like Bracovirus ankyrins H4 and N5, that bind to Dorsal and Dif, with different affinities [31], Ichnoviral Vankyrins appear to have differential affinities for NF-κB proteins. Our data suggest that I3 may have a preference for Dif, whereas P1 may bind more strongly to Dorsal in the absence of Dif.
Parasitoid wasps make up thousands of species. Using Drosophila and its natural parasitic wasps we have shown that hematopoiesis and the cellular egg encapsulation response are tightly-linked via NF-κB signaling. NF-κB signaling is active in the niche in the absence of wasp infection, but wasp infection activates NF-κB signaling further and reprograms hematopoiesis for wasp egg encapsulation. Our data suggest that even highly identical Ichnoviral ankyrins, I2-vank-3 and P-vank-1, perturb cellular and humoral immunity with remarkable specificity to contribute to the success of their wasp, C. sonorensis. Our results predict that parasitoid infections activate both immune arms of their natural insect hosts, NF-κB-IκB interactions underlie this activation, and successful immune-suppression targets both immune arms. The Drosophila model system can be used to explore molecular functions of additional immune-suppressive molecules critical to species survival and evolution of natural communities.
Targeting immune pathways of enemies by immune-suppressive molecules is a general strategy for success among Hymenopteran insects [36]. Components in the bee venom protect bees against arthropod and vertebrate predators. Bee venom is an ancient therapy for chronic inflammation and pain relief [39]. The context-independent inhibitory effects of Vankyrins on NF-κB signaling reported here provide one clear mechanism by which anti-inflammatory effects of Hymenopteran products may be realized. In addition, inhibition of NF-κB signaling continues to be a significant area of research for strategic development of drug targets for human diseases [40], [41]. Detailed structural studies coupled with rational design of IκB-family ankyrin repeats have potential for the treatment of inflammation-based human diseases from arthritis to cancer. They can also provide the means to weaken the immune system of insect pests to improve agriculture and human health [42].
All D. melanogaster stocks were raised on standard medium at 25°C. Standard crosses were performed to obtain the desired genetic backgrounds. y w; Ubc9/CyO y+ [14] lines were used for anti-Dif and anti-Dorsal staining. Gal4 lines: Cg-Gal4, UAS-GFP (expressed in fat body, lymph gland and hemocytes), Hml-Gal4, UAS-GFP (expressed in hemocytes and lymph gland), y w; UAS-mcd8GFP; Antp-Gal4/TM6 Tb (expressed in the lymph gland niche; abbreviated as Antp>GFP), Srp-Gal4, GFP-dl (expressed in hemocytes). UAS lines: UAS-dl (S. Tanda, Ohio State Univ.), UAS-Toll10b (constitutively active Toll receptor), UAS-dlRNAi (TRiP), UAS-DifRNAi (VDRC, transformant 30579), UAS-GFP (Bloomington Stock Center) and the UAS-vankyrin and UAS-Dif lines as below. Reporter strain for Dorsal/Dif activity, D4-lacZ (A. Courey) contains four tandemly repeated Dorsal/Dif binding sites [27].
To induce flp-out clones [43], developmentally-synchronized 4-day old larvae with the hybrid flip-out and Gal4 activation system [hsp70-flp; Actin>CD2>Gal4] and UAS-GFP transgenes or those with an additional UAS-GFP-dl transgene, were heat-shocked at 37°C in a water bath for 15 min.
To examine their effects on embryonic development, Vankyrins were expressed maternally using the dl1/CyO; Mat-Gal4 strain (A. Courey, UCLA). Percentage of eggs (n = 300 or more) hatched from females with 0, 1, or 2 copies of either P1 or I3 transgenes was recorded.
Mutant strains: b dl8/CyO b; y w; Df(2L)TW119/CyO y+ and y w; Df(2L)J4/CyO y+ [14].
Vankyrin cDNAs from C. sonorensis, P-vank-1 (P1, Accession: AAX56953.1, 171 amino acids) and I2-vank-3 (I3, Accession AAX56959.1, 171 amino acids) (kindly provided by Dr B. Webb, University of Kentucky [16]), were amplified by PCR using forward primers containing a FLAG tag and an EcoR1 restriction site and a common reverse primer containing a Xba1 restriction site (Text S1).
Both cDNAs were cloned into the P-element containing vector pUAST. This vector contains five GAL4 binding sites, allowing GAL4 inducible expression of vankyrins [44]. Constructs were injected into y w embryos (Rainbow Transgenic Flies, Camarillo, California, USA). A strain bearing the UAS-Dif transgene was constructed by inserting the full-length Dif cDNA [45] into the pUAST vector and injections were done in-house.
L. victoriae or L. boulardi adults were exposed to developmentally-synchronized larvae. Two days after infection, fly larvae were dissected to score for infection and encapsulation.
Fly larvae from a 48-hour egg lay expressing either vankyrin cDNAs were infected by Leptopilina victoriae for 24 hours. After 48 hours, larvae were dissected and the number of live and encapsulated wasp larvae was recorded. To test dose response, larvae with either one copy (Cg>P1 or Cg>I3) or two (Cg>P1, P1 or Cg>I3, I3) copies of vankyrin transgenes were used. Cg-Gal4, UAS-GFP flies were used as controls.
To evaluate immune suppressive effects of Vankyrins on the Toll pathway, Cg>Toll10b (control) and Cg>Toll10b, P1 or Cg>Toll10b, I3 larvae were examined for tumor penetrance and expressivity. Third instar larvae from a 6 hour egg-lay were dissected and blood cells, aggregates and tumors from their hemolymph were stained for DNA (Hoechst 33258 – Molecular Probes) and F-actin (rhodamine phalloidin – Invitrogen). Images were acquired using a Zeiss Axioscope 2 Plus fluorescence microscope. The size and number of tumors were recorded using AxioVision LE 4.5 software.
Developmentally synchronized larvae were collected, washed and dissected for either hemolymph or lymph gland according to methods described previously [23]. Antibodies and dilutions used are as follows: β-galactosidase (chicken anti-β-Gal, 1∶200; Immunology Consultants Laboratory, Inc.), rabbit anti-phosphohistone H3 (1∶200 Upstate), mouse anti-prophenoloxidase (anti-ProPO, 1∶10; Dr. T. Trenczek, University of Giessen) or rabbit anti-ProPO2 [46], mouse anti-Dorsal (anti-dorsal 7A4, 1∶4; DHSB, Iowa [47]), rabbit anti-Dif (1∶500; Dr D. Ferrandon – IBMC, Strasbourg), mouse anti-Antp (8C11, 1∶20; DSHB, Iowa), and mouse anti-L1/Atilla (1∶10, I. Ando [48]), mouse anti-Integrin β PS (CF6G11, 1∶10; DSHB, Iowa), and mouse anti-FLAG (1∶1000; Sigma). Secondary antibodies were Cy5, Cy3 or Alexa647 anti-mouse (1∶200; Jackson Immunological and 1∶1000, Invitrogen, respectively), Cy3 anti-chicken (1∶500; Jackson Immunological) and FITC or Cy3 conjugated anti-rabbit (1∶200; Jackson Immunological). All samples were counterstained with Hoechst 33258 (Molecular Probes). Rhodamine phalloidin (Invitrogen) was used where indicated.
The cuticle patterns of embryos from wild type females or haplo-insufficient for dl and expressing 0, 1 or 2 copies of either vankyrin gene were visualized after dechorionation, clearing and mounting in Hoyer's mountant. Images were acquired using dark field optics using a Zeiss Axioscope 2 Plus fluorescence microscope.
RNA was extracted from 7 to 10 pooled third instar larvae using Trizol (Invitrogen). RNA concentration and quality was checked before treating the samples with DNase (Turbo DNA-Free, Ambion). The first cDNA strand was then synthesized using ProtoScript M-MuLV First strand cDNA synthesis kit (NEB). The volume was then completed to 50 µl. RNA was stored at −80°C while cDNA was aliquoted and stored at −20°C.
For quantitative PCR, iQ SYBR Green supermix kit (Biorad) was used as per the manufacturer's recommendations except the reactions were done in 20 µl. Primers used and PCR conditions are described in Text S1. For the qPCR, three technical and 3–4 biological repeats were performed. Transcripts levels were normalized using the ribosomal rp49 gene. Melting curves were analyzed and quantification was made by using the ΔΔCT method.
All analysis except for qPCR were performed using the R software [49]. All data were tested for normality. The non-normal data were transformed when possible or a non-parametric test was applied. qPCR data were analyzed by Student t-test using the trial version of GenEx software (http://www.biomcc.com/genex-software.html).
|
10.1371/journal.pgen.0030077 | Genetic and Physical Mapping of DNA Replication Origins in Haloferax volcanii | The halophilic archaeon Haloferax volcanii has a multireplicon genome, consisting of a main chromosome, three secondary chromosomes, and a plasmid. Genes for the initiator protein Cdc6/Orc1, which are commonly located adjacent to archaeal origins of DNA replication, are found on all replicons except plasmid pHV2. However, prediction of DNA replication origins in H. volcanii is complicated by the fact that this species has no less than 14 cdc6/orc1 genes. We have used a combination of genetic, biochemical, and bioinformatic approaches to map DNA replication origins in H. volcanii. Five autonomously replicating sequences were found adjacent to cdc6/orc1 genes and replication initiation point mapping was used to confirm that these sequences function as bidirectional DNA replication origins in vivo. Pulsed field gel analyses revealed that cdc6/orc1-associated replication origins are distributed not only on the main chromosome (2.9 Mb) but also on pHV1 (86 kb), pHV3 (442 kb), and pHV4 (690 kb) replicons. Gene inactivation studies indicate that linkage of the initiator gene to the origin is not required for replication initiation, and genetic tests with autonomously replicating plasmids suggest that the origin located on pHV1 and pHV4 may be dominant to the principal chromosomal origin. The replication origins we have identified appear to show a functional hierarchy or differential usage, which might reflect the different replication requirements of their respective chromosomes. We propose that duplication of H. volcanii replication origins was a prerequisite for the multireplicon structure of this genome, and that this might provide a means for chromosome-specific replication control under certain growth conditions. Our observations also suggest that H. volcanii is an ideal organism for studying how replication of four replicons is regulated in the context of the archaeal cell cycle.
| Haloferax volcanii is a member of the archaea, which are renowned for thriving in extreme environments. Archaea have circular chromosomes like bacteria but use enzymes similar to those found in eukaryotes to replicate their DNA. Few archaeal species have systems for genetics, and this has limited our understanding of DNA replication. We used genetics to map the chromosomal sites (origins) at which DNA replication initiates in H. volcanii. This species has a multipart genome comprising one main chromosome, three secondary chromosomes, and a plasmid. Five DNA replication origins were found and confirmed to function in vivo. All are adjacent to genes for the initiator protein Cdc6/Orc1, a common feature of archaeal replication origins. Two of the sequences are located on the main chromosome, confirming that multiple origins are often used to replicate circular chromosomes in archaea. Intriguingly, one of the origins from a secondary chromosome appears “dominant” to the principal chromosomal origin, suggesting either a hierarchy or differential usage of origins. This might reflect the different replication requirements of their respective chromosomes. Given the ease of genetic manipulation, H. volcanii holds great promise for studying how replication of four chromosomes is regulated in the context of the archaeal cell cycle.
| In all prokaryotic organisms, and in certain unicellular eukaryotes, DNA replication is thought to initiate at well-defined chromosomal sites. These origins of replication serve as assembly sites for the protein machinery that unwinds the DNA duplex and initiates bidirectional DNA synthesis [1,2]. Bacterial chromosomes typically carry a single replication origin (oriC) and cognate initiator protein (e.g., DnaA), whereas eukaryotic chromosomes contain large numbers of replication origins, which are bound by a multiprotein origin recognition complex (ORC). Archaea use DNA replication proteins similar to those of eukaryotes but have circular chromosomes like bacteria [3]. Relatively little is known about origin utilization in archaea, and the available data suggest major differences in how chromosomes are replicated in the key archaeal groups. In particular, the chromosome of Pyrococcus abyssi (Euryarchaeota) is replicated from a single oriC [4], whereas three different oriCs are used to replicate the single chromosome of Sulfolobus species (Crenarchaeota) [5,6].
Archaeal replication origins consist of a long intergenic sequence containing an A/T-rich duplex unwinding element (DUE), which facilitates the local duplex opening required for replication fork assembly. The intergenic region is typically located upstream of a cdc6/orc1 gene, which encodes a putative initiator protein that is homologous to both a subunit (Orc1) of eukaryotic ORC and the helicase loader Cdc6. Protein complexes formed by archaeal initiator proteins could therefore have a dual function in origin recognition and loading of minichromosome maintenance (MCM) helicase at the origin. However, available biochemical data are consistent with a role in origin recognition only. The intergenic region of the replication origin also carries multiple conserved sequence elements (origin recognition boxes, ORB) that are bound by Cdc6/Orc1 initiator proteins [6,7]. Binding of initiator proteins at the archaeal origin has been shown to proceed in a cooperative manner [8], suggesting that in archaea a defined multimeric initiator protein complex forms at the origin. Direct support for this idea has come from recent experiments indicating that higher-order assembly of Aeropyrum pernix Orc proteins results in structural and topological changes in the origin DNA [9].
Current knowledge of archaeal DNA replication is based on biochemical observations, and genetic studies are needed to test the extant models. In this respect, haloarchaea are ideal model organisms since they are easily cultured and amenable to genetic manipulation [10,11]. The genome sequences of Halobacterium sp. NRC-1 and Haloarcula marismortui have revealed that haloarchaea contain multiple replicons, each with essential genes [12,13]. However, the mechanisms that coordinate the replication of these multiple replicons remain unknown.
The genome of Haloferax volcanii has a multireplicon structure consisting of a main chromosome of 2.9 Mb and four smaller replicons (pHV1 [86 kb], pHV2 [6.4 kb], pHV3 [442 kb], and pHV4 [690 kb]) [14]. Like other haloarchaea, the H. volcanii genome encodes numerous putative cdc6/orc1 genes that are distributed amongst the different replicons (except pHV2) (Table 1). This suggests that lineage-specific duplication of replication origins and/or initiator proteins has driven dynamic genome evolution in haloarchaea. However, previous experimental studies have identified only one likely replication origin on the main chromosome of the related haloarchaeon Halobacterium sp. NRC-1 [15].
Analysis of the genome sequence of H. volcanii (Hartman et al., unpublished data) suggested the presence of numerous replication origins, prompting us to search for autonomously replicating sequence (ARS) elements corresponding to each replicon. Replication initiation point (RIP) mapping was used to confirm that these ARS elements are functional in their chromosomal context. Our genetic data suggest a genome-wide hierarchy of some ARS elements, raising the possibility of chromosome-specific origin regulation in halophilic archaea. This study also provides a framework for investigating how the coordinated replication of four replicons is achieved in halophilic archaea.
Almost all origins of DNA replication identified in archaea to date are adjacent to genes coding for a homologue of eukaryotic Cdc6 and Orc1 proteins [16]. In order to identify putative replication origins in H. volcanii, we performed a TBLASTN search of the genome sequence of H. volcanii strain DS2 (http://www.tigr.org), using the consensus Cdc6/Orc1 sequence (COG1474.1) as a query. A total of 14 potential cdc6/orc1 genes were found on the five replicons (Table 1), although one (orc11) is missing sequences coding for the C-terminal winged-helix domain that is required for DNA binding [17]. The large number of cdc6/orc1 genes found in the H. volcanii genome parallels the situation in other halophilic archaea; ten cdc6/orc1 genes are found in Halobacterium sp. NRC-1 (three replicons) and 17 are present in H. marismortui (nine replicons) [12,15]. Three of the ten putative origins in Halobacterium sp. NRC-1 were examined in a previous study, and only one (associated with orc7) proved capable of supporting autonomous replication of a plasmid [15]. Thus, it is unlikely that all the cdc6/orc1 genes in H. volcanii are adjacent to active origins of DNA replication.
To avoid any a priori bias regarding the location of replication origins, we carried out a genetic screen for ARS elements from H. volcanii. A partial HpaII digest was performed with genomic DNA from H. volcanii strain WR340 [18], which like all laboratory strains of H. volcanii lacks the smallest replicon pHV2 [19,20]. DNA fragments of between 4 and 8 kb were cloned in the nonreplicating pyrE2-marked plasmid pTA131 [21] and used to transform the ΔpyrE2 ΔradA strain H49 to prototrophy for uracil (pyrE2+). H. volcanii radA mutants such as H49 are defective in homologous recombination [22]; this precaution was taken to prevent integration of the plasmid into the genome by homologous recombination.
A total of 35 transformants were obtained. Plasmid DNA from seven transformants was sequenced, and in all cases, the insert corresponded to the same sequence on contig number 454 (Figure 1A). Since H. volcanii has more than one replicon, it was striking that only one ARS element was recovered. To ensure this was not a technical artifact of the screen, we repeated the library construction using AciI to generate the partial digest of H. volcanii genomic DNA and transformed H49 as before. Plasmid DNA from six AciI library transformants was sequenced, and all contained the same region of the H. volcanii genome isolated in the initial screen. This region comprises two divergently transcribed genes including orc10, separated by a 910-bp intergenic region featuring a 104-bp A/T-rich putative DUE (68% A/T versus 33% for overall genome) and several direct repeats (Figure 1A). Similar features are encountered at almost all characterized archaeal origins of DNA replication (e.g., [4]). However, it is notable that the nucleotide sequence of the direct repeats (Figure 1B) bears only little similarity to the ORB identified by Robinson et al. [6].
To determine the minimal sequence needed for DNA replication activity, we carried out a partial AciI digest of a representative 5-kb ARS insert that was cloned in plasmid pTA194 (Figure 1A). Fragments of 1–1.3 kb, 1.3–1.6 kb, 1.6–2 kb, 2–3 kb, and 3–4 kb were excised separately from an agarose gel and ligated with pTA131 to construct an ARS subclone library. Transformants of H. volcanii H49 were obtained with DNA from each size range. Plasmid DNA from six transformants in the 1–1.3-kb range was sequenced, and in all cases, the insert included the 910-bp intergenic region (Figure 1A). We were able to delimit the minimal origin further by amplifying a 633-bp fragment of the intergenic region using PCR (Figure 1A). This fragment was subcloned in pTA131 to generate pCN12 and used to transform the ΔpyrE2 ΔradA strain H112. Thus, the minimal origin is located in this region, and in contrast to what had been observed for Halobacterium sp. NRC-1 [15], the cdc6/orc1 gene is not required in cis for ARS activity in H. volcanii.
We used the ARS insert in pTA194 to probe a Southern blot of intact H. volcanii DNA displayed on a pulsed field gel (PFG). Two bands were observed (Figure 1C), which correspond in size to replicons pHV4 (690 kb) and pHV1 (86 kb) [14]. The intensity of the two bands was similar, suggesting that this replication origin is present on both pHV1 and pHV4 and that both replicons are present in all cells. We employed the minimal ARS element from the AciI subclone library (in pTA250) to generate low copy–number shuttle vectors with pyrE2, trpA, hdrB, and leuB selectable markers [21] (Figure S1). The principal advantage of these shuttle vectors over existing plasmids based on the origin of replication from pHV2 is their smaller size (∼4.5 kb for pHV1/4-based vectors versus ≥7.5 kb for pHV2-based vectors).
The fact that only one ARS element was recovered in the initial genetic screen suggests that this sequence might be “dominant” and thereby prevent the isolation of other origins. We hypothesized that deleting this sequence from pHV1 and pHV4 by a gene knockout system [18,21] would allow the isolation of ARS elements corresponding to origins of DNA replication on the main chromosome and pHV3. The pHV1/4 replication origin and adjacent genes were subcloned to generate pTA252, and a 1-kb fragment containing the intergenic region necessary for ARS activity was replaced by a trpA selectable marker for tryptophan biosynthesis [21]. This plasmid (pTA266, Figure 1D) was used to transform the ΔpyrE2 ΔtrpA strain H53, and integration at the orc10 locus was verified by Southern blot (Figure 1E, strain H220). Counter-selection with 5-fluoroorotic acid (5-FOA) was used to ensure loss of integrated pTA266 by intramolecular recombination and to yield a trpA-marked deletion of the intergenic region (Figure 1D). While origin deletion events (trpA+ 5-FOA-resistant cells) were obtained, they were outnumbered >1,000-fold by events leading to restoration of the wild-type (trpA– 5-FOA-resistant cells), indicating a strong bias for maintenance of the replication origin. The deletion was verified by Southern blots of a genomic DNA digest (Figure 1E, strain H230) and intact DNA displayed on a PFG (Figure 1C). The pHV1 and pHV4 replicons are still present (unpublished data) and must therefore be using alternative replication origins. Since the deletion strain H230 did not show any obvious growth defects, these alternative origins presumably act as sites of efficient DNA replication initiation. This observation prompted us to search for other ARS elements.
We repeated the screen outlined above using the deletion strain H230 as a source of genomic DNA. AciI was used for a partial digest of H230 genomic DNA, fragments of 3–5 kb were cloned in pTA131, and H49 was transformed as before. Plasmid DNA from nine transformants was sequenced, and in all cases the insert localized to a single region of the H. volcanii genome (contig number 455, which corresponds to the main chromosome). This region (Figure 2A) comprises two divergently transcribed genes including orc1, separated by a 1,360-bp intergenic region featuring two A/T-rich DUEs and several direct repeats. The nucleotide sequence of the direct repeats (Figure 2B) is 89% identical to the ORB sequence (euryarchaeal consensus 5′-GTTCCAGTGGAAAC-AAA‐‐‐‐GGGGG-3′) [6]. In addition to orc1, a number of other genes related to DNA replication and repair is found in the vicinity of the ARS (Figure 2A), encoding the DP1 exonuclease subunit of the archaeal D family DNA polymerase [23], the Hef helicase/endonuclease [24], a homologue of the bacterial UvrC nucleotide excision repair protein, an NAD-dependent DNA ligase [25], and the Hel308 helicase [26].
The ARS insert in one plasmid (pTA313) was used to probe a Southern blot of intact H. volcanii DN5A on a PFG. As expected, one band was observed (Figure 2C) that corresponds in size to the main chromosome (2.9 Mb) [14]. A 1.1-kb Sau3AI-HindIII fragment of pTA313, comprising the intergenic region only (Figure 2A), was subcloned in pTA131 to generate pTA441 and used to transform H49. Transformants were obtained with high efficiency indicating that, as with the pHV1/4 origin, the cdc6/orc1 gene is not required in cis for ARS activity. To delimit the minimal origin further, a 692-bp fragment of the intergenic region was amplified by PCR (Figure 2A), subcloned in pTA131 to generate pCN11, and used to transform the ΔpyrE2 ΔradA strain H112. The sequence was able to maintain the plasmid.
To explain why only one ARS element was recovered in the initial genetic screen, we determined whether the replication origin located on pHV1 and pHV4 might be “dominant” to other origins. The ΔpyrE2 ΔradA strain H112 was transformed with an equimolar mixture of pTA194 and pTA313 (0.5 μg each). These are ARS plasmids from the initial and secondary genetic screens, which carry the pHV1/4 and chromosomal origins, respectively. A total of 23 transformants were analyzed by Southern blotting for the presence of each ARS plasmid (Figure 3). The vast majority (21/23) contained the ori-pHV1/4 plasmid pTA194 and in most cases (17/23) this was the sole plasmid detectable. Only six transformants contained the oriC plasmid pTA313, and in most cases (4/6) pTA194 was also present. To determine the fate of ARS plasmids in transformants containing both pTA194 and pTA313, cells were propagated by restreaking, both while maintaining selection for the pyrE2 marker and without selection. In all three cases analyzed (Figure 3B, transformants 3, 5, 10), only pTA194 remained after further propagation. This was not due to greater stability of pTA194 relative to pTA313, since transformants propagated without selection showed a marked loss of pTA194, whereas pTA313 was well maintained in the transformant where it was present exclusively (Figure 3C, transformant 8). These results recapitulate the outcome of the genetic screens and indicate that the ori-pHV1/4 plasmid pTA194 is dominant to the oriC plasmid pTA313.
The stability of ARS plasmids in the absence of selection was investigated in a quantitative manner. Cultures of H112 transformed with pTA250 or pTA441, which carry the intergenic region of ori-pHV1/4 and oriC, respectively, were propagated for ∼25 generations in nonselective Hv-YPC broth. At regular intervals, the fraction of uracil+ cells (indicative of ARS plasmids) was determined by plating on selective and nonselective media. As shown in Figure 3D, the ori-pHV1/4 plasmid pTA250 is significantly less stable than the oriC plasmid pTA441; plasmid loss was calculated to be 16% per generation for pTA250 and 9% per generation for pTA441.
The leading and lagging strands of bacterial and archaeal genomes differ in their base composition [27,28], and a surplus of G over C is usually found on the leading strands of replication. Thus, abrupt changes in the strand-specific nucleotide compositions indicate the presence of a replication origin or terminus. We analyzed the G/C disparity of the largest contig (number 455) of the H. volcanii genome sequence, which corresponds in size to the main chromosome (2.9 Mbp). Two different algorithms, ORIGINX [29] (unpublished data) and Z-CURVE (Figure 4A) [30] gave similar results. The “GC-skew” of H. volcanii suggests the presence of multiple origins on the main chromosome, similar to what has been proposed for Halobacterium sp. NRC-1 [30,31]. The amplitude of the calculated H. volcanii GC-skew was similar to that of H. marismortui, but substantially smaller than was observed for the normalized GC-skews of the main chromosomes of three other halophilic archaea and a non-halophilic archaeon (Figure 4B). Whereas the oriC of the halophilic bacterium Salinibacter ruber has a well-defined GC-skew minimum at the origin, the four halophilic archaea show a GC-skew with inverted polarity, when compared to bacteria and thermophilic archaea. Therefore, the leading strand of replication contains an excess of C in haloarchaea, similar to what has been suggested for Mycoplasma species [32]. While the mechanisms that establish these skews at a molecular level are still poorly understood, it is of note that their formation seems to be independent of GC% of genomes [32]. For example, the chromosome of the haloarchaeon Haloquadratum walsbyi, which is not GC-rich (48% GC), also shows an inverted polarity.
The H. volcanii G/C disparity curve has three peaks for the main chromosome (Figure 4A): (1) A peak near the ARS element identified in genomic libraries from H230, which is associated with the orc1 gene (oriC1); (2) A peak associated with the orc11 and orc14 genes. This chromosomal region has a base composition different than the rest of the genome and contains putative viral genes such as HNHc endonuclease, VirB4, and tragVirD4 helicases or bacteriophage T4-like integrase (unpublished data). Moreover, it does not coincide with any ARS elements isolated in this work; and (3) A peak associated with the orc5 gene, suggesting the presence of a second replication origin on the main chromosome (oriC2). GC, AT, and G + C content disparities were also carried out for contigs corresponding to the smaller replicons and predicted the location of DNA replication origins for pHV1 (confirming the origin identified in the first ARS screen), pHV3, and a second origin on pHV4 (Figure 4A); all of these peaks are located near cdc6/orc1 genes. Strikingly, ori-pHV1/4 sequence could only be found on contig number 454 (pHV1) and not on contig number 452 (which should correspond in size to pHV4).
To test the possibility of a second chromosomal origin and to identify origins on the smaller replicons, we directly cloned candidate ARS elements. A region carrying the putative second chromosomal replication origin (oriC2) was isolated as a 9.3-kb NotI fragment from a genomic DNA library and was cloned in pTA416 (Figure 5A). This fragment contains two divergently transcribed genes coding for Orc5 and a putative Rad25/Xpb-related helicase, separated by a 2,094-bp intergenic region featuring an A/T-rich DUE and several direct repeats with similarity to the ORB consensus (Figure 5B). A 2.4-kb fragment comprising this intergenic region was amplified by PCR, cloned in pTA131 to generate pTA612, and used to transform the ΔpyrE2 ΔradA strain H112. Transformants were obtained with high frequency, indicating that this region supports ARS activity (Figure S2). However, after 15 d of incubation the colonies were significantly smaller than those seen with H112 transformed with pTA250 (pHV1/4 origin) or pTA441 (oriC1). This suggests that the efficiency of this putative second chromosomal origin (or the stability of the plasmid) is significantly lower than that of the pHV1/4 and orc1-associated origins. Moreover, the intergenic sequence upstream of orc5, which contains oriC2, can be deleted efficiently by using the same gene knockout procedure as was used to delete the pHV1/4 origin [21]. The resulting strain CN28 shows no growth defect (unpublished data).
The 793-bp intergenic sequence on contig number 453 (corresponding to pHV3) is located between genes coding for a putative TATA box-binding protein and Orc6, and features two DUEs and several direct repeats with similarity to ORB consensus sequence (Figure 6A and 6B). To test if this intergenic region confers ARS activity, a 693-bp PCR fragment including the DUEs and repeats was inserted in pTA131 to generate pCN26 and used to transform H112. Transformants were obtained with high frequency (Figure S2). When probed with the ori-pHV3 sequence, a PFG confirmed that this origin is located on the pHV3 replicon (Figure 6C).
The 688-bp intergenic region on contig number 452 (which should correspond to pHV4) is located between a putative translation initiation factor and orc3 genes and features a DUE and multiple direct repeats that are different from the ORB consensus sequence (Figure 6D and 6E). This putative origin was tested for ARS activity by cloning a 592-bp PCR fragment of the intergenic sequence in pTA131 to generate pCN27 and used to transform H112. Significantly fewer transformants were obtained with pCN27 (ori-pHV4–2) than with pCN26 (ori-pHV3) or pTA250 (ori-pHV1/4), indicating that the second origin of pHV4 is not used efficiently (Figure S2). Moreover, a PFG probed with ori-pHV4–2 sequence showed that this origin is actually located on the 2.9-Mb main chromosome (Figure 6F), suggesting that the strains used in our experiments and/or sequenced by TIGR may have undergone genome rearrangements.
The stability of ARS plasmids pTA612 (oriC2), pCN26 (ori-pHV3), and pCN27 (ori-pHV4–2) was determined in a qualitative manner. Six transformants of H112 containing each ARS plasmid were grown on nonselective agar and then restreaked on selective medium. After restreaking, none of the transformants showed growth on selective agar, indicating complete loss of the plasmid (Table 2). In comparison, transformants with pTA441 (oriC1) showed normal growth on selective agar, indicating efficient maintenance of the plasmid. Transformants with pTA250 (ori-pHV1/4) showed growth on selective agar, but the number of colonies was significantly reduced relative to growth on nonselective agar, indicating some loss of the plasmid.
As a control, we tested whether the intergenic region adjacent to orc4 could function as an ARS element; this region does not coincide with a peak of G/C disparity (Figure 4A) and does not feature a DUE. A 1.7-kb ApaI-BglII genomic DNA fragment of this intergenic region was cloned in pTA131 to generate pTA611 and used to transform H112. No transformants were observed, indicating that not every cdc6/orc1 gene is associated with a potential DNA replication origin.
RIP mapping [33] was used to determine the exact positions of DNA replication initiation at the origins we had identified. A preparation of enriched nascent DNA strands was used in primer extension reactions using 32P-labeled primers hybridizing to the leading strand either side of the DUEs of each ARS element (Table S1). The sizes of the amplification products were determined using electrophoresis under denaturing conditions (Figure 7A), allowing identification of the shortest amplification products that mark the start site of the leading strand. Clear transition points between leading and lagging strand replication within or near the DUEs were revealed for both strands of each origin (Figure 7B). It is of note that the detected transition points colocalize with chromosomal regions where DNA is predicted to be bent. Amplification products obtained in negative control experiments, which used linearized plasmids containing the origin sequences that were isolated from Escherichia coli, are indicated in Figure 7B using dots. As expected, they do not coincide with the extension products obtained using H. volcanii replication intermediates. These results indicate that all the ARS elements isolated during this work are functional in their chromosomal context.
Initiation of DNA replication at an origin depends on cis-acting sequence elements. In bacteria and unicellular eukaryotes, these have been well-characterized using genetics. In contrast, archaeal replication origins were identified relatively recently [2], and only limited genetic methods have been used to study DNA replication in archaea [15]. In this work, we isolated several ARS elements that use chromosomally encoded factors for their replication. RIP assays were used to confirm that these ARS elements correspond to functional replication origins in their chromosomal context. The origins are distributed on the different replicons of H. volcanii, including two on the main chromosome. Previous work in Sulfolobus solfataricus and recent studies in A. pernix have shown that Crenarchaeota can use more than one origin to replicate a circular chromosome [5,6]. Bioinformatics has suggested that Halobacterium sp. NRC-1, a euryarchaeon like H. volcanii, might have two chromosomal replication origins [34], but attempts to identify the second origin experimentally were not successful [15]. Our work provides the first proven example from the Euryarchaeota of multiple origins per replicon.
The general characteristics of the H. volcanii replication origins are similar. All contain AT-rich sequences required for unwinding and/or bending of the origin DNA. These are surrounded by repeated sequence motifs that correspond to the classic ORB elements found at other archaeal origins. However, some of these repeats show significant differences to the archaeal mini-ORB consensus. Many archaeal ORBs also feature a characteristic “G-string” element (Figure 2) that contributes to Cdc6/Orc1 binding at the origin (S. Bell, personal communication). While “G-string” elements are found at all origins of H. volcanii, the two ORBs surrounding the primary DUE of oriC1 also contain an extended “G-string” that appears to be specific to halophiles (Figure 2). Furthermore, long GT-stretches are found after some repeats, such as the two ORBs surrounding the DUE of ori-pHV1/4 (Figure 1B). The subtle variations found in halophilic ORB sequences might contribute to origin-specific binding by different Cdc6/Orc1 proteins. Alternatively, these halophile-specific features could promote DNA bending at high intracellular salt concentrations, thus favoring the formation of a higher order complex between the origin and initiator proteins [9].
The number of Cdc6/Orc1 proteins is almost certainly higher than the number of origins in H. volcanii; for example, the intergenic region adjacent to orc4 does not show ARS activity. Furthermore, we were able to delete several cdc6/orc1 genes (orc1, orc5, and orc10, unpublished data), suggesting that their functions at least partially overlap. It is possible that some Cdc6/Orc1 proteins might promote the initiation of “routine” DNA replication, while other initiators are adapted to function under specific physiological conditions. For example, our observation that the second chromosomal origin (oriC2) appears to function less efficiently in laboratory conditions suggests that it is only used under certain circumstances. This could be due to location of the origins on ARS plasmids, as opposed to their native chromosomal loci. On the other hand, it is noteworthy that the overall pI values of the different Cdc6/Orc1 proteins vary between 3.98 and 5.59 (Table 1), with the oriC1-associated Orc1 being the most acidic, raising the possibility that they might function optimally under different salt concentrations.
A surprising result that emerged from our initial genetic screen for ARS elements was that only one origin (ori-pHV1/4) was isolated. This is unlikely to be due to technical limitations of the method, since extensive libraries were constructed using different restriction enzymes. Moreover, we were able to recapitulate the outcome of the screen by cotransforming H. volcanii with a mixture of ori-pHV1/4 and oriC1 ARS plasmids (Figure 3). The majority (74%) of transformants contained only the ori-pHV1/4 plasmid, and less than 9% contained only the oriC1 plasmid (Figure 3A). This disparity is unlikely to be due to differences in plasmid establishment, since the two ARS constructs show the same transformation efficiency (Table 2) and confer a similar phenotypic load on their host (Figure S2). Instead, it would appear that the ori-pHV1/4 ARS plasmid is dominant to the oriC1 ARS plasmid. Evidence in favor of this suggestion emerged from the examination of transformants containing both plasmids (Figure 3B). Upon further propagation, only the ori-pHV1/4 ARS plasmid remained while the oriC1 ARS plasmid was lost. This result is all the more intriguing given that ori-pHV1/4 plasmids are significantly less stable than oriC1 plasmids (Figure 3C and 3D). Furthermore, the copy number of ori-pHV1/4 plasmids is lower than that of oriC1 plasmids (unpublished data). Thus, our results cannot be explained by mere incompatibility of ARS plasmids, since this would favor the retention of the oriC1 plasmid.
Instead, we suggest that in H. volcanii there is a functional hierarchy of replication origins, which arises from competition for common replication factors. “Dominant” origins such as ori-pHV1/4 might be bound more efficiently by Cdc6/Orc1 or other proteins such as MCM helicase and primase that participate in replication initiation. Other factors might also contribute to differential usage of origins. ARS elements that initiate efficiently can be selectively enriched from a genomic library of Saccharomyces cerevisiae [35], and reduced initiation efficiency was found to correlate with transcriptional activity directed towards the ARS element. This suggests that transcriptional interference with the prereplication complex is a determinant of origin hierarchy in eukaryotes. It is noteworthy that four of the five origins identified in this work (and almost all other archaeal origins) feature transcription units that are orientated away from replication initiation sites (Figures 1, 2, 5, and 6). Therefore, transcriptional interference might also regulate origin firing in archaea.
Our genetic observations indicate that while the functions of initiator proteins at least partially overlap in H. volcanii, the replication origins oriC1 and ori-pHV1/4 play a key role in the replication of their corresponding chromosomes. This suggests that duplication of replication origins, and not of the initiator genes, could have allowed the H. volcanii genome to develop toward a multireplicon structure. This process might still be continuing, since there are important differences between our data and the assembled genome sequence. For instance, both pHV1 and pHV4 carry a similar (if not identical) replication origin. A comparable situation is found with Halobacterium sp. NRC-1 plasmids pNRC100 and pNRC200, which share a large region of around 150 kb including long (33 and 39 kb) inverted repeats [13]. An additional example is ori-pHV4–2, which in the current version of the assembled genome sequence is located on the contig corresponding to pHV4, while in our laboratory strains is carried on the main chromosome. Detailed manual examination of the genome sequencing data (including shotgun data and closure results) indicate that the assembly is an accurate representation of the strain used for sequencing. Though it is possible the assembly is inaccurate, we believe it is more likely that there are genomic differences between the strain sequenced and strains used in our experiments. For instance, the H. volcanii DS2 strain was used to isolate DNA for sequencing (Table 3), whereas genetic and biochemical experiments used derivatives of H. volcanii WFD11 or DS70 [19,20].
These discrepancies suggest that H. volcanii, whose genome is considered relatively stable compared to other halo-archaea, undergoes periodic genome arrangements that are possibly mediated by recombination at the replication origins. In this respect, it is noteworthy that the two rRNA operons are located close to the two chromosomal origins described in this work. The rrnA operon is 200 kb away from oriC1, and rrnB is only 6 kb away from oriC2. Homologous recombination between identical sequences at these rrn operons could provide a mechanism for origin movement throughout the genome. Intriguingly, the rrn operons are both oriented away from the origins, presumably to avoid collisions between replication fork and rRNA transcription machinery.
It is notable that the H. volcanii GC-skew plots do not delineate origins as clearly as the plots of the other sequenced haloarchaea (Figure 4). For example, the peak in contig 454 does not match the ARS isolated in the initial screen. Conversely, there is a peak in contig 455 that has higher amplitude than the two origins detected but is not associated with an experimentally determined origin (Figure 4A); this peak is most likely due to a recent integration by an AT-rich prophage. These quirks could indicate either that the H. volcanii “origin signal” (i.e., the skew caused by origin usage) is weak, or that the origin is mobile and has left behind a residual signal at its previous location. Finally, the fact that haloarchaeal genomes present “inverted” GC-skews indicates that the haloarchaea may not follow the same rules for genome structure patterning as in other species.
Given the apparent plasticity of the H. volcanii genome, we believe that pHV4, pHV3, and pHV1 should be considered as secondary chromosomes and not plasmids, in the sense that a plasmid is an extrachromosomal element and is often characterized by “selfish” behavior. Prokaryotic chromosome biology has been dominated by the archetype of E. coli, but there are many examples of bacteria with more than one chromosome [36]; like H. volcanii, they often show evidence of dynamic rearrangements between the replicons (e.g., [37]). The distinction between mini-chromosomes and mega-plasmids is usually made on the basis of replicon size and the presence of essential (housekeeping) genes. In this regard, both pHV4 and pHV3 are large (690 and 442 kb, respectively), and while genes predicted to be essential are only found on the former, H. volcanii strains that have lost pHV3 grow very slowly and filament [20]. More relevant to this work, pHV4, pHV3, and pHV1 all use Cdc6/Orc1-dependent origins for replication, as does the main chromosome. By contrast, pHV2 is very small (6.4 kb), lacks homology with chromosomal sequences, is easily cured, and uses a distinct (presumably Rep-dependent) replication origin [19]. Thus it is a plasmid.
In conclusion, we have shown that the four chromosomes of H. volcanii are replicated in an analogous manner using Cdc6/Orc1-dependent origins. Since efficient tools now exist to characterize the cis-acting requirements for replication initiation in archaea, H. volcanii will serve as an interesting model system for studies on the regulated replication of complex archaeal genomes.
Unless stated otherwise, chemicals were from Sigma (http://www.sigma-aldrich.com) and enzymes from New England Biolabs (http://www.neb.com). Standard molecular techniques were used [38].
H. volcanii strains are shown in Table 3, plasmids in Table 4, and oligonucleotides in Table S1.
H. volcanii strains were grown at 45 °C on either complete (Hv-YPC) or casamino acids (Hv-Ca) agar, or in Hv-YPC or casamino acids broth, as described previously [21,39].
Transformation of H. volcanii and isolation of total genomic DNA were carried out as described previously [21]. To isolate plasmid DNA, 2 ml of a saturated culture (grown in Hv-Ca broth) was centrifuged at 3,300 ×g for 8 min and the cells were resuspended in 50 μl of 1 M NaCl, 20 mM Tris HCl (pH 7.5); thereafter, a QIAprep miniprep kit (Qiagen, http://www.qiagen.com) was used according to the manufacturer's instructions. To isolate crude DNA, cells were resuspended in 400 μl of water and incubated at 70 °C for 10 min.
H. volcanii DNA was digested with HpaII, AciI, or TaqI for 30 min at the recommended temperature, using ∼0.2 units of enzyme/μg DNA in suboptimal buffer (e.g., New England Biolabs buffer 1 for AciI). DNA fragments of the desired size were excised from agarose gels and ligated with plasmid pTA131 [21], which had previously been cut to completion with ClaI and the DNA ends dephosphorylated. The plasmid library was used to transform an E. coli dam– strain, and DNA was prepared directly from colonies to avoid differential amplification. Coverage of the libraries was >99.5% (>5,000 E. coli colonies were used in DNA preparation and H. volcanii has a genome size of ∼4 Mb). DNA was used to transform H. volcanii pyrE2 radA mutants H49 or H112, and transformants were selected on Hv-Ca plates lacking uracil.
A 3-kb BamHI-StuI fragment of pTA194 with the pHV1/4 replication origin was subcloned to generate pTA252, and a 1-kb BsmI-XbaI fragment of the intergenic region was replaced by a trpA selectable marker [21] to generate pTA266 (Figure 2D). pTA266 was used to transform ΔpyrE2 ΔtrpA strain H53, and transformants were selected on Hv-Ca plates lacking uracil and tryptophan. One transformant (H220) was grown without selection for ∼30 generations and plated on Hv-Ca + 5-FOA (5-fluoroorotic acid) to select for loss of the integrated plasmid. Around 0.4% of cells in the culture were 5-FOA-resistant but only 0.13% of these were Trp+ and therefore had deleted the pHV1/4 replication origin.
H112 containing pTA250 or pTA441 (maintained by selection on Hv-Ca agar) was used to inoculate a culture in Hv-YPC broth and grown at 45 °C. At regular intervals aliquots were plated Hv-YPC agar, and colonies patched on Hv-Ca to determine the fraction of uracil+ cells. % plasmid loss per generation (l) was calculated using the formula
where n is the number of generations and ura+ is the fraction of uracil+ cells.
Intact H. volcanii DNA was prepared in agarose plugs. 2 ml of culture (OD650nm of 1.0) was pelleted at 3,300 ×g for 10 min, 4 °C, resuspended in 1 ml of cold spheroplasting solution (15% sucrose, 1 M NaCl, 27 mM KCl, 50 mM Tris-HCl [pH 8.5]) + 0.1% NaN3, and pelleted again. Cells were gently resuspended in 80 μl of spheroplasting solution, transferred to 42 °C, mixed with 100 μl of 1.5% low-melt agarose (in 0.5× spheroplasting solution, 100 mM EDTA) and pipetted into plug moulds (Bio-Rad, http://www.bio-rad.com). Plugs were incubated in 5 ml of lysis solution (1% sarkosyl, 500 mM EDTA, 20 mM Tris-HCl [pH 8.8]) + proteinase K (0.5 mg/ml) for 3 h at 52 °C, then transferred to fresh lysis buffer + proteinase K + RNaseA (30 mg/ml) and incubated overnight at 52 °C. Plugs were washed three times in 10 ml of 100 mM EDTA, 25 mM Tris-HCl (pH 7.5) at 37 °C, equilibrated in 0.5× Tris-borate-EDTA (TBE) for 90 min at 20 °C and exposed to 50 Gy of γ radiation (137Cs, 375 Gy/s) to linearize circular DNA molecules. Plugs were loaded onto a 1% agarose 0.5× TBE gel and electrophoresis was performed at 14 °C in a CHEF mapper (Bio-Rad) using 0.5× TBE buffer, voltage gradient of 6 V/cm, switch angle of 120°, and switch times of 0.47 s (initial) to 1 min 33.83 s (final). Total run time was 20 h 18 min.
H. volcanii DS2 was diluted to an OD650nm of 0.15 in 100 ml of Hv-YPC media, grown to OD650 nm of 0.3, and pelleted. Cells were resuspended in 4 ml of lysis buffer (25 mM Tris-HCl [pH 7.5], 20 mM EDTA, 100 mM NaCl, 200 μg ml−1 proteinase K, 1% SDS). After 1 h incubation at 50 °C, 4 g of CsCl and 100 μl of Hoechst-33342 (5 mg ml−1) were added and the refractive index adjusted to 1.410 with CsCl (1 g ml−1). DNA was purified by CsCl gradient ultracentrifugation. To enrich for replicating intermediates, total DNA was passed down a BND-cellulose column pre-equilibrated with NET buffer (10 mM Tris-HCl [pH 8.0], 1 mM EDTA, and 1 M NaCl) to selectively bind single-stranded DNA. After washing with NET buffer, bound DNA was eluted with NET buffer + 1.8% caffeine at 50 °C. DNA was isopropanol-precipitated and resuspended in TE buffer (10 mM Tris-HCl [pH 7.5], 1 mM EDTA) at 1 μg μl−1. After phosphorylation of 5′-OH ends with T4 polynucleotide kinase (Promega, http://www.promega.com), DNA was treated with λ-exonuclease to digest 5′ nicked DNA ends; replication intermediates protected by RNA primers are unaffected by this treatment. Primer extension reactions used ∼500 ng of enriched replicating intermediates, 25 ng of radiolabeled primer (labeled using [γ32P]-ATP, and T4 polynucleotide kinase) and 2 units of Deep Vent (exo-) DNA polymerase. After 30 cycles of reaction (60 s at 94 °C, 60 s at 70 °C, and 90 s at 72 °C), amplification products were separated on a 6% polyacrylamide gel under denaturing conditions. Radioactive material was detected using a Phosphorimager system (Amersham, http://www.gehealthcare.com). Control experiments using linearized plasmid DNA isolated from E. coli were performed under similar conditions.
Nucleotide representation disparities were calculated using either ORIGINX (http://www.cbs.dtu.dk/services/GenomeAtlas/suppl/origin) or ZPLOTTER (http://tubic.tju.edu.cn/zcurve) programs. Publicly available sequences released in June 2006 were used to calculate local minima and maxima in nucleotide disparities. DNA curvature was calculated using a BEND.IT server (http://hydra.icgeb.trieste.it/∼kristian/dna/bend_it.html), which predicts in qualitative terms a curvature propensity of a given DNA sequence using DNase I based bendability parameters [41] and the consensus bendability scale [42]. The resulting data of each program were plotted using Origin Pro 7.5 software (OriginLab Corporation, http://www.originlab.com). |
10.1371/journal.pgen.1003473 | Scavenger Receptors Mediate the Role of SUMO and Ftz-f1 in Drosophila Steroidogenesis | SUMOylation participates in ecdysteroid biosynthesis at the onset of metamorphosis in Drosophila melanogaster. Silencing the Drosophila SUMO homologue smt3 in the prothoracic gland leads to reduced lipid content, low ecdysone titers, and a block in the larval–pupal transition. Here we show that the SR-BI family of Scavenger Receptors mediates SUMO functions. Reduced levels of Snmp1 compromise lipid uptake in the prothoracic gland. In addition, overexpression of Snmp1 is able to recover lipid droplet levels in the smt3 knockdown prothoracic gland cells. Snmp1 expression depends on Ftz-f1 (an NR5A-type orphan nuclear receptor), the expression of which, in turn, depends on SUMO. Furthermore, we show by in vitro and in vivo experiments that Ftz-f1 is SUMOylated. RNAi–mediated knockdown of ftz-f1 phenocopies that of smt3 at the larval to pupal transition, thus Ftz-f1 is an interesting candidate to mediate some of the functions of SUMO at the onset of metamorphosis. Additionally, we demonstrate that the role of SUMOylation, Ftz-f1, and the Scavenger Receptors in lipid capture and mobilization is conserved in other steroidogenic tissues such as the follicle cells of the ovary. smt3 knockdown, as well as ftz-f1 or Scavenger knockdown, depleted the lipid content of the follicle cells, which could be rescued by Snmp1 overexpression. Therefore, our data provide new insights into the regulation of metamorphosis via lipid homeostasis, showing that Drosophila Smt3, Ftz-f1, and SR-BIs are part of a general mechanism for uptake of lipids such as cholesterol, required during development in steroidogenic tissues.
| Steroid hormones are cholesterol derivates that control many aspects of animal physiology, including development of the adult organisms, growth, energy storage, and reproduction. In insects, pulses of the steroid hormone ecdysone precede molting and metamorphosis, the regulation of hormonal synthesis being a crucial step that determines animal viability and size. Reduced levels of the small ubiquitin-like modifier SUMO in the prothoracic gland block the synthesis of ecdysone, as SUMO is needed for cholesterol intake. Here we show that SUMO is required for the expression of Scavenger Receptors (Class B, type I). These membrane receptors are necessary for lipid uptake by the gland. Strikingly, their expression is sufficient to recover lipid content when SUMO is removed. The expression of the Scavenger Receptors depends on Ftz-f1, a nuclear transcription factor homologous to mammalian Steroidogenic factor 1 (SF-1). Interestingly, the expression of Ftz-f1 also depends on SUMO and, in addition, Ftz-f1 is SUMOylated. This modification modulates its capacity to activate the Scavenger Receptor Snmp1. The role of SUMO, Scavenger Receptors, and Ftz-f1 on lipid intake is conserved in other tissues that synthesize steroid hormones, such as the ovaries. These factors are conserved in vertebrates, with mutations underlying human disease, so this mechanism to regulate lipid uptake could have implications for human health.
| Larval molting and metamorphosis in Drosophila melanogaster relies on pulses of ecdysteroid hormones. During the larval stages, the prothoracic gland (PG) is the tissue responsible for the synthesis of the steroid hormone ecdysone that is secreted to the hemolymph and converted to 20-hydroxyecdysone (20E) in target tissues [1]. Other tissues releasing ecdysteroids in the adult are the gonads, ovaries and testes. Cholesterol is the precursor of all steroid hormones. In arthropods, which are unable to synthesize cholesterol, ecdysteroids are synthesized from dietary cholesterol or phytosteroids. Cholesterol is converted to 20E through a series of enzymatic reactions that involve the cytochrome P450 enzymes coded by the Halloween genes spook, spookier, phantom (phm), disembodied (dib), shadow, shade and the Rieske non-heme iron oxigenase gene neverland [2], [3]. A transcriptional cascade triggered by 20E occurs at the onset of metamorphosis that leads to the sequential expression pattern of the transcription factors DHR3, Ftz-f1, E74 and E75 [4]. A similar transcriptional cascade is required during embryogenesis and could also be required for each larval molting [5]. However, many of the aspects involved in the regulation of ecdysteroid biosynthesis remain unknown.
The conjugation of SUMO (Small Ubiquitin-related MOdifier) to target proteins is a reversible post-translational modification highly conserved in all eukaryotic organisms. SUMOylation regulates diverse cellular processes including cell survival and proliferation, nuclear import, intracellular trafficking, transcriptional regulation and maintenance of genomic and nuclear integrity [6]. In addition, Smt3, the only SUMO homologue in Drosophila, has a role in the regulation of ecdysteroid levels during post-embryonic development [7]. Smt3 is required in the PG to produce the ecdysteroid peak necessary for the larval to pupal transition. Interestingly, smt3 knockdown PG cells results in reduced intracellular channels and, as a consequence, exhibit low levels of lipid and sterol droplets indicating that impaired cholesterol uptake could contribute to the low ecdysteroid levels observed.
The nuclear hormone receptor superfamily function as transcription factors that regulate several functions such as metabolism, development and homeostasis. Recent studies have implicated the Drosophila nuclear receptors DHR96 and dHNF4 in cholesterol and triacylglycerol homeostasis and in lipid mobilization and fatty acids β-oxidation [8]–[11]. However, it is unknown whether the nuclear receptors might regulate cholesterol homeostasis in the PG. The mammalian NR5A2 Liver receptor homolog 1 (LRH-1), member of the Ftz-f1 subfamily of nuclear receptors, has been shown to be involved in lipid absorption and homeostasis [12]. In addition, LRH-1 and its close relative Steroidogenic Factor 1 (SF-1 or NR5A1) are modified by SUMO and also bind phospholipids [13]–[20]. Recently, the disruption of SF-1 SUMOylation in mice showed the inappropriate activation of target genes that led to endocrine abnormalities and changes in cell fate [21]. Interestingly, SF-1 regulates the expression of proteins related to sterol uptake and/or mobilization such as the Scavenger Receptor Class B type I (SR-BI), which belongs to the Cluster of Differentiation 36 (CD36) family [22], [23]. SR-BI, in addition to its role in the selective uptake of High Density Lipoprotein cholesteryl ester, is required for the formation of the microvillar channels in the mammalian adrenal gland [24]. SF-1 and its Drosophila orthologue Ftz-f1 control the transcriptional regulation of cytochrome P450 enzymes involved in sterol conversion, and therefore could play similar roles in the activation of steroid synthesis [25], [26].
In order to clarify the role of SUMOylation in the mechanism of cholesterol uptake, we analyzed the function of the Drosophila CD36 family and Ftz-f1 in the PG during steroidogenesis at the onset of pupariation. We show that the Drosophila CD36 family member Sensory neuron membrane protein (Snmp1) is necessary for lipid uptake in the PG, downstream of Smt3 function. We also show that SUMO is required for ftz-f1 expression and, in addition, Ftz-f1 is modified by SUMO in vitro and in vivo. More importantly, reduced levels of Ftz-f1 in the PG leads to impaired pupariation with reduced levels of lipid droplets, similar to the smt3 knockdown phenotype. Conversely, overexpression of ftz-f1 is able to rescue Snmp1 expression in smt3 knockdown PGs. Finally, extending our observations in the PG, we saw that Smt3, Ftz-f1 and the Scavenger Receptors have a role in lipid uptake during ovarian steroidogenesis. Our results suggest similar requirements for cholesterol uptake in various steroidogenic tissues.
In Drosophila, ftz-f1 encodes for two protein isoforms with distinct temporal expression patterns: αFtz-f1 expressed early in embryogenesis and βFtz-f1 expressed later in embryogenesis and during larval, prepupal and early pupal stages [27], [28]. βFtz-f1 has previously been implicated in regulating ecdysteroid titers during post-embryonic development, specifically at the prepupa to pupa transition [26], [28]. To investigate whether Ftz-f1 contributes to the smt3 knockdown phenotype (herein referred to as smt3i), we silenced both isoforms of ftz-f1 in the PG using UAS-ftz-f1i lines and the phm-Gal4 driver. The resulting larvae will be referred to as ftz-f1i. Reduced levels of ftz-f1 led to arrested development at larval stages (Figure 1A, 1B). Similar to the smt3i phenotype, ftz-f1i larvae arrested at third instar (L3) before pupariation, as shown by the mouth hook morphology (Figure 1B). These larvae did not pupariate and survived as L3 for several weeks. In contrast to the smt3i phenotype, we also observed larvae that arrested development at the transition from the second instar (L2) to L3, as shown by the double mouth hooks. These larvae die at 120 hours after egg lying (AEL; Figure 1B). The two larval phenotypes could reflect the silencing of the two ftz-f1 isoforms.
The developmental arrest at L3 suggested that ftz-f1i larvae were not able to synthesize normal levels of ecdysteroids at the onset of pupariation. Accordingly, and similarly to other low ecdysteroid mutants including the smt3i larvae, ftz-f1i larvae fed with exogenous 20E pupariated (Figure 1C).
smt3i PG cells show a reduction in lipid droplets and sterol levels, in addition to changes in steroidogenic enzymes and transcription factors [7]. ftz-f1i PGs also showed reduced levels of lipid droplets per cell (Figure 1D, 1E), as well as reduced levels of the steroidogenic enzyme Dib (data not shown). Taken together, these results show that Ftz-f1 is required in the PG at the end of L3 to acquire appropriate levels of cholesterol and to process it into ecdysone.
It was previously reported that Ftz-f1 protein is reduced in smt3i PGs [7], which could explain why ftz-f1 silencing phenocopies SUMO downregulation. SUMO could be necessary for ftz-f1 transcription, for Ftz-f1 protein modification, or both. To clarify this point we analyzed ftz-f1 expression in PGs from 120 hours AEL blue-gut wandering larvae (i.e. 5–12 hours before pupariation), and 120 hours AEL clear-gut larvae (i.e. 1–6 hours before pupariation). ftz-f1 transcription is upregulated in clear-gut larvae compared to blue-gut (Figure 1F, 1G). However, ftz-f1 expression is lower in smt3i PGs and it does not get upregulated in older larvae (Figure 1H, 1I). This indicates that SUMO is involved in ftz-f1 transcriptional regulation.
It has been reported that the mammalian homologues of Ftz-f1 are modified by SUMOylation [14], [18]–[21]. Therefore, we decided to test whether Ftz-f1 can be modified by SUMO. We determined the potential SUMOylation sites in Ftz-f1 using the SUMOplot analysis program and the Phosida Posttranslational Modification Database (Figure 1J). The two isoforms share the C-terminal region, but contain different N-terminal regions (Figure 1J). The in silico analysis showed that αFtz-f1 and βFtz-f1 share four SUMO consensus sequences, all of them conserved among insects, of which three have high prediction scores (Figure 1J and data not shown). The SUMOylation consensus sites are located in the DNA binding domain, in the hinge region and two of them in the ligand binding domain (Figure 1J and Figure S1). Those located in the DNA binding domain and in the ligand binding domain, are conserved between insects and NR5A2 (LRH-1) but not NR5A1 (SF-1; Figure 1J and Figure S1).
Using an in vitro SUMOylation assay, our results showed that Ftz-f1 protein is modified in the presence of SUMO, appearing as additional slow-migrating bands (Figure 1K). In order to analyze whether Ftz-f1 can also be SUMOylated in vivo, we developed a SUMOylation assay in cultured Drosophila cells. Smt3 was expressed as a fusion with a biotinylation-target peptide (bioSUMO, see Materials and Methods), along with Ftz-f1 and BirA enzyme. In this assay, biotinylated SUMO-conjugated proteins were bound to NeutrAvidin beads allowing for the specific isolation of SUMOylated material. Our results show that full-length αFtz-f1 is SUMOylated in Drosophila S2R+ cells (Figure 1L), with the estimated molecular weight for the main band (low asterisk) corresponding to the addition of one bioSUMO moiety. An additional but weaker band was observed at a higher molecular weight (upper asterisk in Figure 1L), suggesting the possibility that Ftz-f1 could also be modified by more than one bioSUMO moiety in vivo.
Taken together, these results show that SUMOylation regulates Ftz-f1 in two ways. On one hand, it is necessary for ftz-f1 expression and, on the other hand, the protein Ftz-f1 is modified by SUMO in vitro and in vivo, suggesting that the post-translational modification o Ftz-f1 could potentially contribute to Ftz-f1 function in the PG.
The CD36 family of Scavenger Receptors is necessary for lipid uptake in various mammalian cell types [29]. We hypothesized that members of this family could be necessary for sterol uptake in the PG and could mediate some of the functions of Smt3 and/or Ftz-f1. Fourteen members of the CD36 family were identified in Drosophila melanogaster [30]. By in situ hybridization it was recently shown that three of these receptors, peste (pes), croquemort (crq) and Snmp1 were upregulated in the PG at the onset of pupariation [31]. We confirmed this upregulation by quantitative real-time PCR (qPCR) of cDNA samples from precisely staged PGs at 96 hours AEL, 120 hours AEL blue-gut wandering larvae and 120 hours AEL clear-gut larvae (Figure 2A).
By immunostaining using specific antibodies we observed that Snmp1 was expressed in Drosophila PG cells at 120 hours AEL (Figure 2B, 2C). Interestingly, Snmp1 expression was upregulated from L3 blue-gut (Figure 2B) to L3 clear-gut larvae (Figure 2C). The same results were observed when antibodies against Crq were used (Figure 2D, 2E), suggesting a requirement for high levels of the Scavenger Receptors at the end of L3. The Scavengers expression upregulation from blue- to clear-gut larvae coincides with an increase in the content of lipid droplets of the clear-gut PGs (Figure 2F, 2G).
As reported previously, the expression of Snmp1 is upregulated at the level of mRNA from blue- to clear-gut larvae (Figure 3A, 3B) [31]. Interestingly, we observed that the levels of expression of Snmp1 are reduced, and are not upregulated, in smt3i PGs (Figure 3C, 3D). Similar results were obtained when specific antisense probe for crq was used (Figure S2A–S2D). However, pes expression is still present in smt3i PGs (Figure S2E–S2H), indicating different regulatory requirement for pes respect to crq and Snmp1.
At the level of proteins, we observed that Snmp1 expression was reduced in smt3i and in ftz-f1i PG cells (Figure 3E, 3F, 3G). Although basal levels of expression were evident in smt3i 120 hours larvae (Figure 3E), the upregulation of Snmp1 in smt3i PGs was never observed, even when we analyzed 144 hours AEL larvae (Figure 3F, compare to Figure 2C). The same results were observed when anti-Crq antibodies were used (Figure 3H, compare to Figure 2E).
These results indicate that SR-BIs are downstream of Smt3 and Ftz-f1 in the PG and suggest that SR-BIs might be required to mediate the role of Smt3 and Ftz-f1 in steroidogenesis during the larval to pupal transition.
To test the implication of the three Drosophila SR-BI members expressed in the PG in cholesterol uptake and steroidogenesis, we used the phm-Gal4 line to silence crq, pes or Snmp1 specifically in the PG. Interestingly, at 25°C Snmp1 knockdown in the PG (herein called Snmp1i) led to L3 developmental arrest (Figure 4B). The Snmp1i L3 larvae survived for approximately ten days, darkened in the anterior part but failed to form a cuticle, the pseudo-pupae dying at this stage (Figure 4B). Unlike smt3i or ftz-f1i, the long-lived Snmp1i larvae had normal levels of expression of the cytochrome P450 enzyme Dib, indicating that the receptor does not have a major role in regulating this steroidogenic enzyme (Figure 4D, 4E). However, similarly to smt3i and ftz-f1i, Snmp1i PGs had a reduced content of lipid droplets, reaching between 5 to 16% of droplets per cell in comparison to controls (Figure 4C, 4F, 4G). Intriguingly, these droplets were abnormally big, being about three times bigger in Snmp1i cells compared to controls (Figure 4C, 4F, 4G). When Snmp1i L3 larvae were fed with exogenous 20E, 100% of the animals pupariated, and 66% led to adult flies (n = 30).
In contrast, knockdown of crq or pes in the PG at 25°C did not lead to any obvious phenotype. However, we observed L3 developmental arrest when silencing pes at 29°C (data not shown). Also in this case, the lipid droplets in the PG were less in number and larger in size than in controls (data not shown), which suggest that these receptors are involved in lipid uptake as well as in lipid droplet mobilization and/or the control of lipid droplets size.
Besides PG, we wondered whether the molecular mechanism of lipid transport would be conserved in other steroidogenic tissues. The ovary is one of the sources of ecdysteroids in female adult insects. In cockroaches and locusts it has been shown that the follicle cells can synthesize and secrete ecdysone [32]–[34]. In Drosophila, in vitro synthesis and secretion of ecdysteroids has also been described in the ovary [35], [36]. smt3 and other SUMOylation genes are expressed during oogenesis in Drosophila [37]. We examined Smt3 protein expression, which was localized mainly to the nucleus, and observed the highest levels in the germarium and in the follicle cells at stages 2–8 egg chambers, with weaker expression also evident in nurse cells at these stages (Figure 5A, 5B). At later stages of oogenesis, Smt3 expression in the follicle and nurse cells was maintained (data not shown). To ask whether the Smt3 requirement for sterol uptake is a common feature for steroidogenic tissues, we silenced smt3 in the follicle cells using UAS-smt3i lines and the follicle-specific T155-Gal4 driver (Figure 5C) [38].
The ovary expresses several members of the cytochrome P450 enzyme family involved in ecdysteroid synthesis such as Phm, Dib, Shadow, Shade, Neverland and Dare [39]–[43]. We have focused our study from stage 8 until stage 10–11 of oogenesis, when the highest expression levels of these enzymes have been observed. In control female ovaries Dib expression started at stage 8 in the follicle cells, with the highest levels at stages 9–10, in correlation with the peak of ecdysone synthesis in the ovary (Figure 5D and data not shown) [44], [45]. Similarly to what happens in the PG [7], the expression levels of Dib in smt3 knockdown follicle cells were drastically reduced (Figure 5E). In addition, Ftz-f1, which was expressed both in nurse and follicle cells, was also reduced in the smt3 silenced follicle cells (Figure 5F, 5G), as shown in the PG [7].
At the stages analyzed, the follicle and nurse cells and the oocyte show high number of lipid droplets (Figure 6A). Interestingly, we observed a strong reduction in the number of lipid droplets in smt3i follicle cells, which correlate with the observations shown in the PG (Figure 6B and data not shown). However, the oocyte was not completely depleted of lipids as it received lipids from the nurse cells throughout the ring canals (Figure S3A), which might explain why these oocytes are able to give rise to viable embryos. The reduction of Ftz-f1 levels in the follicle cells produced a more severe phenotype with death of many ovarioles. However, those that survived showed also a reduction in the number of lipid droplets in the follicle cells, this reduction being not as strong as in the smt3i phenotype (Figure 6C).
To follow up our observations in the PG, we analyzed whether the SR-BI family could be involved in the lipid uptake in the ovary. Indeed, we observed that Snmp1 knockdown in the follicle cells clearly reduced the lipid droplet content in these cells (Figure 6D). Interestingly, the few droplets observed in some cells were located in the basal part, suggesting a mobilization impairment of droplets from the basal to the apical surface of follicle cells (Figure S3B).
During the stages 8–10 of oogenesis, follicle cells accumulate over the oocyte and become columnar, showing numerous microvilli on their apical surface [46]. These microvilli can be detected by the expression of Cad99C, a cadherin involved in the regulation of the microvilli length [47], [48]. We detected Cad99C protein on the apical plasma membrane of the follicle cells surrounding the oocyte (Figure 6E). However, in smt3i follicle cells Cad99C expression was highly reduced, which suggests a microvilli malformation in these cells (Figure 6F). DE-Cad, which marks adherent junctions and is required for centripetal cell migration, was also greatly reduced in smt3i follicle cells (Figure 6G, 6H) [49]. We observed a slight delay in the centripetal migration of follicle cells and very small gaps in the vitelline membrane that might be attributed to the reduction in the expression level of cadherins (data not shown). Even with these changes, egg fertility was only slightly affected and viable embryos were obtained.
These results show that reduced levels of Smt3 in follicle cells affect the levels of Dib and Ftz-f1, alter lipid droplet size and distribution, and affect membrane surface area (in this case, by altering microvilli), suggesting that Smt3 performs parallel functions in the PG and in the ovary. Moreover, Snmp1 could play similar roles in lipid uptake and mobilization in the PG and in the follicle cells.
Our results showed that silencing Snmp1 phenocopies the impairment of lipid uptake of smt3 or ftz-f1 knockdowns. To test the role of Snmp1 in the smt3i phenotype in the PG, we over-expressed Snmp1 in an smt3i background. Interestingly, we observed the rescue of the number of lipid droplets per cell in 51% of the larvae (n = 29; Figure 7A, 7B). The lipid droplets were comparable to the controls in size, indicating that the overexpression of Snmp1 was able to restore the uptake of lipids and their mobilization in the PG. Furthermore, Snmp1 was able to rescue the lipid droplets content in smt3i follicle cells (Figure 7C). These results demonstrate the role of Snmp1 in sterol uptake in both the PG and the ovary and underline the relevance of the Drosophila SR-BI family in steroidogenesis.
Our results suggest that SR-BIs are downstream of Ftz-f1. To study whether Snmp1 transcription is mediated, at least in part, by Ftz-f1 we analyzed its promoter region. Interestingly, the Snmp1 locus contains two putative Ftz-f1 binding sites (TCAAGGTgG, position −1410 from the initial methionine; CCAAGGgCA, position +1666) that only differ in one nucleotide from the consensus sequence 5′-PyCAAGGPyCPu-3′. In addition, an atypical SF-1 binding site (TttGGGCCA, position −1974) that contains the core of the consensus sequence TCAGGGCCA [50], is present in the promoter and could also be implicated in Snmp1 regulation. We examined whether Ftz-f1 could activate Snmp1 transcription using a reporter gene cloned downstream of a 2 Kb fragment located at 5′ of the Snmp1 transcription initiation point. Interestingly, α and βFtz-f1 activated luciferase activity significantly in S2R+ Drosophila cells (Figure 8A). A mutation in position −1971, which eliminates the atypical SF-1 binding site reduces luciferase activity in presence of αFtz-f1 (Figure 8B), suggesting that activation depends of Ftz-f1 binding to that site.
Fusions of SUMO with the protein of interest can be used as a model of constitutive SUMOylation without the pleiotropic effects of overexpressing SUMO in the cells [51]. In addition, fusions of Ubc9, the E2 SUMO conjugating enzyme, are also used successfully as model for constitutive SUMOylation [52]. To examine the effect of Ftz-f1 modification on Snmp1 activation, we analyzed the transcriptional activity of Smt3-βFtz-f1 and Smt3-αFtz-f1 fusion proteins (Figure 8A). Cotransfection assays showed that while α or βFtz-f1 were able to increase the luciferase activity compared with the control vector, Smt3-αFtz-f1 or Smt3-βFtz-f1 caused a reduction in the level of transcription (Figure 8A). This effect was reproduced using the Ftz-f1 proteins fused to Lesswright, the Drosophila homologue to Ubc9 (data not shown). These results suggest that SUMOylation reduces Ftz-f1 activation capacity on Snmp1.
According to our previous data, ftz-f1 expression depends on SUMO. In turn, Snmp1 expression depends of Ftz-f1. In fact, α or βFtz-f1 overexpression in the PG rescues Snmp1 expression in an smt3i background in 100% of the PGs analyzed (n = 37 and n = 36, respectively; Figure 8C, 8E, 8G, compare with Figure 3E). We then examined in vivo the differences of activity between α or βFtz-f1 and Smt3-α or Smt3-βFtz-f1. Smt3–αFtz-f1 or Smt3–βFtz-f1 overexpression only recovered Snmp1 expression in 8 or 25% of the PGs, respectively (n = 25 and n = 39; Figure 8D, 8F, 8G), suggesting a difference in the activity of Ftz-f1 depending on its SUMOylation status.
Here, we have investigated the effect of SUMOylation on SR-BI and Ftz-f1 activities in the context of steroidogenesis. We show that SUMOylation has a dual role on Ftz-f1 function. On one hand, ftz-f1 transcription depends on SUMO. On the other hand, SUMO modifies Ftz-f1, reducing its capacity to activate Snmp1 transcription. In addition, we demonstrated that Drosophila SR-BI family is involved in steroidogenesis by regulating the cholesterol uptake in the PG required for the synthesis of ecdysone. Our results also show that Snmp1 is involved in cholesterol uptake, acting downstream of SUMO and Ftz-f1, and is able to recover the lipid content in smt3i PGs. Furthermore, we showed that SUMO and the Scavenger Receptors are also involved in lipid capture and mobilization in the ovarian follicle cells.
SUMO modification, a highly conserved pathway throughout evolution, is known to impact the activity, interactions, localization and stability of proteins [53]. A number of studies during the last years have clearly established the essential role of SUMOylation during development. In Drosophila melanogaster, the single SUMO protein Smt3 is expressed during development and is highly enriched during embryogenesis and in adult females [54]–[56]. At later stages it is predominantly expressed in the central nervous system and in the gonads [37], [56], [57]. Components of the Drosophila SUMO conjugation pathway have also been implicated in diverse processes such as embryogenesis, wing morphogenesis, central nervous system development, neurodegeneration, photoreceptor development and immune response (reviewed in [58], [59]. In addition, Smt3 is required in the PG for the correct function of the ecdysteroid biosynthetic pathway at the time of puparium formation, which derives from defects in cholesterol uptake and formation of intracellular channels [7]. We have now shown that Smt3 requirement for lipid uptake is a common feature for steroidogenic tissues by analyzing its function in the ovary, a tissue that also requires cholesterol for the synthesis of ecdysone. In addition, we show that SUMO is required to activate ftz-f1 transcription. During the late larval pulse of ecdysone, the transcription factors E75, DHR3 and the Nitric Oxide Synthase (NOS) are known to regulate Ftz-f1 expression. Ftz-f1 is a direct target of DHR3 [60], [61] and E75 suppresses this DHR3 mediated expression of βFtz-f1 [62]. NO, produced by NOS, prevents the E75 function as suppressor of DHR3 [63]. Therefore, these factors could be responsible for the SUMO dependent ftz-f1 transcriptional activation. Noteworthy, NOS is modified by SUMO [64], [65] and its downregulation in the PG prevents pupariation [63]. An interesting question to be analyzed in the future will be the study of the biological role of NOS SUMOylation in the regulation of ftz-f1 expression.
Although several Drosophila Smt3-modified proteins have been identified, the effect of SUMOylation remains unknown for most of them. A proteomic study in early Drosophila embryos identified 140 SUMOylation targets that confirmed the role of this pathway in Ras signaling, cell cycle and pattern formation [66]. Smt3 also regulates negatively JNK signaling through sequestering HipK in the nucleus [67]. Further identification of SUMO targets at different developmental stages will be particularly important. We are especially interested in PG proteins modified by SUMO during the larval-pupal transition since Smt3 function is required in this crucial developmental window. In that context, the SUMO modification of Drosophila Ftz-f1 described here, to our knowledge, for the first time is particularly exciting. Ftz-f1 hypomorphic regulatory mutants show defects at the prepupal-pupal transition, such as failure of head eversion and salivary gland cell death, and suggest a function for this transcription factor in muscle contraction events at this transition [28], [68], [69]. βFtz-f1 expression in late first and second instar larvae and its role in molting has also been described in Drosophila and other insects [28], [70], [71]. However, the precise expression pattern and the role of Ftz-f1 in Drosophila PGs from third instar larvae were not completely understood. Our results show that disruption of Ftz-f1 in the PG by RNAi impairs development at late L3, which clearly proves that Ftz-f1 is required at the larval-pupal transition. Interestingly, Ftz-f1 knockdown in the PG results not only in reduction of Dib expression as expected, but also in diminution of Snmp1 expression and a significant decrease of lipid levels. Therefore, our study implicates Ftz-f1 in sterol homeostasis in the PG as well as in the ovary, suggesting that this role could be extended to more steroidogenic tissues. Interestingly, Nhr-25, the only C. elegans member of the NR5A family, controls the larval to adult transition by regulating an endocrine program of lipid uptake and synthesis [72], [73].
A growing number of nuclear receptors are also known to be targets of SUMOylation (reviewed in [74], [75]. In mammals, the nuclear receptor coregulator KLF5 (Krüppel-like transcription factor 5) uses SUMOylation as a molecular switch to repress or activate genes involved in lipid catabolism [76]. Other transcription factors modified by SUMO and involved in energy metabolism include PPAR-γ, C/EBPs and SREBPs [77]–[79]. In addition, a deSUMOylating enzyme, SeNP2 plays a critical role in the control of adipogenesis [80]. SUMOylation of SF-1, as suggested for other transcription factors, attenuates its transcriptional activity [18]. However, only a subset of SUMO sensitive targets seems to be affected [81]. On the other hand, Androgen receptor interacting protein 4, which interacts with SUMOylated SF-1, suppresses SF-1 mediated transcription [82]. Recently, the elimination of SF-1 SUMOylation in mice has been described [21]. UnSUMOylable SF-1 mutants activated Sonic hedgehog signaling and altered other potential SUMO sensitive targets, leading to endocrine abnormalities and changes in cell fate. SUMO modification has also been associated with increased transcriptional activity of nuclear receptors, as reported for retinoid acid receptor related orphan receptor α (ROR α) and estrogen receptor α (ERα). Interestingly, SUMOylation at the hinge region of both nuclear receptors has been associated to transcriptional activation [83], [84]. SUMOylation of Ftz-f1, as occurs with its orthologue SF-1, could be an important mechanism to control its activity and probably a correct ratio of SUMOylated to unmodified Ftz-f1 must be maintained for proper development. As for SF-1, SUMOylation of Ftz-f1 seems to reduce its capacity of transcriptional activation on Snmp1. What could be the biological function of Ftz-f1 SUMOylation? One possibility might be that SUMOylation attenuates Ftz-f1 function after pupariation. As the first peak of ecdysone production subsides, perhaps Ftz-f1 SUMOylation and reduced levels of SR-BI contribute to this downregulation, separating it from the second ecdysone peak that drives pupation itself (10–12 hours after puparium formation).
Cholesterol, a main component of the cell membranes, is also important for the synthesis of steroid hormones. Steroid hormone biosynthesis requires, in addition to correct cholesterol uptake, appropriate intercellular and intracellular transport. Insects, which are incapable of synthesizing cholesterol, incorporate it from the diet through intestinal absorption and then transport it to different tissues via open circulation in the hemolymph associated with the lipoprotein lipophorin [85]. Several ultrastructural changes have been described in active PG cells, such as increased agranular ER, mitochondria and increased intracellular channels and nuclear folding that correlate with the sterol uptake and/or release of ecdysone [7], [86]–[88]. However, the mechanisms used to incorporate cholesterol in the PG and secrete ecdysone are still largely unknown.
Two main pathways have been described for cellular uptake of lipoprotein-cholesteryl esters: the low-density lipoprotein (LDL) receptor mediated endocytic uptake of LDL-cholesterol and the “selective” cholesteryl ester uptake pathway mediated by SR-BI. In insects, proteins related to the mammalian LDLR, lipophorin receptors, have been identified [89]. Recently, the function of Drosophila Lpr1 and Lpr2 for neutral lipid uptake in imaginal disc cells and oocytes have been described; however, the phenotype of lpr1 and lpr2 mutants does not suggest a role for these receptors in the PG or in the larval-pupal transition [89].
In mammalian steroidogenic tissues the SR-BI “selective pathway”, without endocytic uptake, seems to be the main one used to satisfy the cholesterol requirements for hormone synthesis. [90]–[92]. Interestingly, SR-BI is also necessary for the formation of the microvillar channels of the adrenal gland, as shown by the reduction of channels in SR-BI null mice and the increased formation of these channels after overexpression of SR-BI [24], [93]–[96]. SR-BI is also expressed in the rat ovary where, similar to adrenocortical cells [95], it is detected on microvilli and membranes of microvillar channels that contained trapped lipoproteins [97]. The expression of SR-BI is regulated by the nuclear receptors SF-1 and LRH-1, supporting the significance of these receptors in lipid capture for steroidogenesis [22], [23], [98]. Other factors such as the hormones ACTH, estrogen or gonadotropin induce SR-BI expression [99], [100].
The Drosophila CD36 gene family consists of 14 genes [30], [101]. Recently, the expression of Snmp1, Crq and Peste in steroidogenic tissues such as PG, ovaries and testes was described [31], pointing to a role for these receptors in these tissues. Significantly, the expression of these receptors in the PG was upregulated at the moment of pupariation when high levels of ecdysteroids are required [31] and this work). Silencing Snmp1 or pes in the PG produced the developmental arrest at L3 prior to pupariation. These results indicate that these receptors are not functionally equivalent and the presence of one of them cannot substitute for the other in the PG. Alternatively, reduction of the levels of one of the receptors might be enough to lower the total content of Scavengers and, therefore, the total capacity of the cell for capturing lipids. Interestingly, overexpression of Snmp1 is able to rescue the lipid content of smt3i PG cells. However, this rescue of lipid content is not sufficient to allow the larval-pupal transition, suggesting that the cells are still unable to produce a threshold level of ecdysone. There could be several explanations for this. For instance, overexpression of one of the receptors would be enough for lipid capture but not for lipid mobilization. In this respect, abnormally large lipids droplets were observed in Snmp1i PG cells, which suggest an additional role in sterol mobilization or a function in the regulation of the lipid droplet size. Several proteins have been shown to affect the size of the lipid droplets such as Rab small GTPases, sterol regulatory element binding protein cleavage activating protein (SCAP) and isoforms of phosphocholine cytidylyltransferase [102], [103]. Mutants for other proteins that promote intracellular transport of lipids in the PG, such npc1a mutants, have abnormal accumulation of intracellular sterol and are unable to molt due to low levels of ecdysone [104]–[106].
We showed that the expression of SR-BIs increases at the onset of pupariation, coinciding with an increase in PG's lipid content. Moreover, our results showed that SR-BIs are regulated by Ftz-f1. Interestingly, the Snmp1 locus contains two putative Ftz-f1 binding sites and an atypical SF-1 binding site that could be implicated in Snmp1 regulation. Indeed, experiments in cultured cells and in vivo showed that Ftz-f1 is able to activate Snmp1 promoter. Snmp1 might not be the only Scavenger Receptor regulated by Ftz-f1 in the PG. Furthermore, in addition to SUMOylation influencing the capacity of Ftz-f1 to regulate SR-BIs expression (Figure 8H), we observed that Snmp1 contains two high score putative SUMOylation sites (data not shown). Is Snmp1 modified by SUMO and could this modification affect its function in cholesterol uptake/transport during the larval to pupa transition? Does SUMOylated Ftz-f1 affect the regulation of other SR-BIs in clear-gut larvae? We cannot discard the possibility that other proteins involved in ecdysone synthesis or transport are SUMOylated. The fact that viability is not rescued by Snmp1 overexpression, suggests that this is indeed the case. These questions remain unanswered and will be addressed in the future.
In summary, we demonstrated that Drosophila SR-BI family and Ftz-f1 participate in steroidogenesis downstream of SUMOylation by regulating the lipid uptake in the PG required for the synthesis of ecdysone. The participation of these factors in lipid uptake is conserved in other steroidogenic tissues, suggesting a general role for SUMO, Ftz-f1 and SR-BI in lipid uptake. Our data provide new insight into the lipid homeostasis of the organism.
Flies were raised on standard Drosophila medium at 25°C. The wild-type (WT) control strain was Vallecas. Gal4 strains were phm-Gal4,UAS-mCD8::GFP/TM6B,Tb (called phm-Gal4, obtained from P. Leopold and C. Mirth) [107], [108] and P(GawB)T155-Gal4 (Bloomington Drosophila Stock Center- BDSC). UAS-RNAi lines were: UAS-smt3i [7]; UAS-ftz-f1i (Vienna Drosophila RNAi Center- VDRC- #2959, which recognizes α and βftz-f1 isoforms); UAS-pesi (VDRC #33155); UAS-crqi (VDRC #45883); UAS-Snmp1i (VDRC #04210) and UAS-Snmp1i (NIG-FLY #7000R-3). Overexpression UAS lines used were obtained from BDSC: y1w67c23;P(EPgy2)crqEY14489 (#20939); w;P(UAS-Snmp1.B)217.1/CyO;TM2/TM6B,Tb1 (#25044); w*;P(UAS-Snmp1.YFP(2)273.4/CyO;TM2/TM6B,Tb1 (#25046) and w*;P(UAS-Snmp1-EGFP)218.3/CyO;TM2/TM6B,Tb1 (#25045). Information about other strains can be found in FlyBase (http://flybase.bio.indiana.edu).
Full length α and βftz-f1 cDNA sequences (EMBL database accession numbers HE716957 and HE716956, respectively; GeneArt) were cloned into the EcoRI-XbaI sites of pUASTattb vector [109] to generate pUASTattb-αftz-f1 and pUASTattb-βftz-f1, respectively. 3×Flag sequences were inserted into the EcoRI-BglII sites. To generate constitutively SUMOylated forms, degenerated nucleotide smt3 sequence that encodes for WT Smt3 protein (EMBL database accession number FN539078) [110] was introduced into the AscI and PacI sites of the previous vectors to generate pUASTattb-Smt3-αftz-f1 and pUASTattb-Smt3-βftz-f1, respectively.
Renilla and firefly luciferase (Fluc) were amplified from psiCHECK2 (Promega) by PCR and exchanged for GFP in Ac5-STABLE1 (EcoRI-HindIII) [111] to generate Ac5-FFluc-STABLE1. For the construction of pSnmp1-FLuc genomic DNA was used as template to amplify a PCR fragment containing 2 Kb of the Snmp1 upstream region (including 77 bp of 5′UTR) using the oligonucleotides Snmp1(−2000) (5′- GATCAGATCTTGAGCACTTAGGCATTTTCAAAACTATTTGGG -3′) and Snmp1(+1) (5′- GATCGAATTCCTCTGGGCAATGTTTCGATCTCTACTC -3′) (numbering based on initiator methionine). The resulting BglII-EcoRI fragment was used to replace the Actin5C promoter in Ac5-FFluc-STABLE1. Two potential Ftz-f1 binding sites were identified based on published consensus sites. Fragments were prepared for individual and double mutants in these potential Ftz-f1 binding sites using 2-step overlap extension PCR, using the forward and reverse primers Ftz-f1Mut(−1971) (5′- CTTAGGCATTTTCAAAACTATTTGttCCAGCAATAATTGGTAGCAAAC -3′) and Ftz-f1Mut(−1410) (5′-GAGCCCAGTTAGCCGGTCAAttTGGCAGAGCATCTAACTTAAATGG -3′) (mutated nucleotides in lowercase and bold). Resulting fragments were used to replace Snmp1 WT promoter sequence to generate pSnmp1Mut-FLuc plasmids. All constructs were fully sequenced.
To generate pUASTattb-crq and pUASTattB-pes, ESTs RE02070 and RE21078 inserted in the PFLC1 plasmid (Drosophila Genomics Resource Center- DGRC), were digested with EcoRI and BamHI, or MfeI and BglII, respectively. Fragments were inserted in pUASTattb [109] digested with EcoRI and BglII. Transgenic lines were generated following standard transformation procedures [112].
SUMOylation motifs were identified using SUMOplot (http://www.abgent.com/sumoplot) software and Phosida Posttranslational Modification Database (http://www.phosida.com). For the in vitro procedure ftz-f1 cDNA (LD34889, DGRC) was translated using TNT-T7 (wheat germ extract; Promega), to which 35S-methionine was added (Amersham Biosciences and Pierce). This cDNA construct contains three out of the four conserved SUMOylation consensus sites. Translated Ftz-f1 was incubated with an ATP-regenerating system, SUMO-1, Ubc9 and E1 activating enzyme (Biomol). Reactions were incubated at 30°C for 2 h, resolved by SDS-PAGE and exposed.
For cellular SUMOylation assays we developed new technology based on Franco et al. [113]. In brief, a plasmid encoding a form of Smt3, capable of being biotinylated, as well as the enzyme necessary for biotinylation, BirA, were introduced into cells. Any proteins that undergo SUMOylation will also be biotinylated facilitating their recovery using streptavidin beads. BirA was amplified from the UAS-(bioUb)6-birA vector [113] and cloned in pAc5-STABLE2-Neo [111] by substituting GFP, generating pAc5-FLAGmCherry-BirA. To generate pAc5-bioSUMO-BirA, degenerated Drosophila smt3 sequence (EMBL database accession number FN539078) [110] was fused to a biotin tag according to the strategy described [113]. The fusion was cloned into the pAc5-FLAGmCherry-BirA vector by substituting FLAGmCherry.
Drosophila S2R+ cells were obtained from DGRC [114]. Cells were cultured at 25°C in Drosophila Schneider's medium (Invitrogen) supplemented with 10% fetal bovine serum (Gibco) and 1% penicillin/streptomycin (Gibco). Transfections were performed using Effectene (Qiagen) in 6-well plates with 1 µg of pAC5-Gal4 (Addgene #24344) [115], 1 µg of pUASTattB-Flag-βftz-f1 and 1 µg of pAc5-bioSUMO-BirA or pAc5-BirA.
Transfected cells were collected after 3 days, washed with phosphate buffered saline (PBS) 1X and lysed in 200 µl of lysis buffer [8 M urea, 1% SDS, 50 mM N-ethylmaleimide in PBS and protease inhibitor mixture (Roche)]. Pulldowns were done according to [113] using 50 µl suspension of NeutrAvidin-agarose beads (ThermoScientific). For elution, samples were heated 5 minutes at 95°C in 4× Laemmli sample buffer with 100 mM DTT. The eluted sample was separated from the beads using a Vivaclear Mini 0.8 µm PES microcentrifuge filter unit. The recovered volume for both control and experimental samples was 30 µl.
For Western blots we used mouse monoclonal anti-Flag M2 antibody (1∶2000; Sigma), HRP-conjugated secondary antibody (1∶5000; Jackson ImmunoResearch) and HRP-linked anti-Biotin antibody (1∶200; Cell Signaling Technology).
ESTs from the DGRC cDNA collections were used as templates for the synthesis of the RNA probes (IP13851 for Snmp1, RE21078 for pes, RE02070 for crq and LD34889 for ftz-f1). RNA labeling was performed using the DIG RNA labeling Mix (Roche) according to the manufacture instructions, using 1 µg of linearized DNA.
RNA probes were hybridized to larval tissues at 55°C in 50% deionized formamide (Sigma), 5× saline sodium citrate, 50 µg/ml heparin sodium salt (Sigma), 0.1% Tween 20, and 100 µg/ml of phenol extracted sonicated salmon sperm DNA (Amersham Biosciences). Samples were incubated with anti-digoxigenin antibody (1∶2000; AP Fab fragments, Roche) and signal was detected using 4-Nitro blue tetrazolium chloride and 5-Bromo-4-chloro-3-indolyl-phosphate (Roche).
RNA was extracted from isolated ring glands complexed with brain hemispheres placed in RNAlater (Ambion) and frozen in liquid nitrogen. At least two different pools of 50 to 100 specimens were collected per genotype. Total RNA extracts were obtained using the “mirVana miRNA isolation kit” (Ambion) according to the manufacturer's instructions and were quantified using Nanodrop (Thermo Scientific). cDNAs were prepared from 0.2 µg of RNA using the SuperScript III First-Strand Synthesis System for RT-PCR (Invitrogen) at a 10 µl volume per reaction, following manufacturer's instructions.
Oligonucleotides for pes, crq, Snmp1 and RpL32 were designed using NCBI primer blast (http://www.ncbi.nlm.nih.gov/tools/primer-blast). RpL32 was used as control. Oligo sequences were:
Pes(+)384Fwd, 5′-TCGCCGCTGCCTTTAGACTTCGATA-3′;
Pes(−)660Rev, 5′-CACGTCTAGCAGCAGAGTGCGCTAC-3′;
Crq(+)1747Fwd, 5′-GAGCCCGATGACGACTTCGACATAT-3′;
Crq(−)1967Rev, 5′-ACCCACTTTTTCGTCACAGTCAGCG-3′;
Snmp1(+)741Fwd, 5′-ATGGGTCAGGCCAATCACTCGGATT-3′;
Snmp1(−)935Fwd, 5′-CAGGCCATCCTCCTTTTTCAAGCCC-3′;
RpL32(+)365For, 5′-CCTTCCAGCTTCAAGATGACCATCC-3′;
RpL32(−)598Rev, 5′-ATCCGTAACCGATGTTGGGCATCAG-3′.
qPCR was done using FastStart Universal SYBR green Master (Roche). Reactions were performed in 10 µl, adding 2 µl of cDNA, 2× SYBR green and 0.2 µl of each primer (10 µM), in a CFX96-thermocycler (BioRad) with the following protocol: 95°C for 10 min, 40 cycles of 95°C for 10 seconds and 58.5°C for 1∶30 min, and a final extension of 95°C for 1 min. Per each pair of primers a melt curve from 65 to 95°C, with 0.5°C temperature increment every 5 seconds was done. Reactions were done in triplicates and checked by electrophoresis for validation of amplification specificity. RpL32 was used as control.
Drosophila S2R+ cells were seeded in 24-well plates and transfected with 150 ng of pSnmp1-FLuc or pSnmp1Mut-FLuc, either alone or co-transfected with the same quantities of pUASTattb-αftz-f1, pUASTattb-βftz-f1, pUASTattb-Smt3-αftz-f1 or pUASTattb-Smt3-βftz-f1. 150 ng of pAc-Gal4 [116] and 50 ng of pAc-Renilla were added to all the wells. Transcriptional activity was measured 48 hours after transfection using the Dual-Luciferase Reporter Assay System (Promega), following the manufacturer's instructions. Luminescence was measured in a microplate luminometer (Veritas). Results are given as means+S.D. Differences between groups were calculated using Student's t test in Microsoft Excel.
Adults were allowed to lay eggs during 8 hours and wandering larvae were collected 5 days AEL. Larvae and ovaries from adult flies were dissected in PBS, fixed in 4% paraformaldehyde for 20 minutes and washed three times in PBT (PBS, 0.3% triton X-100) for 20 minutes. Samples were blocked in PBT +1% BSA for one hour and incubated with the appropriate antibodies at 4°C overnight. Next day, tissues were washed with PBT three times, for 20 minutes each and incubated with secondary antibodies at room temperature for two hours. The following primary antibodies were used: guinea pig polyclonal anti-Cad99C (1∶3,000) [47]; mouse monoclonal DE-Cad (1∶25; DCAD2, Developmental Studies Hybridoma Bank); rabbit polyclonal anti-Smt3 (1∶500) [117]; goat polyclonal anti-Ftz-f1 (1∶25; Santa Cruz, Sc-27221); rabbit polyclonal anti-Crq (1∶100) [118]; rabbit polyclonal anti-Snmp1 (1∶1000) [119] and rabbit polyclonal anti-Dib (1∶200) [26]. Fluorescence Alexa 568 secondary antibody (Molecular Probes) was used at 1∶200 dilution. DAPI (Roche) was used at 1∶2000 dilution. Phalloidin-TRITC (Sigma) was used 1∶1000. Samples were mounted in Vectashield (Roche) mounting medium. Confocal images were taken with a Leica DM IRE2 confocal microscope.
Ring glands were fixed in 4% paraformaldehyde for 20 minutes, washed twice in PBS and stained with Oil Red O (Sigma) solution at 0.06% in isopropanol for 30 minutes. Samples were washed twice in PBS before mounting in Vectashield. Images were taken with a Leica DM IRE2 confocal microscope.
Quantification of lipid droplets was done on single plane confocal micrographs of Oil Red O stainings using the Analyze Particle tool from ImageJ software. At least 10 independent micrographs were analyzed per genotype. Measurements were analyzed and plotted using Microsoft Excel.
phm-Gal4>UAS-ftz-f1i and phm-Gal4>UAS-Snmp1i larvae were collected at 120 hours AEL and placed in groups of 10 individuals in new tubes. These were supplemented with 20E (Sigma) dissolved in ethanol and mixed with yeast at 1 mg/ml. Control larvae were fed with yeast supplemented with ethanol alone.
|
10.1371/journal.pcbi.1003614 | Reciprocal Regulation as a Source of Ultrasensitivity in Two-Component Systems with a Bifunctional Sensor Kinase | Two-component signal transduction systems, where the phosphorylation state of a regulator protein is modulated by a sensor kinase, are common in bacteria and other microbes. In many of these systems, the sensor kinase is bifunctional catalyzing both, the phosphorylation and the dephosphorylation of the regulator protein in response to input signals. Previous studies have shown that systems with a bifunctional enzyme can adjust the phosphorylation level of the regulator protein independently of the total protein concentrations – a property known as concentration robustness. Here, I argue that two-component systems with a bifunctional enzyme may also exhibit ultrasensitivity if the input signal reciprocally affects multiple activities of the sensor kinase. To this end, I consider the case where an allosteric effector inhibits autophosphorylation and, concomitantly, activates the enzyme's phosphatase activity, as observed experimentally in the PhoQ/PhoP and NRII/NRI systems. A theoretical analysis reveals two operating regimes under steady state conditions depending on the effector affinity: If the affinity is low the system produces a graded response with respect to input signals and exhibits stimulus-dependent concentration robustness – consistent with previous experiments. In contrast, a high-affinity effector may generate ultrasensitivity by a similar mechanism as phosphorylation-dephosphorylation cycles with distinct converter enzymes. The occurrence of ultrasensitivity requires saturation of the sensor kinase's phosphatase activity, but is restricted to low effector concentrations, which suggests that this mode of operation might be employed for the detection and amplification of low abundant input signals. Interestingly, the same mechanism also applies to covalent modification cycles with a bifunctional converter enzyme, which suggests that reciprocal regulation, as a mechanism to generate ultrasensitivity, is not restricted to two-component systems, but may apply more generally to bifunctional enzyme systems.
| Bacteria often use two-component systems to sense and respond to environmental changes, which involves autophosphorylation of a sensor kinase and phosphotransfer to a cognate response regulator. However, despite conservation of this ‘classical’ scheme there exist substantial variations in the mechanism of phosphotransfer among systems. Also, many sensor kinases exhibit phosphatase activity raising the question whether such a bifunctional architecture enables special regulatory properties in the response behavior to input signals. According to previous studies, classical two-component systems are unlikely to produce sigmoidal response curves (ultrasensitivity) if the sensor protein is bifunctional. Here, I argue that this is not necessarily true if the input stimulus (allosteric effector) reciprocally affects multiple activities of the sensor kinase, as it seems to be common for bifunctional enzymes. To this end, I propose and analyze an extension of the experimentally well-supported Batchelor-Goulian model which shows that ultrasensitivity requires a high-affinity effector and saturation of the phosphatase activity. The underlying mechanism involves sequestration of the effector by the sensor kinase which restricts the occurrence of ultrasensitivity to sufficiently low effector concentrations. Hence, this operating regime might be useful to sense effector limitations or to amplify weak input signals.
| Two-component systems (TCSs) are modular signal transduction systems which are utilized by bacteria and other microbes to respond to intracellular or environmental stimuli [1], [2]. ‘Classical’ TCSs consist of a sensor histidine kinase (HK) and a cognate response regulator (RR), which often acts as a transcription factor to activate or repress a particular set of response genes. Upon stimulation, the HK autophosphorylates at a conserved histidine residue and transfers the phosphoryl group to an aspartate residue in the receiver domain of the RR. Often, the unphosphorylated form of the HK also exhibits phosphatase activity towards the phosphorylated form of the RR (RR-P) endowing many HKs with a bifunctional design (Fig. 1). In addition, some RRs exhibit intrinsic phosphatase activity which leads to autodephosphorylation of RR-P with a half-life ranging between seconds to hours [1].
Even though the overall signal flow from the sensor kinase to the response regulator is well-conserved between different systems there exist substantial variations in the particular mechanism through which the phosphoryl group is transferred to the regulator protein [3]. To better understand their regulatory properties it has become a useful strategy to compare different TCS architectures based on their potential input-output behavior. Following that strategy, it has been argued that phosphorelay systems, where the phosphotransfer to the RR does not occur in a single step but via additional intra- or intermolecular reactions [4], may generate ultrasensitivity and robustness against noise [5]. Systems with a split histidine kinase comprise another class of TCSs where a functional HK is generated through binary association between two distinct proteins each of which alone would not be able to phosphorylate the cognate RR(s) [6]. A theoretical study suggested that such systems can potentially exhibit ultrasensitivity and bistability if the phosphatase activity is predominantly located on the free form of one of the proteins making up the split kinase [7]. Yet another study compared TCSs with a mono- and a bifunctional HK arriving at the conclusion that ultrasensitivity and bistability can also occur in classical TCSs if the unphosphorylated forms of the HK and the RR form a dead-end complex and if the dephosphorylation of the RR mainly occurs via an HK-independent phosphatase [8].
In contrast, systems with a bifunctional design are expected to generate graded responses to input stimuli [8]–[10] and to mediate concentration robustness [11]–[13]. The latter property means that the system response (concentration of phosphorylated RR) is invariant with respect to variations of the total RR and HK concentrations, at least in a certain range of concentrations. Moreover, based on theoretical studies of covalent modification cycles with a bifunctional converter enzyme it has been argued that ultrasensitivity is unlikely to occur in such systems if the bifunctional enzyme employs only a single catalytic site for its opposing activities [14], [15]. Based on this argument it, thus, appears unlikely that classical two-component systems with a bifunctional sensor kinase would exhibit ultrasensitivity given that the phosphotransferase and phosphatase activities of the sensor kinase are believed to occur on a single catalytic site in the dimerization domain of the protein [16], [17]. Interestingly, this conclusion does not apply to bifunctional enzymes with two distinct catalytic sites where ultrasensitivity may arise from the formation of a ternary complex between the enzyme and its two substrates [18] as observed experimentally in the uridylylation cycle of the PII protein [19].
In the present study, I wish to argue that ultrasensitivity may still occur in two-component systems with a bifunctional enzyme kinase if the input signal reciprocally affects multiple activities of the sensor kinase. Reciprocal regulatory patterns have been observed in the PhoQ/PhoP system which mediates adaption in response to limitation as well as in the NRII/NRI system which mediates adaptation to nitrogen limitation by sensing the concentration of deuridylylated PII protein in the cytosol. In both cases, binding of an allosteric effector ( or PII) inhibits the autokinase activity and, concomitantly, activates the phosphatase activity of the respective sensor protein (Fig. 2A) [20], [21]. Indeed, based on structural analysis of HK domains it has been argued that reciprocal regulation could be quite common for bifunctional enzymes [17].
In a first step, the impact of reciprocal regulation is analyzed in covalent modification cycles with a bifunctional converter enzyme, which will serve as a ‘toy’ model that allows for an intuitive understanding of the potential mechanism for the generation of ultrasensitivity. In a second step, it will be shown that the same mechanism may also generate ultrasensitivity in classical TCSs with a bifunctional sensor kinase. To this end, an extension of the experimentally well-supported Batchelor-Goulian model (see below) is proposed which assumes that autokinase and phosphatase activities of the HK are reciprocally regulated by an allosteric effector (Fig. 2C). Analysis of this model shows that if the affinity of the effector is low (as in the case of for PhoQ) the system exhibits a graded response to changes in the effector concentration and stimulus-dependent concentration robustness – in agreement with experiments in the PhoQ/PhoP system [22]. In contrast, a high-affinity effector may lead to ultrasensitivity at low effector concentrations, but requires saturation of the sensor kinase's phosphatase activity. Comparison of the model predictions with in vitro experiments suggests that in the NRII/NRI system the occurrence of ultrasensitivity is (partly) suppressed by the intrinsic autophosphatase activity of NRI.
To rationalize the occurrence of concentration robustness in the EnvZ/OmpR system of E. coli, Batchelor and Goulian proposed a simple mathematical model based on the three activities of the bifunctional EnvZ (denoted by HK in Fig. 2B). Guided by the observation that the total OmpR concentration is much larger than that of EnvZ [23] () they have argued that, in the limit , the steady state concentration of OmpR-P (denoted by in Fig. 2B) is determined by a quadratic equation [11], which can be written in the form (SI Text S1)(1)Here, denotes the total OmpR concentration, and the parameters and are proportional to the Michaelis-Menten constants associated with the phosphatase () and phosphotransferase () reactions. Note that Eq. (1) does not depend on the total EnvZ concentration (). Hence, the Batchelor-Goulian model predicts that, in the limit , the concentration of OmpR-P is approximately independent of variations in the total concentration of the sensor kinase, i.e. [OmpR-P] exhibits (concentration) robustness with respect to changes in .
Interestingly, Eq. (1) also predicts concentration robustness of with respect to the total concentration of the response regulator () under certain conditions. To see this more explicitly, it is worth mentioning that a structurally similar equation has been analyzed previously in the context of concentration robustness for covalent modification cycles with a bifunctional converter enzyme [24]. This analysis has shown that the shape of the stimulus-response curve, described by Eq. (1), depends on the relative magnitude between the two parameters and [18]. To this end, it is useful to consider two limiting cases corresponding to and . In the first case, the physiologically reasonable solution of Eq. (1) can be approximated by (SI Text S1)(2)whereas, in the second case, one obtains the approximate solution(3)In any case, from the expressions in Eqs. (2) and (3) it is readily apparent that becomes independent of the total RR concentration if the latter is sufficiently large, i.e. if (Eq. 2) or (Eq. 3). Hence, if , the parameter determines both, the threshold concentration beyond which becomes approximately constant as well as the value of that constant. In contrast, if , the predicted threshold concentration () is much larger than the asymptotic phosphorylation level of the response regulator (). Also, the approach to the asymptotic level is different for the two regimes: If , increases approximately linearly with up to the threshold (Eq. 2) whereas, in the opposite case, it increases hyperbolically (Eq. 3). Due to the linear relationship between and in Eq. (2) the regime has been called ‘signal-transducing’ in Ref. [25].
Together, Eqs. (2) and (3) suggest that there exist two different regimes for the occurrence of concentration robustness and, as will be shown below, there is experimental evidence for either case.
To test the predictions of their model, Batchelor and Goulian measured changes in the transcriptional activity of OmpR-controlled genes using a two-fluorescent reporter strain, which provided indirect evidence for concentration robustness of OmpR-P. Recently, Gao and Stock directly confirmed the predictions of the Batchelor-Goulian model in the PhoR/PhoB system using a Phos-tag based method allowing for a quantification of the PhoB-P levels as a function of total PhoB amounts [26]. Experiments were performed with the wild-type (WT) system as well as with a PhoB mutant () which exhibits reduced interaction strength (affinity) with PhoR. Both measurements could be well described by Eq. (1) with a ratio varying between 0.1–0.2 (Fig. 3A, solid lines). Overlaying the response curves with the respective values (dotted lines) indicates that the PhoR/PhoB system operates in the regime since the threshold concentration (), beyond which PhoB-P becomes constant, is approximately equal to the value of that constant, as expected from Eq. (2). The observed shift of the threshold concentration in the mutant strain results from the reduced affinity of which is associated with a larger value for . Since , increasing leads to an increased value of so that the asymptotically constant phosphorylation level of is reached at higher total PhoB concentrations, i.e. for total (Fig. 3A).
Concentration robustness has also been observed in the reconstituted NRII/NRI system of E. coli under in vitro conditions [27]. However, in that case the shape of the response curve is quite different (Fig. 3B): The dependence between [NRI-P] and total [NRI] does not appear to be linear below the threshold concentration and the asymptotically constant phosphorylation level () is only reached for very large values of total [NRI] (). Together, this indicates that the NRII/NRI system operates in the regime and, indeed, fitting the measurement data to Eq. (3) supports this view (Fig. 3B, solid line). Moreover, since in vivo concentrations of NRI are typically much lower than the threshold concentration of [28] it has been argued that, in the NRII/NRI system, concentration robustness will most likely not play a role under physiological conditions [27].
To understand how ultrasensitivity may arise in TCSs with a bifunctional HK it will be helpful to analyze the consequences of reciprocal regulation in a related, but more simple system first. To this end, the reaction mechanism in Fig. 4A, which describes the reversible phosphorylation of a substrate by a bifunctional enzyme , is considered. The enzyme exhibits both, kinase () and phosphatase () activities, which catalyze the phosphorylation () and dephosphorylation reactions (), respectively. The transition between the two activity states is mediated through binding of an allosteric effector . For simplicity, it is assumed that has no phosphatase activity and, conversely, has no kinase activity so that effector-binding effectively inhibits the enzyme's kinase activity and, concomitantly, activates its phosphatase activity. Note that this system is similar to TCSs with a bifunctional sensor kinase where the autophosphorylation and phosphotransfer reactions are replaced by a covalent modification (cf. Figs. 2C and 4A). Also, the bifunctional converter enzyme is assumed to have just a single catalytic site, which is supposed to mimic the fact that the phosphotransferase and phosphatase activities of the sensor kinase in TCSs are also likely to occur on a single catalytic site [17].
The dynamics of this system is described by the set of ordinary differential equations (ODEs)(4)together with the conservation relations for the total concentrations of substrate (), converter enzyme () and allosteric effector ()(5)(6)(7)
If the substrate concentration is much larger than that of the converter enzyme (), one can neglect the concentrations of the enzyme-substrate complexes (since by Eq. 6) in the conservation relation for the substrate (Eq. 5), and the concentration of unmodified substrate can be expressed as(8)For later comparison, it will be useful to employ the quasi-steady state approximation (QSSA) in order to derive an effective equation for . By construction, the QSSA preserves the steady state structure of the underlying system [29] (which is the main focus here) although, for a better description of the transient dynamics, application of the total QSSA may be advantageous [30]. To apply the QSSA, it is assumed that, after a short transient period, the enzyme-substrate and the enzyme-effector complexes reach a quasi-steady state, defined by , and , which leads to the algebraic relations(9)Here, and denote Michaelis-Menten constants associated with the kinase and phosphatase activities, respectively, whereas denotes the dissociation constant for the enzyme-effector complex.
Using the QSSA condition , it follows that(10)where Eqs. (4), (8) and (9) have been used. In Eq. (10), and have to be found as functions of from the conservation relations (Eqs. 6 and 7)(11)(12)Intuitively, it is clear that if the effector concentration is sufficiently large () the amount of effector that can be sequestered by the enzyme will be small since . Under this condition the conservation law for the effector (Eqs. 7 and 12) always reduces to independent of whether the binding affinity of the effector is high (if is small) or low (if is large). The latter only becomes important when the effector concentration is equal to or smaller than the enzyme concentration (), e.g. under effector-limiting conditions. In the following, it will be shown that the type of effective equation, that is obtained from Eqs. (10)–(12), depends on the ratio which may be regarded as a relative binding affinity for the enzyme-effector complex.
The Batchelor-Goulian model is based on the three activities of the sensor kinase shown in Fig. 2B, i.e. it essentially focuses on the signal transduction layer in the general scheme for two-component signaling depicted in Fig. 1. However, within the context of this model it may become difficult to predict the input-output behavior as a function of the input stimulus, especially if the latter affects multiple enzyme activities as observed in the PhoQ/PhoP/ and NRII/NRI/PII systems (Fig. 2A). Guided by these examples the Batchelor-Goulian model will be extended by incorporating a mechanism that accounts for reciprocal regulation of the sensor kinase's autokinase and phosphatase activities by an allosteric effector. Analysis of this model shows that a low-affinity effector may lead to stimulus-dependent concentration robustness whereas a high-affinity effector may generate ultrasensitivity. In the latter case, the underlying mechanism is essentially the same as for covalent modification cycles (cf. Fig. 4).
To implement reciprocal regulation it is assumed (cf. Fig. 2C) that, in the absence of the effector, the free form of the sensor kinase () can undergo autophosphorylation and mediates the phosphotransfer to the response regulator (step 1 and 2), but does not exhibit phosphatase activity (step 3). The latter is assumed to be activated through effector-binding (step 4), so that the phosphatase activity is carried by the ligand-bound form of the sensor kinase. Since cannot undergo autophosphorylation (and phosphotransfer) binding of the ligand effectively leads to inhibition of the HK's autokinase activity and, concomitantly, activates its phosphatase activity.
The dynamics of the extended model, as shown in Fig. 2C, is described by the five ODEs(22)(23)(24)(25)(26)together with the three conservation relations(27)(28)(29)where , and denote the total concentrations of response regulator, histidine kinase and effector, respectively. Measurements in the PhoQ/PhoP and NRII/NRI systems have shown that the ratio between the total concentrations of RR and HK is large () [22], [28] in which case one can use the simplified conservation relation (cf. Eq. 8)(30)instead of Eq. (27). Similar as in the case of covalent modification cycles (Eqs. 10–12), the steady state behavior of the system, described by Eqs. (22)–(30), depends on the affinity of the allosteric effector () relative to the total enzyme concentration ().
Note that for the derivation of Eqs. (22)–(30) it has been assumed that signal-sensing and the reactions describing the catalytic activities of the sensor kinase take place in the same compartment (the cytoplasm of the cell). Hence, this model directly applies to cytosolic TCSs, such as the NRII/NRI system, but not to systems with a transmembrane sensor kinase where signal-sensing typically occurs in a different compartment. For example, in the PhoQ/PhoP system the sensor kinase PhoQ responds to changes of the concentration in the periplasm [20]. However, since effector-binding does not involve mass transfer the conditions for the occurrence of concentration robustness and ultrasensitivity are essentially the same (up to a factor accounting for the different compartment volumes) as those which are derived below on the basis of Eqs. (22)–(30) (see Methods).
In many two-component systems, the phosphorylation level of the response regulator protein is modified by a bifunctional sensor kinase which, apart from exhibiting autokinase and phosphotransferase activity, also catalyzes the dephosphorylation of the response regulator through a phosphatase activity. In the present study, I have argued that the spectrum of potential input-output behaviors of such bifunctional systems does not only comprise graded responses [8]–[10] and concentration robustness [11], [12], but also ultrasensitivity as it is well-known from phosphorylation-dephosphorylation cycles with distinct converter enzymes [31]. To this end, I have proposed and analyzed an extension of the Batchelor-Goulian model [11] which considers the biologically motivated case where the autokinase and phosphatase activities of the sensor kinase are reciprocally regulated by an allosteric effector (Fig. 2).
The analysis of the extended model showed that there exist two operating regimes under steady state conditions depending on the effector affinity: If the affinity is low compared to the total concentration of the sensor kinase () the system produces a graded response to changes in the effector concentration (Eqs. 33 and 34) and exhibits stimulus-dependent concentration robustness, which means that the maximal phosphorylation level of the response regulator does not only depend on kinetic model parameters (as in the original Batchelor-Goulian model), but also on the effector concentration. Consistent with experiments in the PhoQ/PhoP system [22], the extended model predicts an increase in the maximal phosphorylation level as the effector concentration is lowered (Eq. 32). However, if the effector affinity is sufficiently high () the steady state equation for the extended model (Eq. 35) becomes structurally identical to that for covalent modification cycles with distinct converter enzymes (Eq. 19) so that ultrasensitivity may arise from the zero-order effect [31].
Apart from enzyme saturation due to the zero-order effect, sequestration of a signaling molecule into an inactive complex represents an alternative mechanism for the generation of ultrasensitivity in signal transduction networks [33]–[35]. Often, sequestration involves a reaction of the form(41)where, by definition, is sequestered by into the complex . In this sense, regulation of enzyme activities by an allosteric effector may also be regarded as a form of sequestration. In the case of reciprocal regulation shown in Fig. 2C, the enzyme-effector complex () is not catalytically inactive, but rather has a different activity compared to the free form of the enzyme (). Buchler and Louis have shown that the simple mechanism in Eq. (41) can give rise to ultrasensitivity in the concentrations of and if the stoichiometric binding parameter (where ) exceeds unity, and the degree of ultrasensitivity increases as [36]. In the present study, the stoichiometric binding parameter () plays a different role for the generation of ultrasensitivity since the condition does not guarantee the occurrence of ultrasensitivity per se, but only the validity of the reduced model, described by the steady state equation in Eq. (35). To obtain ultrasensitivity within the reduced model, the (apparent) Michaelis-Menten constants for the phosphotransferase and phosphatase activities of the sensor kinase also have to be sufficiently small (Eq. 37), which distinguishes the mechanism, proposed in the present study, from purely sequestration-based mechanisms.
Interestingly, the idea of reciprocal regulation, as a mechanism to generate ultrasensitivity, does not seem to be restricted to two-component systems as the same mechanism may also apply to covalent modification cycles with a bifunctional converter enzyme (Fig. 4A). In both cases, reciprocal regulation may lead to ultrasensitivity if the stoichiometric binding parameters ( in the case of covalent modification cycles or in the case of two-component systems) are sufficiently large. In this case, almost all free effector molecules are bound to the respective enzyme which leads to a tight partition of enzyme states into those with phosphatase activity and those with kinase activity (cf. Eqs. 20 and 52). As a consequence, the system behaves as if phosphorylation and dephosphorylation were catalyzed by independent enzyme subpopulations, which rationalizes why the corresponding steady state equations (Eqs. 18 and 35) are structurally identical to that for covalent modification cycles with distinct converter enzymes (Eq. 19). However, this mechanism only ‘works’ as long as the enzyme is not saturated by the effector, which restricts the occurrence of ultrasensitivity to effector concentrations that are smaller than that of the respective enzyme (Figs. 4B and 6A).
To assess the potential relevance of reciprocal regulation for the occurrence of ultrasensitivity under physiological conditions one has to evaluate to what extent the requirements for its occurrence (substrate excess, a large stoichiometric binding parameter and saturation of the sensor kinase's phosphatase activity) are satisfied in a particular system in vivo. Based on measurements in the EnvZ/OmpR, PhoQ/PhoP and PhoR/PhoB systems, it seems that the requirement of substrate excess does not represent a limitation for the occurrence of ultrasensitivity as response regulator proteins are typically much more abundant than their respective sensor proteins [22], [23], [26]. In contrast, estimation of the stoichiometric binding parameter appears more difficult due to the limited knowledge on the range of input signals for a particular sensor kinase and their affinities relative to the total enzyme concentration. In general, histidine kinases may sense different signals (such as ions, metabolites, small peptides or auxiliary proteins) with widely different affinities [3]. Hence, it is conceivable that the same system produces a graded response with respect to a low-affinity effector and an ultrasensitive response with respect to another effector with a high affinity. For example, apart from mediating adaptation to -limiting conditions the PhoQ/PhoP system is also involved in the regulation of bacterial virulence. This transcriptional program is initiated by antimicrobial peptides that seem to bind to the same periplasmic site in the sensor domain of PhoQ as , but with a 100-fold higher affinity [37], which could potentially shift the stoichiometric binding parameter into a regime where sigmoidal responses become possible.
The occurrence of ultrasensitivity also requires saturation of the sensor kinase's phosphatase activity which means that the Michaelis-Menten constant, associated with that activity, has to be smaller than the total concentration of the response regulator. Measurements in the EnvZ/OmpR system have shown that the dissociation constant for the EnvZ-OmpR complex is 5-fold smaller than the total OmpR concentration which indicates that enzyme saturation is, in principle, possible under physiological conditions [23]. However, the occurrence of ultrasensitivity can also be compromised by a sufficiently strong, unregulated phosphatase activity which may arise from a basal phosphatase activity of the sensor kinase (Fig. 7D) or from an autophosphatase activity of the response regulator. The latter might explain why the NRII/NRI/PII system exhibits only a weak sensitivity with respect to changes in the effector (PII) concentration (Fig. 8B). Alternatively, it has been speculated that the observed weak sensitivity results from a non-saturable phosphatase activity of NRII [27] which is consistent with the prediction that ultrasensitivity requires the phosphatase activity to operate in the zero-order regime (Eq. 37). On the other hand, it has been shown that single mutations in the dimerization domain of a sensor kinase can substantially affect its interaction strength with cognate and even non-cognate response regulator proteins [26], [38], which suggests that binding affinities between sensor kinases and response regulator proteins are highly evolvable. Hence, it is conceivable that one may employ directed evolution or site-directed mutagenesis to ‘adjust’ these binding affinities in a favorable range for ultrasensitivity to occur. In this sense, the results presented here may also guide the design of synthetic regulatory circuits which aim to implement ultrasensitive response behavior at the level of two-component systems [39].
Under steady state conditions, the right-hand sides of Eqs. (22)–(26) are set to zero so that summation of Eqs. (22) and (26) readily yields(42)Similarly, summation of Eqs. (23) and (26) leads to the steady state relation(43)where denotes the dissociation constant for the enzyme-effector complex. From Eqs. (24)–(26) together with Eqs. (30) and (43) one obtains the expressions(44)where is defined by(45)whereas and denote Michaelis-Menten constants associated with the phosphotransferase and phosphatase activities of the sensor kinase, respectively.
Using the expressions from Eqs. (43) and (44) in Eq. (42) and in the conservation relations, Eqs. (28) and (29), yields the set of algebraic equations(46)and(47)(48)from which the steady state concentrations , and have to be found.
Similar as in the case of a covalent modification cycle with a bifunctional enzyme the type of steady state solution, that is obtained from Eqs. (46)–(48), depends on the affinity of the allosteric effector. If this affinity is low () the concentration of free effector is approximately equal to the total effector concentration (). Replacing by in Eq. (46) readily yields the quadratic equation in Eq. (31) with and defined in Eq. (32).
In contrast, if the affinity of the effector is sufficiently high () the combination of Eqs. (47) and (48) yields a quadratic equation similar to that in Eq. (15)(49)where and denote the rescaled enzyme concentration and the relative binding affinity, respectively. In the limit , the solution of Eq. (49) can be approximated by [40](50)With this approximation the concentration of free effector becomes (cf. Eq. 48)(51)Using this expression for in Eq. (46) yields the equationwhich can be rewritten in the form shown in Eq. (35) of the main text.
Similar to the case of covalent modification cycles it is straightforward to show (using Eqs. 43–45, 50 and 51) that a high-affinity effector leads to a partition of enzyme states according to (cf. Eq. 20)(52)so that and may be regarded as apparent phosphatase and kinase concentrations, respectively.
For TCSs with a transmembrane sensor kinase autophosphorylation, phosphotransfer and dephosphorylation occur in the cytosol whereas signal-sensing typically takes place in the periplasm (for gram-negative bacteria) or directly in the extracellular space (Fig. 1). Hence, a proper model would have to distinguish at least 3 compartments: The cytosol (where the response regulator is located), the plasma membrane (to which the sensor kinase is confined) and the extracellular space (where the effector is located). For gram-negative bacteria one would also have to consider a periplasmic compartment as many sensor kinases seem to respond to signals in the periplasmic rather than directly in the extracellular space [3]. Together, this makes it difficult to propose a generic model for TCSs that are regulated by non-cytosolic effectors which will, therefore, not be attempted here.
Instead, to evaluate the impact of compartmentalization on the conditions for the occurrence of ultrasensitivity and concentration robustness it seems reasonable to consider (as a first approximation) a simplified model where the reactions describing the catalytic activities of the sensor kinase occur in the cytosol (similar as assumed in the original Batchelor-Goulian model) whereas binding of the effector to the regulatory site of the sensor kinase occurs either in the periplasm or in the extracellular space. Because effector-binding does not involve mass transfer between the extracellular space (or the periplasm) and the cytoplasm the equations for such a two-compartment model are essentially the same as those for a single compartment (Eqs. 22–30) if the mass-balance equations are written in terms of average molecule numbers (rather than concentrations). The corresponding ODE system then reads(53)where denotes the average amount of species (measured in ). Compared to Eqs. (22)–(26) the second-order rate constants , and are now measured in units of , i.e. they are independent of the volume of the compartment in which the corresponding reaction occurs. In contrast, first order rate constants (, , , , , and ) have the same unit () as before. Mass conservation is now expressed in terms of molecule number conservation for the total amount of response regulator (), sensor kinase () and effector () as(54)Since the structure of Eqs. (53) and (54) is identical to that of Eqs. (22)–(29) it is clear that the conditions for the occurrence of concentration robustness and ultrasensitivity are identical in both cases if concentration-based quantities are replaced by their respective molar counterparts.
Specifically, ultrasensitivity is predicted to occur if the amount of response regulator is much larger than that of the sensor kinase () and if the affinity of the effector is sufficiently high. The latter condition is now expressed as(55)where the dissociation constant is measured in . Under these conditions, the steady state amount of phosphorylated response regulator is determined by the analog of Eq. (35)(56)where and are defined by the same expressions as in Eq. (36). Similar as , the Michaelis-Menten constants and are measured in units of . Conversely, if the effector has a low affinity () the steady state amount of is determined by the analog of Eq. (31)(57)where the rescaled Michaelis-Menten constants and are defined by the same expressions as in Eq. (31).
To analyze the impact of the compartment sizes on the input-output behavior one has to rewrite Eqs. (56) and (57) in terms of concentration-based quantities. For this purpose, the concentrations of the response regulator and that of the sensor kinase(58)are measured with respect to the cytosolic volume , whereas the effector concentration(59)is measured with respect to the extracellular (or periplasmic) volume . In the case of an extracellular effector, one may think of as the effective volume that is accessible to each cell in a population. In general, the effective volume decreases as the number of cells increases, e.g. due to cell growth. However, for the present purpose will be taken as a constant parameter. In addition, it is assumed that the extracellular space is a well-mixed compartment so that effector-diffusion can be neglected.
Using the definitions in Eqs. (58) and (59), Eqs. (56) and (57) can be written in the form(60)and(61)where(62)denotes the ratio between the cytosolic volume and that of the extracellular (or periplasmic) space. Also, in Eqs. (60) and (61) the dissociation constant and the Michaelis-Menten constants have been rescaled according to(63)which gives them the conventional unit . The rescaling is motivated by the fact that, in a concentration-based description of chemical reactions, second-order rate constants have to be proportional to the volume of the compartment in which the corresponding reaction occurs [41], i.e. , and giving them units of .
Similar as Eq. (35), Eq. (60) predicts that ultrasensitivity may occur at low effector concentrations () if the affinity of the effector is sufficiently high (). The latter condition follows from Eq. (55) using that (Eq. 63) and (Eq. 58). Hence, depending on the volume ratio the occurrence of ultrasensitivity may be favored (if ) or suppressed (if ) compared to a system that is regulated by a cytosolic effector (for which ). For example, if regulation occurs via a periplasmic effector may vary between 1.5 and 4 corresponding to a periplasmic volume fraction of 20–40% of the total cell volume [42]. In contrast, if regulation occurs via an extracellular effector the volume ratio may be substantially smaller than 1 () (especially at low cell densities) which would make the condition less likely to hold and, therefore, suppress the occurrence of ultrasensitivity.
Interestingly, Eq. (61) does not explicitly depend on the volume ratio. Hence, if reciprocal regulation occurs via a low-affinity extracellular effector () the stimulus-response curves predicted by Eq. (61) are identical with those depicted in Fig. 5 if one replaces and by their extracellular (or periplasmic) counterparts and , respectively.
The response curves in Fig. 7C and 7D have been generated using the following set of equations (the corresponding reaction mechanism is shown in Fig. 7A and 7B)(64)where , and have to be replaced using the conservation relations
|
10.1371/journal.pcbi.1007154 | Resurgent Na+ Current Offers Noise Modulation in Bursting Neurons | Neurons utilize bursts of action potentials as an efficient and reliable way to encode information. It is likely that the intrinsic membrane properties of neurons involved in burst generation may also participate in preserving its temporal features. Here we examined the contribution of the persistent and resurgent components of voltage-gated Na+ currents in modulating the burst discharge in sensory neurons. Using mathematical modeling, theory and dynamic-clamp electrophysiology, we show that, distinct from the persistent Na+ component which is important for membrane resonance and burst generation, the resurgent Na+ can help stabilize burst timing features including the duration and intervals. Moreover, such a physiological role for the resurgent Na+ offered noise tolerance and preserved the regularity of burst patterns. Model analysis further predicted a negative feedback loop between the persistent and resurgent gating variables which mediate such gain in burst stability. These results highlight a novel role for the voltage-gated resurgent Na+ component in moderating the entropy of burst-encoded neural information.
| The nervous system extracts meaningful information from natural environments to guide precise behaviors. Sensory neurons encode and relay such complex peripheral information as electrical events, known as action potentials or spikes. The timing intervals between the spikes carry stimulus-relevant information. Therefore, disruption of spike timing by random perturbations can compromise the nervous system function. In this study we investigated whether the widely-distributed voltage-gated sodium (Na+) ion channels important for spike generation can also serve as noise modulators in sensory neurons. We developed and utilized mathematical models for the different experimentally inseparable components of a complex Na+ channel current. This enabled phenomenological simplification and examination of the individual roles of Na+ components in spike timing control. We further utilized real-time closed-loop experiments to validate model predictions, and theoretical analysis to explain experimental outcomes. Using such multifaceted approach, we uncovered a novel role for a resurgent Na+ component in enhancing the reliability of spike timing and in noise modulation. Furthermore, our simplified model can be utilized in future computational and experimental studies to better understand the pathological consequences of Na+ channelopathies.
| Real-time signal detection in uncertain settings is a fundamental problem for information and communication systems. Our nervous system performs the daunting task of extracting meaningful information from natural environments and guides precise behaviors. Sensory neurons for instance use efficient coding schemes such as bursting that aid in information processing [1]. Mathematical models of bursting have helped explain the basic structure of an underlying dynamical system as one in which a slow process dynamically modulates a faster action potential/spike-generating process, leading to stereotypical alternating phases of spiking and quiescence [2, 3]. The so-called recovery period of the slow process governs the intervals between bursts which is often susceptible to random perturbations. Uncertainty in spike/burst intervals can alter the timing precision and information in a neural code [4]. Consequently, ionic mechanisms that modulate the recovery of membrane potential during spike/burst intervals, can play a role in maintaining the stability of these timing events and aid neural information processing. Here we examined a candidate mechanism involving neuronal voltage-gated Na+ currents for a role in the stabilization of burst discharge (durations and intervals) and noise modulation.
Voltage-gated Na+ currents are essential for spike generation in neurons [5]. The molecular and structural diversity of Na+ channels and the resultant functional heterogeneity and complexity, suggest their role beyond mere spike generation [6]. For instance, in addition to the fast/transient Na+ current (INaT) mediating action potentials, a subthreshold activated persistent Na+ current (INaP) participates in the generation of subthreshold membrane oscillations (STO) (e.g., see [7]). These oscillations can lead to membrane resonance by which a neuron produces the largest response to oscillatory inputs of some preferred frequency [8, 9]. Neurons utilize this mechanism to amplify weak synaptic inputs at resonant frequencies [10]. The slow inactivation and recovery of INaP further provides for the slow process required for rhythmic burst generation [11–13] and therefore can contribute to efficient information processing in multiple ways. However, during ongoing activity, random membrane fluctuations can alter the precision and order of burst timings which can distort/diminish the information in neural code. Here, we provide evidence that a frequently observed resurgent Na+ current (INaR), often coexistent with INaP might be a mechanism by which neurons stabilize burst discharge while maintaining its order and entropy. The INaR, in neurons and other excitable cells is an unconventional Na+ current which physiologically activates from a brief membrane depolarization followed by repolarization, such as during an action potential [11, 14–17]. In the well-studied neuronal Nav1.6-type Na+ channels, such a macroscopic INaR is biophysically suggested to occur from an open-channel block/unblock mechanism [18, 19]. Consequently, INaR is known to mediate depolarizing after-potentials and promote high-frequency spike discharge in neurons [14, 20–24]. Sodium channels containing the Nav1.6 subunits carry all three types of sodium currents and are widely distributed in the central and peripheral neurons and participate in burst generation [14, 25]. Sodium channelopathy involving alteration in INaR and INaP, and its association with irregular firing patterns and ectopic bursting in disease (e.g., [26–29]), prompted us to investigate distinct roles for these Na+ currents in regulating bursting in sensory neurons.
Lack of suitable functional markers and experimental tools to dissociate the molecular mechanism of INaR from INaP, led us to use computational modeling and dynamic-clamp electrophysiology to examine a role for INaR in burst control; however see [19, 23, 24]. Although existing Markovian models model a single channel Nav1.6 type INa using a kinetic scheme (e.g., [30, 31]), they have limited application for studying exclusive roles of INaR and INaP in the control of neural bursting; however see [13, 23, 32, 33]. Here, we developed a novel mathematical model for INaR using the well-known Hodgkin-Huxley (HH) formalism which closely mimics the unusual voltage-dependent open-channel unblocking mechanism. We integrated the model INaR into a bursting neuron model with INaP that we previously reported to study their exclusive roles in burst control in the jaw proprioceptive sensory neurons in the brainstem Mesencephalic V (Mes V) nucleus [8]. To validate model predictions, we used in vitro dynamic-clamp electrophysiology, and theoretical stability analysis. Using these approaches, we identify a novel role for INaR in stabilizing burst discharge and noise modulation in these sensory neurons (see Fig 1 for a workflow and approaches used).
We first developed a novel HH-based model for the resurgent component of a Nav1.6-type Na+ current and combined it with our previous HH-type models for the transient and persistent Na+ currents [8]. Although the total Na+ current arises from a single channel [11, 30], we formulated the model INa as a sum of the three components as shown in Fig 2A (also see Methods): the transient, INaT, the resurgent, INaR, and, the persistent, INaP current. This enabled easy manipulation of individual components to study their exclusive role in burst control. We ensured that each of these components mimic the distinct voltage-dependencies and kinetics observed during voltage-clamp recording in Mes V neurons [11, 34]. The classic INaT with fast inactivation kinetics (order of 1 ms) mediates spike generation, the persistent INaP with slow inactivation/recovery (order of 1000 ms) mediates STO and provides for the slow process responsible for rhythmic bursting in Mes V neurons [8, 35, 36], and, the resurgent INaR which mimics the phenomenology of an open channel unblock mechanism with a peak time of ~6 ms and decay kinetics on the order of 10 ms [11, 19, 30] (also see Fig 3).
We compared our model INa and its components to the macroscopic INa observed in voltage-clamp recordings in Mes V neurons [11, 14] to ensure qualitative and quantitative similarities (see Fig 3B). We further ensured that our INa model closely resembled the INa generated in Markovian models [30], which follow a kinetic scheme and do not formulate the three components separately (see S1 Fig). Our model INa also satisfied previously described contingency of Nav1.6-type Na+ channels carrying resurgent current: activation of INaR only up on brief depolarization followed by hyperpolarization to ~ -40 mV, and the correspondence between maximum channel availability and peak time of resurgent current30 (see S2 Fig).
We incorporated the model INa into a conductance-based single-compartment neuron model (see schematic and membrane voltage trace in black in Fig 2B). Together with a minimal set of Na+, K+ and leak conductances, the model neuron faithfully reproduced an expected rhythmic burst discharge observed in Mes V sensory neurons; the total INa generated during action potentials is shown in expanded time in the figure (red trace). Fig 2C shows the dynamic-clamp experiment in an intrinsically bursting Mes V neuron. The control burst was generated by simply driving the neuron with a step depolarization, following which we blocked action potential generation by bath application of tetrodotoxin (1 μM TTX) (see black horizontal bar in Fig 2C). Subsequently, we introduced the real-time model INa using dynamic-clamp during TTX application and by adjusting the conductances of the three INa components suitably, we were able to regenerate action potential bursts (see Methods on the choice of conductance values). The dynamic-clamp INa generated during action potentials is shown in expanded time in the figure for comparison with the model simulation in Fig 2B (red trace).
As noted earlier, the total INa in our model has the novel resurgent component, INaR; the transient and persistent components are similar to our previous report [8]. Fig 3A (left panel) illustrates a well-accepted mechanism of Na+ resurgence [30], wherein a putative blocking particle occludes an open channel following brief depolarization such as during an action potential; subsequently upon repolarization, a voltage dependent unblock results in a resurgent Na+ current. Our INaR formulation recapitulates this unusual behavior of Na+ channels using nonlinear ordinary differential equations for a blocking variable (br) and a competing inactivation (hr) (see Methods). Different from a traditional activation variable of an ionic current in HH models, we formulated the INaR gating using a term (1−br) to enable an unblocking process (see INaR equation in Methods and in Fig 3A). Here, the model variable br reflects the fraction of channels in the blocked state at any instant, and ‘1’ denotes the maximum proportion of open channels, such that (1−br) represented the fraction of unblocking channels. Such a formulation enabled mimicking the open-channel unblocking process as follows: Normally the br is maintained ‘high’ such that (1−br) term is small and therefore no INaR flows, except when a spike occurs, which causes br to decay in a voltage-dependent manner. This turns on the INaR which peaks as the membrane voltage repolarizes to ~ -40 mV, following which the increasing inactivation variable hr gradually turns off the INaR. The steady-state voltage dependency of unblock, (1−br∞(V)), and the competing inactivation (hr∞(V)) in the model are shown in Fig 3A (middle panel), along with the equation for INaR; the magenta shaded region highlights the voltage-dependency of INaR activation as posited to occur during open-channel unblocking. In Fig 3A (right panel), we show simulated INaR (in magenta), peaking during the recovery phase of spikes (in black). In Fig 3B (I.), we reproduced experimentally observed INa during voltage-clamp recording and highlight the resurgent component in both model (left) and experiment (right), (inset shows experimental protocol; also see legend and Methods). A comparative current-voltage relationship for the model and experiments is shown in Fig 3B (II.); also see S3 Fig for detailed kinetics of model INaR. Taken together, the above tests and comparisons ensured the suitability of our model for further investigation of INaR mediated burst control.
Given that INaR is activated during the recovery phase of an action potential, physiologically, any resulting rebound depolarization may control the spike refractory period, and increase spike frequency and burst duration [22, 37]. We tested this by selectively increasing the maximal resurgent conductance gNaR in our model neuron simulation and validated the predictions using dynamic-clamp experiments as shown in Fig 4A and 4B. In parallel, we also exclusively modified the maximal persistent conductance gNaP using model simulations and verified the effects using dynamic-clamp experiments as shown in Fig 4C and 4D. We only focused on rhythmically bursting Mes V neurons and quantified the burst timing features including the inter-burst intervals (IBIs), burst duration (BD) and inter-spike intervals (ISIs) as illustrated in the boxed inset in Fig 4. Figure panels 4e –j show a comparison of the exclusive effects of gNaR versus gNaP on each of these burst features. These experimental manipulations using model currents and quantification of resulting burst features revealed significant differences and some similarity between the action of gNaR and gNaP in burst control. First, increases in gNaR resulted in longer IBIs, which was in contrast with the effects of the persistent Na+ conductance, gNaP, which decreased IBIs (effect highlighted with red double arrows in Fig 4A–4D and quantified using box plots in Fig 4E and 4H). Alternatively, increasing gNaR reduced ISIs, while gNaP had the opposite effect on these events, resulting in an overall increase in ISIs with gNaP increases (see Fig 4F and 4I); whereas gNaR and gNaP had similar effect in increasing BDs (see Fig 4G and 4J). Box plots in Fig 4E–4J show 1st, 2nd (median) and 3rd quartiles; error bars show 1.5x deviations from the inter-quartile intervals. For each case (or cell), n values represented events (ISI, IBI and BD) and reported here are within-cell effects for different gNaR and gNaP applications during 20 sec step-current stimulation. For the gNaR applications (or series), IBI mean ± std were 210.49 ± 69.33 (control, n = 7), 450.29 ± 116.56 (1x, n = 26) and 1074.06 ± 199.17 (2x, n = 11), and for gNaP applications, these values were 1448.71 ± 450.92 (control, n = 4), 910.10 ± 527.27 (1x, n = 8) and 644.35 ± 234.19 (1.5x, n = 6). For gNaR series, ISI mean ± std were 20.26 ± 1.65 (control, n = 59), 12.56 ± 2.25 (1x, n = 646) and 10.56 ± 1.81 (2x, n = 783), and for gNaP series these values were 12.76 ± 1.13 (control, n = 159), 12.57 ± 1.05 (1x, n = 611) and 13.56 ± 1.42 (1.5x, n = 503). For gNaR series, BD mean ± std were 176.36 ± 69.57 (control, n = 8), 300.56 ± 77.90 (1x, n = 27) and 671.16 ± 285.96 (2x, n = 12), and for gNaP series these values were 346.59 ± 88.39 (control, n = 4), 571.30 ± 197.56 (1x, n = 8) and 1334.20 ± 592.59 (1.5x, n = 7). Treatment effects and group statistics for all the replicates showing the above effects across six Mes V bursting neurons, each from a different animal are summarized in S2 Table. A one-way ANOVA was used to test the treatment effects of gNaR and gNaP applications, and when significant, a post hoc two-sample Student t-test was used for group comparisons between control and 1x, 1x and 2x, and control and 2x for gNaR cases, and between control and 1x, 1x and 1.5x, and control and 1.5x for gNaP cases. Asterisks above box plots between groups in Fig 4E–4J indicate p<0.05 using two-sample Student t-tests for post-hoc comparisons. Furthermore, as shown in Fig 4, the model simulations predicted consistent effects (note white circles denote predicted values in panels e—j) with dynamic-clamp experiments, making the model suitable for further analysis of INaR/INaP mediated mechanism of burst control (see white circles in panels 4e - j). In two additional bursting neurons, we also conducted gNaR and gNaP subtraction experiments which showed consistent reverse effects of additions (see S4 Fig and legend). In the gNaP subtraction experiment, note that a -2x gNaP resulted in the abolition of bursting and subthreshold oscillations as shown in the figure. Such an effect was reproduced in the neuron model by setting gNaP = 0, as shown in S4C Fig.
To test whether INaR and INaP differentially modulate the regularity and precision of spike timing we observed the inter-event intervals (or IEIs) during real-time addition of gNaR and gNaP. As shown in Fig 5A and 5B, addition of gNaR improved the regularity of the two types of events: the longer IBIs and the shorter ISIs, whereas, addition of gNaP did not show such an effect. Note that application of real-time gNaR improved both the timing precision and regularity of IBIs (compare left and right panels of Fig 5A), while application of dynamic-clamp gNaP did not seem to affect either of these features (compare left and right panels of Fig 5B). Furthermore, in Fig 5C we highlight that gNaR application also improved spike-to-spike regularity of ISIs within bursts (shown are two representative bursts for gNaR (Fig 5C, left panel) versus gNaP (Fig 5C, right panel) application from Fig 5A and 5B (see figure legend). Taken together, these opposite effects of persistent and resurgent Na+ currents arising from a single Na+ channel may act in concert to offer a push-pull modulation of burst timing regulation.
The effect of persistent gNaP in reducing IBIs can be explained by its subthreshold activation [11], wherein increases in gNaP promotes STO and burst initiation which in turn increases burst frequency and reduces IBIs. Additionally, a high gNaP together with its slow inactivation helps maintain depolarization which can increase BDs. During spiking, INaP inactivation slowly accumulates and can contribute to any spike frequency adaptation as a burst terminates. Further increases in gNaP can accentuate such an effect and lead to increases in spike intervals within a burst. In contrast, the rebound depolarization produced due to INaR can decrease ISIs and promote spiking which prolongs the BDs. An effect which is not immediately obvious is an increase in IBI due to increases in gNaR since IBIs are order of magnitude slower events than rise/decay kinetics of INaR. Moreover, this current is not active during IBIs. It is likely that INaR which promotes spike generation produces its effect on lengthening the IBIs by further accumulation of INaP inactivation during spiking and in turn reducing channel availability immediately following a burst. We examined this possibility using model analysis of INaR’s effect on modulating the slow INaP inactivation/recovery variable, hp.
The simulated membrane potential (grey traces) and the slow INaP inactivation/recovery variable, hp (overlaid magenta traces) of the model neuron under three conditions shown in Fig 6A–6C: 1) with default values of gNaR and gNaP (Fig 6A), 2) an 1.5x increase in gNaP (Fig 6B), and, 3) a 2x increase in gNaR (Fig 6C). The peak and trough of the slow inactivation/recovery variable, hp correspond to burst onset and offset respectively. Comparing these traces in the three panels, we note that an increase in gNaR effectively lowers the hp value at which burst terminates (see curvy arrow in Fig 6C and legend). This observation was further supported by estimations of theoretical thresholds for burst onset and offset for increasing values of gNaR (Fig 6D) and, similar thresholds for increasing values of gNaP are provided for comparison in Fig 6E (see legend and S1 Text for details). Note that changes in gNaR did not alter the burst onset thresholds, consistent with a lack of resurgent current before spike onset (see brown arrow indicating burst onset threshold in Fig 6D). In contrast, increasing gNaR, consistently lowered the threshold values of slow inactivation/recovery for burst offset (see highlighted dashed box with arrows pointing to the burst offset thresholds decreasing with increasing gNaR values in Fig 6D). The net effect is longer recovery time between bursts and therefore prolonged IBIs. Additionally, in Fig 6D, an increase in gNaR extended the range of slow inactivation/recovery for which stable spiking regime exists (marked by the green circles). This gain in stability is indicative of a negative feedback loop in the Na+ current gating variables which model the slow inactivation and the open-channel unblock process. As shown in Fig 6F (see boxed inset), during a burst, presence of a channel unblocking process and the resulting resurgent Na+ can lead to further accumulation of slow inactivation (a positive effect) which eventually shuts off the unblocking events with further inactivation (negative feedback), and terminates a burst. The schematic on the left in Fig 6F summarizes a negative feedback loop between the unblocking and slow inactivation processes of Na+ currents which could mediate the stabilization of burst discharge as described above.
Next, we examined whether the stability of burst discharge offered by the presence of INaR can also contribute to noise tolerance. During quiescence/recovery periods between bursts, the membrane voltage is vulnerable to perturbations by stochastic influences, which can induce abrupt spikes and therefore disrupt the burst timing precision. We introduced a Gaussian noise to disrupt the rhythmic burst discharge in the model neuron when no gNaR was present as shown in Fig 7A and 7B. Subsequent addition of gNaR restored burst regularity (Fig 7C). Model analyses in the (hp,V) plane provides insight into the mechanism of INaR mediated noise tolerance. Briefly, we project a portion of a (hp,V) trajectory corresponding to the termination of one burst until the beginning of the next (see expanded insets in Fig 7A, 7B and 7C) to the (hp,V) diagrams shown in Fig 7D, 7E and 7F respectively (see S1 Text for details). In Fig 7D, beginning at the magenta circle, the (hp,V) trajectory (magenta trace) moves to the right as hp recovers during an IBI, until a burst onset threshold is crossed; point where the blue circles meet the red and black curves (see S1 Text for details), and eventually bursting begins; see upward arrow marking a jump-up in V at the onset of burst. During a burst, while V jumps up and-down during spikes, hp moves to the left as slow inactivation accumulates during bursting (left arrow). Finally, when hp reduces sufficiently, (hp,V) gets closer to the burst offset threshold (points at which the green and blue circles meet), and the burst terminates (down arrow). What is key in this figure is that the IBI is well-defined as the time period in which the (hp,V) trajectory moves along the red curve of steady states during the recovery process and moves past the burst onset threshold until a burst begins. However, when stochastic influences are present, the recovery period near-threshold is subject to random perturbations in V and can cause abrupt jump-up/spikes during the recovery period (see expanded inset in Fig 7B). Projecting (hp,V) during this period on to Fig 7E, we note that the near-threshold noise amplitudes can occasionally push the (hp,V) trajectory (magenta) above a green region of attraction and this results in such abrupt spikes. Now, when gNaR is added, the apparent restoration of burst regularity (see Fig 7C) can be attributed to an expansion in this green shaded region as shown in Fig 7F (see arrow pointing to a noise-tolerant region). In this situation, near-threshold random perturbations have less of an effect during the recovery process to induce abrupt spikes. This way, the net effect of INaR on slow Na+ inactivation prevents abrupt transitions into spiking regime following burst offset and in turn contributes to burst refractoriness and noise tolerance. We suggest that such a mechanism can make random fluctuations in membrane potential less effective in altering the precision of bursts and therefore aid information processing.
The spike/burst intervals, their timing precision and order are important for information coding [38–42]. Given our prediction that INaR can offer noise tolerance and stabilize burst discharge, we examined whether it can reduce uncertainty in spike/burst intervals and restore order in burst discharge. We tested this using model simulations and dynamic-clamp experiments as shown in Fig 8A–8D. In both cases, as shown by spike raster plots in Fig 8E and 8F, we disrupted the inter-event intervals (IEIs) by additive White/Wiener noise input while driving rhythmic burst discharge using step depolarization (also see Methods). Subsequent addition of gNaR conductance restored the regularity of rhythmic bursting. We used Shannon’s entropy as a measure of uncertainty in IEIs and show that increases in entropy due to noise addition was reduced to control levels by subsequent increases in gNaR as shown in Fig 8G (see Methods). We also quantified the Coefficient of Variation (CV) and noted that adding noise which shortened IBIs, indeed decreased the CV, due to a reduction in the standard deviation (s.d.) of the IEI distribution. Subsequent addition of INaR, which significantly lengthened the IBIs, resulted in increases in CV values due to an increase in IEI s.d. Taken together, INaR moderated burst entropy and improved the regularity of spike/burst intervals.
Using a combination of mathematical modeling and simulations, theory and dynamic-clamp electrophysiology, we demonstrate a novel role for voltage-gated Na+ currents in burst control, noise modulation and information processing in sensory neurons. While the subclasses of Na+ currents presented here are experimentally inseparable, our phenomenological simplification using the HH-formalism with in silico knock-in of each Na+ current component using dynamic-clamp experiments, allowed examination of their individual contributions to burst control. Additionally, theoretical analysis suggested a putative negative feedback loop between the persistent inactivation and resurgent unblocking processes of the Nav1.6 Na+ channels. These results show the apparent consequences on burst control and signal processing capacity of neurons when these currents are present.
In contrast with INaP, which drives near threshold behavior and burst generation, INaR facilitated slow channel inactivation as bursts terminate (also note [11]). Increased channel inactivation due to INaR in turn prolonged the recovery from inactivation required to initiate subsequent burst of activity. Such an interaction between open-channel unblock process underlying INaR, and, the slow inactivation underlying INaP, offer a closed-loop push-pull modulation of ISIs and IBIs. Specifically, presence of INaR facilitates slow Na+ inactivation as shown by our theoretical analyses of model behavior; such enhanced slow channel inactivation can eventually shut off channel opening and unblocking. This resulted in burst stabilization. Theoretically, this represented an enlarged separatrix (or boundary) for transitioning from a sub-threshold non-spiking behavior to spiking behavior (see enlarged green shaded region in Fig 7F), and the neuron becomes refractory to burst generation and hence offers noise tolerance.
Is this apparent effect of INaR physiologically plausible? Biophysical studies indicate that recovery from fast inactivation is facilitated in sodium channels that can pass resurgent current [30]; as shown here, this appears to be true for recovery from slow inactivation as well. Consistently, in the SCN8a knockout Med mouse, which lack the Nav 1.6 sodium channel subunit, recordings from mutant cells showed an absence of maintained firing during current injections, limited recovery of sodium channels from inactivation, and failure to accumulate in inactivated states. This is attributed to a significant deficit in INaR [11, 20, 37]. Furthermore, maintained or repeated depolarization can allow a fraction of sodium channels in many neurons to enter inactivation states from which recovery is much slower than for normal fast inactivation (reviewed in [43]). Here, our simulations and model analyses predict that the presence, and increase in INaR conductance, provides for a such a physiological mechanism to maintain sustained depolarization and promote fast and slow Na+ inactivation.
Neuronal voltage-gated Na+ currents are essential for action potential generation and propagation [5]. However, to enable fight-or-flight responses, an overt spike generation mechanism must be combined with noise modulation to extract behaviorally relevant inputs from an uncertain input space. Here we show that, the voltage-gated Na+ currents can serve an important role in neural signal processing (see summary in Fig 9). As shown in the figure, a sub-threshold activated persistent Na+ current contributes to membrane resonance, a mechanism of bandpass filtering of preferred input frequencies [9]. We call this type of input gating, which is widely known to be important for brain rhythms [9, 41], a tune-in mechanism (see figure legend). In some cases ambient noise or synaptic activity can amplify weak inputs and promote burst generation [44, 45]. This way, a tune-in mechanism such as the persistent Na+ current can contribute to weak input detection and promote burst coding [46, 47]. Then again, during rhythmic bursting, presence of resurgent Na+ maintains the order and precision of the timing events of bursts while preventing abrupt transitions into spiking phase due to stochastic influences as shown here. During ongoing sensory processing, such burst timing regulation can provide for noise cancellation or what we call a tune-out mechanism, which can mitigate random irregularities encoded in bursts (see Fig 9 and legend). Whether this leads to improved sensory processing in the presence of natural stimuli and/or sensorimotor integration during normal behaviors needs to be validated. Our biological prediction here that a sensory neuron can utilize these voltage-gated Na+ currents as a tune-in-tune-out mechanism to gate preferred inputs, attenuate random membrane fluctuations and prevent abrupt transitions into spiking activity supports such a putative role.
The conductance-based Mes V neuron model that we used to investigate the physiological role for INaR and INaP components of INa in burst discharge, incorporates a minimal set of ionic conductances essential for producing rhythmic bursting and for maintaining cellular excitability in these neurons [8]. These include: 1) a potassium leak current, Ileak, 2) sodium current, INa as described above, and, 3) a 4-AP sensitive delayed-rectifier type potassium current (IK) [8, 48]. The model equations follow a conductance-based Hodgkin-Huxley formalism [5] and are as follows.
V′=(−INa−IK−Ileak+Iapp)/C
ht′=ht∞(V)−htτt
hp′=hp∞(V)−hpτp(V)
br′=αb(1−br)br∞(V)−kbβbr(V)br
hr′=αhr(V)hr∞(V)−0.8βhr(V)hr
n′=n∞(V)−nτn
In what follows, we provide the formulation for each of the ionic currents and describe in detail, the novel INaR model.
In vitro action potential clamp studies in normal mouse Mes V neurons, and voltage-clamp studies in Nav1.6 subunit SCN8a knockout mice have demonstrated existence of three functional forms of the total sodium current, INa, including the transient (INaT), persistent (INaP) and resurgent (INaR) components [11, 14]. Each of these currents is critical for Mes V electrogenesis including burst discharge, however, their exclusive role is yet unclear. Lack of suitable experimental model or manipulation to isolate each of these TTX-sensitive components, led us to pursue an alternative approach involving computational model development of the physiological INa. To further allow model-based experimental manipulation of individual components of the INa, we designed a conductance-based model as follows. Although a single Nav1.6 channel can produce all three INa components observed experimentally, we used a set of three HH-type conductances, one for each of the transient, the persistent, and the resurgent components. This allowed us to easily manipulate these components independently to test their specific role in neural burst control. The equation for the total sodium current can be written as:
INa=INaT+INaR+INaP
where,
INaT=gNaT(mt∞(V)ht)(V−ENa)
INaR=gNaR((1−br)3hr5)(V−ENa)
INaP=gNaP(mp∞(V)hP)(V−ENa)
The maximal persistent conductance, gNaP was set 5–10% of the transient, gNaT [49] and the resurgent was set to 15–30% of gNaT, based on the relative percentage of maximum INaR and INaT as revealed by voltage-clamp experiments shown in Fig 3; ENa is the Na+ reversal potential.
Based on experimental data, the gating function/variable, mt∞(V), and ht, for INaT, and, mp∞(V), and, hP, for INaP are modeled as described in [8]. The rate equations for the inactivation gating variables ht, and, hP, model the fast and slow inactivation of the transient and persistent components respectively. The activation gates are steady-state voltage-dependent functions, consistent with fast voltage-dependent activation of INa.
Steady-state voltage-dependent activation and inactivation functions of transient sodium current respectively include:
mt∞(V)=11+e(−(V+35)4.3);ht∞(V)=11+e((V+55)7.1)
Steady-state activation, inactivation and steady-state voltage-dependent time constant of inactivation for persistent sodium current respectively include:
mp∞(V)=11+e(−(V+50)6.4);hp∞(V)=11+e((V+52)14);τp(V)=100+100001+e((V+60)10)
The novel INaR formulation encapsulates the block/unblock mechanism using a block/unblock variable (br), and, a second hypothetical variable for a competing inactivation, which we call, hr. We call this a hybrid model, to highlight the fact that the model implicitly incorporates the history or state-dependent sodium resurgence, following a transient channel opening, and combines this into a traditional Hodgkin-Huxley type conductance-based formulation. In the br′ and hr′ rate equations for br, and, hr, the block/unblock variable, br increases or grows according to the term, αb(1−br)br∞(V), and decays as per the term, kbβbr(V)br, described as follows:
αb(1−br)br∞(V): In this growth term, we incorporate state-dependent increase in br, as follows; we assume that the rate of increase in br is proportional to the probability of channels currently being in the open state, with a rate constant, αb which we call ‘rate of unblocking’; such probability is a function of the membrane voltage given by, br∞(V), defined as below:
br∞(V)=11+e((V+40)12)
The term (1−br∞(V)), models the steady-state voltage-dependency guiding the unblocking process. The channels being in open state is represented by the term, (1−br). Note that if (1−br) is close to 1, this means that larger proportion of channels are in an open state, and therefore br grows faster, promoting blocking. We modeled br∞(V) as a decreasing sigmoid function, such that, at negative membrane potentials, channels have a high probability to enter future depolarized states and therefore, (1−br)~0, in turn, br does not grow fast.
kbβbr(V)br: In this decay term, we assume that the rate of decay of br, is proportional to the probability of channels being in the blocked state, with a constant of proportionality kb, and, this probability is given by a voltage-dependent function, βbr(V), defined as below:
βbr(V)=21+e(−(V−40)8)
Note that, βbr(V) gives a high probability at depolarized potentials, indicating a blocked state and enables decrease in br in subsequent time steps.
Taken together, br, represents a phenomenological implementation of a previously described block/unblock mechanism of a cytoplasmic blocking particle [19] (see schematic of channel gating in Fig 3A). Additionally, a hypothetical competing inactivation variable, hr, sculpts the voltage-dependent rise and decay times and peak amplitude of sodium resurgence at -40 mV following a brief depolarization (i.e., transient activation), as observed in voltage-clamp experiments (see Fig 3B). The functions, αhr(V),βhr(V) and hr∞(V) are defined as voltage-dependent rate equations that guide the voltage-dependent kinetics and activation/inactivation of the INaR component as given below.
The steady-state voltage-dependency of the competing inactivation necessary to generate a resurgent Na+ current is defined as follows:
hr∞(V)=11+e((V+40)20)
The voltage-dependent rate functions of such inactivation is defined by two functions as follows:
αhr(V)=11+e(−(V+40)8);βhr(V)=0.51+e(−(V+40)15)
The steepness of the voltage-dependent sigmoid functions for activation and inactivation were tuned to obtain the experimentally observed INaR activation (see Fig 3; also see [11, 14, 30]). To obtain the kinetics (rise and decay times) of INaR comparable to those observed during voltage-clamp experiments (see S3 Fig), the model required three units for the blocking variable ((1−br)3) and five units for the inactivation variable (hr5) (see INaR equation). Together, the modeled INa reproduced the key contingencies of the Nav1.6 sodium currents (see S2 Fig) [18, 30, 50].
Sensitivity analyses was conducted for the key parameters of INaR gating including αb, and, kb. Note that these two parameters control the rate of blocking. As expected, increasing αb, that controls rate of increase in br, decreased the peak amplitude of INaR, similar to an experimental increase in block efficacy by a β-peptide (e.g., [19]). On the other hand, kb also moderates br, and increasing kb, enhances br decay rate, that significantly enhanced INaR, and, therefore burst duration (not shown). Large increases in kb significantly enhanced INaR, and indeed transformed bursting to high frequency tonic spiking. However, the effects of INaR on bursting described in the results section were robust for a wide range of values of these parameters (>100% increase from default values), and, for our simulations, the range of values, αb = 0.08 to 0.1, kb = 0.8 to 1.2, were used to reproduce Mes V neuron discharge properties. To reproduce experimentally observed spike width, we additionally tuned the inactivation time constant, τt = 1.5±0.5, for INaT.
The 4-AP sensitive delayed-rectifier type potassium current, IK, and the leak current, Ileak were modeled similar to [8] as below; also see [48].
IK=gKn(V−EK)
Ileak=gleak(V−Eleak)
where, the steady-state voltage-dependent activation function for the gating variable, n is given as:
n∞(V)=11+e(−(V−43)3.9)
EK and Eleak are K+ and leak reversal potentials respectively. Model parameter values used are provided in S1 Table.
All animal experiments were performed in accordance to the institutional guidelines and regulations using protocols approved by Animal Research Committee at UCLA. Experiments were performed in P8-P14 wild-type mice of either sex. Mice were anesthetized by inhalation of isofluorane and then decapitated. The brainstem was extracted and immersed in ice-cold cutting solution. The brain-cutting solution used during slice preparation was composed of the following (in mM): 194 Sucrose, 30 NaCl, 4.5 KCl, 1.2 NaH2PO4, 26 NaHCO3, 10 glucose, 1 MgCl2. The extracted brain block was mounted on a vibrating slicer (DSK Microslicer, Ted Pella) supported by an agar block. Coronal brainstem sections consisting of rostro-caudal extent of Mes V nucleus, spanning midbrain and pons were obtained for subsequent electrophysiological recording.
To obtain direct experimental data to drive INaR model development, we performed voltage-clamp experiments on Mes V neurons and recorded Na+ currents by blocking voltage-gated K+ and Ca2+ currents similar to [11]. The pipette internal solution contained the following composition (in mM): 130 CsF, 9 NaCl, 10 HEPES, 10 EGTA, 1 MgCl2, 3 K2-ATP, and 1 Na-GTP. The external recording solution contained the following composition (in mM): 131 NaCl, 10 HEPES, 3 KCl, 10 glucose, 2 CaCl2, 2 MgCl2, 10 tetraethylammonium (TEA)-Cl, 10 CsCl, 1 4-aminopyridine (4-AP), and 0.3 CdCl2. The voltage-clamp protocol consisted of a holding potential of -90 mV followed by a brief voltage pulse (3 ms) of +30 mV, to remove voltage-dependent block, followed by voltage steps between -70 mV to -10 mV, in steps of 10 mV for ~ 100 ms to activate INaR, and then returned to -90 mV. A 1 μM TTX abolished the Na+ current and the residual leak current was subtracted to isolate evident sodium currents. Recordings with series resistance Rseries>0.1Rm were discarded, where Rm is the input resistance of the neuron; we did not apply any series resistance compensation.
Dynamic-clamp electrophysiology and in vitro current-clamp recording were used for testing the physiological effects of Na+ currents on burst discharge as well as noise-mediated entropy changes corrected by INaR [51]. We selected neurons responding with a bursting pattern in response to supra-threshold step current injection in the Mes V nucleus in brainstem slice preparation for our study; >50% of neurons showing other patterns (e.g., tonic or single spiking cells) were discarded. Dynamic-clamp was successfully performed in bursting cells (n = 10). For dynamic-clamp recording, slices were placed in normal ACSF at room temperature (22–25°C). The ACSF recording solution during patch-clamp recording consisted of the following (in mM): 124 NaCl, 4.5 KCl, 1.2 NaH2PO4, 26 NaHCO3, 10 glucose, 2 CaCl2, 1 MgCl2. Cutting and recording solutions were bubbled with carbogen (95% O2, 5% CO2) and maintained at pH between 7.25–7.3. The pipette internal solution used in current clamp experiments was composed of the following (in mM): 135 K-gluconate, 5 KCl, 0.5 CaCl2, 5 HEPES (base), 5 EGTA, 2 Mg-ATP, and 0.3 Na-ATP with a pH between 7.28–7.3, and osmolarity between 290 ± 5 mOsm. Patch pipettes (3–5 MΩ) were pulled using a Brown/Flaming P-97 micro pipette puller (Sutter Instruments). Slices were perfused with oxygenated recording solution (~2ml/min) at room temperature while secured in a glass bottom recording chamber mounted on an inverted microscope with differential interface contrast optics (Zeiss Axiovert 10). Current clamp (and dynamic-clamp) data were acquired and analyzed using custom-made software (G-Patch, Analysis) with sampling frequency: 10 kHz; cut-off filter frequency: 2 kHz.
The Linux-based Real-Time eXperimental Interface (RTXI v1.3) was used to implement dynamic-clamp, running on a modified Linux kernel extended with the Real-Time Applications Interface, which allows high-frequency, periodic, real-time calculations [52]. The RTXI computer interfaced with the electrophysiological amplifier (Axon Instruments Axopatch 200A, in current-clamp mode) and the data acquisition PC, via a National Instruments PCIe-6251 board. Euler’s method with step size 0.05 ms was used for model integration resulting in a computation frequency of 20 kHz.
The model INaR current used for dynamic clamping into Mes V neuron in vitro was developed as discussed above. The ionic conductance gNaR was set to suitable values to introduce model INaR current into a Mes V neuron during whole-cell current-clamp recording. For dynamic-clamp experiments involving gNaR mediated noise modulation, two approaches were used to model random noise Inoise, generated in RTXI and injected as pA current:
We used stochastic current injection as an external input (additive noise) in order to produce irregularities/uncertainties in burst discharge. Our choice of the noise model was to experimentally disrupt spike timing regularity [45] and is not directly based on any known noise characteristics in Mes V neurons. The jaw muscle spindle afferent Mes V neurons are not known to have spontaneous synaptic events and we did not characterize Na+ channel fluctuations in these neurons [53]. Nonetheless, the stochastic noise we used as shown in the representative example in Fig 8 most closely matched a diffusive synaptic noise model with Gaussian distribution [54]; Fig 8C illustrates the temporal features of the noise inputs described above. A Gaussian white noise generated in MATLAB with zero mean and unit standard deviation or a Wiener noise generated in XPPAUT were used to disrupt firing patterns in the model neuron.
Model simulation and all the analyses were performed using MATLAB (Mathworks) Model code available up on request. Model bifurcation analyses were performed using XPPAUT/AUTO [55]. A variable step Runge-Kutta method ‘ode45’ was used for current-clamp simulations and ‘ode23s’ was used for voltage-clamp simulations.
Inter-event intervals (IEI) between spikes in dynamic-clamp recordings were detected using Clampfit 9.0 software and were classified post hoc as ISIs and IBIs based on a bi-modal distribution of IEIs. Typically, IEI values < 40 ms were considered as ISIs within bursts and IEI values ≥ 40 ms were considered as IBIs. Any occasional isolated spikes were eliminated from analyses for burst duration calculations.
To calculate Shannons’ entropy [56] in the inter-event intervals (IEIs), we generated histograms IEI and calculated the probabilities for each bin of the underlying IEI distributions for each 10 sec spike trains. The probability of kth IEI bin from a distribution of n equal size bins was calculated from the bin counts, N(k) as:
p(k)=N(k)∑k=1nN(k)
The entropy, H was calculated using the following formula:
H=−∑k=1npklog2pk
where, n is the total number of IEI bins, each with probability, pk.
The coefficient of variation (CV) in IEIs was calculated as follows:
CV=sx¯
where, s is the IEI sample standard deviation, and, x¯ is the sample mean.
|
10.1371/journal.ppat.0030110 | Mycobacterium tuberculosis nuoG Is a Virulence Gene That Inhibits Apoptosis of Infected Host Cells | The survival and persistence of Mycobacterium tuberculosis depends on its capacity to manipulate multiple host defense pathways, including the ability to actively inhibit the death by apoptosis of infected host cells. The genetic basis for this anti-apoptotic activity and its implication for mycobacterial virulence have not been demonstrated or elucidated. Using a novel gain-of-function genetic screen, we demonstrated that inhibition of infection-induced apoptosis of macrophages is controlled by multiple genetic loci in M. tuberculosis. Characterization of one of these loci in detail revealed that the anti-apoptosis activity was attributable to the type I NADH-dehydrogenase of M. tuberculosis, and was mainly due to the subunit of this multicomponent complex encoded by the nuoG gene. Expression of M. tuberculosis nuoG in nonpathogenic mycobacteria endowed them with the ability to inhibit apoptosis of infected human or mouse macrophages, and increased their virulence in a SCID mouse model. Conversely, deletion of nuoG in M. tuberculosis ablated its ability to inhibit macrophage apoptosis and significantly reduced its virulence in mice. These results identify a key component of the genetic basis for an important virulence trait of M. tuberculosis and support a direct causal relationship between virulence of pathogenic mycobacteria and their ability to inhibit macrophage apoptosis.
| The infection-induced suicide of host cells following invasion by intracellular pathogens is an ancient defense mechanism observed in multicellular organisms of both the animal and plant kingdoms. It is therefore not surprising that persistent pathogens of viral, bacterial, and protozoal origin have evolved to inhibit the induction of host cell death. M. tuberculosis, the etiological agent of tuberculosis, has latently infected about one third of the world's population and can persist for decades in the lungs of infected, asymptomatic individuals. In the present study we have identified nuoG of M. tuberculosis, which encodes a subunit of the type I NADH dehydrogenase complex, as a critical bacterial gene for inhibition of host cell death. A mutant of M. tuberculosis in which nuoG was deleted triggered a marked increase in apoptosis by infected macrophages, and subsequent analysis of this mutant in the mouse tuberculosis model provided direct evidence for a causal link between the capacity to inhibit apoptosis and bacterial virulence. The discovery of anti-apoptosis genes in M. tuberculosis could provide a powerful approach to the generation of better attenuated vaccine strains, and may also identify a new group of drug targets for improved chemotherapy.
| Tuberculosis is an infectious disease of enormous and increasing global importance. Currently, about one third of all humans are latently infected with its etiologic agent, Mycobacterium tuberculosis (Mtb), and an estimated 2.5 million people die of tuberculosis annually [1]. After infection of a mammalian host, Mtb is able to resist innate host defenses sufficiently to increase the local bacterial burden and disseminate throughout the body. With the onset of the adaptive immune response, however, the bacterial numbers are controlled in over 90% of infected individuals. Nevertheless, the host is not able to completely clear the bacterial burden, thus leading to persistence of Mtb within the lungs and other tissues of healthy individuals. These latent infections can be reactivated to generate full-blown disease, a process that is accelerated by immunocompromised states resulting from senescence, malnutrition, and co-infection with HIV, which is a major source of morbidity and mortality associated with the current HIV epidemics in many countries [2–5].
Programmed cell death (apoptosis) plays an important role in the innate immune response against pathogens and comprises an evolutionarily conserved defense strategy that extends even into the plant world [6,7]. It is therefore essential for persisting intracellular pathogens to have strong anti-apoptosis mechanisms [8–12]. While a few studies have suggested that under some conditions Mtb may induce host cell apoptosis [13–16], a substantial body of evidence points strongly to the expression of strong anti-apoptotic mechanisms by Mtb and other closely related virulent bacteria. Furthermore, this capacity is not found in avirulent species, suggesting a causal link between virulence and inhibition of macrophage apoptosis [17–19]. This hypothesis is supported by the recent discovery that the genetic predisposition of different inbred mouse strains to mycobacterial infections is linked to the capacity of their macrophages to undergo apoptosis or necrosis upon infection, with the former response imparting a resistant and the latter a susceptible host phenotype [20].
Further confirmation of the findings that Mtb inhibits host cell apoptosis is provided by a number of studies that have addressed its molecular mechanism. The importance of Mtb-induced upregulation of anti-apoptosis genes in infected macrophages for apoptosis inhibition was supported by functional data using either anti-sense oligonucleotides to knock down mcl-1 expression [19] or A1 knock-out mice lacking the anti-apoptosis gene A1 [21,22]. These results implicate the intrinsic (mitochondria-mediated) apoptosis pathway as a target for Mtb-mediated apoptosis inhibition, because mcl-1 and A1 are both members of the large family of Bcl-2–like proteins that localize prominently to mitochondria. However, this is contradicted by the finding that overexpression of Bcl-2 (another mitochondrial anti-apoptotic protein) could not rescue cells from undergoing apoptosis after infection with nonvirulent mycobacteria, thus suggesting that the extrinsic pathway (death receptor–mediated) is involved in the infection-induced apoptosis [23]. Consistently, virulent Mtb strains could inhibit FasL-induced apoptosis in Fas-expressing cells [18]. The same group reported very recently that lipoglycans of the Mtb cell wall stimulate the activation of NF-kB via TLR-2 and that the subsequent upregulation of cellular FLIP leads to inhibition of FasL-mediated apoptosis [24]. Furthermore, it was suggested that Mtb stimulated the secretion of soluble TNF-R2, which led to the reduction of bioactive TNF-α in the medium and therefore less stimulation of the TNF-R1 [25]. Altogether, it seems that virulent Mtb is able to inhibit induction of host cell apoptosis via multiple pathways, and probably encodes mechanisms to interfere with both intrinsic and extrinsic pathways for initiation of programmed cell death.
The inhibition of macrophage apoptosis by Mtb is believed to provide a number of advantages to the bacterium in its struggle to resist the host immune response. These include preservation of a favorable host cell environment during growth and persistence [26,27], evasion of apoptosis-linked bactericidal effects [28,29], and avoidance of efficient cytotoxic T cell priming via the detour pathway of antigen cross-presentation [15,30–32]. This last point is of potential importance to the improvement of tuberculosis vaccines, because attenuated mycobacterial strains that induce higher levels of host cell apoptosis would be expected to stimulate more robust cellular immunity, as suggested by a recent study using recombinant M. bovis Bacille Calmette-Guérin (BCG) expressing listeriolysin [33,34]. Therefore, the identification of mycobacterial genes required for prevention of apoptosis could lead to specific strategies for designing more efficacious forms of BCG or other attenuated mycobacterial vaccine strains.
In order to clarify the role of mycobacteria in host cell apoptosis and to address its importance for bacterial virulence, we sought to identify anti-apoptosis genes via a gain-of-function genetic screen. Using this approach, we successfully identified two independent genomic regions of virulent Mtb (strain H37Rv) that mediate the inhibition of host cell apoptosis by the facultative pathogen M. kansasii. The analysis of a defined set of bacterial mutants of M. kansasii in immunocompromised (SCID) mice demonstrated a causal relationship between inhibition of apoptosis and virulence. These findings were confirmed via a loss-of-function strategy using the newly identified anti-apoptosis gene, nuoG, and demonstrated attenuation of Mtb nuoG mutants in immunocompromised and immunocompetent mice. Altogether our findings allowed, to our knowledge for the first time, the demonstration of a causal relationship between inhibition of host cell apoptosis and virulence of mycobacteria.
To identify genes in Mtb responsible for anti-apoptotic effects, we established a gain-of-function genetic screen using the nonpathogenic M. smegmatis mc2155 strain, which is a fast-growing mycobacterium that is extremely efficient for transformation [35]. The human cell line THP-1 was chosen for use as the host cells for this screen because published work indicates that these cells provide an accurate model for the apoptotic response of Mtb-infected primary human alveolar macrophages [23]. Our initial studies established that M. smegmatis infection of THP-1 cells induced strong apoptosis after 1.5 d of infection when compared to BCG-infected macrophages, as assessed by disruption of the cell monolayer and by staining for DNA strand breakage using the TUNEL assay (Figure 1A). The failure of BCG to induce apoptosis seems to contradict the report by Kean et al. [17] in which seven different species of mycobacteria were compared for their capacity to induce apoptosis in primary human alveolar macrophages, and BCG was among the apoptosis-inducing mycobacterial species. Nevertheless, in that study, no fast-growing species were included, and all apoptosis assays were performed after 5–7 d of infection. In an independent study, we demonstrated that M. smegmatis and other nonpathogenic mycobacteria like M. fortuitum have a very strong capacity to induce apoptosis with rapid kinetics, even when compared to facultative pathogenic mycobacteria like M. kansasii and BCG (A. Bohsali and V. Briken, unpublished data). Therefore, the induction of apoptosis by M. smegmatis had to be analyzed very early after infection (16–36 h; Figure 1A and 1B), and at this time point, little or no induction of macrophage apoptosis by BCG could be observed.
The gain-of-function screen was performed using a library of 312 M. smegmatis clones containing Mtb genomic DNA fragments on an episomal cosmid. Two clones (designated J21 and M24) containing cosmids with separate nonoverlapping Mtb genomic DNA inserts gave significantly reduced levels of apoptosis and were selected for detailed study. To confirm that the observed effects on apoptosis were due to the cosmids contained in these clones, the episomal cosmid DNA was extracted and re-transfected into M. smegmatis. In both cases, the re-transformed clones had the same phenotype as the original clones, showing an approximately 50% reduction of apoptosis of infected THP-1 cells (Figure 1B). It is important to note that our screening was designed to emphasize specificity rather than sensitivity, as it focused on the clones showing the strongest and most reproducible suppression of apoptosis. Thus, it is very likely that other regions with anti-apoptosis capacity may remain to be identified.
Although the effects of cosmids M24 and J21 on reducing the apoptosis induced by M. smegmatis were highly reproducible, the magnitude of this effect was relatively modest. Most likely, this reflected the very strong capacity of M. smegmatis to induce apoptosis when compared to other mycobacteria (Figure 1 and unpublished data), and we assessed this by testing the effects of the Mtb cosmids on apoptosis induction by another mycobacterial species, M. kansasii. This opportunistic pathogen is known to be a strong inducer of apoptosis, but it shows slower host cell killing than M. smegmatis, with significant levels of apoptosis being observed only after 5–7 d of infection [17]. We transformed the cosmids into M. kansasii to generate clones Mkan-J21 and Mkan-M24, and also generated an M. kansasii control strain using the empty cosmid vector pYUB415 (Mkan-CO). Based on FACS analysis of TUNEL staining, wild-type M. kansasii and Mkan-CO induced comparable high levels of apoptosis at 5 d post infection (95% and 86% apoptosis, respectively), whereas Mkan-J21 and Mkan-M24 showed markedly reduced levels (16% and 19% apoptosis, respectively) (Figure 1C). The transfection of the two cosmid clones did not affect the in vitro growth of either M. smegmatis or M. kansasii (unpublished data).
Apoptosis of infected macrophages has been reported to directly kill ingested bacteria [29], and killing of bacteria within apoptotic bodies is also facilitated as a result of enhanced phagocytosis by activated bystander macrophages [28]. Therefore, we hypothesized that inhibition of apoptosis is important for mycobacterial evasion of the host's innate immune response. This was tested by infecting groups of SCID mice (BALB/c background) with the M. kansasii cosmid transformants or with Mtb H37Rv. As expected, H37Rv was highly virulent in these mice that lack adaptive immunity (median survival 15 d), whereas Mkan-CO showed only modest virulence even in these immunodeficient animals (median survival >200 d) (Figure 2A). Remarkably, Mkan-J21 and Mkan-M24 showed significantly increased virulence in SCID mice, with median survival times of 44 d or 60 d, respectively (p = 0.0002 compared to Mkan-CO for both survival times by log-rank test). Tissue bacterial burdens in SCID mice infected with the various mycobacteria were consistent with the survival data, based on colony-forming units (CFUs) in the lung, liver, and spleen at various time points after infection (Figure 2B). Histopathology of the lungs revealed that after 35 d, airways in the lungs of Mkan-J21–infected mice were almost completely consolidated, and in Mkan-M24–infected mice the lungs showed significant infiltration of inflammatory cells. In contrast, lungs of mice infected with Mkan-CO had normal morphology (Figure 2C). In order to correlate the virulence of the different strains with their capacity to inhibit apoptosis, lung sections obtained 14 d after infection were stained for apoptotic cells using a TUNEL-based assay and analyzed by microscopy as described in the Materials and Methods section. This revealed minimal levels of apoptosis in situ for lung tissue infected with H37Rv (2%), Mkan-J21 (5%), and Mkan-M24 (7%) compared to significantly higher levels in Mkan-CO–infected (27%) lungs (Figures 2D and S1). Overall, these results confirmed the anti-apoptotic activity of genes contained in cosmids J21 and M24, and strongly supported the importance of apoptosis inhibition for the virulence of mycobacteria.
Cosmids J21 and M24 both contained about 30 mycobacterial genes, and therefore it could not be completely excluded that the effect on bacterial virulence was in part or totally caused by another gene or genes linked to those responsible for the anti-apoptosis effects. As Mkan-J21 showed the strongest enhancement of virulence in SCID mice, we selected cosmid J21 for additional studies to determine the precise gene or genes required for inhibition of apoptosis by M. tuberculosis. Sequencing of J21 cosmid DNA (unpublished data) showed that its insert corresponded to bp 3511794 through bp 3545572 of the Mtb genome, according to the standard annotation for strain H37Rv [36]. This interval contains the intact open reading frames of 31 annotated genes, including a large operon that encodes the 14 subunits of the Mtb type I NADH dehydrogenase complex (NDH-1), along with other genes that encode a variety of different known or predicted functions (Figure S2).
To identify which gene or genes in this region were important for the anti-apoptosis activity, a series of deletion mutants spanning different regions within the genomic interval corresponding to J21 was created in Mtb H37Rv using specialized transduction [37] (Figures S2 and S3). These mutants were analyzed by using SCID mice to test for reduction in bacterial virulence (Figure S4A), and by using THP-1 cells to test for loss of apoptosis inhibition (Figure S4B). Although several of the deletions showed modest effects, only one mutant, Mtb ΔRv3151, which corresponded to the deletion of the nuoG subunit of NDH-1, gave both statistically significant extension of survival in SCID mice and enhanced apoptosis in THP-1 cells relative to parental Mtb. The gain-of-function experiments described in Figures 1 and 2 using the episomal cosmids relied on the proper function of the endogenous gene promoters in transfected M. smegmatis or M. kansasii. To confirm that nuoG was actually transcribed on the J21 cosmid in M. smegmatis, reverse transcription–PCR with primers specific for Mtb nuoG was performed and clearly demonstrated that the nuoG gene was transcribed (Figure S4C). These experiments implicated nuoG, and potentially the complete functional NDH-1 complex, in mediating most or all of the anti-apoptotic properties of cosmid clone J21.
To confirm that nuoG deletion was responsible for the pro-apoptotic phenotype and to exclude significant polar effects of the deletion on other nearby genes, the mutant was complemented with a plasmid carrying a copy of nuoG behind a constitutively active promoter that integrates into the Mtb chromosome at the unique attB site [38]. This gave full complementation for the in vitro apoptosis assays (Figure 3), although there was a residual increase in apoptosis induction observed in vivo (Figure 4B and 4F) that may have been due to some minor polar effects on the transcription of other members of the nuo-operon. The mutant (MtbΔ), complemented mutant (MtbΔC), and wild-type Mtb were analyzed for the capacity to grow in vitro, which demonstrated that the nuoG deletion had no effect on aerobic growth rate (Figure 3A), confirming a previous report that NDH-1 is not essential for mycobacterial growth in culture [39]. This is most likely due to the fact that Mtb has two additional NADH dehydrogenases, which are the non-proton pumping type II NADH dehydrogenases encoded by the ndh and ndhA genes [40].
To confirm the pro-apoptotic effect of nuoG deletion, we compared levels of apoptotic cell death in cultures of differentiated THP-1 cells and extended the analysis to include detection of apoptosis in bone marrow–derived macrophages (BMDMs) from BALB/c mice following infection. The Mtb nuoG mutant induced significantly more apoptosis than the complemented strain or the wild-type Mtb in human and mouse macrophages (Figure 3B and 3C). This phenotype of the nuoG mutant was also confirmed in BMDMs from C57BL/6 mice (unpublished data). The difference in apoptosis induction was not due to a reduced phagocytosis of the nuoG mutant, which was demonstrated by comparing rates of infection via acid-fast staining (Figure 3D) and CFU determination (unpublished data). These results show that nuoG was necessary for Mtb to inhibit apoptosis of primary murine macrophages or the human macrophage-like THP-1 cells. The nuoG Mtb mutant still induced less apoptosis (Figure 3B) when compared to M. kansasii–infected cells (Figure 1C), which probably reflects the fact that virulent Mtb expresses multiple anti-apoptosis genes. Therefore, the deletion of only one gene does not completely abolish the capacity of the bacterium to inhibit apoptosis. Since nuoG is part of a multi-subunit NDH-1 complex that is present in the cosmid J21, it was of interest to determine if nuoG alone could confer the gain-of-function in M. kansasii or if, in contrast, expression of the whole nuo-operon was needed. Therefore, Mtb-nuoG was constitutively expressed in M. kansasii and the capacity to induce host cell apoptosis was analyzed (Figure 3E). This demonstrated that expression of Mtb-nuoG alone very significantly reduced apoptosis induction by M. kansasii when compared to that of wild-type and empty vector-transfected bacteria (Figure 3E).
Most of the sequenced genomes of mycobacteria contain a nuoG gene within a nuo-operon containing 14 genes that code for the NDH-1. The one exception to date is M. leprae, in which the whole operon is deleted except for a nuoN pseudogene [41]. nuoG of M. kansasii was cloned using PCR and sequenced in order to allow protein sequence comparison of all the nuoGs in the mycobacterial strains used in our study. Comparison of the nuoG protein sequences among virulent mycobacteria revealed a high degree of homology (99% identity, Figure S9). Interestingly, nuoG of BCG is also highly homologous (99%) to nuoG of virulent mycobacterial species, which is consistent with our unpublished data demonstrating that the deletion of nuoG in BCG also increases the potential of the bacteria to induce apoptosis. These findings suggest that the vaccine strain BCG retained this virulence mechanism from its parental M. bovis strain, although overall it may still induce more apoptosis than fully virulent mycobacteria. In contrast, nuoG of M. smegmatis is only 70% identical (Figure S9) and nuoG of M. kansasii is only 34% identical to nuoG sequences of virulent mycobacterial species. The latter nuoG protein is truncated due to a stop codon introduced at codon 295, and thus nuoG of M. kansasii is missing about 512 of the C-terminal amino acids. Interestingly, the M. kansasii nuoG is quite homologous up to amino acid 288 (94%). In conclusion, it seems likely that both nuoG of M. smegmatis and M. kansasii have lost their apoptosis-inhibiting function within the NDH-1 complex, and therefore the overexpression of Mtb nuoG is able to restore the capacity of the bacteria to inhibit host cell apoptosis. It will be of great interest to explore this hypothesis further by examining the expression of the various nuoG proteins in the nuoG deletion mutant of Mtb and analyzing the apoptosis induction of these complemented bacteria.
How does nuoG mediate apoptosis inhibition? For nuoG to have a direct effect on host cell apoptosis pathways, one would assume that it needs to be secreted in order to interact with host cell proteins or lipids. However, if the structure of the NDH-1 complex of Mtb is similar to that of other bacterial NDH-1 complexes, nuoG will be located in the cytosol of the bacterium [42]. To determine whether Mtb nuoG is secreted experimentally, we created a phoA-nuoG fusion protein. phoA can convert a colorless substrate into a blue product, but only if it is secreted by the bacterium [43]. This assay failed to detect secretion of nuoG (Figure S7), which is consistent with the absence of a signal peptide and the predicted cytosolic localization of this component of NDH-1. Altogether, it thus seems unlikely that nuoG is secreted, and a direct effect of nuoG onto the host cell can also be judged to be unlikely.
The disruption of the NDH-1 system in the nuoG mutant might have a very profound impact on the metabolism and proteome of the mycobacterium, which might result in an indirect effect on host cell apoptosis induction. Nevertheless, the absence of an in vitro growth defect in the nuoG mutant (Figure 3A) would argue against a profound effect of the deletion on bacterial metabolism. In order to address the effect of the nuoG mutation on the proteome of the mycobacteria, the proteins of wild-type and mutant Mtb were separated via 2-D gel electrophoresis. This revealed that the nuoG mutation did not induce a major change in the proteome (Figure S8). Thus, we found that the deletion of nuoG did not have a major impact on the general bacterial metabolism or its proteome. Instead, we propose that nuoG exerts its anti-apoptotic and virulence-promoting function via the enzymatic activity of the NDH-1 complex in a more specific way.
Regardless of the remaining questions about the potential mechanism of the nuoG/NDH-1–mediated apoptosis inhibition, our identification of an apoptosis-inducing mutant of Mtb allowed us to analyze the importance of apoptosis inhibition for bacterial virulence. First, the importance of host cell apoptosis inhibition in innate immune defense was analyzed by infecting immunodeficient SCID mice. The median survival times for mice infected with Mtb wild-type or the complemented strain were not significantly different (14 d and 16.5 d, respectively). In contrast, mice infected with ΔnuoG Mtb survived twice as long as those infected with wild-type bacteria (median survival of 27.5 d, p < 0.0001, log-rank test; Figure 4A), even though all mice received similar initial bacterial doses as confirmed by CFU counts at day 1 after infection (Figure S10). Consistently, the amount of apoptosis induced in lung sections of these mice was significantly increased from less than 1% with wild-type bacteria to about 13% in mutant bacteria (p < 0.05) (Figures 4B and S5). The complemented bacteria still showed increased apoptosis (6%), but this was significantly reduced compared to the nuoG mutant (p < 0.05). Therefore, the deletion of nuoG significantly increased the induction of apoptosis in the lungs of SCID mice. These results corroborated the findings presented in Figure 2 by using a loss-of-function approach, and together both sets of experiments point towards an important role of infection-induced apoptosis in the innate immune response. This statement is supported by the recent results linking the capacity of host cell macrophages to undergo apoptosis upon mycobacterial infection to the susceptibility of different mouse strains [20].
The importance of nuoG-mediated apoptosis inhibition for bacterial virulence in immunocompetent mice was analyzed using BALB/c mice (Figure 4C–4F). Again, the mutant significantly (p < 0.004) delayed the death of infected mice when compared to wild-type (median survival 175 d) and complemented bacteria (median survival 193 d, p = 0.16 compared to wild-type) (Figure 4C). Measurement of bacterial CFUs showed that the growth of the mutant in the lungs of BALB/c mice was significantly reduced by approximately 0.8 log at week 20, after similar initial growth during the first 3 wk of infection (Figure 4D). In contrast, the bacterial loads in spleen and liver were not significantly different at 3, 10, or 20 wk post infection (unpublished data). Comparison of the histopathology of lung sections of infected BALB/c mice demonstrated an obvious reduction in granulomatous inflammation at week 20 in animals infected with the ΔnuoG mutant (Figure 4E). Although lung histopathology appeared similar in wild-type and ΔnuoG mutant–infected BALB/c mice at 3 wk (Figure 4E), staining of lung sections at this time point for TUNEL reactivity revealed a significant increase of greater than 10-fold in apoptotic cells in mice infected with the ΔnuoG mutant compared to mice that were uninfected, or infected with wild-type or complemented bacteria (Figures 4F and S6).
Taken together, our results demonstrate that nuoG is an anti-apoptosis gene of Mtb that is important for bacterial virulence in both immunocompromised and immunocompetent mice and thus strongly support the general hypothesis that the inhibition of host cell apoptosis is important for virulence of mycobacteria. The challenge ahead is to determine the molecular mechanism by which a bacterial NADH dehydrogenase can manipulate host cell apoptosis induction. It is intriguing to speculate that perhaps the NDH-1 complex of virulent bacteria has taken on a separate function in modifying apoptotic responses of infected macrophages from its original purpose of energy generation, which is now mainly performed by the type II dehydrogenases ndh and ndhA, which are both essential genes. One of the unique features of NDH-1, as opposed to the NDH and NDHA dehydrogenases of M. tuberculosis, is its capacity to pump protons across the bacterial membrane. We therefore hypothesize that these protons, in conjunction with the secreted bacterial superoxide dismutases (SodA and SodC), could serve to neutralize the superoxide anions generated within the phagosome by an activated NOX2 complex to generate hydrogen peroxide, which is then further catabolized to water and oxygen by bacterial catalase (KatG). Superoxide anions are a known trigger for apoptosis in a variety of biological systems, so the involvement of NDH-1 and nuoG in their elimination may interrupt a critical signal that initiates the host cell apoptosis response. Our hypothesis would predict that other NDH-1 subunits involved in proton translocation, such as nuoL and nuoM [42], will have the same apoptosis phenotype as the nuoG mutant, a prediction that has not yet been tested experimentally. This proposed general mechanism for inhibition of apoptosis is further supported by other studies implicating SodA as an anti-apoptotic factor in Mtb ([44,45]). The discovery of nuoG and NDH-1 as anti-apoptosis factors encoded by specific Mtb genes suggests new strategies for improving currently used and novel tuberculosis vaccines, and could also provide targets for development of antimicrobial drugs for treatment of persistent disease.
M. smegmatis (mc2155) has been previously described [35], and M. kansasii strain Hauduroy ( ATCC 12478) and Mtb strain H37Rv (ATCC 25618) were obtained from the American Type Culture Collection (http://www.atcc.org/). M. bovis BCG Pasteur strain was obtained from the Trudeau Culture Collection (Saranac Lake, New York, United States). GFP-expressing BCG and M. smegmatis were generated by subcloning the enhanced GFP gene (Clontech, http://www.clontech.com/) into the mycobacterial episomal expression vector pMV261. The resulting plasmid (pYU921) was transfected into competent cells by electroporation as previously described [35]. M. smegmatis was cultured in LB broth with 0.5% glycerol, 0.5% dextrose, and 0.05% TWEEN-80. Mtb H37Rv and M. kansasii were grown in 7H9 broth with 0.5% glycerol, 0.5% dextrose, 0.05% TWEEN-80, and 10% OADC enrichment (DIFCO, http://www.bd.com/ds/). For selective media, 50 μg/ml hygromycin or 40 μg/ml kanamycin were added.
Human myelomonocytic cell line THP-1 (ATCC TIB-202) was cultured and differentiated using phorbol myristate acetate (PMA) (Sigma, http://www.sigmaaldrich.com/) as described [46]. Bone marrow macrophages were derived from the femur and tibia of BALB/c mice as described [46]. Bacteria were grown to an OD600 ranging from 0.5 to 0.8, sonicated twice for 20 s using a cup horn sonicator, and allowed to settle for 10 min. The infection was carried out at a multiplicity of infection (MOI) of 10:1 (10 bacilli to 1 cell) for 4 h in triplicate wells, after which extracellular bacteria were removed by four washes with phosphate buffered saline (PBS). The cells were incubated in DMEM (Invitrogen, http://www.invitrogen.com/) with 20% human serum (Sigma) and 100 μg/ml gentamicin (Invitrogen), and an apoptosis assay was performed after the indicated periods of culture.
The TUNEL assay was performed to reveal apoptosis-induced DNA fragmentation in either tissue culture cells or lung sections of infected mice using the In Situ Cell Death Detection Kit, Fluorescein (for cultured cells) or In Situ Cell Death Detection Kit, POD (for lung sections) (Roche Applied Science, http://www.roche-applied-science.com/). The assay was carried out as described by the manufacturer and the percentage of stained cells was analyzed using flow cytometry for cultured cells or quantification via light microscopy for the animal tissue sections.
The strategy for generation of the Mtb genomic library in cosmid vector pYUB415 has been previously described [47]. Briefly, Mtb (strain Erdman) genomic DNA was purified and partially digested with Sau3A. DNA fragments of about 40 kbp were selected by agarose gel purification and ligated into arms of cosmid vector pYUB415 digested with BamH1 as previously described [47]. DNA was packaged in vitro with Gigapack XL (Stratagene, http://www.stratagene.com/) and Escherichia coli were transduced and selected on LB plates containing 100 μg/ml ampicillin. Over 105 independent clones were pooled, and DNA for transformation was obtained using standard alkaline lysis method.
Transformations were performed by electroporation of competent mycobacteria as described [35]. For the initial screen, M. smegmatis was transformed with the genomic DNA cosmid library described above, and 312 cosmid clones were picked and grown in liquid medium containing 50 μg/ml hygromycin. Assuming random distribution of Mtb sequences among the cosmid transformants and an average insert size of about 40 kbp, the 312 cosmid clones represented 3-fold coverage of the entire Mtb genome. After three successive screens using bright field microscopy or flow cytometry to assess levels of cell death, 12 clones were selected for quantitative assessment using TUNEL staining followed by flow cytometry. This identified three clones of greatest interest (M24, J21, and I16), and their cosmid DNA was purified and screened by restriction digest (not shown). This revealed that the inserts of M24 and I16 were identical, but different from the insert of J21. For cosmid J21, the 5′ and 3′ ends of the insert DNA were sequenced and aligned with the published genomic Mtb DNA sequence (Figure S2), and subsequently the whole insert was sequenced to confirm that is corresponded to the sequence published data.
Specific genes of Mtb were disrupted using specialized transduction as described [37]. To create the nuoG::hyg-null allele, the hygromycin resistance cassette was introduced between the first 4 bp of the nuoG 5′ end and the last 163 bp of the 3′ end of the open reading frame. The successful deletion of the gene was demonstrated by Southern blotting as described previously (Figure S3). For complementation of the ΔnuoG mutation, nuoG was amplified by PCR and cloned behind the hsp60 constitutive promoter into the plasmid pMV361, which allows integration of a single copy into the genome of Mtb [38].
BALB/c or SCID/Ncr (BALB/c background) mice (4- to 6-wk-old females) were infected intravenously through the lateral tail vein with 1 × 106 bacteria. For survival studies, groups of ten mice were infected, and after 24 h, three mice per group were sacrificed to determine the bacterial load in the organs. In order to follow the bacterial growth, an additional three mice per time point were infected. The organs (lung, spleen, liver) were homogenized separately in PBS/0.05% TWEEN-80, and colonies were enumerated on 7H10 plates grown at 37 °C for 3–4 wk. For histopathology, tissues were fixed in 10% buffered formalin and embedded in paraffin; 4-μm sections were stained with hematoxylin and eosin. TUNEL staining was performed on the paraffin-embedded tissue sections using the In Situ Cell Death Detection Kit, POD (Roche Applied Science) per the manufacturer's protocol. Quantification was performed on coded specimens by a blinded observer by counting the number of apoptotic nuclei per ∼200 total nuclei in eight separate areas of two lung sections for each of the three mice per group. All animals were maintained in accordance with protocols approved by the Albert Einstein College of Medicine and University of Maryland Institutional Animal Care and Use Committees.
The GenBank (http://www.ncbi.nlm.nih.gov/Genbank/index.html) accession numbers for the sequences of the nuoG proteins from the following mycobacteria are M.bovis AF2122/97 (CAD95267), M. bovis BCG-Pasteur (YP979258), M. kansasii Hauduroy (EF607211), M. smegmatis mc2155 (YP886418), M. tuberculosis CDC1551 (AAK47578), and M. tuberculosis H37Rv (CAB06288). |
10.1371/journal.pcbi.1001046 | Decrypting the Sequence of Structural Events during the Gating Transition of Pentameric Ligand-Gated Ion Channels Based on an Interpolated Elastic Network Model | Despite many experimental and computational studies of the gating transition of pentameric ligand-gated ion channels (pLGICs), the structural basis of how ligand binding couples to channel gating remains unknown. By using a newly developed interpolated elastic network model (iENM), we have attempted to compute a likely transition pathway from the closed- to the open-channel conformation of pLGICs as captured by the crystal structures of two prokaryotic pLGICs. The iENM pathway predicts a sequence of structural events that begins at the ligand-binding loops and is followed by the displacements of two key loops (loop 2 and loop 7) at the interface between the extracellular and transmembrane domain, the tilting/bending of the pore-lining M2 helix, and subsequent movements of M4, M3 and M1 helices in the transmembrane domain. The predicted order of structural events is in broad agreement with the Φ-value analysis of α subunit of nicotinic acetylcholine receptor mutants, which supports a conserved core mechanism for ligand-gated channel opening in pLGICs. Further perturbation analysis has supported the critical role of certain intra-subunit and inter-subunit interactions in dictating the above sequence of events.
| Pentameric ligand-gated ion channels are a family of membrane proteins that open/close an ion-conducting channel in response to the binding of specific ligands. Some members of the family, including nicotinic acetylcholine receptors, play key physiological roles in signal transduction at synapses. Despite many experimental and computational studies of the gating transition of these pentameric ion channels, the structural basis of how ligand binding couples to channel opening remains uncertain. In particular, the all-atom computer simulation of the gating transition is limited to nanosecond ∼ microsecond time scales while the entire transition takes tens of microseconds. In this study, we have employed a highly efficient coarse-grained modeling method to dissect the sequence of structural events underlying the gating transition. The model predictions are in broad agreement with the kinetic analysis of mutants of nicotinic acetylcholine receptors. This study has established a useful computational framework to simulate the functional dynamics of pentameric ligand-gated ion channels.
| Pentameric ligand-gated ion channels (pLGICs) are a family of membrane proteins that open/close an ion-conducting channel in response to an increase/decrease in the binding affinity for specific ligands [1], [2], [3], [4]. Some members of the family, including nicotinic acetylcholine receptors (AChRs, [5]), play key physiological roles in signal transduction at synapses.
The pLGICs share the common structural architecture of a pentamer with each subunit consisting of an extracellular ligand-binding domain (ECD) and a transmembrane channel domain (TMD). The ligand-binding sites lie at the interfaces between adjacent ECDs and the TMD of each subunit is comprised of four transmembrane helices, M1–M4. Recent structural investigations have yielded several atomic or near-atomic structural models of pLGICs, including a 4 Å-resolution refined model of the Torpedo AChR obtained by electron microscopy [6], [7], the crystal structures of acetylcholine-binding proteins (AChBP) [8], [9], [10], the ECD of mouse AChR α subunit [11], and bacterial pLGICs from Erwinia chrysanthemi (ELIC) and the cyanobacterium Gloebacter violaceus (GLIC) [12], [13], [14], [15]. The crystal structures of ELIC [12] and GLIC [13], [14] may represent the low-affinity, closed-channel and high-affinity, open-channel conformations of the pLGICs, respectively. Despite their moderate sequence similarity (<20% sequence identity), the two proteins are highly similar in both secondary and tertiary structures [13], [14]. A comparison of the ELIC and GLIC structures offers the possibility of a detailed view to the global and local structural changes associated with the gating transition of pLGICs despite their variation in bound ligand (ELIC is gated by an unknown ligand, and GLIC is gated by proton instead of a neurotransmitter).
Various mechanistic models for the gating transition of pLGICs have emerged from a wealth of experimental data and structure-based simulations. It has been suggested that agonist binding initiates various conformational changes, including the movements of binding site loops A and B [16], loop C [17], [18], [19], [20] and loop F [20], a quaternary twist motion [21] and a tertiary deformation within the ECD [7]. These structural changes are thought to propagate to the TMD and cause either rotation [21], [22] or tilting [13], [14], [23], [24] of the pore-lining M2 helix, which leads to the opening of the physical gate formed by the bulky side chains of hydrophobic residues in the equatorial region of the M2 helix [12]. Although multiple interface loops, secondary structure elements, and key residues have been implicated in the signal transmission from ECD to TMD (see [19], [25], [26], [27], [28]), the full details of the signaling pathway are not known with certainty. To explore the molecular mechanism of signaling in AChRs, single-channel kinetic and rate-equilibrium free energy relationships (Φ-value analysis) of mutant AChRs have been analyzed [3], which has led to the proposal that the gating occurs as a conformational cascade that propagates from the ligand-binding site to the channel pore via sequential, coupled movements of rigid-body blocks with distinct Φ-values [29]. The nature of these structural motions is thought to be stochastic Brownian motions [30] although the details remain to be worked out. It is likely that one or more of the intermediate states of this conformational pathway has been detected in high-resolution patch-clamp experiments [31].
The gating transition of pLGICs has been studied extensively by a variety of computational methods, including equilibrium molecular dynamics (MD) simulation [17], [18], [32], [33], [34], [35], [36], [37], targeted MD simulation [27], Brownian dynamics simulation [38], and normal mode analysis (NMA) [21], [39], [40], [41]. Nevertheless, atomistic MD simulations of protein dynamics are limited to a time range of nanoseconds ∼ microseconds [42] despite fast advancing computing technology. Although MD simulations ranging from tens of nanoseconds (see [27], [32]) to one microsecond [37] have revealed interesting conformational changes that may lead to channel opening/closing, the simulation times remain far less than the 10∼20 µs time range necessary for the activation of neuromuscular AChRs [43].
To overcome the time-scale barrier for MD simulations, a variety of coarse-grained models [44] have been developed to simulate protein conformational dynamics with greater efficiency. Of particular interest to the present study is the elastic network model (ENM) [45], [46], [47], which represents a protein structure as a network of Cα atoms with neighboring ones connected by springs with a uniform force constant [48]. The normal mode analysis (NMA) of ENM often yields a handful of low-frequency modes that dominate the large-scale conformational changes observed between two protein crystal structures [47], [49]. Numerous studies have established ENM as an efficient and robust means to tease out the functionally relevant conformational dynamics from protein structures with no limit in time scale or system size (for reviews, see [50], [51], [52]). Indeed, ENM has formed the basis of several computational methods for modeling protein conformational transitions [53], [54], [55], [56]. In an earlier study [55], one of us developed the mixed-ENM technique to generate a transition pathway between two given conformations using a double-well mixed-ENM potential, which is built from two ENM potentials constructed based on the two given conformations. A similar approach (plastic network model) was proposed by Maragakis and Karplus [54]. In another related study by Delarue and coworkers, a transition pathway was generated by minimizing an ENM-based action function [56]. Recently, Zhu and Hummer applied the mixed-ENM method to the gating transition of the TMD of pLGICs [57]. They found that the conformational transition involves a concerted tilting of helices M2 and M3, and M2 changes its bending state, which results in an early closure of the channel pore during the open-to-close transition [57]. Despite the above insights, the conformational transition of the full pLGIC (including both ECD and TMD) remains to be simulated to determine the sequence of structural events that couple ligand binding to channel gating.
Recently, one of us has developed an interpolated-ENM (iENM) method based on the mixed-ENM method to predict a likely transition pathway from the beginning conformation to the end conformation of a transition [58]. Compared with the mixed-ENM method, which is based on an approximate solution of saddle points of the mixed-ENM potential [55], the iENM method solves the saddle points exactly and efficiently by iteratively calling a sparse linear-equation solver [58]. Such improvement has led to better prediction of the order of local and global structural changes as validated by experimental structural data [58]. We have used iENM to compute a possible transition pathway from the closed-channel conformation to the open-channel conformation of pLGICs as captured by the crystal structures of ELIC and GLIC, respectively. The iENM pathway predicts a sequence of structural events beginning with the movements of ligand-binding loops, and is followed by the displacements of loop 2 and loop 7 at the TMD-ECD interface, the tilting/bending of pore-lining M2 helix, and subsequent movements of M4, M3 and M1 helices. The predicted order of structural events is in general agreement with the Φ-value analysis of AChR mutants, which supports a conserved core mechanism for ligand-gated channel opening in pLGICs. Further perturbation analysis has supported the critical role of certain intra-subunit and inter-subunit interactions in dictating the above sequence of events.
We will first discuss the results of ENM-based NMA on the ELIC structure, which will motivate the modeling of the gating transition of pLGICs beyond single-mode description. Next, we will perform the iENM-based transition pathway modeling of the gating transition, and compare the results with the Φ-value analysis. Finally, we will employ perturbation analysis to identify the key interactions that dictate the specific order of structural events predicted by iENM.
Previous NMA studies have found that the lowest normal mode captures a quaternary twist motion of the homo-pentameric α7 nAChR with opposing rotations of the ECD and TMD, which is accompanied by reorganizations within subunits and opening of the channel pore [21], [40]. To explore if similar conformational changes are favored by the ELIC crystal structure, we have performed ENM-based NMA (the cutoff distance Rc is chosen to be 10 Å, which maximizes the cumulative overlap between the lowest 1% normal modes and the observed conformational change from ELIC to GLIC structures).
Indeed, the observed conformational change overlaps significantly with the first normal mode (overlap = 0.54), which describes a quaternary twist motion of ECD relative to TMD [14]. To evaluate if this mode facilitates channel opening, we have generated a new ELIC conformation after displacing the Cα atoms along the direction of the eigenvector of this mode by an RMSD of 3 Å (note: the RMSD between ELIC and GLIC structures is ∼3.1 Å). Then we use the HOLE program [59] to calculate the radius profiles of the channel pore formed by the Cα atoms only (each Cα atom is assigned an atomic radius of 3 Å following [57]). The minimal pore radius is found to be nearly unchanged after the displacement along the first mode (∼0.007 Å), which indicates no opening of the ion-conducting channel. Therefore, unlike α7 nAChR, the first normal mode alone does not support a coupling between the quaternary twist motion and the opening of the channel pore in ELIC. The same observation was made in another NMA study of ELIC structure based on an all-atom force field [36]. Therefore, the single-mode description of ELIC dynamics does not fully support the “twist-to-open” model of the gating transition of pLGICs [21]. Indeed, much of the observed conformational change from ELIC to GLIC structures is not captured by one or a few lowest modes (only 44% is captured by the lowest 1% or 45 normal modes) (see [14]). Therefore, it is necessary to incorporate more normal modes to accurately model the conformational transition that leads to channel opening in pLGICs.
The iENM method [58] enables the simulation of a conformational transition between two given conformations by implicitly utilizing all normal modes from NMA. The iENM method generates a possible transition pathway by solving a set of saddle points for an interpolated potential function constructed from the two ENM potentials based at the beginning and end conformations of a transition (see Methods). We have applied iENM to the conformational transition from the ELIC structure to the GLIC structure to simulate the ligand-gated transition of pLGICs. The resulting iENM pathway consists of 54 intermediate conformations (sampled at an RMSD increment ≥0.1 Å, for a movie see Video S1 of Supporting Information). To dissect the motional order of individual residues, we have calculated a parameter () for each residue (low/high means early/late movement, see Methods). The residues of ELIC are colored according to (residues with low/medium/high are colored red/white/blue, see Fig. 1a). The distribution of supports the following motional order of structural elements: loops A and Cloop 2loop 7, M2 helixM4 helixM3 helixM1 helix (listed in the order of increasing, see Table 1). Therefore, we can deduce a sequence of structural events beginning with the closing of ligand-binding loops in ECD (including loops A and C), followed by the displacements of loop 2 and loop 7 at the TMD-ECD interface, then the tilting/bending of M2 helix, and later the movements of M4, M3 and M1 helices.
To validate the iENM modeling results, we have compared our prediction of sequence of structural events with the Φ-value analysis of the α subunit of AChR mutants [3]. Despite the tremendous differences between the two approaches (the former is based on coarse-grained structural simulation, while the latter is based on kinetic analysis of mutated proteins), we have found broad agreement between their predicted order of structural events during the close-to-open transition of pLGICs summarized as follows (note: unlike , low/high Φ value implies late/early motion):
To further quantify the comparison between the theoretical and experimental Φ values, we have averaged and 1-Φ over residues of 10 secondary structure motifs (loops A, B, C, 2, 7, 9, and helices M2, M3, M4, and M2–M3 linker, see Fig. 2). We do not include residues of M1 helix for the lack of experimental Φ values. The cross-correlation coefficient between the average and 1-Φ is 0.73. As seen from Fig. 2, both of them follow a series of “ascending staircases” from loop A to M3 helix, with the only significant disagreement at loop B (if we remove loop B, the cross-correlation coefficient jumps to 0.95). Possible reasons for this disagreement are: 1. the ligand-binding loops (including loop B) are not well conserved between AChRs and ELIC, so their dynamics may differ; 2. there is a gap in the structural alignment of loop B between ELIC and GLIC (at residue 132, see Fig. S1 of Supporting Information), which may cause inaccuracy in modeling. Further studies are needed to resolve the above possibilities. Additionally, the variations of and Φ values do not seem to agree (see Fig. 2), especially in M2 helix where large scatter in Φ values were found [66]. This disagreement may be attributed to either the model limitation (such as the lack of sidechain and solvation) or the functional divergence between ELIC and AChRs.
To explore the intra-subunit and inter-subunit interactions that may dictate the sequence of structural events predicted by iENM, we have combined iENM with a perturbation analysis ― namely, we perform iENM after turning off the elastic interactions between selected sets of residues, and then analyze how the values of key structural elements change in response to such perturbation. The results are summarized as follows (see Table 1).
First, after turning off the intra-subunit interactions between TMD and ECD, the values of all transmembrane helices (M1–M4) increase significantly, which support the importance of the TMD-ECD interactions in facilitating the motions of transmembrane helices following ligand binding. To further identify the key TMD-ECD interactions, we have turned off those TMD-ECD interactions which involve loop 2, loop 7 and loop 9, respectively. We have found the values of M2 helix increase significantly following the perturbation to loop 2 (but not loop 7 or 9, see Table 1). This finding supports the primary role of loop 2 in coupling ECD with the pore-lining M2 helix, while loop 7 and loop 9 may play some auxiliary role.
Second, after turning off the inter-subunit interactions between the ECDs of adjacent subunits, the values of M2 helix increase significantly, while that of M1 and M4 helices decrease significantly (see Table 1). As a result, the motional order M2M4M3M1 is changed to M4M1M3M2. Therefore, the inter-subunit interactions of ECDs are critical in controlling the sequential motions of transmembrane helices, which may allow ligand binding at the inter-subunit interfaces of ECD to activate or inhibit channel opening.
As revealed by structural comparison, the M2 and M3 helices in GLIC have tilted relative to ELIC as a rigid unit by about 9°. This rotation results in an outward movement of the helix pair away from the pore axis on the extracellular side and an inward movement towards the pore axis on the intracellular side of the channel (see Fig. 3a). Because the channel pore is lined by M2 helices, the pore constriction is shifted from the extracellular side to the intracellular side of the channel (see Fig. 3b). We have compared the intermediate pore conformations predicted by iENM and mixed-ENM method [57]. The mixed-ENM modeling of the TMD found that both the intracellular side and the extracellular side of the pore are closed at the middle of the transition pathway (with = 0.5), and both M2 and M3 helices undergo concerted tilting during the transition [57]. However, the iENM modeling of the entire pLGICs has found that the intracellular side of the pore is closed while the extracellular side of the pore is half-open at the middle of the transition pathway (see Fig. 3b), and the M2 helix moves earlier than the other transmembrane helices (M1, M3 and M4) (see Fig. 1a). Therefore, the modeling of both ECD and TMD is needed to elucidate how ligand binding facilitates the outward tilting of M2 helix followed by the motions of other transmembrane helices.
Our finding that M2 helix moves earlier than the rest of TMD implies a key coupling between M2 helix and ECD that forms early during the gating transition. This result agrees with previous proposals that the inner β-sheet of ECD is significantly correlated to the movement of M2 helix [21], and M2 helix moves independently from the other transmembrane helices [67].
Previous studies have suggested that several conserved loops (including loop 2, loop 7, loop 9, pre-M1 region, and M2–M3 linker, see [68]) at the ECD-TMD interface are involved in the signal transmission from ECD to TMD. It was proposed that loop 2 functions as an actuator that acts on the M2–M3 linker, while loop 7 may serve as a stator to bracket the rotation of M2 and M3 helices [18], [40]. Alternatively, both loop 2 and loop 7 may act together to coordinate the communication between ECD and TMD [69]. Based on the Φ-value analysis of AChR mutants, others suggested that a combination of side-chain interactions at several positions between loop 2 and M2 helix, and loop 7 and M2–M3 linker (specifically, P272 in the AChR) allows energy to be transferred from ECD to TMD [19], [25], [29].
Our finding supports the importance of the early-formed coupling between loop 2 and M2 helix, which is followed by structural rearrangements of loop 7 and M2–M3 linker. This result agrees with a recent MD simulation of ELIC [36], which found that the correlation between residues from loop 7 and M2–M3 linker is most prominent, while the correlation between loop 2 and loop 7 or M2–M3 linker is much weaker. The importance of loop 2 was also suggested by a Targeted MD simulation [27], which found that the closing of loop C transmits to the lower part of the β10 strand, which subsequently displaces loop 2 via the interaction between R209 and E45 (see Fig. 1c), and eventually drives the opening of channel pore [19], [40]. Our finding, rather than pinpointing a signaling path from the ligand-binding site to loop 2 via a chain of interactions [70], supports the collective involvement of a cluster of low- residues in the inner β-sheet (see Fig. 1a), which agree with the proposals that emphasized the collective motion of inner β-sheet [7], [13] and the involvement of a network of interactions including salt bridges [33] and electrostatic interactions [26] in controlling the gating process.
The coupling between loop 2 and M2 helix involves a conserved residue P253 (corresponding to P272 of Torpedo AChR or P269 of α7 AChR, see Fig. 1c). The conformational transition from ELIC to GLIC involves an inward displacement of the tip of loop 2 toward the pore center accompanied by an outward motion of the C-terminus of M2 helix (see Fig. 3a). In a targeted MD simulation of α7 AChR [27], the motion of loop 2 was sterically obstructed by M2–M3 linker (including P269). So the removal of the steric obstruction between these residues permits a rotation of the M2–M3 linker during the gating transition [27]. Similarly, Ref [6] proposed a ‘pin-into-socket’ model via a contact between loop 2 and the hydrophobic pocket formed by the end residues of the apposing M2 helix (αS269–α272 of Torpedo AChR, see Fig. 1c).
Although our modeling favors loop 2 over loop 7 as the primary element in coupling ECD with TMD, we cannot rule out possible loop-7-mediated coupling specific to eukaryotic pLGICs. Functional divergence of loop 7 is conceivable because this loop differs substantially between ELIC and nAChRs in both structure and sequence (see [36]).
The underlying structural picture of iENM modeling differs from the conformational cascade scenario proposed earlier [3]. The former involves a continuous energy-based interpolation between the ELIC and GLIC structures which features highly collective motions of various protein parts at different pace (as characterized by the parameter), and the latter postulates Brownian motions of various protein parts in a discrete and stochastic fashion. The map of reaction progress obtained by using iENM modeling shares some general characteristics with the experimental map of Φ values in AChRs, but it does not show large spread in Φ values that have been revealed by kinetic analysis (see Fig. 2). The iENM and Brownian cascade models of gating represent extreme representations of the transition state ensemble. The former posits a single, frictionless barrier devoid of intermediate states while the latter holds that the barrier has rugged energy landscape that is populated by multiple, metastable intermediates. Future studies are needed to resolve the applicability of these two alternative mechanisms.
High-resolution protein structures are critical for meaningful simulations of protein dynamics. Until recently, for the lack of high-resolution structures of full pLGICs, many MD and NMA simulations were conducted using homology models of pLGICs (see [18], [40]) with uncertain accuracy (see [33]). A main advantage of coarse-grained methods like iENM is that they are insensitive to atomic details and inaccuracy of initial structures. Additionally, the transition pathways predicted by iENM are independent of the specific form of the double-well potential function [58], and the biological relevance of the iENM-predicted pathways has been validated recently by structural data [58]. Therefore, iENM offers highly robust and efficient predictions for the dynamics of protein conformational transitions, including the gating transition of pLGICs. Compared with previous NMA studies based on a single normal mode [21], [39], [40], the iENM method has implicitly utilized all normal modes [58] to explore the conformational transition from ELIC to GLIC, which cannot be accurately described by one or a few normal modes [14]. Therefore, it offers the possibility of dissecting the sequential motions of residues underlying the coupling between ligand binding and channel opening.
The iENM modeling does not explicitly include any bound ligand, which can be justified in light of the recent finding that the conformational pathway of the gating transition of nAChR is essentially unchanged whether or not agonists occupy the ligand-binding sites [71]. On the other hand, the lack of atomic details and solvent modeling would prevent iENM from probing the full details of channel gating dynamics (such as the hydration/dehydration of the pore).
Besides iENM, several alternative computational techniques [54]–[56] may be used to model the pathway of the gating transition. We have tried one of them (MinActionPath) [56], which seems to predict a different order of structural events than iENM (see Table S1). A systematic comparison between iENM and alternative methods will be desirable in the future.
In this study, we assume that the ELIC (GLIC) crystal structure captures the closed-channel (open-channel) form of pLGICs, although further studies are needed to establish the physiological relevance of the ELIC and GLIC structures. Notably, the TMD of ELIC is significantly different from that of the functionally closed structure of nAChR determined by electron microscopy [7]. Surprisingly, the latter resembles the TMD of GLIC. Indeed, an MD simulation of the nAChR structure found the channel pore to shrink further, which suggests that it is not at a fully closed state [35]. It is possible that the closed-state ensemble of ELIC is comprised of multiple conformations as represented by the ELIC structure [12] and the nAChR structure [7].
It is encouraging that we have found remarkable agreements between the iENM modeling based on ELIC/GLIC structures and the Φ-value analysis of α subunit of nAChR mutants [3], although the complexity and richness of the Φ-value analysis results is not reproduced by the iENM. Together, they support a conserved structural mechanism for ligand-gated channel opening in pLGICs. Nevertheless, given the sequential and structural differences between ELIC/GLIC and nAChRs, one should be cautious when using ELIC/GLIC as modeling system to guide functional studies of nAChRs. In the future, we will test the modeling results by performing Φ-value analysis directly on GLIC.
The intermediate conformations predicted by iENM obey the five-fold symmetry which is present in both ELIC and GLIC structures. It is, however, conceivable that structural fluctuations away from the minimal-energy iENM pathway may lead to asymmetric conformations as observed in a recent MD simulation of the GLIC structure [37]. Additionally, as a hetero-pentamer, the motions of five subunits of nAChRs are unlikely to follow the five-fold symmetry. A detailed modeling of the asymmetric motions in nAChRs awaits the solution of open- and close-channel conformations of nAChRs.
Our modeling is based on two crystal structures with different sequences, so a structural alignment is used to model the open form of ELIC using the GLIC structure. Although the uncertainty in alignment does not seem to significantly affect the results of our modeling, it is highly desirable to perform modeling using both closed and open forms of the same protein in the future.
In an ENM, a protein structure is represented as a network of Cα atoms whose minimal-energy conformation is given by a crystal structure. A harmonic potential accounts for the elastic interaction between two Cα atoms that lie within a cutoff distance Rc (set to 10 Å following [55]). The potential energy function of ENM is [48] (1)where is the distance between the Cα atoms i and j, is the value of as given in a crystal structure, N is the number of Cα atoms, and is the Heaviside function. is the force constant of the spring between the Cα atoms i and j. is set to 10 for chemically bonded residues [72] (), and 1 otherwise (the unit of can be arbitrarily chosen without changing the modeling results).
The ENM potential energy can be expanded near a given conformation to the second order: (2)where , is the gradient of at , andis the 3N3N Hessian matrix given by(3)where the elements of 3N3N matrixare given by (4)where () is the x, y, z component of the Cartesian coordinates of the Cα atom (). Note that the matrix elements of are nonzero only if ( = and = ), or ( = and = ), or ( = and = ), or ( = and = ).
From the hessian matrix H1 computed at ( represents the Cα coordinates of the beginning conformation of a transition), we can solve 3N normal modes: the eigenvalue () and eigenvector () of mode m satisfy . To evaluate the similarity between and the observed conformational change from to( represents the Cα coordinates of the end conformation of a transition which is superimposed on ), we compute the overlap coefficient for mode m, and the cumulative overlap for the lowest M modes (after excluding the six translational/rotational modes). () gives the percentage of the observed conformational change captured by mode m (the lowest M modes).
We consider an arbitrary double-well potential function with two minima at the beginning and end conformations of a transition. It satisfies: if, and if, where and are two single-well potentials. Remarkably, the transition pathways generated by iENM (see below) are independent of the mathematic form of which varied in previous studies [54], [55]. The saddle points (SP) of are solved as follows (5)which is equivalent to solving the following equation (after setting )(6)where λ is a parameter of interpolation that varies from 1 to 0 (assuming and). Therefore, the problem of solving SP for the double-well potential function is converted to the problem of minimizing a linearly interpolated potential function. Alternatively, Eq. 6 gives a set of minimal-energy crossing points between and where = is at minimum.
Following the above general formulation, we have proposed an iENM protocol [58] based on a double-well potential function, where and are two ENM potential functions (see Eq. 1) based at the beginning and end conformations of a transition, and is a steric collision energy defined as follows:(7)where = 10, is the minimal distance between the Cα atoms of non-bonded residues in the beginning and end conformations of the transition (∼3 Å for the ELIC and GLIC structures). The chemically bonded residue pairs () are excluded from the summation in Eq. 7. The addition of penalizes steric collisions between residues whose Cα atoms are within a distance of. For the gating transition studied here, steric collisions are not serious so the addition of is not essential in determining the transition pathway.
With the addition of the collision energy, the SPs are solved by setting which is equivalent to solving the following SP equation (the SP is represented by):(8)As λ varies from 1 to 0, traces a pathway that connects the beginning and end conformations of a transition. Because this pathway passes all possible SPs, it gives a ‘universal’ minimum-energy path regardless of the mathematic form of[55], [58]. iENM will output the above pathway as the predicted pathway for the given transition.
We solve Eq. 8 by finding the minima of the linearly interpolated potential function using the Newton-Raphson algorithm (for details, see [58]):
Following [57], a fractional progress parameter () is defined for an intermediate conformation along a transition pathway:, where l is the length of the part of the pathway from the beginning conformation to the intermediate conformation, while L is the total length of the pathway from the beginning conformation to the end conformation. The length of a pathway is computed approximately by summing up RMSDs between consecutive conformations along the pathway.
To quantify the motional order of individual residues along the iENM pathway, we use the following procedure [58]: first, we determine for each residue its ‘crossover conformation’ on the iENM pathway where the residue's Cα atom is at equal distance from its beginning and end positions of the transition (see Fig. 4); next, we assign to each residue the value of its crossover conformation. Residues with low (high)values, as colored by red (blue) in Fig. 1a, move early (late) during the transition.
We use the DALI server [73] to perform the structural alignment of the ELIC structure (PDB code: 2VL0) and the GLIC structure (PDB code: 3EHZ). 282/306 residues in each ELIC subunit are structurally aligned with Z score 22.0 (see Fig. S1 of Supporting Information). Only 24 residues in the ECD of ELIC cannot be aligned to GLIC ― including residues 59–62, 67–70, 132 (in loop B), 151–157 (in loop 9), 176–183 (in loop C). Most of them correspond to insertions in the ECD of ELIC compared with GLIC. For the lack of Cα coordinates for these unaligned residues in the open conformation, we do not include their non-bonded interactions in the ENM potential function (EENM2) constructed from the open conformation.
To check the dependence of iENM modeling on structural alignments, we have tried two alternative structural alignment techniques (SSAP [74], CE [75]), which have obtained slightly different alignments in ECD than DALI. We have got very similar results in the motional order of residues by using these alternative structural alignments.
We have also applied iENM modeling to another GLIC crystal structure (PDB code: 3EAM) and obtained essentially same results.
We structurally align the entire pentamer (except the above mentioned unaligned residues) to account for the motions of all parts equally (both within and between ECD and TMD domains).
|
10.1371/journal.pntd.0001532 | Fatal Dengue Hemorrhagic Fever in Adults: Emphasizing the Evolutionary Pre-fatal Clinical and Laboratory Manifestations | A better description of the clinical and laboratory manifestations of fatal patients with dengue hemorrhagic fever (DHF) is important in alerting clinicians of severe dengue and improving management.
Of 309 adults with DHF, 10 fatal patients and 299 survivors (controls) were retrospectively analyzed. Regarding causes of fatality, massive gastrointestinal (GI) bleeding was found in 4 patients, dengue shock syndrome (DSS) alone in 2; DSS/subarachnoid hemorrhage, Klebsiella pneumoniae meningitis/bacteremia, ventilator associated pneumonia, and massive GI bleeding/Enterococcus faecalis bacteremia each in one. Fatal patients were found to have significantly higher frequencies of early altered consciousness (≤24 h after hospitalization), hypothermia, GI bleeding/massive GI bleeding, DSS, concurrent bacteremia with/without shock, pulmonary edema, renal/hepatic failure, and subarachnoid hemorrhage. Among those experienced early altered consciousness, massive GI bleeding alone/with uremia/with E. faecalis bacteremia, and K. pneumoniae meningitis/bacteremia were each found in one patient. Significantly higher proportion of bandemia from initial (arrival) laboratory data in fatal patients as compared to controls, and higher proportion of pre-fatal leukocytosis and lower pre-fatal platelet count as compared to initial laboratory data of fatal patients were found. Massive GI bleeding (33.3%) and bacteremia (25%) were the major causes of pre-fatal leukocytosis in the deceased patients; 33.3% of the patients with pre-fatal profound thrombocytopenia (<20000/µL), and 50% of the patients with pre-fatal prothrombin time (PT) prolongation experienced massive GI bleeding.
Our report highlights causes of fatality other than DSS in patients with severe dengue, and suggested hypothermia, leukocytosis and bandemia may be warning signs of severe dengue. Clinicians should be alert to the potential development of massive GI bleeding, particularly in patients with early altered consciousness, profound thrombocytopenia, prolonged PT and/or leukocytosis. Antibiotic(s) should be empirically used for patients at risk for bacteremia until it is proven otherwise, especially in those with early altered consciousness and leukocytosis.
| Fatality rate and causes of fatality in dengue-affected patients greatly varied from one reported series to another. A better understanding of the clinical and laboratory manifestations of fatal patients with dengue hemorrhagic fever (DHF) is important in alerting clinicians of severe dengue and improving management. In a retrospective analysis of 10 adults who died of and 299 survived (controls) DHF, dengue shock syndrome (DSS) alone was found in only 20% of dengue-related death, while intractable massive gastrointestinal (GI) bleeding was found in 40%, and DSS with concurrent subarachnoid hemorrhage, intractable massive GI bleeding with concurrent bacteremia, bacterial sepsis/meningitis, and sepsis due to ventilator associated pneumonia each were found in 10%. Early altered consciousness (developed ≤24 h after hospitalization), GI bleeding/massive GI bleeding and concurrent bacteremia were significantly found among the deceased patients. Our data suggest that hypothermia, leukocytosis and bandemia at hospital presentation may be warning signs of severe dengue. Clinicians should be alert to the potential development of massive GI bleeding, particularly in patients with early altered consciousness, profound thrombocytopenia, prothrombin time prolongation and/or leukocytosis. Antibiotic(s) should be empirically used for patients at risk for bacteremia until it is proven otherwise, especially in those with early altered consciousness and leukocytosis.
| Dengue is the most prevalent mosquito-borne viral infection in the world [1]. Clinically, dengue ranges from asymptomatic, nonspecific febrile illness, classic dengue, to dengue hemorrhagic fever/dengue shock syndrome (DHF/DSS) [1]. Fatality rate and causes of fatality in dengue-affected patients greatly varied from one report to another [1]–[13]. While DSS was the major cause of fatality in patients with dengue illness reported in some series [1]–[13], causes other than DSS were predominantly responsible for fatality reported in others [2], [10], [12]–[14]. However, only a small number of dengue-attributed mortality cases were included for analysis in each of these series [2], [8], [10]–[12]. A better description of the clinical and laboratory presentations of cases with fatal outcome may lead clinicians to an earlier recognition of the warning signs of severe dengue resulting in timely and improved management. To achieve this, the importance of continuous analysis of relevant findings in fatal patients from dengue-affected populations cannot be overemphasized.
Among the major dengue epidemics in Taiwan over the past 3 decades, a large dengue outbreak caused by DENV-1 occurred in 1987–1988 in southern Taiwan, followed by another one caused by DENV-2 in 2002 in the same geographic area [15]. During the 2002 dengue epidemic in southern Taiwan, more than 5000 dengue cases were reported, and most of them were DHF that developed in adults [15], [16]; of note, dengue-related fatality was found in 10 adults admitted at Kaohsiung Chang Gung Memorial Hospital (KSCGMH), a 2500-bed facility serving as a primary care and tertiary referral center in this area. In this study, we retrospectively compared the clinical and laboratory features of dengue-affected adults who turned out to be fatal and who survived, and analyzed the fatal dengue cases aiming at understanding the causes of mortality and clarifying the clinical and laboratory evolutions preceding mortality.
The data in this work were analyzed anonymously, and the study was conducted with a waiver of patient consent approved by the Institutional Review Board of KSCGMH (Document No. 99-2671B).
Patients with the diagnosis of dengue admitted to KSCGMH between June and December 2002 were potentially eligible for inclusion in this retrospective study. All clinically diagnosed dengue cases were serologically confirmed by at least one of the following criteria: (i) a positive reverse transcriptase-polymerase chain reaction (RT-PCR), (ii) a positive enzyme-linked immunosorbent assay for specific immunoglobulin M antibody for dengue virus in the acute-phase serum, and (iii) at least fourfold increase in dengue-specific hemagglutination inhibition titers in the convalescent serum when compared to that in acute-phase serum [17]. The diagnosis of DHF was established based on the presence of fever, hemorrhage, thrombocytopenia (<100×109 cells/L) and clinical evidence of plasma leak (i.e., presence of hemoconcentration, pleural effusion, ascites and/or hypoalbuminemia) indicating increased vascular permeability [17]. Hemoconcentration referred to >20% increase in hematocrit calculated as: (maximum hematocrit - minimum hematocrit)×100/minimum hematocrit. The severity of DHF in serologically confirmed dengue patients was stratified based on the World Health Organization (WHO) criteria. Grade I referred to a positive tourniquet test result being the only hemorrhagic manifestation, while grade II referred to spontaneous bleeding. Grade III referred to a circulatory failure manifested by a rapid and weak pulse, as well as a narrowing pulse pressure (≦20 mmHg), whereas grade IV referred to a profound shock, with an undetectable pulse or blood pressure [17]. Grades III and IV of DHF were grouped as DSS [17].
All fatal DHF patients in this series resulted in dengue-related mortality which referred to the death that occurred within three weeks after hospitalization because of DHF. Hypothermia referred to a temperature <36°C detected at least twice from the ear drum of a dengue-affected patient. Massive gastrointestinal (GI) bleeding was defined as the passage of large amount of tarry or bloody stool coupled with hemodynamic instability and/or rapid decrease in hemoglobin level to ≦7.0 g/dL. Acute renal failure was defined as a rapid increase in serum creatinine (Cr) level ≧0.5 mg/dL when compared to that found at the patient's hospital presentation. Acute hepatic failure was defined as a raise in serum alanine aminotransferase (ALT) level ≧400 U/L (reference value, <40 U/L). Leukocytosis was defined as a peripheral white cell count >12000/µL. Bandemia referred to the presence of band-form granulocytes in the peripheral blood. Profound thrombocytopenia referred to a platelet count <20000/µL. Prolongation of prothrombin time (PT) was defined as a PT≥3 seconds than a control, and prolongation of activated partial thromboplastin time (APTT) as an APTT ≥20% than a control. Concurrent bacteremia was defined as a positive bacterial growth from blood that was sampled for culture within 72 h after the patient was hospitalized for dengue.
Demographic, clinical, laboratory and imaging information of the included DHF patients were retrieved from the retrospective review of their medical charts for analyses. Initial laboratory data referred to data detected from the dengue-affected patients upon their arrival at KSCGMH. Pre-fatal laboratory data were those detected from the blood specimens of the fatal patients sampled within the immediate 48 h before fatality.
The 309 DHF patients included for analyses were separated into two groups: those who were fatal (fatal group, N = 10) and those survived (control group, N = 299). The survived patients were those with detailed information available. We compared the demographic, clinical, imaging characteristics and initial laboratory data of the fatal patients and those of the controls, as well as the pre-fatal laboratory data and the initial laboratory data of the fatal patients. Mann-Whitney U test was used in comparison of continuous variables, while the Fisher's exact test was used to in assessment of dichotomous variables. A 2-tailed P<0.05 was considered statistically significant.
A total of 714 adults with dengue illness were found at KSCGMH during the study period, and among them, 10 (8 men and 2 women; median age, 63.5 years [range, 33–78]) with DHF (7 grade II DHF and 3 DSS) turned out to be fatal, accounting for a dengue-related mortality rate of 1.3% (details are shown in Table S1). Of these fatal patients, the time lapses between dengue onset and hospital presentation ranged from 1 to 6 days (median, 2 days), between hospital presentation to fatality 2 to 18 days (median, 4.5 days), and between dengue onset to fatality 4 to 21 days (median, 7.5 days). With the exception of patient 2 in whom the dengue diagnostic test was carried out from the blood specimen collected on the day 3 of his hospitalization, all patients had their blood sampled for dengue diagnosis within 24 h after admission. The median from dengue onset to the definitive diagnosis made was 5 days (range, 4–11 days). Infection with DENV-2 in all fatal patients was confirmed by RT-PCR.
Manifestations indicating plasma leak in these fatal patients included hemoconcentration (patients 2, 5–10), presence of pleural effusion (patients 1, 3–6, 8 and 10) and hypoalbuminemia (patients 1, 2 and 4). Seven patients (patients 1–4, 6, 7 and 9) with grade II DHF experienced shock resulting from bacterial sepsis (patients 1 and 4), concurrent bacterial sepsis and massive GI bleeding (patient 9), and massive GI bleeding (patients 2, 3, 6, and 7). DSS alone was found in 3 (patients 5, 8 and 10) patients. Shock, regardless of cause, developed 1 to 16 days (median, 3 days) after their hospitalizations, and 4 and 17 days (median, 6.5 days) after dengue onset. Among the 3 DSS patients, DSS was recognized on day 3 (patients 8 and 10) and day 6 (patient 5) of their hospitalization, respectively. Patient 2 experienced 2 episodes of massive GI bleeding with hypovolemic shock on day 8 and day 16 of his hospital stay, respectively. Patient 4 with an underlying lung cancer suffered septic shock on day 15 of his hospitalization. The demographic, clinical and laboratory information of the fatal patients and controls is summarized in Tables 1 and 2.
A variety of clinical manifestations were found in each of these 10 fatal patients (Table 1). The leading ones, in decreasing order, were fever (>38°C) (90%), GI bleeding (90%), pleural effusion (70%), bone pain and cough (each 60%). The previously reported early warning signs for severe dengue [11], [17], [18], namely, persistent vomiting was found in 4 (40%) patients, and sustained abdominal pain in 2 (20%). Pulmonary edema developed in 3 (patients 4, 5 and 8) patients; 2 of them with DSS experienced acute pulmonary edema emerged on day 5 (patient 5) and day 6 (patients 8) after dengue onset, respectively, while the other one (patient 4) with lung cancer and hypoalbuminemia (serum albumin, 1.4 g/dL [normal range, 3.0–4.5 g/dL]) experienced septic shock on day 15 of hospitalization (day 17 after dengue onset) thus receiving fluid resuscitation, and pulmonary edema was found on following day. Acute renal failure was found in all of 8 patients (patients 1, 2, 4–9) with data available, and acute hepatic failure in 4 (patients 1, 6–8) (57.1%) of 7 patients with data available. Concurrent bacteremia was noted in 3 (patients 1, 4 and 9) (37.5%) of 8 fatal patients from whom blood was sampled for bacterial culture within 72 h after their hospitalization (Table S1). Of these 3 bacteremic patients, one (patient 1) experienced Klebsiella pneumoniae meningitis, while the other two experienced primary K. pneumonia (patient 4) and Enterococcus faecalis bacteremia (patient 9), respectively. Of a total of 9 patients (patients 2–10) with GI bleeding, 4 (patients 3, 7, 9 and 10) (44.4%) developed GI bleeding at their arrival; 5 (patients 2, 3, 6, 7 and 9) (55.5%) experienced massive GI bleeding, and 3 (patients 3, 7 and 9) (33.3%) developed massive GI bleeding within 24 h after admission. Among the 5 patients with massive GI bleeding, E. faecalis bacteremia was found in one (patient 9); active bleeding was endoscopically found in another (patient 3) with a gastric ulcer, and in the other (patient 7) with hemorrhagic gastritis. Only patients 3 and 7 received endoscopic examination. Among the 5 patients with massive GI bleeding, acute renal failure developed in 2 (patients 2 and 9), and concurrent acute renal and hepatic failure in the other 2 (patients 6 and 7).
Of the 5 fatal patients with consciousness disturbance, 4 (patients 1, 2, 7 and 9) were found to developed altered consciousness within 24 hours and one (patient 5) in the day 4 of his hospital stay. All of these 5 patients had blood sampled for bacterial culture, and one of them (patient 1) had additional cerebrospinal fluid sampled for bacterial culture. Among the 4 patients with early altered consciousness, massive GI bleeding alone (patients 7), uremia and massive GI bleeding (patient 2), E. faecalis bacteremia and massive GI bleeding (patient 9), and K. pneumoniae meningitis and bacteremia (patient 1) each were found in one. Hypernatremia (serum sodium >170 meq/L [normal range, 134–148 meq/L]) was additionally found in patient 2 in day 8 of his hospital stay. Altered consciousness abruptly developed in patient 5 on day 4 of his hospitalization which resulted from subarachnoid hemorrhage disclosed by a brain computed tomography, and a cerebral angiographic study was deferred because of his critical condition and acute renal failure in particular; his blood culture for bacteria was negative, and although hyperkalemia (serum potassium, 7.9 meq/L [normal range, 3.6–5.0 meq/L]) was found on day 7, hemodialysis was not carried out as it was hemodynamically unstable. Neither hyperglycemia nor hypoglycemia was found in the 10 fatal patients in this series. Hyponatremia was not found in our series. Serum calcium level was not assayed in these fatal patients.
Hypothermia was noted in 2 (20%) patients (patients 6 and 9). One patient (patient 9) with hypothermia detected at arrival experienced concurrent primary E. faecalis bacteremia, while the other (patient 6) experienced abrupt change in temperature with a rapid switch from fever to hypothermia on day 4 of her hospital stay, and her blood bacterial culture was negative.
All patients experienced respiratory failure that necessitated mechanical ventilatory support. The mean time from patient's hospital presentation to starting mechanical ventilation was 3 days (range, 1–6 days), and the root causes of respiratory failure included massive GI bleeding (patients 3, 6, 7 and 9), sepsis (patient 1), DSS (patients 8 and 10), subarachnoid hemorrhage (patient 5), persistent drowsiness occurred within 24 h after hospitalization (patient 2) and lung cancer with pleural effusion (patient 4).
Intravenous fluid including 0.9% saline, Ringer's lactate, and 5% dextrose in 0.9% saline was administered at infusion rates ranging from 0.6 mL/Kg BW/h to 2.7 mL/Kg BW/h for the 10 fatal patients before development of shock and/or severe GI hemorrhage. In addition, transfusion of platelets and/or other blood component(s) (i.e., packed red blood cells and/or fresh frozen plasma) was given for these fatal patients. Intravenous fluid replacement and blood transfusion were detailed in Table S1.
Prior to shock development, intravenous fluid supplements with 0.9% saline for the 3 patients with DSS was 1.6 mL/Kg BW/h (patients 5), 1.3 mL/Kg BW/h (patient 8) and 0.8 mL/Kg BW/h (patient 10), respectively. Markedly elevated hemoglobin levels were found in patients with DSS on the day shock developed. Only platelet transfusion was given for these 3 patients before development of DSS. Among the 5 (patients 2, 3, 6, 7 and 9) patients with massive GI bleeding, intravenous fluid (0.9% saline or Ringer's lactate) supplement was infused at rates ranging from 1.4 mL/Kg BW/h to 2.5 mL/Kg BW/h before development of hypovolemic shock, and 2 to 8 units of packed red blood cells were transfused on the day the massive GI bleeding emerged (Table S1).
Because superimposing bacterial sepsis could not be excluded in these critically ill patients, all of them received intravenous antibiotic(s) within 72 h after admission. Upon hospitalization, the 3 patients (patients 1, 4 and 9) with concurrent bacteremia received empirical antibiotic(s) (i.e., ceftriaxone for patient 1, piperacillin and gentamicin for patient 4, and ceftriaxone and penicillin for patient 9) to which the subsequently isolated bacteria were susceptible in vitro.
When it comes to causes of fatality in these10 fatal patients, intractable massive GI bleeding with hypovolemic shock was found in 4 (40%) (patients 2, 3, 6 and 7), DSS alone in 2 (20%) (patients 8 and 10), while DSS with subarachnoid hemorrhage (patient 5), K. pneumoniae bacteremia and meningitis with septic shock (patient 1), sepsis due to mechanical ventilation associated pneumonia (patient 4), as well as concurrent E. faecalis bacteremia and intractable massive GI bleeding with shock (patient 9) each (10%) were found in one patient. None of the fatal patients underwent autopsy.
Pre-fatal leukocytosis was found in 6 (patients 4, 5, 6, and 8–10) (66.7%) of the 9 patients (patients 1, 2, and 4–10) with data available, and bandemia in 4 (patients 1, 2, 5 and 7) (66.7%) of the 6 patients (patients 1, 2 and 4–7) in whom the differential count of peripheral white blood cells was available. Of the 6 patients with development of pre-fatal leukocytosis, 2 (patients 4 and 10) had leukopenia upon their arrival, 4 had their blood sampled for bacterial culture and E. faecalis bacteremia (patient 9) was found in one patient (25%), all experienced GI bleeding, and 2 (patients 6 and 9) (33.3%) developed massive GI bleeding. Prolongation of pre-fatal PT was found in 6 (75%) (patients 2, 4–6, 8 and 9) of the 8 (patients 1, 2, 4–6, and 8–10) patients with data available. Of note, all the 6 patients with pre-fatal PT prolongation developed GI bleeding, and of them, 3 (50%) experienced massive GI bleeding. Pre-fatal profound thrombocytopenia was found in 6 (60%) fatal patients (patients 1, 3, 5 and 8–10); of them, 5 (patients 3, 5 and 8–10) (83.3%) developed GI bleeding, and 2 (33.3%) (patients 3 and 9) experienced massive GI bleeding. Pre-fatal hyperkalemia was found in only 1 (patient 5) of the 6 (patients 2, 4, 5, 7–9) patients with data available.
Significant differences in demographics and clinical manifestations between fatal patients and controls included male gender (80% vs. 44.1%, P = 0.047), hypovolemic shock due to massive GI bleeding (50% vs. 0.7%, P<0.001), concurrent bacteremia (37.5% vs. 3.9%. P = 0.010), concurrent bacteremia with shock (25% vs. 1.3%, P = 0.022), DSS (40% vs. 2.3%, P<0.001), pulmonary edema (30% vs. 2.8%; P = 0.005), acute renal failure (100% vs. 2%; P<0.001), acute hepatic failure (57.1% vs. 4.4%; P<0.001), hypothermia (20% vs. 0%, P = 0.001), GI bleeding (90% vs. 16%, P<0.001), massive GI bleeding (50% vs. 0.7%; P<0.001), subarachnoid hemorrhage (10% vs. 0%, P = 0.032) and early altered consciousness (40% vs. 0%, P<0.001) (Table 1).
Significant higher proportion of bandemia (37.5% vs. 1.8%; P = 0.001) from initial laboratory data between the fatal patients and the controls, and significant higher proportion of pre-fatal leukocytosis (66.7% vs. 10%; P = 0.020) and lower pre-fatal platelet count (median, 17000 cells/µL vs. 35000 cells/µL; P<0.001) as compared to the initial laboratory data of the fatal patients were found (Table 2).
The time interval from the dengue onset to patients' arrival at KSCGMH between the fatal and control groups did not differ significantly (median, 2 days vs. 3 days; P = 0.055) (Table 1). The timing of admission in both the fatal and control groups allowed us to evaluate the critical evolutionary changes in the dengue-affected patients because critical events (e.g., dropped blood pressure and circulation collapse) usually occur between day 3 and day 7 of the disease course [17], [18].
Dengue case fatality rate was reported to vary from 0.5% to 5.0% [2]–[4], [7], [9], [10], [12]. However, once DSS developed, the case fatality may soar to as high as 12–44% [3]–[5]. Our series showed that of all dengue-related deaths, DSS alone accounted for only 20%, while intractable massive GI bleeding alone for 40%, and DSS with concurrent subarachnoid hemorrhage, intractable massive GI bleeding with concurrent bacteremia, bacterial sepsis with meningitis, and sepsis due to ventilator associated pneumonia each were responsible for 10%.
DSS is characterized by severe plasma leak that leads to rapidly developed shock, and timely volume replacement is the cornerstone of therapy for the affected patients [17]. Notably, the volumes of intravenous fluid supplement prior to the full blown development of DSS in the 3 patients (patients 5, 8, 10) in our series were obviously suboptimal [17], [18]; in spite of the subsequent fluid resuscitation and blood/blood component transfusion, they died of profound shock and multi-organ failure between day 6 and day 7 after the onset of illness. The pulmonary edema developed in the 2 patients (patients 5 and 8) with DSS on day 5 and day 6 after the dengue onset, respectively, was accompanied by a concurrent marked hemoconcentration (see Table S1 for details) suggested continuous fluid leakage from the intravascular compartment to the extravascular compartment and the lung alveolar space in particular, leading to profound shock and pulmonary edema. In contrast, the pulmonary edema developed on day 16 in patients 4 who had an underlying lung cancer with malignant pleural effusion obviously resulted from fluid overload.
The latest WHO scheme classified dengue in terms of clinical severity as severe dengue (i.e., presence of severe bleeding, severe plasma leak and/or severe organ involvement) or non-severe dengue; for practical reasons, patients with non-severe dengue were further separated into those with warning signs (i.e., abdominal pain, persistent vomiting, clinical fluid accumulation, mucosal bleeding, lethargy/restless, liver enlargement, and increase in hematocrit in concurrent with rapid decrease in platelet count) and those without them [18]. Severe dengue patients with aggravated plasma leak and/or bleeding necessitate aggressive fluid resuscitation and additional blood transfusion as necessary, while non-severe dengue patients with warning sign(s) require strict observation, appropriate medical intervention and intravenous hydration, as they are at high risk for evolving into critical phase–severe dengue [18]. In addition to the aforementioned ones, our data suggest that leukocytosis, bandemia and hypothermia may be warning signs of severe dengue. From the pathophysiological point of view, leukocytosis and/or bandemia indicates a superimposing bacterial infection and/or other stressful stimuli [19]. Our data suggested that massive GI bleeding (33.3%) and bacteremia (25%) be the major causes of DHF patients' pre-fatal leukocytosis. Significantly, leukocytosis was found in the deceased patients before their death, and bandemia was found at the fatal patients' hospital presentation (Table 2); the latter suggest that bandemia may be an early warning parameter of severe dengue.
Mucosal bleeding may occur in any patient with dengue, and if the patient remains stable with fluid resuscitation/replacement, the mucosal bleeding should be considered a minor one [18]. Minor mucosal bleeding in dengue patients often results from diapedesis of erythrocytes around blood vessels with little inflammatory reaction [20]. If major bleeding occurs, it is usually from the GI tract [2], [12], [21], [22], and one of the risk factors for major GI bleeding is the existence of a peptic ulcer [21], which is unfortunately not uncommonly develops in patients under stress [23]. One dengue series with 30 fatal dengue cases included disclosed that 80% of the fatal patients experienced GI bleeding, and severe bleeding with shock accounted for 30% of fatality [2]. The previously reported data [2], [12], [22] and ours suggest that even minor or moderate GI bleeding should be regarded as a warning sign of severe dengue and the patient in question needs close monitoring, as it potentially evolved into life-threatening intractable massive GI bleeding. Gastric ulcer and hemorrhagic gastritis each endoscopically found in one fatal patient with massive GI bleeding in our series raises the question of whether a H2-blocker or proton-pump inhibitor should be used in patients with severe DHF patients for prevention of massive GI bleeding. Further study is needed to answer this question.
Of note, massive GI bleeding was found in 75% of patients who experienced early altered consciousness (Table 1); 50% of patients with pre-fatal PT prolongation and 33.3% of patients with pre-fatal profound thrombocytopenia experienced massive GI bleeding. These data suggested that clinicians be alert to the potential development of severe GI bleeding when facing DHF patients with altered consciousness, and persistent PT prolongation and thrombocytopenia, and thereby initiate a timely management as necessary.
Abdominal pain and persistent vomiting, the previously reported clinical warning signs of severe dengue [11], [17], [18], did not differ between the fatal patients and controls in this series. In contrast, hypothermia significantly found in the fatal patients suggested that it should be considered a warning sign of severe dengue. Dengue-affected patients with hypothermia should therefore be intensively monitored, and aggressive workup is needed to clarify the potential cause(s) so that an effective treatment can be started timely.
It is noteworthy that 50% of our patients presented with early altered consciousness suffered concurrent bacterial sepsis (K. pneumoniae meningitis and E. faecalis bacteremia, respectively), highlighting the need for an immediate empirical antibiotic administration for dengue-affected patients with altered consciousness for the presumably superimposing bacterial sepsis until it is proven otherwise.
The bacteria (2 K. pneumoniae and 1 E. faecalis isolates) grew from culture of blood of 3 patients (two of them each with the underlying hypertension and lung cancer) sampled within 48 h after their admission were of normal intestinal flora. Our observation and previously reported concurrent bacteremia in patients with DHF caused by the members of Enterobacteriaceae [12], [24] suggested that DHF patients are vulnerable to bloodstream invasion by microbes from the intestinal tract where they normally inhabit. These findings are consistent with the development of portal of entry for bacteria in bowels by disintegration of intestinal mucosal barriers in DHF patients reported previously [25], [26]. Of the fatal bacteremic patients in this series, one patient with K. pneumoniae bacteremia and the other with simultaneous K. pneumoniae bacteremia and meningitis clearly experienced septic shock, while the shock in the patients who suffered massive GI bleeding and concurrently E. faecalis bacteremia might result from both hypovolemia and sepsis in view that E. faecalis has relatively low clinical virulence [27]. Nevertheless, our data suggest that when it comes to empirical use of antibiotic for suspicious concurrent bacteremia in dengue-affected patients, it is reasonable to cover bacteria from the intestinal tract.
It is not surprising that acute renal failure (100%) and acute hepatic failure (57.1%) exclusively developed in fatal DHF patients in our series, as severe plasma leakage, massive bleeding and/or profound shock would lead to tissue hypo-perfusion, potentially rendering acute renal failure and hepatic failure [16], [28], [29].
There are some limitations in the present study. First, the fatalities in this series may be biased by patients' severity resulting from patient selection and referral pattern in a single medical center. Second, the lack of a standardized treatment protocol for severe dengue cases might bias patients' clinical outcomes in this retrospective analysis; this study thus addressed the pre-fatal clinical and laboratory evolutions in the deceased DHF patients, but not the appropriateness of treatment for them. Third, the small number of fatal cases made the statistical power quite small.
In summary, our report highlights the causes of fatality other than DSS in patients with severe dengue, and suggested that in addition to those mentioned by the WHO 2009 scheme, hypothermia, leukocytosis, and bandemia may be warning signs of severe dengue. Early altered consciousness and GI bleeding/massive GI bleeding were significantly found among deceased DHF patients in this series. Dengue-affected patients should be closely monitored and appropriately treated once GI bleeding emerges, as it potentially evolved into massive GI bleeding; once massive GI bleeding develops, patients are at high risk for mortality, and this may be particularly true in patients with early altered consciousness, leukocytosis, profound thrombocytopenia and PT prolongation. Antibiotic(s) should be empirically added for patients at risk for developing concurrent bacteremia, especially in those with early altered consciousness and emergence of leukocytosis. Our data suggest that bandemia at hospital presentation may be a warning parameter for severe dengue, and monitoring the potential emergence of leukocytosis and persistence of thrombocytopenia may be helpful in evaluation of the progressive dengue severity. Further study is needed to confirm our observations. The findings of the suboptimal fluid resuscitations and blood/blood component transfusions in some of the fatal cases in this series underscores the importance of a timely effective volume replacement by fluid infusion and blood/blood component transfusion for patients with a severe dengue.
|
10.1371/journal.pgen.1005020 | microRNAs Regulate Cell-to-Cell Variability of Endogenous Target Gene Expression in Developing Mouse Thymocytes | The development and homeostasis of multicellular organisms relies on gene regulation within individual constituent cells. Gene regulatory circuits that increase the robustness of gene expression frequently incorporate microRNAs as post-transcriptional regulators. Computational approaches, synthetic gene circuits and observations in model organisms predict that the co-regulation of microRNAs and their target mRNAs can reduce cell-to-cell variability in the expression of target genes. However, whether microRNAs directly regulate variability of endogenous gene expression remains to be tested in mammalian cells. Here we use quantitative flow cytometry to show that microRNAs impact on cell-to-cell variability of protein expression in developing mouse thymocytes. We find two distinct mechanisms that control variation in the activation-induced expression of the microRNA target CD69. First, the expression of miR-17 and miR-20a, two members of the miR-17-92 cluster, is co-regulated with the target mRNA Cd69 to form an activation-induced incoherent feed-forward loop. Another microRNA, miR-181a, acts at least in part upstream of the target mRNA Cd69 to modulate cellular responses to activation. The ability of microRNAs to render gene expression more uniform across mammalian cell populations may be important for normal development and for disease.
| microRNAs are integral to many developmental processes and may 'canalise' development by reducing cell-to-cell variation in gene expression. This idea is supported by computational studies that have modeled the impact of microRNAs on the expression of their targets and the construction of artificial incoherent feedforward loops using synthetic biology tools. Here we show that this interesting principle of microRNA regulation actually occurs in a mammalian developmental system. We examine cell-to-cell variation of protein expression in developing mouse thymocytes by quantitative flow cytometry and find that the absence of microRNAs results in increased cell-to-cell variation in the expression of the microRNA target Cd69. Mechanistically, T cell receptor signaling induces both Cd69 and miR-17 and miR-20a, two microRNAs that target Cd69. Co-regulation of microRNAs and their target mRNA dampens the expression of Cd69 and forms an incoherent feedforward loop that reduces cell-to-cell variation on CD69 expression. In addition, miR-181, which also targets Cd69 and is a known modulator of T cell receptor signaling, also affects cell-to-cell variation of CD69 expression. The ability of microRNAs to control the uniformity of gene expression across mammalian cell populations may be important for normal development and for disease.
| The complexity of developmental processes in metazoans relies on mechanisms that confer a degree of robustness against environmental and genetic variation [1]. microRNAs are small non-coding RNAs that negatively regulate gene expression at the post-transcriptional level by reducing mRNA stability and/or translation. Their role in dampening gene expression makes microRNAs potential building blocks for gene regulatory circuits that can stabilize gene regulatory networks [2–5].
Gene expression is subject to intrinsic stochasticity associated with mRNA transcription and translation, as well as extrinsic noise such as fluctuations in upstream regulators. Gene expression noise is not restricted to protein coding genes: the expression of primary microRNA transcripts, their processing into pre-microRNAs, nuclear export, processing into mature microRNAs, association with RISC components, etc., presumably all have stochastic components. The participation of microRNAs in the regulation of protein-coding genes could therefore add noise contained in both microRNA and protein-coding systems. Feed-forward loops (FFLs) are recurrent network motifs that can reduce gene expression noise by buffering fluctuations in upstream regulators [6]. Placing the expression of a microRNA and its target mRNA under the control of common upstream regulators can link the production of mRNAs to the production of microRNAs that target the mRNAs. Theoretical considerations [2] and computational simulations [7, 8] suggest that this circuit topology, which resembles an incoherent FFL, allows microRNAs to buffer protein expression against fluctuations in the activity of upstream regulators [9]. In silico models predict that FFL regulation enables microRNAs to reduce not only the level of target gene expression, but also cell-to-cell variability [7, 8]. Data from synthetic circuits indicate that co-expression of microRNAs and target mRNAs can reduce temporal fluctuations and in some cases cell-to-cell variability in reporter gene expression [7, 10].
Emerging experimental evidence supports a role for microRNAs in biological robustness [2]. microRNAs affect several phenotypic traits in Drosophila, for example by stabilizing the regulation of the enhancer of split transcription factor to guide sensory organ development under conditions of environmental flux [11]. Loss of microRNAs can increase the variation of primordial germ cell numbers [12, 13] and sensory bristles [14], and quantitative phenotypic traits in the Drosophila cuticle [15]. These data demonstrate that microRNAs can buffer variation in phenotypic traits, but it is not clear whether this is achieved by reduced variation in the expression of microRNA target genes or the operation of thresholds for phenotypic outcomes [16]. Zebrafish miR-26b and ctdsp2 mRNA are encoded by the same primary transcript, and ctdsp2 mRNA is a target of miR-26b [17]. The processing of miR-26b is developmentally controlled during neuronal differentiation, effectively initiating a microRNA-mediated incoherent FFL but the consequences for cell-to-cell variation in the expression of ctdsp2 have yet to be established [17]. microRNAs can dampen temporal oscillations in gene expression in C. elegans [18] and reduce fluctuations in the average expression of reporter constructs in mammalian cells [19]. Measurements at the population level, but not in individual cells, showed that methyl CpG-binding protein 2 (MeCP2) acts through BDNF to induce the neuronal miRNA miR-132, which feeds back to repress MeCP2 [20]. However, simple negative feedback loops like this may increase noise as determined experimentally and computationally [7]. The miR-17-92 family forms a complex network with Cyclin D1 in neuronal progenitors, and the variability of Cyclin D1 expression was increased by heterozygosity in Dicer [21]. The relationship between microRNAs and variability of target gene expression is complicated in this system, since miR-17-92 is required for the differentiation of mouse cortical neuronal progenitors [22], and reduced microRNA expression affects the frequency of proliferating neuronal progenitors as well as the expression of Ccnd1 within them [21, 22]. That loss of microRNAs can also result in reduced variability in the expression of pluripotency markers was recently demonstrated for mouse ES cells [23].
Here we address the impact of the microRNA biogenesis pathway on cell-to-cell variability of endogenous gene expression in mouse thymocytes (developing T cells). This system offers a number of key advantages. T cell development proceeds in a series of discrete developmental steps that are defined by the expression of cell surface markers [24]. This allows for (i) the precise definition and isolation of cell populations at specific developmental stages for the molecular characterization of microRNA and target mRNA expression, (ii) the use of developmentally regulated Cre transgenes for the synchronous deletion of conditional alleles of the RNase III enzyme Dicer, (iii) the verification of reduced microRNA expression at defined developmental stages, and (vi) like-for-like comparisons between control and Dicer-deficient cells at the same developmental stage. Thymocytes readily form cell suspensions that are ideally suited for analysis and sorting by flow cytometry, and high-quality reagents are available to enable quantitative flow cytometry at the single cell level [25]. Using this approach we demonstrate that microRNAs can reduce cell-to-cell variation of target gene expression in mammalian cells. The activation-induced microRNA target CD69 was regulated by microRNAs in two different ways. miR-181a affected variation by modulating the responsiveness of thymocytes to activation signals, acting at least in part upstream on the target mRNA Cd69. Members of the miR-17-92 cluster were co-regulated with the target mRNA Cd69, resembling an activation-induced incoherent FFL.
We previously characterised an experimental system where a developmentally regulated Lck-Cre transgene deletes a conditional Dicer allele in developing mouse thymocytes [26]. As a result, the expression of Dicer-dependent microRNAs was reduced by ∼90% at the CD4 CD8 double positive (DP) stage of development (Fig. 1A) [26]. miReduce analysis [27] of 3'UTR motifs associated with post-transcriptional de-repression in Lck-Cre DP thymocytes (see GSE57511) showed enrichment for microRNAs miR-181, miR-17 and miR-142 (Fig. 1B).
We evaluated flow cytometry as an approach to determine protein expression by individual cells. To estimate technical noise we examined CD8a and CD8b, which are expressed as obligate heterodimers in DP thymocytes. The mean expression and the coefficient of variation (CV) of CD8a and CD8b were very similar for control and Dicer-deficient DP thymocytes (Fig. 1C, left and centre), as was the ratio of CD8a/CD8b expression for individual control and Dicer-deficient DP thymocytes (Fig. 1D, right). The CV of these ratios defines the upper bound of technical noise [25]. Based on published criteria for quantitative flow cytometry [25] we identified antibodies directed against putative microRNA targets including CD44, a predicted target of miR-21 (www.targetscan.org) and established target of miR-34 [28] as well as the predicted miR-181 targets Ly6a and H2-K1 (www.targetscan.org; Fig. 1D). As expected based on elevated mRNA expression (see GSE57511), Dicer-deficient thymocytes showed higher average expression of CD44, Ly6a and H2-K1 than control cells. Interestingly, the cell-to-cell variation of CD44, Ly6a and H2-K1 expression was also increased in Dicer-deficient thymocytes (Fig. 1D, E) while there was no change in the negative staining control, MHC class II (H2-Ab1; Fig. 1D).
We used the CV as a stringent measure of variation because in contrast to the standard deviation (SD), the CV is expected to decrease as the mean expression increases (CV = standard deviation/mean). An increase in the CV at the same time as an increase in mean protein expression therefore unambiguously indicates an increase in cell-to-cell variation (Fig. 1D, E). As the CV is expected to decrease with the level of expression, the finding of an increased CV in Dicer-deficient cells that expressed higher protein levels prompted several control experiments. First, we asked whether the increased CV was explained by residual microRNA-retaining cells. Experimental mixing and computational modeling experiments indicated that this was highly unlikely (S1 Fig.). Second, the apparent impact on the CV could be due to technical limitations in the detection of low levels of protein expression: if microRNAs reduce expression below the sensitivity or our instrumentation we would detect expression—and associated noise—only in Dicer-deficient cells but not in wild type cells. To address this concern we asked whether the level of protein expression detected in wild type and Dicer-deficient cells was biologically meaningful. We sorted control and Dicer-deficient thymocytes according to the level of CD44 and Ly6a protein detected by flow cytometry and carried out quantitative reverse transcriptase (RT)-PCR for Cd44 and Ly6a transcripts (Fig. 1F). The data correlated CD44 and Ly6a protein expression with the abundance of Cd44 and Ly6a mRNA in both control and Dicer-deficient DP thymocytes. Although Cd44, Ly6a and H2-K1 mRNAs are not confirmed direct microRNA targets in thymocytes, these data demonstrate that our instrumentation discriminates meaningful levels of protein expression.
To unambiguously determine the impact of microRNAs on cell-to-cell variability of target gene expression on a direct microRNA target in thymocytes we focused on CD69, which is inducibly expressed in response to T-cell activation [29]. CD69 controls cell migration and sphingosine 1-phosphate signaling [30], and the Cd69 mRNA is a well-characterised target of miR-181 and other microRNAs [31–33]. In response to activation signals through the T cell receptor (TCR), DP thymocytes initiated the expression of CD69 (Fig. 2A), and graded activation signals induced a proportional increase of Cd69 mRNA and CD69 protein (S2 Fig.). As expected for an established microRNA target, average CD69 levels were higher in Dicer-deficient than in control DP thymocytes (S2B Fig.). In addition, Dicer-deficient DP thymocytes showed an increase in the CV of CD69 expression (Fig. 2B), and this increase was seen over a range of activation conditions (Fig. 2C). The broader distribution of CD69 expression among Dicer-deficient DP cells was due in part to a greater fraction of CD69hi cells (Fig. 2D, characterized by the co-expression of CD25 Fig. 2A). In addition, Dicer-deficient DP thymocytes showed an increased CV of CD69 expression within the CD69hi CD25+ subset (Fig. 2E, F). Hence, Dicer-deficient DP thymocytes showed increased cell-to-cell variability in the expression of the microRNA target CD69. This was true whether the CV was assessed for the entire DP thymocyte population, or separately for the CD69+ population or the CD69high CD25+ subset (Fig. 2G).
Taken together, our results indicate that microRNAs can shape not only the level but also the cell-to-cell variability of protein expression in developing thymocytes. To investigate the underlying mechanisms we next identified endogenous microRNAs that target Cd69 in DP thymocytes.
The Cd69 3'UTR contains predicted sites for miR-181, miR-130 and miR-17/20 (http://www.targetscan.org) and there is firm experimental evidence for Cd69 regulation by miR-181a, miR-130 and the miR-17-92 cluster (which encodes the microRNAs miR-17, -18, -19a, -19b, -20a, and -92 [34] in T lymphocytes [31–33].
To evaluate the impact of endogenous microRNAs on the expression of proteins linked to the 3'UTR of Cd69 we developed a dual fluorescence reporter construct. The construct encodes the two fluorescent reporter proteins, eGFP and mCherry, under the control of separate retroviral long terminal repeat (LTR) and mouse Pgk promoters, as well as a cloning site 3’ of the eGFP transcript (Fig. 3A). In a manner similar to luciferase reporter constructs, 3’ UTRs of interest can be cloned into this site to measure their impact on the expression of GFP relative to mCherry. In contrast to heterologous reporter assays, however, this system allows to delineate the biological activity of endogenously expressed microRNAs after retroviral gene transfer of the reporter construct into primary cells. We characterised this dual fluorescence reporter system in mature CD4+ T cells that were isolated from lymph nodes and activated in vitro to render them receptive to retroviral gene transfer (S3A Fig.). To determine the impact of Dicer on the expression of eGFP linked to the 3'UTR of Cd69 we transduced wild type and CD4Cre Dicerlox/lox (Dicer-deficient) [35] CD4+ T cells with the reporter construct containing the entire Cd69 3'UTR (Fig. 3A, Cd69 3'UTR). Fig. 3B shows a representative dot plot of mCherry and eGFP-Cd69 3'UTR expression in control (black) versus Dicer-deficient CD4+ T cells (red). Compared to the empty control vector, wild type CD4+ T cells expressed eGFP-Cd69 3'UTR at a lower level and Dicer-deficient CD4+ T cells expressed eGFP-Cd69 3'UTR at a higher level (Fig. 3B, C), indicating that as well as repressive miRNA-binding sites, the CD69 3’ UTR may contain sequences that enhance expression. Mutation of the miR-181 site in the Cd69 3'UTR did not measurably affect the expression of eGFP in mature CD4+ T cells (Fig. 3C), which express only low levels of the developmentally regulated miR-181 [31]. However, deletion of the miR-130 and particularly the miR-17/20 site resulted in increased eGFP expression in wild type CD4+ T cells (Fig. 3C).
Next, thymocytes were transduced with Cd69 3'UTR reporter constructs and maintained for 24 hours in reaggregation thymic organ cultures until the expression of fluorescent reporters by CD4+ CD8+ DP thymocytes was recorded by flow cytometry. In contrast to mature CD4+ T cells, the miR-181 site affected eGFP-Cd69 3'UTR expression in CD4+ CD8+ DP thymocytes, which express maximal levels of the developmentally regulated miR-181 [31]. The predicted sites for miR-130 and miR-17/20 within the Cd69 3' UTR also affected eGFP reporter gene expression in DP thymocytes.
Taken together, these results show that eGFP-Cd69 3'UTR expression was Dicer-dependent in mature CD4+ T cells (Dicer-deficient thymocytes could not be used successfully for retroviral gene transduction and subsequent reaggregate organ cultures) and that the impact of predicted microRNA binding sites reflected the developmental regulation of microRNAs [31]. We focused our subsequent analysis on miR-181a and the miR-17-92 cluster.
To explore the influence of miR-181 on the CV of CD69 expression we analysed DP thymocytes deficient in mir-181ab1, which accounts for most of the miR-181a and -b copies in DP thymocytes [36]. Following activation, miR-181-deficient DP thymocytes showed increased mean CD69 expression (control = 245 ± 17, mean miR-181 ko = 278 ± 10, n = 26, P<10–10, 2-tailed T-test). Interestingly, CD69 expression in miR-181-deficient DP thymocytes also showed an increased CV (Fig. 4A) over a range of activation conditions (Fig. 4B). The increased CV was due mainly to a higher fraction of CD69hi cells among miR-181-deficient DP thymocytes (Fig. 4C). The CV of CD69 expression within the CD69hi subset was only mildly affected (Fig. 4D).
These results show that miR-181 is an important determinant of cell-to-cell variability in CD69 expression in activated DP thymocytes, and is required to restrict the fraction of CD69hi DP cells. This is consistent with a role for miR-181 as a modulator of TCR signaling [36–38] (Fig. 4E).
miRNA expression responds to T-cell activation signals [34, 35, 39–45]. Many microRNAs are downregulated upon T-cell activation [40–43], but the expression of the miR-17-92 cluster is upregulated in activated mouse and human T cells [45]. Since the miR-17-92 cluster encodes microRNAs that target the Cd69 3'UTR, including miR-17 and miR-20a (Fig. 3), we investigated how the expression of miR-17 and miR-20a was affected by the activation of DP thymocytes.
We applied graded stimuli (0, 0.1, 1 and 10μg H57/ml) to induce a graded increase in Cd69 mRNA expression (Fig. 5A). Interestingly, this graded increase in Cd69 mRNA was accompanied by a proportional upregulation of miR-17 and miR-20a expression (Fig. 5A).
Next, we asked how miR-17 and miR-20a expression was related to the range of responses by individual cells to a uniform extracellular signal. When stimulated with a fixed concentration of TCR antibody, DP thymocytes expressed a range of CD69 protein, from undetectable to high (see Fig. 2A). We applied a uniform stimulus (1μg H57/ml) and sorted DP thymocytes that expressed no detectable CD69 protein (CD69 neg), low levels of CD69 (CD69 lo), intermediate levels of CD69 (CD69 med) or high levels of CD69 (CD69 hi). Increasing expression of CD69 protein correlated with increasing Cd69 mRNA levels, and with incremental expression of miR-17 and miR-20a (Fig. 5B).
Hence, activation signals of increasing strength induce a proportional upregulation of the microRNAs miR-17 and miR-20a and the target mRNA Cd69. Furthermore, cells exposed to a uniform stimulus show a range of responses, and the induction of microRNAs and mRNA target is coordinated with the expression of the protein encoded by the target mRNA in individual cells. These findings suggest that miR-17, miR-20a and Cd69 are co-regulated. Mechanistically, the transcription factor Myc provides a link between thymocyte activation and the coordinated regulation of Cd69 and the miR-17-92 cluster. Myc expression is upregulated by signals that drive lymphocyte activation and mediates downstream transcriptional responses [46] (Fig. 5B). Myc and Cd69 are induced by shared signaling pathways downstream of the TCR [47], and Myc directly activates transcription of the miR-17-92 cluster [48]. These data indicate that the microRNA target mRNA Cd69 and microRNAs of the miR-17-92 cluster form an incoherent feed-forward loop in response to TCR signaling (Fig. 5C).
The coordinated regulation of miR-17, miR-20a and Cd69 in response to TCR signaling provides a potential mechanism for restricting cell-to-cell variability of microRNA target gene expression. To explore this idea further we implemented computational models of noise regulation by microRNAs. In one model, a microRNA and target mRNA are induced together and the microRNAs inhibits the translation of the mRNA as part of an incoherent feedforward loop [8] (S4A Fig.). In an alternative model, a co-regulated pair of microRNA and mRNA interact to induce mRNA degradation [49] (S4B Fig.). Both models predict that microRNA feedforward regulation reduces the mean and the CV of target expression. To implement a more specific model of CD69 regulation we estimated the mRNA copy numbers for Cd69 and the microRNA copy numbers for miR-17 and miR-20a in resting and activated T cells (see legend Fig. 5D). This model predicts that thymocyte activation results in mean CD69 expression of 887 with a CV of 10.2% when Cd69 mRNA and miR-17/miR-20a are induced together (activated with microRNA FFL, filled grey histogram in Fig. 5D). In contrast, induction of Cd69 mRNA without upregulation of miR-17/miR-20a results in a higher mean (1300) and increased CV (14.6%, activated without microRNA FFL, filled red histogram in Fig. 5D, P<10–4). This result is consistent with our experimental data where the mean and CV of activation-induced CD69 expression were significantly elevated in Dicer-deficient thymocytes: in the absence of a functional microRNA biogenesis pathway, the activation-driven increase in Cd69 mRNA was not balanced by increased miR-17 and miR-20a expression.
microRNAs are essential for mammalian development [50] due to their diverse range of regulatory roles in gene expression. They facilitate developmental transitions by the reciprocal regulation of microRNAs and their targets in cell types derived from a common progenitor [23] and by participating in regulatory circuits with switch-like functions [5], buffer against environmental and genetic variation [2–5], limit intrinsic transcriptional noise (by allowing mRNA 'overproduction' and post-transcriptional removal of excess transcripts) [2, 51] and reduce extrinsic noise as part of FFLs [2, 7, 8, 10], as demonstrated in the current study for a mammalian developmental system.
The inducible expression of the established microRNA target Cd69 [31–33] allowed us to explore molecular mechanisms by which microRNAs affect the cell-to-cell variability of target gene expression in thymocytes. miR-181 is a known modulator of TCR signal transduction [36–38] and our data show that the deletion of mir-181ab1 affected the CV of CD69 expression mainly by altering the proportion of thymocytes that expressed CD69 at high levels. The expression of miR-181a is downregulated as thymocytes mature [31] and in this way may account for developmentally regulated changes in the responsiveness of thymocytes to TCR signaling. Our data are consistent with this model and further suggest that developmental regulation of miR-181a reduces cell-to-cell variability of thymocyte responses to TCR signaling. A different mechanism applies to the regulation of CD69 by miR-17 and miR-20a, two microRNAs of the miR-17-92 cluster. Our data show that the expression of these microRNAs is induced together with Cd69 mRNA in response to TCR signals, and that the expression of CD69 protein, Cd69 mRNA and miR-17/miR20a is proportional in thymocytes. This co-regulation of microRNAs and target mRNA has the potential for feed-forward regulation. While the specific circuitry that places Cd69 and miR-17-92 under the shared control of TCR signaling remains to be elucidated, Myc and Cd69 are induced by shared signaling pathways downstream of the TCR [47], and Myc directly activates transcription of the miR-17-92 cluster [48].
Computational and experimental data suggest that FFLs can confer microRNA-mediated robustness of target gene expression by reducing noise that originates upstream of the transcription of the target mRNA itself [8, 10]. Modeling the impact of microRNA feedforward regulation either by translational inhibition or mRNA degradation predicted a reduction in the mean and CV of target expression [8, 49]. This was confirmed by modeling the experimentally estimated copy numbers of Cd69 mRNA and miR-17 and miR-20a in resting and activated T cells. Of note, while all models captured the ability of microRNAs to reduce both the average expression and the CV of microRNA targets, they nevertheless overestimated the actual impact of microRNA-mediated feed-forward regulation. Neither model fully predicted the complexity of the data, specifically the experimentally observed skewing of expression at the top end of the expression spectrum. This indicates that the current models do not fully capture the integration of microRNAs into biological circuits and their impact on gene expression.
TCR signaling drives developmental decisions in thymocytes according to a specific set of rules: too little signal results in a failure to differentiate ('neglect'), too much signal results in activation-induced cell death ('negative selection') [24]. Intermediate signals induce thymocyte differentiation ('positive selection') towards CD4-expressing T helper and CD8-expressing cytotoxic T cells. The nature and strength of signals also directs differentiation towards specialized T cell subsets such as regulatory T cells (Treg cells) and natural killer T cells (NK-T). The functionality of CD4, CD8, Treg and NK-T cells depends on their TCR specificity and it is therefore critical that signal strength and lineage choice are appropriately matched [24, 52]. microRNAs are intimately involved in T cell lineage choices [26, 35, 36–38, 53–57]. The ability to mount predictable responses to extracellular signal is therefore as important for T cell development as it is for other developmental decisions and we speculate that the exploration of microRNA-mediated regulation of cell-to-cell variation in gene expression in other cell types will prove relevant for understanding normal development and disease.
Mouse work was done according the UK Animals (Scientific Procedures) Act under the authority of project licences issued by the Home Office, UK. LckCre Dicer [26], CD4Cre Dicer [35] and mir-181ab1-deficient mice [36] have been described. Fixation and intracellular staining of thymocytes were done as described [39], Antibodies used were RM4-5 (anti-CD4), 53-6.7 (anti-CD8a), 53-5.8 (anti-CD8b), PC61 (anti-CD25), H1.2F3 (anti-CD69), IM7 (anti-CD44), E13-161.7 (anti-Ly6a), AF6-88.5 (anti-H2-K1), and 11-5.2 (anti-H2-Ak; Becton Dickinson) and cells were analysed and sorted on FACS Calibur, LSR II and FACS Aria instruments (Becton Dickinson, Oxford, UK).
Mature CD4+ T cells were activated with anti-CD3 and anti-CD28 for 24 hours, thymocytes were activated with the indicated concentrations of plate-bound T cell receptor beta antibody H57-597 and 2ug/ml of anti CD28 (37.51) for 18 hours.
Dual Fluorescence reporter constructs were based on pMSCVpuro plasmids (Clontech) and contained cDNAs for the fluorescent reporter proteins eGFP under the control of the retroviral LTR and mCherry under the control a separate Pgk promoter, as well as a cloning site in the 3’ UTR of eGFP for the introduction of 3’ UTRs. 3’ UTR fragments were cloned from lymphocyte cDNA and microRNA site mutations introduced by PCR. Retrovirus was produced and activated mature CD4+ T cells or newborn thymocytes were transduced by spin infection as described [58]. Cells were reaggregated with dissociated stromal cells from deoxyguanosine-treated embryonic thymi as described [59], recovered 24 hours later and reporter fluorescence was assessed by flow cytometry. To model the relationship between GFP and mCherry we used orthogonal linear regression, with the relative level of eGFP to mCherry calculated as the slope of the fitted line. These ratios of eGFP expression to cherry expression are normalised to the eGFP/mCherry ratio of the empty vector, to quantify the change in eGFP expression in experimental vectors compared to the empty vector. By comparing eGFP expression from control and Dicer-deficient cells the level of miRNA-dependent repression can also be observed.
RNA was extracted from three biological replicates of Dicerlox/lox and DicerΔ/Δ DP thymocytes, and processed for Affymetrix Mouse Genome 430 2.1 array hybridisation as described [58]. Gene expression array data have been deposited at Geo under accession number GSE57511. Array data were analysed using dChip (http://www.dchip.org). Microarray probe sets were mapped to Refseq transcripts [60]. microRNA sequences were from miRBase [61]. 3' UTR nucleotide motifs were identified using miReduce [27].
Total RNA was isolated using RNAbee (Tel-Test, Friendswood, TX) and reverse transcribed. PCR reactions included 2x SYBR PCR Master Mix (Qiagen), 300nM primers and 2 μl of cDNA as a template in 50μl reaction volume. Cycle conditions were 94°C for 8 min, 40 cycles of 94°C for 30 sec, 55°C for 30 sec, 72°C for 1 min, followed by plate read. All primers amplified specific cDNAs with at least 95% efficiency. Data were normalized to the geometrical average of two housekeeping genes, using the CT method as outlined in the Applied Biosystems protocol for reverse transcriptase-PCR. Primer sequences were (5' to 3'):
Ywhaz fw CGTTGTAGGAGCCCGTAGGTCAT rev TCTGGTTGCGAAGCATTGGG
Ube fw AGGAGGCTGATGAAGGAGCTTGA rev TGGTTTGAATGGATACTCTGCTGGA
Computational modeling of microRNA effects on target gene expression was done as described [8, 49].
|
10.1371/journal.ppat.1003525 | The Smallest Capsid Protein Mediates Binding of the Essential Tegument Protein pp150 to Stabilize DNA-Containing Capsids in Human Cytomegalovirus | Human cytomegalovirus (HCMV) is a ubiquitous herpesvirus that causes birth defects in newborns and life-threatening complications in immunocompromised individuals. Among all human herpesviruses, HCMV contains a much larger dsDNA genome within a similarly-sized capsid compared to the others, and it was proposed to require pp150, a tegument protein only found in cytomegaloviruses, to stabilize its genome-containing capsid. However, little is known about how pp150 interacts with the underlying capsid. Moreover, the smallest capsid protein (SCP), while dispensable in herpes simplex virus type 1, was shown to play essential, yet undefined, role in HCMV infection. Here, by cryo electron microscopy (cryoEM), we determine three-dimensional structures of HCMV capsid (no pp150) and virion (with pp150) at sub-nanometer resolution. Comparison of these two structures reveals that each pp150 tegument density is composed of two helix bundles connected by a long central helix. Correlation between the resolved helices and sequence-based secondary structure prediction maps the tegument density to the N-terminal half of pp150. The structures also show that SCP mediates interactions between the capsid and pp150 at the upper helix bundle of pp150. Consistent with this structural observation, ribozyme inhibition of SCP expression in HCMV-infected cells impairs the formation of DNA-containing viral particles and reduces viral yield by 10,000 fold. By cryoEM reconstruction of the resulting “SCP-deficient” viral particles, we further demonstrate that SCP is required for pp150 functionally binding to the capsid. Together, our structural and biochemical results point to a mechanism whereby SCP recruits pp150 to stabilize genome-containing capsid for the production of infectious HCMV virion.
| Human cytomegalovirus (HCMV) causes birth defects in newborns and life-threatening complications in immunocompromised individuals, such as AIDS patients and organ transplant recipients. The smallest capsid protein (SCP) – only 8 kDa molecular mass as compared to the 155 kDa major capsid protein – has been demonstrated to be essential for HCMV growth, but is dispensable in herpes simplex virus type 1. These seemingly contradictory observations have been a paradox. Here, we solve this paradox by high resolution cryo electron microscopy (cryoEM), in conjunction with functional studies using ribozyme inhibition. Our structural comparisons of HCMV virion and capsid reveal molecular interactions at the secondary structure level and suggest that SCP might contribute to capsid binding of pp150, an essential, cytomegalovirus-specific tegument protein. SCP-deficient particles generated by ribozyme inhibition of SCP-expression in HCMV-infected cells show no pp150 tegument density, demonstrating that SCP is required for the functional binding of pp150 to the capsid. Our results suggest that SCP recruits pp150 to stabilize the HCMV nucleocapsid to enable encapsidation of the genome, which is more densely packaged in HCMV than in other herpesviruses. Overall, this study not only resolves the above paradox, but also illustrates the passive acquisition of a new, essential function by SCP in the production of infectious HCMV virions.
| Human cytomegalovirus (HCMV), the prototype of betaherpesvirus subfamily of the Herpesviridae, is a leading viral cause of birth abnormalities and a major life-threatening pathogen in AIDS and organ transplant patients [1]. HCMV virion shares a common architecture with other herpesviruses and consists of a polymorphic envelope, a tegument compartment and an icosahedral nucleocapsid enclosing a linear dsDNA genome. The HCMV genome is the largest amongst that of all human herpesviruses, and encodes a remarkable number of conserved proteins, as well as unique envelope and tegument proteins that lack homologs in alpha- or gammaherpesviruses [2]. The HCMV capsid shell, similar to those of herpes simplex virus type 1 (HSV-1) and Kaposi's sarcoma-associated herpesvirus (KSHV), is composed of four major proteins: the major capsid protein (MCP; encoded by UL86) [3], the minor capsid protein (mCP; encoded by UL85), the mCP binding protein (mC-BP; encoded by UL46) [4], and the smallest capsid protein (SCP; encoded by UL48.5) [5], [6]. All herpesvirus capsids studied to date share a T = 16 icosahedral assembly with pentons (MCP pentomers), hexons (hexamers of MCP), connecting triplexes (heterotrimers of two mCP and one mC-BP), and SCP attached to the tip of each MCP [7], [8], [9], [10], [11]. While the other three capsid structural proteins are conserved, SCP is very divergent in size, amino acid sequence and function among different herpesviruses. In HCMV, SCP was shown to be essential for virus growth [12], but its function is still unknown.
CryoEM reconstruction also revealed different patterns of association between capsid and overlying tegument proteins in CMV and HSV. In HCMV, a layer of highly organized filamentous density of tegument proteins is attached to the pentons, hexons and triplexes of the underlying nucleocapsid [13]. The three-dimensional (3D) reconstruction of the simian cytomegalovirus (SCMV) capsid isolated from the cytoplasm of infected cells also revealed tegument proteins attached to the capsid [8], similar to HCMV. In contrast, HSV-1 ordered tegument proteins only bind pentons and those triplexes surrounding pentons [14], [15], [16]. These observations indicate that viral proteins overlying the conserved capsid, such as tegument and envelope proteins, have evolved to have virus-specific structural and functional roles. Recently, biochemical and structural studies have assigned pp150 to the ordered filamentous tegument densities of CMV virion and suggested its function in stabilizing the dsDNA-filled C capsid [17], [18], but more structural details are needed to fully understand the molecular interactions between pp150 and capsid proteins.
Here, we report the 3D structures of HCMV capsid and intact HCMV virion at 6 Å and 9 Å resolution, respectively (Figure S1). Comparison of the capsid and virion structures reveals, at a secondary-structure level, that SCP mediates interactions between the capsid and tegument protein pp150. By constructing a ribozyme that inhibits SCP expression in HCMV-infected cells, and cryoEM reconstruction of the resulting “SCP-deficient” viral particles, we further demonstrate that SCP is required for pp150 binding to capsid and its absence results in only viral particles devoid of the DNA genome, thus revealing why SCP is essential for HCMV infection.
Due to the large size of CMV particles and the difficulties to purify them, resolutions of previous cryoEM reconstructions were limited [8], [10], [13], [19]. In this study, highly purified HCMV capsid (no tegument, Figure 1A) and the intact virion (containing tegument proteins, Figure 1B) were obtained and imaged in a 300 kV Titan Krios high-resolution electron microscope. 3D reconstructions of HCMV capsid and virion were obtained at 6 Å and 9 Å resolution, respectively (Figure S1). The improved structures are a result of an exhaustive effort of processing more than 37,000 capsid images and 56,000 virion images. At these sub-nanometer resolutions, secondary structural elements, particularly α-helices, can be identified, as exemplified by the close-up views of a hexon in the capsid reconstruction (Figure 1C–E). Molecular boundaries can be established, allowing us to describe the interactions between the ordered tegument proteins and capsid proteins at secondary-structure level for the first time.
The 6 Å reconstruction of the HCMV capsid reveals the molecular boundaries among the 150 hexons, 12 pentons and 320 triplexes in the T = 16 icosahedral particle (Figure 1C), allowing identification of individual molecules. In particular, the upper domain of an MCP monomer was extracted from a central hexon (Figure 1F) and superimposed with the crystal structure of the HSV-1 MCP upper domain (MCPud) (PDB accession code 1N07) (Figure 1G) [20]. Except for very minor differences at the tip and the outer surface of the subunits (arrow and arrowhead in Figure 1G), excellent match of all the α-helices between HCMV and HSV-1 MCPuds is observed, indicating that the bulk of MCPud structure is conserved between the two viruses. This match also demonstrates the high quality of the map.
In addition, the fitting reveals that SCP molecules were lost in this highly purified capsid sample, probably due to the use of detergent during all purification steps (see Experimental Procedures). In gently prepared capsid preparations, SCP binds MCP at the upper domain as shown in previous capsid reconstructions we obtained [10], [19]. The absence of SCP in the capsid reconstruction provides the advantage to identify the molecular boundary between SCP and MCP in the virion reconstruction (see below).
The 9 Å reconstruction of HCMV virion shows a layer of filamentous tegument proteins bound to the capsid in an icosahedrally ordered fashion, like a net enclosing the entire capsid (Figure 1H). Three of these tegument densities sit on top of each triplex, forming a ‘group-of-three’, and extend to the top of the nearest subunits in the three surrounding capsomers. The location and appearance of these tegument densities are similar to those decorated by anti-pp150-antibodies [17], [18]. In each asymmetric unit of the herpesvirus capsid, there are six quasi-equivalent triplexes, Ta, Tb, Tc, Td, Te and Tf (Figure 1I), following the nomenclature of [21]. The group-of-three tegument densities on triplexes Tb, Td and Te are the most similar in structure. We averaged the densities within three cubes, each of which contains one of these group-of-three tegument densities (e.g., the region encompassing Te is outlined by the dashed square in Figure 1I) to improve the signal/noise ratio (Figure 2A). Helices in the tegument proteins can be resolved in the averaged density (as illustrated in Figure 2E). The same cubic regions from the capsid reconstruction were averaged for comparison (Figure 2B). This comparison allowed us to differentiate densities of MCP and triplexes from densities attributable to SCP and tegument proteins, and subsequently to segment out SCP and the tegument densities. The boundary between SCP and pp150 was established by referring to our pervious SCP-containing capsid reconstruction [19].
The three tegument densities in the averaged group-of-three exhibit a high level of similarity and appear nearly identical when displayed side by side (Figure 2D). This structural similarity among the three tegument densities suggests that three copies of the same tegument protein or protein complex associate with each triplex, which differs from the situation in SCMV, where only two copies of tegument densities were interpreted to bind to each triplex [8]. In each tegument density of HCMV, we resolved two helix bundles, an upper one and a lower one, joined by a long central helix (∼67 Å in length). The upper helix bundle (UHB) is composed of the central helix and five shorter surrounding helices. The lower helix bundle (LHB) only has three short helices surrounding the central helix (Figure 2E).
Previous studies of HCMV and SCMV particles have suggested pp150 as one of the candidates for the capsid-interacting tegument densities [8], [13], [18]. Secondary structure prediction indicates that the C terminal half of pp150 is almost entirely coils, in contrast to the N terminal half, which contains many helices (Figure 2F). Among these predicted helices, the longest one has 47 residues from a.a.195 to a.a.241 (Figure 2F), which would span ∼70 Å as each amino acid in an α-helix gives an axial distance of 1.5 Å. This ∼70 Å length of the longest predicted helix and the measured ∼67 Å length of the central helix resolved in the cryoEM density correlate well with each other, and both are more than twice the length of any other predicted or resolved helices. Moreover, there are eight other predicted major helices with more than 3 helical turns (each turn = 3.6 a.a.) and their lengths (12–22 a.a.) also correlate with those of the eight shorter helices resolved in the tegument density. Therefore, we conclude that the resolved tegument density is contributed only by the N-terminal half of pp150 molecule, while the C-terminal half of pp150 may be disordered or flexible. This conclusion is consistent with previous biochemical data showing that N-terminal segment of SCMV pp150 was both necessary and sufficient to bind either SCMV or HCMV capsid in vitro [22].
We further identified the interface between the tegument density and the capsid. At one end of pp150, its LHB has direct contacts with the triplex (Figure 3A–C). At the other end, pp150 UHB interacts with the capsomer through one of its five short helices (arrowhead in Figure 3D and 3E). This interaction appears to be mediated by the 8 kDa SCP molecule, which is situated in the cleft formed by the pp150 UHB and the upper domain of MCP (Figure 3D–E). The direct contact between the densities assigned to pp150 and SCP suggests direct binding of the two molecules, although at this resolution, one can't rule out the unlikely possibility that binding of SCP to MCPud can in theory change conformation of MCPud, causing it to bind to pp150 directly.
To assess the functional significance of SCP in mediating pp150 binding to the capsid, we constructed a cell line expressing a ribozyme that inhibits the expression of SCP when the cell line is infected by HCMV. Then, we determined the consequence of this inhibition on pp150-binding to the capsid by cryoEM analyses of viral particles harvested from this cell line.
We constructed a ribozyme, called SCP1, by covalently linking the 3′ terminus of a previously established M1GS ribozyme variant (V482) [23] to an 18-nt guide sequence complementary to the targeted HCMV SCP mRNA sequence. Two other ribozymes, SCP2 and TK1, were also designed and used as controls. SCP2 contains the same guide sequence as SCP1 but has multiple point mutations at the catalytic P4 domain that abolish its catalytic activity [24], thus serving as a control for the antisense effect in our experiments. TK1 targets the mRNA of thymidine kinase (TK) of HSV-1 and serves as a control to determine whether M1GS RNA with an incorrect guide sequence could target SCP mRNA in tissue culture. We subsequently constructed cell lines expressing each of these three M1GS ribozymes and carried out the following three experiments.
First, we analyzed SCP mRNA expression in HCMV-infected cells by Northern blotting, using the level of viral immediate-early (IE) 5-kb mRNA as an internal control (Figure 4A). Based on radioactivity of 32P-labeled probes, we estimated that target mRNA expression level was reduced by 98±8%, 7±4%, and 3±3% (average of three experiments) in cells expressing SCP1, SCP2 and TK1, respectively. Furthermore, the protein level of SCP, as determined by Western analyses with the MCP protein level as the internal and loading control, was reduced by 97±9%, 8±5%, and 2±1% in cells expressing SCP1, SCP2, and TK1, respectively (Figure 4C, lanes 9–12). Thus, targeted cleavage of SCP mRNA by ribozyme SCP1 significantly reduced SCP expression in cells expressing SCP1, but not in cell lines expressing both control ribozymes. The low level of inhibition observed in SCP2-expressing cells was probably due to an antisense effect, as SCP2 has a target-binding affinity similar to that of SCP1 but is catalytically inactive.
Second, we assessed the effect of SCP-inhibition in viral yield by measuring viral titers of stocks from HCMV-infected cells that express the ribozymes. At 5 days post-infection, viral yields were reduced by at least 10,000-fold in cells expressing SCP1, whereas no significant reduction was observed in cells expressing SCP2 or TK1 (Figure 4B).
Third, to uncover the structural basis of the reduction of viral yield due to SCP inhibition, we imaged viral particles isolated from SCP1-expressing cells by cryoEM and compared its 3D structure with that of the wild-type HCMV virion. Using MCP as the internal and loading control, Western analyses showed that SCP was hardly detected in HCMV particles isolated from SCP1-expression cells but was readily found in viral particles isolated from cells that did not express any ribozymes or expressed control ribozymes SCP2 or TK1 (Figure 4C, lanes 13–15). CryoEM images of wild-type HCMV virion have the characteristic “fingerprint” appearance (Figure 1B), which is a hallmark of encapsidated genomic dsDNA [14], [25]. In contrast, none of the cryoEM images of particles harvested from the SCP1-expressing cell culture media shows a fingerprint pattern (Figure 5A), indicating that they do not contain viral DNA genome and are thus non-infectious. The existence of non-infectious enveloped particles (NIEPs) in this preparation indicates that the inhibition of SCP expression does not prevent capsid assembly and envelopment. Furthermore, 3D reconstruction at 20 Å resolution of these SCP-deficient particles shows a structure with no visible tegument densities bound to the capsid (Figure 5C). In contrast, reconstruction of wild-type virion at the same resolution clearly shows tegument densities interacting with the underlying capsid (Figure 5B). This result clearly demonstrates that SCP is required for the functional binding of pp150 to capsid. Considering that pp150 may function in stabilizing the dsDNA-filled C capsid [17], [18], we reason that, in the absence of SCP, pp150 can no longer form the stabilizing network of density surrounding the capsid, thus preventing the formation of DNA-containing virion (Fig. 5A). However, we cannot rule out the possibility that the lack of DNA in the SCP-deficient particle and failure to bind pp150 are two unrelated, downstream consequences of lacking SCP. It is also noteworthy that, although the absence of SCP prevents pp150 from binding to the capsid with icosahedral symmetry, it does not necessarily eliminate binding of pp150 to the capsid triplex in a non-icosahedrally ordered fashion, which could have also produced a cryoEM map without visible pp150 densities.
As mentioned above, among all human herpesviruses, HCMV has the largest dsDNA genome contained within a capsid of similar size. As a result, the distance between adjacent dsDNA duplex in HCMV capsid is 23 Å [26], as compared to 26 Å and 25 Å for those in alphaherpesvirus [14] and gammaherpesvirus [27], respectively. It is conceivable that the electrostatic repulsion of the more densely packed genome in HCMV would exert higher pressure to the capsid shell, possibly rendering the DNA-containing capsid (i.e., “C capsid”) unstable. Indeed, throughout our cryoEM imaging of the HCMV capsid preparation, not a single intact DNA-containing capsid was observed among the over 30,000 particle images we examined (Figure 1A), in stark contrast to the situations of alphaherpesvirus [25] and gammaherpesvirus [27] where C capsids can be readily purified from the nuclei of infected cells. Upon tegumentation, including the addition of pp150, DNA-containing nucleocapsids are stabilized and thus are routinely found in HCMV virion preparations (Figure 1B).
Of the capsid structural proteins, SCP is the least conserved across different herpesviruses in size, amino acid sequence, and function. For example, the HSV-1 SCP has a molecular weight of 12 kDa and is dispensable for virus growth in cell culture [28]. The 16 kDa KSHV SCP is the largest and is essential for capsid assembly [29], [30]. The 8 kDa HCMV SCP is the smallest and is essential for virus growth [12], but its functional role is a long-standing mystery. Here, by using SCP-targeting ribozyme and cryoEM reconstruction, we provide the first evidence that SCP is required to stabilize DNA-containing HCMV capsids, and that it may do so by directly or indirectly mediating pp150 binding to the capsid. To our best knowledge, this role is the only function of HCMV SCP identified to date. This result, when considered together with the absence of a pp150 homolog in both alpha- and gammaherpesviruses, indicates that SCP of herpesviruses has diverged in function though its location in different herpesviruses is conserved. Perhaps, SCP has a yet unknown, non-essential function conserved across all herpesviruses, but in HCMV, it is re-utilized by pp150 as a partner to stabilize DNA-containing capsid, an essential process for HCMV infection.
Taken into consideration the relatively small size and essential function, HCMV SCP clearly provides a potential target for intervention against HCMV infection. One possible way is to design SCP-mimicking peptides that act as competitive inhibitors of pp150 binding and functioning, thus preventing infectious viral particle formation.
Human fibroblast MRC-5 cells were cultured in Dulbecco's Modified Eagle Medium (DMEM) plus 10% fetal bovine serum (FBS). 20 flasks (175 cm2) of cells were grown to 90% confluence and infected with HCMV strain AD169 at a multiplicity of infection (MOI) of 0.1–1. At 6 days post infection, when half of the cells were lysed, the media were collected and centrifuged at 10, 000 g for 15 min to remove cell debris. The clarified supernatant was collected and centrifuged at 80, 000 g for 1 hr to pellet HCMV virions. Pellets were resuspended in a total volume of 2 ml phosphate buffered saline (PBS, pH 7.4) and loaded on a 15%–50% (w/v) sucrose density gradient and sedimented at 100,000 g for 1 hr. Usually we observe three light-scattering bands – top, middle and bottom – containing mainly NIEPs, virions and dense bodies, respectively. The middle band was collected and diluted in PBS to a total volume of 13 ml. Virion particles were pelleted again at 80,000 g for 1 hr and resuspended in 30 µl PBS for cryoEM sample preparation.
To obtain HCMV capsids, we infected 90% confluent MRC-5 cells at MOI = 5. At 3 days post infection, when cytopathic effect reached 100%, cells were collected, pelleted by low-speed centrifuge at 1000 g for 10 min, and washed with PBS. The pellet was then resuspended in PBS containing 0.5% NP-40 (w/v) and incubated on ice for 5 min. The mixture was centrifuged at 1000 g for 10 min to pellet cell nuclei. To break nuclear membrane, the pellet was then resuspended in PBS, subjected to three cycles of freezing (−80°C, 10 min), thawing (37°C, 3 min) and vortexing, passed through a 23 gauge hypodermic needle for 20 times, and incubated in PBS with 2% NP-40 overnight at 4°C. The lysate was centrifuged at 1500 g for 5 min to remove large debris and then sedimented through a 30% sucrose cushion at 100,000 g for 1 hr. The pellet was resuspended in PBS containing 2% NP-40, diluted to a final volume of 13 ml, and centrifuged again at 70,000 g for 1 hr. The pelleted capsids were resuspended in 30 µl PBS and used for cryoEM sample preparation.
Plasmids V482, pFL117 and pC102 contain the DNA sequences coding for variant V482 RNA, M1 RNA and mutant C102, respectively, driven by the T7 RNA polymerase promoter [19], [24], [31]. Mutant ribozyme C102 contains several point mutations at the catalytic domain (P4 helix). The DNA sequence coding for ribozyme TK1, which targets the mRNA of thymidine kinase of HSV-1, has been described [31]. The DNA sequence encoding ribozyme SCP1 was constructed by PCR with V482 as the template. The 5′ and 3′ PCR primers were AF25 (5′-GGAATTCTAATACGACTCACTATAG-3′) and M1SCP1 (5′-CCCGCTCGAGAAAAAATGGTGCTGAGCAAGTATACGCGTGTGGAATTGTG-3′), respectively. The DNA sequence coding for ribozyme SCP2 was constructed by introducing into the DNA sequence coding ribozyme SCP1 with the point mutations (A347C348→C347U348 and C353C354C355G356→G353G354A355U356) that were found in C102 and were shown to abolish the ribozyme activity [19], [24], [31]. The procedures for in vitro cleavage and binding analyses were carried out as described previously [24].
The DNA sequences encoding the ribozymes were subcloned into retroviral vector LXSN and placed under the control of the U6 RNA promoter. The retroviral DNA containing the ribozyme sequence was transfected into human U373MG cells, using protocols modified from Miller and Rosman [32]. After 48–72 h of transfection, cells were incubated in culture medium containing 600 µg/ml neomycin. Cells were subsequently selected in the presence of neomycin for 2 weeks, and neomycin-resistant cells were cloned [24].
Cells (n = 1×106) were either mock-infected or infected with HCMV at an MOI of 0.05–5 in 1.5 ml DMEM supplemented with 1% FBS. After 2 h incubation, the inoculum was replaced with DMEM supplemented with 10% (v/v) FBS. The infected cells were incubated for 4–72 h and total cellular RNA or protein was isolated from the cells as described previously [24]. Protein samples were also prepared from HCMV particles purified from the infected cells. The RNA fractions were separated in formaldehyde-containing 1% agarose gels, transferred to a nitrocellulose membrane, hybridized with 32P-radiolabeled RNA probes containing the HCMV sequences, and analyzed with a STORM840 PhosphorImager. The RNA probes used to detect M1GS RNA, HCMV IE 5-kb RNA transcript, and SCP mRNA were synthesized from plasmids pFL117, Cig27, and pSCP, respectively [24], [33]. RNA probes were in vitro synthesized and radiolabeled using an in vitro RNA synthesis kit (Promega, Inc, Madison, IN).
In Western analysis experiments, protein samples were separated on SDS/7.5% polyacrylamide gels cross-linked with N,N″methylenebisacylamide, and then transferred electrically to nitrocellulose membranes. We stained the membranes using the antibodies against HCMV proteins in the presence of a chemiluminescent substrate (GE Healthcare), and analyzed the stained membranes with a STORM840 phosphorimager [24]. Quantitation was performed in the linear range of RNA and protein detection.
Cells (n = 1×105) were infected with HCMV at MOI values specified in the Results section. The cells and medium were harvested at 1-day intervals throughout the 7 days after infection. Viral stocks were prepared by adding an equal volume of 10% (v/v) skim milk, followed by sonication. The titers of the viral stocks were determined by infecting 1×105 human foreskin fibroblasts and counting the number of plaques 10–14 days after infection [24]. The values obtained were averages from triplicate experiments.
Ribozyme SCP1-expressing U373MG cells were infected with wild type HCMV at MOI = 1–5. After 2 hr incubation, the medium was replaced with fresh DMEM plus 10% FBS to remove any free, extracellular viral particles. At 4 days post infection, viral particles were purified using the same procedure as described above for the wild type HCMV virion. Due to the significantly lower viral yield in SCP1-expressing cells, no clear light-scattering bands were visible in the density gradient. We therefore collected the fraction of the gradient corresponding to the range encompassing the three bands visible in the wild type virion purification. This fraction was diluted in PBS and centrifuged at 80,000 g for 1 hr to pellet SCP-deficient viral particles. The pellet was then resuspended in 10 µl PBS, verified by negative staining electron microscopy to contain viral particles, and used for cryoEM sample preparation. The remainder gradient was also collected in fractions of 3 ml each. Each fraction was diluted with PBS, pelleted, resuspended in 10 µl PBS, and checked individually with negative staining electron microscopy to confirm the absence of viral particles.
An aliquot of 2.5 µl purified sample was applied to a 300 mesh Quantifoil R1.2/1.3 grid, blotted with filter paper, and plunge-frozen in liquid ethane. CryoEM images were collected at liquid nitrogen temperature in an FEI Titan Krios cryo electron microscope operated at 300 kV with parallel illumination. The wild type HCMV virion and SCP-deficient HCMV particle images were recorded on a Gatan 4k×4k charge-coupled device (CCD) camera at an effective magnification of 97, 498× (nominal magnification 59, 000× on film plane), corresponding to an effective pixel size of 1.538 Å/pixel at the specimen level. The HCMV capsid images were recorded on Kodak SO-163 films at a magnification of 59,000× and micrographs were digitized with Nikon Coolscan 9000ED scanner at a step size of 6.35 µm/pixel, giving a pixel size of 1.076 Å/pixel on specimen. In all cases, the electron dosage used in cryoEM imaging was ∼25e−/Å2. The defocus values were determined with CTFFIND [34] and are in the range of 0.5 µm to 3 µm underfocus.
Data processing and 3D reconstructions were accomplished with IMIRS [35], [36]. Orientation and center parameters of each particle were refined against projections computed from 3D reconstructions in an iterative procedure until no further improvement in the reconstruction was obtained. Particles were selected based on the phase residues between the images and the projections. 3D reconstruction was obtained using the symmetry-adapted spherical harmonics method [36]. The final capsid and virion reconstructions were obtained by averaging 20,502 particles (selected from 37,460 capsid images) and 11,863 particles (selected from 56,297 virion images), respectively.
Visualization and averaging of density maps were carried out with UCSF Chimera [37]. Density regions to be averaged were segmented out as density cubes of similar size. These density cubes were then first manually aligned and subsequently computationally aligned by the “fit in map” function of Chimera. Averaged density was produced by executing the “vop add” command on the above aligned density cubes.
Secondary structure prediction of pp150 was performed with PSIPRED using the Protein Structure Prediction Server [38].
The cryoEM density maps of the capsid, the virion, and the SCP-deficient particles have been deposited in the Electron Microscopy Data Bank (EMDB) (accession code 5695, 5696 and 5697, respectively).
|
10.1371/journal.pntd.0001707 | Geographical Distribution of Trypanosoma cruzi Genotypes in Venezuela | Chagas disease is an endemic zoonosis native to the Americas and is caused by the kinetoplastid protozoan parasite Trypanosoma cruzi. The parasite is also highly genetically diverse, with six discrete typing units (DTUs) reported TcI – TcVI. These DTUs broadly correlate with several epidemiogical, ecological and pathological features of Chagas disease. In this manuscript we report the most comprehensive evaluation to date of the genetic diversity of T. cruzi in Venezuela. The dataset includes 778 samples collected and genotyped over the last twelve years from multiple hosts and vectors, including nine wild and domestic mammalian host species, and seven species of triatomine bug, as well as from human sources. Most isolates (732) can be assigned to the TcI clade (94.1%); 24 to the TcIV group (3.1%) and 22 to TcIII (2.8%). Importantly, among the 95 isolates genotyped from human disease cases, 79% belonged to TcI - a DTU common in the Americas, however, 21% belonged to TcIV- a little known genotype previously thought to be rare in humans. Furthermore, were able to assign multiple oral Chagas diseases cases to TcI in the area around the capital, Caracas. We discuss our findings in the context of T. cruzi DTU distributions elsewhere in the Americas, and evaluate the impact they have on the future of Chagas disease control in Venezuela.
| Chagas disease is caused by a protozoan parasite called Trypanosoma cruzi. T. cruzi infects a wide variety of mammal species in Latin America as well as man, and is spread by multiple species of blood sucking triatomine insect vectors. The presence of genetic diversity in T. cruzi in the Americas is well established, with six different major genetic types in circulation. The genetic diversity of T. cruzi in Venezuela is relatively poorly understood. In this work we present the results from the genotyping of over seven hundred isolates from 17 of the 24 states. Our dataset comprises strains isolated from wild and domestic animals, several species of triatomine vector, as well as from human Chagas disease cases, including those associated with oral transmission of T. cruzi. Amongst other findings, our data reveal a surprisingly high frequency of atypical genotypes in humans, particularly TcIV, which has rarely been reported. We evaluate our findings in the context of T. cruzi diversity elsewhere in the Americas, and assess the impact they have on the future of Chagas disease control in Venezuela.
| Trypanosoma cruzi, the etiological agent of Chagas disease, infects approximately 8 million people in Latin America [1]. A further 20 million people are at risk of infection. Chagas disease is widely dispersed across 21 countries in the Americas, with a natural distribution (in wild transmission cycles) from the Southern States of the USA [2] to Central Argentina [3]. Chagas disease is a vector-borne zoonosis, and transmission is generally achieved via the infected faeces of various triatomine bug species, evacuated during a blood meal. Infection is maintained in wild transmission cycles by numerous mammalian reservoir hosts, especially opossums (Didelphis sp.) and armadillos (Dasypus sp.) [4]. Human infection occurs at foci throughout the natural distribution of T. cruzi where triatomines have adapted to exploit the domestic setting, but also orally (via ingestion of triatomine contaminated foodstuffs) in endemic countries, as well as via blood transfusion, organ transplantation and congenital infection in and outside of areas of traditional endemicity [1].
T. cruzi is likely to be ancient and indigenous to the Americas [5], [6]. Indeed, the parasite demonstrates considerable genetic diversity as initially revealed by multilocus enzyme electrophoresis (MLEE) [7]–[9]. These early studies supported the typing of the T. cruzi into three main groups or zymodemes, called Z1, Z2 and Z3. The implementation of further molecular techniques in combination with MLEE, allow the division of the T. cruzi species in six groups or discrete typing units (DTU), denoted TcI, TcIIa, TcIIb, TcIIc, TcIId and TcIIe [10]. More recently, in a meeting of experts held in Brazil [11], a new nomenclature was recommended for the intraspecific classification of T. cruzi discrete typing units (DTUs) into TcI, TcII, TcIII, TcIV, TcV and TcVI. However, while the T. cruzi DTUs are relatively genetically stable in space and time, their evolutionary, ecological and epidemiolgical significance is far from clear [4], [12]. Some limited patterns emerge (reviewed in [13]), TcII, TcV and TcVI seem largely restricted to domestic transmission cycles south of the Amazon basin, where they cause considerable human disease. TcIII is infrequent from domestic sources, strongly associated with Dasypus novemcinctus in terrestrial transmission cycles, and found throughout South America. TcIV is enigmatic, so far uncommon among humans, broadly limited to Amazonia and Northern South America and most commonly reported from primates; TcI is the most abundant of all T. cruzi lineages in silvatic transmission cycles, where it primarily infects arboreal marsupials and triatomines in lowland tropical South America and terrestrial rodents and triatomines in arid rocky ecotopes. TcI is the major cause of human disease in northern South America, but also reported from chagasic patients sporadically throughout the Southern Cone.
In Venezuela, previous studies revealed TcI in humans, triatomine bugs, wild and domestic mammals [8], [14]–[21]. T. cruzi genotype TcIV has been reported infecting humans, triatomine bugs and the primate Saimiri sciureus [8], [14]. Infection with TcIII has been found in D. novemcinctus and associated Panstrongylus sp. nymphs [22]. Overall, however, reports of T. cruzi DTUs in Venezuela are focal and fragmented. In the present study we report T. cruzi genotype data from 778 T. cruzi strains systematically genotyped using multiple molecular markers in our laboratory over the past 12 years. These include samples obtained from 17 of Venezuela's 24 states, acute and chronic chagasic patients, seven species of triatomine bugs, nine species of wild and domestic mammals, and representatives of silvatic, peridomestic and domestic cycles. These isolates have been classified to DTU level using biochemical and molecular techniques. Our data represent a uniquely comprehensive record of the six T. cruzi DTUs in Venezuela and a valuable addition to our understanding of the parasite's genetic diversity in South America.
All procedures including use of laboratory reared mice and wild mammals, have been conducted following the regulations for the use of animals in Scientific Research included in the Code of Ethics for Life, of the National Fund for Scientific, Technological and Innovation; Ministery for Science and Technology, with the approval of the Commission on Ethics, Bioethics and Biodiversity, documents N° 7513, 07 Dec. 2007 and N° 4403, 28 Oct. 2010, based on the communication of approval of protocols by the Scientific Ethical Committee of the Institute of Tropical Medicine, Faculty of Medicine, Universidad Central de Venezuela, document N° CEC-IMT 19/2009, 13 Dec. 2005, 20 Jun. 2007, 22 Nov. 2009. In the same way, all studies involving patients and inhabitants at endemic communities, have been conducted according to the regulations for the research in humans, stated in the Code of Ethics for Life, of the National Fund for Scientific, Technological and Innovation, Ministery for Science and Technology, with the approval of the Commission on Ethics, Bioethics and Biodiversity, documents N° 7513, 07 Dec. 2007 and N° 4403, 28 Oct. 2010, based on the communication of approval by the Scientific Ethical Committee of the Institute of Tropical Medicine, Faculty of Medicine, Universidad Central de Venezuela, document N° CEC-IMT 19/2009, 20 Jun. 2007, 13 Dec. 2005, 22 Nov. 2009. All subjects were asked for their voluntary participation in this study by providing a written informed consent under the supervision and approval of the above mentioned Ethical Committees. After being sure that the informed consent was clearly understood by each individual, it was signed by every person, indicating the citizen identification card number (C.I.), in every particular case. The search for triatomine bugs inside the houses, peridomestic and surrounding areas was done with the owners/residents permission. The tests for Chagas disease on domestic mammals were carried out with the owner's permission and the procedure was approved by the Scientific Ethical Committee of the Institute of Tropical Medicine, Faculty of Medicine, Universidad Central de Venezuela, document N° CEC-IMT 19/2009, 20 Jun. 2007, 13 Dec. 2005, 22 Nov. 2009.
T. cruzi isolates were obtained from chagasic outpatients from different geographical areas of Venezuela attending the Instituto de Medicina Tropical (IMT) of the Universidad Central de Venezuela (UCV) as well as chagasic patients living in rural areas of Venezuela where Chagas disease is endemic. Another group of patients were from urban areas of Caracas, the Capital city, and the neighbor State Vargas (see Table 1), all of them in the acute phase of the disease, presumably infected via oral transmission. All patient isolates were collected under informed consent following the ethical permissions of the Research Ethics Commission of the Institute of Tropical Medicine, Faculty of Medicine, Universidad Central de Venezuela. The second group of T. cruzi isolates was obtained from seven different species of triatomine bug and originated from insects brought to the IMT by members of the public and those obtained during fieldwork. Triatomines were identified to species level according to Lent and Wygodzinksy, 1979 [23] and in some case via molecular methods (as part of Fitzpatrick et al., 2008 [24]) as detailed in Table 1. A third group of T. cruzi isolates was found infecting nine species of wild and domestic mammal (see Table 2), captured during multiple field expeditions to endemic and urban regions.
Parasites were isolated via several different techniques. Briefly: parasites from chagasic patients were obtained by indirect xenodiagnosis, by hemoculture of peripheral blood, or by i.p. inoculation of Balb/c mice with peripheral blood. From wild and domestic mammals, parasites were isolated by direct xenodiagnosis, by hemoculture from peripheral or cardiac blood, or by i.p. inoculation of Balb/c mice with cardiac blood. From triatomine bugs, naturally infected or used in the xenodiagnosis, the parasites were isolated by direct culture of feces in blood agar or by i.p. inoculation of Balb/c mice with bug faeces. To achieve xenodiagnosis, we used 12 to 15 instar nymphs of Rhodnius prolixus, 3rd or 4th stage, reared in the laboratory. Initially the parasites were grown in biphasic medium blood-agar followed by culture in supplemented RPMI 1640 medium as described by Miles (1993) [25].
A phenotypic analysis was initially done using the isoenzyme technique as described by Miles et al,. (1977) [7]. For this analysis we used phosphoglucomutase (E.C.2.7.5.1, PGM) and glucose phosphate isomerase (E.C.S.3.l.9, GPI) enzymes (Figure 1). They were examined by thin-layer starch gel electrophoresis as described by Carrasco et. al. (1996) [26]. Random Amplified Polymorphic DNA (RAPD) genotyping (Figure 2) was performed as in Carrasco et al., (1996) [26]. PCR reactions for RAPD typing were achieved using primers A1, A2, L4 and L5 (Table 3). Each reaction took place in a 20 µL final volume containing 10 mM Tris HCl (pH 8.8) buffer, 0.2 Mm of each dNTP, 20 pg of primer, 1.0 unit of Taq DNA polymerase (Invitrogen, Brazil) and included 5 ng of whole genomic DNA. Reaction conditions were as follows: two cycles at 95°C for 5 min, 30°C for 2 min and 72°C for 1 min, 32 cycles at 95°C for 1 min, 40°C for 2 min, and 72°C for 1 min, and a final extension cycle at 72°C for 5 min. Primer sequences are listed in Table 3. PCR restriction fragment length polymorphism (PCR-RFLP) genotyping (Figure 3) targeted two loci: Glucose phosphate isomerase (GPI) and Heat Shock Protein 60 (HSP60) genes were amplified and cut using restriction enzymes HhaI and EcoRV respectively, following protocols set out in Westenberger et al. 2005 [27].
All PCR products were visualised on 2.5% agarose gels (Invitrogen, USA) using appropriate molecular weight markers. Several DTU reference strains were included for comparison and are listed in Table 4.
In total we genotyped 778 isolates to DTU level. Images from selected electrophoretic gels for the various genotyping techniques are displayed in Figures 1–3. T. cruzi isolates genotyped are distributed across 17 endemic states in Venezuela. The genotype analysis of all the isolates shows that 732 belong to TcI group (94.1%); 24 isolates to TcIV group (3.1%) and 22 to TcIII group (2.8%). We recovered and genotyped 95 isolates from humans. Among these, 20 isolates were designated to TcIV (21.0%), with a further 75 typed as TcI (79.0%). Interestingly, TcIV was widely distributed across Venezuela as a secondary agent of human infection. Samples are presented by genotype, state and abundance in Figure 4 and Table 1 and 5. TcIII, although underrepresented in the total dataset (N = 22), was nonetheless conspicuous in its absence from humans. Full details of sample codes and genotypes are included in Tables S1, S2 & S3.
Among nine mammal species identified with T. cruzi infection, TcI was the DTU most frequently encountered (102; 88.7%; Table 2 and 5). TcIII was also apparent (13; 11.3%), however, uniquely among nine banded armadillos (Dasypus novemcinctus). By state, TcIII was found with greatest frequency in Barinas and Anzoategui. These states were also the only ones from which D. novemcinctus was sampled and the presence of TcIII cannot therefore be discounted elsewhere. TcIV was absent from the infected silvatic mammals captured as part of this study, and we are thus far unable to establish the natural reservoir host of this lineage in Venezuela.
Triatomines yielded the vast majority of isolates examined (N = 568). Correspondingly, the greatest diversity of distinct T. cruzi DTUs (TcI, TcIII, TcIV) was also encountered among the triatomines. As previously, TcIII and TcIV represented a minority of the total number of genotypes sampled (2.3%). Among seven species of triatomine, TcIII was only recovered from P. geniculatus. This triatomine bug also yielded a single TcIV strain in Vargas state, close to the Distrito Capital (near Caracas), where TcI and TcIII were also identified in circulation among the same species. Interestingly, TcIV was also present among domestic R. prolixus in Portuguesa, where one insect showed single infection with TcIV and two others presented mixed infection of TcI and TcIV (see Table S2), in this case reflecting the high proportion of human TcIV cases in this state.
The 778 genotype records from humans, mammals and triatomine vectors presented in this study dramatically expand our understanding of the geographical distribution of T. cruzi genotypes in Venezuela.
Perhaps most significant is the frequent occurrence of TcIV among human Chagas disease cases in Venezuela. So far reports of this DTU in humans are sparse. They include at most half a dozen cases across Northern Brazil [8], [28], as well as some historical cases from Venezuela [8]. However, given the continuity of the ecotopes and major vector distributions (e.g. R. prolixus in Venezuela) in the areas from which these cases originate, especially the lowland Llanos region which lies between Venezuela and Colombia, we suspect that the distribution of human TcIV cases is likely underreported. We characterised TcIV from domestic R. prolixus at one study site in Portuguesa, and from domestic P. geniculatus in Vargas State (Table 1, Figure 4). As with TcI, therefore, this DTU may be actively maintained in domestic cycles in Venezuela. The risk of epizootic transmission events cannot be defined until the silvatic abundance and niche of TcIV can be established. In Brazil, and Bolivia, silvatic TcIV has been isolated primarily from primates and Rhodnius species triatomines [14], [28]. There are also limited records of this genotype from Panstrongylus species and the coati Nasua nasua [28]–[30]. In theory, TcIV should also be primarily located in arboreal cycles in Venezuela, associated with primates and Rhodnius species. Indeed, there is a single TcIV record from the squirrel monkey Samairi scuireus in Venezuela [14]. Targeted capture efforts should improve our understanding of enzootic TcIV in Venezuela, as well as help identify whether it shares the same risk factors for epizootic transmission as TcI (e.g. [15], [31])
T. cruzi is an extremely successful parasite. Evidence to support this assertion lies in its continental distribution and the sheer variety of reservoir hosts it naturally infects. TcI in Venezuela is perhaps typical of this success, with nine different species infected, including highly atypical hosts like the collared peccary (Dicotyles tajacu), and white tailed deer (Odocoileus virginianus). The epidemiological importance of atypical infections is debatable, either in terms of maintaining wild parasite transmission, or in representing a risk to human populations. Of critical relevance to human transmission in Venezuela are ecotopes dominated by palms (e.g. Attalea sp.), R prolixus vectors, and Didelphis marsupialis reservoir hosts. As ever, we isolated the great majority of wild TcI from D. marsupialis, well known as the primary host of this genotype [4] and for its tendency to aggregate around human communities [32]. Wild R. prolixus readily invades houses [24], establishing domestic colonies and propagating disease among rural communities. Risk factors for transmission are well established [31], and control strategies can be designed to maximise successful interruption of transmission.
TcIII, by comparison with TcI, is a less promiscuous DTU. In common with other studies through South America [22], [33], we isolated this genotype almost exclusively from D. novemcinctus and its associated triatomine vector P. geniculatus. In non-human cases, we isolated TcIII with similar global frequency to TcIV (Table 1 and 5). By comparison, no TcIII infection was observed in man, while TcIV was common. Nonetheless, we did find a TcIII infected P. geniculatus as a primary domestic disease vector at a peri-urban focus and it seems remarkable no TcIII was isolated from man. Similarly, TcIII is largely absent from humans throughout the rest South America, with only one confirmed report [34]. Together, these data suggest that the restricted host range of TcIII may be related to more than just transmission ecology. Detailed genetic, biochemical and biological characterisation of experimental in vitro and in vivo infections could shed light on more fundamental constraints on TcIII infectivity.
TcI transmission across Latin America is widespread [13], [35]. Several vector - reservoir host – ecological niche cliques are relevant in terms of human disease. Transmission around Caracas is an important example of the emerging importance of peri-urban transmission in the impoverished districts of several Latin American cities [17]. In Caracas rodents (Rattus rattus) are the primary synantropic host and P. geniculatus the vector (Table 1). Similarly, TcI transmission is maintained by murid rodents (although via Triatoma species vectors), in hyper-endemic arid sub-Andean valleys that impinge on the city of Cochabamba [36]. Peri-urban transmission in Arequipa, Peru, accounts for high levels of seropositivity even among children [37]. In this case, however, relevant reservoir host and vectors are less well characterised. Nonetheless, in Venezuela and elsewhere, control of disease transmission in an urban environment represents a very different challenge to that at rural foci. National authorities could benefit from the sharing of experience in relation to peri-urban Chagas disease control.
The great majority of human isolates from Caracas characterised in this manuscript originate from several oral Chagas disease outbreaks in the city. The largest oral outbreak so far recorded in the city occurred at a school in 2007 [21], [38]. It is thought that over 1000 were exposed, mostly children, among whom 103 developed infection and one died. Classic epidemiological approaches indentified a contaminated batch of Guava juice as the likely source, and three isolates typed from patients and nearby triatomines were TcI. Data presented here do not include the genotypes of isolates from the 2007 outbreak. Nevertheless, we have included isolates obtained from P. geniculatus and R. rattus from the site where the Guava juice was prepared, which were also TcI. In addition, we did include data from two further outbreaks, and all human genotypes also correspond to TcI. The existence of these strains, and accompanying non-human isolates from the same sites, opens the door to high resolution molecular epidemiological work to pinpoint the actual source of these oral cases. Rigorous molecular epidemiological studies can complement and enhance control recommendations for oral disease outbreaks, which are currently limited to food hygiene measures [38], to help prevent future outbreaks and perhaps shed light on the elevated case mortality rates associated [1].
Several anecdotal reports exist to suggest that human Chagas disease mega-syndromes are more common in the Southern Cone region of Latin America [12], [13]. This aspect of differential disease presentation between northern and southern South America is often circumstantially attributed to the presence of TcII, TcV and TcVI in the south [13]. Consistent with several current and historical studies, however, we observed severe cardiac forms of disease in Venezuela among TcI cases (Nessi et al., manuscript in preparation). To date, however, we have not detected digestive forms of the disease. Using high resolution microsatellite markers, we recently demonstrated a substantial reduction in genetic diversity among 15 TcI isolates from humans in Venezuela, by comparison to their silvatic (wild) counterparts [19]. Our analysis indicated that most human infections originate from the same genetically depauperate clade, while incursion of strains from the local silvatic environment was a far rarer event. The remaining 60 human TcI isolates that are uncharacterised by high resolution markers not only offer considerable scope to test the robustness of the TcI human clade, but, in conjunction with clinical history, may also allow us to test the strength of association between TcI sub-DTU level diversity and disease presentation. Importantly, we can confirm that a number of human TcIV cases in this study were symptomatic (Nessi et al., manuscript in preparation) and this DTU can be considered an epidemiologically important secondary agent of Chagas disease. High resolution analyses of TcIV isolates from human cases promise to reveal whether these isolates also represent a genetically restricted clade.
Chagas disease is potentially re-emergent in Venezuela [39]. The data presented in this manuscript are especially important to understanding the eco-epidemiology of infection locally as well as in the context on renewed efforts to interrupt transmission in rural and urban settings. Vitally, they also lay the groundwork for future, hypothesis driven research aimed at discovering the epidemiological/biological relevance of genetic diversity within the T. cruzi DTUs. For example, it is now technically possible to identify the sources of emergent peri-urban and oral transmission. Also, in conjunction with detailed longitudinal clinical data it may be possible to investigate the impact parasite genetic diversity has on the outcome of human disease.
|
10.1371/journal.pcbi.1002689 | Spike-Timing Dependence of Structural Plasticity Explains Cooperative Synapse Formation in the Neocortex | Structural plasticity governs the long-term development of synaptic connections in the neocortex. While the underlying processes at the synapses are not fully understood, there is strong evidence that a process of random, independent formation and pruning of excitatory synapses can be ruled out. Instead, there must be some cooperation between the synaptic contacts connecting a single pre- and postsynaptic neuron pair. So far, the mechanism of cooperation is not known. Here we demonstrate that local correlation detection at the postsynaptic dendritic spine suffices to explain the synaptic cooperation effect, without assuming any hypothetical direct interaction pathway between the synaptic contacts. Candidate biomolecular mechanisms for dendritic correlation detection have been identified previously, as well as for structural plasticity based thereon. By analyzing and fitting of a simple model, we show that spike-timing correlation dependent structural plasticity, without additional mechanisms of cross-synapse interaction, can reproduce the experimentally observed distributions of numbers of synaptic contacts between pairs of neurons in the neocortex. Furthermore, the model yields a first explanation for the existence of both transient and persistent dendritic spines and allows to make predictions for future experiments.
| Structural plasticity has been observed even in the adult mammalian neocortex – in seemingly static neuronal circuits structural remodeling is continuously at work. Still, it has been shown that the connection patterns between pairs of neurons are not random. In contrast, there is evidence that the synaptic contacts between a pair of neurons cooperate: several experimental studies report either zero or about 3–6 synapses between neuron pairs. The mechanism by which the synapses cooperate, however, has not yet been identified. Here we propose a model for structural plasticity that relies on local processes at the dendritic spine. We combine and extend the previous models and determine the equilibrium probability distribution of synaptic contact numbers of the model. By optimizing the parameters numerically for each of three reference datasets, we obtain equilibrium contact number distributions that fit the references very well. We conclude that the local dendritic mechanisms that we assume suffice to explain the cooperative synapse formation in the neocortex.
| The structure of neocortical networks of neurons changes in time: new synapses are formed, maturate, and eventually are pruned again, in the adult as well as in the developing animal [1], [2], for recent reviews see [3], [4], [5]. The majority (about ) of excitatory synaptic contacts terminate on dendritic spines [6], and dendritic spines almost always () form a synapse [7]. The synapses on dendritic spines are highly dynamic [8], [9], for example [10] found an average spine turnover of in primary visual cortex and of in somatosensory cortex. Yet in the adult animal, the statistics of the numbers of synapses are preserved over time, indicating that synapse creation and pruning balance each other [11], [12], [13]. According to theoretical studies on associative networks, structural plasticity enhances the memory capacity of a network substantially [14], [15], and has been shown to be related to motor learning in the brain [16].
The three studies [17], [18], [19] reported the distributions of numbers of synaptic contacts for different intra-cortical synapses in rat somatosensory cortex. Fares et al. [20] subsequently analyzed whether the reported distributions could result from random and independent synaptic contact formation, given a set of potential sites (close appositions) between axons and dendrites of reconstructed cells. As they showed, independent formation of synaptic contacts alone cannot explain the distributions. In addition a cooperative pruning mechanism, by which synaptic contacts that constitute a single synapse stabilize each other, is required to explain the observed distributions.
Here we build on grounds of this work and go beyond it in two aspects: Primarily, we consider synaptic processes that operate continuously in time. Secondly, we investigate an explicit candidate mechanism for the cooperation between synaptic contacts: Local correlation detection at the dendritic spines and thus dependent pruning and maturation of spines.
Recently Kasai et al. [21] summarized known properties of the plasticity of dendritic spines. Their model [22] describes the dynamics of the volume of dendritic spines. Here we restrict this model to three distinct categories of synapse states and introduce an explicit spike-timing dependence. Other models of structural plasticity [23], [24], [25] are based on the firing rate of the neurons. Consequently, in these models spike-timing and correlations of the spiking activity do not play a role, so they cannot show the mechanism of synaptic cooperation that we hypothesize here. The relative timing of pre- and postsynaptic activity indeed influences structural plasticity at the dendritic spine [26]. In contrast to previous models, the model of Helias et al. [27] is sensitive to the spike-timing of the pre- and the postsynaptic cell and describes structural plasticity in biophysical terms of protein kinetics in response to synaptic input. Here we choose an intermediate scale by still describing single synaptic contacts, but with a higher level of abstraction than previous work [22], [27]. The goal of the present work is to demonstrate the potential of local correlation detection at the spine, while making minimal assumptions about the involved biophysical processes. The assumptions entering our model, as introduced in detail in Methods, required to qualitatively explain the experimental results, are: a) presynaptic release of glutamate causes postsynaptic depolarization at excitatory synapses, b) depolarization electrically spreads within the dendrite, c) there is a correlation sensing mechanism sensitive to the relative time of presynaptic and postsynaptic firing (e.g. NMDA receptors) that causes downstream effects on the evoked synaptic amplitude in a spike timing dependent plasticity (STDP, [28]) like manner, d) synapses with small amplitude are more likely to be pruned than strong ones. Because of its analytical tractability, we can compute the steady state of our model and match its parameters to experimental reference data, analogous to Fares et al. [20]. Our results show that no direct signaling between synaptic contacts is necessary to explain cooperative synapse formation. In contrast, it suffices that distinct synaptic contacts cooperate in exciting the postsynaptic neuron, and thereby indirectly affect spike-timing dependent structural plasticity at other synaptic contacts.
In this section we introduce a model of structural plasticity and describe the optimization procedure to fit the model to the experimental reference data.
Let us first introduce a simple model for the correlation detection at the postsynaptic dendritic spine. An action potential of the postsynaptic neuron causes a depolarization at the site of each dendritic spine. The spine has the biophysical substrate to maintain a signal that depends on the time of the action potential in relation to the time when a presynaptic impulse arrived [27]. Here, we call this signal the correlation trace and assume a phenomenological model: if the presynaptic neuron spiked shortly before the postsynaptic one, the correlation trace is increased by . We call this a causal event. For the opposite relative timing, called an anti-causal event, the trace is decreased by . The trace therefore counts causal and anti-causal combinations of pre- and postsynaptic spikes. Further we assume that the correlation trace is forgetful: it has a leak with the time constant , and we also assume that there is some additive noise in the process. The dynamics of the correlation trace at the synapse is given by the stochastic differential equation [29](1)where is the spike train of the postsynaptic neuron, with the spike times , the factor specifies if the particular spike is counted as a causal or anti-causal event, and is an additional white noise with mean zero and infinitesimal variance . Mathematically, the trace is identical to a shot noise [30] with the exponential kernel driven by the stochastic input process .
Let us now introduce a minimal model of correlated spiking of the neurons. For each postsynaptic spike, we speak of a causal event at a given synaptic contact if the closest spike of the presynaptic neuron occurred prior to the postsynaptic one (because it could have caused the postsynaptic spike). If the closest presynaptic spike occurred after the postsynaptic one, the event is called anti-causal. Suppose that the probability for a causal event is given by . If both of the neurons fire independently, then . We define with probability and with probability for each postsynaptic spike in .
Strictly speaking, the process defined by (1) is unphysical, since through it depends on events in the near future (because the time of the next presynaptic spike has to be known). We hence consider (1) as an effective, adiabatic description of the correlation trace, since we are only interested in the statistics of the trace on long timescales. A process like (1) could result from several biophysical implementations that do in fact respect causality. For example the synaptic weights in phenomenological models of spike timing dependent plasticity, for which causal implementation are known [31] follow dynamics similar to (1). An example of a cellular mechanism to implement (1) is the number of activated CaMKII macro-molecules [27] or long-term potentiation [32], [33], [34].
Now let us further assume that postsynaptic spikes occur according to a Poisson point process with rate . The firing rate comes about through integration of thousands of synaptic inputs, and the particular synaptic connection modeled here only provides a small contribution to . Since structural plasticity is known to be a slow process compared to the activity of neurons, and since the time constant of possible candidate mechanisms for the correlation trace can be considerable [35], a large integration time constant is reasonable, such that . Then the equilibrium probability distribution of is a normal distribution with mean and variance ,(2)(3)Eqs. (2, 3) can be obtained by considering two independent stochastic processes and with and . Then and mean and variance of follow from summing the respective statistics of and , which can be obtained using standard techniques [30]. Note that is independent of .
The probability of causal spike pairings depends on the number of active synapses connecting the presynaptic neuron to the postsynaptic one, because each excitatory synapse increases the chance of the presynaptic neuron to make the postsynaptic neuron fire. As demonstrated for integrate-and-fire neurons in [36], the probability of a spiking response to a presynaptic spike is proportional to the synaptic weight of the input spike for a wide range of magnitudes of the synaptic strength. If the membrane potential of the postsynaptic neuron integrates the inputs linearly, the synaptic weight of the input from the presynaptic neuron is proportional to the number of active synaptic contacts between the neurons. So the probability of a spike response, and so of a causal event, rises proportionally to the number of active synaptic contacts. Effectively we thus assume , where is the number of active synaptic contacts between the presynaptic to the postsynaptic neuron, is the response probability per synapse, and . The two components of the probability can be interpreted as the probability for a causal event by chance due to the Poisson firing of the postsynaptic neuron with rate and the probability exceeding chance level triggered by the arrival of the presynaptic spike.
In order to obtain an estimate of consider a single synaptic contact between two neurons that, upon activation, causes an excitatory postsynaptic potential (EPSP) with amplitude . For a leaky integrate-and-fire model neuron in the asynchronous state [37] resembling cortical activity, we can read off the response probability of a neuron to such a voltage jump from Fig. 4D in [36]. So we have with . The respective values of the EPSP size per contact have been reported along with the reference datasets in [17], [18], [19], and we list them among the model parameters in Tab. 1. We thus arrive at the estimate for the probability of a correlated pairing(4)and thus with (2) the mean of the stationary distribution of is(5)Note that the linear model (4) for the probability of causal spike pairings may for large and yield values of , which are nonsensical. A consistent definition of should saturate when reaching the value . Taking into account this saturation at large , however, would not make a difference for the models considered here, because all solutions for synapse distributions found below exhibit vanishing probably throughout at such large values of .
So far, we hypothesized a generic correlation detection mechanism at each synaptic contact and computed its equilibrium statistics (5) and (3) for multiple excitatory contacts between two neurons. In our model the values of the correlation trace of each synaptic contact follow a normal distribution, specified by the its mean and variance , which depend on the parameters . Note that only the mean depends on the number of active contacts, whereas is a constant. To reduce the amount of free parameters of the model, we further set which is a reasonable choice for neocortical neurons.
The synaptic correlation trace can guide structural plasticity. Because of a lack of detailed knowledge about the biomolecular mechanisms involved [38] we again employ a simple effective model. As structural plasticity is a slow process compared to the spiking activity of neurons, we assume that the distribution of the correlation trace at each synaptic contact is effectively in stochastic equilibrium throughout. This is also known as an adiabatic approximation. Let us now assume that a structural change of the synaptic contact is initiated when the correlation trace crosses a boundary value . A biochemical mechanism underlying this assumption could be the activation of a signaling pathway when a specific number of activated CaMKII molecules is reached [39]. Given the random trajectory of the correlation trace , we need to know how long it takes until crosses the boundary upon which the synaptic contact makes the transition. This is known as a first passage time problem with an absorbing boundary. The inverse of the mean first passage time is called the escape rate. For simplicity, let us approximate as a Brownian motion with the same infinitesimal mean and variance as the actual process (1). Then, according to the Arrhenius approximation [40], the escape rate is(6)in the case that the values of are far from the boundary , such that . Here the proportionality constant is called the Arrhenius constant. Now what happens to the rate of structural changes when approaches ? If we take the model of a boundary crossing process seriously, then the escape rate should diverge as . However, in a biological system it is more plausible that the rate of structural changes converges to a certain maximum rate which the cellular machinery can achieve. Based on this argument we construct our model for the rate of structural changes by extrapolating from the Arrhenius approximation (6), forcing it to eventually converge to a plateau,(7)with . Depending on the sign of , either approaches or departs from the plateau for increasing .
As our assumptions about the biophysical implementation are quite general, they can model maturation, shrinkage and pruning of a synaptic contact alike. However, these are distinct processes that take place during different stages in the life cycle of a synaptic contact. For example, we cannot assume that the correlation detector noise has the same magnitude for small and large dendritic spines since the number of channels mediating the signal might be different for the two. Therefore we use the model (7) for maturation, shrinkage and pruning, but choose a different set of transition parameters and correlation trace noise for each case. We decorate quantities associated to maturation with , those associated to shrinkage with and pruning with . So the rate of maturation transitions is defined as(8)the rate of shrinkage transitions as(9)and the rate of pruning as(10)The correlation trace parameters are assumed to be identical for both thin spines (inactive synaptic contacts) and large spines (active synaptic contacts), which will be defined below. The pruning rate uses the same parameters as shrinkage, except for the noise magnitude of maturation, because pruning is assumed to take place in thin spines.
Apart from the activity dependent transitions, the model also includes intrinsic fluctuations as in [22], see Fig. 1B. We assume that random maturation (enlargement), shrinkage and pruning of a spine occurs constantly with the rate , and the creation of new thin spines with rate .
Let us summarize the structural plasticity model we have introduced. At each of the synaptic contacts between a pair of neurons, a correlation trace is formed by counting causal and anti-causal pre-post spike pairings. The distribution of the values of the correlation trace depends on the number of active synaptic contacts since they all contribute to firing the postsynaptic neuron. We have further assumed that the activity-dependent structural changes of synaptic contacts depend on the correlation traces. Finally we also included intrinsic fluctuations of the synapse configuration.
Above we defined a model for structural plasticity for the synapse between a pre- and a postsynaptic neuron. Although in this model the individual synaptic contacts may continuously change, the state of the synapse develops towards a stable steady state. A synapse typically consists of many individual synaptic contacts, as depicted in Fig. 1A. The neocortex is densely packed with only very limited unoccupied extracellular space. Accordingly, pairs of neurons cannot form arbitrary numbers of synaptic contacts [41]. Fares et al. [20] investigated reconstructed cortical tissue and counted the numbers of close appositions between pairs of neurons. At such a close apposition a synaptic contact may form, but is not necessarily present. Describing these results statistically, a probability distribution for the number of close appositions between two neurons can be obtained [20].
At each of the close appositions, the neurons may form a synaptic contact; in our model, we treat the different volumes of spines and EPSP amplitudes in a coarse-grained fashion, distinguishing just three different states for each contact to occupy – active, inactive, or unrealized. An active synaptic contact here describes a larger dendritic spine that contains both AMPA and NMDA receptors. An inactive contact models a thin, either newly formed or recently shrunk dendritic spine that has much less AMPA receptors [42], [43] and contributes little to firing the postsynaptic neuron. An unrealized contact, finally, is a close apposition where no contact has formed, but might be formed in the future. It is a close apposition without an established synaptic contact and corresponds to the potential synapse in [20]. A similar model has previously been proposed in the context of associative networks [44], [45].
We denote the numbers of synaptic contacts in these three states by , and respectively. Since at any time , the state of a synapse is unambiguously defined by the combination of the number of active and inactive contacts . Now consider an ensemble of independent synapses, each with the maximum contact number . The probability of a synapse to be in the state evolves in time according to the Master equation [46](11)The first term sums up the rate of leaving the state by all possible transitions. The second and third terms sum up all possibilities to go into state from other states by shrinkage and pruning, and the fourth term by maturation. The last term considers the transitions due to the creation of inactive synapses, the rate of which is given by (see also Fig. 1B). The steady state distribution does not depend on the time scale of the transition rates, so we can consider the constants in units of . The time scale of the structural plasticity is then set by . In (25) below we will see how the value of can be determined by experimental data.
To determine the steady state configuration of the synapse, let us introduce a numbering of all the possible synaptic states , such that the probability of each state is represented by the value , with a one to one correspondence between indices and states . Then (11) can be written as(12)where the entries of the matrix can be read off the Master equation. Since describes a Markov process it is column-stochastic (which means all columns sum up to zero). Since the process is irreducible, according to the Perron-Frobenius theorem there is only one stationary solution. We can determine the stationary probability distribution by solving under the constraint that . We implemented the construction of the matrix efficiently using Cython [47] and solved for the stationary solution using Scientific Python [48].
The stationary distribution depends on the number of close appositions . [20] have estimated the distribution for the three types of intra-cortical connections that we consider. We incorporate this by determining for each separately, and subsequently compute the averaged distribution(13)Fares et al. [20] provide the distribution for up to .
For comparison with the reference datasets (see below), we are merely interested in the marginal probability of a certain total number of synaptic contacts, disregarding whether they are active or inactive. The marginalization can be obtained from by summing over all states with , or more conveniently phrased as(14)where the function returns the value of of the state with index , and equals if is true and otherwise. Analogously, the marginal average distributions of the number of active and inactive synapses are(15)(16)
In the experimental studies [17], [18], [19] a set of occurrence frequencies of numbers of synaptic contacts for several pairs of neurons was obtained, each for three different types of intra-cortical projections. Complementing this, for the same three datasets, the probability of a pair of neurons to be connected with at least one active contact can be estimated [20]. The distribution of the numbers of active synaptic contacts serve as reference data in our study. For each of the three datasets, we transform the reported data to the probability mass function(17) we evaluate (14) and obtain the residuals(18)for . The residuals are scaled by the maximum of the reference distribution to enable comparison of the quality of the fits across reference datasets. We minimize the sum of squared residuals(19)using the Levenberg–Marquardt algorithm, applying its implementation from Scientific Python [48]. We call the error of the model. The optimization problem has several local minima, so we initialize the optimization procedure at many points in the -dimensional parameter space and compare the values of to which the optimization converged. Specifically, we choose four different initial values in each parameter dimension, which makes a total of distinct optimization runs per reference dataset. The parameter sets which resulted in a minimal value of are shown in Tab. 1, along with additional information on the model, and the resulting equilibrium distributions are shown in Fig. 2.
We also obtained parameter sets which yield good fits to two reference distributions simultaneously. The resulting distributions for the connections L4-L4 and L5-L5 are displayed in Fig. 3. Here the fit error was defined as the sum of the errors (19) of both distributions, . With respect to this the same optimization procedure was performed.
Here we compute the average lifetime of an inactive synaptic contact and of an active contact in the equilibrium state of the synapse model, see Fig. 1a for the possible transitions. We define the lifetime as the expected time until the contact is pruned. Consider an active contact in a synapse. It may become an inactive contact either through intrinsic or activity dependent shrinkage. The mean time up to the transition from active to inactive is . An inactive contact, on the other hand, might make a transition to the active state (maturation), which would take the time , or to the unrealized state (pruning) in the time . The mean time until the first transition, either maturation or pruning, is . Either of the two transitions happens with a probability given by the fraction of rates involved, and analogously . If the inactive contact becomes active, then it will become inactive eventually, and subsequently might be pruned or become active again. Accounting for the possible paths the inactive contact may take upon its first transition we obtain the expected lifetime of the inactive contact as(20)where is the lifetime of an active contact in a synapse that has active contacts. In turn, starting from an active contact just adds one active to inactive transition, so(21)Inserting (21) and the definitions above into (20) yields(22)We average the lifetimes across the equilibrium probability distribution of synapse states and obtain(23)(24)
To match the time scale of structural development of our model to what is known from in-vivo studies we compute the spine turnover ratio as it is defined in [10],(25)Here and are the numbers of gained and lost spines during a given period of time, and is the number of spines observed. In our model, the expectation values of these quantities are given aswhere and are given in units of . So we obtain in units of from (25). In rat somatosensory cortex [10] found . Accordingly (25) sets the time scale of the model to .
In this section we consider a simplified version of our model which does not include inactive synaptic contacts (thin spines). In that model, at a close apposition there can be either no synaptic contact or an active synaptic contact. Between these two states transitions are allowed just as between the inactive and the active state in the full model (see Fig. 1B), but here we call them (creation) and (pruning), which may yet be arbitrary functions. Assume there are close apposition between a pair of neurons. Then the state of the synapse is defined by the number of active connections . Let us denote the probability of the state by here. In stochastic equilibrium the probability fluxes into and out of the state must balance, so for it must hold thatfrom which follows that(26)Now consider the case of . In all three reference datasets, , as can be seen in Fig. 2a1–a3. According to (26) this requires , which can only be achieved by or . In contrast, around the secondary peak of the reference distributions , (26) entails that and must be of comparable magnitude. More specifically, the right hand side of expression (26) has to change from values larger than to values smaller than as passes the secondary peak from below. To satisfy these requirements, even only approximately, demands highly non-monotonous choices of the functions and that are difficult to justify biophysically. Once the model includes the intermediate state of the inactive synaptic contact, however, it is possible to find biologically plausible parameter sets to explain the reference distributions, as described in the rest of the paper.
Kasai et al. [22] monitored the temporal evolution of the volume of dendritic spines and described it as a random walk with volume dependent drift and diffusion components. According to their findings, newly formed dendritic spines are small, and accumulate AMPA receptors as the spine volume increases. Thus a small spine can grow or disappear, and a large spine can shrink. Spines of all volumes, however, were found to contain NMDA receptors [42]. The study by Holtmaat et al. [10] suggests that thin spines are more readily pruned than thick spines, and that they may be of a lower efficacy or NMDA receptor-only (inactive) synapses. It was previously suggested 43,49 that small spines might correspond to silent synaptic contacts. In our model, we distinguish between only three states that each synaptic contact can occupy: active, inactive and unrealized, without considering the spine volume and channel density of the dendritic spines explicitly. These states correspond, respectively, to large spines, thin spines and close appositions with no spine, as illustrated in Fig. 1b. Note that we do not claim that the functional distinction between thin and large spines is actually as clear-cut as assumed in the model – the model merely represents a coarse-grained spine state. In this model, transitions between the three morphological/functional states are possible. Following [22] and [50], such transitions can occur either due to intrinsic fluctuations, or depending on the activity of the pre- and postsynaptic neuron. Note also that, although the inactive synaptic contact is allowed as a transitional state, it turns out to rarely occur in the optimized models that will be discussed below.
The basic idea of the model put forward in this study is the following: As described in [27] dendritic spines have the biomolecular capability of detecting correlations in the relative spike timing of the pre- and postsynaptic neuron. If there are several active excitatory synaptic contacts from a presynaptic neuron to a single postsynaptic cell, all these synaptic contacts contribute to elicit spikes in the postsynaptic neuron. Hence each of the contacts increases the correlation between the two cells, measurable at each of the corresponding dendritic spines. So even if there is no direct communication between the synaptic contacts, they affect each other indirectly by increasing the correlation of pre- and postsynaptic spikes. Spike-timing dependence of structural plasticity is thus a candidate mechanism for the cooperation between synaptic contacts.
According to [27] the magnitude of calcium influx into the dendritic spine depends on the proximity of pre- and postsynaptic spikes in time. The calcium influx activates or deactivates CaMKII macro-molecules and thus leaves a local memory. We call such a memory of the spike-timing correlation a correlation trace. The activation of the CaMKII subunits can be preserved for a long time [51]. The model we consider here, however, does not rely on the biophysical details of CaMKII activation, but just assumes a correlation trace is available. For the purpose of this study, the correlation trace could also come about by other mechanisms.
Employing this correlation trace, we introduce a phenomenological model for activity dependent maturation, shrinkage and pruning of spines depending on the correlation of the spike-timing of pre- and postsynaptic cell, as described in detail in Methods. The model is based on [22] and incorporates the basic properties of structural plasticity [21], activity-independent creation and pruning of spines, intrinsic fluctuations of spine volume, and activity-dependent spine remodeling. The set of all synaptic contacts connecting a given pair of neurons constitute a synapse, see also Fig. 1a. The state of a synapse is defined by the number of active contacts (large spines) and inactive contacts (thin spines) . The time-evolution of the synapse state is then described as a Markov process. For a given parameter set, we solve for the stationary probability distribution of the states .
The parameters of the model were then optimized so that the distribution of the total number of synaptic contacts reproduces the experimental reference data, shown in Fig. 2a1–c1, along with the respective transition rates of the model (6) in a2–c2. For each of the three reference datasets (a, b, c), we show the best model that resulted from the optimization. The models can reproduce the experimental distributions of synapse numbers.
The existence of such a stationary distribution means that the average numbers of inactive, active and potential sites of the synapse do not change in time. This is so despite the constant creation and pruning of synaptic contacts since these processes compensate each other in equilibrium. Implicitly, the model allows that inactive, active and potential sites coexist between a given pair of neurons.
The parameter sets for the displayed models are given in the second section of Tab. 1. The fit of the connection L4–L23 takes very different parameter values than the others. Nonetheless for all three modeled connections, the time constant of the correlation trace is large compared to the time scale of fluctuations of neuronal activity, in agreement with our assumption about the distribution of the correlation trace. Concerning the parameters of the activity dependent structural plasticity, we find qualitatively similar results across datasets: In all three cases, both maturation and shrinkage/pruning rates decrease with increasing active synapse number, granting long-term stability to established synapses.
The models of the intralaminar connections L4-L4 and L5-L5 show remarkable similarities. Both have a comparable and the rate of intrinsic, activity independent transitions is low, although this was not an a priori assumption. The parameter values for and are difficult to interpret individually. Across all the models, inactive synaptic contacts are rare, as indicated by the fraction of which ranges between and and . A fit of both connections L4-L4 and L5-L5 with a single parameter set is displayed in Fig. 3. Although the model distributions in Fig. 3a1,b1 do not follow the reference data as closely as in Fig. 2, a good agreement of the distributions and the reference is achieved.
For each of the three reference datasets we obtained many models with a comparable fit error . Fig. 2a4–c4 and Fig. 3a4,b4 show the error (circles) of the best parameter sets obtained, ordered by the value of . We also investigate how the model distribution changes in response to an increase in the baseline probability of causal spike pairing. Some models decrease their contact number, while other models increase it, as can be seen from the derivative (squares in Fig. 3, 1, row 4). This quantity can take very different values for comparable fit errors . We call a model Hebbian if the number of active contacts grows upon an increase of causal spike pairings (). Conversely we call a model anti-Hebbian if the number of contacts decreases (). This diversity indicates that the plasticity model used here is general enough to implement Hebbian and anti-Hebbian learning, depending on the parameters. In Tab. 1 the value of the derivative is given along with other properties of the selected models. For each dataset, we selected the best model irrespective of it being Hebbian or anti-Hebbian. Another characteristic property of a model is the joint distribution of active synapses and inactive synapses (Fig. 2 and 2, row 3). Especially in the model connections L4-L4 and L5-L5 and tend to be strongly correlated. While the expectation value of is largely determined by the reference data, the smaller expectation value of indicates that only a small proportion of spines are small and functionally weak. The marginal distribution of active and inactive synapses are shown for the three best models in Fig. 4a. In all selected models, across the datasets, the expectation values of and do not sum up to the expectation value of the number of contacts , which means that many unrealized synapses (close appositions without inactive or active contacts) are present, consistent with experimental findings [21].
Fig. 4b addresses the question whether a homeostasis of the neuronal firing rate can be achieved by this structural plasticity model. Here, a homeostasis means that an increase in firing rate leads to pruning of input synapses, thus lowering the firing rate in effect. Conversely a decrease in firing rate should lead to synapse maturation. If that is the case, the plasticity rule establishes a homeostatic control of the firing rate to a fixed value. In our model, a negative derivative means that the plasticity rule acts as a firing rate homeostasis. From Eq. (5) we conclude that a change in can either increase or decrease the value of for a given , depending on the value of the baseline correlation . Previously, we arbitrarily set to . However, given a parameter set, we can change the value of to without changing the transition rates (and all equilibrium properties of the model) if we also shift the thresholds to(27)because then all the distances to the threshold are preserved. Thus is effectively a free parameter of the model and can be adjusted to set , as is shown in Fig. 4b. Hence the structural plasticity model we propose can establish a firing rate homeostasis.
Furthermore we derived the expected lifetime of a synapse in the model, which is also shown in Tab. 1. Here the lifetime is defined as the expected time until the synapse is pruned. Before being pruned, it can go back and forth between the states inactive and active several times (fluctuate in volume). The lifetime is very different for inactive and for active synapses, the latter exceeding the former by about one order of magnitude or more. This is due to the fact that typically several active contacts coexist and mutually stabilize, which entails small rates , and (cf. Fig. 2 row 2). If an active synapse becomes inactive, the rate to go back to the active state also increases, which promotes going back and forth through these states. This behavior matches nicely with the volume fluctuations of large dendritic spines described in [22]. Through (25) finally we can relate the time scale of the models to experimental data [10]. The values for are listed in Tab. 1and confirm our assumption of a time scale separation of structural plasticity and neuronal activity. Using this estimate of the timescale, the lifetimes of inactive contacts are about a couple of days, while the lifetimes of active contacts span from a month up to years. [10] called spines with a lifetime of less than days transient, and spines with longer lifetimes persistent. This distinction roughly applies to the lifetimes of inactive and active contacts in our model.
We propose a model of structural plasticity to explain cooperative synapse formation [20]. The transitions of the states of synapses are assumed to depend on a signal locally available to a spine that depends on the correlation between pre- and postsynaptic activity, the correlation trace. There is strong evidence that a correlation trace could indeed be implemented in the dendritic spine through phosphorylation of the macromolecule CaMKII [35], [27], [34]. CaMKII has also been shown to be necessary for structural and long-term plasticity [52], [53], [4], and may also drive presynaptic changes [54]. Here we assume an abstract, effective correlation trace instead of explicitly modeling the dynamics of CaMKII. This makes our results independent of the specific mechanisms employed at the synaptic contact, since also other processes may be available to form the correlation trace. We assume the correlation trace at the spine is forgetful, such that it integrates causal and anti-causal spike pairing events like a leaky integrator with a certain time constant. This time constant affects the location of the equilibrium probability distribution of the correlation trace and its variance. Across the datasets L4-L4 and L5-L5, the time constants are comparable. If the correlation trace is implemented biologically by the cycle of expression, activation and degradation of CaMKII, these time constants will be observable in experiments. The optimized values for the time constant are well in the range of possible values that sustained CaMKII activation can show [51] for all three reference datasets. The differences in the model parameters of the connection L4–L23 compared to the other two intralaminar connections might be explained by the finding that most synaptic contacts of this connection are formed on dendritic shafts rather than on spines [55]. At dendritic shafts functionally similar plasiticity mechanisms could be at work, but our model might be less appropriate for this type of connection. However, although in early postnatal development more shaft synapses exist, in later stages synapses on spines dominate [56], [57].
The rates of structural changes at the synapse are assumed to be a function of the equilibrium correlation trace distribution. To model this dependence mathematically we chose a versatile functional form (7). This is necessary since a comprehensive quantitative description of the correlation dependence of structural plasticity is not known to date. Our optimization results for the transition rates show a strong selectivity for specific numbers of active contacts in a synapse: Transition rates are much higher in case there are few active contacts between two neurons, and many active contacts stabilize the system in all of the three modeled intra-cortical synapse types. Future experiments could investigate whether synaptic contact number (or EPSP amplitude) correlates with calcium transient amplitudes at the spines and with rates of spine maturation, shrinkage and pruning.
Using the optimized models we also computed the expected lifetimes synaptic contacts. The lifetime of active contacts is about ten to one hundred times larger than the lifetime of inactive contacts across our models. This can be understood given the experimental references' results that an active contact is always accompanied by several others. For such synapses, our models predict a vanishing rate of activity dependent transitions, which lets the synapses stay in the active state for a long time. Thus persistent spines here correspond to active contacts, and transient spines to inactive contacts. Our finding constitutes a statistical explanation of the existence of these two distinct classes of spines [10].
Our best-fit models show functional differences. Most notably, the models can be either Hebbian or anti-Hebbian, in the sense that an increase in the frequency of causal spike pairing leads to either increased or decreased numbers of active contacts. Both Hebbian and anti-Hebbian connections have been observed in the neocortex [58]. For all connections we found comparably good fits of both types. Furthermore our model predicts a joint probability distribution of active and inactive contacts which goes beyond current experimental references. Future experiments which determine both of these numbers for many neuron pairs will allow further evaluation of our model. A possibility to optically distinguish and monitor active and inactive synapses in experiments might be to use fluorescent markers for AMPA and NMDA receptors. Synaptic contacts that what we call “inactive” should show less AMPA than “active” ones, but the inactive ones also include those synapses with few AMPA receptors.
Previous models of structural plasticity have assumed a homeostasis of the firing rate [59], [3], in the sense that if neuronal activity increases beyond an a-priori chosen set-point, synaptic contacts are pruned to decrease the excitatory drive, and the reverse for activity below the set-point. Indeed the correlation dependent structural plasticity model [27] shows this behavior. We have investigated whether our models show firing rate homeostasis by computing how the expected number of active contacts changes with the firing rate. This dependency can be chosen arbitrarily by adjusting a free parameter of the model (see Fig. 4b). Our model hence is capable of providing the proposed firing-rate homeostasis for properly chosen parameters.
To obtain a simple Markov process, we used the discrete categories “unrealized”, “inactive” and “active” to describe the state of a synaptic contact. Technically our model is similar to the cascade synapse model of [60] but adds the morphological interpretation of the synaptic states. The inactive contact might be closely related to silent synapses, but in the actual biological system such a clear-cut distinction between functional states can probably not be made, see for example [38]. Busetto et al. [61] found that silent synapses are abundant in the developing animal but vanish in the adult. However, only spines that were morphologically mature were included in their study, making no claim about existence of thin spines with small heads. Quantal EPSC analysis in the adult neocortex showed that close to all synaptic contacts of the connection L4–L23 are functional [55]. Our model of this connection also shows no inactive synapses in expectation, which renders them unobservable in practice. Further [62] find in cultured hippocampal slices that newly formed spines contain AMPA receptors. Small spines, however, are generally easy to miss, since they are often smaller than the resolution limit of optical microscopy [10], [61], and they may also be pruned again quickly after formation [63]. After all there is ample evidence that newly formed spines are small [21] and that AMPA receptor density correlates with volume [22]. We thus follow [43] and approximate thin, small spines as inactive synaptic contacts, and large spines as active ones as described above in detail.
As a consequence of the coarse-grained description of the state of synaptic contacts, all active synaptic contacts in our model produce an EPSP of a fixed amplitude . However, in biology this amplitude varies from contact to contact. Including a fine grained description of synaptic amplitudes in a structurally similar model as the one presented here would result in a massive increase of the dimension of the state space and is therefore potentially unfeasible. Such a dispersion of synaptic amplitudes would result in a different functional dependence of the mean (2) and variance (3) of the correlation trace on the number of active contacts . However, at a given synapse the mean would still be monotonically increasing with . On a population level, the dispersion of synaptic amplitudes thus results in an additional contribution to the width of the distribution of the correlation trace in (3). We can think of part of the noise added to as representing this contribution. This reduces the precision of correlation detection at the dendritic spine. In a model with dispersion of synaptic amplitudes, we therefore expect to find qualitatively similar fits for our coarse grained model at a correspondingly reduced additional noise.
We defined that inactive synaptic contacts host NMDA receptors. The conductance of NMDA receptors increases upon a postsynaptic depolarization if the magnesium block is removed. At negative voltages NMDA channels have a smaller but non-vanishing conductance and hence mediate excitatory postsynaptic currents (EPSC). However, the time scale of NMDA activation is much slower than that of AMPA channels. A postsynaptic action potential partially caused by NMDA currents of one synaptic contact would thus occur much later than the presynaptic glutamate release. The postsynaptic depolarization is therefore less efficient in opening the NMDA receptors at another synaptic contact of the same synapse. This, however, is the crucial mechanism that allows correlation detection and cooperation in our model. Hence one may assume that NMDA currents contribute much less to the correlation trace, and thus have vanishing impact on the cooperative plasticity of our model. We therefore use the term “inactive” here in a functional sense.
In neonatal rat hippocampus also presynaptically silent synapses have been observed, which show a very low probability of transmitter release [64], [65]. However, even a low probability of release enables the formation of a postsynaptic correlation trace at the dendritic spine as in our model. Moreover, even presynaptic changes of the transmitter release have been reported to depend on such a correlation trace in a similar way [54]. The dependence of maturation and shrinkage/pruning on the correlation trace that we use here is a sufficiently generic model to also include these presynaptic mechanisms, although we do not intend to model them here explicitly.
The term structural plasticity describes a broad range of phenomena, many of which have not been addressed here. Competition between synapses from distinct neurons to a common postsynaptic neuron has been shown to be important for the emergence of cortical network structure [66]. In the more detailed models of structural plasticity in neuronal networks based on the activity of CaMKII [27], [67], cooperation and competition between synaptic contacts necessarily occurs. Here we assumed that synapses between different pairs of neurons develop independently, so inter-synaptic competition effects were not considered. Furthermore, structural plasticity also includes changes to the network structure that can come about by migration of axons on much longer time scales. Our model rather describes the steady state of the adult cortex, during which spines form and retract, but the axonal arborization can be assumed to be constant [13]. In lesion studies it has been shown that the steady state can become unstable and axons again begin to migrate [68].
Although simple and abstract in its description of complex cellular phenomena, our model can explain the cooperation of synaptic contacts in the adult neocortex, postulated in [20]. The model shows how continuously active structural plasticity can lead to the global configuration of synaptic contact numbers that was observed experimentally. The key ingredient of the model which mediates the necessary cooperation is a trace of the spike-timing correlations of the pre- and postsynaptic neuron. The resulting synaptic learning rule is local (it solely requires mechanisms at the synaptic contacts) but can nonetheless explain cooperative synapse formation.
|
10.1371/journal.pntd.0000433 | Early Exposure of Infants to GI Nematodes Induces Th2 Dominant Immune Responses Which Are Unaffected by Periodic Anthelminthic Treatment | We have previously shown a reduction in anaemia and wasting malnutrition in infants <3 years old in Pemba Island, Zanzibar, following repeated anthelminthic treatment for the endemic gastrointestinal (GI) nematodes Ascaris lumbricoides, hookworm and Trichuris trichiura. In view of the low intensity of worm infections in this age group, this was unexpected, and it was proposed that immune responses to the worms rather than their direct effects may play a significant role in morbidity in infants and that anthelminthic treatment may alleviate such effects. Therefore, the primary aims of this study were to characterise the immune response to initial/early GI nematode infections in infants and the effects of anthelminthic treatment on such immune responses. The frequency and levels of Th1/Th2 cytokines (IL-5, IL-13, IFN-γ and IL-10) induced by the worms were evaluated in 666 infants aged 6–24 months using the Whole Blood Assay. Ascaris and hookworm antigens induced predominantly Th2 cytokine responses, and levels of IL-5 and IL-13 were significantly correlated. The frequencies and levels of responses were higher for both Ascaris positive and hookworm positive infants compared with worm negative individuals, but very few infants made Trichuris-specific cytokine responses. Infants treated every 3 months with mebendazole showed a significantly lower prevalence of infection compared with placebo-treated controls at one year following baseline. At follow-up, cytokine responses to Ascaris and hookworm antigens, which remained Th2 biased, were increased compared with baseline but were not significantly affected by treatment. However, blood eosinophil levels, which were elevated in worm-infected children, were significantly lower in treated children. Thus the effect of deworming in this age group on anaemia and wasting malnutrition, which were replicated in this study, could not be explained by modification of cytokine responses but may be related to eosinophil function.
| Infants and very young children commonly become infected with intestinal nematode infections. However, the worm burdens are generally very light, so a beneficial effect of deworming on wasting malnutrition and anaemia in this age group which we have demonstrated was unexpected and the mechanism unclear. To investigate this, we have, for the first time, determined whether such worm infections in infants induce significant immune reactions which might be detrimental to nutrition and growth e.g. by inducing inflammation in the gut or by cytokine effects on erythropoiesis. We also determined if such responses are modulated by regular deworming over a 9 month period. Peripheral blood cells from infants infected with Ascaris and hookworms in particular responded to stimulation with worm antigens, producing predominantly Th2 cytokines. Although the Th2 cytokine responses in the periphery were not significantly altered by deworming, the levels of eosinophils, which are regulated by the Th2 cytokine, IL-5, were lower after treatment. It is possible that eosinophils play a role in gut pathology leading to wasting malnutrition and anaemia in the very young and that this effect is reduced by deworming.
| In endemic countries infants are exposed to gastrointestinal (GI) nematode infections soon after birth and infection intensity increases during childhood. Helminth infections in children are associated with malnutrition [1],[2], linear growth stunting [3] as well as iron deficiency anaemia [4],[5], effects related to the intensity of infection [6] and generally attributed to the direct and indirect effects of the worms on the gut i.e. blood loss, mucosal damage, secondary infection, malabsorption [7]–[10]. Several studies have now documented that deworming leads to improvements in nutritional outcomes e.g. anaemia and wasting malnutrition and in development of school-aged children who often harbour the highest intensities of these worms [11]–[13]. However, a recent study in Pemba found significant benefits of anthelminthic treatment on growth, anaemia and appetite in children <30 months of age who harboured very light GI nematode infections [14]. Indeed the benefits of mebendazole were significant only in this younger age group and not in children 30–71 months old. In view of this unexpected effect of deworming very lightly infected children, it was suggested that the nutritional benefit may be related to prevention of the indirect effects of the worms such as on the immune responses they induce rather than to their direct effects.
Various aspects of the immune response to the initial/early exposure to GI nematodes might contribute to anaemia and malnutrition, and this may be alleviated by worm treatment. Pro-inflammatory cytokines and acute phase proteins can suppress appetite [15],[16], induce protein loss [17] and raise the levels of resting energy expenditure [18],[19], as well as affect anaemia (the anaemia of chronic disease) [15]–[17],[19],[20]. One possibility therefore is that primary exposures to GI nematode infections induce inflammatory (Th1-mediated) responses in a significant proportion of the infants resulting in anaemia and malnutrition. Although older humans in endemic areas generally develop Th2 dominated cytokine responses to GI nematode infections, characterised functionally by IgE and eosinophilia [21]–[23], some studies have shown that Th1 cytokines are induced [24]. The responses to initial exposures in infants has not been studied previously but studies of gut nematode infections in mice have shown that, depending on the worm species and host genotype, the response to primary infection can be polarized to either Th1 or Th2 [25] and influenced by infection intensity [26]. By analogy, it may be hypothesised that light primary exposures in the human population may result in Th1 responses, in at least a proportion of individuals.
Th2-mediated responses may also affect gut function leading to impaired nutrition. Studies of GI nematode infections in mice have shown that the barrier function of the mucosa can be profoundly altered by the action of Th2 cytokines on epithelial cells and/or mast cells resulting in increased mucosal permeability, reduced glucose absorption, increased ion secretion and intra-luminal fluid accumulation [27]–[30]. Infections of Ascaris suum in pigs, which are considered a relevant model for human ascariasis, cause similar effects coincident with upregulated expression of Th2 cytokines [31]. In humans, T. trichiura can induce mast cell infiltration and an immediate hypersensitivity response in the colon of infected children leading to release of histamine [32]. During T. suis infection in pigs, a good model for human trichuriasis, upregulation of expression of Th2 cytokines in the mucosa parallels mucosal hypertrophy characterised by infiltration of mast cells and eosinophils which may play an immunopathological role [33]. Similarly in humans, hookworm infections result in eosinophil infiltration [34], Charcot-Leyden crystal production [35] and, following infection with Ancylostoma caninum, eosinophilic enteritis [36]. Worm-induced Th2 cytokines can also induce increased smooth muscle contractility in mice [28],[37],[38] and pigs [31] and, in mice, can induce epithelial cell turnover [39], goblet cell hyperplasia and mucus secretion [40],[41].
The current immunological study was a trial within a larger field based randomised treatment trial (manuscript in preparation) to confirm the beneficial effects of treating intestinal helminth infections in early childhood on anaemia and malnutrition [14]. The immunological study was a primary aim of the project designed to investigate possible immunological mechanisms involved in the pathogenesis of these early infections and the amelioration of this by deworming. The specific aims were to establish (i) if measurable immune responses to worm infections (cytokines, acute phase proteins) could be demonstrated in very young (6–24 months) children harbouring light infections (ii) if so, whether such infants made predominantly Th1 or Th2 cytokine responses or some one and some the other; (iii) whether such responses were altered by periodic (3-monthly) anthelminthic treatment which might explain the benefits afforded by such treatment in this age group.
This study was nested within a community-based treatment trial designed to test whether periodic mebendazole treatment in 6–24 month old infants would decrease rates of severe anaemia and protein-energy malnutrition (International Standard Randomised Controlled Trial Number 83988447). The study was performed between September 2003 and October 2004 at the Public Health Laboratory-Ivo de Carneri, Pemba Island, Zanzibar, United Republic of Tanzania. Pemba Island is densely populated and mostly rural, with subsistence farming as the main economic activity. Plasmodium falciparum malaria is holoendemic, as are the geohelminths, Ascaris lumbricoides, Trichuris trichiura, Ancylostoma duodenale, and Necator americanus.
Initially 2664 children aged 6–23 months were screened for helminth infection. During the screening process, age-matched triplets of infants (comprising 2 infected (matched for infection species) and 1 uninfected infant) were formed and randomised for treatment stratified by age (3 groups 6–11, 12–17 and 18–23 months) and by infection status (Ascaris, Trichuris, Ascaris and Trichuris, hookworm with or without any other infection). These children formed the immunology study cohort; 335 infants were randomised to placebo and 318 to mebendazole treatment. The CONSORT protocol is in Figure S1 and Protocol S1. All children screened but not selected for the immunology study were subsequently randomised in the main community-based treatment trial with random allocation to treatment or placebo groups. The immunology study children were still involved in the randomised treatment trial (manuscript in preparation). Age-matched selection into the immunology study was essential to the design, because the probability of infection was very strongly related to infant age. Without age-matched selection, the infected children would naturally have been older than the uninfected children, creating a biased comparison with regard to infection status. Having created the age-matched samples of infected and non-infected children, we then analysed the data without regard to the original matching. This is a valid approach for matched follow-up (cohort) studies [42]. At baseline blood was taken for the immunological investigations after which the infants were treated with a 3 day course of mebendazole, 100 mg twice daily, or identical placebo that was repeated every 3 months over a study period of 12 months. A blood sample was again taken for immunological studies 1 month after the 3rd treatment round to allow time for any possible effects of worm reductions on cytokine responses to develop. The study was approved by the ethical review committees of the London School of Hygiene and Tropical Medicine, Johns Hopkins Bloomberg School of Public Health, Cornell University, and the Ministry of Health of Zanzibar. Because of the high rate of illiteracy amongst parents, verbal informed consent was obtained from the mothers or from the guardians of all enrolled infants, documented by signature of a literate witness, following the ethical review committees' approval.
Stool samples were collected on 2 consecutive days and stored at 4°C. Individual Kato-Katz slides were prepared from both samples and the means taken [43]. The two samples were then combined and 2 g were used for assessment by a sedimentation technique [44]. In a small proportion of the Ascaris positive stool samples (6.6%), the Ascaris egg counts were very high and egg counts in individual Kato-Katz slides were stopped at 999 (i.e. 23, 976 epg). The percentage of egg reduction induced by treatment (ERR) was estimated as 100[1−exp(−D)]%], where D was the mean difference for a particular treatment.
The somatic hookworm antigen from adult Necator americanus, maintained in a hamster life cycle was prepared as described elsewhere [35]. Necator americanus worms were kindly provided by Prof. J Behnke, Prof. D Pritchard and Dr A Brown of Nottingham University. The Trichuris suis and Ascaris suum were kindly provided by Dr Dolores Hill and Dr Joseph Urban Jnr of the United States Department of Agriculture. The Ascaris and Trichuris antigens were prepared as described elsewhere [45],[46]. In brief, the Ascaris antigen was derived from adult A. suum that were homogenized, extracted in 1× Dulbecco's PBS overnight at 4°C, spun at 20,000 g, concentrated and dialysed against 10 mM TBS. The supernatant was filter-sterilised, aliquoted and stored at −80°C. Adult T. suis worms were cultured for 36 hours and culture fluid used for the ES antigen. Somatic antigen from T. suis was prepared as for A. suum above. Protein concentrations were determined using the Bio-Rad protein assay.
The Whole Blood Assay (WBA) was carried out as described elsewhere [35]. Heparinised venous blood was used no later than 4 hours after venepuncture. Preliminary validation of the helminth antigens to induce recall cytokine responses in the WBA was carried out in endemic helminth infected teenagers before use in the infant studies. Ascaris and Trichuris antigens were used at a final concentration of 30 µg/ml, whilst a pool of somatic hookworm antigens was used at 20 µg/ml. Phytohemagglutinin and purified protein derivative concentrations as well as haematology procedures for differential cell counts were carried out as described elsewhere [35].
Culture supernatants were stored at −80°C. Matched monoclonal antibody pairs from Pharmingen (Oxford, UK) or R & D Systems (Abingdon, UK) were used according to the manufacturer's instructions (IL-5: TRFK5 and JES1-5A10, IL-10: JES3-9D7 and JES3-12G8, IL-13: JES10-5A2 and B69-2 from Pharmingen; IFN-γ from R & D Systems). When a new kit was introduced it was validated by testing in parallel with the previous kit using a large batch of positive culture supernatant which was used throughout the study.
All analyses were performed using the STATA statistical analysis software package (version 9; Stata Corp). Medium alone negative control values were subtracted from all results that were above the lower limit of assay detection (i.e. 15 pg/ml). All ELISA plate readings were standardised for each cytokine by use of the positive control supernatant run on each plate in duplicate. Where data are presented as percentage responders, a response was defined as a cytokine concentration of >31.24 pg/ml, derived from a comparison of the frequency of responses in worm negative infants with worm positive infants, where a bimodal distribution was observed. Contingency tables and Pearson's Chi-squared tests were used to compare proportions of responders. Fisher's exact test was used for small sample sizes. T test or ANOVA were used with adjustment for multiple comparisons by Bonferroni procedure. Effects of age and sex were investigated using ANOVA. Also in view of the report of effects of malaria infection on helminth cytokine responses [24], malaria infection status was included in the analysis. Non-normally distributed variables were transformed, or non-parametric tests (Mann Whitney test, Wilcoxon signed rank test or Kruskal-Wallis test) were used. To examine the relationships between two variables non-parametric regression (Lowess) was used. If the relationship was approximately linear, Pearson correlation coefficients were calculated and a linear regression model was fitted. If the relationship was non-linear, the data were transformed. If this failed to produce a linear relationship, then the non-parametric correlation Spearman test was used. Bootstrap was used to infer variances of regression coefficients, P-values and 95% confidence intervals when the data were not normal. Regression models were used to identify predictors for cytokine response adjusting for age, sex and malaria. Cure rates, percentage reduction in prevalence and egg reduction rates were calculated as described [47]. Cytokine responses measured from stimulation of whole blood with Trichuris ES and somatic antigen were minimal in infants, and mean response to the two antigens were used in analysis.
The 666 subjects were selected on the basis of being (any) worm positive or negative at a ratio of approximately 2∶1 (70.7% were worm positive) with age-matching of infected and uninfected infants. Amongst the worm positive infants, 42.0% had Ascaris infection (±other worms), 16.8% were positive for hookworm (±other worms) and 71.5% were positive for Trichuris (±other worms) (Table 1 and Figures S2, S3, and S4). Mean intensities (eggs per gram [epg]) were 1061 for Ascaris, 213 for hookworm and 213 for Trichuris (see also Figure S5) and there was no significant effect of co-infection with one of the other worms on intensities of infection. According to WHO categorisation [48] the majority of the infants harboured “light” infections (80.4% Ascaris, 1–4999 epg, 96.8% hookworm, 1–1999 epg, 89.9% Trichuris, 1–999 epg). Egg counts were significantly positively associated with age for Ascaris egg positive infants (n = 184, β = 1.06, P = 0.047, 95% CI 1.00–1.12) but this was not significant for Trichuris or hookworm egg positive infants. There was no significant association between intensity and sex. The prevalence of Plasmodium species infection in this age group was 24.9%.
The frequency of subjects making cytokine responses above the cut off (i.e. >31.24 pg/ml, see Materials and Methods) at baseline for the Ascaris, and Necator antigens is shown in Figure 1.
Amongst Ascaris egg positive infants, 47% and 46% respectively made IL-5 and IL-13 responses whilst only 20% and 12% respectively made IFN-γ and IL-10 (Figure 1A). A significantly higher proportion of infants in the Ascaris positive group made IL-5 and IL-13 responses compared to both worm negative infants and Ascaris negative/other worm positive infants. The percentage of IL-5 and IL-13 responders was also significantly higher in Ascaris negative/other worm positive infants compared to worm negative infants. The percentage of IFN-γ or IL-10 responders was similar between the infection groups.
Similarly, 54% of hookworm egg positive infants made IL-5 and IL-13 responses to the homologous Necator antigen but only 6% and 7% made IFN-γ and IL-10 responses respectively (Figure 1B). A significantly higher percentage of hookworm positive infants made antigen-specific IL-5 and IL-13 positive responses compared to both worm negative infants and to hookworm negative/other worm positive infants. Hookworm negative/other worm positive infants also had a significantly higher percentage of responders compared to worm negative infants. The percentage of IFN-γ and IL-10 responders was very low in all groups.
There were minimal cytokine responses to the Trichuris antigen amongst the infants, with a very low percentage of responders (<5%), similar between the infection groups (data not shown).
The frequency of responses to phytohemagglutinin (PHA) and purified protein derivative (PPD) which were included in all assays were: IL-5, 85%; IL-13, 91%; IFN-γ, 52% and IL-10, 56% for PHA and the responses amongst BCG-vaccinated infants (BCG scar-positive) to PPD were:- IL-5, 48%, IL-13, 50%; IFN-γ, 78% and IL-10, 24%.
The mean cytokine concentrations produced in cultures to the helminth antigens for responders (>31.24 pg/ml) are shown in Figure 2 and reflect the data on frequency of responses.
The mean levels of IL-5 and IL-13 to Ascaris antigen were higher albeit not significantly in both Ascaris positive and Ascaris negative/other worm positive infants compared to worm negative infants (Figure 2A). The mean levels of IL-5 and IL-13 to hookworm antigen were significantly different between the infection groups (IL-5: ANOVA F(4, 189) = 4.58 P = 0.011, IL-13: ANOVA F(4, 179) = 5.98 P = 0.003). Mean IL-5 and IL-13 levels in the hookworm positive responders were significantly higher compared to worm negative infants (IL-5: P = 0.023, IL-13: P = 0.027), and IL-13 levels for the hookworm positive responders were also significantly higher compared to hookworm negative/other worm positive infants (P = 0.011) (Figure 2B). The mean responses to Trichuris antigen (<200 pg/ml) were not significantly different between the different infection groups (data not shown).
The above data showed that Th2 cytokine responses to Ascaris and hookworm antigens predominate amongst infected infants. To demonstrate the reliability of the IL-5 and IL-13 results as an indicator of overall Th2 responsiveness we plotted the correlation between the levels of these two cytokines for the Ascaris antigen stimulations. IL-5 and IL-13 responses to Ascaris antigen in worm positive infants have a significant positive association (n = 471, repetitions = 1000, Bootstrap coefficient = 1.31, SE = 0.11, P<0.001, 95% CI 1.09–1.53, r2 = 0.8438). A similar correlation holds for hookworm antigen responses (data not shown).
Although there was no significant association between Th2 (IL-5 and IL-13) responses and IFN-γ responses in worm positive infants, the relatively few infants who made elevated IFN-γ responses to Ascaris and hookworm antigens also made elevated Th2 responses, whilst many infants who did not make IFN-γ responses made high IL-5 responses. Thus there was not a subset of infants who made a Th1 biased response.
Figure 3 shows the levels of Th2 cytokine responses to Ascaris and hookworm antigen respectively in relation to age and infection status.
The parasitology data following the last round of 3-monthly mebendazole treatment are shown in Figure 4A and 4B. In the mebendazole treated group the prevalence of any worm infection at follow-up (40.9%) was reduced by 42% compared to baseline (70.3%) (z = 7.20 P<0.001, data not shown) and was 41% lower compared to the placebo group (69.9%) (Chi2 = 50.7237 P<0.001). In the placebo group there was no significant difference between prevalence at baseline (68.9%) and follow-up (69.9%). The reduced prevalence in the mebendazole compared with the placebo treated infants was also seen when stratified by worm species (Ascaris 2.7 vs 19.9% z = 6.6352 P<0.0001, Trichuris 36.8 vs 64.7% z = 6.8413 P<0.001, and hookworm 8.4 vs 14.7% z = 2.3954 P = 0.0166). The egg reduction rate followed a similar pattern with mebendazole causing a greater reduction in egg count for Ascaris infections, followed by Trichuris and then hookworm infections (Figure 4B).
In view of the correlation between worm infection and levels of IL-5, which controls eosinophil production, maturation, migration and persistence in the tissues [49]–[52], the pattern of blood eosinophilia was of interest. As seen in Figure 6, worm positive infants had a significantly higher mean eosinophil count compared to worm negative infants at baseline (Figure 6A) and at follow-up (Figure 6B). Also the mean count was slightly but significantly lower (P = 0.0039) in mebendazole treated infants compared with placebo (Figure 6C). By comparison, basophil counts in infants were minimal (<1%) throughout.
In this study the goals at baseline were i) to establish that GI nematode specific cytokine responses could be measured in infants and ii) to determine the levels and balance of Th1/Th2 cytokines induced by initial exposure of infants to infections with A. lumbricoides, hookworm or T. trichiura. The prevalence of infection with these helminths in 5–11 month old infants in Pemba at the time of this study was only 26.5% [44] but previous studies in Pemba have shown that the prevalence reaches 90% by the age of 3–5 yr [14]. Transmission on the island is year round and so it is likely that many of the infections detected in the 6–24 month old infants in this study would represent recent primary exposures to infection. Despite this and the fact that the majority of infections were very light, cytokine responses to Ascaris and to hookworm antigens could be demonstrated in significant proportions of the infants. Notably cytokines of the Th2 subset predominated amongst responders and there was no evidence of a subset of individuals who made Th1 polarised responses. This is in contrast to primary exposures of mice to the nematode T. muris which induces Th2 responses in certain inbred strains but Th1 responses in others [39].
Although T. trichiura was a common infection in this age group minimal cytokine responses were seen to Trichuris antigen. Low cytokine responses to Trichuris antigens have also been reported in some studies of older humans [53],[54] while others have reported higher responses [55],[56]. We do not consider that the failure to detect cytokine responses to T. trichiura infection in the infants was due to the use of the heterologous Trichuris suis antigen since we showed in preliminary studies with the WBA that this antigen was able to stimulate cytokine production from blood of T. trichiura infected Pemban teenagers. Furthermore, we also found that heterologous antigen from T. muris (kindly provided by Prof J Bradley, University of Nottingham, UK), which has been shown to induce cytokine responses in older humans in other studies [55],[56] also failed to stimulate cytokine production from our T. trichiura infected infants (data not shown). It is possible that the apparently greater sensitization to Ascaris and hookworm antigen compared to Trichuris is due to the fact that, unlike Trichuris, Ascaris and hookworms have a larval migratory phase which may have a major role in immune stimulation as was reported in an experimental hookworm infection [35].
The cytokine responses to Ascaris and hookworm antigens were significantly positively associated with age at baseline and also consistently increased between baseline and follow-up. These differences between baseline and follow-up were not due to technical differences since there was overlap between the testing of the samples from the baseline and follow-up and validation of all cytokine assays over the course of the study using a pool of positive control supernatant which was included on all plates. Since, the cytokine responses to PHA or PPD did not show this consistent increase at follow-up (data not shown) we conclude that the increased response reflects increased worm exposure over time. The greater frequency and level of Th2 compared with Th1 responses to Ascaris and hookworm antigen seen at baseline were maintained over a year of further exposure. Such a Th2 bias is also apparent following prolonged exposure to Ascaris and Trichuris infections [22],[55] but a more balanced Th1/Th2 cytokine response has been reported in hookworm infection in children and adults [24],[57],[58] although these studies employed purified peripheral blood cell in culture rather than whole blood which may have influenced the cytokine profiles demonstrated.
Amongst the Ascaris or hookworm egg positive infants there were a higher proportion of responders to the homologous antigen than amongst egg negative infants or infants with a different species of worm indicating some degree of specificity in the antigen responses to particular worms. However, a proportion of infants lacking Ascaris or hookworm infections but harbouring one or both of the other worms also responded to Ascaris or hookworm antigens respectively. This could be due to antigen-specific sensitization by prepatent infections in these individuals or to a failure of parasitological detection of infection. However, it may also reflect a degree of cross-reactivity in the responses to the worm antigens as previously suggested [22],[59]. Antigenic cross-reactivity is also supported by the work of Jackson et al 2004 [56] who reported that cytokine responses to somatic T. trichiura, T. muris and A. lumbricoides antigens in WBA were strongly intercorrelated even though the majority of people in the study area had single T. trichiura or A. lumbricoides infections. Another possible explanation for positive responses in parasitologically negative infants could be prenatal priming to helminth antigens in helminth infected mothers [60].
A proportion of infants who were infected with Ascaris and hookworm did not make detectable antigen-specific cytokine responses. This was not correlated with intensity of infection and so does not seem to be due to a sub-threshold level of immune priming. A possible explanation is that the larval phase of infection rather than the persisting egg-laying adult worms may be largely responsible for the cytokine production [35] and so responders may be the more frequently/more recently exposed individuals.
With regard to our starting hypothesis, that the immune response to the worms may contribute to anaemia and wasting malnutrition in infected infants it is clear that the idea of Th1 sensitisation leading to pro-inflammatory cytokines such as TNF-α and IL-6 affecting nutrient metabolism, erythropoiesis and appetite is not supported by this data. The analysis of acute phase proteins and nutritional indicators will be reported elsewhere (manuscript in preparation). It is conceivable that the GI nematode-specific Th2 cytokines demonstrated in infants could be responsible for impaired nutrition due to effects on gut function as demonstrated in mice and pigs [27]–[31] and/or increased nutritional demand due to the generation of immune components. However, we found no evidence that periodic anthelminthic treatment reduced the level of systemic Th2 responses although it again led to reduced anaemia and wasting malnutrition (manuscript in preparation). The only response correlating with worm infection status which was significantly altered by treatment was the decline in peripheral blood eosinophilia. Eosinophil infiltration local to sites of worm infestation has been shown in humans harbouring light T. trichiura infections [61] and eosinophils have been implicated in the enteritis induced by zoonotic hookworms [36]. So perhaps eosinophils are involved in mediating gut inflammation and impairing nutrition. Other locally generated responses could impact on gut inflammation and function e.g. helminth-infected infants make more pronounced inflammatory cytokine responses to generic TLR ligands [62]. It should be pointed out that immune responses local to the worms in the gut may differ from recall responses seen in the periphery e.g. in pigs T. suis induces a much higher frequency of IL-4 positive cells in ileo-caecal lymph node lymphocytes compared to PBMCs [63]. So reduction in the numbers of worms by chemotherapy may significantly reduce local immunopathological effects in the gut even in the face of unaltered systemic immune responses.
Following the implementation of various helminth control programmes in Pemba Island, the prevalence and intensity of infections in the infants in this study were low and Trichuris predominated. Similar studies in areas of higher transmission and with other species predominating would be of interest.
|
10.1371/journal.pbio.1001360 | Regulation of DNA Replication within the Immunoglobulin Heavy-Chain Locus During B Cell Commitment | The temporal order of replication of mammalian chromosomes appears to be linked to their functional organization, but the process that establishes and modifies this order during cell differentiation remains largely unknown. Here, we studied how the replication of the Igh locus initiates, progresses, and terminates in bone marrow pro-B cells undergoing B cell commitment. We show that many aspects of DNA replication can be quantitatively explained by a mechanism involving the stochastic firing of origins (across the S phase and the Igh locus) and extensive variations in their firing rate (along the locus). The firing rate of origins shows a high degree of coordination across Igh domains that span tens to hundreds of kilobases, a phenomenon not observed in simple eukaryotes. Differences in domain sizes and firing rates determine the temporal order of replication. During B cell commitment, the expression of the B-cell-specific factor Pax5 sharply alters the temporal order of replication by modifying the rate of origin firing within various Igh domains (particularly those containing Pax5 binding sites). We propose that, within the Igh CH-3′RR domain, Pax5 is responsible for both establishing and maintaining high rates of origin firing, mostly by controlling events downstream of the assembly of pre-replication complexes.
| Each time a mammalian cell duplicates its genome in preparation for cell division it activates thousands of so called “DNA origins of replication.” The timely and complete duplication of the genome depends on careful orchestration of origin activation, which is modified when cells differentiate to perform a specific function. We currently lack a universally accepted model of origin regulation that can explain the replication dynamics in complex eukaryotes. Here, we studied the mouse immunoglobulin heavy-chain locus, one of the antibody-encoding portions of the genome, where origins change activity when antibody-producing B cells differentiate in the bone marrow. We show that multiple aspects of DNA replication initiation, progression, and termination can be explained mathematically by the interplay between randomly firing origins and two independent variables: the speed of progression of replication forks and the firing rate of origins along the locus. The rate of origin firing varies extensively along the locus during B cell differentiation and, thus, is a dominant factor in establishing the temporal order of replication. A differentiation factor called Pax5 can alter the temporal order of replication by modifying the rate of origin firing across various parts of the locus.
| During the S phase, mammalian chromosomes replicate in a precise temporal order, with the timing of replication typically changing gradually across hundreds of kilobases. Cell differentiation induces regional changes in the order of replication which can affect 45%, or more, of the mouse genome [1]. Various studies have examined how the temporal order of replication is established and modified at specific gene loci, but provided discordant explanations about the role played by DNA origins of replication. For example, within a 340 kb portion of the Igh locus, changes in replication timing have been linked to modifications in the distribution of active origins and in their firing efficiency (see definitions in Table 1) [2]. In contrast, within the beta-globin locus, changes in replication timing can occur without significant changes in origin distribution, or firing efficiency, and have been ascribed to modifications in the timing of origin firing [3]–[5]. Does this mean that the temporal order of replication is determined by multiple mechanisms? Are origin distribution, firing efficiency, and the timing of origin firing regulated independently? Which aspect of origin activation is controlled by cell differentiation? These are some of the questions addressed in this study.
Answering these questions requires a quantitative understanding of the dynamics of origin firing. Based on measurements of average origin activity across entire genomes, various stochastic models of origin firing have been recently used to explain specific aspects of eukaryotic DNA replication, such as the duration of S phase [6]–[13]. If origin firing can occur stochastically anywhere along the genome and at any time during S phase, origin distribution and the timing of origin firing cannot be responsible for establishing the temporal order of replication [14]. Recent observations indicate that the profile of replication timing of the budding yeast genome can be explained by differences in the firing rate of individual origins and stochastic origin firing [15]. However, yeast differs from metazoans in many aspect of DNA replication (e.g., S. cerevisiae has well-defined origins of replication, lacks the developmental control of the temporal order of replication, shows no correlation between gene expression and the temporal order of replication, has a short S phase, etc.). In addition, previous studies have mostly relied on the measurement of individual parameters of DNA replication which can be modeled with limited detail to determine the dynamics of origin firing (e.g., the timing of replication) and can produce misleading results if applied to complex genomes [16]. Hence, testing this hypothesis rigorously in mammalian cells requires the measurement of multiple parameters of DNA replication.
Any large portion of a mammalian genome can be used to test the stochastic firing of origins provided enough information is available about DNA replication. For this reason, we used the assay called single molecule analysis of replicated DNA (SMARD) [17] to collect unbiased information about all aspects of DNA replication initiation, progression, and termination across a 1.4 megabase region encompassing the mouse immunoglobulin heavy-chain (Igh) locus (Figure 1A). The experimental data sets collected by SMARD included the temporal order of replication, the steady-state distribution of replication forks, the time required to replicate the region, the average speed of replication forks, the distribution of initiation and termination events, the percentage of replicating molecules containing initiation and termination events, and the average number of events per molecule.
Using a novel mathematical procedure [18], we established that the experimental data sets collected by SMARD are fully consistent with the stochastic firing of origins as defined in Table 1. We also show that many aspects of DNA replication (including the temporal order of replication) can be explained by variations in the rate of origin firing I(x,t) along the Igh locus (see definition in Table 1). According to the nomenclature proposed by others [19], this rate indicates the number of initiation events occurring per length of unreplicated DNA, over a given period of time, as mathematically defined in Materials and Methods.
Our results point to significant differences in the regulation of DNA replication between the mouse Igh locus and yeast chromosomes. In S. cerevisiae, each origin of replication appears to be characterized by a specific firing rate, which differs for different origins [15]. Within the Igh locus, initiation events lack focal points corresponding to individual origins. Instead, the locus comprises large domains (spanning tens or hundreds of kilobases) where the firing rate of multiple origins is virtually uniform and similarly regulated. Along the locus, changes in the firing rate occur abruptly at the border between different Igh domains, and the firing rate of each domain is not affected by deletions that span multiple origins. It is the combined effect of different domain sizes and firing rates that determines the temporal order of replication. This organization remains valid for cells at different stages of B cell development (e.g., bone marrow pro-B cells blocked at the uncommitted and committed stages of differentiation by homozygous mutations of the Pax5 and Rag2 genes; Figure 1B). We also show that the changes in DNA replication that occur during B cell commitment can be quantitatively explained by substantial changes in the firing rate of origins within specific domains of the Igh locus. Therefore, the rate of origin firing is the parameter that is being regulated across large sections of the locus during cell differentiation. The role of the developmental regulator Pax5 in this process and its mechanism of action are also examined.
Bone marrow pro-B cells isolated from Pax5−/−Rag2−/− mice retain the ability to proliferate but are blocked at the uncommitted stage of differentiation and maintain the Igh locus in germline configuration [20]. In order to perform SMARD, we sequentially labeled a population of exponentially growing cells with 5′-iodo-2′-deoxyuridine (IdU) and 5′-chloro-2′-deoxyuridine (CldU), for 3–4 h (see Table S1, column b). Under these conditions, each labeling period is long enough to allow the complete replication of large sections of the Igh locus, resulting in DNA molecules variably substituted with the halogenated nucleotide (see example in Figure S1) [2],[17],[21],[22]. We then digested the genomic DNA with restriction enzymes that cut infrequently within the locus (PmeI, or PacI, or SwaI) and isolated four of the resulting restriction fragments by pulsed-field gel electrophoresis (gray bars in Figure 1A). These fragments were stretched on microscope slides, hybridized with specific DNA probes, and analyzed by fluorescence microscopy to detect the incorporation of the nucleotide analogs along individual DNA molecules.
The molecules fully substituted with halogenated nucleotides provide a vast amount of information about the process of DNA replication [2],[17],[21],[22]. In Pax5−/−Rag2−/− cells, the analysis of ∼4,000 hybridization signals yielded 1,304 fully substituted DNA molecules, 764 of which met the standards required to perform precise measurements (Table S1, columns c–f). This population includes molecules that incorporated only one type of halogenated nucleotide (single-labeled molecules), as well as molecules that incorporated both of them (double-labeled DNA molecules). As explained in previous publications [17],[22], the ratio between single- and double-labeled molecules is proportional to the time required to replicate each restriction fragment (Table S1), which is linked to the average number of replication forks participating in the replication of each restriction fragment and to their average speed (Table S1). Thus, these values can be directly determined from the experimental data as described in the legend of Table S1. This allowed us to determine that, in uncommitted pro-B cells, the average speed of replication forks is similar within the four restriction fragments (between 2.25 and 3 kb/min). A similar value was also measured at the genomic level using a different assay (∼2.5 kb/min; IV and PN, unpublished observation). This suggests that replication forks move at comparable speeds throughout the Igh locus.
In the double-labeled DNA molecules, IdU-CldU transitions mark the positions of replication forks at the time of the label switch. Initiation events appear as IdU-labeled regions surrounded by CldU, while fork collisions display the complementary pattern (e.g., Figure 1C–D). Figure 2A summarizes the location of each initiation event (red bars) and fork collision (green bars) detected in 158 individual restriction fragments. The normalized frequency of the events across the four fragments is also indicated (expressed as number of events scored per 100 double-labeled DNA molecules, per 100 kb). The values before normalization are presented in Table S1. Note that initiation events and fork collisions spanning adjacent restriction fragments are not shown in the figure since these events are scored as simple IdU-CldU transitions (a fact that is taken into consideration by our mathematical model and in Figure S2C). Overall, these results indicate that initiation events are more frequent near the DH-JH and middle-VH gene families than in other parts of the locus. In these regions, the presence of non-overlapping events implies the existence of clusters of active origins that are reminiscent of initiation zone [23],[24]. In contrast, fork collisions are more frequent between the origin clusters (termination regions in Figure 2A).
It is important to point out that SMARD experiments are designed to detect only a fraction of the active origins (the number of double-labeled DNA molecules required to reach mapping saturation increases with the number of potential origins). Moreover, reaching mapping saturation is increasingly unlikely where origin efficiency is low, and near the end of the restriction fragments. Thus, Figures 2A and S2 are expected to largely underestimate the actual density of origins along the locus (see definition in Table 1). Even so, the distance between the midpoints of initiation event within the DH-JH region suggests an origin density higher than one potential origin per 10 kb (Figure S2E, PacI#3). In Figure 2A, origin density appears lower near the middle-VH region genes. However, within this section of the locus, the midpoints of initiation events overlap only rarely, indicating that that origin mapping is far from saturation (Figure S2E, PmeI#5). Lack of saturation is even more pronounced across the proximal-VH region (none of the detected initiation events overlap; Figure 2A, PmeI#4). We conclude that potential origins of replication are likely to be present at a relatively high density across most of the Igh locus.
Since cells grow asynchronously during DNA labeling, the IdU-CldU transitions depict all stages of DNA replication, reflecting the dynamics present at the steady-state of growth (right panel in Figure S1) [2]. This means that the population of double-labeled DNA molecules can be used to determine the steady-state distribution of replication forks across the Igh locus (arrowheads in Figure 2B–C), and the average number of forks participating in the replication of each restriction fragment (Table S1, column l). Similarly, the temporal order of replication can be obtained from the level of IdU substitution of these molecules, with peaks marking the regions that replicate first and valleys marking the regions that replicate last (Figure 2D). Overall, these results indicate that, in uncommitted pro-B cells, DNA replication tends to begin near the DH-JH and middle-VH gene families, as well as at origins located 3′ and 5′ of the Igh locus. From there, replication forks proceed to replicate the locus until they collide with an oppositely moving fork (predominantly within the termination regions).
The data presented above show that the replication of the Igh locus follows a precise temporal program. However, this program represents a population average. At the level of individual DNA molecules, initiation events do not seem to occur in any particular order, with different origins firing in different molecules. Initiation events were also detected within the portions of the Igh locus that replicate last (e.g., the termination region within PmeI#4 and the region between Cδ and Cγ3; Figure 2A), taking place as the corresponding molecules were at various stages of their replication. Such events are not easily explained by deterministic models of origin activation. A domino activation of origin firing [13] also seems unlikely. The initiation events detected in the population of double-labeled DNA molecules are rarely associated with externally generated replication forks, and when such forks are present their location is tens to hundreds of kilobases from the firing origins (e.g., 40 and 100 kb in the example shown in Figure S2A). Hence, it is possible that Igh origins fire according to stochastic dynamics.
To determine whether the results obtained by SMARD can be quantitatively explained by stochastic origin firing (see definition in Table 1), we used a mathematical formalism and a simulation procedure that we recently developed for this purpose [18]. As briefly described in Materials and Methods, this procedure allows us to fit many of the data collected by SMARD (namely, the distribution of the replication forks, Figure 2B–C, the temporal order of replication, Figure 2D, and the replication time of the restriction fragments, Tr, Figure S2B) to computer-generated curves calculated from a series of rate equations.
In the simplest possible scenario, the curves were calculated assuming that bidirectional origins of replication fire stochastically throughout the Igh locus and the S phase (see Table1 for a precise definition), generating forks that move at a constant speed. In these calculations, the rate of origin firing was the only parameter allowed to vary freely across the Igh locus, while remaining constant in time (meaning that, within a genomic region, initiation events continue to occur at the same rate from the beginning of S phase until the region is replicated in the entire population of cells). It is important to point out that the rigid constrains imposed by this scenario do not accurately reflect physiological conditions. For example, it is known that modest changes in the speed of replication forks can indeed occur along the genome and during the S phase [25]. However, by limiting the number of free variables, this scenario allows us test the stochastic firing hypothesis more stringently.
Strikingly, we found that this simple scenario is sufficient to reproduce all experimental data sets collected by SMARD (Figures 2B–D and S2B–C). This scenario can even reproduce data sets that were not used during the fitting procedure (such as the location of initiation events and fork collisions, the number of molecules containing such events, the average number of events per molecules, and the average speed of replication forks). Since the fit was performed simultaneously for all restriction fragments, the presence of local discrepancies is not particularly surprising (e.g., PmeI #4 in Figures 2D and S2B). This is likely to reflect the constrains imposed by this scenario (e.g., within some portions of the locus fork speed may deviate from the average and origin density may also vary). Nevertheless, the high quality of the fit (reduced chi-square, 1.18) indicates a close match between calculated and experimental data sets. Hence, the results collected by SMARD are fully compatible with the stochastic firing of origins throughout the Igh locus.
In addition to the changes in firing rate along the locus, scenarios involving a larger number of variables were also considered (e.g., allowing for a variable speed of replication forks or for changes in the rate of origin firing during the S phase). In principle, changing these parameters could have a major effect on the replication dynamics of the Igh locus. However, these scenarios improved the quality of the fit only marginally (M.G.G., J.B., and P.N., unpublished observation). This means that variations in parameters other than the firing rate of origins along the genome, while possible, have a limited impact on the replication dynamics and the temporal order of replication of the Igh locus in uncommitted pro-B cells.
The high quality of the fit calculated above indicates that the computer-generated data set provides a good approximation of the firing rate of origins across the Igh locus. Figures 2E and S2D show that the firing rate is very low throughout most of the locus (4.5×10−6 initiation events per kb per minute). Two exceptions are the DH-JH region (55±13.2 kb in size) and of the middle-VH region (281±43.1 kb in size), where the firing rate is up to 77-fold higher. This variation implies that origins located in different parts of the Igh locus have a very different tendency to fire.
Another striking feature of these results is that the firing rate changes abruptly at a few specific locations (transitions), while remaining virtually uniform across large sections of the locus (plateaus). Attempts at fitting bell-shaped curves (which lack both plateaus and sharp transitions) produced fits of lower quality compared to box-shaped curves of the kind shown in Figure 2E (M.G.G., J.B., and P.N., unpublished observation). Therefore, the Igh locus appears to be organized in precisely defined domains where origins have a similar rate of firing (suggesting a significant degree of coordination among the origins of individual domains). We conclude that the temporal order of replication is a consequence of the combined effect of domain sizes and firing rates.
Equation 3, in Materials and Methods, also allows us to determine the efficiency of origin firing for specific sections of the Igh locus, according to the definition provided in Table 1. The values obtained for the DH-JH and middle-VH regions are only 1.1 and 1.5 initiation events, per allele, per S phase (IAS, gray dashed lines in Figure 2E). These regions are known to contain multiple origins of replication (Figure 2A). Hence, most of the origins remain silent during each cycle of replication (firing is inefficient). Notably, the portion of Igh locus spanning the proximal-VH genes produces 0.4 IAS even if its firing rate is about 2 orders of magnitude lower than the DH-JH and middle-VH regions. The nonlinear relationship between firing rate and origin efficiency can be explained by the fact that the proximal-VH genes occupy a very large genomic region and tend to replicate after adjacent portions of the locus (Figure 2D). This provides more time for the origins therein located to fire. Hence, efficiency is not an intrinsic property of individual origins, but rather the result of the distribution of the firing rate throughout the Igh locus.
The computer-generated data set allows us to draw a few additional conclusions. Although Igh origins fire inefficiently, they are responsible for the replication of 87% of the locus, with only minor contributions from external origins (Figure S2F). This means that inefficient origin firing extends beyond the margins of the locus into the surrounding regions. In addition, various portions of the locus have a very high probability of being replicated by forks moving in one particular direction (peaks and valleys in Figure S2G). Since the molecules analyzed in this experiment originated from both Igh alleles, the strong bias in fork direction indicates that both alleles follow a similar replication program. Finally, we calculate that it takes approximately 4 h to complete the replication of this portion of the genome in 100% of the cell population (Figure 2F). Hence, these experiments provide a complete description of DNA replication, within the Igh locus, over a broad portion of the S phase.
Bone marrow pro-B cells from Pax5−/− mice are blocked at the uncommitted stage of differentiation but, unlike Pax5−/−Rag2−/− cells, they can undergo DH-JH recombination. The resulting loss of various sections of the DH-JH origin cluster (the ∼55±13.2 kb region containing the origins with the highest firing rate; Figure 2E) provides us with a tool to study how the firing rate is regulated. Specifically, we can use ex vivo propagation to obtain clonal populations of cells carrying particular DH-JH deletions. Here, we studied a heterozygous pro-B cell clone that carries a 65 kb deletion on its 129/Sv allele and a 25 kb deletion on the C57BL/6 allele. The different sizes of the resulting restriction fragments allowed us to use SMARD to investigate the effect of each deletion on DNA replication.
Consistent with the loss of the entire DH-JH origin cluster, the 65 kb deletion reduced the number of initiation events occurring within fragment PacI#3 to an undetectable level (left portion of Figure 3A and Figure S3B). The direction of replication fork movement (Figure 3B–C) and the uniform slope of the temporal order of replication (Figure 3D) indicate the passive replication of the region by forks originating 3′ of the Igh locus. This is similar to results previously obtained in non-B cells, where the Igh locus is part of a replication timing transition region [2]. Besides for the loss of the DH-JH origins, the rate of origin firing in adjacent portions of the locus remains unchanged (compare PacI#3 and SwaI#2 in Figures 2E and 3A). This rate is very low compared to the origins located at 3′ of the Igh locus. Hence, even if the 5′-end of PacI#3 now replicates 3 h later compared to unrearranged pro-B cells (compare Figure 3F at 280 kb to the corresponding portion of Figure 2F), this delay is not sufficient to produce large numbers of initiation events within PacI#3 and SwaI#2 (infrequent initiation events may still occur within this portion of the Igh locus but their visualization would require the analysis of a much larger sample of double-labeled DNA molecules).
In contrast, the 25 kb deletion removes less than half of the DH-JH origin cluster. On this allele, initiation events and fork collisions continue to occur at the same locations described for Igh locus in germline configuration (compare the right portion of Figure 3A with Figure 2A). However, there is a strong right-to-left bias in the direction of fork movement (Figure 3B–C), and the IdU content shows a steady decrease in the same direction with an inflection point at the site of the remaining DH-JH origins (Figure 3D). These factors indicate the passive replication of the region in a fraction of the cell population. Accordingly, there is a 40% reduction in the number of initiation events occurring within this section of the locus during each S phase (0.7 IAS, Figure 3E), compared to the level detected in unrearranged pro-B cells (1.1 IAS, Figure 2E). These results indicate that the decrease in initiation events is proportional to the reduction in size of the DH-JH origin cluster. Hence, the firing rate of the remaining Igh origins and the location of the firing rate transitions are unaffected by the 25 kb deletion (Figure 3E). This is reminiscent of results obtained by studying deletions of the DHFR locus by 2D-gel electrophoresis (although in that case conclusions were based on measurement of origin efficiency, which is context dependent, and not of firing rate) [26],[27]. We conclude that firing rate of origins is regulated independently in different sections of the locus and that origin activity at one location is not affected by the presence, or absence, of neighboring origins.
In order to study how the replication of the Igh locus changes during cell differentiation, we isolated B cell progenitors from the bone marrow of a Rag2−/− mouse (129/Vs.). These cells efficiently undergo B cell commitment but maintain the Igh locus in germline configuration, which prevents them from developing any further. In these cells, initiation events and fork collisions are distributed across most of the Igh locus (Figure 4A). Only a few of these initiation events are centered at the same genomic location (Figure S4), indicating that the experiment is far from mapping saturation (leading to an underestimate of origin density). This suggests that potentially active origins are present at high density across most of the Igh locus, perhaps every 10–20 kb, although larger gaps may exist at a few locations. Despite the widespread activation of origins, it is still possible to distinguish portions of the locus where replication forks have a preferred direction of movement (e.g., PacI#3; Figure 4B–C) and regions that replicate first and last (Figure 4D). Thus, our results indicate that the level of origin activity is not uniform across the Igh locus and point to major differences in DNA replication between committed and uncommitted pro-B cells.
Once more, we found that the data obtained by SMARD are fully compatible with the stochastic firing of origins (reduced chi-square, 1.03; Figure S4C). The nearly perfect fit confirms that the temporal order of replication is mostly determined by the regulation of a single variable along the locus (the firing rate of origins). Computer calculations also show that the rate of origin firing is not uniform (blue line in Figure 4E). A comparison with the results obtained for uncommitted pro-B cells (red dotted line) reveals that B cell commitment is accompanied by changes in the rate of origin firing across most of the Igh locus. However, the largest changes (whether positive or negative) occur in regions where a previous genomic screening identified the presence of Pax5 binding sites in committed pro-B cells [28]. For example, the firing rate increases nearly 50-fold across the CH-3′RR region (∼216 kb in size and containing two Pax5 binding sites) and decreases 10-fold throughout the DH-JH origin cluster (∼55 kb in size and containing one Pax5 binding site). These results indicate that the rate of origin firing is the parameter that is being regulated during cell differentiation. In addition, they suggest that Pax5 participates in regulating the firing rate of origins during B cell commitment (although the number of known binding sites for this factor is far smaller than the number of potential origins affected by its expression).
As a result of B cell commitment, changes in origin efficiency were also observed throughout the Igh locus (compare gray dotted lines in Figures 2E and 4E). However, the overall efficiency of origins across the 1.4 Mb region increases only from 3.2 IAS in uncommitted pro-B cells (Figure S2D) to 7.4 IAS in uncommitted pro-Bs (Figure S4C). Thus, the firing rate can increase by 1–2 orders of magnitude across most of the Igh locus but produce variations of only 2.3-fold in the total number of initiation events (and replication complexes) involved in the replication of the locus. Given the high density of potential origins throughout the Igh locus, these results also suggest that the firing efficiency of individual origins is below 10%. Therefore, following B cell commitment, origins continue to fire inefficiently even if the firing rate increases across most of the Igh locus.
Reconstituting Pax5 expression in Pax5−/− pro-B cells induces B-cell commitment [20],[29] and can be used to study the role of this protein in origin regulation. For this purpose, we transduced bone marrow pro-B cells from a Pax5−/− mouse (129/Sv-C57BL/6) with a retroviral vector containing the expression cassette Pax5ER-IRES-GFP [30],[31]. We then sorted and expanded the GFP+ cells to obtain a polyclonal population expressing Pax5ER (KO-Pax5ER pro-B cells; Figure S5A–B). This protein is the fusion product of Pax5 and the hormone-binding domain of the estrogen receptor, which becomes biologically active in the presence of 4-hydroxy-tamoxifen (4-OHT). In KO-Pax5ER pro-B cells, 4-OHT is able to induce commitment-specific changes in the methylation of the 3′RR DNA, indicating that Pax5ER is able to interact with at least one of the Pax5 binding sites of the CH-3′RR region [32]. In KO-Pax5ER pro-B cells, the occurrence of B-cell commitment was monitored by the Pax5-dependent expression of the cell-surface-marker CD19. Before the addition of 4-OHT, we consistently found that only 4% of KO-Pax5ER pro-B cells are CD19+ (Figures 5A and S5C). However, 65%–90% of cells become CD19+ after the addition of 4-OHT. Thus, in our inducible system, the activity of Pax5ER is modest in the absence of 4-OHT but increases dramatically after induction, leading to B cell commitment.
Can the induction of Pax5ER increase the firing rate of origins within the CH-3′RR region? To answer this question, we studied KO-Pax5ER pro-B cells before and after induction with 4-OHT for 28 h (see labeling scheme in Figure 5B). This induction time was chosen because it allows enough time for the synthesis and turnover of the gene products regulated by Pax5 and for cells to become fully committed to the B lineage [30]. In the absence of 4-OHT, we detected a limited number of initiation events within the CH-3′RR region, which is consistent with the modest activation of Pax5ER described above (first and third column in Figure 5C). However, this portion of the locus continues to be passively replicated in the majority of cells, as indicated by the replication fork distribution and the temporal order of replication (Figures 5D–F and S5H–I). In contrast, 4-OHT induction profoundly alters all aspects of DNA replication within the CH-3′RR region, suggesting a strong increase in origin activity (second and fourth column in Figures 5C–F and S5H–I). Numerical calculations confirm that the firing rate of origins reaches the level detected in committed pro-B cells (compare the 129/Sv results in Figures 5G and 4E). Hence, inducing B cell commitment in vitro produces changes in the firing rate of Igh origins that are similar to those observed in pro-B cells isolated from mice.
Additional experiments confirmed that these effects are specific. For example, the firing rate of CH-3′RR origins does not increase when 4-OHT is provided to uncommitted pro-B cells that do not express Pax5ER (Table S1). In addition, the induction of Pax5ER in KO-Pax5ER pro-B cells does not affect origin activity broadly and non-specifically. For example, we did not observe significant changes in cell proliferation and cell-cycle profile (Figure S5L) or the appearance of markers of DNA damage and DNA damage checkpoint activation (Figure S5M). We conclude that the change in origin activity observed in KO-Pax5ER pro-B cells results from the activation of Pax5ER and the induction of B cell commitment.
B cell commitment involves changes in the expression of hundreds of tissue-specific genes [33], raising the question of whether Pax5 is directly responsible for regulating the firing rate of origins. After the addition of 4-OHT, Pax5ER requires only a few minutes to translocate from the cytoplasm to the nucleus, where it changes the expression of its target genes over a period of several hours [30]. We suspected that the firing rate would change very slowly if its regulation depended on the activation or silencing of these genes. Therefore, we decided to determine the speed at which the firing rate increases within the CH-3′RR region following the addition of 4-OHT to KO-Pax5ER pro-B cells.
The labeling scheme for the time-course experiment is presented in Figure 6A. Our experimental design takes advantage of the fact that SMARD detects the occurrence of initiation events in the population double-labeled DNA molecules. These molecules begin their replication during a precise interval of time Tr that preceded the label switch, meaning that they originate from cells that entered S phase sometimes before each molecule started replicating [2]. As a result, we can study how DNA replication initiates, progresses, and terminates across the genomic region of interest, knowing that these events occurred in cells that crossed the G1/S transition during a specific window of time from the addition of 4-OHT (this procedure also avoids cell synchronization and experimental manipulations that may generate experimental artifacts).
Surprisingly, we found that the replication of the CH-3′RR region changes very rapidly after the addition of 4-OHT. For all induction times between 1 and 24 h, the distributions of replication forks and the temporal orders of replication are similar to the experiment performed at 28 h (compare Figure 6B–E with Figure 5D–F). To accurately determine how the firing rate changes over time, we simultaneously fit all experimental data sets obtained for the 129/Sv allele (including those presented in Figure 5). This resulted in a reduced chi-square of 1.02 (Figure S6A–C). As summarized in Figure 6H, the firing rate of the CH-3′RR origins reaches a level comparable to committed pro-B cells within the first hour of induction. Given that this induction time is similar to the replication time of the restriction fragment (approximately 1 h; Table S1), we conclude that the firing rate of origins began to increase within minutes after the addition of 4-OHT. This rapidity implies that the changes in firing rate do not depend on modifications in gene expression and are likely to be caused by a different mechanism (e.g., transcription, chromatin remodeling, etc.).
But how is the regulation of the firing rate of origins achieved during cell differentiation? It is important to note that the population of double-labeled DNA molecules originates from cells that are in S phase at the time of the label switch, having crossed the G1/S transition 0 to 80′ before each molecule started replicating (Figure S7). This means that the 7-fold increase in the firing rate observed after 1 h of induction (Figure 6H) takes place in cells that were first exposed to 4-OHT at a time in the cell cycle close to the G1/S transition (between 60′ before G1/S and 80′ into S phase; Figure S7). We know that, in mammalian cells, the replicative helicase is loaded between the end of mitosis and early G1, completing the process of origin licensing and the formation of pre-replication complexes (pre-RCs) as reviewed in [34]. We can therefore conclude that Pax5 controls the rate of origin firing, and the temporal order of replication, by affecting a stage in origin activation downstream of pre-RC assembly. A corollary to this conclusion is that pre-RCs can form across the CH-3′RR region even in the absence of 4-OHT. Therefore, pre-RCs can form efficiently even in genomic regions characterized by a low firing rate of origins.
Are the effects of Pax5ER induction limited to the late stages of origin activation? Figure 6H shows that the firing rate of the CH-3′RR origins changes over time, peaking at 6 h induction and returning to the level of committed-pro-B cells thereafter (at 12, 24, and 28 h). We know that the cell cycle of KO-Pax5ER pro-B cells is approximately 14 h and that their G1 lasts about 6 h (A.D. and P.N., unpublished observation). This means that the 14-fold increase in firing rate that we observe during the first 6 h of induction is linked to events taking place during the preceding G1 (mostly after the assembly of the pre-RCs). In contrast, the firing rate measured at 12 h of induction is the result of events taking place not only after pre-RC assembly but also before (during the previous M, G2, and S phases). Since this value is higher than the pre-induction level but 63% lower than the firing rate measured at 6 h (Figure 6H), we conclude that some inhibition of the firing rate takes place during the previous M, G2, and S phases. However, this inhibition is insufficient to fully compensate for the increase in firing rate that takes place during G1. We conclude that the firing rate of the CH-3′RR origins is affected by the induction of Pax5ER at various stages of the cell cycle but is dominated by events that occur after pre-RC assembly.
Are the changes in firing rate induced by Pax5 stable? Or do they need to be reestablished at every cell division? To address this question, we took advantage of the fact that the activation of Pax5ER can be slowly reversed by OHT withdrawal. We first grew KO-Pax5ER pro-B cells in the presence of 4-OHT for 36 h to induce full B cell commitment and to increase the rate of origin firing within the CH-3′RR region. We then isolated the population of committed pro-B cells (CD19+ fraction), expanding it in the absence of 4-OHT for 48 h before labeling the replicating DNA and performing SMARD. Results indicate that, following 4-OHT withdrawal, origin activity and the temporal order of replication revert to pre-induction levels (right portion of Figure 6 and Figure S6). Therefore, the activity of Pax5 is continuously required to maintain high rates of origin firing within the CH-3′RR region (these changes are not epigenetically inherited). This also means that a low firing rate represents the default state for these Igh origins, an observation that is consistent with the low level of origin activity observed in all types of non-B cells examined to date [2].
In this study, we provide a model of origin activation that can explain multiple parameters of DNA replication initiation, progression, and termination throughout a significant portion of the S phase. Parameters, such as the location of initiation events and fork collisions, the efficiency of origin firing, the number and distribution of replication forks, the timing of origin firing, and the temporal order of replication can be quantitatively explained by the interplay between two independent variables: the speed of replication forks and the firing rate of origins. Only the latter varies extensively along the locus representing the dominant variable. The firing rate of origins is also the main variable that is being regulated during cell differentiation.
For each individual experiment, both independent variables can be considered constant in time (an approximation that is sufficiently accurate to describe the events taking place within the Igh locus during the first few hours of S phase but does not exclude the possibility of variations over the entire S phase). Other aspects of DNA replication behave as dependent variables that do not require dedicated regulatory pathways. This means that the extensive literature concerning the regulation of origin efficiency and the timing of origin firing can be reexamined in terms of changes in the firing rate of origins. For example, the apparent discrepancies mentioned in the introduction regarding some earlier studies [2],[3],[5] could be reconciled by changes in the firing rate of origins during cell differentiation. Hence, our model can provide a simple conceptual framework to explain the dynamics of DNA replication at a specific mammalian locus.
Our study suggests that potentially active origins are present at high density throughout the Igh locus in both committed and uncommitted pro-B cells (Figures 2A, 4A, S2E, S4D). Active origins were detected even within portions of the locus that replicate last (e.g., PmeI#4 in Figure 2). This interpretation is consistent with recent observations showing the presence of active origins of replication in portions of the human IGH locus previously thought to contain only silent ones [13],[35]. However, based on the analysis of a larger set of parameters of DNA replication (representing all stages of replication), we showed that the initiation events detected experimentally can be explained by an entirely stochastic mechanisms of activation (meaning that origin firing can occur in any unreplicated portion of the Igh locus, at any time during the S phase).
Our study also suggests a significant degree of coordination in the level of origin activity across large sections of the Igh locus (tens to hundreds of kilobases in size). Within these domains, the firing rate appears to be relatively uniform and is modulated synchronously, over a broad range, as a result of changes in cell differentiation. The temporal order of replication is determined by the combined effect of variable domain sizes and firing rates. Consistent with studies performed at other mammalian loci [26], we showed that groups of origins can be deleted from the Igh locus without significantly affecting the firing rate of the remaining origins, or the location of the firing rate transitions (these two qualities of chromosomal domains are not significantly affected by the activity of individual origins of replication). In contrast, clustering of yeast origins with similar characteristics has been observed in various studies, but the firing rate of each origin is different and appears to be individually regulated [15],[36],[37]. Hence, our results point to significant differences in the way origin activity is regulated in the two organisms, perhaps reflecting a role of domain boundaries in the modulation of mammalian DNA replication (see below) [38].
During B cell commitment, the largest changes in firing rate occur in Igh domains that contain Pax5 binding sites and depend on the continuous expression of this protein. This suggests that cell-specific factors participate in controlling the temporal order of replication by regulating the firing rate of origins within specific chromosomal domains. Pax5 modifies the firing rate of origins using at least two separate mechanisms that operate at different times during the cell cycle. One mechanism acts upstream of pre-RC assembly. It produces a stable decrease in the peak rate of origin firing across the CH-3′RR region, suggesting it affects a limiting step in origin activation (perhaps the formation of pre-RCs). However, the limited magnitude of the change means that this mechanism can only marginally affect the temporal order of replication. In contrast, within the CH-3′RR region, the rate of origin firing is mostly determined by events occurring in late G1/early S (after the loading of ORC and MCM proteins). This second mechanism produces a rapid increase in the firing rate of origins that continues for a few hours after the induction of Pax5ER (cells need to be exposed to Pax5 for most of G1 in order to reach the maximum level of origin activity). These results are consistent with recent observations suggesting that temporal order of replication is strongly affected by events downstream of the assembly of pre-RCs [39],[40].
Given that Pax5 affects origin activity in multiple ways, dissecting its activity will require additional studies. However, it is important to point out that chromosomal loops comparable to the CH-3′RR, DH-JH, proximal-VH, and middle-VH domains have been recently identified by 4C and 3C assays in C57BL/6 committed pro-B cells [41]. This suggests that the sharp transitions in firing rate detected in our study may occur at the bases of chromosomal loops; with plateaus making the body of each loop (adjacent loops with a similar firing rate may not be distinguishable). Interestingly, the loop organization of the Igh locus can change during B cell differentiation [42]. In addition, Pax5 has been shown to alter the chromosomal architecture of various portions of the locus, inducing locus contraction [43], and promoting chromatin looping within the CH-3′RR region [44]. Hence, it seems plausible that chromosomal topology and domain boundaries play a role in modulating the firing rate of origins within the Igh locus. In this context, the role of Pax5 could be helping to form new chromosomal loops or altering the topology of preexisting ones. Similar mechanisms could provide a way to affect the activity of multiple origins of replication across large sections of the genome (perhaps by changing their accessibility to rate-limiting factors or by facilitating DNA unwinding at origins). In fact, preliminary evidence indicates that the deletion of short regulatory elements can affect the firing rate of origins across entire Igh domains (P.N., unpublished result).
One of the appeals of the model of origin regulation presented above is that physiological variations in a range of parameters of DNA replication can automatically be compensated without necessarily requiring the intervention of checkpoints. For example, a localized decrease in replication fork progression will provide more time for origins to fire, resulting in the activation of additional origins (proportionally to their firing rate). This can limit the probability that sections of the genome will remain unreplicated for extended periods of time. A similar compensatory mechanism can operate at the genomic level as a result of cell-to-cell variations in origin licensing, or due to factors that are rate-limiting for origin firing [45]. Even in the presence of a drastic reduction in the global level of origin activity (e.g., following the activation of DNA damage checkpoints), residual initiation events are expected to continue to occur according to the differential in firing rate of origins in the various portions of the genome (therefore preserving the temporal order of replication). These predictions are consistent with experimental observations such as those referring to the activation of dormant origins during replicative stress [46],[47]. Thus, this model of origin regulation can help us to explain experimental observations that are not fully understood.
Short-term bone marrow pro-B cell cultures were prepared as previously described [48]. Briefly, bone marrow cells were isolated from an 8-wk-old Rag2−/− mouse (129/Sv) and plated at 106 cells/ml in RPMI supplemented with 5% heat-inactivated fetal calf serum (FCS; Invitrogen), 2 mM L-glutamine, 1× penicillin-streptomycin, 50 μM 2-mercaptoethanol, and 10 ng/ml IL-7 (R&D). Cells were kept in culture for 7 d before labeling the replicating DNA with halogenated nucleotides. FACS analysis performed before DNA labeling indicated that more than 99% of the cells in culture were CD19+ (corresponding to committed pro-B cells). In contrast, pro-B cells from Pax5−/− and Pax5−/− Rag2−/− mice can be grown ex vivo for extended periods of time [49]. Long-term uncommitted pro-B cell cultures were prepared from the bone marrows of 1-wk-old Pax5−/− Rag2−/− (129/Sv) and Pax5−/− (129/Sv-C57BL/6) mice, as previously described [20]. These cells were expanded ex vivo on irradiated feeder cells in IMDM medium supplemented with 2% heat inactivated FCS, 0.15% primatone, 2 mM L-glutamine, 1× penicillin-streptomycin, 50 μM 2-mercaptoethanol, and 1 ng/ml IL-7. The mouse stromal cell line ST2 was used to prepare the feeder layers for pro-B cell cultures. These cells were grown in IMDM medium supplemented with 10% heat-inactivated FCS, 2 mM L-glutamine, 1× penicillin-streptomycin until 60% confluent, then irradiated (11 Gy) and used as feeders. All pro-B cell cultures were maintained as polyclonal populations. However, a clonal population was also derived from the bone marrow of a Pax5−/− mouse (129/Sv-C57BL/6) and used for some of the experiments presented in this study.
The location of Pax5 binding sites in committed pro-B cells, as well as their DNA sequence, chromatin features, and transcriptional activity, is presented for the entire Igh locus in a previous publication [28]. Briefly, within the 1.4 Mb region considered in the current study (from strain 129/Sv), the DNA sequences for six of these sites are present at 11 different locations (marked from 1 to 11 in Figure 4E). The sites marked from 1 to 4 have unique DNA sequences that align to the reference mouse genome (strain C57BL/6) at the following positions: assembly mm8_ chr12:<113676000113676290; chr12:113768900-113769250; chr12:113895270-113895500; and chr12:113972530-113972740. In contrast, sites 5 and 11 have the same DNA sequence and align to the reference strain at position mm8_chr12: 114630500-114630900, while sites 6–10 align at position mm8_chr12:114541600-114542000. With the exception of site number 4, all other sites display active chromatin marks in committed pro-B cells (H3K9ac and H3K4me2 [28]). However, these marks depend on Pax5 binding only at sites 1 and 2. Hence, there is no simple correlation between the firing rate of origins and the presence of active chromatin marks. Moreover, unlike Pax5 binding sites located in the distal portion of the Igh locus, none of these sites is associated with PAIR elements and Pax5-dependent antisense transcription [28].
A portion of the uncommitted pro-B cell culture was transduced with the retroviral vector Pax5ER-IRES-GFP that we previously used to reconstitute Pax5 expression in uncommitted pro-B cells [20],[30],[31],[50]. In these experiments, GFP expression allowed us to identify the cells successfully transduced with the retroviral vector. The populations of GFP+ cells were isolated by FACS in sterile conditions using a FACS MoFlo apparatus (sorting buffer: 1× PBS, 4% FBS, 1× penicillin-streptomycin). We have previously shown that the fusion protein Pax5ER can respond within minutes of the addition of 4-OHT (1 μM), resulting in the efficient expression of target genes (usually within a few hours from induction [30]). Pax5ER expression was monitored by immunoblot using antibodies raised against the amino terminus of Pax5 (Santa Cruz), or against the carboxy terminus of the estrogen receptor alpha (Santa Cruz), to determine the total level of expression in 3×105 cells, using the expression of beta-actin as a reference (Santa Cruz). The induction of Pax5ER was monitored by FACS using anti-mouse-CD19 antibody (Pharmingen). Following 4-OHT induction, CD19+ and CD19− pro-B cells were isolated using an AutoMACS cell separator and anti-CD19 conjugated magnetic beads (Miltenyi), according to the manufacturer's procedure. Cell-cycle analysis before and after 4-OHT induction was performed by FACS following propidium-iodide staining of the cells (50 mg/L) in hypotonic buffer (0.1% sodium citrate). The lack of DNA-damage checkpoint activation was determined on total cell lysates from 3×105 cells in the presence of a protease and phosphatase-inhibitors cocktail (Thermo Scientific), using antibodies raised against p53 (Cell Signaling), phospho-p53, Ser15 (Cell Signaling), phospho-histone H2A.X, Ser139 (Cell Signaling), phospho-Chk1, Ser345 (Cell Signaling), phospho-Rad17, Ser645 (Cell Signaling), and phospho-ATM, Ser1981 (Santa Cruz).
A detailed procedure for SMARD is presented in previous publications [17],[21],[22]. Briefly, cells in the exponential phase of growth were sequentially labeled with IdU and CldU (25 μM each) for periods of time between 3 and 4 h at a density between 3×105 and 8×105 cells/ml. The duration of each of the labeling periods was chosen to be longer than the replication time of the fragments analyzed but short enough to prevent multiple rounds of replication according to the principles described in previous publications [17],[21]. Total genomic DNA was prepared in agarose plugs (about 106 cells/plug) and digested with either PacI, SwaI, or PmeI (40 units/plug). Restriction fragments were purified by pulsed-field gel electrophoresis, recovered by agarase digestion, and stretched on silanized microscope slides by capillary action. During SMARD, the restriction fragments of interest were detected with the following probes: SwaI #2 was detected with several PCR products spanning the positions 14,606–25,310 and 35,667–46,206 of the Igh clone BAC199M11 (GeneBank: AF450245), together with a plasmid clone containing the switch region γ1 (pγ1/EH10; [51],[52]); PacI #3 was detected with the plasmids UUGC2M0237A15 and UUGC2M0215F07 (GeneBank: AZ966978, AZ966626/AZ951255, AZ950883), and with PCR products spanning, respectively, positions 9530909–9543550, 9560850–9571463, and 9644195–9652184 in the mouse chromosome 12 contig (GeneBank: NT_039553); PmeI #4 was detected with PCR products spanning, respectively, positions 1204509–1206646, 1206954–1208600, 1209395–1211337, and 1212606–1213479 in GeneBank: AJ851868; and PmeI #5 was detected with the plasmids UUGC1M0532J03, UUGC1M0259G14, UUGC2M0265N20, UUGC1M0166E11, and UUGC2M0258M14 (spanning positions 211454–219261, 413828–421730, 591439–598359, 631282–640007, and 647267-663409 in GeneBank: AJ851868). The same probes were used for the detection of both the 129/Sv and the C57BL/6 alleles. Microscopy image alignment to the maps of the restriction fragments and data analysis were performed as previously described [22]. Only molecules unbroken, fully-substituted with halogenated nucleotides, and sufficiency stretched to be aligned were considered.
In order to determine how origin firing is regulated in mammalian cells, we developed a new mathematical formalism and a simulation procedure which allow us to rigorously analyze the experimental results obtained by SMARD. The numerical simulations used in this study allow us to reproduce SMARD data sets from any given known replication scenario (i.e., a defined firing rate and fork speed), making it possible for us to study the statistics of the experimental data sets. As reported in the Results section, we performed numerical simulations of all experiments presented in this study. The simulation procedure is described in detail in reference [18] but can be summarized as follows:
In principle, numerical simulations could also be used to determine the most likely set of parameters responsible for generating specific experimental data sets obtained by SMARD. However, the need to explore a large number of variables makes this fitting procedure impractical and time consuming when the replication scenario is not pre-determined. Therefore, we also developed an analytical formalism that allows us to rapidly model the results obtained by SMARD [18]. The formalism leads to rate equations that describe changes in the number of replication forks as a function of both the location along the genome (x) and the time elapsed since the beginning of S phase (t). More specifically, we define a set of three coupled partial differential equations that describe the space-time evolution of the replication fraction, f(x, t), and the two densities of forks, ρ±(x, t), moving rightward (+) and leftward (–). Equation 1 gives the rate of change of the probability that a specific location x along the genome is replicated at time t,:where v±(x, t) is the speed of the replication fork at the position x at time t. This first equation states that the rate of replication of a specific location is proportional to the number of forks at the same location times their respective replication time. The rate of change of the fork densities (i.e., the number of forks moving in a given direction per kb) is given by Equation 2:where I(x, t) is the firing rate that corresponds to the number of initiation events per time per unreplicated DNA length. The two terms on the right-hand side of the latter equation represent initiation and termination events, respectively, while the terms on the left-hand side express the space-time-propagation of the replication forks. This second equation defines the rate of change of the probability of observing a replication fork at a specific location and time in the form of a classic transport equation. If both the firing rate and the fork speed space-time functions are known, these equations can be numerically integrated to obtain f(x, t) and ρ±(x, t). Once our set of equations is solved, the results can be converted to profiles of IdU content and fork distributions, as they are observed in the SMARD experiment (for details, see reference [18]). The solution can also be used to calculate the average number of initiation events occurring within a specific section of the locus, per allele, during each S phase as:where x1 and x2 are the boundaries of the region of interest (Equation 3). This calculation approach provides us with a faster and more precise way to describe the average replication kinetics for a specific genomic region, and its results were validated using simulations of known replication scenarios, as described in the previous section. In addition, simulations are still needed in order to study the statistical fluctuations of f(x, t) and ρ±(x, t).
This calculation technique was then used to fit the experimental data sets to parameterized functional forms for I(x, t) and v±(x, t). This means that starting with a series of parameters defining both the firing and the velocity profiles, we used our calculation technique to produce the corresponding SMARD-like results. The parameters were then iteratively adjusted in order to obtain the best fit between the calculated values and the experimental measurements. For the fits performed in this study, we assumed homogeneous fork speeds throughout both space and time (i.e., v±(x, t) = v for left- and right-moving forks). We also assumed that the firing rate is constant throughout time but spatially inhomogeneous, I(x, t) = I(x). A larger firing rate implies a shorter average replication time and vice versa. Time-dependent firing rates were considered but were not needed to obtain a satisfactory representation of the data. In other words, at the hundred-kilobase-to-megabase scale considered in this study, the replication timing is dominated by spatial variations in the firing rate. Our calculation technique also provided us the distribution of times required to duplicate a given section of the genome.
The SMARD data used in the fit are shown as empty symbols in Figures 2B–D and S2B. For each experiment, we performed a global fit of the two replication fork distributions (solid lines in Figure 2B–C) of the IdU content (solid lines in Figure 2D), and of the replication times for all restriction fragments (filled symbols in Figure S2B). Error bars on the experimental data were obtained from numerical simulations where each simulated experiment uses the best-fit parameters to collect the same number of labeled DNA fragments analyzed in the real experiment. Figure S2D shows the free parameters obtained from the fits, with the firing rate of origins I(x) also graphically illustrated in Figure 2E. For the purpose of the calculations, the firing rate was considered to be composed by a constant background of initiation events (one parameter), to which additional events are added where initiations are more likely to occur (active origins clusters). For simplicity, these initiation clusters were assumed to have the shape of a rounded box defined by three parameters (position and width of the cluster, and the firing rate within the cluster). While this is likely to represent an oversimplification of the conditions existing in vivo, we found that more complex shapes were not required to obtain meaningful representations of the experimental data sets. In contrast, Gaussian curves were tested but resulted in fits that do not accurately reproduce the observed number of initiation events and their location. Hence, curves where the firing rate changes sharply along the genome provide a better representation of the experimental data sets than curves where the firing rate changes gradually. In Figure 2E, two additional parameters show the level of origin firing outside the genomic region investigated. The value of these parameters may be seen as an effective constant firing rate profile outside the studied region. Finally, one parameter was used to fit the velocity of fork propagation (meaning that the speed of replication forks was assumed to be constant throughout the region analyzed). Among the other results of the fit, Figure S2F shows the probability for each section of the genome to be replicated by forks that originated within the modeled region, while Figure S2G shows the probability of being replicated by replication forks moving leftward or rightward. Finally, Figure 2F shows the average replication kinetics of the modeled genomic region. The color value on this space-time graph indicates the probability for each genomic position to replicate at a specific time in S phase. The transition from unreplicated (orange) to replicated (blue) occurs at a different speed in a different portion of the genome. Three contour lines are shown to illustrate which regions replicate faster than others (showing when 10%, 50%, and 90% of cells have replicated the corresponding portion of the genome).
The images collected by SMARD represent all stages of DNA replication. As a result, the number of initiation and termination events scored on individual DNA molecules (Table S1, columns g and h) is lower than the actual number of events taking place during every replication cycle. For example, initiation events can be easily scored when the size of the IdU patch is small, but no longer so at more advanced stages of replication (when one of the outward-moving forks runs off the fragment or coalesces with another fork). In contrast, our calculation method determines the actual number of initiation events and fork collisions occurring, on average, every replication cycle. Hence, in order to compare the outcomes of these two procedures, we performed additional simulations of the SMARD experiments to determine how many of initiations/collisions would be observed experimentally from the results of each fit. Figure S2C shows that the number of events obtained from the fit is comparable to the number observed experimentally.
|
10.1371/journal.pntd.0001673 | Sporotrichosis in Sub-Himalayan India | Sporotrichosis is endemic in the Sub-Himalayan belt, which ranges from the northern to the north-eastern Indian subcontinent. Similar to many parts of the developing world, sporotrichosis is commonly recognized clinically in this region however consolidated epidemiological data is lacking. We report epidemiological, clinical and microbiological data from a hundred culture positive cases of sporotrichosis. Out of 305 clinically suspicious cases of sporotrichosis, a total of 100 isolates were identified as Sporothrix schenckii species complex (S. schenckii) on culture. Out of the culture proven cases 71% of the cases presented with lymphocutaneous type of lesions while 28% had fixed localized type and 1% had disseminated sporotrichosis. Presentation with lesions on hands was most frequently seen in 32% with arm (23%) and face (21%) in that sequence. The male to female ratio was 1∶1.27. Age ranged from 1 ½ years to 88 years. Mean age was 43.25 years. Disease was predominantly seen in the fourth to sixth decade of life with 58% cases between 31 and 60 years of age. Since the first report from the region there has been a steady rise in the number of cases of sporotrichosis. Seasonal trends reveal that most of the patients visited for consultation in the beginning of the year between March and April. This is the first study, from the most endemic region of the Sub-Himalayan belt, to delve into epidemiological and clinical details of such a large number of culture proven cases over a period of more than eighteen years which would help in the understanding of the local disease pattern of sporotrichosis.
| Sporotrichosis is a sub-acute or chronic granulomatous fungal infection involving mainly the skin and subcutaneous tissue with neighbouring lymphatics. It is caused by thermally dimorphic fungus, Sporothrix schenckii, which is prevalent worldwide. Consequent to trauma, the fungus establishes itself in skin and subcutaneous tissue of gardeners, forestry workers, farmers, carpenters, and others who are involved in outdoor activities. Sporotrichosis is the most frequently encountered sub-cutaneous mycosis in the sub-Himalayan belt and all previous reports are based on clinical review. It is imperative to understand the epidemiology of sporotrichosis as an accurate diagnosis, and prompt initiation of the appropriate treatment would prevent chronic debility. This is the first extensive compilation of culture proven cases from Himachal Pradesh in India. It thus justifies our interest in analysing data from our institute seen over 18 years and 7 months, with the objective to compare the demographic factors, briefing the clinical aspects and the culture characteristics of the isolates obtained from the clinical material.
| Sporotrichosis is a sub-acute or chronic granulomatous mycotic infection involving primarily the skin and subcutaneous tissue with neighbouring lymphatics. It is caused by thermo-dimorphic fungus, Sporothrix schenckii which is prevalent worldwide growing on timber, sphagnum moss, plant detritus, soil, etc. Consequent to trauma, the fungus establishes itself in skin and subcutaneous tissue of gardeners, forestry workers, farmers, carpenters, and others who are involved in outdoor activities. S. schenckii is known to thrive at high humidity of 92–100% and a mean temperature ranging between 25°C–37°C [1], [2]. An uncommon mode of transmission includes bites and scratches from animals [3].
The disease characteristically occurs sporadically though large outbreaks have been reported. Since its first recognition by Benjamin Schenck in 1898 from Johns Hopkins Hospital in Baltimore, most cases have been reported from tropical and sub-tropical countries like Mexico in North America, Brazil and Peru in South America and Japan [1], [3], [4]. The first case in India was reported from Assam by Ghosh in 1932, followed by majority reports from Assam, West Bengal and Andhra Pradesh. Recent reports are from Chandigarh, Delhi, Sikkim, Uttar Pradesh, Punjab and Himachal Pradesh [1], [5]–[8].
Himachal is a small North Western hilly state of India with altitudes ranging between 350–6975 meters above the sea level [9]. The geo-climatic conditions are conducive for growth of S. schenckii as a saprophyte in the environment. Singh et al. described the first case of sporotrichosis from Himachal Pradesh in 1980 [6]. Since then, a large number of cases have been reported [5], [6], [10]–[13].
In 1908, Splendore identified asteroid bodies which are stellate forms consisting of a central spore surrounded by a mass of eosinophilic material in human tissue and named the fungus S. asteroides. Small oval cells resembling cigars in the tissue sections were described by de Beurmann and Gougerot [14]. Both these forms are rarely seen in direct microscopy or histopathological examination of the tissue. Serodiagnosis using complement fixation test, latex agglutination, precipitation test, fluorescence techniques or intradermal sporotrichin skin test, do not distinguish between cutaneous infections and uninfected individuals in endemic areas. Fungal culture remains the gold standard for establishing a definitive diagnosis of the disease [4], [15].
Sporotrichosis is the most frequently encountered sub-cutaneous mycosis in Himachal and all previous reports are based on clinical review. This is the first extensive compilation of culture proven cases from Himachal Pradesh in India. It thus justifies our interest in analysing data from our institute seen over 18 years and 7 months, with the objective to compare the demographic factors briefing the clinical aspects and the culture characteristics of the isolates obtained from the clinical material.
A retrospective review of mycological records of culture proven sporotrichosis cases since January 1992 till July 2010 over a duration of 18 years and 7 months was conducted in our institute. Permission was taken from the Heads of Departments of Microbiology and Dermatology to access the records as was the study protocol. The demographic data of all patients' were noted. All the patients' data was anonymized. All samples were collected after taking written informed consent from the patients. This was noted onto the biopsy register maintained in the minor operation theatre in the Department of Dermatology where samples are collected. The skin biopsies, scrapings and pus samples submitted with the clinical possibility of sporotrichosis were processed in the routine mycology laboratory for direct microscopy and culture [2]. Demographic data of patients', whose diagnosis of sporotrichosis was established on the basis of fungal culture, were analysed using statistical methods (Chi-square method).
Total number of samples received in the mycology laboratory of the Microbiology department over duration of 18 years and 7 months was 2356. Out of these, 305(12.94%) samples were received with the clinical possibility of sporotrichosis. A total of 100 isolates were identified as S. schenckii on the basis of growth characteristics, macroscopic and microscopic morphology, and thermal dimorphism. Out of the hundred culture positive cases 71% cases presented with lymphocutaneous type, 28% had fixed localized lesions (Figure 1) and 1.0% of disseminated cutaneous form. The period prevalence of clinically suspected samples received for laboratory confirmation was 12.9%. However, etiology on basis of culture was established in 100 patients giving laboratory prevalence of 4.24%. Presentation with lesions on hands was most frequently seen in 32%, with arm (23%) and face (21%) in that sequence. Other presentations included 2 cases each with lesions on neck and abdomen and 1 case each with rare presentations on breast and disseminated sporotrichosis.
Out of the 100 culture proven cases, 56 patients were females and 44 were males. The male to female ratio was 1∶1.27. Age ranged from 1 ½ years, the youngest child to 88 years, the oldest male patient. Mean age was 43.25 years. Seven children were <15 years old representing 4.21% of the study group. Although the disease was predominantly seen in the fourth to sixth decade of life, with 58% cases between 31 and 60 years of age, none of the age groups (decades) was significantly higher, [P value>0.05]. The average number of cases was 5.4 per year. To evaluate annual trends, the study period was compared in three parts. The number of culture proven cases in the first (1992–1997), second (1998–2003) and third (2004–July 2010) parts were 2, 24 and 74 respectively, giving a ratio of 1∶12∶37 (Figure 2). Seasonal trends reveal that most of the patients visited for consultation in the beginning of the year, with a highly significant number of cases presenting in March and April clubbed together [P value = 0.0002] with 63% of the visits occurring from February to June (Figure 3). In the present study, 90% of cases had agriculture as their main or spare-time job probably exposing them to contaminating injuries. Out of 100 cases, 61 were residents of rural areas subjecting them to environmental exposure. History of trauma with thorns, hay, wood splinters, needles, blades or injury due to fall was given in 47 patients out of which samples of 24 (51.06%) cases showed growth of S. schenckii.
Direct microscopy of one specimen revealed Asteroid bodies (1%) whereas no sample showed yeast cells. Only seven samples were assessed by PAS staining but results were inconclusive. Fungal cultures were observed twice a week up to 30 days and morphological appearance of characteristic growth ascertained using micro-slide culture technique. Growth characteristics showed few mm to one cm white yeast like colony in 3–6 days, yeast like, buff mycelial form in 7–10 days, buff mycelial form to black leathery growth in two weeks and black mycelial, leathery growth after 2–3 weeks. Retrospective data of growth characteristics of 79 isolates were retrievable. In this study, we observed that 90.6% (68/75) of the isolates showed growth within two weeks of inoculation and incubation at 25°C. Out of 100 isolates, 31/71(43.66%) of lymphocutaneous type showed culture positivity within the 2nd week. While 11/28(39.28%) cases of fixed cutaneous form showed growth within 14 days. Dimorphism was demonstrated by yeast conversion on brain heart infusion blood agar after incubation at 37°C and 91.30% (63/69) showed yeast conversion within two weeks (Table 1). Patients were treated with saturated solution of potassium iodide (SSKI) therapy. In cases of intolerable side-effects due to SSKI, itraconazole and fluconazole were the alternate anti-fungal agents.
Himachal Pradesh is a small hill state in North-West India bounded between 30°22′ to 33°12′ North latitude and 75°47′ to 79°04′ East longitude lying entirely in the Western Himalayas. The whole terrain is mountainous with altitude varying from 350 m to 6975 m above the mean sea level. Himachal Pradesh encompasses wide range of agro-climatic conditions conducive to horticulture and agriculture. These are integral to the economy and sustainability of people with 90% of the population dependent on this primary sector. Out of a total of 60,77,900 population, rural community constitutes 54,82,319 and 8,63,000 population are primarily farmers by occupation [9].
S. schenckii is a thermally dimorphic fungus existing as a mould at 25°C and as yeast in host tissue at temperature of 37°C [2]. It is widely distributed in nature as a common soil saprophyte occurring on decaying wood, corn stalks and sphagnum moss as has been reported by Mehta et al. from Himachal Pradesh [12]. Professionals involved in agricultural activities like gardeners, farmers, orchardists, florists, foresters or persons whose occupational or recreational activities bring them in contact with ecotypes favourable with saprophytic growth of the fungus and are associated with traumatic contact with plant material and soil are at greater risk of acquiring sporotrichosis.
Age, sex and race have no epidemiological significance in disease causation. Different studies report male predominance [1], [16] or females [1], [6], [8], [17], [18] being more frequently affected. The present study shows a female preponderance (M∶F = 1∶1.27). The disease is related to outdoor practices. In our state, both men and women cultivate and plough the land and collect firewood and fodder from forests along with rearing of livestock. Thus, women too are at high risk of acquiring the disease accounting for a significant number of cases. Previous reports document sporotrichosis in infants as young as 87 days in a study and16 months in another and persons as old as 71 and 87 years [17], [19].The age ranged from 18 months to 88 years in the present study. A rising incidence was observed with increasing age peaking at 4th and 5th decades and declining thereafter. These trends correlate with the active years of the rural inhabitants when they are maximally exposed to occupational trauma and subsequent infection. Children less than 15 years accounted for a meagre 8% which is in sharp contrast with the study from Peru where 60% infections were reported in this age range [3]. The regional variations in the age and sex distribution of cases are usually attributable to different exposure conditions related to economy and literacy. Our findings are however corroborative with other studies [4] where 15% children were <10 years. Previous reports from Himachal reveal similar occurrence [6].
Humidity and rainfall are prime environmental factors predisposing to disease. The temperature between 25°C and 37°C and high humidity of 92–100% helps growth of S. schenckii in the environment [2]. The lower hills of the state known as the Sub-Himalayas have annual rainfall of 1500 mm to 1800 mm with average rainfall of 152 cm accounting for high humidity. Widespread rain occurs from late June to September and part of the state experiences semi-tropical to semi-arctic conditions. In paradox to the expected, we observed 63% of hospital visits between February and June which is the cooler, drier part of the year with sporadic rains and minimum cases during fall and peak winter. Ghosh et al. [5] have reported rain and humidity as two significant factors for sporotrichosis. Similarly, in South Africa, the temperature of 26°C–27°C and 92–100% humidity were known to favour occurrence of sporotrichosis. Contrary to this, in Mexico, greater frequency of infection has been reported during the dry and cooler part of the year similar to our observations [5]. With our experience of this endemic region, we comprehend that though the villagers acquire infection during rainy season, they usually visit peripheral health facilities close at hand where diagnostic modalities for confirmation of fungal infections are not available. Keeping in view infections other than sporotrichosis with similar looking lesions, the therapy instituted usually does not cover fungal infections. Patients do not respond to the treatment at these centres and are referred to tertiary health care hospitals, to a dermatologist after an average duration of disease of 6 to 10 months. The suitability of time when the tilling activities in the land holdings are minimum labour intensive is another local factor accounting for regional trends.
The disease prevalence has perceptibly increased over the years. Perhaps, the increased education, awareness and the literacy rate of 76.05% existent in Himachal Pradesh [9] makes the people realise the need to visit the hospital for ailments. Furthermore, the disease is frequently suspected by the treating physician at the tertiary health care facility thereby accounting for rising incidence. There was a significant rise in number of culture confirmed cases in 2003 and again in 2010 which cannot be accounted for by the above factors alone. The reasons cannot be elucidated in absence of more elaborate epidemiological studies and further identification of the isolates. To validate hypothesis like increase in environmental distribution of S. schenckii or change in virulence of the agent, extensive studies need be done.
Clinical cases of sporotrichosis have been classified into lymphangitic or lymphocutaneous lesions, localized or fixed type, multifocal or disseminated and extracutaneous categories by Sampaio and Lacaz [20]. The classic form is lymphocutaneous accounting for nearly 70% of cutaneous sporotrichosis cases [2]. Our findings corroborated fully with this as we also reported 71% of the cases to have lymphocutaneous lesions. Lesions on hands were the most frequent presentation followed by arm, face, legs and foot in that order [2]. We came across rare sites like lesions on breast, ear lobe and disseminated presentations during our study.
The lesions at inoculation site can be multiple, may remain as such or nodules appear along the lymphatics. Hematogeneous spread or multiple traumatic implantations of the fungus is perhaps responsible for cutaneous dissemination and may even be seen in individuals apparently without any predisposing factors for immunosuppression [2]. Extracutaneous forms like pulmonary, osteoarticular, ocular, central nervous system disease, etc., though reported both in immunocompetent and immunosuppressed cases have not been encountered during the study period. It has also been experienced that only 5.06% (4/79) of the isolates showed growth at 37°C thereby implying that the local strain in Himachal probably is unable to thrive at internal body temperature thus unlikely to cause extracutaneous disease (Table 1).
Histopathology is usually non-specific and rarely diagnostic as the cigar-shaped, budding yeast-like cells are scanty in the pus or tissue samples, rarely demonstrable and invariably non-contributory in confirmation of diagnosis [6], [19]. Asteroid bodies (Splendore-Hoeppli phenomenon) in pus were originally described by Splendore in 1908 [16]. They consist of a central round cell of 5–13 microns, occasionally budding, surrounded by club-shaped multiple hyaline projections. Its demonstration serves only as an adjuvant diagnostic test as eumycotic mycetoma, actinomycosis and granules associated with Pseudomonas aeruginosa may show presence of Asteroid structures. Researchers have variously reported presence of Asteroid bodies as 40% to 85.7%; however, many researchers point out the rarity of these forms in human lesions [16]. In the present study, direct microscopy was performed in 58 samples and Asteroid bodies were revealed in a solitary case (1%) whereas yeast- like cells were not observed in any sample. Special stains like PAS are usually not helpful in recognition of sporotrichosis due to paucity of yeast in tissue and none of the specimens stained by PAS contributed in diagnosis in our study. Lack of experience in visualizing fungal forms in these stains together with rarity of yeast cells in specimens are reasons apprehended for inability to recognise fungal elements.
The chances of isolating S. schenckii from sporotrichosis cases are very good [2] giving a confirmed diagnosis in 2 weeks as seen in our study. The culture is characterised by initial white yeast- like growth appearing within 10 days followed by buff coloured mycelial forms in another 4–5 days. The colour changes to greyish-black with a leathery to velvety texture with occasional aerial mycelium over 2 weeks. S. schenckii usually does not grow straight from the human specimen, but after growing at 25°C, there is a conversion at 37°C.
In the present account, we observed that 81% (64/79) of the isolates showed growth in the second week after inoculation and incubation at 25°C. Out of 100 isolates, 71 were lymphocutaneous of which 31 (43.66%) showed culture positivity within the 2nd week. While out of the 28 cases of fixed cutaneous form of sporotrichosis, only 11 (39.28%) showed growth within 14 days. This curious aspect revealed while studying growth of isolates is corroborated by facts cited previously in literature [2], [13], [21].
Other immunological techniques having little diagnostic significance in healthy people and are not recommended in endemic areas thus were not undertaken. Patients were treated with saturated solution of potassium iodide (SSKI) therapy as it is cheap, effective and gives consistently effective results in a dermatologist's experience. Favourable response within 2 weeks in most cases with healing in 4–8 weeks or up to 32 weeks was observed. To achieve mycological clearance, therapy may be continued for another 4–8 weeks. In cases of intolerable side-effects due to SSKI, itraconazole and fluconazole were the alternate anti-fungal agents used and encouraging results were seen as also documented by other workers [6].
Our present observation demonstrates and emphasises the importance of culture of samples from suspected cases. Fungal culture is gold standard test of great diagnostic significance especially in cases of clinical ambiguity where lesions are not differentiable from lupus vulgaris, cutaneous leishmaniasis, nocardial infections, chromoblastomycosis, syphilis, tuberculosis verrucosa cutis, foreign body granuloma, etc. Culture is undoubtedly the modality of choice for diagnosis, especially because sometimes histopathology gives inconclusive results and usually direct microscopy is non-contributory. We also conclude that there is a significant rise in incidence of sporotrichosis in the sub-Himalayan region.
|
10.1371/journal.ppat.1000281 | Estimation of Transmission Parameters of H5N1 Avian Influenza Virus in Chickens | Despite considerable research efforts, little is yet known about key epidemiological parameters of H5N1 highly pathogenic influenza viruses in their avian hosts. Here we show how these parameters can be estimated using a limited number of birds in experimental transmission studies. Our quantitative estimates, based on Bayesian methods of inference, reveal that (i) the period of latency of H5N1 influenza virus in unvaccinated chickens is short (mean: 0.24 days; 95% credible interval: 0.099–0.48 days); (ii) the infectious period of H5N1 virus in unvaccinated chickens is approximately 2 days (mean: 2.1 days; 95%CI: 1.8–2.3 days); (iii) the reproduction number of H5N1 virus in unvaccinated chickens need not be high (mean: 1.6; 95%CI: 0.90–2.5), although the virus is expected to spread rapidly because it has a short generation interval in unvaccinated chickens (mean: 1.3 days; 95%CI: 1.0–1.5 days); and (iv) vaccination with genetically and antigenically distant H5N2 vaccines can effectively halt transmission. Simulations based on the estimated parameters indicate that herd immunity may be obtained if at least 80% of chickens in a flock are vaccinated. We discuss the implications for the control of H5N1 avian influenza virus in areas where it is endemic.
| Outbreaks of highly pathogenic H5N1 avian influenza in poultry first occurred in China in 1996. Since that time, the virus has become endemic in Asia, and has been the cause of outbreaks in Africa and Europe. Although many aspects of H5N1 virus biology have been studied in detail, surprisingly little is known about the key epidemiological parameters of the virus in its avian hosts (the length of time from infection until a bird becomes infectious, the duration of infectiousness, how many birds each infectious bird will infect). In this paper we show, using experimental transmission studies with unvaccinated and vaccinated chickens, that H5N1 avian influenza induces a short duration of infectiousness (∼2 days) and a very short period of time from infection until infectiousness (∼0.25 day) in unvaccinated chickens. Furthermore, while transmission was efficient among unvaccinated birds, no bird-to-bird transmission was observed in vaccinated chickens. Our results indicate that it may be difficult to curb outbreaks by vaccination after an introduction in a flock has been detected. On the other hand, preventive vaccination could be effective in preventing virus introductions and limiting the size of outbreaks.
| Highly pathogenic avian influenza virus strains of the H5 or H7 subtypes are noted for being highly contagious among various bird species and inducing high mortality rates in poultry. Although outbreaks of highly pathogenic avian influenza have been reported since the 1950s the current focus is on the H5N1 subtype. The first outbreaks of H5N1 were reported in Hong Kong in 1997 [1]–[2]. Since then the virus has spread to South East Asia, Africa, and Europe. The outbreaks in Europe were controlled by rapid depopulation of infected premises, pre-emptive culling of neighbouring farms, movement restrictions, and zoo-sanitary measures [3]–[5]. In Asia, however, the disease has become endemic, and control by means of culling in conjunction with movement restrictions and zoo-sanitary measures is both infeasible socio-economically and unlikely to result in elimination [6]–[10]. Therefore, vaccination is the most widely used containment strategy. For instance, in Indonesia alone more than 400 million of vaccine doses have been administered since 2004.
Despite the fact that aspects of H5N1 avian influenza biology have been studied in detail, ranging from molecular studies of host range factors, phylogenetic analyses aimed at unravelling the virus' evolutionary pathways, surveillance of H5N1 in wild birds, studies into the clinical course of H5N1 infections in humans, and vaccine efficacy and safety studies, there is scant information of the basic epidemiological characteristics of H5N1 viruses in their avian hosts. Specifically, little is known about the infectious period of H5N1 in various host species, the duration of the latent period, and the transmissibility of the virus from bird to bird. For a proper understanding of the transmission dynamics of the virus and to be able to assess the potential impact of control measures such as vaccination, however, this information is crucial. For instance, it is well-known that both the invasion prospects of the virus as well as the number of individuals ultimately infected are critically affected by the (distribution of the) infectious period and transmission parameter. The (distribution of the) period of latency is also of importance since it is a key factor affecting the initial growth rate and duration of an epidemic [11]–[13].
Here we present and analyze experimental transmission studies with highly pathogenic H5N1 avian influenza virus (A/Chicken/Legok/2003) in chickens to obtain quantitative estimates of key epidemiological parameters. Specifically, we performed experiments in which an artificially infected chicken was placed in a cage with a susceptible contact bird, and in which the transmission chain was monitored by taking daily samples from the trachea and cloaca [14]–[16]. The samples were subsequently tested for the presence of virus by egg-culture. In addition, blood samples were taken weekly to determine the antibody response to infection. In all, two experiments, each containing 11 trials, were carried out with unvaccinated chickens, two experiments of 11 trials were performed using an H5N1 inactivated oil emulsion vaccine which contains a strain that is identical to the challenge virus (A/Chicken/Legok/2003), and four experiments of 11 trials were carried out with two heterologous H5N2 inactivated oil emulsion vaccines (A/Turkey/England/N28/73 and A/Chicken/Mexico/232/94/CPA) that are both genetically and antigenically distant from the challenge virus.
The experiments are analyzed by tailored statistical methods based on a SEIR (susceptible-exposed-infectious-removed) epidemiological model. In this way all estimated parameters have a clear-cut biological interpretation (mean and variance of the latent and infectious period, transmission rate, reproduction number). Here we use two different methods of analysis. The first uses final size data i.e. the number of birds that are ultimately infected, and is aimed at estimation of the reproduction number [17]. The second approach uses all available information and is based on Bayesian inference that relies on Markov Chain Monte Carlo (MCMC) techniques [18]–[21]. This allows one to estimate not only the reproduction number, but also other epidemiological parameters of interest. The main advantage of our controlled experimental setup over field studies [6],[7],[22] is that the parameters of interest can be estimated with high precision using a limited number of birds. In addition, our controlled experimental setup makes it possible to ascribe differences between control and treatment groups directly to the treatment without having to take into account the potential effect of confounding variables (e.g., age and size of the birds, stocking density, feeding status).
All inoculated unvaccinated birds (Tables 1 and 2) showed signs of infection (depression, labored breathing), shed virus from both the trachea and cloaca (apart from a single bird in Table 2), and died within a few days after infection (range: day 2–day 3). Furthermore, all contact birds died on day 4 or on day 5 after infection of the inoculated bird, indicating rapid infection as well as rapid progression of the disease towards death. In the experiments with unvaccinated birds 8 out of 22 birds escaped infection. These birds did not show signs of disease, did not shed detectable virus, and remained serologically negative when tested in the HI assay at days 7 and 14.
In the experiments with vaccinated birds no contact birds were infected and only a few of the inoculated birds shed virus on just a few days. In fact, only 7 out of 66 inoculated birds shed virus for a total of 12 days. Of these, virus was isolated from the trachea only on 11 days and from the trachea and cloaca on a single day. None of the vaccinated birds died in the course of the experiments, and no signs of disease were observed in any of the vaccinated birds. Details of the vaccination experiments are given in Tables S2, S3, S4, S5, S6, and S7.
With regard to transmission, the final size analyses indicate that there are significant differences between the experiments with unvaccinated and vaccinated birds. Table 3 summarizes the results. For the low-dose experiment with unvaccinated birds estimates of the reproduction number are 9.0 (95% confidence interval (CI): 1.9–86) in case of an exponentially distributed infectious period and 3.4 (95%CI: 1.3–7.6) in case of a fixed infectious period. For the high-dose experiment the estimates of the reproduction number are 1.7 (95%CI: 0.40–6.6) in case of an exponentially distributed infectious period and 1.2 (95%CI: 0.37–2.9) in case of a fixed infectious period. Although the difference between the two experiments is not statistically significant (p = 0.23) [23], it does hint at the possibility of a role of the inoculation dose in impacting on the transmission dynamics. If all experiments with unvaccinated birds are combined the outcome is that 14 out of 22 initially susceptible contact birds are infected. In this case the estimates of the reproduction number are 3.5 (95%CI: 1.4–9.6) and 2.0 (95%CI: 1.0–3.5), assuming exponentially distributed and fixed infectious periods respectively, indicating that the virus is able to spread epidemically in unvaccinated populations.
No transmission was observed in all six experiments with vaccinated birds, resulting in a maximum likelihood estimate of the reproduction number of 0. The (two-sided) 95% confidence interval ranges from 0–0.80 or 0–0.67, depending on the assumptions regarding the distribution of the infectious period. Furthermore, the null-hypothesis that the reproduction number is larger than the threshold value 1 can safely be rejected (p = 0.011 in case of an exponentially distributed infectious period, and p = 0.0041 in case of a fixed infectious period). Hence, it is unlikely that an epidemic can occur in vaccinated populations.
The experiments with unvaccinated birds are analyzed using Bayesian methods to obtain estimates of the transmissibility of the virus and the distributions of the latent and infectious periods. Table 4 and Figures 1–3, S1, S2, S3, S4, and S5 summarize the main findings. In the low-dose experiment (Table 1) as well as the high-dose experiment (Table 2) the estimated mean of the latent period (or, more precisely, the median of the marginal posterior distribution of the parameter determining the mean) is small (0.20 (day) or 0.44 (day)), as is the variance of the latent period (0.044 or 0.078) (Figures S1 and S2). The estimated means of the transmission parameter are also comparable in the two experiments, ranging from 0.74 (day−1) in the high-dose experiment to 0.80 (day−1) in the low-dose experiment. With regard to the infectious period, however, there appear to be differences between the low- and high-dose experiments, with the birds in the low-dose experiment having a substantially longer infectious period. In fact, the estimated mean of the infectious period is 2.5 (day) (95%CI: 2.2–2.8 (day)) for the low-dose experiment, and 1.3 (day) (95%CI: 0.92–1.8 (day)) for the high-dose experiment. In both experiments the estimated variance of the infectious period is small (0.16 or 0.13), indicating that the infectious period distribution is narrowly centered around the mean.
We then analyzed the data of the low- and high-dose experiments simultaneously to obtain more precise estimates of the parameters of interest. We considered three scenarios (labeled by B, C, and D) that differ with regard to assumptions on the latent and infectious periods (see Methods). Analysis of the pooled data (scenario B) verifies the earlier indications (scenarios A1–A2) that both the estimated mean and variance of the latent period are small, while the estimated mean of the infectious period (2.1 (day)) lies between the estimated means of the infectious period in the analyses of the low- and high-dose experiments (Figure 1). In comparison with the separate analyses of the low- and high-dose experiments the estimated variance of the infectious period increases (0.16 and 0.13 in the low- and high-dose experiments versus 0.33 in scenario B), probably because of the need to accommodate both short (∼1.5 (day)) and long (∼2.5 (day)) infectious periods. Alternatively, if the infectious period distributions are allowed to differ between the low- and high-dose experiments (scenario C), then the estimated mean infectious periods as well as the corresponding variance estimates revert to values close to those in the separate analyses of the low- and high-dose experiments (Figure 2). Based on Bayes factor (see Methods) the model that allowed for differences in the infectious periods has substantially higher support (BF = 21 for the pair of simulations of competing models with the smallest difference in marginal likelihoods) than the model in which the infectious period distributions in the low- and high-dose experiments are assumed to be equal. Finally, if the latent and infectious period distributions are allowed to differ between inoculated and contact birds (scenario D; Figure 3) there is some evidence that, overall, the infectious period of the contact infected birds was somewhat longer than that of the artificially infected birds (mean 2.5 (day)(95%CI: 1.9–3.3) versus mean 1.7 (day)(95%CI: 1.4–2.1)). Figure 3 furthermore shows that the variances of the latent and infectious period distributions of the contact infected birds could not be estimated with precision. An extended analysis including alternative informative prior distributions and an artificially extended dataset indicate that this is indeed the case, and that the experiments of Table 1 and 2 do not contain sufficient information to estimate the variance of the latent and infectious periods of the contact infected birds (unless substantial prior information is added)(results not shown).
Two derived epidemiological measures of interest are the reproduction number R and the generation interval Tg [11]–[13]. In our setting the reproduction number is given by the product of the infectious period and the transmission rate, while the generation interval is defined as the moment of infection of the contact bird, relative to the time at which the inoculated bird was returned to the cage following inoculation. Overall, the generation interval ranges from an estimated mean of 1.2 (day) in scenario A1 and scenario D to 1.8 (day) in scenario A2, with limited variation around these estimates. This indicates that the generation interval is short, and lies in the range of 1–2 days. With regard to the reproduction number, we find substantial differences in the reproduction number between the low- and high-dose experiments. In fact, the estimated reproduction number is 2.0 (95%CI: 0.96–3.6) in the low-dose experiment, and 0.99 (95%CI: 0.38–2.1) in the high-dose experiment. This difference can be ascribed to differences in the mean infectious period in the low- versus high-dose experiments (Table 4). If the data of the low- and high-dose experiments are pooled and assumed to have the same infectious period distribution (scenario B), the estimated reproduction number lies between the above extremes (1.6; 95%CI: 0.90–2.5). Alternatively, if the data are pooled but the infectious period distributions are allowed to vary between the low- and high-dose experiments, the (infection-type specific) estimated reproduction numbers are 1.8 (95%CI: 1.1–3.0) and 1.2 (95%CI: 0.71–2.0) in the low- and high-dose experiments, respectively.
To explore the implications of the parameter estimates for the dynamics of H5N1 avian influenza in large populations of poultry we have performed stochastic simulations of an SEIR model using the parameter estimates presented in Table 4. The parameters determining the latent and infectious periods can directly be plugged into the model, but some care should be taken with the transmission parameter as it is not obvious how the parameter determining transmission between two individuals should be extrapolated to larger populations. The two common assumptions are that each individual makes a fixed number of contacts per unit of time regardless of population size (the frequency dependent transmission assumption), or that each individual makes a fixed number of contacts with each of the other individuals in the population per unit of time (the density dependent transmission assumption) [24]–[25]. Under the frequency dependent transmission assumption the total number of contacts that an individual makes per unit of time does not depend on total population size, while under the density dependent transmission assumption the number of contacts that an individual makes per unit of time increases linearly with total population size [24]. It is plausible that for small to moderately sized populations the transmission rate increases monotonically with increasing population size and that this increase flattens off as population size becomes large (birds cannot increase their activity levels indefinitely). Here we perform simulations of populations of 10,000 birds. In the simulations we first use the transmissibility estimates presented in Table 4, which we subsequently multiply by a factor 2. This implies that in our simulations birds in a population of 10,000 are either as active as birds that are kept in pairs, or twice as active as birds in pairs.
Figure 4 shows two representative simulations of an epidemic in a population of 10,000 individuals using the parameter estimates of the low-dose experiment (Table 4). The top panel shows the time course of the epidemic in case of low transmissibility (leading to a reproduction number of R = 2.0), while the bottom panel shows the dynamics if the transmission rate parameter is increased twofold (implying a reproduction number of R = 4.0). The figure shows that the epidemic unfolds in about a month (top panel) to approximately two weeks (bottom panel), depending on whether the transmission parameter is small or large. Furthermore, the figure shows that the peak prevalence is about 25% of total population size if transmissibility is low, and approaches 65% if transmissibility is high. Increasing the virus' transmissibility from twofold to, say, tenfold leads to minor changes in the infection dynamics as every susceptible individual is already very quickly (within a time span of a week) infected in the high transmissibility scenario (results not shown). It is of note that in comparison with standard stochastic models that assume exponentially distributed latent and infectious periods the epidemics in Figure 4 are considerably more peaked, while their durations are substantially shorter (results not shown) [12].
Rapid detection of outbreaks of H5N1 highly pathogenic avian influenza virus in poultry is of paramount importance for efficient control within poultry flocks and to be able to minimize the opportunities of virus transmission between flocks [26]–[28]. If we assume that avian influenza can be detected with high specificity if mortality is at least 0.5% on two consecutive days [26],[29], then an outbreak will be detected in our simulations between days 11 and 12 after introduction if transmissibility is low, and between days 7 and 8 if transmissibility is high (see the blue arrows in Figure 4). In case of low transmissibility, this gives a window of opportunity of at most ten days to reduce the infectious output of the flock (Figure 4A). If, however, transmissibility is high, circulation of the virus will only be detected near the moment of peak infectivity, and there is a window of opportunity of at most five days for control measures to be effective in reducing the infectious output of infected flocks once they are detected (Figure 4B). Overall, our simulations indicate that control of H5N1 avian influenza in poultry flocks once an outbreak has been detected may be more difficult than hitherto thought [22], [26]–[27].
To further investigate the potential for control by vaccination we have carried out simulations using estimates of the epidemiological parameters (Table 4) and efficacy of vaccination (Table S1). Because of the reasons discussed above, it is highly unlikely that an outbreak can be controlled by vaccination once it has been detected. Adding to this is the fact that it may take 7–10 days for vaccination to become effective in interfering with transmission [14]–[16]. However, it may still be possible to prevent or curb outbreaks by preventive vaccination.
Figure 5 gives an overview of the fraction of outbreaks that yield a major outbreak (numbers near circles), the size of the major outbreaks (circles), and the duration of the epidemics (squares) as a function of the fraction of birds that is vaccinated prior to introduction of the virus. If transmissibility is low (cf. Figure 4A)(blue lines), the probability of a major outbreak as well as the size of the major outbreaks decrease with increasing vaccination coverage. The duration of major outbreaks, however, increases with increasing vaccination coverage [11]–[12]. Major outbreaks cannot occur for the parameters presented in Table 4 if coverage is at least 60%. If, on the other hand, pathogen transmissibility is high (cf. Figure 4B)(red lines), then the probability of a major epidemic and final size of the epidemics increase in comparison with the low-transmissibility scenario, while the duration of the epidemics decreases [11]. Still, both the probability of a major outbreak as well as the size of the outbreak decrease with increasing vaccination coverage, and major outbreak cannot occur if vaccination coverage is at least 80%. Summarizing, our simulations indicate that it is possible to attain a state of herd immunity by incompletely vaccinating flocks of chickens even if birds are assumed to make twice as many contacts per unit of time as estimated in our transmission experiments.
In this study we have attempted to fill the remarkable void of quantitative information on key epidemiological parameters of H5N1 highly pathogenic avian influenza in chickens. Our results indicate that H5N1 virus induces a short period of latency and a short infectious period. In fact, our estimate of the mean of the latent period varies from 0.20 days (95%CI: 0.049–0.43 days) in scenario A1 to 0.44 days (95%CI: 0.14–0.87 days) in scenario A2 (Table 4). Likewise, the mean infectious period varies from 1.3 days (95%CI: 0.92–1.8 days) in scenario A2 to 2.5 days (95%CI: 2.2–2.8 days) in scenario A1. Estimates of the variance of the infectious period are generally low, much lower than the corresponding means (Table 4). This implies that the distributions of the infectious periods are fairly narrow. Similar results were reported by Carrat and colleagues [30] who found that shedding of human influenza viruses increased sharply 0.5–1 day after infection, while the infectious period was centered narrowly around five days.
Our estimates of the transmission parameter are remarkably similar across the different datasets and model scenarios. The estimate of the transmission parameter is lowest if the data of all experiments are combined (median: 0.73 per day; 95%CI: 0.43–1.2 per day) and highest if the analysis allows for differences between inoculated and contact birds (median: 0.81 per day; 95%CI: 0.44–1.3 per day). In combination with the estimates of the mean infectious period these estimates yield estimates of the reproduction number varying from 0.99 (95%CI: 0.38–2.1) in the high-dose experiment (scenario A2) to 2.0 (95%CI: 0.96–3.6) in the low-dose experiment (scenario A1).
In view of the generally held belief that highly pathogenic avian influenza viruses spread easily and rapidly among chickens [14]–[16], [22], [26]–[27] our estimates of the reproduction number may seem low. In this respect a number of points are worth of discussion. First, we have assumed frequency dependent transmission, which assumes that each bird makes a fixed number of contacts per unit of time, regardless of the size of the population [24]. This is convenient since it allows one to directly extrapolate from small to large populations. The reason is that under this assumption the reproduction number does not depend on total population size. There is, moreover, evidence that a frequency dependent transmission model provides a better description of the pathogen dynamics than a density dependent model in farm animals that are generally held at a constant stocking density [31]. Still, some uncertainty remains as to how our estimates of the transmission parameter and infectious period should be combined into an estimate of the reproduction number. To address this potential problem we have in our simulations included a high transmissibility scenario (Figures 4 and 5) that in essence assumes that birds in large populations are twice as active as birds in our transmission experiments with pairs of birds.
Second, it is not straightforward to extrapolate our results that were obtained in an experimental setting to the situation in the field. This is especially so for estimates of the transmission parameter, which are the result not only of an autonomous process of viral replication and interaction of the pathogen with the immune system within a single host, but also of an interaction between different individuals. Ambient temperature, stocking density, feeding status of the birds, etcetera could all impact on this interaction and critically affect estimates of the transmission parameter. To counter this we have tried to match the conditions in our experiments to those in commercial laying chicken farms. Reassuringly, a recent analysis of transmission of H5N1 in the field [7] also indicates that the reproduction number of H5N1 virus among chickens is fairly low, ranging from 2.0 to 3.5. This suggests that our estimates of the reproduction number obtained using pairs of birds are low but not unreasonable.
A third point that deserves attention is the fact that housing systems of layer flocks vary from floor systems in which birds can mingle freely to caged systems in which no direct contact between (groups of) birds is possible. In principle, our study is aimed at quantifying transmission in a situation where there is direct contact between birds, corresponding to a floor system. However, the lone study that focused on within-flock transmission (mostly backyard flocks) did not find differences between different housing systems, suggesting that if there are differences in the transmission dynamics they cannot be large [7]. Nevertheless, more information on the infection dynamics in the field would be highly welcome to help bridging the gap between findings obtained in experimental studies and the situation in the field.
While it is not straightforward to extrapolate from our experimental setting to the field situation, experimental transmission studies also have distinct advantages over field studies. In particular, while field studies often suffer from various sources of bias and confounding, this is not the case in an experimental setting. This allows one to directly ascribe differences between control and treatment groups directly to the treatment (e.g., vaccination) since all other animal and environmental conditions are held constant. Moreover, an experiment has the added advantage over a field study that far fewer birds are needed and that the birds can be sampled more often and efficiently than in a field study. This has allowed us to obtain precise estimates of the key epidemiological parameters of H5N1 highly pathogenic avian influenza in unvaccinated chickens using no more than 50 birds.
Our results show remarkable differences between experiments in which the inoculated bird received a low infection dose (0.2*105 EID50) and experiments in which the inoculated bird received a high dose (0.2*106 EID50). Specifically, while 9 out of 11 birds were infected in case of a low infection dose (Table 1), only 5 out of 11 were infected in case of a high infection dose (Table 2). This is an interesting and counterintuitive result, which is likely to result from the fact that the infectious period in the experiments in which the inoculated bird received a high inoculation dose is significantly smaller than in the experiments in which the inoculation dose was low or in which the infectious period of the naturally infected birds was estimated separately (low dose: mean 2.5 days (95%CI: 2.2–2.8); high dose: mean 1.3 days (95%CI: 0.92–1.8); contact birds only: mean 2.5 days (95%CI: 1.9–3.3)). Earlier experimental transmission studies with H7N7 highly pathogenic avian influenza virus (A/Chicken/Netherlands/621557/03) in a variety of birds and H5N1 highly pathogenic avian influenza virus (A/Chicken/China/1204/04, also designated A/Chicken/GxLA/1204/04) in ducks used an infection dose of 0.2*106 EID50 since this yielded comparable infections in inoculated and naturally infected animals [14]–[16]. The finding that the infection dose is of importance in determining the duration of infection is of both theoretical and practical relevance as it suggests that the infection pressure in the population may not only determine the incidence of infection but also the course of infection. If it is typical that a low infection dose is associated with a long infectious period while a high infection dose generally leads to infections that are of short duration, then this would necessitate a rethinking of the critical determinants of H5N1 avian influenza transmission in populations of birds, and it could potentially have profound implications for optimal control and containment strategies.
To investigate the implications of our parameter estimates for the dynamics of H5N1 avian influenza virus in large groups of chickens we have carried out stochastic simulations. Since it is not obvious how the transmission parameter as estimated between pairs of chickens can be extrapolated to large populations, we considered a low and high transmissibility scenario (Figure 4). The simulations indicate that, even if we assume that the transmission parameter is small, the epidemic usually unfolds in about a month, and that once the epidemic has taken off it only takes about two weeks to come to an end. If, as appears more likely, the transmission rate is larger in large population than in populations of two birds, then the epidemic takes off more quickly after a primary introduction, and also comes to an end more quickly. For control purposes this implies that it will be very difficult, if not impossible, to effectively control an outbreak once it has been detected. It may even prove difficult to reduce transmission opportunities from an infected population (a farm, say) to susceptible populations, as the number of dead birds may start to rise just before peak infectivity (Figure 4). This suggests that perhaps other indicators of infection, such as lethargy, reduced feed or water intake should be added to the mortality indicator to obtain a sensitive syndrome-reporting system [32].
While H5N1 virus spreads rapidly among unvaccinated chickens, no transmission was observed at all in the experiments with inactivated oil emulsion vaccines (Tables S1, S2, S3, S4, S5, S6, and S7). This was true not only for an H5N1 vaccine virus which had 100% homology to the challenge virus, but also for genetically distant heterologous viruses that contained inactivated H5N2 viruses. These findings indicate that it is possible, at least in principle, to reduce transmission by vaccination to the extent that no epidemics can occur. This suggestion is corroborated by our simulations which indicate that a vaccination coverage as low as 60%–80% may already be sufficient to obtain herd immunity (Figure 5). Of course, it should be borne in mind that in our experiments all birds received two vaccination doses, that the timing of challenge (two weeks after the last vaccination bout) was probably ideal, and that in the field there are various factors that may interfere with vaccination (concurrent infections, immune depression by various causes) [33]. Still, our results and those of others [34]–[36] provide a proof-of-principle that herd immunity can be obtained with currently available inactivated oil emulsion vaccines. The finding that H5N1 avian influenza virus has a lower transmissibility than hitherto believed [26] also implies that outbreaks may be easier to prevent than previously thought, since the reproduction number is already relatively close to the threshold value of 1.
All experiments were carried out in PT Medion laboratories in Bandung, Indonesia, which have high containment facilities (BSL3). In all experiments, specific pathogen-free (SPF) layer chickens from the animal unit of Medion were used. The birds were hatched and housed in one group until 4 weeks of age. At that age, pairs of birds were housed in cages. Three rooms were available to house the various vaccinated and unvaccinated pairs of birds. Two rows with three levels of cages on top of each other were available in each room. The rows with cages were separated by a corridor of approximately 1 m width. The various rooms as well as the rows with the cages had separate ventilation systems. Each cage had a separate feeding and drinking system. The floor and walls of each cage were covered with plastic to prevent spread of manure or other materials between cages. When sampling the birds, animal caretakers used a new pair of gloves for each cage. Unvaccinated sentinel birds were placed at regular distances between the cages used in the experiments to ensure that no transmission had taken place between cages. All sentinels survived and remained seronegative during the course of the experiments.
The challenge strain used in the experiments was A/Chicken/Legok/2003 H5N1, a highly pathogenic H5N1 strain isolated in Indonesia in 2003 which is genetically very close to strains that circulate in Indonesia in 2008. The strain has been used in experiments carried out at Medion and is able to induce infection, typical signs of disease, and high mortality rates in chickens.
Inactivated oil emulsion vaccines were available from three different manufactures: PT Medion (Bandung, Indonesia), PT Vaksindo (Bogor, Indonesia) and Intervet (Mexico). The vaccines contained either an H5N1 or H5N2 virus strain. The H5N1 vaccines contained A/Chicken/Legok/2003 H5N1, i.e. the vaccine and challenge strains were identical. The H5N2 vaccines contained either A/Turkey/England/N28/73 H5N2 or A/Chicken/Mexico/232/94/CPA H5N2. The protein homologies of the antigenic part of the hemagglutinin (HA1) of the challenge strain to the H5N2 A/Turkey/England/N28/73 and H5N2 A/Chicken/Mexico/232/94/CPA vaccine strains are 92% and 86%, respectively.
All vaccines were re-vialed in coded bottles, and the identity of the vaccines was not known to the staff involved in the experiment. In this manner the experiments were double blinded.
Because the size of a natural infection dose is unknown the inoculum consisted of diluted allantoic fluid containing either 105 EID50 per ml (low inoculation dose) or 106 EID50 per ml (high inoculation dose). The birds were inoculated both intranasally (0.1 ml) and intratracheally (0.1 ml). Virus titres were confirmed before and after inoculation by titration on embryonated SPF eggs.
Each experiment consisted of a set of 11 trials. In each of the trials an inoculated bird was placed in a cage with an uninfected contact bird, and the transmission chain was monitored daily by virus isolation performed on swabs taken from the trachea and cloaca. In all, a total of eight experiments were carried out. Unvaccinated birds were used in two experiments. In the first of these the inoculated birds received a low infection dose, while in the second the inoculated birds received a high infection dose. The remaining six experiments with vaccinated birds differed with respect to the vaccine used, the manufacturer, and the inoculation dose. Tables 1 and 2 show the data of experiments with unvaccinated birds, and Tables S1, S2, S3, S4, S5, S6, and S7 give an overview of the experiments with vaccinated birds.
At 4 weeks of age all birds of the vaccination experiments received their first vaccination dose. A second vaccination was carried out at 7 weeks of age. At 10 weeks of age (day 0) one bird was chosen at random per cage, taken from the cage, and infected intratracheally and intranasally. To avoid direct infection of the contact bird by the inoculum the artificially infected birds were placed back in their cages only after a delay of 8 hours.
Tracheal and cloacal swabs were taken daily for 10 days after challenge from all birds. Swabs were incubated for 1 h in one ml of PBS medium containing antibiotics. The medium was subsequently stored at −70°C until testing. Three embryonated SPF chicken eggs were injected with 0.2 ml of the swab medium per egg. After culture for 4 days or when embryos died, the allantoic fluid was harvested and a hemagglutination (HA) assay was performed following standard procedures (www.oie.int). When at least one of the eggs was positive in the hemagglutination assay the swab was considered to be positive.
The serological status of the birds was determined just before vaccination, at the start of the experiments just before inoculation (day 0) and, for birds that survived, at the end of the experiments (day 14). Serum blood samples were taken from all birds by puncturing the wing vein. Blood samples were centrifuged and serum was stored at −20°C until tested. The sera were tested in the hemagglutination inhibition (HI) test according procedures described in the Manual of Diagnostic Tests and Vaccines for Terrestrial Animals of the OIE (www.oie.int) using 4 HA units (HAU) of A/Chicken/Legok/03 H5N1 as antigen. Titres were expressed as 2log of the serum dilution that caused complete inhibition of agglutination, as specified by OIE guidelines.
Clinical signs of disease were recorded daily for a period of up to 10 days after challenge.
As a first step we estimated the reproduction number R by final size methods [14]–[17]. Since each trial contains only one inoculated bird and one susceptible contact bird, the likelihood function takes the following simple form:(1)In this equation N and n are the number of trials per experiment and the number of infected contact birds, while represents the Laplace transform of the infectious period probability distribution when the mean infectious period is scaled to 1. Hence, in case of an exponentially distributed infectious period, and in case of a fixed infectious period. Table 3 provides estimates of the reproduction number with corresponding 95% confidence intervals, as well as p-values of the null-hypothesis that the reproduction number is greater than or equal to the threshold value of 1 [23].
In a second step, we estimated all parameters of interest by Bayesian methods [18]–[21]. In the following we denote by the transmission rate parameter, by and the parameters determining the latent period probability distribution, and by and the parameters of the infectious period probability distribution. We assume that the latent and infectious periods are gamma distributed, and that and , and and represent the means and variances of these distributions. The corresponding probability densities are denoted by and .
Further, , , and are N-dimensional vectors which contain the time points of the S→E, E→I, and I→R transitions for inoculated () and contact () birds in the N trials. Hence, we have by definition, while all other transition times are unknown. The unknown transitions are added in the analyses by Bayesian imputation. We adopt the convention that denotes the exact time at which the contact bird in experiment j is infected, that denotes the exact time that the inoculated bird in experiment j became infectious, etcetera.
With these notational conventions, the contribution of trial j to the likelihood is given by(2)In the above equation and denote the infection hazard in trial j at time t and the probability that the contact bird in trial j remains uninfected up to time t, respectively. If we let […] denote the indicator function, the infection hazard is given by(3)where the parameter represents the delay between the moment of inoculation and the placing back of the inoculated birds in their cages, and the function marks the beginning of the at-risk period for the contact bird. In all trials and experiments, the delay is 8 hours, i.e. (day). The probability that the contact bird in trial j remains uninfected up to time t can be expressed in terms of the infection hazard as follows(4)Using Equations (2)–(4) the likelihood function is given by the product of the contributions of the individual trials:(5)where P represents the set of trials. Equations (2)–(5) form the basis of the analyses in Figures 1, S1, S2, and S3.
The likelihood contribution in Equation (2) assumes that the latent and infectious periods of inoculated and infectious birds are identically distributed. To investigate the validity of these assumptions we also considered a model which allows for differences between the inoculated and contact birds. In this case, the likelihood contribution becomes(6)where and are the probability density functions of the latent and infectious periods of the inoculated birds () and contact birds (). The results of the analyses based on Equation (6) are given in Figure 3 and Figure S5. In a similar manner, the likelihood contribution in Equation (2) is adapted to allow for differences in the infectious period in the low- versus high-dose experiments. The results of these analyses are summarized in Figure 2 and Figure S4.
Notice that, since the transmission rate in Equations (1), (2), and (5) is divided by the total size of the population (i.e. 2), the above model assumes frequency dependent transmission (as opposed to density dependent transmission) [24]. For the present experimental setup with one inoculated bird and one contact bird, the value of the transmission parameter of the density dependent model is simply given by the transmission rate parameter of the frequency dependent transmission model divided by 2 (the size of the population). In case of a frequency dependent transmission model the (basic) reproduction number is given by the product of the transmission rate parameter and the mean infectious period: . In case of a density dependent transmission model the reproduction number is a function of population size, and it is given by , where denotes the transmission parameter of the frequency dependent model with two birds [25].
As in earlier papers [18]–[21] the epidemiological parameters of interest (, , , , and ) were estimated by Bayesian methods of inference using Markov Chain Monte Carlo. Throughout, all prior distributions of the parameters were uniformly distributed on the interval (0.001–5). As an alternative we also considered vague gamma prior distributions, and obtained comparable results (results not shown).
In our simulations the epidemiological parameters and unobserved transitions were updated by a random-walk Metropolis algorithm. We used Normal proposal distributions with the current value as mean, and a standard deviation of 0.025, 0.05, or 0.1. The transmission parameters and unobserved transitions were updated in blocks, in the order , , , , , and . Notice that updating of the individual transition vectors needs to take into account the infection data of Tables 1 and 2 and the information contained in the other transition vectors, as these specify the admissible intervals of the various transitions. The above updating scheme yielded chains that converged quickly and showed satisfactory mixing. In all analyses we took a burn-in of 25,000 cycles and a simulation length of 200,000 cycles. Thinning was applied by taking only each 100th cycle as a sample from the posterior distribution. We performed four replicate simulations to check the precision of the parameter estimates obtained by the above procedures. These simulations yielded parameter estimates and 95% credible intervals that were close to those given in Table 4.
To choose between models of different complexity we made use of Bayes factors (BF) [37]. To this end the marginal likelihoods of competing models were estimated by importance sampling using the harmonic means of the posterior likelihood values [37]. The BF converged slowly, possibly because of the high dimensionality of the model (86 unobserved transition events plus 5–9 epidemiological parameters), the mutual dependencies of the unobserved transitions, and the fact that the likelihood is strongly affected by the parameters defining the variances of the latent and infectious periods if those are small. However, this did not appear to be a major practical problem as differences between competing models were usually large. When reported in the text we calculated the BF of the pair of simulations that had the smallest difference in marginal likelihoods.
A suite of Bayesian analyses were performed for the experiments with unvaccinated birds. First, we analyzed the low- and high-dose experiments of Tables 1 and 2 separately (scenarios A1 and A2). Second, we pooled the data of the low- and high-dose experiments (scenario B). We then considered an integrated analysis of the two experiments that allowed for differences in the infectious periods in the low- versus high-dose experiments (scenario C). Finally, we considered a scenario which allowed for differences in the epidemiological characteristics of the inoculated and contact birds (scenario D).
To explore the implications of the parameters estimated by the above procedures for the pathogen dynamics in large groups of birds, we performed simulations of the stochastic SEIR model using the Sellke construction [16]. In the simulations we assumed gamma distributed latent and infectious periods, and used the medians of the parameter estimates of Table 4 as input values. The programs for the MCMC analyses and simulated epidemics were written in Mathematica 6.0 (www.wolfram.com).
|
10.1371/journal.pntd.0001592 | Population Structure and Transmission Dynamics of Plasmodium vivax in the Republic of Korea Based on Microsatellite DNA Analysis | In order to control malaria, it is important to understand the genetic structure of the parasites in each endemic area. Plasmodium vivax is widely distributed in the tropical to temperate regions of Asia and South America, but effective strategies for its elimination have yet to be designed. In South Korea, for example, indigenous vivax malaria was eliminated by the late 1970s, but re-emerged from 1993. We estimated the population structure and temporal dynamics of transmission of P. vivax in South Korea using microsatellite DNA markers.
We analyzed 255 South Korean P. vivax isolates collected from 1994 to 2008, based on 10 highly polymorphic microsatellite DNA loci of the P. vivax genome. Allelic data were obtained for the 87 isolates and their microsatellite haplotypes were determined based on a combination of allelic data of the loci. In total, 40 haplotypes were observed. There were two predominant haplotypes: H16 and H25. H16 was observed in 9 isolates (10%) from 1996 to 2005, and H25 in 27 (31%) from 1995 to 2003. These results suggested that the recombination rate of P. vivax in South Korea, a temperate country, was lower than in tropical areas where identical haplotypes were rarely seen in the following year. Next, we estimated the relationships among the 40 haplotypes by eBURST analysis. Two major groups were found: one composed of 36 isolates (41%) including H25; the other of 20 isolates (23%) including H16. Despite the low recombination rate, other new haplotypes that are genetically distinct from the 2 groups have also been observed since 1997 (H27).
These results suggested a continual introduction of P. vivax from other population sources, probably North Korea. Molecular epidemiology using microsatellite DNA of the P. vivax population is effective for assessing the population structure and transmission dynamics of the parasites - information that can assist in the elimination of vivax malaria in endemic areas.
| Vivax malaria is widely prevalent, mainly in Asia and South America with 390 million reported cases in 2009. Worldwide, in the same year, 2.85 billion people were at risk. Plasmodium vivax is prevalent not only in tropical and subtropical areas but also in temperate areas where there are no mosquitoes in cold seasons. While most malaria researchers are focusing their studies on the parasite in tropical areas, we examined the characteristics of P. vivax in South Korea (temperate area) temporally, using 10 highly polymorphic microsatellite DNA (a short tandem repeat DNA sequence) in the parasite genome, and highlighted the differences between the tropical and temperate populations. We found that the South Korean P. vivax population had low genetic diversity and low recombination rates in comparison to tropical P. vivax populations that had been reported. We also found that some of the parasite clones in the population were changing from 1994 to 2008, evidence suggesting the continual introduction of the parasite from other populations, probably from North Korea. Polymorphic DNA markers of the P. vivax parasite are useful tools for estimating the situation of its transmission in endemic areas.
|
Plasmodium vivax, the second most prevalent species of the human malaria parasite, is widely distributed around the world, especially in Asia and South America; it ranges from tropical to temperate areas [1], [2]. In these countries, the proportion of P. falciparum cases is gradually decreasing due to the impact of global malaria control programs such as “The Roll Back Malaria Partnership” and “The Global Fund to Fight AIDS, Tuberculosis and Malaria” as well as local control programs. In contrast, the proportion of P. vivax cases is gradually increasing [1], and therefore deserves more attention than it has previously received [3].
Understanding the genetic characteristics of the malaria parasite population is important for monitoring the transmission pattern and evaluating the effectiveness of malaria control in endemic areas [4]–[7]. Recently, the population structure and transmission dynamics of P. vivax have been reported in some tropical and subtropical areas where the parasites are prevalent throughout the year or seasonally prevalent but not discontinuous during the year [8]–[13]. However, little is known about these characteristics in temperate areas where vivax malaria is only seasonally prevalent and discontinuous during the year.
In the Republic of Korea (South Korea), which is in the temperate zone of the continent of Asia, indigenous vivax malaria had been successfully eliminated by the late 1970s thanks to an effective program conducted by the National Malaria Eradication Service of the South Korean government with the support of the WHO [14]–[16], but has re-emerged since 1993 [17]. At the beginning of the re-emergence, the patients were only South Korean soldiers, veterans, and soldiers from the US military who were serving in the border area between North and South Korea in the western Demilitarized Zone (DMZ) [18]–[20]. Gradually, however, the number of infected civilians who lived in or near the area increased [18], suggesting local transmission of P. vivax between humans and Anopheles mosquitoes in the country. The number of vivax malaria cases increased steadily until 2000 (4,183 cases), then began to decrease gradually until 2004 (864 cases) (Fig. 1) [1], [16]. In spite of continuous malaria control measures implemented by the South Korean government, the numbers of reported cases fluctuated between 1,000 and 2,000 cases per year from 2005 to 2009 [1]. The WHO reports that vivax malaria was more prevalent in the Democratic People's Republic of Korea (North Korea), where there were 296,540 cases in 2001 and 14,845 cases in 2009 [1], [19].
We previously conducted genetic epidemiological surveys of the P. vivax population in South Korea using DNA sequences of some antigenic molecules of the parasite (circumsporozoite protein, Duffy binding protein, apical membrane antigen 1, merozoite surface protein-1) and found that there were 2 genotypes in the country's parasite population [21]–[25]. The advantage of using antigenic molecules of the parasites for genetic epidemiology is that they could be vaccine candidates; however such antigenic molecules are under strong selective pressure from the host immune system, so the variation in the molecules might be biased due to this [26]. In previous studies, the isolates that were used were collected from vivax malaria patients in a single year so temporal changes in the parasite population could not be examined. In the present study, we examined the population structure and the transmission dynamics of P. vivax in South Korea temporally using 10 highly polymorphic neutral DNA markers of the parasite collected from 1994 to 2008 and compared these characteristics with those reported in tropical and subtropical areas. Based on these data, we provide a possible explanation as to why it has not been possible to eliminate vivax malaria in South Korea in spite of a continuous governmental effort.
A total of 255 P. vivax samples isolated from South Korean soldiers or veterans who had served in the DMZ from 1994 to 2008 were used in this study. These patients were also diagnosed by microscopic examination of peripheral blood smears when they contracted malaria. The patient blood samples were collected and preserved at −30°C until use. This study was performed according to the Ethical Guidelines for Clinical Research issued by the Ministry of Health, Labour and Welfare of Japan on July 31, 2008, and the Ethical Guidelines for Epidemiological Research issued by the Ministries of Health, Labour and Welfare, and of Education, Science, Culture, and Sports of Japan on December 1, 2008. Because of the long-term prior collection of widely distributed samples, written or oral informed consent from the patients for the specific purpose of this study could not be obtained at each sample collection. However, no author of the study was involved in gathering patient samples and the individual information of the donors was disconnected from the authors. Thus, all the samples were anonymized, and indeed it is most unlikely that the results obtained from the analysis of the isolated parasites would result in a breach of donor privacy.
Parasite DNA was extracted from frozen whole blood samples by phenol-chloroform extraction after proteinase K digestion [27] or by QIAamp DNA Mini Kit (Qiagen, Valencia, CA, USA).
Ten microsatellite DNA loci were amplified by PCR. The loci were as follows: MS1 (chromosome 3), MS4 (chromosome 6), MS5 (chromosome 6), MS6 (chromosome 11), MS7 (chromosome 12), MS8 (chromosome 12), MS9 (chromosome 8), MS12 (chromosome 5), MS15 (chromosome 5) and MS20 (chromosome 10). The PCR primer sets and amplification conditions were consistent with the protocol of Karunaweera et al. [28]. Sizes of fluorescently-labeled PCR products were measured on an Applied Biosystems Prism Genetic Analyzer 3130xl using GeneMapper(R) version 4.1 with a 500 ROX size standard (Applied Biosystems, CA, USA).
Amplified different-sized PCR products using the same primer sets were considered to be individual alleles within a locus, as size variation among isolates is consistent with the repeat number in a microsatellite locus [5]. The electropherogram shows peak profiles for the microsatellite loci, based on the fluorescence intensity of the labeled PCR products in this analysis. Multiple alleles per locus were scored if minor peaks were taller than at least one-third the height of the predominant allele for each locus. Multiple-genotype infections (MGIs) were defined as those in which at least one of the 10 loci contained more than one allele [5].
Population genetic analyses were performed based on allele frequencies of the 10 microsatellite loci of the population. The level of genetic diversity of the P. vivax population in South Korea was assessed by allele number per locus (A) and expected heterozygosity (HE). HE values for each locus were calculated using HE = [n/(n−1)] [1−Σpi2], where n corresponds to the number of isolates examined and pi is the frequency of the ith allele. The statistical differences among those values were evaluated by Welch's t-test.
Multilocus linkage disequilibrium (LD) was assessed using the standardized index of association (IAS) [29], [30]. This analysis was performed using the LIAN 3.5 Web interface [31]. IAS was calculated using the formula IAS = (VD/Ve−1)/(l−1) with permutation testing of the null hypothesis of complete linkage equilibrium (IAS = 0), where VD is the observed mismatch variance, Ve is the expected mismatch variance, and l is the number of examined loci. Significances of the observed IAS values were calculated by Monte-Carlo simulation, using 10,000 random permutations of the data. This statistic is a variation of the method proposed by Maynard-Smith et al [29]. The results were standardized by the number of loci, to enable a comparison of different data sets [30]. This test was applied to the data sets from each population in two ways. First, the mixed-clone infections were excluded so that only the single-clone infections were analyzed, giving absolute confidence in the haplotype profile. Second, any multilocus genotype found in more than one isolate was only counted once in the analysis, i.e. unique haplotypes only, reducing the sample size slightly and thereby removing the possible effect of recent epidemic expansion of particular clones [5].
Microsatellite haplotypes of the isolates were determined based on a combination of the allelic data of the 10 loci. The relationships among the haplotypes were estimated by eBURST analysis [32].
The allelic data of the 10 microsatellite loci were obtained from 87 of the 255 (34%) isolates that were used in the study. They were not available in the remaining 168 isolates (66%) due to failure in acquiring PCR products of some loci by PCR-based genotyping. Failure was possibly due to there being only a small amount of DNA for PCR amplification or to the DNA being of low quality after multiple times of freeze-thaw.
When different sizes of alleles were observed in one locus, we regarded this as multiple genotype infections (MGIs). MGIs were observed in some of the 10 microsatellite loci in 85 of the 87 isolates (97.7%). The frequencies of MGIs varied among the 10 loci (0.00 to 0.84; average: 0.29) (Table 1). We also examined the number of MGI loci per isolate. In the 87 isolates with 10 loci, the highest frequency of MGI loci per isolate was 2 (25 isolates) and the frequencies decreased gradually according to the increase in the number of MGI loci (Fig. 2). The highest number of MGI loci per isolate was 8 (one isolate). The major alleles in each locus were used for population genetic analysis.
In the 10 loci, the number of alleles (A) for each locus was 2 to 7 (average: 4.3). The expected heterozygosity (HE) for each of these loci was 0.05 to 0.66 (average: 0.43) (Table 2).
Next, the P. vivax population was divided into 2 groups: one comprised of the 47 isolates collected from 1994 to 2000 when the numbers of vivax malaria cases increased; the other comprised of the 40 isolates collected from 2001 to 2008, when the numbers of cases decreased until 2004 and then increased slightly. The level of genetic diversity was reassessed for each group. For the first group, the averages ± SE of A and HE were 2.70±0.26 and 0.36±0.06, respectively. For the second group, the averages ± SE of A and HE were 3.80±0.57 and 0.50±0.10, respectively (Fig. 1). The levels of genetic diversity were relatively higher in the second group, with P values at 0.11 and 0.24 for average A and average HE, respectively.
Furthermore, we also divided the population into 3 groups, each covering 5-year periods: 1994 to 1998 (33 isolates), 1999 to 2003 (36 isolates) and 2004 to 2008 (18 isolates). The level of genetic diversity was reassessed for each group (Fig. 3). For the first group, the averages ± SE of A and HE were 2.50±0.27 and 0.31±0.05, respectively. For the second group, the averages ± SE of A and HE were 3.00±0.42 and 0.42±0.09, respectively. For the third group, the averages ± SE of A and HE were 3.80±0.57 and 0.56±0.10, respectively. The levels of genetic diversity gradually increased with P values at 0.06 and 0.05 if we compared the difference of the average A between the first group (1994–1998) and the third group (2003–2008) and the difference of the average HE between the first and the third group, respectively.
Likewise, the analysis of genetic diversity, IAS values were also calculated for the two populations: one comprised the isolates collected from 1994 to 2000 and the other comprised the isolates collected from 2001 to 2008, with permutation testing of the null hypothesis of IAS = 0 (equilibrium of multilocus frequencies) (Table 3). When the single-clone haplotype was used in the analysis, the IAS values of the former (1994–2000) and the latter (2001–2008) were 0.529 and 0.218, respectively, whereas when the unique haplotypes were used in the analysis, those of the former and the latter were 0.346 and 0.173, respectively. Significant linkage disequilibrium was observed in both populations (P<0.001).
Similar to the analyses of genetic diversity, we also divided the population into 3 groups covering 5-year periods: 1994 to 1998 (33 isolates), 1999 to 2003 (36 isolates) and 2004 to 2008 (18 isolates). The IAS values were also calculated for each group. When the single-clone haplotype was used in the analysis, the IAS values of the first (1994–1998), the second (1999–2003) and the third (2004–2008) groups were 0.584, 0.315 and 0.140, respectively, whereas when the unique haplotypes were used in the analysis, those of the first, the second and the third were 0.408, 0.231 and 0.153, respectively (Table 4). Significant linkage disequilibrium was observed in both populations (P<0.001).
Microsatellite haplotypes of the 87 isolates were determined based on a combination of the allelic data of the 10 microsatellite loci; 40 haplotypes (H1–H40) were observed (Table 5). There were 2 major haplotypes (H16 and H25): H16 was observed in 9 isolates (10%) out of the 87 isolates in samples collected from 1996 to 2005; H25 was observed in 27 isolates (31%) out of the 87 isolates in samples collected from 1995 to 2003. H16 and H25 share only 3 alleles in the loci, MS8, MS12 and MS20, but those in 7 other loci were different from each other.
The relationships among the 40 haplotypes were estimated by eBURST analysis [32] with the following criterion: when 2 isolates shared more than 7 identical loci out of the 10 loci, they were connected with a branch (Fig. 4). Again, two major groups were found: Group 1 was composed of 36 isolates (41%) including the isolates with H25; Group 2 was composed of 20 isolates (23%) including those with H16. Some new or isolated haplotypes, namely H5, H6, H7, H8, H11, H12, H19, H31, H32, H33, H35, H36, that were not included in the 2 major groups or connected to any other haplotypes, have also been observed since 1998. H6 and H7 were not shown in Figure 2 because these haplotypes were quite different from the other haplotypes.
This is the first 15-year-long longitudinal study on P. vivax population genetics using highly polymorphic neutral markers. The present study demonstrated that the level of genetic diversity of the P. vivax population in South Korea was remarkably lower than the levels in tropical and subtropical areas reported by Karunaweera et al. [28] and Orjuela-Sánchez et al. [33] (Table 2). The 10 microsatellite loci used in the present study were a subset of the 14 loci used in the previous studies by other groups [28], [33]. Imwong et al. also reported that the mean values of HE of P. vivax populations from Thailand (n = 28), India (n = 27) and Colombia (n = 27) were 0.77, 0.76 and 0.64, respectively [11], using 11 other microsatellite loci in the genome. These values reported by Imwong et al. were also higher than those in South Korea.
Sample size (n) and sampling conditions such as the size of sampling area and the length of sampling period may affect levels of genetic diversity of living organisms. In comparison to other studies the sample size of the present study (n = 87) was relatively large and the sampling period (15 years: from 1994 to 2008) was relatively long [11], [28], [33]. Generally, one would expect to see an increase in the level of genetic diversity when these conditions (a large number of samples and a long sampling period) are present. However, the South Korean P. vivax population showed low levels of genetic diversity, suggesting that the effective size of the re-emerged P. vivax population in South Korea might be small.
Microsatellite variation is strongly dependent on the length of repeat arrays [34]. Studies of numerous organisms have shown higher levels of variation in loci with long repeat arrays than those with short repeat arrays [35]. In the present study, however, even the locus with a long repeat array (MS20) showed low levels of variation in the South Korean population. Also in the same population, MS8 and MS12 showed low levels of variation, although the loci were not short repeat arrays. The genetic diversities of the loci from Sri Lankan and Brazilian populations (Table 2) were higher than those from the South Korean population. Therefore, the uniqueness of the diversity would not be solely dependent on the characteristics of the loci. In fact, mutations of microsatellite loci are generally considered to be neutral. However, if the loci are in a certain gene or close to a certain gene, the mutation may not be strictly neutral. Indeed, 6 of the 10 loci examined in this study (MS4, MS5, MS8, MS9, MS15, MS20) were in a gene coding a hypothetical protein or a known protein (Table 2). One of the 10 loci (MS7) was between a gene coding a hypothetical protein and a gene coding a merozoite surface protein-7 which is expected to be under strong selective pressure. Therefore, the mutation of those 7 loci may not be strictly neutral. The allelic data suggested that the frequencies of strand-slippage events of the microsatellite loci during mitotic replication in the South Korean P. vivax population were very low because identical alleles in the known loci have been found for 10 years or longer in this population.
Multiple genotype infection (MGI) is one of the important indexes of population genetics and epidemiology of malaria parasites because MGI is the first step in recombination of the parasite genome between different clones. In the case of P. falciparum, the rate of MGIs per population is basically associated with the endemicity [4], [5]. That is, the MGI rate of P. falciparum population is higher in high transmission areas and lower in low transmission areas. However, this is not the case with P. vivax populations because high MGI rates were observed among the P. vivax populations in low transmission areas [11], [12]. This feature could be attributed to relapse owing to hypnozoites in the liver of a vivax malaria patient. Although MGI is an important index, the methods or criteria of determining MGI is problematic. When any locus of the 10 loci showed more than 1 allele, we regarded the isolate as an example of MGI. Using this method, 85 (97.7%) out of the 87 isolates were MGIs. Focusing on each locus, the MGI rate per locus varied from 0.0% to 83.9% (average 29.1%) (Table 1). Focusing on the number of MGI loci per isolate, we found an interesting distribution pattern, similar to an F-distribution (Fig. 2). In the present study, the highest frequency of MGI loci per isolate was 2 (found in 25 isolates). The frequency decreased gradually, that is, 3 MGI loci; 19 isolates, 4 MGI loci; 12 isolates, 5 MGI loci; 6 isolates, and so on. We suspect that this distribution pattern may vary in each endemic area with different endemicity.
In the case of P. falciparum populations, the levels of genetic diversity are normally associated with the levels of malaria endemicity. That is, the levels of genetic diversity of the parasite populations are higher in high transmission areas and lower in low transmission areas [4], [5], although some exceptions have been reported [36].
We suspect that there will also be some association between the levels of genetic diversity and the levels of malaria endemicity in P. vivax populations, even though the correlation between these factors is not clearly understood at the time of writing. In P. vivax populations, the levels of genetic diversity tend to be high even in low transmission areas [11]–[13], [28]. This tendency can likely be attributed to unique biological features of P. vivax, such as early gametocytogenesis and relapse. Early gametocytogenesis may enhance the efficiency of transmission to Anopheles mosquitoes, allowing transmission to occur before symptoms appear – or, more importantly, before antimalarial drugs are administered. Relapses may also enhance the transmission and increase the genetic diversity of P. vivax populations, because the relapse will increase the probability of the coexistence of multiple genotype clones in a single patient, which are subsequently sucked up by an Anopheles mosquito in a single meal. Thus, the levels of genetic diversity of P. vivax populations could be higher than those observed in P. falciparum populations even in low transmission areas.
In the present study, the levels of genetic diversity of the South Korean population between 1994 and 2000 (when the number of malaria cases increased) were relatively lower than the levels of genetic diversity between 2001 and 2008 (when the number decreased). On the contrary, the levels of multilocus LD in the population between 1994 and 2000 were relatively higher than those between 2001 and 2008. These results suggested that the latter population was more genetically diverse and had less inbreeding. Furthermore, we divided the population into 3 groups each covering 5-year periods (1994–1998, 1999–2003 and 2004–2008) and reexamined the levels of genetic diversity and multilocus LD. Then, we again observed that the levels of genetic diversity in the populations had gradually increased, whereas the levels of multilocus LD had gradually decreased even though there was still strong multilocus LD in the most recent population (2004–2008). This result was surprising to us because we expected that the effective population size of the latter population would have decreased due to the reduction in the number of alleles in the population. However, this was not the case. In the South Korean populations, the association between the diversity and the endemicity of the P. vivax population is elusive.
There are, however, at least two possible explanations for this result. One is that the levels of genetic diversity of the P. vivax population increased in North Korea from 2001 to 2002, while the number of vivax malaria cases was very high (296,540 cases in 2001, 241,190 cases in 2002) [1]. Some of the isolates might have then been introduced to South Korea from North Korea by Anopheles mosquitoes. The other possible explanation is that the genetic diversity began accumulating in the South Korean population after the re-emergence in 1993. If the latter hypothesis is correct then the malaria control program conducted by the South Korean government might not have affected the parasite population structure.
One of the clear differences between the P. vivax population in South Korea and populations in tropical and subtropical areas is the pattern of transmission: in South Korea, vivax malaria is seasonally prevalent with a peak during July and August and no transmission in the winter season [37] and very long incubation periods with 8 to 13 months [14], suggesting that the chance for the recombination of the genome is limited to specific time periods within the year, possibly once or at most twice a year. In fact, we found strong LD in the South Korean population, suggesting that the frequency of recombination in this population would be very limited. However, these results might be associated with the location of the examined MS DNA loci: the 6 loci are in a gene coding a protein and another locus is between 2 genes coding respective proteins. In tropical and subtropical areas, on the other hand, vivax malaria is prevalent throughout the year, and thus recombination may occur throughout the year; this would lead to an increase in the levels of genetic diversity in tropical and subtropical areas. Indeed, in the populations from the Brazilian Amazon, identical haplotypes were rarely observed two years in a row, even in the same endemic area [12]. This would suggest that frequent recombinations occurred between the clones in the population.
The present study showed evidence of a low recombination rate and low frequencies of strand-slippage events of the microsatellite loci during mitotic replication in the P. vivax population of South Korea in comparison to populations in tropical and subtropical areas [12], and demonstrated that the 2 dominant haplotypes (H16 and H25) had been transmitting for several years (H16; 1996, 1998–2001, 2005, H25; 1995–2001, 2003) (Table 5). This continuous existence of the same haplotypes for several years is definitive evidence of a low recombination rate in the South Korean P. vivax population.
This continuous existence of the same haplotypes could be explained by a local adaptation to vector species. According to Joy et al. [38], for example, P. vivax in southern Mexico was genetically differentiated into 3 populations. They suggested that this differentiation would be the result of adaptation to different Anopheles species. On the other hand, in South Korea, Anopheles sinensis is a main vector of P. vivax and the other Anopheles species are very minor. Therefore, continuous existence of the predominant haplotypes could not be explained by a local adaptation to certain vector species in this country. There might be some other advantages of these haplotypes, or simply, the variation in the P. vivax population on the Korean peninsula had been very small owing to an effective national eradication program conducted by the National Malaria Eradication Service under the operation of the South Korean government with the support of the WHO in the 1970s [14]–[16].
Although the predominant haplotypes (H16 and H25) and their relatives had been transmitting in the DMZ for a long time, their transmission ended in 2005. We speculate that these predominant haplotypes were probably eliminated by the malaria control programs conducted by the North Korean government. In fact, according to the WHO World Malaria Report 2010, the number of vivax malaria cases in North Korea decreased substantially (2001: 296540 cases, 2005: 11507 cases). The reason for this reduction was not mentioned in detail, however this is probably due to the effect of mass drug administration by the North Korean government supported by South Korea. We suspect that the population structure of P. vivax in North Korea was changed dramatically and that these predominant “old” haplotypes were eliminated completely or became very minor in both the North Korean and South Korean populations. We suspect that the South Korean P. vivax population is a subpopulation of the North.
Our previous genetic epidemiological analyses of the South Korean P. vivax population using antigenic molecules [21]–[25] and the mitochondrial genome [39] showed that there were 2 types (or groups) of parasite populations in the endemic area. In these previous studies we examined groups of isolates collected from vivax malaria patients in the DMZ in 1997 [21], 1998 [22]–[24], 1999 [39]. In the present study, we examined isolates collected from patients in the DMZ between 1994 and 2008 using 10 highly polymorphic microsatellite loci. Once again, we observed two types of parasite populations (Fig. 4). However, some other haplotypes (clones) have been observed in the endemic area since 1998. The new haplotypes were genetically different to the 2 major groups that have been transmitted since the beginning of the re-emergence (Fig. 4). This finding was consistent with the results of analyses by Choi et al. using the DNA sequences of 2 antigenic molecules (circumsporozoite protein, merozoite surface protein-1) of isolates collected in the DMZ from 1996 to 2007 [40]. They also reported that new genotypes have been observed since 2000 and that the new genotypes had been rapidly disseminated in the endemic area.
The genetic differences between the 2 major groups and the new haplotypes in our data suggested two possibilities: the new haplotypes could have arisen in the DMZ in South Korea through recombination between existing clones in the population; or their emergence could be attributed to a continual introduction of P. vivax from other population sources, probably from North Korea. The present study suggested a low recombination rate in the South Korean population and would seem to indicate that the latter possibility is more likely.
A less likely possibility is that all of the isolates examined in this study were continually introduced from North Korea because all the isolates were collected from South Korean soldiers who served in the DMZ. These patients were normally treated by chloroquine within 4 days of the onset of the symptoms, and then treated by primaquine as a radical cure. The recurrence rate (both new infection and relapse may be included) of vivax malaria among them is 1.6% (62 cases of 3881 cases) and the definitive relapse rate is only 0.2% (8 cases of 3881 cases) [41]. In addition, the incubation period of P. vivax on the Korean peninsula is very long (8 months to 13 months) [14] and the transmission is mainly in summer [38]. Moreover, the period of conscription is about 2 years. Therefore, it might have been very difficult to transmit continuously among the South Korean soldiers in the DMZ, leading to the high recombination rate of the genome within the parasite population of the study area.
There are a number of sampling limitations to the present study. The number of isolates per year was relatively small (2 to 14 isolates, average: 5.8 isolates/year), and the sample size during the years 2004–2008 was particularly small (2 to 7 isolates, average: 3.6 isolates/year). Moreover, all of the isolates used in this study were collected only from South Korean soldiers or veterans and not from civilians, whose proportion among vivax malaria patients in South Korea has been gradually increasing [18]. In order to overcome these limitations and more accurately estimate the current status of the parasite population in South Korea, it will be necessary to include new isolates collected from civilians in the endemic areas and to increase the sample size of recent years.
Although travel between South and North Korea is basically restricted and the malaria control programs in the two countries may not be the same, we suspect that the South Korean P. vivax population is a subpopulation of the North Korean population because the majority of malaria patients live near the border [42]. Anopheles mosquitoes can fly over the DMZ, and South Korean travelers are allowed to visit some parts of North Korea, such as Kaesong and Kumgang-san, which are very famous for sightseeing. Furthermore, from 2001 to 2009 the number of vivax malaria cases in North Korea ranged from twice as high as the number in South Korea to many times higher, indicating that the size of the parasite population in North Korea is probably larger. Thus, the inclusion of North Korean isolates in the analyses would greatly enhance the accuracy of the estimation of the parasite population structure and the transmission dynamics and provide a more complete picture of the P. vivax population in the Korean peninsula; unfortunately the feasibility of doing this is low.
In conclusion, molecular epidemiology using highly polymorphic DNA markers of the P. vivax population is a very useful tool for assessing the population structure and transmission dynamics of the parasites, the knowledge of which may lead to the effective control of vivax malaria in the respective endemic areas.
|
10.1371/journal.pntd.0006590 | A novel cell-free method to culture Schistosoma mansoni from cercariae to juvenile worm stages for in vitro drug testing | The arsenal in anthelminthic treatment against schistosomiasis is limited and relies almost exclusively on a single drug, praziquantel (PZQ). Thus, resistance to PZQ could constitute a major threat. Even though PZQ is potent in killing adult worms, its activity against earlier stages is limited. Current in vitro drug screening strategies depend on newly transformed schistosomula (NTS) for initial hit identification, thereby limiting sensitivity to new compounds predominantly active in later developmental stages. Therefore, the aim of this study was to establish a highly standardized, straightforward and reliable culture method to generate and maintain advanced larval stages in vitro. We present here how this method can be a valuable tool to test drug efficacy at each intermediate larval stage, reducing the reliance on animal use (3Rs).
Cercariae were mechanically transformed into skin-stage (SkS) schistosomula and successfully cultured for up to four weeks with no loss in viability in a commercially available medium. Under these serum- and cell-free conditions, development halted at the lung-stage (LuS). However, the addition of human serum (HSe) propelled further development into liver stage (LiS) worms within eight weeks. Skin and lung stages, as well as LiS, were submitted to 96-well drug screening assays using known anti-schistosomal compounds such as PZQ, oxamniquine (OXM), mefloquine (MFQ) and artemether (ART). Our findings showed stage-dependent differences in larval susceptibility to these compounds.
With this robust and highly standardized in vitro assay, important developmental stages of S. mansoni up to LiS worms can be generated and maintained over prolonged periods of time. The phenotype of LiS worms, when exposed to reference drugs, was comparable to most previously published works for ex vivo harvested adult worms. Therefore, this in vitro assay can help reduce reliance on animal experiments in search for new anti-schistosomal drugs.
| Schistosomiasis remains a major health threat, predominantly in developing countries. Even though there has been some progress in search of new drugs, praziquantel remains the only available drug. Probably the most important advance in the search for new drugs was in vitro transformation of cercariae and their subsequent culture. However, hit identification in compound screenings is exclusively tested in skin stage parasites and is only confirmed for more mature worms in a subsequent step. This is in part due to the lack of an easy culture system for advance-stage parasites. We present here a reliable and highly standardized way to generate LiS worms in vitro in a cell-free culture system. The inclusion of in vitro drug tests on advanced-stage parasites in initial hit identification will help to identify compounds that might otherwise be overlooked. Furthermore, the ability to continuously observe the parasite’s development in vitro will provide an important platform for a better understanding of its maturation in the human host. Taken together, this opens up new avenues to investigate the influence of specific cell types or host proteins on the development of Schistosoma mansoni and provides an additional tool to reduce animal use in future drug discovery efforts (3Rs).
| Schistosomiasis, a chronic and debilitating helminthic disease, is one of the most important neglected tropical diseases (NTD). The WHO estimates that more than 206 million people are currently infected and in need of chemotherapy world-wide [1]. Moreover, over 200,000 people die each year due to the sequelae of the disease [2, 3]. Among the parasitic diseases, schistosomiasis is often considered only second in importance to malaria [2] and thus a major public health menace. Therefore, the WHO aims to eliminate schistosomiasis as a public health problem globally by 2025 [4, 5]. Implementation of safe water, sanitation and hygiene (WASH) strategies [6], intensified case management, veterinary public health, vector control and mass drug administrations (MDAs) are all crucial in reducing the disease burden [5]. Of all these approaches, MDA dominates national control programs thanks to the excellent safety and efficacy profile of praziquantel (PZQ), the only currently available drug [7, 8], as well as its low cost per treated individual [9]. However, the reliance on PZQ also raises concerns about emerging resistance should the drug pressure increase. Resistance against PZQ has already been observed in experimental models [10] while first instances of decreased drug efficacy have been observed in the field [7, 11, 12]. To be prepared for the emergence of resistant strains of schistosomes and to support the elimination of schistosomiasis, new drugs and complementary strategies, such as vaccines, are of imminent importance. To identify new drugs and new vaccine targets, in vitro assays of larval and adult worm stages are paramount for high throughput testing and simultaneously for reducing reliance on in vivo or ex vivo experiments in accordance with the 3Rs (replacement, reduction and refinement) of animal testing [13].
The current cultivation protocols for in vitro generated larval stages such as the schistosomula rely on the supplementation of fetal calf serum (FCS) for short-term culture [14, 15] and on the supplementation with FCS, human serum (HSe), erythrocytes and peripheral blood mononuclear cells (PBMCs) for long-term culture and in vitro juvenile worm development [16]. Such non-standardized culture conditions are prone to variability of serum batches, rely on the continuous supply of fresh human blood, and make it difficult to isolate pure schistosomula-derived soluble antigens during in vitro culture or to investigate the role of specific serum proteins in the development of the parasite. In addition to these hindrances and to avoid inter-assay or inter-laboratory fluctuations, well-defined and standardized culture conditions independent of serum and cell supplementations are needed. Up to now, long-term culture has been reported to rely on costly, non-commercially available and highly complex culture media such as Basch medium 169 which are difficult to implement in routine use and prone to batch variability [17]. Thus, simplifying culture conditions as well as the generation and handling of advanced-stage parasites opens new possibilities in the search for new drugs and facilitates the upscaling of existing drug screening strategies.
To establish a robust in vitro assay, it is important to continuously imitate the parasite’s in vivo development. This development within the final host is quite complex and occurs over a period of six to seven weeks. After penetration of the skin, cercariae transform to schistosomula which remain in the skin for about three days before migrating through the host’s vasculature. These skin stage schistosomules (SkS) traverse the capillaries of the lung where the majority of the parasites can be found on day 7 after infection [18]. To facilitate their migration, these SkS become longer, more slender and active and are hence called lung stage schistosomula (LuS). These continue their journey to the portal and mesenteric veins following the bloodstream [19, 20] where they continue to undergo morphological changes. The bifurcated gut is fused, the parasites initiate feeding, and continue to grow. This early liver stage (early LiS) is followed by the late LiS characterized by a drastic increase in length and prominent oral and ventral suckers. These juvenile worms (late LiS) then start to pair up and become fertile upon which the oviposition starts approximately 35 days after infection [21].
Treatment with PZQ, although initially with good efficacy, does not diminish the high reinfection rates encountered in the field [22]. This is partly due to the inability of PZQ to efficiently target early larval stages of the parasite [23, 24]. In the current search for new anti-schistosomal drugs, a two-step strategy is used. Firstly, SkS are tested immediately after their transformation from cercariae and then, once an active compound has been identified, it is tested mainly on ex vivo cultured adult worms that have been isolated from infected hamsters or mice [14, 25–27]. Other larval stages like the LuS, early and late LiS are omitted as potential drug targets. Therefore, future compounds with an activity predominantly directed against juvenile and adult stages might be overlooked. Thus, a highly standardized and robust way to generate advanced larval stages of Schistosoma mansoni would provide an opportunity to incorporate initial advanced-stage schistosomula testing into the current drug screening strategies to meet the desired target candidate profiles (TCP).
We disclose here a new serum- and cell-free cultivation method of newly transformed schistosomula (NTS) up to the LuS (7 days p.t.) and a cell-free cultivation method up to late LiS worms (starting to show 28 days after transformation) of S. mansoni. Our culture system allows the detection of stage-dependent differences in the activity of drugs with known anti-schistosomal properties, such as PZQ, Oxamniquine (OXM), the standard drug to treat schistosomiasis caused by S. mansoni prior to the advent of PZQ, and two antimalarial drugs that have recently been described to have anti-schistosomal properties, Mefloquine (MFQ) and Artemether (ART) [16, 28]. Therefore, this highly standardized, straightforward and reliable culture method is a basis for integrating drug screenings of advanced larval stages into initial hit identification in the search of novel schistosomicidal drugs.
Cercariae of an in-house Brazilian or imported NMRI strain of S. mansoni were harvested from infected Biomphalaria glabrata snails and used for mechanical transformation into NTS as described before [29]. Biomphalaria glabrata snails infected with the NMRI strain were provided by the NIAID Schistosomiasis Resource Center of the Biomedical Research Institute (Rockville, MD) through NIH-NIAID Contract HHSN272201700014I for distribution through BEI Resources and used for experiments as indicated due to temporary limited availability of the Brazilian strain. Briefly, cercariae were incubated 30 min on ice then centrifuged at 1932 x g for 3 min at 4°C. The pellet was resuspended in Hank’s balanced salt solution (HBSS) (Cat. No. H6648, Sigma-Aldrich, Germany) supplemented with 200 U/ml Penicillin and 200 μg/ml Streptomycin (Cat. No. P4333, Sigma-Aldrich) and transformed by mechanical stress applied by pipetting and vortexing, which was confirmed by microscopy (10x magnification). Separation of tails and cercarial bodies was accomplished by repeated sedimentation in ice-cold HBSS. The NTS were then transferred to the various culture media and kept overnight to complete transformation.
NTS (~100 NTS in 150 μl) were maintained in HybridoMed Diff 1000 (HM) (Cat. No. F 8055/1, Biochrom GmbH, Germany), Medium 199 (M199) (Cat. No. M4530), Dulbecco’s Modified Eagle Medium (DMEM) (Cat. No D5796) or RPMI 1640 (Cat. No. R8758, Sigma-Aldrich, Germany) supplemented with 200 U/ml penicillin, and 200 μg/ml streptomycin (Sigma-Aldrich) in a 96-well flat bottom tissue culture plate (Cat. No. 353075, Corning Incorporated, USA) and incubated at 37°C in 5% CO2 and humidified air for up to 4 weeks. For each condition, experiments were performed in triplicates. The medium was replaced every 7 days. Scoring was performed on day 1, 3 and 7 post-transformation (p.t.) and then again weekly until week 4 p.t.
The viability of NTS was scored using an Axiovert10 microscope (Zeiss, Germany). The scoring system was adapted from the Swiss TPH [29] and the WHO-TDR [27]. For the scoring, three main criteria were assessed: motility, morphology and granularity. The score was applied ranging from 0 (dead parasites, no movement, heavy granulation, blurred outline, rough outer tegument and blebs) to 1 (very reduced motility, rough outer tegument with some blebs) to 2 (reduced motility or increased uncoordinated activity, slight granularity, intact tegument with slight deformations), and finally 3 (regular smooth contractions, no blebs and a smooth outer surface, no granulation with clear view of internal structures which are visible under bright field microscope). Single NTS die during the mechanical transformation and are, thus, present from the beginning. The amount of dead parasites is taken into consideration when applying the viability score, thereby lowering the overall score. The score represents an overall impression of the visual appearance of all NTS in a well. In order to capture any subtle changes in the appearance of all schistosomula per well, viability scores 0–3 were further subdivided into 0.25 steps (e.g. 0, 0.25, 0.50, 0.75). For determination of larval development stage, morphological characteristics were used and based on previously published works [17, 21]. The skin stage was characterized by resembling the cercarial head in shape and undirected regular movements, the lung stage by an increase in length and decrease in diameter. The early LiS was characterized by growth of the parasite initially in diameter and then further in length and the clear visibility of the gut. The late LiS showed clearly identifiable oral and ventral suckers and a further increase in length, especially of the body past the ventral sucker.
Blood sampling of HSe was prepared from blood of consenting healthy volunteers with no previous history of schistosomiasis upon written consent. Fresh blood was left at room temperature for 30 min to clot, then centrifuged at 1845 x g for 20 min and serum was collected and pooled from 6 individuals and stored at -20°C until further use. NTS were incubated (100 in 150 μl) in Basch-Medium 169 [30] kindly provided by Prof. C. Grevelding and Dr. T. Quack (Universität Giessen, Germany), DMEM and HM supplemented with 200 U/ml Penicillin and 200 μg/ml Streptomycin and with HSe in different concentrations (1–50%) or 20% heat-inactivated FCS (Cat. No. F7524, Sigma-Aldrich, Germany) at 37°C for 8 weeks in a 96-well plate. Medium without HSe supplementation served as controls. The medium was changed weekly, and viability was scored on day 1, 3 and 7 p.t. and then again, every 7 days. Developmental stages were determined by bright field microscopy using an inverted Axiovert 10 microscope (Zeiss). For each condition experiments were performed in triplicates.
PZQ (kindly provided by Merck, Germany), OXM (Cat. No. PH002704), MFQ (Cat. No. M2319) and ART (Cat. No. A9361, Sigma-Aldrich, Germany) were dissolved in DMSO depending on drug solubility (PZQ 10 mg/ml, OXM 5 mg/ml, MFQ 33.3 mg/ml, ART 10 mg/ml) and stored at 4°C until use. NTS were cultured in HM supplemented with (SkS, LuS and LiS) or without (SkS and LuS) 20% HSe. To test the drug sensitivity of the distinct developmental stages, we incubated SkS (24-hour-old NTS), LuS (7-day-old NTS) and LiS (6-week-old NTS) parasites with PZQ, OXM, MFQ and ART at different concentrations (1, 10, 100 μg/ml). Before the addition of the drugs, a medium exchange was performed. For each condition experiments were performed in triplicates. Scoring was performed before (0 h) and 3, 24, 48, 72 and 168 h after treatment (a.t.).
For statistical analysis of the experiments to determine the optimal culture medium as well as the HSe concentration, a Kruskal-Wallis test was performed on day 1 and 3, week 1, 4 and 8 p.t. and if significant (p ≤ 0.05), the data was further analyzed by employing Mann-Whitney U tests comparing pure medium with serum-supplemented medium followed by a Bonferroni correction. For the statistical analysis of the experiment comparing HM to Basch medium 169, testing was performed in the same manner except that Mann-Whitney U testing was used to compare HM to Basch medium 169 for each of the shown conditions. For statistical analysis of the drug treatment experiments, Mann-Whitney U testing was performed to determine statistically significant differences between the DMSO control and the treated groups. Statistical testing was done with IBM SPSS Statistics 24 (IBM). Graphs were made with PRISM 5 (GraphPad).
To generate and maintain NTS under serum- and cell-free conditions, we cultivated NTS immediately following transformation in different highly standardized and commercially available culture media such as HM, DMEM, RPMI and M199 (Fig 1A and 1B) supplemented only with 200 U/ml penicillin and 200 μg/ml streptomycin. Over a period of four weeks, regular visual viability scoring of the parasites was performed following defined criteria adapted from previously published works [27] (S1 Fig). In DMEM, NTS survived well for at least four weeks. Their viability peaked on day 3 p.t., (2.6 ± 0.1), characterized by an increased motility and hardly any granularity. Then viability declined slightly (2.0 ± 0.0) due to the death of single NTS and slight internal granulation on day 7 after which the viability stabilized throughout the remainder of the experiment (Fig 1A and 1B). In HM, NTS initially scored (2.0 ± 0.3) on day 3 p.t., slightly lower compared to those in DMEM. However, from one week p.t. onwards, NTS in HM had already surpassed those kept in DMEM in the viability scoring (2.7 ± 0.3) due to the absence of granularity and increase of regular, steady movement. NTS in HM stayed viable throughout the experiment (four weeks) (Fig 1A and 1B). Compared to these two media (HM and DMEM), M199 and RPMI supported viability of NTS rather poorly. In M199, viability already started to drop (1.8 ± 0.3) on the third day p.t., characterized by the majority of the NTS being heavily damaged or already dead. By week 3 p.t., all NTS had died (Fig 1A and 1B). In RPMI, the viability declined even faster with the majority of the NTS already heavily damaged or dead on day 3 p.t. and the death of all NTS by week 2 p.t. (Fig 1A and 1B). In all conditions, NTS were initially oval shaped on day 1 to day 3, resembling the SkS. Skin-stage schistosomula were characterized by a stumpy looking oval outline with regular elongations and contractions either straight or to the side and were 105.7 ± 7.8 μm in length (in HM) (S2 Fig). By one week p.t., the surviving NTS had developed LuS characteristics. LuS NTS were characterized by a more elongated and slender form with an increased activity compared with the SkS. The parasites’ movements were still characterized by regular elongation and contraction in changing directions with an average length of 184.4 ± 45.5 μm (Fig 1B and S2 Fig) [31]. In all media, however, development of surviving NTS halted after the first week of culture and, therefore, remained at the LuS stage for the remainder of the experiment. Dead NTS, in M199 or RPMI medium, displayed massive granulation internally and an irregular outer tegument. In contrast, healthy NTS, in HM and DMEM, were elongated and slender with a clear interior and a well-contrasted outline (Fig 1B). Taking all of this into consideration, HM is the most suitable medium for long-term culture of LuS NTS without any serum or cell supplementation; however, development under these conditions is halted in the LuS.
To test whether, in the absence of any cell supplementation, LuS NTS could develop further into the LiS by adding serum of the most important definite host, namely the human [2, 32], we supplemented HM and DMEM with 20% HSe. In both serum-free controls, we observed a decline in viability after the fourth week of observation as before (Fig 2A). More specifically, in HM and DMEM, all SkS that survived, developed to the LuS by the first week p.t. Until the fourth week p.t., around 49% (48.7 ± 4.2 dead out of 99.7 ± 10.6 total parasite count/well) of the NTS died in DMEM compared to only around 21% (39.0 ± 9.0 dead out of 185.7 ± 14.0 total) in HM. Following the fourth week p.t., a steady increase in dead NTS was present in both conditions (Fig 2B and 2D). However, the addition of 20% HSe to the culture halted the loss of viability and greatly reduced NTS-death within the first 4 weeks of transformation. Interestingly, the next developmental stage, the early LiS started to develop from 2 weeks p.t., as evident by the further increase in size (growing wide and stumpy rather than in length) and the gut becoming clearly visible under the light microscope. By week 4 p.t., the parasite had developed further into the late LiS, characterized by clearly visible ventral and oral suckers and elongation of the aboral part of the parasite’s body. Its activity could be either restricted to the oral or aboral part of the body or could encompass the entire worm. This stage was characterized by a steady increase in length (1289.6 ± 247.0 μm) after 6 weeks of culture (S2 Fig). The time point of this development was the same for HM and DMEM. Even though the percentage of the early LiS in DMEM (29.4% or 34.0 ± 2.8 out of 115.5 ± 6.4 total) was higher than that in HM (6.1% or 9.3 ± 1.2 out of 153.0 ± 20.0), the percentage of the late LiS was comparable between DMEM (11.5 ± 5.0 out of 115.5 ± 6.4 total) and HM (15.3 ± 3.2 out of 153.0 ± 20.0) with around 10% having developed after 8 weeks of incubation (Fig 2C and 2E). Importantly even though the initial time point for the occurrence of the late LiS in HM and DMEM was the same, the overall growth rate of the parasites in HM supplemented with HSe was increased compared with that in DMEM since the late LiS in HM (week 4 and after) grew to be remarkably bigger (Fig 2F, arrows). Strikingly, despite the addition of HSe, a larger number (~ 40%) of NTS had died in DMEM by 8 weeks (Fig 2F) as compared to HM, where only ~10% of NTS were dead.
Since one of the media currently considered to be the “gold standard” to raise advanced developmental stages is Basch medium 169, we cultured NTS (derived from the S. mansoni NMRI strain) with Basch medium 169 and HM medium without further serum supplementation. Interestingly, NTS raised in HM again did not develop past the LuS, whereas in Basch medium 169, the development to the early LiS stage was observed starting from week two onwards. Furthermore, in Basch medium 169 the viability score (2.3 ± 0.1 at week 3 p.t.) was around 0.5 score points higher than in HM (1.9 ± 0.4 at week 3 p.t.) for the first 3 weeks of the experiment. This was followed by a slight drop in viability in Basch medium 169 in the fourth week with a final score of 1.9 ± 0.1 for Basch medium 169 and 1.8 ± 0.3 for HM. Upon addition of FCS, the currently most commonly used serum supplement in in vitro culture of S. mansoni [33–35], we could observe a drastic decline in viability for both media within the first week after transformation starting after the first three days of culture (0.8 ± 0.1 in HM and 0.7 ± 0.1 in Basch medium 169) (Fig 3A). In Basch medium 169 supplemented with 20% FCS the rate of development was reduced and delayed with only 8.7% early LiS by week 4 compared to 58.9% in unsupplemented medium. (Fig 3B and 3C). FCS supplementation to HM lead to an increase in dead NTS (from 38.7% to 79.4% dead parasites on week 4 p.t.) and was unable to promote development past the LuS. (Fig 3E and 3F). The addition of HSe to Basch medium did not further increase the viability score in contrast to the addition of HSe to HM which resulted in an overall increase of ~0.5 viability score points (2.3 ± 0.1 score points on week 4 p.t.) (Fig 3A). HSe was able to promote development to the late LiS in both media, and the first late LiS were observed 3 weeks p.t. in Basch medium 169 and 4 weeks p.t. in HM. However, the percentage of developing late LiS was much higher in HM (46.7%) compared to Basch-medium 169 (23.9%) (Fig 3D and 3G) indicating that addition of HSe to HM had a more pronounced effect on promoting larval development towards juvenile worms.
Next, we investigated whether HSe induced development of advanced larval stages of S. mansoni increases in a concentration-dependent manner. Therefore, we cultured NTS in HM supplemented with 1, 5, 10, 20 or 50% of HSe. Even at the lowest concentration (1%), HSe successfully prevented the death of NTS which otherwise started to occur in the 5th week p.t. (Fig 4A) and survival rates remained steadily above 67% (49.0 ± 7.4 dead in 151.7 ± 15.3 total) throughout the experiments (8 weeks) in all HSe-supplemented conditions (Fig 4B–4D). Surprisingly, however, at a concentration of 50% HSe, the survival rate (69.1% or 44.7 ± 10.1 dead out of 144.7 ± 21.2 total) as well as viability (2.2 ± 0.6) started to decline around week 8 (Fig 4D and 4E). Nevertheless, the final viability score and survival rate was still higher than in the control (viability score of 0.3 ± 0.0 and survival rate of 1.8% or 160.3 ± 7.8 dead out of 163.3 ± 8.5 total) (Fig 4A and 4E). However, development of NTS up to early and late LiS was only observed at higher concentrations (single early LiS starting in 5% HSe and late in 20% HSe) of HSe (Fig 4C, 4D and 4F). Early LiS development could be observed starting 2 weeks p.t. in all concentrations but their number LiS increased concentration-dependently. The first parasites in the late LiS were detected at 4 weeks p.t. in 20% HSe and single worms even one week earlier in 50% HSe. The overall percentage of late LiS parasites increased with rising serum concentrations, and, finally, around 10% LiS had developed (15.3 ± 3.2 in 153.0 ± 20.0 total parasite count/well) in 20% HSe compared to approx. 18% (25.3 ± 4.2 in 144.7 ± 21.2 total) in 50% HSe-supplemented HM (Fig 4C and 4D). Since viability decreased in 50% HSe-supplemented medium after the 7th week p.t. and sufficient numbers (13–19 late LiS/well) of late LiS worms were generated at 20% HSe [26, 28], we decided to use 20% HSe supplementation for the generation of parasites for drug screening experiments.
To test the different stages (SkS, LuS and late LiS) generated with the new method in HM and 20% HSe supplementation for the in vitro drug testing, we chose drugs with known anti-schistosomal properties [28, 36]. Late LiS (6-week-old schistosomula) (Fig 5A, 5C, 5E and 5G), LuS (7-day-old NTS) (Fig 5B, 5D, 5F and 5H) and SkS (24-hour-old NTS) (S3A–S3D Fig) NTS were cultured in the presence or absence of different concentrations (100, 10 and 1 μg/ml) of PZQ (Fig 5A and 5B and S3A Fig), OXM (Fig 5C and 5D and S3B Fig), MFQ (Fig 5E and 5F and S3C Fig) or ART (Fig 5G and 5H and S3D Fig) and assessed by regular viability scoring.
PZQ treatment was detrimental to LiS and LuS at all concentrations, which was clearly evident from the drop in viability as soon as 3h a.t. and the death of almost all parasites on day 1 a.t. (Fig 5A and 5B), characterized by contracted parasites, damaged tegument and no detectable motility (S4 Fig). No such pronounced effect of PZQ was observed on SkS, however, only high concentrations of PZQ slightly reduced the observed viability starting at 3h a.t., but NTS viability then stabilized for the remaining part of the experiment (S3A Fig). In serum-free conditions, PZQ activity was comparable to serum-supplemented HM as observed on day 3 a.t. At that time point, the susceptibility of the LiS and LuS to PZQ compared with that of the SkS. Serum-free conditions did not alter the activity profile of PZQ on the SkS or LuS (Fig 6A). OXM and MFQ both had similar effects on the LiS at high concentrations. In particular, MFQ had a clear and strong anti-schistosomal activity at 100 μg/ml. In contrast to MFQ, OXM was also potent at 1 μg/ml. even more so than at 10 μg/ml (Fig 5C and 5E). MFQ showed similar activity on the SkS and LuS compared to the LiS (Fig 5E and 5F and S3C Fig). Overall, OXM was potent in reducing the viability of the parasites in the SkS at all tested concentrations. Even though the LuS was the least susceptible stage to OXM, a reduction in the viability could still be observed which was most prominent at 1 μg/ml (Fig 5D). The stage-dependent vulnerability of the schistosomula was observed quite well on day 3 a.t. It is worth mentioning that LuS and SkS NTS were slightly more susceptible to OXM in serum-free culture in the concentration of 100 μg/ml. Also, MFQ exerted an increased activity against both SkS and LuS NTS in serum-free compared to HSe-supplemented culture (Fig 6B and 6C). Closer and careful observations revealed that the highest concentration of OXM induced a hyperactive state directly following the addition of the drug at all stages, which was followed by heavy granulation and tegumental damage in the SkS and LiS already starting on day 1 a.t. and, in the LuS, slightly delayed starting one week a.t. Morphologically, MFQ caused almost instantaneous contraction as well as heavy granulation with a blurred, disintegrated outline of the parasite at 100 μg/ml (S4 Fig). ART did not show any effect on viability or a morphological alteration in LiS or LuS (Fig 5G and 5H and S4 Fig), but a slight reduction in viability of the SkS at 100 μg/ml (S3D Fig). On day 3 a.t., we could not detect a drop in viability or any morphological changes in any of the tested stages in HSe-supplemented medium (S4 Fig) or in serum-free medium. However, in the serum-free medium we could detect a clear schistosomicidal effect at 100 μg/ml, with death following an initial paralysis of the parasites. However, upon addition of the drug to serum free-culture, the previously dissolved drug precipitated to a small extent (Fig 6D). Taken together, we could show the varying stage-dependent activity profiles of selected compounds already known for their anti-schistosomal properties and no schistosomicidal activity for artemether.
In the London Declaration of 2012, the world, represented by partners in governments, academia, NGOs, pharmaceutical companies and more, committed itself to accelerate the control, elimination and eradication of 10 NTDs by 2020 [37]. Despite successes in several areas, progress towards the elimination of schistosomiasis has remained rather limited [38]. Reasons for this lack of progress range from challenges in schistosome vector control to limited developments in drug discovery programs. Rather than developing new compounds, the scarce resources are focused towards drug repurposing. The limited interest of pharmaceutical companies in researching entirely new compounds is due to the resource- and time-consuming process to get an approval by the health authorities reviewed in 2014 by Panic et al. [39]. Another limitation in in vitro screening used for testing compound libraries is the availability of in vitro-generated advanced stage parasites. Therefore, current in vitro compound screenings of S. mansoni still mostly focus on SkS schistosomula observed for up to 72 h after treatment for initial hit identification or on adult worms retrieved from infected mice for hit confirmation [14, 40, 41]. This focus leaves a “blind spot” for drug efficacy toward intermediate developmental stages of the parasite. The value of monitoring drug efficacy for initial hit identification at all consecutive stages of parasite development in parallel is clearly illustrated by the stage-specific efficacy of PZQ, which shows decreased efficiency between 1 day and 5 weeks of development with minimal worm burden reduction at the age of 4 weeks in vivo [42, 43], but still represents the most effective and widely used schistosomicidal drug available [44]. Studies on ex vivo harvested LuS schistosomula sometimes manage to bridge this gap, but are challenging and require a living host (invoking practical as well as ethical aspects) [45].
In vitro S. mansoni culture systems obviously can circumvent some of these challenges and allow for the continual observation of consecutive larval stages. In vitro transformation and culture techniques have first been established in the 1970s and 1980s [17, 46] and since then different culture media have rarely been evaluated [47]. Until today, those techniques rely on the presence of serum (mainly FCS) for early larval stages [48, 49] and the addition of human blood cells to generate advanced larval and juvenile worm stages [16, 50]. Viability scoring relies on visual assessment of the larvae in culture by bright field microscopy, which still is the gold standard for drug efficacy tests [27, 28]. While the method has its merits and can be highly effective for dedicated drug efficacy testing, it is very labor intensive. At the same time, microphotographic-based automated analysis [51] using algorithms is complicated by the presence of large numbers of RBCs and/or PBMCs that overlay or co-localize with larvae, making reliable assessment of tegument damage, for example, something for the “trained eye” only. Moreover, the repeated addition of fresh cells from human donors is a factor that is difficult or impossible to fully standardize.
In this study, we generated a novel cell-free in vitro assay that allows the development and long-term observation of S. mansoni larvae. Mechanically transformed NTS [25, 28] were successfully cultured in a serum- and cell-free culture medium for up to four weeks, allowing for their development to LuS schistosomula but not further. Supplementation with 20% FCS decreased the viability after the third day of cultivation and the developmental block persisted throughout the period of cultivation. Most recently published in vitro NTS assays that have reported the use of FCS as supplement in either Medium 169 or Medium 199 do not score viability after 3 days of culture [33–35]. Thus, the relative loss in viability in FCS supplemented HM after three days needs to be explored further since batch differences in FCS are a well-known phenomenom in cell culture assays and could explain our observations as well. However, the addition of HSe to the culture broke the developmental block at the LuS and propelled development to LiS worms, and survival for at least eight weeks. Thus, this cell-free culture for advanced stage S. mansoni represents important progress in existing tools that rely on the addition of “non-standardized” human blood cells [16]. Instead, the commercially available HM, normally used to cultivate hybridomas and other cell lines, is highly standardized, supplemented with transferrin, insulin, bovine serum albumin (BSA)/oleic acid complex, absorbable amino acids, D-glucose, vitamins and minerals. The presence of albumin and insulin may be critical for the prolonged survival (≥ 4 weeks) of parasites under serum-free conditions, and indeed, earlier findings showed that schistosomula can ingest and digest albumin and IgG in vitro, utilizing them as nutrient sources [52]. In addition, it was shown that S. mansoni has two insulin receptors (SmIR1 and SmIR2), essential for the survival of the pathogen in vitro as well as in vivo [53]. Nevertheless, and in contrast to the natural development in vivo, NTS development halted in serum-free insulin-containing media at a stage that, in terms of size and morphology, closely resembles that found in the lung vasculature [21]. This indicates that other factors than insulin or albumin contribute to the further development beyond the lung-stage. This is possibly not so surprising considering that the parasite is blood-dwelling and therefore permanently surrounded by the blood components. Nevertheless, FCS which is widely used in in vitro culture [33–35] could not see development past the LuS. Cattle, in contrast to humans, is an atypical definitive host for S. mansoni, and nevertheless an important alternative definitive host [2, 32, 54]. HSe did not only improve overall viability but also induced development of NTS to LiS parasites in a concentration-dependent manner. When compared to ‘gold standard’ Basch medium 169, which is a non-commercial mixture of 15 components which needs to be freshly prepared before use [17], HM supplemented with HSe yielded comparable viability scores but increased the development of LiS schistosomes in vitro. Furthermore, we observed a difference in the rate of development between two strains of S. mansoni, whereby late liver stage parasites were detected in a higher proportion in the NMRI compared to the Brazilian strain starting from the 4th week p.t. (Fig 3A and 3B). Differences in the dynamics of development of several S. mansoni strains within a host have been previously demonstrated using an in vivo mouse model [55, 56].
Juvenile worms could be maintained for extended periods of time (more than a year although no pairing or egg production could be observed) through the addition of HSe to HM alone in contrast to non-supplemented controls. However, concentration-dependent advances in stage development (in 50% HSe) came at the price of increased larval mortality at seven weeks and onwards. This may be explained by the increased numbers of parasites reaching the LiS, their increased size and corresponding increased rate in medium nutrient depletion. This would be in line with the fact that the parasite’s feeding and cell divisions are limited, both indicators of low metabolism rates until it reaches the LiS about 15 days after infection in vivo [21]. We found supplementation with 20% HSe to provide an optimum between HSe consumption and frequency of medium changes, generating a reasonable number of late LiS worms for further studies. This allows the incorporation of all developmental stages in initial hit identification strategies for future drug screenings, making it possible to identify compounds with an activity profile predominantly targeting lung to liver stage parasites. Such drugs would probably otherwise be missed. Taken together, we show that in contrast to cercariae that are susceptible to the complement system present in serum [57], schistosomula not only tolerate the presence of HSe but require it to progress to late larval stages within their definitive host, the human.
Another aim of this study was to investigate the suitability of this method to screen compound libraries for schistosomicidal activity against, in particular, advanced stages of S. mansoni in addition to the mostly used early stages [28, 51]. NTS showed comparable stage-specific susceptibility to four drugs known to possess anti-schistosomal properties as seen previously with other in vitro assays [16, 28], confirming the validity of our novel culture method. For example, PZQ, undoubtedly the most important tool to control schistosomiasis in the field, [16, 58] was described to have reduced efficacy starting 24h after infection and lasting until day 35 after infection when susceptibility is increasing again [42, 43, 59]. We could confirm this in our novel culture with the SkS (day 1) and LuS parasites (day 7) being less vulnerable compared to late LiS worms (day 42) which is in accordance with previous publications although to a slightly lesser extent. This could be attributed on the one hand to the difference in the media that were employed (e.g. M199, RPMI 1640 or DMEM as compared to HM) [33–35] and on the other hand to the difference in S. mansoni strains used to generate NTS for in vitro drug testing [28, 60, 61]. The substantial tegumental damage and anti-schistosomal activity in our assay induced by OXM, the drug used against S. mansoni before the discovery of PZQ [62, 63], was also comparable to that observed by others [16, 28]. Specifically, morphological changes of the tegument as well as schistosomicidal activity, which was more pronounced against LiS and SkS schistosomula at high concentrations, could be observed. The increased activity in 1 μg/ml compared with 10 μg/ml of OXM was surprising. A similar observation has been made in a previous study [28], but the reason for this is still unknown. MFQ, a well-known anti-malarial drug, that was shown to also be active against S. mansoni in vivo and in vitro [64], is thought to impact the heme detoxification as well as glycolysis in the parasite [65]. Interestingly, the stage-dependent activity profile of MFQ is different to that of PZQ and more potently targets NTS than mature worms [29], something we were able to confirm in our novel assay as well. ART is another anti-malarial drug that was previously shown to act against S. mansoni [36] and is thought to act via toxication by the hemin byproduct of the parasite’s digestion of hemoglobin. Indeed, in the absence of a cellular source of hemoglobin and comparable to our cell-free system, ART lost activity, supporting the notion that hemin is required for the efficacy of the drug [66]. In the absence of serum during early developmental stages, we found strongly enhanced schistosomicidal activity of MFQ, slightly increased potency of ART, but mostly unchanged activity of PZQ and OXM. Bioavailability of MFQ is known to be strongly dependent on binding to plasma proteins (which can be as high as 98% [67]), and for ART plasma-binding lies between 92–98%, in line with our observed increase in activity in serum-free cultures [68]. On the other hand, PZQ binds to a lesser extent to plasma proteins (~80%) in concentrations of 10–100 μg/ml and in vitro as low as 50% [69], something that is also reflected in the limited increase in toxicity we observed in non-serum supplemented cultures. Importantly, our assay allows researchers to continually observe toxicity effects and inhibition of maturation from the SkS (day 1–3 p.t.) parasite over LuS (from day 7 p.t.) and early LiS (from day 14 p.t.) to the late LiS (from day 28 p.t.) in settings ranging from compound or drug testing, but also to screen for natural factors from non-permissive hosts or, reversely, to identify growth-promoting compounds from the host-adapted, parasite friendly environment.
Taken together, since late LiS worms generated in our cell-free assay showed drug-specific phenotypes for all drugs except artemether and responded similarly to ex vivo harvested worms from infected laboratory animals [28], this assay has the potential to reduce the reliance on in vivo generated worms by replacing ex vivo harvested worms in the initial hit identification and to thereby reduce costs and labor for large-scale drug screening assays. We are, however, aware that hit confirmation in ex vivo harvested worms might remain a necessity and that further studies such as comparative gene-expression of in vivo and in vitro generated parasite stages are ultimately necessary to clarify the extent of similarity. Eventually, the independence from host blood cells facilitates the automated assessment of larval viability in large-scale assays, due to a lack of visual interference by host cells. In addition, the high level of standardization will allow researchers to investigate and identify components within HSe that are exploited by the parasite for its development in the dominant definite human host and thus define mechanisms that underlie the host-specificity of this parasite. Ultimately, such understanding will pave the way for the identification of new drug and vaccine targets.
|
10.1371/journal.pgen.1007782 | Telomere length-dependent transcription and epigenetic modifications in promoters remote from telomere ends | Telomere-binding proteins constituting the shelterin complex have been studied primarily for telomeric functions. However, mounting evidence shows non-telomeric binding and gene regulation by shelterin factors. This raises a key question—do telomeres impact binding of shelterin proteins at distal non-telomeric sites? Here we show that binding of the telomere-repeat-binding-factor-2 (TRF2) at promoters ~60 Mb from telomeres depends on telomere length in human cells. Promoter TRF2 occupancy was depleted in cells with elongated telomeres resulting in altered TRF2-mediated transcription of distal genes. In addition, histone modifications—activation (H3K4me1 and H3K4me3) as well as silencing marks (H3K27me3)—at distal promoters were telomere length-dependent. These demonstrate that transcription, and the epigenetic state, of telomere-distal promoters can be influenced by telomere length. Molecular links between telomeres and the extra-telomeric genome, emerging from findings here, might have important implications in telomere-related physiology, particularly ageing and cancer.
| Telomeres (special DNA-protein assemblies that protect chromosome ends) affect ageing and diseases such as cancer. Although this has been recognized for many years, biological processes that connect telomeres to ageing, cancer and other cellular functions remain to be fully understood. Certain proteins, believed to be only telomere-associated, engage DNA outside telomeres. This raises an interesting question. Does telomere length influence how telomere-binding proteins associate with DNA at regions distal from telomeres. If so, how does this impact function? Motivated by these questions, in the present studies we tested if extra-telomeric binding of the well-known telomere-repeat-binding-actor-2 (TRF2) depends on telomere length. Our results show that the level of DNA-bound TRF2 at telomere-distal sites changes as telomeres shorten or elongate. Consequently, TRF2-mediated gene regulation affects many genes. Notably, histone modifications that dictate chromatin compaction and access to regulatory factors, at sites distant from telomere ends also depended on telomere length. Together, this links the state of telomeres to gene regulation and epigenetics directly in ways not previously appreciated that might impact a more complete understanding of molecular processes underlying ageing and cancer.
| The shelterin complex comprises a group of proteins that confer stability to telomeres[1–6]. Components of shelterin and other telomeric proteins, for example, RAP1, TRF1 and TRF2, and POT1 have been extensively studied in telomere homeostasis[7–10], DNA replication[11–15], repair and genomic stability[16–21]. Previous findings implicate shelterin members in non-telomeric functions related to their presence at distal regions outside telomeres[22–29]. Telomere-independent RAP1 was reported to influence global gene expression through associations with extra-telomeric DNA[23] and also by directly associating with the IkappaB kinase complex thereby inducing NFKappaB-dependent gene expression[24,30]. In addition to RAP1 engaging non-telomeric DNA, another shelterin component TRF2[31] was detected to bind DNA in about 200 sites genome wide[25,26]. Furthermore, TRF2-mediated transcription regulation was reported for PDGFR-beta[27] and recently promoter-binding of TRF2 was reported to repress the cyclin-dependent kinase p21 (CDKN1A/CIP1/WAF1) through TRF2-dependent recruitment of the REST-LSD1 repressor complex[28].
Together these implicate functions of shelterin proteins beyond telomeres. They also, contextually, raise the question about whether and how shelterin components might link telomeres to non-telomeric cellular functions. Herein we ask this question specifically focusing on TRF2. We examined telomeric and non-telomeric TRF2 occupancy in human cells with short vis-à-vis long telomeres. Proportional increase in TRF2 occupancy at telomeres was evident in cells with enhanced telomere length consistent with a previous report[32]. Importantly, TRF2 occupancy was significantly depleted at many non-telomeric promoter sites across the genome in cells with elongated telomeres. The promoters were located at varying distances from telomere ends ranging from ~3 to 60 Mb. In addition, reduced promoter TRF2 occupancy in cells with elongated versus short telomeres resulted in altered gene expression. We noted that epigenetic histone modifications at the promoters, though remote from telomeres, were dependent on telomere length; that is, increase (or decrease) in activation marks H3K4me1/H3K4me3, (along with reduction (or increase) in H3K27me3) in respective promoters consistent with activation/repression of the gene was observed in all cases. Functionally, this shows TRF2-mediated transcription dependent on TRF2 promoter occupancy in cells with elongated/short telomeres. Taken together, these findings suggest that epigenetic changes and transcription at promoters remote from telomeres are telomere-length dependent.
We sought to test TRF2 occupancy at sites distal to telomeres in cells with short vis-à-vis long telomeres. For this 8 gene promoters with extra-telomeric TRF2 binding reported by us and others were taken[26–28,33].In addition, another 15 gene promoters with putative TRF2 binding sites were randomly selected from TRF2 ChIP-seq in HT1080 cells such that the genes were located at varying distances from telomeres (~2 to 60 Mb from nearest telomere) and had TRF2 occupancy within 1.5 Kb of transcription start sites (TSS) (Fig 1A and S1 Table). Significant TRF2 occupancy was first confirmed at all the 23 promoters by ChIP-qPCR in fibrosarcoma HT1080 and normal human fibroblast MRC5 cells, with the exception of LINC01136 promoter, which was not enriched for TRF2 occupancy in MRC5 cells (Fig 1B and 1C). No TRF2 occupancy was found within 7 negative control regions (CTCF, GAPDH, b-actin, Synapsin and STAT2 promoters and 3’UTRs of p21 and SAMD14) in HT1080 and MRC5 cells (Fig 1B and 1C and S1A Fig).
TRF2 silencing by siRNA resulted in significantly altered expression of most target genes in both the cell types. Expression of all the 23 target genes changed in HT1080 cells (Fig 1D and S1B Fig); 20 genes were up-regulated and two genes PDGFR β and WRNIP1 were down-regulated (OPN4 did not change significantly; PDGFR β decrease on TRF2 silencing in cancer cells was also noted earlier[27]). In MRC5 cells, out of 22 target genes (due to lack of promoter TRF2 occupancy in MRC5 cells LINC01136 was excluded) expression of 18 genes altered significantly (Fig 1E and S1B Fig; differential expression of CHRM2, OPN4, PDGFR β and PSMA8 was not significant). WRNIP1 down-regulation was consistent in both cells lines. Increase in p21 and RPA2 expression in both cell lines on TRF2 silencing was consistent with previous reports[26,28] and OPN4 did not show significant change in expression in either cell line. Expression of negative control genes CTCF, b-actin, Synapsin and STAT2 did not change significantly upon TRF2 silencing in both cell types (Fig 1D and 1E and S1C Fig).
We first checked whether cells with long telomeres have more telomere-bound TRF2. The isogenic line of HT1080 cells with enhanced telomeres reported earlier[32] was used (designated as cells with long telomeres HT1080-LT cells in following text). We confirmed HT1080-LT cells had elongated telomeres compared to HT1080 cells (from ~4.5 Kb in HT1080[32,34,35] cells to average telomere length of ~8–9 Kb in LT cells; S2A and S2B Fig). Expression of both hTERT and hTERC genes, and telomerase activity was enhanced in HT1080-LT cells reaffirming the reported telomere elongation phenotype (S2C Fig). HT1080-LT cells with elongated telomeres had relatively more TRF2 occupancy at telomeres compared to HT1080 cells consistent with the earlier finding[32] (Fig 1F and 1G, S2D Fig). Total cellular and chromatin-bound TRF2 was roughly similar in HT1080-LT and HT1080 cells, and TRF2 in the nucleoplasm was significantly low compared to chromatin-bound TRF2 for the same amount of protein lysate suggesting nuclear TRF2 was mostly bound to chromatin (S2E Fig). TRF2 expression in whole cell lysate was found to be similar in HT1080 and HT1080-LT cells (S2F Fig).
Binding of TRF2 at the 23 target promoter sites validated earlier (Fig 1B) was compared in HT1080 cells with short versus long telomeres. TRF2 occupancy was significantly altered in cells with long telomeres in 17 of the 23 promoters. Notably, in all the 17 cases promoter TRF2 occupancy was depleted in HT1080-LT vis-à-vis HT1080 cells (Fig 1H). TRF2 occupancy did not change detectably in case of KCNH2, LINC01136, PSMC2, PTPN11, RYR2 and SMAD7; CTCF and GAPDH promoters with no TRF2 binding were used as negative controls.
To further test this we next used non-cancerous normal fibroblast MRC5 and corresponding isogenic cells with longer telomeres (made using a different mode of telomere elongation[36]). Repeated treatment of MRC5 cells with G-rich terminal oligonucleotides (GTR) over multiple passages (S3A Fig) resulted in telomere elongation as reported earlier[36]. We found ~2-3-fold elongation of telomeres in cells sequentially fed with GTR for either 7 or 14 cycles (oligo-fed OF7 or OF14 cells, respectively). Telomere elongation was from about 9 Kb in case of MRC5 cells[37] to average telomere length of either 18 Kb (MRC-OF7) or 27 Kb (MRC5-OF14) (Fig 2A and 2B and S3B Fig). Accordingly, increase in expression of hTERT and hTERC and enhanced telomerase activity was observed (S3C and S3D Fig).
Increase in telomere-bound TRF2 was found in the cells with long telomeres (OF7, OF14) relative to untreated MRC5 cells (Fig 2C and 2D, S3E Fig). Chromatin-bound TRF2 did not show significant difference in MRC5, MRC5-OF7 and MRC5-OF14 cells. In addition, similar to HT1080 cells, we noted that free TRF2 in the nucleoplasmic fraction in all the three lines was low compared to chromatin-bound TRF2 for the same amount of protein lysate suggesting nuclear TRF2 was largely bound to chromatin (S3F Fig). However, we noted ~15–20% increase in nucleoplasmic and total TRF2 in telomere-elongated MRC5 cells (S3F and S3G Fig).
Next, we asked whether non-telomeric TRF2 occupancy varied in cells with short versus long telomeres in MRC5 cells. We tested the 22 target promoters with significant TRF2 occupancy validated earlier in MRC5 cells (Fig 1C). In 19 of the 22 promoters TRF2 binding was significantly depleted in cells with long telomeres compared to untreated MRC5 cells with short telomeres (Fig 2E), similar to the observations in HT1080 cells. In case of OPN4, PSMC2 and PTPN11 altered TRF2 occupancy was not significant. Together, results obtained in HT1080 and MRC5 cells suggest that TRF2 occupancy at sites distal to telomeres was influenced by the length of telomeres.
We recently reported that p21 was transcriptionally repressed by TRF2[28]. Here we asked if altered TRF2 binding in cells with elongated telomeres affected p21 expression. Because of reduced occupancy of TRF2 at the p21 promoter in HT1080-LT cells (Fig 1H), we observed that p21 promoter activity as well as mRNA and protein levels were enhanced in HT1080-LT compared to HT1080 cells with shorter telomeres (Fig 3A). Transient over expression of hTERT and hTERC in HT1080 cells did not affect TRF2 occupancy at the p21 promoter or p21 expression (S4A–S4C Fig). Therefore, it is unlikely that change in p21 expression was due to any indirect effect of hTERT and/or hTERC expression used for inducing telomere elongation in HT1080 cells. Increase in p21 promoter activity, mRNA expression, and p21 protein levels was also found in MRC5-OF7 and OF14 cells with long telomeres relative to untreated MRC5 cells (Fig 3B) consistent with reduced TRF2 occupancy at the p21 promoter in MRC5 cells with long telomeres (Fig 2E), supporting similar observations in HT1080 cells.
Together these demonstrate that regulation of p21 by TRF2 is dependent on whether cells have long or short telomeres showing expression of a gene remote from telomeres (~36 Mb) can be affected by telomere length.
TRF2-dependent recruitment of the RE-1 silencing transcription factor (REST) and lysine-specific demethylase 1 (LDS1) at the p21 promoter was observed by us earlier[28]. Along with reduced TRF2 occupancy, REST and LSD1 binding at the p21 promoter was significantly depleted in HT1080-LT when compared to HT1080 cells (Fig 3C and S5A and S5B Fig). On the other hand, REST occupancy at the Synapsin promoter (S1A Fig) was not different in HT1080/HT1080-LT cells (S5A Fig) indicating that TRF2-independent REST binding was not affected by telomere length. The CTCF promoter with no reported REST occupancy was used as a negative control (S5A Fig).
Because loss of REST and LSD1 was expected to affect the epigenetic state we analyzed the presence of activating (H3K4me1 and H3K4me3) and silencing (H3K27me3) histone modifications at the p21 promoter. In HT1080-LT cells there was a significant increase in both activation marks H3K4me1 and H3K4me3; and reduction in suppressor mark H3K27me3 (Fig 3D). Together these argue for TRF2-dependent epigenetic changes at the p21 promoter that are sensitive to telomere length.
We checked whether telomere length altered the expression of genes other than p21. Expression of 13 out of 16 genes (excluding p21 described above), where promoter TRF2 occupancy was earlier observed to be depleted in HT1080-LT (Fig 1H), was significantly up or down regulated in HT1080-LT versus HT1080 cells (Fig 4A). This was consistent with the effect expected from TRF2 silencing in HT1080 cells (Fig 1D) and, therefore, likely from reduced promoter TRF2 occupancy in cells with long telomeres. In five promoters where TRF2 occupancy did not change in HT1080-LT/HT1080 cells expression of corresponding genes also remained unaltered (KCNH2, LINC01136, PTPN11, RYR2 and SMAD7). Altered PSMC2 expression was observed though TRF2 binding remained unchanged in HT1080-LT/HT1080 cells whereas CHRM2 and PDE3A expression did not change despite altered TRF2 binding in short/long telomere HT1080 cells (Fig 1H). In contrast to the other target genes, PDGFRβ and WRNIP1 were down-regulated on TRF2 silencing (Fig 1D) suggesting TRF2-mediated activation. This was consistent with loss of promoter TRF2 occupancy in HT1080-LT relative to HT1080 cells giving down-regulation of PDGFRβ and WRNIP1 (Fig 4A).
To test whether the differential expression of 14 genes (including p21) in HT1080-LT cells was sensitive to hTERT/hTERC we overexpressed hTERT and/or hTERC. Expression of none of the 14 genes was altered by hTERC whereas two genes ANXA2 (~25% up-regulation) and TBC1D2B (~20% down-regulation) were altered by hTERT (S6 Fig). In HT1080-LT cells ANXA2 expression was up-regulated by ~3-fold (Fig 4A) suggesting a combined effect of both increased hTERT and loss in TRF2 promoter occupancy. TBC1D2B expression, on the other hand, was up-regulated >3-fold in HT1080-LT cells (Fig 4A).
Next, we overexpressed TRF2 in HT1080-LT cells to see if the effect of reduced promoter TRF2 occupancy could be reverted. Fourteen genes with both altered TRF2 occupancy and gene expression in HT1080/HT1080-LT cells were tested. TRF2 overexpression resulted in suppression of 11 of the 14 genes in HT1080-LT cells (S7A Fig). PDGFR-beta and WRNIP1 were up-regulated–consistent with TRF2-mediated up-regulation noted for these genes. In case of THRA no significant change in expression was observed. Furthermore, we compared levels of gene expression in both HT1080/HT1080-LT cells following TRF2 overexpression. In 10 out of the 14 genes expression levels were similar (S7B Fig). Higher expression of ANXA2 in HT1080-LT cells was likely due to increased hTERT (as noted independently on hTERT overexpression, S6 Fig).
In case of MRC5 cells, out of 21 targets (excluding p21) 19 genes were differentially expressed in MRC5-OF cells with elongated telomeres relative to untreated MRC5 cells (Fig 4B). Similar to HT1080 cells, for all except two (OPN4 and PMSC2) of the 19 genes, TRF2 promoter occupancy was also significantly reduced in long telomeres (MRC5-OF) compared to untreated MRC5 cells (Fig 2E). CHRM2 expression did not change though TRF2 binding was altered in short/long MRC5 cells, as was noted in HT1080 cells. PTPN11 expression remained unchanged as expected from unaltered TRF2 promoter occupancy in MRC5-OF versus MRC5 cells, which was consistent across both the telomere elongation models.
Based on the histone modifications observed at the p21 promoter (Fig 3D), we tested promoter histone modifications for all the differentially expressed genes in HT1080 cells with long/short telomeres. For all the 11 genes where telomere elongation resulted in activation (Fig 4A) enrichment in activation histone modifications (either H3K4me1 or H3K4me3, or both) and/or depletion in the levels of the repressor modification H3K27me3 was found in HT1080-LT relative to HT1080 cells (Fig 5A–5C). On the other hand, in case of two genes down-regulated on telomere elongation—PDGFRβ and WRNIP1—activation histone marks H3K4me1 and H3K4me3 were either significantly depleted (WRNIP1) or the repressor mark H3K27me3 was enriched (PDGFRβ) within promoters in HT1080-LT compared to HT1080 cells. Taken together, in all the 13 genes differentially regulated on telomere elongation (Fig 4A), which also had reduced promoter TRF2 occupancy in HT1080-LT cells (Fig 1H), we found altered activation and/or repressor histone modifications within promoters to be consistent with TRF2-mediated activation or repression of the gene.
TRF2 silencing in HT1080 cells resulted in modifications in H3K4me1, H3K4me3 and H3K27me3 profiles on gene promoters (S8A–S8C Fig) that was largely consistent with promoter modifications observed in HT1080-LT relative to HT1080 cells (Fig 5). The histone modifications were also consistent with the differential gene expression on TRF2 silencing in HT1080 cells (Fig 1C). We further checked occupancy of REST at the 13 promoters that were sensitive to telomere length and TRF2 occupancy (in addition to p21). Eight out of the 13 promoters had REST occupancy (S9A Fig) and in case of 6 (of the 8) promoters REST occupancy was altered in HT1080-LT cells with elongated telomeres (S9B Fig).
Here we show that transcription of genes remote from telomeres depends on whether telomeres are short or elongated. This is mediated through non-telomeric TRF2 binding. Our results demonstrate that occupancy of TRF2 at telomere-distal gene promoters was relatively depleted when telomeres were long–consequently, TRF2-mediated transcription was affected. Notably, the epigenetic state of p21 and many other promoters was dependent on telomere length, in a fashion consistent with TRF2-mediated up or down regulation of the gene.
TRF2 binding at non-telomeric sites, particularly the extent of occupancy being dependent on state of telomeres (long/short) appears to be a key factor. As noted earlier, telomere-bound TRF2 was enhanced in cells with long telomeres[32]. However, total chromatin-bound TRF2 remained largely unaltered in cells with short versus long telomeres (in both HT1080 and MRC5 cells–S2E Fig and S3F Fig). Therefore, we postulated, enhanced telomeric TRF2 binding in cells with long telomeres may deplete TRF2 from non-telomeric sites. In contrast, in cells with short telomeres non-telomeric sites are likely to have more TRF2 occupancy relative to long telomeres. This was the case in both HT1080 and MRC5 cells (Fig 1H and Fig 2E). These argue for a model where TRF2 occupancy is partitioned between telomeric and non-telomeric sites. As a result, telomeric sequestration of TRF2 in cells with long telomeres restricts TRF2 binding at non-telomeric sites (Fig 6).
The number of TRF2 molecules/cell was reported to be ~50000 to 140000 depending on the cell type [38]. For HT1080 cells (mean telomere length of ~4.5 Kb[34,35]) ~69000 telomeric TRF2 binding sites are expected (S1 Text), which appears to be in the range of estimated TRF2 molecules/cell. Upon telomere elongation (to ~10 Kb) about 2-fold more TRF2 molecules are likely to be required for telomeric binding. However, TRF2 levels remained roughly similar in HT1080/HT1080-LT cells (S2D Fig). The affinity of TRF2 for telomeric/non-telomeric sites was also similar (S10 Fig). Despite this, we observed telomere-bound TRF2 increases in cells with longer telomeres consistent with an earlier report[32]. A possible explanation could be that TRF2 binds to telomeres as part of a larger complex with TRF1, RAP1, TIN2 and POT1[9,18,38–41]. In addition, RAP1 and TIN2 enhance association of TRF2 to telomeres[41,42]. This, along with the increased number of telomeric TRF2 binding sites in cells with elongated telomeres might help sequester more TRF2 to the telomeres.
Furthermore, TRF2 was primarily chromatin-bound in HT1080 and MRC5 cells (S2E Fig, S3F Fig), as also noted earlier[23]. Nucleoplasmic TRF2, though several folds lower than chromatin-bound, remained largely invariant in HT1080/HT1080LT cells (S2E Fig). In MRC5 cells, however, TRF2 in the nucleoplasm increased by ~15%, a small but reproducible observation (S3F Fig). It is not known if this is biologically significant but interaction of TRF2 with nucleoplasmic proteins such as lamin[29] and other nucleolar factors[43] has been reported. Further, nuclear lamin A was noted to be positively correlated with telomere length[44,45]. Therefore, it is possible that the higher nucleoplasmic TRF2 in telomere elongated MRC5 OF cells might be due to more lamin-bound TRF2.
Decrease in non-telomeric promoter TRF2 occupancy when telomeres are long induces, in most cases studied here, permissive chromatin (enriched H3K4me1/me3 and reduced H3K27me3 modifications; Fig 5A–5C, Fig 6). TRF2-mediated recruitment of the histone modification factors REST/LSD1 resulting in repression of p21 was noted earlier[28]. Based on this, loss of TRF2 binding in cells with elongated telomeres gave reduced REST/LSD1 occupancy at the p21 promoter (Fig 3C). As a result, active histone modifications at the p21 promoter increased (and repressor modifications decreased) in cells with long vis-à-vis short telomeres (Fig 3D). In addition to p21, 8 of the 13 other promoters (sensitive to TRF2 occupancy and telomere length) had REST occupancy. In 6 (of the 8) promoters binding of REST decreased in HT1080-LT cells with elongated telomeres (S9A and S9B Fig) consistent with the histone modifications (Fig 5A–5C). It is possible that at the other promoters studied here histone changes result through chromatin modifications factors engaged in TRF2-dependent or independent ways. Although further work will be required to fully decipher the underlying mechanisms behind telomere-dependent distal promoter modifications and gene expression, these results provide early mechanistic support for our observations.
Telomere-length dependent expression of several genes spread across the genome, and in two distinct cell types, observed here suggested transcriptome-wide changes may be linked to telomeres. To test we analyzed two independent microarray datasets[33,46] from short or elongated telomeres derived from isogenic cell lines. Isogenic background was necessary to limit confounding results from cell type-specific variations. In human pancreatic cancer PC-3 cells[46] (telomere elongation within tumors developed in mouse xenografts) analysis of microarray data across four replicates showed 1149 (out of 1461, ~78%) significantly differentially expressed genes (2-fold, p<0.05) were located beyond 10 Mb of telomeres (S11 Fig). Similarly, analysis of telomere-elongated versus control myoblast cells33 revealed >80% of the differentially regulated genes were distal to telomeres (2-fold, p<0.05; S11 Fig). These support our results obtained from a selected number of genes suggesting telomere length-dependent expression of genes are likely to be genome wide.
Looping of telomeres to interstitial sites, referred to as interstitial t-loops, mediated through TRF2 and lamin associations has been reported[29,47]. More recently, looping of the chromosome 5p telomere to the telomerase (hTERT) promoter 1.1 Mb from the telomeric end[48] was shown to result in presence of telomere-bound TRF2 at the hTERT promoter. Accordingly, TRF2 occupancy at the hTERT promoter was enriched in cells with long telomeres and reduced when telomeres were short and therefore less likely to have looping interactions. Both looping models (t-loop and hTERT-telomere) suggest enhanced non-telomeric TRF2 binding in case of long telomeres. In contrast, in the present study, we found loss in promoter TRF2 occupancy in cells with long telomeres. Conversely, for short telomeres, TRF2 occupancy at promoters was enriched. Since almost all the promoters were distal (tens of Mb mostly) to telomeres, this suggests TRF2 binding at non-telomeric sites may be distinct from the telomere looping mechanisms proposed earlier.
Telomere-dependent transcription of genes distal to telomeres shown here also appear distinct from the epigenetic phenomenon called telomere positioning effect (TPE) through which promoters close to the sub-telomeric regions remain silenced[49–51]. Extensively studied in the budding yeast TPE was reported to be primarily due to deacetylation of sub-telomeric nucleosomes by Rap1-mediated recruitment of SIR proteins[51]. DUX4, C1S, ISG15 and SORBS2 expression, located up to 10 Mb from telomeres, were also shown to depend on TPE but by a slightly different mechanism possibly involving telomeric looping to chromatin near these genes called TPE-OLD (over long distances)[33,52–54]. Accordingly, we noted TRF2 binding was enriched at the ISG15 and C1S promoters in cells with relatively long telomeres (HT1080-LT and MRC5-OF cells; S12A and S12B Fig). This, again, was in contrast to loss of TRF2 occupancy found at the distal promoters in cells with elongated telomeres–and, therefore, unlikely to be due to telomere looping.
Binding of TRF1 was reported at non-telomeric sites and in several instances in association with TRF2[25]. We checked for TRF1 binding within the 14 promoters where both change in TRF2 promoter occupancy and altered gene expression was observed in HT1080-LT cells with elongated telomeres. While TRF1 occupancy at the reported non-telomeric sites (HS3ST4 and CLIC6)[25] was retained, no significant TRF1 binding was observed within the 14 promoters in both HT1080 and HT1080-LT cells (S13 Fig).
Another shelterin factor RAP1 was also found to bind at non-telomeric sites[23,26]. Moreover, TRF2-RAP1 association has been reported[24,38,39,42,55]. Therefore, we asked whether the non-telomeric function of TRF2 reported here was RAP1 dependent. Expression of 14 genes that were sensitive to TRF2 was tested following RAP1 silencing. In all the 14 genes RAP1 silencing did not significantly affect TRF2-mediated expression (S14 Fig).
The consensus TRF2 binding site identified by the motif search algorithm MEME[56] in the promoters studied here showed a G-rich motif (S15 Fig). Though constructed from relatively few promoters, presence of the GGG trimer residues was consistent with the consensus TTAGGG motif within interstitial TRF2 binding sites reporter earlier[25]. Association with interstitial TTAGGG motifs was also reported for the shelterin protein RAP1 based on ChIP-seq[23]. For both–TRF2 and RAP1 –extra-telomeric binding, including RAP1-mediated genome wide transcription changes, was reported[23],[25]. However, whether extra-telomeric binding (and expression changes) was influenced by telomere length was not tested.
With relative increase in p21 levels in cells with elongated telomeres, we did not notice much difference in percentage of cells in different phases of cell cycle (S16 Fig). One explanation for this is that in cells with elongated telomeres enhanced levels of hTERT, which is known to induce cell proliferation[57], might counter the effect of p21 expression. Moreover, although ageing primary cells with shortened telomeres have increased p21 expression (and decreased proliferation/senescence) in case of cancer cells, p21 was observed to promote proliferation and oncogenicity in several studies[58–62]. Therefore, further work will be required to understand how telomere elongation/shortening impacts TRF2-mediated p21 expression and resultant proliferation/senescence.
Following telomere replication during S phase, TRF2 is recruited to the newly formed telomeres to prevent telomeric DNA damage–consistent with the presence of TRF2 in S phase[63–65]. Therefore, it is possible that during S phase redistribution of TRF2 binding takes place as telomere length changes in the cells that were used in the present studies. However, further experiments will be required to test this.
In summary, our results show evidence of the telomeric shelterin protein TRF2 regulating expression of genes distal to telomeres in a telomere length-dependent way. While gene regulation by telomeric factors was reported, whether long or short telomeres had any impact on gene transcription at distances remote from telomeres was not studied[23,28,33]. Although based on a selected number of genes, our findings describe involvement of telomeres, in a mechanistic way through telomere-binding proteins such as TRF2. In addition, we observed epigenetics and transcriptional changes across the genome that had not been reported previously. A more complete understanding of this new regulatory mechanism of telomere binding proteins, might lead to an improved understanding of the molecular processes of how telomeres impact cellular physiology, particularly in ageing and cancer.
HT1080 fibrosarcoma cell line was purchased from the NCCS, Pune. Immortalized MRC5 cells were received as a gift from NII, New Delhi. HT1080, MRC5 cells and corresponding telomere elongated cells were maintained in Modified Eagle’s medium (MEM) supplemented with 10% Fetal Bovine Serum (FBS). All cultures were grown in incubators maintained at 37°C with 5% CO2.
The Flow-FISH assay for telomere length detection was performed using DAKO Telomere PNA Kit/FITC codeK5327. Manufacturer’s’ guidelines were followed for assays.
ChIP assays were performed as per protocol provided by Upstate Biotechnology with modifications as suggested in Fast ChIP protocol. ChIP assays were performed using anti-TRF2 antibody (Novus Biologicals NB110-57130), anti-REST (Millipore), ani-LSD1 (CST), anti-H3K4me1, anti-H3K4me3, anti-H3K27me3 (Abcam). anti- TRF1 (TRF 78 Santa-Cruz),Anti-Rabbit IgG/Anti-mouse IgG was used for isotype control in all cell lines.
HT1080 cells/ MRC5 cells were transfected with TRF2 siRNA oligonucleotides (synthesized from Eurogenetics Pvt. Limited) /RAP1 pooled siRNA (Santacruz) using lipofectamine 2000 (Invitrogen) transfection reagent according to manufacturer’s instructions. Silencing was checked after 48 hr of transfection. Pooled SCR siRNA was used as control.
TRF2 WT (myc/DDK-tag), hTERT (Flag-tagged) and hTR cDNA cloned in mammalian expression vector pCMV6 was transfected into HT1080 cells that were 60% confluent using Lipofectamine 2000 transfection reagent (following the manufacturers’ protocol 2 μg of plasmid was used for transfection in a 35 mm well for each case. Expression was checked after 48 hr of transfection.
Plasmid (pGL4.73) containing a CMV promoter driving Renilla luciferase was co-transfected as transfection control for normalization. After 48h, cells were harvested and luciferase activities of cell lysate were recorded by using a dual-luciferase reporter assay system (Promega).
Total RNA was isolated using TRIzol Reagent (Invitrogen, Life Technologies) according to manufacturer’s instructions. Relative transcript expression level for genes was measured by quantitative real-time PCR using SYBR Green form Takara.
For dot blot analysis, Genomic/ ChIP DNA was denatured at 95°C and dot blotted on N+ hybond membrane (Amersham) in pre-wetted in 2X SSC buffer. Rapid-Hyb buffer (Amersham) was used for blocking and hybridization as per manufacturer’s protocol.
Chromatin fractionation assay was carried out as described earlier[66].The nuclear proteins are extracted by allowing cells to swell in hypotonic buffer and then disrupting the cells this is followed by removal of cytoplasmic fraction and using various combinations of low and high salt buffers nuclear proteins are released from nuclei.
For western blot analysis, protein lysates were prepared by suspending cell pellets in 1X cell culture lysis buffer (Promega). Protein was separated using 12% SDS-PAGE and transferred to polyvinylidene difluoride membranes (Immobilon FL, Millipore). Primary antibodies- anti-TRF2 antibody (Novus Biological), anti-p21antibody (Cell signaling technology) and anti-β-actin/anti-GAPDH antibody (Sigma), anti-H2A (abcam). Secondary antibodies, anti-mouse and anti-rabbit alkaline phosphatase conjugates were from Sigma. The blot was finally developed by using Thermo Scientific Pierce NBT/BCIP developing reagents.
The assays were performed using TELO TAGG kit from ROCHE with adherence to manufacturer’s protocol. TRAP assay was performed using TeloTTAGG PCR ELISA kit from ROCHE catalog no.-11854666910. In this assay test cell lysate was used as the source for telomerase is provided PCR conditions allowing telomerase activity on biotinylated TS template. This reaction is followed by overall amplification. The amplified product is quantified by ELISA using Anti-DIG POD antibody performed on Streptavidin coated plate provided within the kit.
Recombinant TRF2 was purified following expression in E. coli. Briefly, transformed cells were inoculated into 5 ml culture with Ampicillin (100 μg/ml) at 37°C overnight in a shaker incubator. 1 ml culture was inoculated with 500 ml fresh LB/Ampicillin and allowed to grow till OD of 0.6–0.8 units (at 600 nm wavelength). Following induction with 0.1 mM final concentration overnight at 18°C the culture was pelleted and sonicated in lysis buffer. 200 ul his-pure nickel NTA beads (Thermo Scientific) were added and incubated at 4°C on a rotatory shaker. Beads were washed consecutively with a 20 ml solution of 20–60 mM imidazole, and protein was eluted with 4 ml of 250 mM imidazole solution. Protein was concentrated along with buffer exchange to remove imidazole using Millipore 15 ml, 30 KDa concentrator columns. Purified protein was quantified by the BCA method (Thermo scientific BCA kit).
384-well streptavidin coated pre-blocked plates from Thermo Scientific (Pierce) were used for ELISA assay. Biotinylated oligonucleotides (IDT) were diluted to 5 picoM in 1X PBST buffer and loaded into each well, incubated at 37°C for 2 hours and washed 3 times with 1X PBST buffer. Purified TRF2 diluted in 1X PBST buffer was incubated with oligonucleotides for 2 hours at 4°C, washed 5 times with 1X PBST buffer, anti-TRF2 antibody (Novus NB110-57130) was added 1:1000 dilution (30 ul per well) and incubated for 1 hr at room temperature. Wells were washed five times with 1X PBST, 10 ul BCIP/NBT substrate was added to each well and absorbance was recorded at 610 nm using TECAN multimode reader. Two controls were used in ELISA assay to subtract background binding of antibody and protein. Protein negative control: except TRF2 protein all other reagents were added to determine the background binding of antibodies. Oligonucleotide negative control: except oligonucleotide, all other reagents along with increasing concentration of protein were added to determine background binding of the protein. The absorbance obtained from control wells were used for normalization and data analyzed using GraphPad Prism7.
Assay was performed using FITC BrdU Flow Kit from BD Pharmingen using manufacturer provided protocol.
|
10.1371/journal.ppat.1007313 | Promyelocytic leukemia (PML) nuclear bodies (NBs) induce latent/quiescent HSV-1 genomes chromatinization through a PML NB/Histone H3.3/H3.3 Chaperone Axis | Herpes simplex virus 1 (HSV-1) latency establishment is tightly controlled by promyelocytic leukemia (PML) nuclear bodies (NBs) (or ND10), although their exact contribution is still elusive. A hallmark of HSV-1 latency is the interaction between latent viral genomes and PML NBs, leading to the formation of viral DNA-containing PML NBs (vDCP NBs), and the complete silencing of HSV-1. Using a replication-defective HSV-1-infected human primary fibroblast model reproducing the formation of vDCP NBs, combined with an immuno-FISH approach developed to detect latent/quiescent HSV-1, we show that vDCP NBs contain both histone H3.3 and its chaperone complexes, i.e., DAXX/ATRX and HIRA complex (HIRA, UBN1, CABIN1, and ASF1a). HIRA also co-localizes with vDCP NBs present in trigeminal ganglia (TG) neurons from HSV-1-infected wild type mice. ChIP and Re-ChIP show that vDCP NBs-associated latent/quiescent viral genomes are chromatinized almost exclusively with H3.3 modified on its lysine (K) 9 by trimethylation, consistent with an interaction of the H3.3 chaperones with multiple viral loci and with the transcriptional silencing of HSV-1. Only simultaneous inactivation of both H3.3 chaperone complexes has a significant impact on the deposition of H3.3 on viral genomes, suggesting a compensation mechanism. In contrast, the sole depletion of PML significantly impacts the chromatinization of the latent/quiescent viral genomes with H3.3 without any overall replacement with H3.1. vDCP NBs-associated HSV-1 genomes are not definitively silenced since the destabilization of vDCP NBs by ICP0, which is essential for HSV-1 reactivation in vivo, allows the recovery of a transcriptional lytic program and the replication of viral genomes. Consequently, the present study demonstrates a specific chromatin regulation of vDCP NBs-associated latent/quiescent HSV-1 through an H3.3-dependent HSV-1 chromatinization involving the two H3.3 chaperones DAXX/ATRX and HIRA complexes. Additionally, the study reveals that PML NBs are major actors in latent/quiescent HSV-1 H3.3 chromatinization through a PML NB/histone H3.3/H3.3 chaperone axis.
| An understanding of the molecular mechanisms contributing to the persistence of a virus in its host is essential to be able to control viral reactivation and its associated diseases. Herpes simplex virus 1 (HSV-1) is a human pathogen that remains latent in the PNS and CNS of the infected host. The latency is unstable, and frequent reactivations of the virus are responsible for PNS and CNS pathologies. It is thus crucial to understand the physiological, immunological and molecular levels of interplay between latent HSV-1 and the host. Promyelocytic leukemia (PML) nuclear bodies (NBs) control viral infections by preventing the onset of lytic infection. In previous studies, we showed a major role of PML NBs in favoring the establishment of a latent state for HSV-1. A hallmark of HSV-1 latency establishment is the formation of PML NBs containing the viral genome, which we called “viral DNA-containing PML NBs” (vDCP NBs). The genome entrapped in the vDCP NBs is transcriptionally silenced. This naturally occurring latent/quiescent state could, however, be transcriptionally reactivated. Therefore, understanding the role of PML NBs in controlling the establishment of HSV-1 latency and its reactivation is essential to design new therapeutic approaches based on the prevention of viral reactivation.
| Herpes simplex virus 1 (HSV-1) is a human pathogen with neurotropic tropism and the causal agent of cold sores and more severe CNS pathologies such as encephalitis [1]. After the initial infection, HSV-1 remains latent in neuronal ganglia with the main site of latency being the trigeminal (or Gasserian) ganglion (TG). Two transcriptional programs are associated with HSV-1 infection, the lytic cycle and latency, which differ by the number and degree of viral gene transcription. The lytic cycle results from the sequential transcription of all viral genes (approximately 80) and leads to the production of viral progeny. The latency phase, occurring exclusively in neurons, is limited to the abundant expression of the so-called Latency Associated Transcripts (LATs), although physiologically a transitory expression of a limited number of lytic genes is not excluded, making latency a dynamic process[2–4].
Following lytic infection of epithelial cells at the periphery, the viral particle enters the axon termini of the innervating neurons by fusion of its envelope with the plasma membrane. The nucleocapsid is then carried into the neuron body by retrograde transport, most likely through the interaction of viral capsid components [5] with microtubule-associated proteins such as dynein and dynactin [6–10]. Once the nucleocapsid reaches the cell body, the virus phenotype changes from the one at the axon termini because most of the outer tegument proteins, including VP16, a viral transactivator that is essential for the onset of lytic infection, remain at the axonal tip [11–13]. Hence, when the viral DNA is injected into the neuron nucleus, it does not automatically benefit from the presence of VP16 to initiate transcription of lytic genes. Rather, the balance between lytic and latent transcriptional programs most likely depends on stochastic events and on undescribed neuron-associated factor(s) able to initiate the transcription of VP16 through the activation of neuro-specific sequences present in the VP16 promoter [14]. Without VP16 synthesis, transcription of the viral genes encoding ICP4 (the major transactivator protein) and ICP0 (a positive regulator of viral and cellular gene transcription) is hampered. Hence, ICP4 and ICP0 gene transcription is unlikely to reach the required level to produce these two proteins above a threshold that would favor onset of the lytic cycle. Therefore, in neurons, commitment of the infectious process towards the lytic cycle or latency depends on a race between opposing infection-prone viral components and cellular features with antiviral activities.
Promyelocytic leukemia (PML) nuclear bodies (NBs) (also called ND10) are proteinaceous entities involved in the control of viral infection as part of the cell and nucleus-associated intrinsic antiviral response but also through innate immunity associated with the interferon (IFN) response [15]. Our recent studies have shown that PML NBs tightly associate with incoming HSV-1 genomes in the nucleus of infected TG neurons in mouse models and in primary TG neuron cultures [16,17]. Hence, PML NBs reorganize in structures called viral DNA-containing PML NBs (vDCP NBs), which are formed at early times during the process of HSV-1 latency establishment and persist during latency per se in a large subset of latently infected neurons in a mouse model of infection [16]. The entrapment of incoming wild type HSV-1 genomes by PML NBs is not a unique feature of latency, because it has recently been shown to occur prior to the onset of lytic infection, as part of the intrinsic antiviral response. HSV-1 genomes trapped in the vDCP NBs are transcriptionally repressed for LATs production [16]. It is known that HSV-1 latency, at least in the mouse model and possibly in humans, is heterogeneous at the single neuron level for the expression of LATs [16,18–25]. Therefore, although at the entire TG level HSV-1 latency could be a dynamic process from a transcriptional perspective, at the single neuron level, a strict, transcriptionally silent, quiescence can be observed, and vDCP NB-containing neurons are major contributors of this latent/quiescent HSV-1 state. In humans, vDCP NB-like structures have also been observed in latently infected TG neurons [17], suggesting that vDCP NBs are probably molecular hallmarks of the HSV-1 latency process, including in the natural host.
Another essential feature of HSV-1 latency is the chromatinization of its 150-kb genome, which enters the nucleus of the infected cells as a naked/non-nucleosomal dsDNA [26–28]. Once the viral genome is injected into the nucleus of the infected neuron, it circularizes, associates with nucleosomes to become chromatinized, and remains as an episome that is unintegrated into the host cell genome [29]. Although latent viral genomes sustain chromatin regulation, essentially through post-translational modifications of associated histones [30–34] not much is known about the mechanisms that induce their chromatinization and which specific histone variants are associated with these latent genomes. In mammals, specific H3 histone variants that differ by a few amino acid residues can influence chromatin compaction and transcriptional activity of the genome. The histone variant H3.3, a specific variant of the histone H3 that is expressed throughout the cell cycle, is deposited in a replication-independent manner, in contrast to H3.1 ([35] and for review [36]). Interestingly, death domain associated protein 6 (DAXX) and α-thalassemia mental retardation X-linked protein (ATRX), initially identified as a transcriptional repressor and a chromatin remodeler, respectively, are constitutively present in PML NBs, and have now been identified as H3.3-specific histone chaperones [37–39]. The other histone H3.3 specific chaperone complex is called the HIRA complex, which is composed of Histone cell cycle regulator (HIRA), Ubinuclein 1 (UBN1), Calcineurin-binding protein 1 (CABIN1), and Anti-silencing function protein 1 homolog A (ASF1a) [35]. The HIRA complex does not normally accumulate in PML NBs except upon entry of the cell into senescence [40,41]. The histone variant H3.3 itself localizes in PML NBs in proliferating and senescent cells, linking PML NBs with the chromatin assembly pathway independently of replication [42–44]. Because vDCP NBs contain DAXX and ATRX [16,17,45], their involvement in the chromatinization of incoming HSV-1 genomes and/or long-term maintenance of chromatinized HSV-1 genomes is thus plausible.
Human primary fibroblasts or adult mouse primary TG neuron cultures infected through their cell body with a replication-defective HSV-1 virus, in1374, which is unable to synthesize functional ICP4 and ICP0 under specific temperature conditions, enable the establishment of a latent/quiescent state for HSV-1 [17,45–47]. The latent/quiescent state of HSV-1 in human primary fibroblasts has also been reproduced using engineered HSV-1 unable to express major immediate early genes [48,49]. We have shown that this latent/quiescent state is linked to the formation of vDCP NBs, mimicking, at least concerning this particular structural aspect, the latency observed in a subset of neurons in mouse models and in humans [16,17]. Here, using the in1374-based in cellula model of infection, we showed that vDCP NBs contained not only the DAXX and ATRX proteins but also all the components of the HIRA complex and H3.3 itself. HIRA was also found to co-localize with vDCP NBs in neurons of TG harvested from HSV-1 wild type infected mice. Both DAXX/ATRX and HIRA complex components were found to interact with multiple viral loci by chromatin immunoprecipitation (ChIP). Using the same approaches, we showed that latent/quiescent viral genomes were almost exclusively chromatinized with H3.3, itself modified on its lysine (K) 9 by trimethylation (H3.3K9me3). Most interestingly, we found that H3.3 chromatinization of the viral genomes was dependent on intact PML NBs, demonstrating that PML NBs contribute to an essential part of the chromatinization of the latent/quiescent HSV-1 genomes. Overall, this study shows that the chromatinization of latent HSV-1 involves a PML NB/histone H3.3/histone H3.3 chaperone axis that confers and probably maintains chromatin marks on viral genomes.
The formation of vDCP NBs is a molecular hallmark of HSV-1 latency, and vDCP NBs are present in infected neurons from the initial stages of latency establishment to latency per se in mouse models [16,17]. Using a previously established in vitro latency system [46] consisting of human primary fibroblast cultures infected with a replication-deficient virus (hereafter called in1374) unable to express functional VP16, ICP4 and ICP0, we and others were able to reproduce the formation of vDCP NBs [17,45]. We first verified that vDCP NBs induced in BJ and other human primary cells infected with in1374 at a non-permissive temperature of 38.5°C, contained, in addition to PML, the proteins constitutively found in the PML NBs, i.e., Sp100, DAXX, ATRX, SUMO-1 and SUMO-2/3 (S1Ai to S1vi Fig, and S1 Table). The DAXX/ATRX complex is one of the two chaperones of the histone variant H3.3 involved in the replication-independent chromatinization of specific, mostly heterochromatic, genome loci [39]. Interestingly, HSV-1 enters the nucleus of the infected cell as a naked/non-nucleosomal dsDNA and remains during latency as a circular chromatinized episome unintegrated in the host genome [29,50]. It is thus tempting to speculate that the presence of DAXX/ATRX in the vDCP NBs could be linked to a process of initiation and/or maintenance of chromatinization of the latent/quiescent viral genome. The other H3.3 chaperone is known as the HIRA complex and was initially described as specific for the replication-independent chromatinization of euchromatin regions [35,51]. Remarkably, proteins of the HIRA complex are able to bind in a sequence-independent manner to a naked/non-nucleosomal DNA [52], suggesting that the HIRA complex could also participate in the recognition and chromatinization of the incoming naked HSV-1 genome. We thus investigated the localization of all members of the HIRA complex and found that they co-localized with the latent/quiescent HSV-1 genomes at 2 days post-infection (dpi) in BJ and other human primary cells (Fig 1Ai to 1iv, S1 Table). To confirm that the co-localization of members of the HIRA complex with the latent/quiescent HSV-1 could be reproduced in neuronal cells, adult mouse TG neuron cultures were infected with in1374 for 2 days before performing immuno-FISH. Mouse Hira, which was the only protein of the HIRA complex detectable in mouse cells, showed a clear co-localization with a subset of viral genomes (Fig 1B). To analyze whether this co-localization was also reproducible in vivo, immuno-FISH was performed on TG samples from HSV-1-infected mice. Hira was found to co-localize with HSV-1 genomes with the “multiple acute”/vDCP NB pattern (see [17,53,54]) in TG neurons from infected mice at 6 dpi (Fig 1C) but not with the “single”/vDCP NB pattern (see [16,53,54]) at 28 dpi (Fig 1D), suggesting a dynamic association of this protein with the vDCP NBs.
To analyze this dynamic association, co-localization between incoming HSV-1 genomes and proteins of the PML NBs or of the HIRA complex was quantified at early times from 30 min pi to 6 hpi using a synchronized infection procedure (Fig 1E and S2 Table). Except for the proteins of the HIRA complex, the percentages of co-localization increased with time. Interestingly, at 30 min pi, the percentage of co-localization of HSV-1 genomes with HIRA was significantly higher than with PML (41±7% vs 23±5%, p value = 0.03, Student’s t-test, S2 Table). Although DAXX and ATRX also showed, on average, a greater percentage of co-localization with HSV-1 genomes (36±7% and 34±5% at 30 min, respectively) compared with PML, the data were not significant (S2 Table). Moreover, a recent study showed the interaction of at least PML, SUMO-2, and Sp100 with incoming HSV-1 genomes as soon as 1 hpi, which supports our data [55]. The absence of co-localization of mouse Hira with viral genomes with the “single”/vDCP NB pattern in mouse TG neurons at 28 dpi suggested that longer infection times could lead to loss of proteins of the HIRA complex from the vDCP NBs. Infection of BJ cells were reiterated as above, but this time quantifications were performed from 24 hpi to 7 dpi. Strikingly, whereas all the proteins permanently present in the PML NBs remained co-localized with a maximum of 100% of the latent/quiescent HSV-1 genome from 2 dpi until 7 dpi, proteins of the HIRA complex peaked at 2 dpi, and then their co-localization decreased at longer times pi, confirming the temporary association of the HIRA complex with the vDCP NBs (Fig 1F, and S3 Table).
To definitively show that proteins of the HIRA complex were present in vDCP NBs, immuno-FISH were performed on BJ cells infected for 2 days with in1374 to detect a member of the HIRA complex, HSV-1 genomes, and PML. Strikingly, while proteins of the HIRA complex showed predominant nucleoplasmic staining in non-infected cells (Fig 2i, 2iii, 2v and 2vii), in infected cells all the proteins clearly and systematically accumulated in PML NBs (Fig 2ii, 2iv, 2vi and 2viii). The accumulation of HIRA in PML NBs following infection by HSV-1 has recently been suggested to be part of an interferon-induced antiviral mechanism [56]. Consequently, HIRA, UBN1, CABIN1 and ASF1a co-localized with the latent/quiescent HSV-1 genomes in vDCP NBs (arrows in Fig 2ii, 2iv, 2vi and 2viii). Altogether, these data show that both DAXX/ATRX and HIRA complexes are present within vDCP NBs in neuronal and non-neuronal cells, suggesting a role for these two complexes in latent/quiescent HSV-1 chromatinization.
The co-localization of proteins of the DAXX/ATRX and HIRA complexes with the incoming HSV-1 genomes and their presence in the vDCP NBs suggested an interaction of these proteins with the viral genome, as shown recently for HIRA on a small subset of viral loci [56]. Since DAXX, HIRA, and UBN1 antibodies were not efficient in the ChIP experiments, we constructed cell lines stably expressing myc-DAXX, HIRA-HA, or HA-UBN1 by transduction of BJ cells with lentiviral- vectors (S2 Fig). Cells were infected with in1374 at 38.5°C and harvested 24 hpi to perform ChIP-qPCR on multiple loci spread over the entire HSV-1 genome, representing promoter or core regions (CDS) of genes of all kinetics (IE/α, E/β, L/γ) (Fig 3A). Cellular glyceraldehyde 3-phosphate dehydrogenase (GAPDH) locus was used as a positive control for enrichment. Significant enrichments compared to controls were detected for all proteins on several viral loci independently of their promoter or CDS status, and with no obvious discrepancy regarding the gene kinetic, confirming the potential interaction of these proteins all along the latent/quiescent HSV-1 genomes. Our immuno-FISH data anticipated a gradual interaction of the four proteins with the incoming viral genomes at early times post infection (see Fig 1E). To verify if this could be measured, ChIP-qPCR were performed at 30 min pi, 2 hpi and 6 hpi, using the same experimental conditions as for the immuno-FISH at early times pi (with synchronization of the infection, see Materials and Methods). The data showed a tendency for a weak interaction with the viral genomes at 30 min pi then an increase at 2 hpi and 6 hpi, although with a lot of variability, probably highlighting the dynamic of the biological events occurring during the initial stages of the infection process (S3B Fig). ATRX showed the more regular increase in its interaction with viral genomes from 30 min to 24 hpi. Overall, the ChIP data correlate with the immuno-FISH, and suggest a dynamic process for the interaction between HSV-1 genomes and proteins of the DAXX/ATRX and HIRA complexes, initiating early after the viral DNA enters the nucleus, and remaining at later times when vDCP NBs are structured.
The co-localization of the two histone H3.3 chaperone complexes with viral genomes suggested the chromatinization of HSV-1 latent/quiescent genomes with the histone variant H3.3. Histones H3.1 and H3.3 differ by only 5 amino acids, and, in our hands, no suitable antibody is available that can distinguish both histones by IF or IF-FISH. We thus constructed lentivirus-transduced BJ cell lines expressing a tagged version of either histone (e-H3.1 and e-H3.3) (see Materials and Methods, and [43], S4A and S4B Fig). We confirmed that ectopic expression of e-H3.3 led to its accumulation in PML NBs unlike e-H3.1 (S4C Fig) [42,43]. In1374 infection of BJ e-H3.1/3-expressing cells led to the co-localization of viral genomes almost exclusively with e-H3.3 (Fig 4Ai, 4ii and 4B). Importantly, e-H3.3 co-localized with HSV-1 genomes together with PML in vDCP NBs (Fig 4C). The lack of co-localization of viral genomes with e-H3.1 was in agreement with the absence of any of the H3.1 CAF-1 chaperone subunits (p150, p60, p48) in the vDCP NBs (Fig 4D, S1 Table). To confirm that e-H3.3, unlike e-H3.1, interacted with HSV-1 genomes, ChIP-qPCR were conducted on the same loci as those analyzed above. As expected, e-H3.3, but not e-H3.1, was highly enriched on the viral genome independently of the examined locus (Fig 4E). Several cellular loci were analyzed as controls for specific enrichments with H3.3 (Enhancer 1 (Enh.1) on chromosome 9, [57]), or H3.1 (leucine-zipper-like transcriptional regulator 1 (LZTR1) on chromosome 22, GEO accession number GSM1135044). Similar data were obtained for all other canonical histones (S5 Fig), confirming that H3.3 association with latent/quiescent HSV-1 genomes is in a nucleosomal context. To confirm that the discrepancy between the binding of e-H3.3 and e-H3.1 to viral genomes was not due to the ectopic expression of histones, we performed similar experiments using antibodies against native proteins. One specific antibody for H3.3, and suitable for ChIP experiments has previously been described [58]. We performed ChIP using antibodies against native H3.1/2 or H3.3 in normal BJ cells infected for 24 h by in1374. The results were similar to those obtained in infected BJ e-H3.3 using the anti-HA antibody (S6 Fig). These data confirmed that no bias was introduced in the ChIP experiments due to the use of tagged histones, and that latent/quiescent HSV-1 genomes are chromatinized with H3.3. The gradual interaction of the four proteins of the H3.3 chaperone complexes with the incoming viral genomes anticipated similar changes in the interaction of H3.3. ChIP-qPCR were performed at 30 min pi, 2 hpi and 6 hpi, using the same experimental conditions as above. The data showed an overall weak or lack of, H3.3 association with the viral genomes at 30 min pi, followed by an increased interaction at 2 hpi and 6 hpi. These data show that the H3.3 chromatinization of the incoming HSV-1 genomes is progressive and follows a kinetic that matches that observed with the proteins of the H3.3 chaperone complexes. The data also fit with recently published data showing the interaction of incoming viral genomes with canonical histones by 2 hpi [55].
Both constitutive (H3K9me2, H3K9me3) and facultative (H3K27me3) heterochromatin marks have been found on various loci on latent HSV-1 genomes in vivo [31,33,34]. To analyze the association of these marks with vDCP NBs-associated latent/quiescent HSV-1 genomes, ChIP were performed targeting H3K9me3, H3K27me3 and one euchromatic mark H3K4me2 as a control (Fig 5). HSV-1 genomes were exclusively associated with H3K9me3 (Fig 5A), matching previous results obtained using quiescent viruses [59,60]. In contrast H3K27me3 (Fig 5B) or H3K4me2 (Fig 5C) marks were not detected. Cellular genes previously described for their association with either marks were analyzed for the specificity of the antibodies used (Zinc-finger protein 554 (ZNF554)/H3K9me3 [61]); myelin transcription factor 1 (MYT1)/H3K27me3 ([62]; Actin/H3K4me2). To confirm that the K9me3 modification is present on H3.3 associated with the HSV-1 genomes, Re-ChIP was performed targeting first H3K9me3 then e-H3.3 in infected BJ and BJ e-H3.3 (Fig 5D). An overall enrichment for H3.3 from samples initially ChIPed with the H3K9me3 antibody was detected only in BJ e-H3.3 and not BJ cells, with 17 viral loci over 31 (55%) showing significant enrichment. The cellular locus, family with sequence similarity 19 member A2 (FAM19A2) specifically enriched with H3.3K9me3 (GEO accession numbers: GSM1358809 (H3.3), and GSM1289412 (H3K9me3)) was used as positive control. These data show that (i) the Re-ChIP experiment is specific of e-H3.3 and (ii) H3.3K9me3 is indeed associated with the vDCP NB-associated HSV-1 latent/quiescent genomes.
To analyze the requirement of the histone H3.3 chaperones for the formation of the vDCP NBs and HSV-1 chromatinization, DAXX, ATRX, HIRA or UBN1 were depleted by shRNAs in normal BJ cells or cells constitutively expressing e-H3.3 prior to infection with in1374 and completion of the experiments. The two tested shRNAs for each protein significantly diminished mRNA and protein quantities in BJ cells (S7A and S7B Fig). None of the shRNA impacted the detection of PML NBs, suggesting that PML NBs were potentially functional when the proteins were individually inactivated (S8 Fig). We first measured the impact of the depletion of each protein on the co-localization of HSV-1 genomes with PML. Both shRNAs for each protein gave similar results, i.e., a significant decrease in the co-localization between HSV-1 genomes and PML and thus a decrease in the formation and/or stability of the vDCP NBs (Fig 6A and 6B, S4 Table). These data show that the inactivation of any of the H3.3 chaperone complex affects to a certain extent the fate of vDCP NBs suggesting a connection between the activity of each H3.3 chaperone complex and the formation and/or maintenance of the vDCP NBs.
We then analyzed the potential impact of the loss of vDCP NB stability on the H3.3-dependent HSV-1 chromatinization. We performed H3.3 ChIP in in1374-infected BJ e-H3.3 cells that had been previously depleted for HIRA, UBN1, DAXX or ATRX using one of the previously validated shRNAs (S9A and S9B Fig). The data showed that overall the inactivation of UBN1, DAXX or ATRX, had a weak impact on the association of H3.3 with the viral loci (1 to 3 loci significantly affected over 31, 3.2 to 9.6%) (Fig 6C). The depletion of HIRA had a relatively greater effect (6/31, 19.4%). To analyze if simultaneous inactivation of both complexes would significantly impact on HSV-1 chromatinization with H3.3, one protein of each complex was inactivated at the same time before performing HSV-1 infection (Fig 7A). Individual inactivation of HIRA and ATRX is known to lead to the functional inactivation of the HIRA and DAXX/ATRX complexes, respectively [35,52,63,64]. We noticed that the inactivation of HIRA by a siRNA was not as efficient as the shRNA on preventing the association of H3.3 with viral genomes (Fig 7B). This is likely due to differences in the efficiency of the siRNA compared to the shRNA (compare WBs of Figs S7B and 7A), and to the transitory effect of the siRNA compared to the stable effect of the shRNA at the time of the infection (see Materials and Methods). Nonetheless, a significant decrease of the association of H3.3 with a large number of viral loci (20/31, 64.5%) was measured by the simultaneous inactivation of HIRA and ATRX compared to their individual inactivation (Fig 7B). These results indicate that the DAXX/ATRX complex may compensate for the loss of the HIRA complex on the chromatinization of latent/quiescent HSV-1 genomes with H3.3, and conversely.
The above experiments were conducted in a context where the cells, although deficient for the activity of one H3.3 chaperone complex at a time, still contained intact PML NBs accumulating e-H3.3 (S7 and S10 Fig). Therefore, we hypothesized that the accumulation of H3.3 within the PML NBs could be one of the key events acting upstream of the H3.3 chaperone complex activity for the induction of chromatinization of the latent/quiescent HSV-1 by H3.3. We analyzed the HSV-1 chromatinization in cells lacking PML NBs. In a previous study conducted in HSV-1 latently infected PML KO mice, we showed that the absence of PML significantly impacted the number of latently infected TG neurons showing the “single”/vDCP NB HSV-1 pattern and favored the detection of neurons containing the “multiple-latency” pattern prone to LAT expression [16,53]. We analyzed the very few neurons showing a “single”/vDCP NB-like pattern in the latently infected PML KO mice for the co-localization of DAXX and ATRX with the viral genomes. We could not detect any of the two proteins co-localizing with the latent HSV-1 genomes (Fig 8Ai to 8vi). Although informative, these in vivo studies did not allow the analysis of the real impact of the absence of PML on the co-localization of the other PML NB-associated proteins with latent HSV-1 genomes, because the neurons showing the “single”/vDCP NB-like pattern were too few to quantify the effect. We thus depleted PML in normal BJ cells using a PML shRNA-expressing lentiviral transduction approach. We verified the efficiency of the shRNAs against PML in normal BJ cells by IF, RT-qPCR and WB (S11A–S11C Fig). PML-depleted BJ cells were superinfected with in1374, and immuno-FISH was performed at 2 dpi to analyze the co-localization of HSV-1 genomes with DAXX, ATRX, HIRA, and UBN1 (Fig 8B). Notably, both PML shRNAs gave similar results. The quantification of the data showed that, similarly to the in vivo situation, the depletion of PML significantly decreased the co-localization of DAXX and ATRX with latent/quiescent HSV-1 genomes, leaving HIRA and UBN1 unaffected for their co-localization (Fig 8C, and S5 Table). Thus, we analyzed whether the failure of DAXX/ATRX to co-localize with the latent/quiescent HSV-1 genomes in the absence of PML NBs, could impact the chromatinization of HSV-1 with H3.3.
We first generated BJ e-H3.3 cells depleted for PML by shRNA-expressing lentiviral transduction similarly to the BJ cells (S11D and S11E Fig). BJ e-H3.3 control or PML-depleted cells were superinfected with in1374 to perform immuno-FISH and analyze the co-localization of HSV-1 genomes with H3.3 (Fig 9A). Quantification of the data showed a significant decrease in the co-localization of latent/quiescent HSV-1 genomes with H3.3 compared with controls (Fig 9B), suggesting an impact of the absence of PML NBs on the latent/quiescent HSV-1 association with H3.3. To complement these results at a more quantitative level, we performed ChIP on e-H3.3. The data showed a major impact of the absence of PML NBs on the H3.3 association with viral genomes, with a significant depletion of H3.3 on multiple loci (21/31, 68%) (Fig 9C). This could not be due to an indirect effect of PML depletion on H3.3 stability because e-H3.3 protein levels were similar in control cells and cells depleted for PML (Fig 9D). Both PML shRNAs gave similar results. To confirm that the absence of PML had an impact on the H3.3 association with latent/quiescent viral genomes, we performed ChIP on in1374-infected control MEF pml+/+ or MEF pml-/- cells previously engineered by lentiviral transduction to express e-H3.3 (Fig 9E). The data confirmed the impaired association of e-H3.3 with latent/quiescent HSV-1 genomes in the absence of PML, with 26/31 (84%) viral loci significantly impacted (Fig 9F). Cellular loci acid Sensing Ion Channel Subunit 2 (Asic2), and Heme Oxygenase 1 (Hmox1) were used respectively as positive and negative controls for deposition of H3.3 in the absence of Pml as described in [44]. To definitively attribute the lack of deposition of H3.3 on viral loci to the absence of PML, MEF pml-/-;e-H3.3 cells were engineered to allow re-expression, under doxycycline induction, of the isoform I of human PML (PML.I) (Fig9G), which was shown to participate to the HSV-1 antiviral restriction mechanism [65]. The formation of PML NBs after induction of PML.I was visualized by IF (Fig 9H). ChIPs were then performed on in1374-infected MEF pml-/-;e-H3.3;myc-PML.I cells previously treated or not with doxycycline (Fig 9I). The data showed that the re-expression of PML.I allowed the re-loading of H3.3 on all the analyzed loci of the latent/quiescent viral genomes with significant results obtained for 21 loci over 31 (68%), demonstrating the essential role of PML/PML NBs in the association of H3.3 with incoming viral genomes. Finally, we wanted to analyze whether the deficit of the H3.3 association with the viral genome in the absence of PML could be compensated by an increase of H3.1 on viral loci. The data from BJ e-H3.1 cells depleted for PML or MEF pml-/-;e-H3.1 cells, and infected with in1374 showed that H3.1 did not replace H3.3 on the viral loci (S12A and S12B Fig). Altogether, these data demonstrate the essential role of PML NBs, probably through the DAXX/ATRX complex activity, in the exclusive H3.3 chromatinization of incoming viral genomes forced to adopt a vDCP NB-associated latent/quiescent pattern due to a deficit in the onset of lytic cycle.
vDCP NB-associated latent genomes have been shown to be transcriptionally silent for the LAT expression in vivo [16], and for the expression of a reporter gene in vitro in mouse TG neuron cultures [17], and in human primary fibroblasts [45]. Moreover, it is known that the viral protein ICP0 induces the destabilization of PML NBs [66] and is essential for HSV-1 reactivation in vivo [67], and for the transcriptional de-repression of a silenced viral genome in vitro [45,59,60]. However, it is not known if the transcriptional recovery is correlated to the destabilization of the vDCP NBs. We analyzed if latent HSV-1 genomes trapped in vDCP NBs were definitively silenced or could resume a transcriptional program leading to replication of viral genomes provided that vDCP NBs were destabilized. ICP0 or its non-functional RING finger mutant (ICP0ΔRF) were expressed from BJ-eTetR/cICP0 or BJ-eTetR/cICP0ΔRF cells harboring vDCP NBs for 4 days (Fig 10). HSV-1 in1374 infected BJ-eTetR cells were used as controls. Expression of ICP0 or ICP0ΔRF was induced for 24 h, 48 h or 72h at the permissive temperature for in1374 replication (32°C) (S13 Fig). Transcription of the reporter (LacZ) gene was measured by RT-qPCR to analyze the transcriptional recovery of the vDCP NB-associated latent/quiescent viral genomes (Fig 10A). The addition of doxycycline in infected BJ-eTetR or BJ-eTetR/c ICP0ΔRF cells did not lead to any significant transcription of the LacZ gene. Only infected BJ-eTetR/cICP0 showed the recovery of LacZ mRNA transcription from 24 h post addition of doxycycline. To analyze if the virus could sustain replication, as suggested by the observation of the BJ-eTetR/cICP0 cell monolayer (S14 Fig), following vDCP NB destabilization, immuno-FISH were performed at 24 h and 48 h post-addition of doxycycline. BJ-eTetR/cICP0 cells (Fig 10B) but not BJ-eTetR/cICP0ΔRF (S15 Fig) showed a clear disappearance of the vDCP NBs. Concomitantly, only BJ-eTetR/cICP0 cells showed the formation of replication compartments (RCs) indicating that the virus is in the process of lytic phase following vDCP NBs destruction by ICP0. To confirm that the lytic transcriptional program was indeed occurring, viral transcripts of all kinetics were analyzed (Fig 10C). Twenty four and 48 h post ICP0 induction, lytic genes were expressed with a clear switch towards the γ genes (UL44/gC and US6/gD) at 48 h confirming the onset of the lytic transcriptional program. Expression of ICP0ΔRF did not enable the re-expression of viral genes. These data show that vDCP NBs-associated latent/quiescent HSV-1 genomes can resume transcription and a lytic program provided that the vDCP NBs are destabilized, suggesting that these genomes are not definitively silenced, and could participate to the reactivation process of HSV-1.
The HSV-1 genome enters the nucleus of infected neurons, which support HSV-1 latency as a naked/non-nucleosomal DNA. Many studies have described the acquisition of chromatin marks on the viral genome concomitantly to the establishment, and during the whole process, of latency. Paradoxically, although it is undisputable that these chromatin marks will predominantly be associated with latency and reactivation, few data are available for the initiation of the chromatinization of the incoming viral genome. Here, we demonstrate the essential contribution of PML NBs in the process of chromatinization of incoming HSV-1 genomes meant to remain in a latent/quiescent state. We showed that PML NBs are essential for the association of the histone variant H3.3 with the latent/quiescent HSV-1.
Two members of the HIRA complex, HIRA and ASF1a, were previously shown to be involved in H3.3-dependent chromatinization of HSV-1 genomes at early times after infection in non-neuronal and non-primary cells favoring the onset of the lytic cycle [68,69]. Moreover a recent study highlighted the interaction of HIRA with quiescent HSV-1 and plasmid DNA in primary human fibroblasts [56]. Our in vivo data in TG neurons and in vitro data in infected human primary fibroblasts or adult mice TG neuron cultures, show that all the proteins of the HIRA complex accumulate within specific nucleoprotein structures called the viral DNA-containing PML NBs or vDCP NBs. vDCP NBs contain transcriptionally silent HSV-1 genome that we previously demonstrated in vivo to be associated with the establishment of latency from the early steps of neuron infection [17]. Additionally, our data show that: (i) the mouse Hira protein, in vivo, and all the components of the HIRA complex, in cultured cells, temporarily accumulate in vDCP NBs, and (ii) significantly greater amount of incoming HSV-1 genomes co-localize with HIRA compared with PML at very early times pi (30 min). These data suggest that the HIRA complex could also be involved to some extent in the establishment of HSV-1 latency by the initial recognition of the incoming naked/non-nucleosomal viral DNA and the chromatinization of non-replicative HSV-1 genomes intended to become latent. In this respect, a recent study suggested an anti-viral activity associated with HIRA against HSV-1 and murine cytomegalovirus lytic cycles [56]. To that extent, although they are both functionally essential for the activity of the HIRA complex [35,51,64,70], our data show that the depletion of HIRA has a greater effect compared to the UBN1 depletion, on the H3.3 association with the viral genomes. This could be simply explained by a better efficiency of the HIRA, compared to the UBN1, shRNAs. Alternatively, HIRA was shown to be recruited to UV-induced DNA damage independently of UBN1 (see figure S2D in [71]), and to participate to the loading of newly synthesized H3.3 on chromatin [72]. Therefore, the depletion of HIRA could indirectly and/or directly impact on two initial events occurring concomitantly to the entry of the viral genomes in the nucleus; first a signaling pathway associated to the detection of DNA breaks present in incoming viral DNA as suggested in [55,73]; and second the chromatinization process per se. If these two events are linked it could explain the differences observed between the HIRA and UBN1 depletion on the loading of H3.3 on the viral genomes. Experiments are in progress to investigate this.
Interestingly, proteins of the HIRA complex have been previously shown to be able to directly bind to naked DNA in a sequence-independent manner, in contrast to DAXX and ATRX [52]. Nevertheless, our ChIP data highlight the interaction of viral genomes with DAXX and ATRX, but we cannot assert that the two proteins directly interact with naked DNA. The gamma-interferon-inducible protein 16 (IFI16), a member of the PYHIN protein family, has been described as a nuclear sensor of incoming herpesviruses genomes, and suggested to promote the addition of specific chromatin marks that contribute to viral genome silencing [74–81]. A proteomic study determining the functional interactome of human PYHIN proteins revealed the possible interaction between ATRX and IFI16 [82]. Thus, it will be interesting to determine in future studies if IFI16 and H3.3 chaperone complexes physically and functionally cooperate in the process of chromatinization of the latent/quiescent HSV-1 genome.
One of the main finding of our study is the demonstration of the essential contribution of PML NBs in the H3.3-dependent chromatinization of the latent/quiescent HSV-1 genomes. A close link between PML NBs and H3.3 in chromatin dynamics has been demonstrated during oncogene-induced senescence (OIS). In OIS, expression of the oncogene H-RasV12 induces DAXX-dependent relocalization of neo-synthesized H3.3 in the PML NBs before a drastic reorganization of the chromatin to form senescence-associated heterochromatin foci [42,43]. Hence, the contribution of the PML NBs in the deposition of H3.3 on specific cellular chromatin loci has also been reported [43,44]. The present study shows that the absence of Pml in HSV-1wt latently infected Pml KO mice, or the depletion of PML by shRNA in BJ cells infected with in1374, significantly affects the co-localization of DAXX and ATRX, but not HIRA and UBN1, with latent/quiescent HSV-1 genomes, confirming previous studies for DAXX and ATRX [45]. Taken together with the impaired association of H3.3 with the viral genomes in the absence of PML NBs, these data suggest that a significant part of the latent/quiescent HSV-1 genome chromatinization by H3.3 could occur through the activity of the DAXX/ATRX complex in association with the PML NBs.
Given the particular structure formed by the latent/quiescent HSV-1 genome with the PML NBs, our study raises the question of the possible acquisition of a chromatin structure within the vDCP NBs. The individual inactivation of DAXX, ATRX, HIRA, or UBN1 significantly impacts the co-localization of the latent/quiescent HSV-1 genomes with PML, and hence the formation of vDCP NBs. However, it only mildly affects the association of H3.3 with viral genomes, suggesting an absence of correlation between the formation of vDCP NBs and H3.3 chromatinization. However, our data show that the depletion of DAXX, ATRX, HIRA, or UBN1 does not modify the accumulation of e-H3.3 at PML NBs, leaving intact the upstream requirement of H3.3 accumulation in PML NBs for H3.3-dependent viral chromatin assembly. We have recently shown that vDCP NBs are dynamic structures that can fuse during the course of a latent infection [17]. It is thus possible that incoming viral genomes can be dynamically associated with vDCP NBs to be chromatinized, and in the absence of any of the H3.3 chaperone complex subunit, this dynamic can be perturbed, resulting in some viral genomes that do not show a co-localization with PML. Given that depletion of none of the four proteins affects the structure of the PML NBs, and considering the essential role of PML NBs in the H3.3 chromatinization of the viral genomes, this possibility cannot be ruled out. The depletion of H3.3, which almost exclusively participates in latent/quiescent HSV-1 genome chromatinization compared to H3.1/2, does not prevent the formation of vDCP NBs (S16 Fig), and is rather in favor of a chromatinization of the viral genome in the vDCP NBs. It is unlikely that canonical H3.1/2 could replace H3.3 for the chromatinization of the incoming HSV-1 genomes prior to the formation of the vDCP NBs. Indeed, our multiple immuno-FISH and ChIP assays failed to detect H3.1/2 and/or H3.1/2 chaperones that associate or co-localize with viral genomes. Nonetheless, we cannot rule out a possible replacement of H3.3 with another H3 variant for the chromatinization of viral genomes before their entrapment by the PML NBs to form vDCP NBs.
Our data show that the vDCP NBs-associated HSV-1 genomes are chromatinized with H3K9me3, and the Re-ChIP assays confirm an association with H3.3K9me3, but not H3K27me3. In vivo, it has been shown that both H3 modifications could be found on latent HSV-1 genomes [31,33,34]. One simple explanation could reside in the heterogeneity of latent genomes distribution within the nuclei of the infected neurons in the in vivo mouse and/or rabbit models of latency [16,17,54], however this would need to be formally demonstrated. Though, vDCP NBs-associated HSV-1 genomes remain compatible with the transcription of lytic genes provided that the vDCP NBs are destabilized by ICP0, a viral protein known to be required for full in vivo reactivation [67], and to erase chromatin marks associated with latent/quiescent viral genomes in vitro [59]. Therefore, vDCP NBs are not a dead end for the virus life cycle, and HSV-1 latently infected neurons containing vDCP NBs are likely to contribute to the process of reactivation.
Altogether, our study demonstrates the essential role of a PML NB/H3.3/H3.3 chaperone axis in the process of chromatinization of viral genomes adopting a vDCP NB pattern, which represents an essential structural and functional aspect of HSV-1 latency establishment. Given the involvement of H3.3 in the chromatinization of other latent herpesviruses belonging to different sub-families than HSV-1, such as EBV [83] and HCMV [58], as well as adenovirus type 5 [84], this pathway of chromatinization is likely to play a major role in the biology of the whole Herpesviridae family, and possibly of other DNA viruses such as adenoviruses, papillomaviruses, hepatitis B virus, and retroviruses.
All procedures involving experimental animals conformed to the ethical standards of the Association for Research in Vision and Ophthalmology (ARVO) statement for the use of animals in research and were approved by the local Ethics Committee of the Institute for Integrative Biology of the Cell (I2BC) and the Ethics Committee for Animal Experimentation (CEEA) 59 (Paris I) under number 2012–0047 and in accordance with European Community Council Directive 2010/63/EU. For animal experiments performed in the USA: animals were housed in American Association for Laboratory Animal Care-approved housing with unlimited access to food and water. All procedures involving animals were approved by the Children’s Hospital Animal Care and Use Committee and were in compliance with the Guide for the Care and Use of Laboratory Animals (protocol number: IAUC2013-0162 of 2/28/2107).
The HSV-1 SC16 strain was used for mouse infections and has been characterized previously [85]. The HSV-1 mutant in1374 is derived from the 17 syn + strain and expresses a temperature-sensitive variant of the major viral transcriptional activator ICP4 [86] and is derived from in1312, a virus derived from the VP16 insertion mutant in1814 [87], which also carries a deletion/frameshift mutation in the ICP0 open reading frame [88] and contains an HCMV-lacZ reporter cassette inserted into the UL43 gene of in1312 [89]. This virus has been used and described previously [17,45]. All HSV-1 strains were grown in baby hamster kidney cell (BHK-21, ATCC, CCL-10) and titrated in human bone osteosarcoma epithelial cells (U2OS, ATCC, HTB-96). In1374 was grown and titrated at 32°C in the presence of 3 mM hexamethylene bisacetamide [90].
PML wild-type, and knockout mice were obtained from the NCI Mouse Repository (NIH, http://mouse.ncifcrf.gov; strain, 129/Sv-Pmltm1Ppp) [91]. Genotypes were confirmed by PCR, according to the NCI Mouse Repository guidelines with primers described in [16].
Mice were inoculated and TG processed as described previously [16]. Briefly, for the lip model: 6-week-old inbred female BALB/c mice (Janvier Labs, France) were inoculated with 106 PFU of SC16 virus into the upper-left lip. Mice were sacrificed at 6 or 28 dpi. Frozen sections of mouse TG were prepared as described previously [16,92]. For the eye model: inoculation was performed as described previously [93]. Briefly, prior to inoculation, mice were anesthetized by intra-peritoneal injection of sodium pentobarbital (50 mg/kg of body weight). A 10-μL drop of inoculum containing 105 PFU of 17syn+ was placed onto each scarified corneal surface. This procedure results in ~80% mouse survival and 100% infected TG.
Primary mouse TG neuron cultures were established from OF1 male mice (Janvier lab), following a previously described procedure [17]. Briefly, 6–8-week-old mice were sacrificed before TG removal. TG were incubated at 37°C for 20 min in papain (25 mg) (Worthington) reconstituted with 5 mL Neurobasal A medium (Invitrogen) and for 20 min in Hank’s balanced salt solution (HBSS) containing dispase (4.67 mg/mL) and collagenase (4 mg/mL) (Sigma) on a rotator, and mechanically dissociated. The cell suspension was layered twice on a five-step OptiPrep (Sigma) gradient, followed by centrifugation for 20 min at 800 g. The lower ends of the centrifuged gradient were transferred to a new tube and washed twice with Neurobasal A medium supplemented with 2% B27 supplement (Invitrogen) and 1% penicillin–streptomycin (PS). Cells were counted and plated on poly-D-lysine (Sigma)- and laminin (Sigma)-coated, eight-well chamber slides (Millipore) at a density of 8,000 cells per well. Neuronal cultures were maintained in complete neuronal medium consisting of Neurobasal A medium supplemented with 2% B27 supplement, 1% PS, L-glutamine (500 μM), nerve growth factor (NGF; 50 ng/mL, Invitrogen), glial cell-derived neurotrophic factor (GDNF; 50 ng/mL, PeproTech), and the mitotic inhibitors fluorodeoxyuridine (40 μM, Sigma) and aphidicolin (16.6 μg/mL, Sigma) for the first 3 days. The medium was then replaced with fresh medium without fluorodeoxyuridine and aphidicolin.
Primary human foreskin (BJ, ATCC, CRL-2522), lung (IMR-90, Sigma, 85020204), fetal foreskin (HFFF-2, European Collection of Authenticated Cell Cultures, ECACC 86031405, kind gift from Roger Everett, CVR-University of Glasgow) fibroblast cells, primary human hepatocyte (HepaRG, HPR101, kind gift from Olivier Hantz & Isabelle Chemin, CRCL, Lyon, France) cells, human embryonic kidney (HEK 293T, ATCC CRL-3216, kind gift from M. Stucki, University Hospital Zürich) cells, U2OS, mouse embryonic fibroblast (MEF) pml+/+, MEF pml-/- cells (kind gift from Valérie Lallemand, Hopital St Louis, Paris), and BHK-21 cells were grown in Dulbecco’s Modified Eagle’s Medium (DMEM) supplemented with 10% fetal bovine serum (Sigma, F7524), L-glutamine (1% v/v), 10 U/mL penicillin, and 100 mg/mL streptomycin. BJ cell division is stopped by contact inhibition. Therefore, to limit their division, cells were seeded at confluence before being infected at a multiplicity of infection (m.o.i.) of 3, and then maintained in 2% serum throughout the experiment. Infections of BJ cells for short times (from 30 min to 6 h) were performed by synchronizing the infection process with a pre-step of virus attachment to the cells at 4°C for one hour. The infection medium was then removed, and the temperature was shifted to 37°C to allow a maximum of viruses to simultaneously penetrate into the cells.
Frozen sections of mouse TG were generated as previously described [92]. Mice were anesthetized at 6 or 28 d.p.i., and before tissue dissection, mice were perfused intracardially with a solution of 4% formaldehyde, 20% sucrose in 1X PBS. Individual TG were prepared as previously described [92], and 10-μm frontal sections were collected in three parallel series and stored at -80°C.
HSV-1 DNA FISH probes consisting of cosmids 14, 28 and 56 [94] comprising a total of ~90 kb of the HSV-1 genome were labeled by nick-translation (Invitrogen) with dCTP-Cy3 (GE Healthcare) and stored in 100% formamide (Sigma). The DNA-FISH and immuno-DNA FISH procedures have been described previously [16,92]. Briefly, infected cells or frozen sections were thawed, rehydrated in 1x PBS and permeabilized in 0.5% Triton X-100. Heat-based unmasking was performed in 100 mM citrate buffer, and sections were post-fixed using a standard methanol/acetic acid procedure and dried for 10 min at RT. DNA denaturation of the section and probe was performed for 5 min at 80°C, and hybridization was carried out overnight at 37°C. Sections were washed 3 x 10 min in 2 x SSC and for 3 x 10 min in 0.2 x SSC at 37°C, and nuclei were stained with Hoechst 33258 (Invitrogen). All sections were mounted under coverslips using Vectashield mounting medium (Vector Laboratories) and stored at 4°C until observation.
For immuno-DNA FISH, cells or frozen sections were treated as described for DNA-FISH up to the antigen-unmasking step. Tissues were then incubated for 24 h with the primary antibody. After three washes, secondary antibody was applied for 1 h. Following immunostaining, the cells were post-fixed in 1% PFA, and DNA FISH was carried out from the methanol/acetic acid step onward.
The same procedures were used for infected neuronal cultures except that the cells were fixed in 2% PFA before permeabilization.
Cells were collected in lysis buffer (10 mM Tris-EDTA, pH 8.0) containing a protease inhibitor cocktail (Complete EDTA-free; Roche) and briefly sonicated. Protein extracts were homogenized using QiaShredders (Qiagen). The protein concentration was estimated by the Bradford method. Extracted proteins were analyzed by Western blotting using appropriate antibodies (see below).
Observations and most image collections were performed using an inverted Cell Observer microscope (Zeiss) with a Plan-Apochromat ×100 N.A. 1.4 objective and a CoolSnap HQ2 camera from Molecular Dynamics (Ropper Scientific) or a Zeiss LSM 800 confocal microscope. Raw images were processed using ImageJ software (NIH).
BJ or MEF cell lines expressing H3.1-SNAP-HAx3 (e-H3.1), H3.3-SNAP-HAx3 (e-H3.3), or Myc-hDAXX were established by retroviral transduction [95]. Briefly, pBABE plasmids encoding H3.1-SNAP-HAx3 or H3.3-SNAP-HAx3 (gift from Dr L. Jansen), pLNCX2 encoding Myc-hDAXX [43], were co-transfected with pCL-ampho (for subsequent transduction of BJ cells, kind gift from M. Stucki, University Hospital Zürich) or pCL-eco (for subsequent transduction of MEF cells, kind gift from M. Stucki, University Hospital Zürich) plasmids [96] by the calcium phosphate method into HEK 293T cells to package retroviral particles [97]. BJ cells stably expressing HIRA-HA and HA-UBN1 or transiently expressing the shRNAs were established by lentiviral transduction. Briefly, pLenti encoding HIRA-HA or HA-UBN1, pLKOneo.CMV.EGFPnlsTetR, pLKO.DCMV.TetO.cICP0, pLKO.DCMV.TetO.cICP0ΔRF (gift from Dr. R. D. Everett, [98]), pLVX-TetOne-Myc-PML.I (issued from pLNGY-PML.I, gift from Dr. R. D. Everett [65]), pLKO empty, pLKO shPML_01, 02, shDAXX_01, 02, shATRX_01, 02, shHIRA_01, 02, shUBN1_01, 02, were co-transfected with psPAX.2 (Addgene #12260) and pMD2.G (Addgene #12259) plasmids by the calcium phosphate method into HEK 293T cells to package lentiviral particles. After 48 h, supernatant containing replication-incompetent retroviruses or lentiviruses was filtered and applied for 24 h on the target BJ or MEF cells in a medium containing polybrene 8 μg/mL (Sigma) [95]. Stable transfectants were selected with Blasticidin S (5 μg/mL, Invivogen), puromycin (1 μg/mL, Invivogen), or neomycin (G418, 1 mg/mL, Millipore) for 3 days, and a polyclonal population of cells was used for all experiments. Target sequences of the shRNA-expressing plasmids are provided in Table 1.
Cells were fixed with methanol-free formaldehyde (#28908, Thermo Fisher Scientific) 1% for 5 min at RT, and then glycine 125 mM was added to arrest fixation for 5 min. After two washes with ice-cold PBS, the cells were scraped and resuspended in “Lysis Buffer” (10% glycerol, 50mM HEPES pH7,5; 140mM NaCl; 0,8% NP40;0,25% Triton; 1mM EDTA, Protease Inhibitor Cocktail 1X (PIC) (Complete EDTA-free; Roche) and incubated for 10 min at 4°C under shaking. The cells were subsequently washed in “Wash buffer”(200mM NaCl; 20mM Tris pH8; 0,5mM EGTA; 1mM EDTA, PIC 1X) for 10 min at 4°C under shaking then were resuspended and centrifuged twice during 5 min 1700g at 4°C in “Shearing Buffer” (10mM Tris pH7,6; 1mM EDTA; 0,1%SDS; PIC 1X). Finally, nuclei were resuspended in 1mL of “Shearing Buffer” and were sonicated with a S220 Focused-ultrasonicator (Covaris) (Power 140W; Duty Off 10%; Burst Cycle 200). Eighty-five μL of the sonication product were kept for the input, 50 μL for analyzes of the sonication efficiency, and 850 μL diluted twice in IP buffer 2X (300mM NaCl, 10mM Tris pH8; 1mM EDTA; 0,1% SDS; 2% Triton) for ChIP. Two micrograms of Ab were added and incubated overnight at 4°C. Fifty microliters of agarose beads coupled to protein A (Millipore 16–157) or G (Millipore 16–201) were added for 2 h at 4°C under constant shaking. Beads were then successively washed for 5 min at 4°C under constant shaking once in “low salt” (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 20 mM Tris HCl pH 8.0, 150 mM NaCl) buffer, once in “high salt” (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 20 mM Tris HCl pH 8.0, 500 mM NaCl) buffer, once in “LiCl” (0.25 mM LiCL, 1% NP40, 1% NaDOC, 1 mM EDTA, 10 mM Tris HCl pH 8.0) buffer, and twice in TE (10 mM Tris pH 8.0, 1 mM EDTA) buffer. Chromatin-antibody complexes are then eluted at 65°C for 30 min under constant shaking with 200 μL of elution buffer (1% SDS, 0.1 M NaHCO3). Input and IP products were de-crosslinked overnight at 65°C with 20 mg/mL of proteinase K (Sigma) and 10 mg/mL of RNAse A (Sigma). DNA was then purified by phenol-chloroform/ethanol precipitation, resuspended in water, and kept at -20°C until use for qPCR.
Quantitative PCR was performed using Quantifast SYBR Green mix (Qiagen) and the MX3005P apparatus (Agilent/Stratagene). Primers were used at a final concentration of 1 μM. Their sequences and target genes are provided in Tables 2 and 3.
Cells were processed similarly to ChIP until addition of the antibody. Two μg of the first antibody were pre-incubating with agarose beads coupled to protein A (Millipore 16–157) in PBS-0.5% BSA overnight at 4°C under shaking. The beads were washed twice with IP Buffer 1X (150mM NaCl, 5mM Tris pH8; 0.5mM EDTA; 0.1% SDS; 1% Triton), then the chromatin was incubated with the antibody/beads overnight at 4°C. Beads were then successively washed for 5 min at 4°C under constant shaking once in “low salt” buffer, once in “high salt” buffer, once in “LiCl” buffer and twice in TE buffer. Chromatin-antibody complexes were eluted at 37°C for 30 min with 100 μL of Re-ChIP elution buffer (1% SDS, 0.1M NaHCO3, 10mM DTT). Fifty μL were kept for analysis of the first capture efficiency, then the other 50 μL were diluted 20 times with IP buffer 1X and incubated overnight at 4°C with the second antibody pre-incubated with agarose beads coupled to protein A. The beads were washed twice with IP Buffer 1X, then successively washed for 5 min at 4°C under constant shaking, once in “low salt” buffer, once in “high salt” buffer, once in “LiCl” buffer and twice in TE buffer. Chromatin-antibody complexes were then eluted at 65°C for 30 min under constant shaking with 200 μL of elution buffer (1% SDS, 0.1M NaHCO3). Input and IP products were de-crosslinked overnight at 65°C a with 20 mg/mL of proteinase K (Sigma) and 10 mg/mL of RNAse A (Sigma). DNA was then purified by phenol-chloroform/ethanol precipitation, resuspended in water, and kept at -20°C until use for qPCR.
Transfections of BJ cells with siRNAs was performed using Lipofectamine RNAiMAX and following the supplier’s procedure (Thermo Fisher Scientific). The following siRNAs were used at a final concentration of 40 nM for 48 h: siRNA_negative control (EUROGENTEC, FR-CL000-005), siHIRA 5’–GGAUAACACUGUCGUCAUC (Dharmacon: J-013610-07) [52]; siH3F3A: 5′-CUACAAAAGCCGCUCGCAA [100]; siH3F3B: 5′-GCUAAGAGAGUCACCAUCA [100].
The antibodies used for immunofluorescence, ChIP and WB are provided in Tables 4–6.
BJ cells were transduced first with pLKOneo.CMV.EGFPnlsTetR to produce BJ-eTetR cell lines stably and constitutively expressing the EGFPnlsTetR protein (selection G418 1 mg/mL). BJ-eTetR cells were then transduced with pLKO.DCMV.TetO.cICP0 or pLKO.DCMV.TetO.cICP0ΔRF to produce BJ-eTetR/cICP0 or BJ-eTetR/cICP0ΔRF expressing ICP0 or its RING finger mutant FXE, respectively (selection puromycin 1 μg/mL). The expression of ICP0 or ICP0ΔRF was induced by the addition of doxycycline (100ng/μL) in the medium. BJ-eTetR, BJ-eTetR/cICP0 or BJ-eTetR/cICP0ΔRF were infected with HSV-1 in1374 for 4 days at 38.5°C to stabilize the formation of vDCP NBs. Then doxycycline was added or not in the medium to induce the expression of ICP0 or ICP0ΔRF. Cells were incubated at 32°C the permissive temperature for in1374 (see section virus). Twenty four hours, 48h or 72h after addition of doxycycline, the cells were fixed to proceed to IF or IF-FISH analyses or treated with FastLane cell SYBR Green RT-PCR (Qiagen 204243) to analyze the LacZ and viral transcripts by RT-qPCR.
|
10.1371/journal.pcbi.1006976 | Fast and robust deconvolution of tumor infiltrating lymphocyte from expression profiles using least trimmed squares | Gene-expression deconvolution is used to quantify different types of cells in a mixed population. It provides a highly promising solution to rapidly characterize the tumor-infiltrating immune landscape and identify cold cancers. However, a major challenge is that gene-expression data are frequently contaminated by many outliers that decrease the estimation accuracy. Thus, it is imperative to develop a robust deconvolution method that automatically decontaminates data by reliably detecting and removing outliers. We developed a new machine learning tool, Fast And Robust DEconvolution of Expression Profiles (FARDEEP), to enumerate immune cell subsets from whole tumor tissue samples. To reduce noise in the tumor gene expression datasets, FARDEEP utilizes an adaptive least trimmed square to automatically detect and remove outliers before estimating the cell compositions. We show that FARDEEP is less susceptible to outliers and returns a better estimation of coefficients than the existing methods with both numerical simulations and real datasets. FARDEEP provides an estimate related to the absolute quantity of each immune cell subset in addition to relative percentages. Hence, FARDEEP represents a novel robust algorithm to complement the existing toolkit for the characterization of tissue-infiltrating immune cell landscape. The source code for FARDEEP is implemented in R and available for download at https://github.com/YuningHao/FARDEEP.git.
| Rapidly emerging evidence suggests that the tumor immune microenvironment not only predisposes cancer patients to diverse treatment outcomes but also represents a promising source of biomarkers for better patient stratification. Different from the immunohistochemistry-based scoring practice, which focuses on a few selected marker proteins, immune deconvolution pipelines inform a previously untapped method to comprehensively reveal the tumor-infiltrating immune landscape. Recognizing the numerous strengths of existing immune deconvolution algorithms, here we show data outliers, which are inevitable in whole tissue sequencing data sets, substantially skew estimation results. Moreover, an estimate related to the absolute amount of each immune subset offers valuable insight into the nature of the host response in addition to percentage information alone. Thus, we engineered a new immune deconvolution pipeline, coined as Fast and Robust Deconvolution of Expression Profiles (FARDEEP), to automatically detect and remove outliers prior feeding data into the deconvolution algorithm and to provide estimates related to the absolute quantity of each immune subset. Utilizing both synthetic and real data sets, we found that FARDEEP returns superior coefficients and offers a robust tool to reveal the immune landscape of human cancers.
| Immune checkpoint blockade has revolutionized the rational design of neoadjuvant cancer therapies. Compelling evidence suggests that a favorable tumor immune microenvironment underpins better clinical responses to radiotherapy, chemotherapy, and immunotherapy [1–3]. Immunohistochemistry (IHC)-based immunoscores, which quantify the number of CD8+ cytotoxic T lymphocytes and CD45RO+ memory T cells, show better prognostic potential than conventional pathological methods in colon cancer patients [4, 5]. Hence, harnessing the composition of intra-tumoral immune cell infiltration is a highly promising approach to stratify tumors [6–11]. The current IHC immunoscoring approach has two limitations. First, the interpretation of immune cell subsets varies among pathologists and institutions, thus lacking a consistent standard for the scoring practice. Second, only a limited number of biomarkers can be assessed simultaneously, which prevents a comprehensive annotation of the immune contexture in the tumor microenvironment (TME). Hence, robust methods for genome data-informed cell type quantitation are in urgent need.
Immunogenomics presents an unprecedented opportunity to resolve the intra-tumoral immune landscape. Cell type deconvolution using leukocyte signature gene expression profiling is a highly promising approach to estimate the global immune cell composition from whole tumor gene expression data [12–17]. However, a significant technical obstacle is that the efficacy and accuracy of gene expression deconvolution are limited by the large number of outliers, which are frequently observed in tumor gene expression datasets [18]. The first step towards enhancing the overall gene deconvolution algorithms is to improve methods for outliers identification and processing. Those outliers include genes with abnormal expression value which may be caused by measurement error, environmental effect, expression from non-immune cells, or natural fluctuations in certain type of immune cells. Notably, the current immune deconvolution gene signature matrix relies on the profiling of differentially expressed genes among different immune subsets. Frequent contamination of transcripts reading from cancer cells may significantly bias the algorithms. In this study, we report a novel FAst and Robust DEconvolution of Expression Profiles (FARDEEP) method that significantly improves the estimation of coefficients.
Let yi be the observed expression value for the ith gene; xi, a p-dimensional vector, be the expected expression of the ith gene for the p different cell types; and X = [x1, …, xn]′ be the signature matrix. The gene-expression deconvolution problem can be formulated as follows,
y i = x i ′ β + ε i , (0.1)
where β ∈ R p is an unknown parameter corresponding to the compositions of p cell types, and εi is a noise term with a mean of 0. Several methods were proposed to solve this deconvolution problem. To enforce the non-negativity of β in (0.1), several algorithms, such as the Non-Negative Least Square (NNLS), Non-negative Maximum Likelihood (NNML) frameworks and the perturbation model (PERT) were developed. They all rely on the signature matrix (X) derived from Microarray experiments [14, 19–24]. To extend this work to RNA-seq data, Finotello et al. [14] proposed a constraint linear model with a signature matrix derived from RNA-seq data. Additionally, the gene expression of each cell may vary depending on its microenvironment and other factors, which will lead to a biased estimation. To address this issue, Microarray Microdissection with Analysis of Differences (MMAD) incorporates the concept of the effective RNA fraction and estimates coefficients using a maximum likelihood approach [25]. To further adapt deconvolution to high-dimensional settings, Altboum et al. [26] proposed a penalized regression framework, Digital Cell Quantifier (DCQ), to encourage sparsity for the estimated β using the elastic net [27]. Cell-type identification by estimating relative subsets of RNA transcripts (CIBERSORT) uses ν-support vector regression (ν-SVR) to enhance the robustness of gene expression deconvolution. CIBERSORT performs a regression by finding a hyperplane that fits as many data points as possible within a tube whose vertical length is a constant ε [12]. The ε-tube provides a region in which estimation errors are ignored. This model does not include an intercept to capture contributions of other contents. Additionally, to increase the computational efficiency, CIBERSORT applies Z-normalization to the data before fitting the regression, which may introduce estimation bias. Based on the CIBERSORT framework, several extensions have been proposed to overcome limitations such as platform inconsistency between signature and mixture matrices and low estimation accuracy for γδ T cell [15–17]. However, the quantitative information of cell proportions of these two approaches is built on CIBERSORT whose performance may be challenged by frequent outliers in whole tumor tissue transcriptomes. To reduce the dependence on the signature matrix, xCell utilizes the concept of single-sample gene set enrichment analysis (ssGSEA) to calculate an immune cell score which could predict the enrichment of immune cells [13]. Despite its robustness, xCell relies much on the ranking of gene expression value which leads to suboptimal solution for the estimation accuracy. Overall, a robust method that determines both the distribution and absolute volume of tumor-infiltrating lymphocytes (TILs) will further improve the current gene deconvolution pipeline.
To handle the heavily contaminated gene expression data and provide absolute cell abundance estimation, we developed a robust method based on the Least Trimmed Square (LTS) framework [28, 29]. LTS finds h observations with smallest residuals, and the estimator β ^ is the least squares fit over these h observations. LTS is an NP-hard problem, and Rousseeuw and Driessen [30] proposed a stochastic FAST-LTS algorithm. Nevertheless, it may give a suboptimal fitting result and get much slower when the sample size and dimension of variables become larger and higher since its accuracy relies on the initial random h-subsets and the number of initial subsets. When n is the sample size and p is the number of coefficients, h is suggested to be the smallest integer that is not less than (n + p + 1)/2 to remove as many outliers as possible while keeping an unbias result. Using the information of only half of the data reduces the power of the estimator because the amount of outliers in the real case cannot be presumed and can be small. Xu et al. [31] proposed an adaptive least trimmed square which is not limited to the randomness of initial subset but only applied the binary dataset. In this study, we extend the adaptive least trimmed square to introduce a model-free method, which could find the number of outliers automatically based on LTS. FARDEEP provides a flexible framework which is suitable for both Microarray and RNA-seq data using LM22 and Immunostate signature matrices respectively. As evidence of high fidelity and robustness, FARDEEP exhibits superior performance in simulated and real-world datasets.
The usual linear deconvolution model can be expressed as below,
y = X β + ε ,
where y ∈ R n is the observed expression data for n immune subset signature genes, X ∈ R n × p denotes a mean gene expression signature matrix for p different cell types, β ∈ R p contains each unknown cell type abundance, and ε ∈ R n is a vector of random errors with zero mean and variance of σ2I. To incorporate outliers, we propose the following model
y = X β + τ + ε , (0.2)
where parameter τ = (τ1, …, τn)′ is a sparse vector in R n with τi ≠ 0 indicating the ith gene is an outlier.
Under the formulation of (0.2), let β ^ ols = ( X ⊤ X ) − 1 X ⊤ y be the Ordinary Least Square (OLS) estimate and H = X(X⊤X)−1X⊤ be the projection matrix. The residuals r = (r1, …, rn) using OLS could be formulated as
r = y − X β ^ ols = ( I − H ) τ + ( I − H ) ϵ . (0.3)
with mean of (I − H)τ and variance of σ2(I − H).
From (0.3), the residuals, ri with the corresponding τi ≠ 0, would deviate from zero, which suggests that the set of outliers can be identified through thresholding as follows
E = { i : | r i | > k × r med } , (0.4)
where E is the set of detected outliers, k is a tuning parameter controlling the sensitivity of the model, and rmed is the median of { | r | i } i = 1 n. We denote the number of elements in set E as |E| and let N be the number of true outliers in the data. First, we can use least squares and formula (0.4) to obtain a rough estimate of E denoted as E ^. Let the cardinality of E ^ be N ¯. Since the model at this moment is inaccurate with contamination of outliers, N ¯ is an overestimation of N which can be used to get an underestimate via N _ = α 1 N ¯ with α1 ∈ (0, 1). With N _, we can then update the least square fitting after removing the N _ samples with the largest absolute value of residuals and obtain an improved estimate of E and the corresponding N ¯. We can improve the model by repeating the procedure, but we need to increase the underestimate of outliers, N _, by a factor of α2 with α2 > 1 for each iteration to force the convergence between N ¯ and N _. In sum, we initialize our algorithm by setting
β ^ ( 0 ) = ( X ⊤ X ) − 1 X ⊤ y , r ( 0 ) = y − X β ^ ( 0 ) ,
which is the OLS solution. For the jth iteration, where j ≥ 1, we update N ¯ ( j ) by
N ¯ ( j ) = { | { i : | r i ( j − 1 ) | > r med ( j − 1 ) } | , j = 1 , min ( | { i : | r i ( j − 1 ) | > k · r med ( j − 1 ) } | , N ¯ ( j − 1 ) ) , j ≥ 2 . (0.5)
where the min(⋅, ⋅) operator guarantees that N ¯ ( j ), an overestimation of N, is non-increasing. Similarly, we update N _ ( j ) through
N _ ( j ) = { ⌈ α 1 N ¯ ( j ) ⌉ , j = 1 , min { ⌈ α 2 N _ ( j − 1 ) ⌉ , N ¯ ( j ) } , j ≥ 2 , (0.6)
where ⌈x⌉ means the ceiling of x ∈ R, α1 ∈ (0, 1) is used to obtain a lower bound for N in the first step, α2 > 1 guarantees the monotonicity of N _ ( j ), and the min(⋅, ⋅) operator guarantees N _ ( j ) is smaller than N ¯ ( j ). Then we update β ^ and r after removing N _ ( j ) outliers by
β ^ ( j ) = ( X ( j ) ⊤ X ( j ) ) − 1 X ( j ) ⊤ y ( j ) , r ( j ) = y − X β ^ ( j ) .
We repeat this procedure until N _ and N ¯ converge.
Hence, we hereby report a novel approach, coined as adaptive Least Trimmed Square (aLTS), to automatically detect and remove contaminating outliers. Our aLTS is an extension of the iterative LTS algorithm proposed by Xu et al. [31] which is designed for binary output such as the comparison between two images or videos.
Because the abundance of cell types are always non-negative, we replaced the OLS regression in the aLTS procedure with non-negative least square regression (NNLS). By applying the modified aLTS to the deconvolution model (0.2) and solving the following problem,
β ^ = argmin β‖ y − X β‖2 2 , subjecttoβ ≥ 0
using Lawson-Hanson algorithm [19], we developed a robust tool, FARDEEP, for cellular deconvolution summarized in Algorithm 1.
One unique advantage of FARDEEP is that it is fast and guarantees to converge within finite steps, which is summarized in the following theorem.
Algorithm 1 FAst and Robust DEconvolution of Expression Profiles
Input: k > 0, 0 < α1 < 1, α2 > 1, y, X
Initialization: solving the following NNLS problem
β ^ ( 0 ) = argmin β‖ y − X β ‖2 2 , subject to β ≥ 0 ; r ( 0 ) = y − X β ^ ( 0 ) .
1: compute N ¯ ( 1 ) and N _ ( 1 ) using (0.5) and (0.6);
2: solving the NNLS problem after removing N _ ( 1 ) genes with largest residuals, and update β ^ ( 1 ), r(1).
3: repeat
4: compute N ¯ ( j ) and N _ ( j ) using (0.5) and (0.6) for j ≥ 2;
5: solving the NNLS problem after removing N _ ( j ) genes with largest residuals, and update β ^ ( j ), r(j);
6: until N ¯ = N _.
Output: Coefficients β ^, Number of outliers N ^, Index of outliers
Theorem 1 Algorithm 1 (FARDEEP) stops in no more than j* steps, where
j
*
=
⌊
−
log
α
1
log
α
2
⌋
+
2
.
Here ⌊⋅⌋ is the largest integer that is less than or equal to x.
Proof. It follows from the fact that the sequence { N ¯ ( j ) } is non-increasing, and { N _ ( j ) } is a geometrically increasing sequence that is bounded by the smallest component of { N ¯ ( j ) }. Specifically, assume that j* steps have been taken in FARDEEP, then j has approached j* − 1, and N _ ( j ) ≥ α 2 N _ ( j − 1 ) for 0 ≤ j ≤ j* − 1, so
N ¯ ( 0 ) ≥ N ¯ ( j * − 2 ) ≥ N _ ( j * − 2 ) ≥ α 2 j * − 2 N _ 0 ≥ α 2 j * − 2 α 1 N ¯ 0 .
which leads to the result.
The β ^ from FARDEEP, denoted as TIL subset score, is the direct estimate of the linear model without any normalization and hence reflects the absolute abundance of TILs. In addition, we can derive the relative TILs abundance from the TIL subset scores through
β ˜ j = β j ^ ∑ k = 1 p β j ^ , (0.7)
where β ^ j is the jth TIL subset score. In practice, the TIL subset score provides important information of patient’s tumor-infiltrating immune landscape, and we have included a discussion in S2 Text.
There are three tuning parameters k, α1, and α2 in FARDEEP. Since α1 is only used in the first iteration, a relatively small α1 is preferred to ensure that FARDEEP does not remove too many outliers at the first step. In practice, FARDEEP is not sensitive to different values of α1, and α2, so we set them to 0.1 and 1.5 respectively by default. However, k controls the number of outliers in each iteration and is critical for the performance of FARDEEP. Thus, we tune k on a case-by-case basis for each sample to preserve meaningful fluctuations of gene expression levels. Effects for different tuning parameters are shown in S1 Table. Since the test group may contain outliers that influence the accuracy of the tuning result, cross-validation is not advised. Instead, we applied the Bayesian Information Criterion (BIC) and assume that the errors follow a log-normal distribution instead of a normal distribution among gene expression datasets as suggest by Beal [32]. We define the modified BIC referring to the setting of She and Owen [33]:
BIC * ( k ) = mlog ∑ i = 1 n 1 { i ∉ E ^ }log 2 ( y i − y ^ i ) 2 m + b ( log ( m ) + 1 ) , (0.8)
where E ^ being the set of detected outliers, b is number of parameters and equals N ^ + p + 1 with N ^ = | E ^ | being the number of outliers, and m equals n − N ^. Then, we choose the value of k associated with the smallest BIC*.
To test the performance of FARDEEP, we compared our approach with the existing methods using numerical simulations and real datasets. Here, we list the outlier genes detected by FARDEEP for real datasets in S4 Table. We use LM22 signature matrix containing 22 immune cell types hematopoietic cells for Microarray data and use quanTIseq signature matrix containing 10 immune cell types for RNA-Seq data. To compare the performance of different methods, we report the sum of squared error (SSE), the coefficient of determination denoted as R-squared (R2) and the Pearson correlation (R) defined as follows
SSE = ∑ j = 1 p ( β j * − β ^ j ) 2 , R 2 = 1 − ∑ j = 1 p ( β j * − β ^ j ) 2 / ∑ j = 1 p ( β j * − β ¯ * ) 2 , β ¯ * = 1 p ∑ j = 1 p β j * , R = ∑ j = 1 p ( β j * − β ¯ * ) ( β ^ j − β ^ ¯ ) ∑ j = 1 p ( β j * − β ¯ * ) 2 ∑ j = 1 p ( β ^ j − β ^ ¯ ) 2 , β ^ ¯ = 1 p ∑ j = 1 p β ^ j ,
where β* is the ground truth, and β ^ is the estimate.
To test the robustness of FARDEEP under different error conditions, we simulated three datasets refer to the setting in [33, 34] with normally distributed errors, heavy tailed errors. The observations were generated based on the linear regression model (0.2). The predictor matrix is X = (x1, …, xn)′ = UΣ1/2, where U i j ∼ U ( 0 , 20 ) and Σ i j = ρ I { i ≠ j } with ρ = 0.5. Consider the proportion of outliers f ∈ {5%, 10%, 20%, 30%}, sample size n = 500, and number of predictors p = 20, we added random errors and outliers to the simulated data as follows:
Random errors: we generated the random error vector from i) standard normal distribution, ii) t-distribution with 3 degrees of freedom.
Vertical outliers: we generated a n dimensional zero vector τ and randomly selected nf elements in τ to be the outliers generated from a non-central t-distribution with 1 degree of freedom and a non-centrality parameter of 30.
Leverage points: we took 20% of the contaminated data as leverage points, that is, replacing the corresponding predictors by the samples from N ( 2 max ( X ) , 1 ).
The coefficients βj were sampled from U ( 0 , 1 ), where j = 1, …, p. Based on the framework above, the dependent variable could be obtained by
y = X β + τ + ε .
We simulated each model 50 times. As shown in Figs 1 and 2, FARDEEP outperforms other methods, evidenced by the SSE, R2 and R values.
To check FARDEEP’s accuracy of outlier detection, we simulated {5%;10%;20%;30%} outliers using the same method as above for a model with both normally distributed and heavy-tailed noise. As shown in Table 1, the tuning parameter k decreases when the amount of outliers becomes larger, and the true positive rates always stay around 1, indicating that the tunning of k is highly effective.
In the supplementary material S3 Text, we also included another outlier construction scheme with X related outliers and a simulation setting with correlated responses. In both scenarios, FARDEEP dominates other methods in terms of SSE, R2 and R values.
Following the similar procedure as in Newman et al., we randomly generated the abundance of different cells from interval [0, 1] [12]. Notably, the sum of cell abundance is not necessarily 1. The measurement errors were sampled from 2 N ( 0 , ( 0 . 1log 2 ( s ) ) 2 ). To incorporate outliers, we randomly selected i/50 of the data and replaced them with data drawn from 2 N ( 10 , ( 0 . 3log 2 ( s ) ) 2 ) where i = 1, 2, …, 25 and s is the standard deviation of original mixtures.
We repeated the procedure nine times and estimated the cell type abundance using FARDEEP, CIBERSORT (without converting to percentage), NNLS, PERT, and DCQ. As shown in S2 Table, we found that the SSE range for FARDEEP is 1.51 × 10−7 to 1.47 × 10−4, R2 and R keeps being 1 regardless of the number of outliers, while Other methods show significantly larger SSE and smaller R2, R.
We used the cell line dataset GSE11103 generated by Abbas et al. [35] that contains gene expression profiles of four immune cell lines (Jurkat, IM-9, Raji, and THP-1) and four mixtures (MixA, MixB, MixC, and MixD) with various ratios of cells. Before analysis, we quantile normalized the mixture data for 54675 probesets and downloaded the immune gene signature matrix with 584 probesets from CIBERSORT website. Then, we applied five deconvolution methods, including FARDEEP, CIBERSORT (without converting to percentage), DCQ, NNLS, and PERT, to calculate the sum of squared errors of the estimated abundance of the four immune cell lines. We also compared with CIBERSORT absolute mode, which is a beta version in CIBERSORT website (S1 Fig). Since the CIBERSORT absolute mode is a beta version and leads to suboptimal results compared with CIBERSORT, we only focused on the comparisons with CIBERSORT. We show that FARDEEP gives the smallest SSE and the largest R2, which indicates the most accurate result (Fig 3).
Both CIBERSORT and FARDEEP are robust deconvolution methods and show advantages in the above datasets, we next sought to compare their performances on mixtures with unknown content. We followed the simulation setting proposed by Newman et al. [12] and downloaded the signature gene file from CIBERSORT website. The mixture file was constructed from the four immune cell lines data, as mentioned in the previous section, and a colon cancer cell line HCT116 (average of GSM269529 and GSM269530 in GSE10650). Cancer cells were mingled into immune cells at different ratios {0%, 30%, 60%, 90%}. Noise was added by sampling from the distribution 2 N ( 0 , ( flog 2 ( s ) ) 2 ), in which f ∈ {0%, 30%, 60%, 90%} and s is the standard deviation of original mixtures. By applying FARDEEP and CIBERSORT (without converting to percentage) on 64 mixtures, we found that FARDEEP remains an accurate estimation, while the tumor contents skew the results of CIBERSORT with larger deviation from the ground truth (Fig 4).
To evaluate the performance of FARDEEP in immune-cell-rich settings that are less affected by outliers, we downloaded and analyzed two gene expression datasets (GSE65135 [12] and GSE20300 [36]) generated from the Affymetrix Microarray, which is the same platform used to generate the signature matrix LM22. The GSE65135 dataset consists of (i) surgical lymph node biopsies of 14 follicular lymphoma patients and (ii) purified B and T cells from the tonsils of 5 healthy controls, and the GSE20300 dataset includes 24 blood samples from pediatric renal transplant patients. Flow cytometry or coulter counter data in these studies, which are presented in relative scales, are treated as ground truth. Thus, we normalized the estimated parameters of each method to the sum of 1 before comparison.
As shown in Fig 5A and 5B for case (i) of GSE65135 and Fig 5D and 5E for GSE20300, FARDEEP outperformed CIBERSORT in terms of R2, R and SSE, which is consistent with our findings with simulated datasets. For case (ii) of GSE65136, we estimated the immune cell composition for purified B and T cells with purity level exceeding 95% and 98%, respectively. For purified B cells, CIBERSORT tends to return non-zero estimates for T cell and a large proportion of other cell types, while FARDEEP gave almost all zero estimates for T cell and on average reduced the estimation error by 61%. Similarly, for the purified T cell, although CIBERSORT had a better performance compared to purified B cell, FARDEEP still significantly improves the estimation accuracy by reducing on average 48% of the estimation error (Fig 5C). Furthermore, as shown in S4 Table, FARDEEP detected gene CD79A and BCL2A1 as outliers for most samples in case (i) of GSE65135. These two genes are known to have high expression levels in follicular lymphoma (B-cell lymphoma) cells [37].
Overall, even in specimens that are rich in immune cells without contamination by non-hematopoietic malignancy, FARDEEP still outperforms CIBERSORT in immune cell deconvolution.
In addition to effectively handling Microarray data, FARDEEP can also deconvolve TILs using RNA-seq data when we replace the signature matrix LM22 with quanTIseq, a signature matrix generated from RNA-seq data containing ten different immune cell types [14]. We applied CIBERSORT and FARDEEP using signature matrix quanTIseq to peripheral blood mononuclear cell (PBMC) mixtures (GSE64655) generated by Hoek et al. [38], and lymph node bulk samples of 4 melanoma patients from GSE93722 [39]. Flow cytometry data in these studies are on a relative scale and are treated as ground truth. We normalized the estimated parameters of each method to a relative scale using (0.7) before comparison. The RNA-seq data are usually less noisy compared to Microarray, and PBMC datasets are usually clean with less unknown contents. Therefore, we expect FARDEEP and CIBERSORT will return comparable results, which is the case in Fig 6A and 6B. However, when dealing with noisier data containing more outliers such as lymph node bulk samples, FARDEEP obtained larger advantage over CIBERSORT as shown in Fig 6C and 6D.
TME of solid carcinomas are different from a lymph node biopsy or peripheral blood, and the highly variable gene expression in cancer cells challenges the accuracy of immune cell deconvolution. It is well-established that immune infiltration profile serves as a promising prognosticator [4, 5]. Hence, we next utilized survival and gene expression data of ovarian cancer (OV) and lung squamous cell carcinoma (LUSC) from The Cancer Genome Atlas (TCGA) database to assess the prognostic relevance of different deconvolution methods. These two datasets were chosen because only LM22 not the RNA-seq based signature matrix quanTIseq includes γδ T cells, and OV and LUSC from TCGA datasets are the only two cancer types with Affymetrix microarray data. Using gene expression data (n = 514 for OV and n = 133 for LUSC), we estimated the immunoscore using ESTIMATE proposed by yoshihara et al. [40], TILs proportion using CIBERSORT, as well as TILs subset scores using CIBERSORT (without converting to percentage) and FARDEEP. Cold tumors typically harbor lower numbers of CD8+ T cells, γδ T cells, M1-like macrophages, and NK cells [11, 41–43]. Thus, we calculated an anti-tumor immune subsets score by the summation of CD8+ T cells, γδ T cells, M1-macrophages, and NK cells. Then, we partitioned the patients into two groups with equal size using the median of either the immunoscore (ESTIMATE) or anti-tumor immune subsets score (CIBERSORT and FARDEEP). We compared the survival curves between the two groups. As shown in Fig 7, FARDEEP most effectively separates patients into high- and low- risk groups with the smallest p-value (p-value = 0.0065 and 0.059 for OV and LUSC respectively). Recently, CIBERSORT website supports a beta-version of an absolute mode for cell deconvolution. We also included CIBERSORT absolute mode in this survival analysis and showed that it returned a better result (p-value = 0.037) compared to the relative mode in the OV dataset. FARDEEP shows a stronger performance with a smaller p-value under this setting (S2 Fig). These results demonstrated that the TIL subset scores could provide additional clinical-relevant information compared to the relative abundance.
In addition, we expected the summation of these TIL subset scores would negatively correlate with tumor purity. To prove this hypothesis, we calculated the summation of 22 TIL subset scores for both OV and LUSC datasets and correlated them with the tumor purity estimated from consensus measurement of purity estimations (CPE) [44]. Even without taking account of stromal cells, as shown in S3 Fig. the summation of TIL subset scores is negatively correlated with tumor purity.
Next, we sought to investigate whether outlier removal reduces contamination by transcripts from cancer cells. We first identified those top outlier-genes, which were consistently removed by FARDEEP in the OV dataset and obtained the average expression values of those outlier-genes from OV cell lines in GSE32474 [45]. As shown in S3 Table, most of these outlier-genes have high expression in cancer cell lines. For example, CXCL10 gene encodes an important chemokine to recruit CD8+ T cells and is also highly expressed in ovarian cancer cells. Thus, although some genes in LM22 may play a role in immune cells, they may be also highly expressed and variable among cancer cells. Such cross-contamination likely skews immune deconvolution analysis. As shown in S3 Table, FARDEEP successfully detected and removed those genes, leading to a more robust and accurate deconvolution analysis.
The cancer immune microenvironment has emerged as a critical prognostic dimension that modulates patient responses to neoadjuvant therapy. However, the current clinical TNM staging system does not have a consistent method to stratify cancers based on their immunogenicity. The recent study shows that the RNA-seq datasets of whole tumors contain valuable prognostic information to assess the cancer-immunity interactions [12, 46]. But the current methods to extract immune signatures are susceptible to the frequent outliers in the datasets, leading to less effective identification of cold tumors. Based on support vector regression, CIBERSORT is one of the most popular robust deconvolution methods. However, this model does not include an intercept to capture possible contribution from other cell types and performs a z-normalization to the data before fitting the regression model, which introduces biases into the output. Discussion of the effect of Z-score normalization for CIBERSORT is included in S1 Text. In this study, we developed a new machine learning tool, FARDEEP, to streamline the removal of outliers and increase the robustness of gene-expression profile deconvolution. Robustness is an indispensable feature to solve a problem of deconvolution because gene expression data are frequently contaminated by a large amount of outliers. FARDEEP solves the deconvolution problem in a robust way because this tool evaluates all outliers across the datasets and then examines the true immune gene signature using non-negative regression. This feature is especially useful to analyze tumors with significant non-hematopoietic tumor components. Interestingly, although FARDEEP and the current robust methods can both tackle immune-cell-rich specimens such as lymph node and PBMCs, FARDEEP exhibits improved prognostic potential when dealing more complex datasets with significant carcinoma cell content.
Although FARDEEP provides a robust computational algorithm to better solve the gene-expression deconvolution problem with noisy datasets, its performance and application rely on the choice of the signature matrix. To avoid estimation bias, it is important to choose the signature matrix derived from the same platform as the mixture matrix. For example, if dealing with gene expression data measured by Affymetrix HGU133A, we should use LM22, but if dealing with RNA-seq data, the signature matrix quanTIseq is preferred. Overall, here we show that FARDEEP is a powerful and rapid machine learning tool that outperforms existing robust methods for gene deconvolution in datasets with significant heavy-tailed noise. FARDEEP provides a new technology to interrogate cancer immunogenomics and more accurately map the immune landscape of solid tumors.
|
10.1371/journal.ppat.1005667 | CD4 T Cell-Derived IFN-γ Plays a Minimal Role in Control of Pulmonary Mycobacterium tuberculosis Infection and Must Be Actively Repressed by PD-1 to Prevent Lethal Disease | IFN-γ–producing CD4 T cells are required for protection against Mycobacterium tuberculosis (Mtb) infection, but the extent to which IFN-γ contributes to overall CD4 T cell-mediated protection remains unclear. Furthermore, it is not known if increasing IFN-γ production by CD4 T cells is desirable in Mtb infection. Here we show that IFN-γ accounts for only ~30% of CD4 T cell-dependent cumulative bacterial control in the lungs over the first six weeks of infection, but >80% of control in the spleen. Moreover, increasing the IFN-γ–producing capacity of CD4 T cells by ~2 fold exacerbates lung infection and leads to the early death of the host, despite enhancing control in the spleen. In addition, we show that the inhibitory receptor PD-1 facilitates host resistance to Mtb by preventing the detrimental over-production of IFN-γ by CD4 T cells. Specifically, PD-1 suppressed the parenchymal accumulation of and pathogenic IFN-γ production by the CXCR3+KLRG1-CX3CR1- subset of lung-homing CD4 T cells that otherwise mediates control of Mtb infection. Therefore, the primary role for T cell-derived IFN-γ in Mtb infection is at extra-pulmonary sites, and the host-protective subset of CD4 T cells requires negative regulation of IFN-γ production by PD-1 to prevent lethal immune-mediated pathology.
| The development of novel tuberculosis vaccines has been hindered by the poor understanding of the mechanisms of host-protection. It has been long-held that IFN-γ is the principle effector of CD4 T cell-mediated resistance to Mtb infection, but Mtb-specific CD4 T cells produce low amounts of IFN-γ in vivo, leading to the possibility that increasing IFN-γ production by Th1 cells might enhance control of Mtb infection. However, the precise contribution of IFN-γ to CD4 T cell-dependent protection and the outcome of increasing IFN-γ production by CD4 T cells have not been evaluated. Here we show that IFN-γ accounts for only ~30% of the cumulative CD4 T cell-mediated reduction in lung bacterial loads over the first 1.5 months of infection. Moreover, we find that increasing the per capita production of IFN-γ by CD4 T cells leads to the early death of the host. Lastly, we show that suppression of CD4 T cell-derived IFN-γ by the inhibitory receptor PD-1 is essential to prevent lethal disease. Therefore, poor control Mtb infection does not result from defective production of IFN-γ, and strategies to selectively boost it are unwarranted. Furthermore, identifying the primary mechanisms of CD4 T cell-dependent control of Mtb infection should be a priority.
| Mycobacterium tuberculosis (Mtb) infection is a leading cause of global morbidity and mortality. In 2014 there were 9.6 million new cases of tuberculosis (TB) and 1.5 million deaths resulting from Mtb infection [1]. The only available vaccine against Mtb infection, Bacillus Calmette-Guérin (BCG), is an attenuated strain of M. bovis that was developed almost a century ago. BCG immunization does prevent severe forms of childhood TB but at best poorly protects against adult disease [2] It is widely accepted that effective vaccination approaches for TB would have an enormous impact on global health; however, efforts in TB vaccine development have been hindered by the lack of mechanistic insight into the cellular and molecular basis of both protective immunity and immunopathology during TB.
CD4 T cells are essential for host resistance to Mtb infection [3] and the protection afforded by various vaccination approaches in experimental animal models is mediated mainly by CD4 T cells [4–8]. Although other cell types may make contributions to vaccine-elicited protection against Mtb infection, it seems likely that a successful vaccination strategy will require the induction of MHC class II-restricted CD4 T cell responses of the sufficient quantity, location, breadth of specificity, and polarized effector capacity. IFN-γ is a key CD4 T cell-derived cytokine and essential for resistance to mycobacterial infections. Mice deficient in IL-12, T-bet, or IFN-γ itself are extremely susceptible to Mtb infection [9]. Humans with inborn errors in the IFN-γ axis are highly susceptible to normally avirulent non-tuberculous mycobacterial (NTM) infections. Deficiencies in IL-12p40, IL-12RI or II, IFN-γR or STAT-1 [10, 11] all result in severe NTM infections early in life. Adults who develop anti-IFN-γ neutralizing autoantibodies are also very prone to mycobacterial infections later in life [12]. Due to the severity of infection in its absence, IFN-γ is often considered the primary mechanism by which the host controls Mtb infection. Although several immune cell types can produce IFN-γ, it has been shown that IFN-γ must be produced by CD4 T cells for the host to survive Mtb infection [13]. However, IFN-γ responses do not correlate with better outcome of Mtb infection [14–16], and a recent vaccine efficacy trial based on a viral vector containing an Mtb antigen was found to generate bacilli-specific CD4 T cells capable of producing high levels of IFN-γ but afforded no protection against the development of TB [17]. It remains unclear why IFN-γ responses are not observed to correlate with resistance to Mtb infection. This may suggest that very small amounts of IFN-γ are needed for optimal protection, or that IFN-γ may even be associated with disease rather than protection during active TB [18]. Also, these findings could indicate that other T cell effector molecules are the primary mediators of bacterial control.
Several studies have found that the amount of IFN-γ produced by Mtb-specific [19, 20] or BCG-specific [21] CD4 T cells in vivo is much lower than expected based on their ability to produce IFN-γ after stimulation with high doses of peptide, with usually <10% of the cells producing IFN-γ in the infected tissue. This has led to the hypotheses that defective production of IFN-γ by Mtb-specific CD4 T cells contributes to the host’s inability to control Mtb. This poor IFN-γ production by Mtb-specific CD4 T cells is at least in part due to poor antigen (Ag) presentation, as these studies have found that increasing Ag concentrations in the tissue enhanced IFN-γ production by CD4 T cells in vivo, and Mtb has evolved mechanisms that interfere with Ag-presentation [22]. The low IFN-γ production may also reflect the degree of effector polarization of the CD4 T cells. Based on several studies, it is now clear that Mtb-specific CD4 T cells that have migrated into the lung tissue parenchyma of Mtb-infected mice (marked by CXCR3 expression) express lower amounts of the key transcription factor T-bet and produce less IFN-γ compared to the KLRG1+CX3CR1+ CD4 T cells that localize to the lung-associated blood vasculature [20] [23] [24] [25] [26]. Despite their reduced IFN-γ producing capacity, this population of CXCR3+KLRG1-CX3CR1- CD4 T cells displayed the best protective capacity against Mtb infection, likely due to their ability to migrate into the lung tissue and interact with infected myeloid cells within the granulomas. Indeed, direct interactions between CD4 T cells and MHCII molecules on infected macrophages is required for the optimal suppression of intracellular growth of Mtb [27]. T cell dysfunction in the form of exhaustion or adaptive tolerance could also play a role in the reduced IFN-γ production by Mtb-specific CD4 T cells. The lung-homing, host-protective population of CD4 T cells expresses high levels of multiple co-inhibitory receptors including programmed cell death-1 (PD-1) [23, 24], which is well known to suppress IFN-γ production [28] [29]. The lungs of Mtb-infected PD-1 KO mice contain greatly increased concentrations of IFN-γ compared to WT mice [30, 31]. However, rather than displaying enhanced control of Mtb, these animals succumb due to CD4 T cell-mediated immunopathology, although precise mechanisms leading to the death of PD-1 KO mice have not been identified. Therefore, there is reason to suggest that defective IFN-γ production by CD4 T cells could account for poor control of Mtb, but also data to indicate that increased IFN-γ production by CD4 T cells might be detrimental under some circumstances.
These observations prompted us to ask several questions regarding the role of CD4 T cell-derived IFN-γ in Mtb infection. First, what is the relative contribution of IFN-γ production to total CD4 T cell-mediated control of Mtb infection? Second, is the inability of the host to better control Mtb infection due to the inadequate production of IFN-γ by the lung parenchymal CD4 T cells? Third, does selectively enhancing the production of IFN-γ by CD4 T cells result in better control of Mtb infection? Lastly, what is the role of PD-1-mediated inhibition of IFN-γ production by Mtb-specific CD4 T cells in protection and pathology? We found that the relative contribution of CD4 T cell-derived IFN-γ to the control of Mtb infection varies dramatically by tissue, with surprisingly small contribution in the lungs compared to the spleen. The amount of IFN-γ production by pulmonary CD4 T cells is more than sufficient for optimal bacterial control in the lung, and increasing the amount of IFN-γ that individual CD4 T cells produce leads to lethal disease. Moreover, we found that host survival of Mtb infection requires PD-1 mediated inhibition of IFN-γ production by parenchymal CXCR3+CX3CR1-KLRG1- effector CD4 T cells, which mediate the best protection against Mtb infection. Therefore, the inability of the host to control Mtb infection does not result from poor IFN-γ production, and selectively increasing IFN-γ production on a per cell basis is detrimental to the outcome of pulmonary infection. These data refine our view of the role of IFN-γ, which is classically considered the primary determinant of anti-mycobacterial immunity, and provide insight into the role of T cell co-inhibition in Mtb infection. Moreover, they show that the same CD4 T cells that mediate best host resistance to Mtb infection are also the most capable of driving lethal disease if not properly regulated.
To examine the role of CD4 T cell-derived IFN-γ in host survival and bacterial control in different tissues, we performed a series of adoptive transfer experiments using RAG1 KO recipients (Fig 1A). Transfer of WT CD4 T cells protected Mtb infected RAG1 KO hosts for >130 days (Fig 1B). In contrast, reconstitution with IFN-γ KO CD4 T cells only slightly extended survival compared to recipients that did not receive T cells (mean survival time, 60 versus 43 days, respectively; Fig 1A and 1B). Thus, production of IFN-γ by CD4 T cell is required for host survival of Mtb infection, consistent with several previous reports [13, 32].
The relative survival of mice containing WT or IFN-γ KO CD4 T cells, however, does not provide quantitative information on the importance of CD4 T cell-derived IFN-γ to bacterial control. We next sought to estimate the relative contribution of IFN-γ to the overall CD4 T cell-mediated suppression of Mtb growth in vivo. To do so, we utilized a similar adoptive transfer model and monitored bacterial loads in the lungs and spleens until 42 days post-infection (p.i.), after which unreconstituted RAG1 KO mice succumb and further comparisons with T cell reconstituted mice cannot be made. Compared to WT CD4 T cells, IFN-γ KO CD4 T cells were able to suppress Mtb replication far better in the lungs compared to the spleen (Fig 1C). Using a simple mathematical model we showed that by comparing the area under the log transformed bacterial growth curves in control RAG1 KO mice vs. recipients of IFN-γ KO or WT CD4 T cells, we could estimate the relative contribution of CD4 T cell-derived IFN-γ to the overall CD4 T cell-mediated control of Mtb growth in the lungs and spleens (see S1 Method for description of the mathematical model). By performing these calculations for two independent experiments, we found that CD4 T cell-derived IFN-γ accounted for ~30% of CD4 T cell-dependent bacterial control in the lungs between day 0 to 42 p.i., but ~80% in the spleens (Fig 1D). Based on our calculation, CD4 T cell-derived IFN-γ contributes surprisingly little to control of pulmonary Mtb infection, but is the principal protective effector molecule in the spleen, at least during the first 6 weeks post-exposure.
The relatively minor contribution of CD4 T cell-derived IFN-γ to control of Mtb in the lungs might be explained by insufficient production of this cytokine by pulmonary T cells [24]. To test this hypothesis, we asked if the amount of IFN-γ produced by CD4 T cells is below what is needed to achieve maximum levels of CD4 T cell-mediated bacterial control in the tissues. To this end, RAG1 KO mice infected with Mtb 7 days prior were reconstituted with naïve CD4 T cells (3 × 106 cells/recipient) at varying ratios of WT to IFN-γ KO CD4 T cells, and on day 42 p.i. lungs and spleens were harvested (Fig 2A). We found that IFN-γ concentrations in the lung homogenates rose steadily as the proportion of WT CD4 T cells in the mixed inoculum was increased (ranging from ~100 pg/ml to ~8 ng/ml; Fig 2B). Mice reconstituted with IFN-γ KO CD4 T cells alone showed a ~60-fold decrease in CFU in the lungs compared to the group not receiving cells (Fig 2C, upper left). As the proportion of WT T cells to IFN-γ KO cells was increased, the bacterial loads in the lungs further decreased ~7-fold. Importantly, the CFU reached a nadir at 40% WT cells, and although the concentrations of IFN-γ continued to rise as more IFN-γ–producing T cells were introduced, the bacterial loads did not further decrease. In fact, mice that received 100% WT CD4 T cells had the same CFU counts in their lungs as mice that received 40% WT + 60% KO T cells. In the spleen, adoptive transfer of IFN-γ KO CD4 T cells alone into RAG1 KO mice resulted in only ~5-fold reduction in CFU compared to the control mice that did not receive cells (Fig 2C, upper right). In contrast to the lungs, the bacterial loads in the spleens continued to decrease as the proportion of WT CD4 T cells increased, reaching a nadir of ~260-fold reduction at 100% WT compared to 100% IFN-γ KO CD4 T cells. Therefore, these data support our previous results showing a tissue-specific role for IFN-γ (Fig 1). Moreover, these data suggest that the limited contribution of CD4 T cell-derived IFN-γ in the lungs cannot be explained by its insufficient production.
To further test the hypothesis that defective IFN-γ production by CD4 T cells does not limit the effectiveness of Mtb-specific CD4 T cells, we performed the in vivo titration experiment with mixtures of IFN-γ–overproducing CD4 T cells and IFN-γ KO CD4 T cells. To do so, we employed CD4 T cells in which the AU-rich element (ARE) in the 3’ untranslated region of the IFN-γ gene is deleted (ARE-Del), resulting in stabilization of the IFN-γ mRNA and increased IFN-γ translation [33]. The ARE-Del CD4 T cells produced ~2-fold higher amounts of IFN-γ in the lung compared to WT CD4 T cells at each incremental ratio (Fig 2B). We confirmed the increased expression of IFN-γ in ARE-Del CD4 T cells by direct ex vivo intracellular staining (DrxICS) in the mice that were received a 1:1 mixture of WT and ARE-Del CD4 T cells (Fig 2D). By this flow cytometry-based assay, there was again ~2 fold increase in IFN-γ expressing CD4 T cells in the ARE-Del compared to WT CD4 T cells. We found an approximately 1.1 fold increase in the geometricMFI of IFN-γ staining in the ARE-Del CD4 T cells compared to WT counterparts in the same lung (1.06 ± 0.01, P = 0.02), but due to the very low staining and lack of separation from the negative cells, inherent in the approach, we cannot be certain of the precise increase in the amount of cytokine produced per cell in the ARE-Del CD4 T cells. Strikingly, IFN-γ–over-producing ARE-Del CD4 T cells were less effective at controlling Mtb growth in the lungs compared to WT CD4 T cells. As IFN-γ–over-expressing CD4 T cells were titrated into the system, CFU counts reached a nadir of ~2 fold reduction (below the CFU amounts found at 100% IFN-γ KO CD4 T cells) at a ratio of 40% ARE-Del to 60% KO CD4 T cells (Fig 2C). As the fraction of ARE-Del CD4 T cells was increased further, the CFU began to rise slightly, so that the levels of bacteria in the lungs of mice that received 100% IFN-γ KO CD4 T cells and 100% ARE-Del CD4 T cells were identical, despite the enormous difference in IFN-γ concentrations in the lung. In the spleens however, the IFN-γ–overproducing CD4 T cells were more effective than WT CD4 T cells at controlling Mtb growth (Fig 2C). Mice receiving 100% ARE-Del CD4 T cells displayed a ~690-fold reduction in bacterial loads in their spleens compared to recipients of 100% IFN-γ KO CD4 T cells versus the 260 fold reduction seen in WT CD4 T cell recipients, indicating that impressive levels of bacterial control can be achieved in the spleen when IFN-γ production by CD4 T cells is amplified beyond normal levels. This experiment allowed us to examine the relationship between IFN-γ in the lung tissues and lung bacterial loads in the setting of WT and elevated IFN-γ production by CD4 T cells (Fig 2E). With WT CD4 T cells, there was a negative relationship between IFN-γ concentrations and CFU counts in the lung that levels off at ~4 ng/ml IFN-γ in the lungs in this particular experimental setting. In contrast, with IFN-γ–overproducing CD4 T cells this negative correlation was lost. These data highlight how the impact of IFN-γ concentrations on bacterial loads depends more on the amount of the cytokine produced per T cell rather than the total amount found in the tissue. Moreover, increasing IFN-γ production by CD4 T cells not only failed to enhance suppression of Mtb growth in the lungs, but actually impaired control.
To examine the outcome of increased IFN-γ production by CD4 T cells on host survival of Mtb infection, we reconstituted Mtb-infected TCRα KO mice with WT, ARE-Del or a 1:1 mixture of WT and ARE-Del CD4 T cells and monitored mouse survival. Mice that did not receive cells all succumbed by day 60 p.i., whereas mice reconstituted with WT CD4 T cells all survived for >150 days (Fig 2F). Notably, mice reconstituted with ARE-Del CD4 T cells all succumbed by day 80 p.i. This increased susceptibility was observed even in the presence of WT CD4 T cells, indicating that ARE-Del CD4 T cells actively promote early mortality after Mtb infection. Collectively, these data show that the inability of the host to control Mtb infection in the lungs does not result from inadequate production of IFN-γ, and increased IFN-γ production by CD4 T cells is, in fact, detrimental to control pulmonary infection and host survival of Mtb infection. In marked contrast, control of Mtb infection in the spleen is highly dependent on CD4 T cell-derived IFN-γ and growth of Mtb is tightly linked to the levels of IFN-γ produced by splenic T cells.
TNF is critical for host resistance to TB [34], and TNF production by CD4 T cells is required for the long-term control of Mtb infection [35] [36] [37]. Given the limited role for CD4 T cell-derived IFN-γ in control Mtb infection in the lung, we next asked if TNF could account for the majority of CD4 T cell-dependent pulmonary bacterial control. To do so, we performed similar in vivo titration experiments with WT and TNF KO CD4 T cells (Fig 3A). We observed that TNF concentrations were the highest in the both lungs and spleens of RAG1 KO mice that did not receive cells and that reconstitution of RAG1 KO mice with 100% TNF KO CD4 T cells leads to a decrease in TNF levels in the lung and spleen homogenates (Fig 3B). Mice that received 100% TNF KO CD4 T cells also displayed a ~190 fold reduction in bacterial loads in the lungs and ~810 fold reduction in the spleens (Fig 3C). Therefore, TNF KO CD4 T cells were able to dramatically restrict bacterial replication in both tissues, likely explaining the decreased amounts of non-T cell derived-TNF in the tissue homogenates. As the proportion of WT cells in the transferred cell population was increased, the levels of TNF in the lung and spleen homogenates gradually rose (Fig 3B). In the spleens, however, increasing the percentage of the T cell inoculum above 40% WT resulted in a decline in total TNF concentrations. Importantly, as the proportion of WT CD4 T cells in the inoculum was increased, the bacterial loads decreased an additional ~ 4 fold below what was observed with 100% TNF KO CD4 T cells in both the lungs and spleens (Fig 3C, left). Therefore, we did not observe a tissue-specific role for CD4 T cell-derived TNF in control of Mtb infection as we did for IFN-γ. These data also argue that TNF is not the major contributor to CD4 T cell-dependent control of Mtb infection in the lung.
Given that IFN-γ makes relatively little contribution to control of pulmonary Mtb infection and that artificial over-expression of IFN-γ promotes disease, we next considered the possibility that IFN-γ production by CD4 T cells must be appropriately suppressed, not enhanced, for host-resistance to Mtb infection. The co-inhibitory receptor PD-1 is well understood to suppress IFN-γ production by T cells [28, 29], and PD-1 KO mice are extremely susceptible to Mtb infection [30, 31, 38] due to CD4 T cell-driven disease [31]. Here we confirmed that PD-1 KO mice succumb rapidly to infection with high levels of IFN-γ in their lungs (Fig 4A and 4B). Moreover, using an intravascular stain (iv-stain) technique that allows discrimination between intravascular and parenchymal T cells [39], we confirmed that I-AbESAT-64−17 and I-AbEsxG46-61-specific CD4 T cells in the lung tissue parenchyma express higher levels of PD-1 than their iv counterparts ([24]; Fig 4C). We next used the iv-stain approach to determine the impact of PD-1 deficiency on the distribution of Mtb-specific CD4 T cells between the lung parenchyma and lung-associated blood vasculature. PD-1 KO mice showed a remarkable increase in accumulation of I-AbESAT-64–17–specific CD4 T cells in the lung parenchyma compared to WT mice (~95% versus ~60%, respectively; Fig 4D). The increased accumulation of lung parenchymal CD4 T cells in the PD-1 KO mice was associated with loss of markers we have previously shown to identify intravascular CD4 T cells (i.e., CX3CR1+, KLRG1+ and T-bethigh) and increased expression of makers that are associated with the protective parenchymal cells including CXCR3 and ICOS but decreased level of CD69 ([20, 24]; Fig 4E). We next compared the ability of lung Mtb-specific CD4 T cells to produce IFN-γ upon in vitro stimulation with ESAT-61–20 peptide, and observed a marked increase in lung parenchymal localization of IFN-γ–producing CD4 T cells in PD-1 KO mice compared to WT mice (~90% versus ~40%, respectively; Fig 4F). To evaluate the amount of IFN-γ production by WT and KO Mtb-specific CD4 T cells in vivo, IFN-γ expression by I-AbESAT-64–17–specific CD4 T cells was measured by DrxICS. We found that ~2-fold more parenchymal tetramer+ PD-1 KO CD4 T cells actively expressed IFN-γ in vivo compared to WT counterparts (Fig 4G). These data show that during Mtb infection PD-1 suppresses the generation of CD4 T cells with the lung-homing phenotype as well as the production of IFN-γ by the CD4 T cells that have migrated into the lung parenchyma.
Although our data suggest that PD-1 has a role in regulating Mtb-specific CD4 T cell responses in the lung parenchyma, we also found a positive correlation between the bacterial load in the lungs and the frequency of lung parenchymal CD4 T cells (Fig 5A). Thus, the increase in CD4 T cell accumulation and IFN-γ production in the lung parenchyma might be secondary to the higher bacterial burden in the lungs of PD-1 mice (Fig 4D and 4F) rather than to a direct effect of the loss of PD-1. To examine the cell-intrinsic role of PD-1 in CD4 T cells, pure naïve (CD25loCD44loCD62Lhi) CD4 T cells were sorted (to >97% purity) from WT or PD-1 KO mice, mixed at 1:1 ratio, and transferred into Mtb-infected congenically disparate WT recipient mice (Fig 5B). In this approach, we cannot track Mtb-specific CD4 T cells with tetramers due to the extremely low frequency of donor cells, so we assume that the transferred CD4 T cells that have upregulated CD44 are responding to the infection. The benefit of this approach, however, is that we can follow a very rare population of polyclonal PD-1 KO CD4 T cells in a WT recipient host. The Thy1.1+ PD-1 KO donor CD4 T cells contained a higher frequency of CD44hi effectors compared to the CD45.1+ WT donor CD4 T cells (Fig 5C and 5D), consistent with the expected role of PD-1 in inhibiting T cell expansion. Similar to what we observed in intact PD-1 KO mice, a greater proportion of the PD-1 KO CD44high donor CD4 T cells were localized in the parenchyma (Fig 5C and 5E), and there were increased frequencies of KLRG1- CD4 T cells among the KO compared to WT donor effector cells in the same lungs (Fig 5C and 5F). As a control, we examined the naïve donor CD4 T cells derived from either WT or PD-1 KO mice and found them similarly localized to the lung parenchyma (Fig 5C and 5G). Importantly, the parenchymal donor effector CD4 T cells derived from PD-1 KO mice actively produced more IFN-γ than WT effectors as detected by DrxICS (Fig 5H). Together, these data indicate that the increased parenchymal CD4 T cell responses in PD-1 KO mice are not solely due to the elevated bacterial loads, and argue that PD-1 plays a cell-intrinsic role in inhibiting the generation of lung-homing CD4 T cells and their IFN-γ production in Mtb infection.
We previously have shown that adoptive transfer of purified lung parenchymal CD4 T cells into T cell–deficient hosts resulted in greater reduction in bacterial loads at 4 weeks p.i. compared to transfer of intravascular cells [24]. To further examine the role of these CD4 T cell subsets in host resistance to Mtb infection, parenchymal iv-stain negative (iv-) or intravascular iv-stain positive (iv+) CD44hi effector CD4 T cells were FACS purified from the lungs of WT mice on day 30 p.i. and adoptively transferred into TCRα KO mice that had been infected with Mtb 7 days prior to transfer (Fig 6A). Although both groups of recipient mice survived much longer than mice not receiving cells (Fig 6B), the recipients reconstituted with iv+ effector CD4 T cells all succumbed by ~day 180 p.i., whereas those reconstituted with iv- effectors all survived for >210 days (Fig 6B). These data confirm that lung parenchymal Mtb-specific CD4 T cells are more protective against Mtb infection compared to intravascular CD4 T cells despite relatively low levels of IFN-γ production in vivo.
We next asked if the CD4 T cells located in the lung parenchyma of PD-1 KO mice were able to drive disease. To do so, iv- CD44high CD4 T cells were FACS sorted from the infected lungs of WT mice or PD-1 KO mice and adoptively transferred into TCRα KO mice that had been infected with Mtb seven days previously (Fig 6C). As shown in Fig 6B, iv- parenchymal effector CD4 T cells derived from WT mice protected susceptible mice from the rapid mortality after Mtb infection (Fig 6D). By contrast, the parenchymal effector CD4 T cells derived from PD-1 KO mice accelerated the mortality of the reconstituted mice (mean survival time, 80 days). Moreover, mice reconstituted with a mixture of iv- CD44hi effector CD4 T cells from WT mice and PD-1 KO mice also rapidly succumbed to infection (mean survival time, 121 days), indicating that in the absence of PD-1 the parenchymal CD4 T cells that are normally protective against Mtb infection drive rapid mortality.
To test if increased IFN-γ production by the parenchymal PD-1 KO CD4 T cells was required to promote disease during Mtb infection, infected TCRα KO mice were reconstituted with WT or PD-1 KO CD4 T cells, or a mixture of WT and PD-1 KO or PD-1/IFN-γ double KO CD4 T cells (Fig 6E). Consistent with the previous finding [31], adoptive transfer of PD-1 KO CD4 T cells led to early mortality of the reconstituted mice, while WT CD4 T cells protected susceptible mice for a long period (mean survival time, 83 versus 217 days, respectively; Fig 6F). Furthermore, mice reconstituted with WT + PD-1 KO CD4 T cells died by day ~130 p.i., indicating that the presence of PD-1 KO CD4 T cells accelerated the mortality of reconstituted mice. Strikingly, mice reconstituted with WT + PD-1/IFN-γ double KO CD4 T cells survived as long as mice that received WT cells alone (mean survival time, 202 days; Fig 6F). Therefore, IFN-γ production by the co-transferred population of PD-1 KO CD4 T cells was required to drive the lethality. Taken together, these data demonstrate that the subset of lung-homing CD4 T cells that mediates host protection requires co-inhibition though PD-1 to inhibit IFN-γ–driven disease, highlighting how the same cells that are most protective against Mtb infection are potentially the most pathogenic if not properly regulated.
Here we used in vivo quantitative approaches to estimate the relative contribution of CD4 T cell-derived IFN-γ production to overall CD4 T cell-dependent control of Mtb infection. We compared the protective capacity of WT versus IFN-γ–overproducing CD4 T cells in different tissues. Our results show that IFN-γ from CD4 T cells accounts for ~30% of the total anti-bacterial effects of CD4 T cells in the lungs early in infection, and this relatively low contribution of IFN-γ is not explained by insufficient production of the cytokine by CD4 T cells. In fact, CD4 T cells that over-produce IFN-γ are less protective in the lungs and eventually drive lethal disease. In contrast, >80% of the protective effect of CD4 T cell responses in the spleen is due to their IFN-γ production, and the more IFN-γ is produced by CD4 T cells the more Mtb growth is suppressed in the spleen. We also found that the lung-homing, CXCR3+KLRG1-CX3CR1- subset of CD4 T cells, that normally mediates the best control of pulmonary Mtb infection, requires inhibition through PD-1 to prevent the detrimental over-production of IFN-γ that leads to the death of the host. Clearly some IFN-γ production by CD4 T cells is essential for host survival of Mtb infection. However, the results presented here collectively show that CD4 T cell-derived IFN-γ has a predominant role in control of bacterial dissemination and/or suppression of Mtb growth at extra-pulmonary sites, but currently we cannot discriminate between these two possibilities. At the same time, production of this cytokine by CD4 T cells also has significant potential for inducing pulmonary pathology.
We have only sampled the spleen as a representative extra-pulmonary site of infection, and do not know the role of CD4 T cell-derived IFN-γ in other tissues. However, the hypotheses that CD4 T cell-derived IFN-γ is essential for control of either dissemination or replication in extra-pulmonary tissues, and that it may also be able to drive disease is consistent with observations on the role of IFN-γ in mycobacterial infection in humans and non-human primates. Individuals with CD4 T cell or IFN-γ deficiencies are well known to be highly susceptible to extra-pulmonary mycobacterial infections [40, 41], and Mtb infected cynomologous macaques depleted of CD4 T cells develop rapidly disseminating disease [42, 43]. Moreover, IFN-γ concentrations in BAL fluid have been found to positively correlate with severity of disease [44]. A few trials testing administration of recombinant IFN-γ to individuals with active TB have been completed. Several found no effects of recombinant IFN-γ treatment on the outcome of Mtb chemotherapy in individuals with drug-resistant infections [45–47], but one study was halted early due to increased mortality of TB patients receiving recombinant IFN-γ [48].
It is important to point out that the lethality we observed when CD4 T cells over-produced IFN-γ was likely due to an increase in the amount secreted by CD4 T cells on a per cell basis and not due to the total amount of IFN-γ found in the tissue nor the total number of cytokine producing CD4 T cells. Indeed, it is well known that increasing the numbers of Mtb-specific CD4 T cells through vaccination or other means results in better control of Mtb infection. However, our results suggest that the generation of highly-polarized T cells that secrete large amounts of IFN-γ on a per cell basis may be counter-productive for control of pulmonary disease in Mtb infection. In contrast, it has been shown that Yeti mice, which similarly to the ARE-Del mice over-produce IFN-γ due to alterations in their mRNA regulation, are highly resistant to Listeria monocytogenes and Leishmania major infection [49]. Therefore, enhanced disease is not a common outcome of increasing the per capita IFN-γ production by CD4 T cells responding to intracellular infections. Andersen and Urdahl have suggested that induction of highly polarized Th1 responses are not ideal for TB vaccines in part due to their poor longevity and inability to home into the lung [50, 51]. Our results support and extend this view. We should, however, emphasize that our estimation of the role for IFN-γ in CD4 T cell-mediated control of Mtb infection was restricted to the first ~6 weeks post-exposure, and we do not know the importance of CD4 T cell-derived IFN-γ for the durability of bacterial control.
The mechanisms underlying the tissue-specific effects of CD4 T cell-derived IFN-γ in control of Mtb infection are not clear. Our data are reminiscent of previous results showing that Mtb infection is exacerbated when TNF is either absent or over-produced [52, 53]. The same paradigm is also true for the IL-1 pathway, as mice deficient in IL-1 or mice that lack iNOS dependent inhibition of IL-1 production are both highly susceptible to Mtb infection [54–56]. The need to optimally balance inflammation to achieve the best bacterial control while not inducing pathology seems to be a major refrain in TB immunology. Indeed, our data are consistent with the damage-response framework Casadevall and colleagues have elegantly put forward, arguing that optimal host benefit during a response to infection is achieved when damage caused by both the microbe as well as the immune response itself are simultaneously minimized [57].
Our results also provide an example of how the class of immune response against a pathogen must be tailored to the tissue in which it occurs, as described by Matzinger [58]. The spleen much more readily accommodates CD4 T cell-mediated IFN-γ production compared to the lung. Mechanistically, there are several testable hypotheses to explain the tissue-specific effects of CD4 T cell-derived IFN-γ in Mtb infection. It has been shown that when macrophages and DCs harbor relatively high numbers of bacilli, high concentrations of IFN-γ induce the necrotic death of the infected macrophages rather than the containment of the bacilli [59, 60]. Although not addressed here, death of the most heavily infected myeloid cells in the lung due to the over-production of IFN-γ might contribute to the impaired control of the infection. The differential IFN-γ sensitivity of the lung and spleen may also represent the cell types present in those tissues. For example, lung epithelial cells have been shown to respond to IFN-γ and be important regulators of inflammation in Mtb infection [61], so high levels of IFN-γ production may induce qualitatively different responses by cell types unique to the tissue. Identification of the specific mechanisms downstream of the IFN-γ receptor that exacerbate Mtb infection in the lung might provide novel targets for host-directed therapy that aim to limit pathology-associated IFN-γ responses during TB.
There is great interest in the role that exhaustion of Mtb-specific CD4 T cells plays in the inability of the host to control Mtb infection. Given its major role in virus and tumor-specific CD8 T cell exhaustion, it was initially hoped that PD-1 dependent inhibition could likewise be targeted in chronic Mtb infection to boost T cell function and enhance bacterial control. However, PD-1 KO mice succumb to CD4 T cell-mediated pathology in Mtb infection [31], and here we identify the specific CD4 T cell subset controlled by PD-1, and implicate IFN-γ as an effector molecule responsible for the immunopathology in the absence of PD-1. These data are cause for concern about therapeutic strategies that boost CD4 T cell function in Mtb infection by targeting inhibitory receptors, now referred to as checkpoint blockade, and at least one report has noted TB reactivation associated with PD-1 blockade treatment of Hodgkin’s lymphoma [62]. However, it should be noted that PD-1 pathway deficient mice also succumb to LCMV infection [63]. Both Mtb and LCMV infected mice succumb near the peak of T cell clonal expansion, the first week in the case of LCMV and approximately the 5th week for Mtb infection. In LCMV infection, the pathology is caused by CD8 T cell-mediated lysis of blood vascular endothelial cells [64], and in Mtb infection by IFN-γ–producing CD4 T cells. Although PD-1 blockade can be detrimental when peak numbers of activated effector T cells are present, blockade later in LCMV infection during the chronic phase results in enhanced viral control [63, 65, 66]. In the setting of Mtb infection, it has been found that PD-1 blockade at late stages of infection has no effect on bacterial loads and does not result in lethal pathology [23, 67]. It is not clear why PD-1 blockade in Mtb infection does not enhance immune responses, but it may be due to technical issues such as poor penetration of blocking antibodies into granuloma or compensation by additional inhibitory pathways that are upregulated as the infection progresses. Nonetheless, in both Mtb and LCMV infection the therapeutic window of checkpoint blockade opens after the peak of the T cell response when the risk of immunopathology decreases. It is possible that under antibiotic therapy PD-1 blockade during active TB may be safe. Therefore, further work is needed to determine if targeting checkpoint pathways could be developed as a safe and effective treatment strategy in Mtb infection. Perhaps other co-inhibitory receptors besides PD-1 that play less of a role in the control of IFN-γ production would be better targets. At the least, our data indicate that impaired IFN-γ production by exhausted Mtb-specific CD4 T cells is unlikely to explain the inability of the host to control Mtb infection. If CD4 T cell exhaustion does play a role in the poor control of Mtb infection, it may be due to the loss of other effector pathways besides IFN-γ.
While our data indicate that PD-1 inhibits the effector functions of Mtb-specific CD4 T cells, we also observed that PD-1 regulates the differentiation of CD4 T cells during Mtb infection. CD4 T cells in Mtb infected PD-1 KO mice displayed a dramatic decrease in KLRG1 expression and expressed lower levels of T-bet and CX3CR1 compared to WT CD4 T cells. Reduced KLRG1 expression and increased parenchymal localization of PD-1 KO CD4 T cells were also observed in our co-transfer experiments in WT recipient mice, indicating that the effect of PD-1 on CD4 T cell differentiation was intrinsic to the CD4 T cells and not secondary to elevated bacterial loads. We previously have shown that KLRG1+CX3CR1+T-betbright CD4 T cells are unable to migrate into Mtb infected lungs, and instead accumulate in the lung-associated blood vasculature. Collectively, our data suggest that in the absence of PD-1 dependent inhibitory signals, highly migratory CD4 T cells with a less differentiated phenotype are preferentially generated leading to an increase in lung parenchymal effector CD4 T cells, and once these cells get into the lung tissue, they over-produce IFN-γ. The role for PD-1 in promoting the generation of more differentiated Mtb-specific CD4 T cells was surprising, as it was expected that the increased signals into CD4 T cells in the absence of PD-1 dependent inhibition would enhance terminal differentiation. In fact, a recent report found that virus-specific PD-1 KO CD8 T cells in mice chronically infected with LCMV are driven into an even deeper state of terminal exhaustion [68]. Therefore, PD-1 plays a major role in regulating both the differentiation and effector functions of Mtb-specific CD4 T cells, but the precise role for PD-1 in the control of Mtb-specific effector CD4 T cell generation remains unclear.
Although several reports have shown that CD4 T cells can mediate IFN-γ–independent control of Mtb infection [32, 37, 69, 70], the relative contribution of these pathways in the overall control of Mtb infection was not quantified. Given the severity of infection in the absence of IFN-γ in animal models and the extreme susceptibility of humans with inborn errors of immunity in the IFN-γ axis to even environmental NTM infections, it is often stated that IFN-γ is the principal protective effector molecule produced by mycobacteria-specific CD4 T cells. Indeed, the host is extremely susceptible to early mortality during Mtb infection if CD4 T cells produce no IFN-γ. However, our data show that the mediators of the majority of bacterial control in the lung remain unknown, and indicate that identification of the pathways utilized by CD4 T cells to control pulmonary Mtb infection should be a top priority. The list of T cell effector molecules that have been shown to contribute to control of Mtb infection is not long. CD4 T cell-derived TNF has been shown to be important in host survival after intravenous [35] and aerosol [36] Mtb infection, but here we found that CD4 T cell-derived TNF plays a minor role in the lungs and spleens even compared to IFN-γ. It is possible that TNF in combination with IFN-γ mediates substantial CD4 T cell-dependent control, but future experiments such as we have described here are needed to directly quantify any potential synergy between these CD4 T cell-derived cytokines in control of Mtb infection. GM-CSF can also suppress the growth of Mtb in macrophages, and is another possible candidate, but GM-CSF derived from CD4 T cells has not been shown to contribute to protection to Mtb infection [71]. It seems likely that as yet unidentified effector pathways account for the largest part of T cell-dependent resistance to Mtb infection. Identification of novel anti-mycobacterial T cell effector molecules would provide mechanism based-correlates of protection to guide vaccine studies, and highlight potential targets for immunological interventions in TB that aim to selectively boost novel host protective pathways.
Studies were performed in accordance with recommendation of the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health (NIH). All experiments involving the use of animals were approved by the National Institute of Allergy and Infectious Diseases (NIAID) Animal Care and Use Committee.
C57BL/6 mice were purchased from Taconic Farms (Germantown, NY). B6.SJL (CD45.1) congenic, IFN-γ KO, TNF KO, TCRα KO and CD45.1 RAG1 KO mice were obtained through a supply contract between the NIAID/NIH and Taconic Farms. ARE-Del mice and have been described previously [33]. PD-1 KO mice [72] were crossed to IFN-γ KO mice in the NIAID animal facility. All animals (8–12 weeks old of sex matched) were housed at the Association for the Assessment and Accreditation of Laboratory Animal Care-approved BSL3 facility at the NIAID according to the National Research Council Guide for the Care and Use of Laboratory Animals. For Mtb infections, mice were exposed to ∼100 CFU of the H37Rv strain of Mtb with an inhalation exposure system (Glas-Col, LLC., Terre Haute, IN). Bacterial loads were measured in tissue homogenates by serial dilution on 7H11 agar plates supplemented with oleic acid-albumin-dextrose-catalase (Difco, Detroit, MI).
Mice were injected intravenously with 2.5 μg of fluorochrome-labeled anti-CD45.2 or anti-CD45 antibody, and after 3 min lungs were harvested [39]. Lungs were minced with scissors and then enzymatically digested for 45 min at 37°C in RPMI-1640 medium supplemented with 1 mg/ml Collagenase D (Roche-Diagnostics, Indianapolis, IN), 1 mg/ml hyaluronidase, 50 U/ml DNase I and 1 mM aminoguanidine (all from Sigma-Aldrich, St. Louis, MO). Digested lung was dispersed by passage through a 100 μm pore size cell-strainer and lung lymphocytes were enriched by 37% Percoll density centrifugation. For direct ex vivo IFN-γ staining, lungs were processed entirely in the presence of brefeldin A at a 1:1000 dilution (eBioscience, San Diego, CA).
For T cell stimulations, cells were incubated with 5 μg/ml ESAT-61–20 peptide for 5 h at 37°C in the presence of brefeldin A and 1 mM aminoguanidine. Cells were stained with various combinations of the following fluorochrome-labeled antibodies: anti-CD4 (RM4-4), CD44 (IM7), CD45 (IM7), CD45.1 (A20), CD45.2 (104), CD69 (H1.2F3), CXCR3 (CXCR3-173), CX3CR1 (polyclonal), Foxp3 (FJK-16s), ICOS (15F9), IFN-γ (XMG1.2), KLRG1 (2F1/KLRG1), PD-1 (29F.1A12), T-bet (eBio4B10), Thy1.1 (OX-7), and Fixable Viability Dye eFluor 780 purchased from BioLegend, eBioscience (San Diego, CA), BD Biosciences (San Jose, CA), and R&D Systems (Minneapolis, MN). I-Ab ESAT-64–17 and I-Ab EsxG46-61 MHC tetramers were produced by the NIAID Tetramer Core Facility (Emory University, Atlanta, GA). For staining with MHC class II tetramers, cells were incubated with tetramer at 1:50 dilution in complete medium containing 10% FCS, 1 mM aminoguanidine and monensin (eBioscience) at a 1:1000 dilution for 1 h at 37°C prior to staining with surface antibodies. All samples were acquired on an LSRFortessa flow cytometer (BD Biosciences) and analyzed with FlowJo software (Tree Star, Ashland, OR).
Mtb-infected WT or PD-1 KO mice were intravenously injected with PE-labeled anti-CD45 antibody on day 30 p.i. Single cell suspensions were prepared from lungs of ~15 mice and pooled. Cells were stained with antibodies specific to CD4 (YTS177.9, Novus Biologicals, Littleton, CO), CD44, KLRG1 and Fixable Viability Dye, and then live intravascular or parenchymal CD44hi CD4 T cells were sorted with FACSAria II (BD Biosciences) under the BSL3 condition. Greater than 97% purity of the sorted populations was achieved in all experiments. For survival experiments, ~4 × 104 cells of each population were adoptively transferred intravenously into TCRα KO mice that had been infected with Mtb 7 days prior to transfer. For adoptive transfer of naïve CD4 T cells, cells were isolated from lymph nodes and spleens of WT, IFN-γ KO, ARE-Del, TNF KO, PD-1 KO or PD-1/ IFN-γ double KO mice using MACS magnetic beads and columns (Miltenyi Biotec. Auburn, CA). Naïve CD4 T cell purity was consistently >90% as determined by flow cytometry. RAG1 KO or TCRα KO mice infected 7 days previously were reconstituted with total 3 × 106 cells of each indicated population. In additional experiments, MACS-isolated WT (CD45.1) or PD-1 KO (Thy1.1) naïve CD4 T cells were further purified to >97% purity by sorting, mixed at a 1:1 ratio and adoptively transferred (2 × 106 total CD4 T cells) into Mtb-infected CD45.2 WT mice.
Levels of cytokines in sterile filtered lung homogenates were determined by Mouse IFN-γ and TNF ELISA kit (eBioscience) according to the manufacturer’s instructions.
The statistical significance of difference between experimental groups was determined by Log-rank (Mantel-Cox) test, paired or unpaired Student’s t test using GraphPad Prism software (version 6). A P-value of <0.05 was considered significant.
For the determination of IFN-γ–dependent CD4 T cell-mediated control of Mtb growth in the tissues, CFU data were log10 transformed and area under the curve (AUC) for each CFU kinetic curve of unreconstituted mice, mice reconstituted with WT or IFN-γ KO CD4 T cells was calculated by the equations as shown in S1 Method. The following equation was used to calculate % CD4 T cell-mediated reduction in CFU that was IFN-γ–dependent:
ϵγ=AUCunreconstituted–AUCreconstitutedwithIFN−γKOCD4TcellsAUCunreconstituted–AUCreconstitutedwithWTCD4Tcells×100%
(1)
The calculations were performed for two independent experiments by averaging the log10 CFU per time point for each experiment. To calculate confidence intervals on the estimated efficacy we resampled CFU per time point for each experiment 10,000 times with replacement and re-calculated the efficacy using Eq (1). The statistical comparison between estimates of efficacy in the lung and spleen was done from 10,000 bootstrapped values [73].
|
10.1371/journal.ppat.1000432 | Genetic Dissection of an Exogenously Induced Biofilm in Laboratory and Clinical Isolates of E. coli | Microbial biofilms are a dominant feature of many human infections. However, developing effective strategies for controlling biofilms requires an understanding of the underlying biology well beyond what currently exists. Using a novel strategy, we have induced formation of a robust biofilm in Escherichia coli by utilizing an exogenous source of poly-N-acetylglucosamine (PNAG) polymer, a major virulence factor of many pathogens. Through microarray profiling of competitive selections, carried out in both transposon insertion and over-expression libraries, we have revealed the genetic basis of PNAG-based biofilm formation. Our observations reveal the dominance of electrostatic interactions between PNAG and surface structures such as lipopolysaccharides. We show that regulatory modulation of these surface structures has significant impact on biofilm formation behavior of the cell. Furthermore, the majority of clinical isolates which produced PNAG also showed the capacity to respond to the exogenously produced version of the polymer.
| Both in the wild and in the clinical setting many bacterial species live within surface-attached communities called biofilms. It is still unclear the extent to which the biofilm lifestyle and its associated phenotypes, such as hyper-tolerance to antimicrobial agents, can be attributed to structural characteristics of the biofilm community or to intrinsic biofilm-specific physiological programs. In order to address this longstanding question, we focused on poly-N-acetylglucosamine (PNAG)–based biofilms, a clinically relevant phenotype of many bacterial pathogens, including E. coli. Instead of working in a biofilm-permissive genetic background, in which the timescale of biofilm formation is slow, we applied the functionally active secreted version of the PNAG exo-polysaccharide (sPNAG) to wild-type E. coli cells, generating robust biofilms on the timescale of hours. In this way, we managed to uncouple upstream regulatory processes and matrix preparatory phase of biofilm formation, focusing specifically on the latter part. By using a powerful genome-wide technology, we identified the genes and pathways involved in sPNAG-based biofilm formation. Our results revealed that structural interactions between sPNAG and surface structures such as lipopolysaccharides are the crucial determinants of biofilm formation and that multiple pathways including acid-tolerance, capsule biosynthesis, and regulation of cell morphology modulate this phenotype.
| Biofilms are an integral component in the life-cycle of many microorganisms. Compared to their planktonic complement, however, bacterial biofilms have remained poorly understood, mostly due to the inherent complexities associated with biofilm studies, including spatial heterogeneity of the biofilm structure, longer generation time, and uncharacterized growth parameters [1].
Bacterial biofilms are characterized by the presence of an extracellular polymeric matrix, which encases the cells. The physicochemical properties of this matrix, including its charge, porosity, and architecture are prominent determinants of biofilm lifestyle. The matrix, for example, could act as a protective barrier by interacting with large, charged, or reactive biocidal agents and neutralizing them [1]. One major component of matrix in various bacterial species is a homopolymer of N-acetylglucosamine. In fact, poly-N-acetylglucosamine (PNAG) is the major virulence factor of Staphylococcus epidermidis [2]. There is increasing evidence that this polysaccharide is produced by a variety of other pathogens including Bordetella, Yersinia, Staphylococcus, Actinobacillai, and certain pathogenic Escherichia coli strains as well. It was reported that enzymatic hydrolysis of poly-N-acetylglucosamines disrupts biofilm formation by Yersinia pestis, Pseudomonas fluorescens, Aggregatibacter actinomycetemcomitans, pathogenic E. coli strains, and various Bordetella species [3]–[7]. This suggests that PNAG is a critical component of the biofilm structure made by all these bacteria. Furthermore, a recent study showed that most E. coli strains isolated from urinary tract and neonatal bloodstream infections possess the pga locus required for PNAG biosynthesis, and almost all of them produce immunologically detectable levels of PNAG [8].
Involvement of PNAG-based biofilms in the pathogenesis of various bacterial species makes it an important phenomenon to study [4], [8]–[11]. Even though the properties of PNAG-based biofilms have been extensively studied in Staphylococcus species [2],[12], the existence of a PNAG-based matrix in biofilm structures from other species, including E. coli, has been reported only recently [3],[13], and is not as well characterized as it is in Staphylococcus species. However, there are some features of PNAG-based biofilm like spatial distribution of the cells in PNAG-based biofilms that are better studied in E. coli [14]. In E. coli K-12, expression of PNAG biosynthesis genes is not high enough to support the formation of a robust biofilm structure under laboratory conditions [15], complicating the analysis of this phenotype. Therefore, in order to study the genetic basis of PNAG-based biofilm formation, we decided to enhance this phenotype in E. coli K-12 by either increasing the level of endogenous PNAG or providing an exogenously produced form of PNAG.
PNAG production in E. coli can be enhanced by manipulation of the genetic elements involved in pga locus regulation [16],[17]. For example, E. coli csrA mutants overproduce the PNAG polymer [13],[17]. However, csrA is a master regulator of the carbon storage system, and csrA mutants show highly pleiotropic phenotypes. An alternative approach for enhancing PNAG-based biofilm formation would be to use a functionally active exogenous source of PNAG to induce biofilm formation in E. coli. In the csrA mutant background, cells secrete PNAG polymer into the growth media [18]. Therefore, the spent media from the csrA mutant culture can be used as a potential source of PNAG for inducing biofilm formation.
We observed that application of exogenously produced PNAG, isolated from ΔcsrA cells saturated culture of csrA mutants, led to robust biofilm formation in the E. coli K-12 background. This provided us a unique opportunity to extensively characterize the underlying genetics of the PNAG-based biofilm formation phenomenon. Given our observations, we favor a model in which electrostatic interactions between this polysaccharide and cell surface structures, such as lipopolysaccharide (LPS), are critical for PNAG-induced biofilm formation by E. coli. We also show that although response to PNAG polymer is a purely structural phenomenon, it can be modulated by multiple pathways including LPS biosynthesis, the acid tolerance system, capsule biosynthesis, and regulation of cell morphology.
Escherichia coli csrA mutants secrete poly N-acetylglucosamine polysaccharide into the culture medium [18]. In order to see whether the secreted polysaccharide is functional, we grew ΔcsrA cells to stationary phase, discarded the cells, and added the cell-free spent media to wild-type MG1655 cells in the presence of fresh media. Interestingly, we observed that the cell-free spent media of saturated ΔcsrA cultures made wild-type cells form a biofilm on a rapid time-scale (Figure 1, compare i and ii). Presence of a carbon source, as expected, was required for formation of mature visible microcolonies by living wild-type cells (Figure 1, compare iv and vii to v and vi). Synthesis of PNAG in E. coli requires the gene products of the pgaABCD operon [13],[17]. To test whether the observed biofilm-inducing activity is associated with the secreted PNAG (sPNAG), we generated four double mutants each harboring csrA deletion together with the deletion of one of the four genes present in the pga locus. There was no detectable biofilm-inducing activity in the cultures of any of the four double mutants. Furthermore, the biofilm-inducing activity was lost after treatment of ΔcsrA culture spent media with Dispersin B, an enzyme that specifically cleaves PNAG [3],[5],[19], confirming that the biofilm inducing factor is an N-acetylglucosamine containing polysaccharide. Differential up-regulation of the pga locus transcription between the ΔcsrA and wild-type cells was also confirmed by a reporter assay (Figure S1).
In order to confirm that the observed biofilm formation phenotype is specific to sPNAG and rule out the possibility of involvement of other biofilm-inducing agents that might be present in the ΔcsrA spent media, we purified sPNAG from spent media. As part of the purification steps, sPNAG was treated with various enzymes, including DNase, RNase, α-amylase, and Proteinase (see Materials and Methods section). The purified sample showed identical biofilm-inducing activity, suggesting that sPNAG is sufficient for inducing biofilm formation. The purified polysaccharide was also characterized by mass spectrometry. As shown in Figure 2, almost all prominent molecules found on the mass spectrum corresponded to N-acetylglucosamine oligomers with different levels of acetylation or to their monomers. In S. epidermidis, deacetylation of PNAG polymer introduces positive charges in the otherwise neutral polymer [20]. Our results also indicate that a considerable fraction of sPNAG building units is deacetylated, which should leave a net positive charge on the polymer. Purified PNAG isolated from Staphylococcus aureus strain MN8m [21] showed similar biofilm-inducing activity when applied to wild-type MG1655 cells, further confirming that PNAG is sufficient for the observed biofilm-inducing activity. The presence of various identical peaks in the mass spectra of PNAG from E. coli and S. aureus (Figure S2) indicates that they are closely related molecular species.
The response of wild-type E. coli cells to sPNAG was so fast that we decided to study the early stages of the process. Using time-lapse microscopy (Video S1), we observed that wild-type cells started seeding microcolony structures on a glass slide in less than an hour of exposure to sPNAG. The pace of microcolony formation observed here was much faster than the previously reported behavior by csrA mutants [15]. The microcolonies expanded in size due to both growth of pre-existing cells and continued incorporation of new cells. No similar activity was observed in the absence of sPNAG (Video S2). These results show that sPNAG enhances both cell-cell and cell-surface interactions. SEM images of the biofilm structures formed by wild-type cells in the presence of sPNAG also confirmed the presence of an extracellular matrix encasing the cells (Figure S3).
We used a microarray-based genetic footprinting strategy [22] to study the genetic basis of sPNAG-mediated biofilm formation. Since formation of biofilm by wild-type cells in the presence of sPNAG is robust and fast, the phenotype is highly amenable to genetic analysis. Starting from a close to saturation Tn5-based library of approximately 5×105 independent transposon insertional mutants, generated in wild-type E. coli MG1655, we devised a selection strategy to enrich for mutants defective in responding to sPNAG. Roughly 1010 cells of the abovementioned library were exposed to sPNAG in LB medium. After 12 hours of incubation, cells which were present in the liquid phase of the culture and were not part of the biofilm were isolated, grown up to log phase, and transferred to a new container with fresh media and fresh sPNAG, in order to enrich for mutants impaired in responding to sPNAG. After four rounds of serial enrichment, no visible biofilm formation activity was present in the enriched population. A schematic representation of the enrichment procedure is shown in Figure 3.
To quantify the contribution of different loci to this impaired biofilm formation phenotype, the insertion sites of the transposon in the enriched population were mapped using a microarray-based approach [22]. The histogram in Figure 4A shows the normalized average output of the hybridization data from two experimental replicates. The z-score for each ORF is indicative of the abundance of transposon insertion events in that ORF (or its vicinity) in the enriched population of mutants. More detailed information regarding the calculation of z-score is provided in the supporting information section (Dataset S1). Interestingly, the majority of the genes that were highly enriched in our selection were involved in two major biological processes: LPS core biosynthesis and regulation of cell shape and morphology. Most of these candidate genes belong to two long operons (Figure 4B). In other words, genetic perturbations caused by transposon insertion in many components of LPS biosynthesis or cell shape regulation made wild-type cells lose their ability to form a biofilm in the presence of sPNAG. The dominance of genes involved in the synthesis and regulation of exposed structural components suggested that physical interaction between sPNAG and these surface structures may be a major determinant of biofilm formation capacity. sPNAG pre-treatment of the cells, however, did not cause any change in the migration of their extracted LPS samples on SDS-PAGE gels (Figure S4).
Incomplete disruption of the targeted genes and polar effects are characteristics of transposon insertion events. In order to get around these complications and get a fine-scale perspective of genetic perturbations that prevent cells from responding to sPNAG, we generated in-frame deletions of some of the candidate ORFs, obtained from our transposon mutagenesis screen, in the MG1655 background [23],[24]. We then studied the behavior of these mutant strains in the presence of sPNAG (Table S2). Among the candidate genes, deletion of rfaY, rafP, or rfaQ, diminished sPNAG-based biofilm formation (Figure 1, compare i and iii). As shown in Figure 4C, the product of these three genes are directly or indirectly involved in addition of phosphate groups to the inner core of LPS [25].
Since the major common point in the LPS structure of rfaY, rfaP, and rfaQ mutants is the lower density of negative charge (phosphate groups) on their LPS outer core, these phosphate groups are likely to be critical for this interaction. Given the positive charge of sPNAG, we favor a model in which electrostatic interaction between the positively charged polysaccharide and the negatively charged phosphate groups on LPS is the major determinant of sPNAG-mediated biofilm formation. Electrostatic interactions were also proposed to be responsible for PNAG-based biofilm formation in S. epidermidis [20].
It is postulated that neighboring LPS molecules can be cross-linked by divalent cations due to the presence of phosphate groups in the LPS structure [25]. Therefore, any presumable electrostatic interaction between phosphate groups and sPNAG should be sensitive to increasing concentrations of divalent cations. As shown in Table S3, response to sPNAG is lost in Ca2+ concentrations higher than 100 µM, which could be considered as an additional support for our electrostatic interaction model. However, changing calcium concentration might also change cell viability. Exposure to this concentration of calcium, however, did not have any effect on the viability of the cells, as measured by viable counting and CFU determination. Identical results were obtained when the experiment was repeated with other divalent cations (manganese or magnesium).
Based on the microarray, transposon insertions in genes involved in regulation of cell shape and morphology should also interfere with sPNAG-base biofilm formation. However, since all the genes in this category are essential, we could not introduce those deletions into the wild-type background and check their phenotype. Transposon insertion events in this operon presumably occurred in either regulatory regions (e.g. promoter) or dispensable parts of essential genes.
As could be inferred from the data, there was no significant correlation between capacity of producing PNAG and responding to it. However, in order to confirm this, we studied sPNAG-induced biofilm formation phenotype of all four ΔcsrA Δpga double-mutants (ΔcsrA pgaA::kan, ΔcsrA pgaB::kan, ΔcsrA pgaC::kan, and ΔcsrA pgaD::kan) and also all four Δpga single-mutants and found it to be indistinguishable from that of wild-type cells.
We were curious to know whether deletion of rfaY, rfaQ, or rfaP abolished the response to sPNAG due to downstream signaling events or the phenotype was a simple consequence of the structural modifications imposed on LPS. Therefore, we decided to systematically identify extra-genic suppressors which can restore biofilm formation capacity of ΔrfaY cells and test whether there are any known or putative signaling pathway components among such suppressors.
To identify suppressors of rfaY deletion, a transposon insertion library was generated in the ΔrfaY background (i.e. a strain with clean deletion of rfaY ORF), and enriched for double mutants that recovered their ability to form a biofilm in the presence of sPNAG (opposite to what was demonstrated in Figure 3). A glass slide was provided as the biofilm formation surface and at the end of each round, the slide was transferred to a new container with sPNAG and fresh media. After four rounds of enrichment, macroscopic microcolony structures could be detected on the glass slide. The transposon insertion sites in the enriched population were mapped by the same footprinting strategy described previously (Dataset S2). Surprisingly, transposon insertions in many of the LPS biosynthetic genes were significantly enriched in this selection (Figure 5A)
To validate our microarray predictions, we generated in-frame deletions of the candidate genes in the ΔrfaY background and studied their behavior in the presence of sPNAG (Table S4). We found that deletion of rfaC, rfaF, rfaI, pgm, galU, rfaH, rfbD, rfbC, and adiY reverted the phenotype of ΔrfaY cells. rfaC, rfaF, rfaI, pgm, rfbC, rfbD, and galU are involved in the synthesis of LPS core structure or its precursors (Figure 5B). rfbC and rfbD are involved in rhamnose biosynthesis which is a component of the second major LPS glycoform in E. coli K12 [26], which is not shown in Figure 5B. rfaH is a transcription anti-terminator which is required for full-length transcription of long operons, including rfaQ-K operon [27]. adiY is the positive regulator of the arginine decarboxylase system and will be discussed later. Overall, these mutants make truncated versions of LPS, an expectation we verified for a subset of them (Figure 5C). We hypothesized that in these truncated structures, inner phosphate groups of lipid A or possibly other negatively charged cell-surface moieties (which tend to be buried by longer LPS chains in the wild-type cells) are now more exposed and available for interaction with sPNAG. Deleting any of the four genes in the pga operon did not restore the biofilm formation capacity of the ΔrfaY cells, as was also inferred from the microarray data.
We also used fluorescence microscopy to characterize the dynamics of biofilm formation in a heterogeneous population composed of cells either capable (represented by ΔrfaY ΔrfaF cells) or defective (represented by ΔrfaY cells) in biofilm formation. To this end, ΔrfaY cells, expressing RFP fluorescent marker, were competed against ΔrfaY ΔrfaF cells expressing GFP with the starting ratio of 1∶1 for making a biofilm on a glass slide in the presence of sPNAG. After 12 hours, the biofilm structure formed on the glass slide was visualized by fluorescence microscopy. As shown in Figure 5D, top row, microcolonies in biofilm structure were mostly formed by ΔrfaY ΔrfaF double mutants (i.e. GFP expressing cells). The same result was obtained by swapping the fluorescent labels (Figure 5D, bottom row).
Transposon insertions typically lead to a loss of function phenotype. In order to complement our transposon insertion based approach, we used an over-expression library in the ΔrfaY background. This library contained ∼2.5×105 independent mutants each carrying a 1–3 kb long genomic fragment of E. coli cloned into the pBR322 plasmid. The over-expression library was enriched for mutants responding to sPNAG, similar to the approach used for studying the ΔrfaY transposon insertion library. After four rounds of enrichment, the over-expressed fragments represented in the enriched population were identified by microarray hybridization (Dataset S3). As expected, when comparing the results from ΔrfaY transposon insertion library with ΔrfaY over-expression library, LPS biosynthesis genes showed the opposite behavior (Figure 6A). Many LPS biosynthetic genes like rfaG, rfaJ, rfaI, rfaP, kdtA, rfbX, rfaZ, and rfaL were found to be among the top 10% highly enriched category in the ΔrfaY background transposon insertion library while in the over-expression library, they belonged to the top 10% most depleted group.
The microarray results also revealed that mutants over-expressing certain genes associated with the acid tolerance system in E. coli were abundant in the enriched population. In order to clarify how this system contributes to biofilm formation, we further characterized the phenotypic consequences of over-expressing these genes in individual cells. To this end, we isolated individual clones from the enriched population. One of the isolated mutants was found to have the genomic region corresponding to three genes, gadW, gadY, and gadX. GadX and GadW are dual regulators of the glutamate-dependent decarboxylase acid-resistance system of E. coli. [28]. GadY is a small RNA which acts as a positive regulator of gadX [29].
Our previous results demonstrated that LPS modification was a dominant mechanism in regulating response to sPNAG. Therefore, we first investigated the effect of gadXYW over-expression on LPS structure. As shown in Figure 6B, ΔrfaY pBR322-gadWYX strain contained some smaller LPS variants as compared to the parental ΔrfaY cell. This suggests that the reversion of ΔrfaY phenotype upon over-expression of gadWYX gene cluster is a consequence of this truncated LPS structure. To test whether this change in LPS structure was due to a transcriptional regulatory event, we measured the transcription of the rfaQ-K operon in both pBR322-gadWYX and pBR322 (empty vector) backgrounds by a β-galactosidase assay and found it to be almost 3-fold lower in the gad-over-expressing cells (Figure 6C). Evidence regarding the existence of a cross-talk between acid tolerance system and LPS regulation has been observed before, and gadE, the transcriptional regulator of the acid resistance system in E. coli, was reported to be a potential activator of the rfaQ-K operon [30]. We also found that over-expression of gadY alone was sufficient to cause the phenotype, although not as strongly as the over-expression of gadXYW.
A strain harboring the pBR322-argR plasmid was also isolated from the enriched ΔrfaY library. However ArgR, the negative regulator of arginine biosynthesis system, acts as a weak suppressor of ΔrfaY biofilm formation deficiency. Putting all these observations together, four of the suppressors found in the transposon insertion and over-expression libraries, adiY, gadX, gadW, and argR, were directly or indirectly associated with the amino acid decarboxylase systems, involved in acid tolerance in E. coli. The biological function of these suppressors (Table S5) and the distribution of acid tolerance genes in the over-expression library (Figure S5) suggest that down-regulation of the acid stress response, or more specifically the amino acid decarboxylase systems, positively contribute to biofilm formation in the ΔrfaY background, presumably due to the changes imposed on LPS.
Overall, our observations support the existence of a physical interaction between sPNAG and LPS. As such, biofilm formation in the presence of sPNAG may be a purely structural phenomenon, occurring as a simple consequence of passive interactions between sPNAG and LPS. Based on this model, even dead cells with intact outer membrane structure should still be capable of responding to sPNAG. To test this, we killed wild-type cells by either UV-irradiation or exposing them to formaldehyde, and visualized their behavior upon exposure to sPNAG. Time-lapse microscopy showed that these dead cells start nucleating microcolony structures, similar to living cells (Video S3). All together, our observations argue that sPNAG-mediated biofilm formation can be considered as a two-step process, starting with the nucleation event which is a purely structural phenomenon, followed by microcolony expansion and maturation which is a growth-dependent process (Figure 1, compare iv and vii to v and vi).
Most natural and clinical isolates of E. coli produce different serotype-specific surface structures including O-antigen and capsular polysaccharide, also known as K-antigen, which are absent in E. coli MG1655 [31],[32]. We were curious to know how variations in composition of these surface antigens might affect the response to sPNAG. Therefore, we chose 11 strains, which were reported to be competent of endogenous PNAG production [8], for further analysis. Among these strains, 7 formed biofilms in the presence of sPNAG (Table S6), and two of the latter group were also O-antigen− (Figure S6). In case of K-antigen, we focused on K1 capsule, a homopolymer of α-(2-8)-linked polysialic acid [33], which is the predominant capsule found in a major subset of these clinical isolates. Three K1+ isolates used in this study were also capable of responding to sPNAG (Table S6).
Natural and clinical isolates of E. coli possess uncharacterized surface structures other than K1, e.g. fimbriae and other capsular polysaccharides. These structures could significantly affect the physicochemical properties of the cell surface. Consequently, cell response to sPNAG in these clinical isolates could not be solely judged based on their O-antigen structure or presence of K1 capsule. Furthermore, the limited number of strains tested in this study and their non-isogenic background make these observations preliminary and they should be followed up with future studies. Therefore, we decided to study the role of O- and K-antigen in the well-characterized K-12 background.
We generated all possible combinations for presence or absence of O16 antigen and K1/K92 capsule in the K-12 background. K92 is a polysialic acid capsule very similar to K1, and its biosynthetic gene cluster can be transferred on a plasmid. As shown in Table 1, only non-capsulated O16+ cells were impaired in sPNAG-based biofilm formation. In the presence of both O16 and K1/K92 antigen, however, cells were capable of responding to sPNAG, which is not surprising considering that capsule is a more exposed surface structure than O-antigen [32]. Since K1 and K92 capsules confer a high density of negative charge to the E. coli cell surface, their presence could contribute to establishing any potential electrostatic interaction with sPNAG. These data suggest that loss of O-antigen (O16) or presence of K1 capsule is associated with sPNAG-induced biofilm formation in the E. coli K-12. However, in order to confirm the involvement of these structures in sPNAG-induced biofilm formation, targeted genetic experiments together with more careful characterization of their role in physiochemical properties of outer membrane are required.
Biofilms can afford protection from a variety of environmental challenges including phagocytosis, extreme pH, and antibiotic exposure [1]. We were curious to characterize some of these biofilm-specific features in sPNAG-based biofilms. In order to investigate their antibiotic tolerance, we challenged the cells with two different antibiotics: ampicillin and polymyxin B. As shown in Figure 7, cells in the context of sPNAG biofilm showed ∼10 fold higher tolerance to polymyxin B compared to planktonic cells, whereas no significant difference was observed in tolerance to ampicillin. Polymyxin B resembles antimicrobial peptides, an integral component of the innate defense system in many organisms, in terms of both structure and mechanism of action [34]. Considering sPNAG as a positively charged matrix encasing the cells, electrostatic repulsion between this polysaccharide and polymyxin B could potentially protect the bacteria by reducing the local concentration of the drug in the vicinity of the cells. However, we still consider the involvement of other as yet unknown sPNAG-induced physiological response in this phenomenon.
Considering other biofilm-specific phenotypes, we found no significant change in the rate of F-plasmid conjugation in sPNAG-based biofilms (Figure S7). We also did not observe any evidence supporting the existence of a phase-variable mechanism regulating sPNAG production in E. coli (Figure S8 and Figure S9).
We were also curious whether sPNAG can induce biofilm formation in species other than E. coli, because this could potentially facilitate interspecies interactions, e.g. conjugation, between E. coli and other microorganisms. Therefore, we exposed Salmonella typhimurium LT2 cells to sPNAG similar to what was done for E. coli, but we did not observe any detectable biofilm formation. We reasoned that if sPNAG-mediated biofilm formation has a structural basis, then a mutant Salmonella with modified surface characteristics may be capable of forming a biofilm in the presence of sPNAG. To test this hypothesis, we applied sPNAG to a transposon insertion library of S. typhimurium LT2 with approximately 105 mutants and enriched for mutants that form biofilms. After four rounds of enrichment, macroscopic microcolony structures were formed by cells. We mapped the transposon insertion sites in 4 of the mutants capable of forming biofilms in a sPNAG-dependent pattern. In all cases, the transposon mapped to different positions in the rfaK gene, involved in LPS biosynthesis. Salmonella rfaK mutants are lacking most of the O-antigen structure [35], providing further support, beyond E. coli, for our physical interaction model between sPNAG and LPS.
Poly-N-acetylglucosamine, the major virulence factor of Staphylococcus epidermidis, has recently been found in many other pathogenic bacteria [7], including E. coli, but PNAG-based biofilms in these pathogens are poorly characterized relative to Staphylococcus species [2]. Here, we carried out a systematic genetic analysis of poly-N-acetylglucosamine-induced biofilm formation in E. coli. However, instead of working in a biofilm-permissive genetic background, in which the time scale of biofilm formation is slow, we applied the functionally active secreted version of PNAG (sPNAG) to wild-type E. coli MG1655 cells and observed rapid and reproducible biofilm formation. The fast kinetics and robustness of sPNAG-induced biofilm formation phenomenon allowed us to comprehensively characterize its underlying genetic basis. Our observations support the notion that electrostatic interaction between positively charged sPNAG and different cell surface antigens with negative charge is responsible for the formation of the biofilm structure. This is consistent with the generally accepted intuition that physicochemical properties of the matrix, including its charge, geometry, ion selectivity, and pore size contribute significantly to biofilm formation [1],[18].
The composition and physicochemical properties of cell surface structures can be modulated by multiple biological pathways and environmental factors. Therefore, although response to sPNAG seems to be the consequence of simple electrostatic interactions, it can be regulated by a complex interplay between LPS structural dynamics, the presence of serotype-specific capsular polysaccharide, the acid tolerance system, and cell morphology as shown in this work. There are also other relevant pathways, not fully active in our laboratory strain, that might contribute to this phenotype, such as addition of positively charged 4-amino-4-deoxy-L-arabinose to lipid A involved in polymyxin B and other cationic antimicrobial peptide tolerance in E. coli and S. typhimurium. Interestingly, a polymyxin B resistant mutant isolated from our transposon insertion library was also impaired in responding to sPNAG (data not shown). Since resistance to polymyxin is usually concomitant with higher density of positive charge on the outer membrane [36], polymyxin resistant mutants are expected to be defective in interacting with sPNAG. Given that antimicrobial peptides generated by the host immune system are a major challenge for the survival of pathogens, encasement in a positively charged matrix or biofilm serves as a protective mechanism against antimicrobial peptides for sensitive cells, while resistant cells can survive without it.
Since in this study, selections were performed with pools of mutants rather than with clonal populations, phenotype of different mutants should be interpreted as a spectrum of different capacities for forming a biofilm rather than a simple binary phenomenon. The results of our selection, carried out on the over-expression library in the ΔrfaY background, strongly support this notion. ΔrfaY cells are defective in responding to sPNAG, so it is reasonable to assume that only a handful of genes in the over-expression library should be capable of suppressing its phenotype and the majority of the library mutants should be equally defective in biofilm formation. However, the fact that a considerable fraction of genes involved in LPS biosynthesis were significantly depleted in that selection (Figure 6A) indicates that even ΔrfaY cells might have residual biofilm formation activity and that over-expression of some LPS biosynthetic genes might reduce this partial activity. The weak biofilm formation activity of ΔrfaY cells could be due to their entrapment in the microcolony structures formed by other mutants in the population. This is in part shown in Figure 5D in which ΔrfaY cells are mostly colocalized with microcolonies rather than being uniformly distributed. Imposed modifications in LPS structures caused by over-expression of some LPS biosynthetic genes [37] might interfere with this partial activity.
In this study, we have demonstrated the involvement of different surface structures and regulatory systems in sPNAG-mediated biofilm formation. However, there might exist a myriad of other uncharacterized regulatory systems, surface structures, and environmental factors that may influence this phenotype through changes in the physicochemical properties of the cell surface. Covalent modifications of lipid A with phosphoethanolamine in a Ca2+-dependent manner, or with 4-amino-4-deoxy-L-arabinose in response to Mg2+ and pH, represent just a few examples [38]. Therefore, even if a natural isolate of E. coli is incapable of responding to sPNAG in the laboratory environment, it may show a different phenotype in its natural environment. There is also the possibility of higher level cooperation between different cell types in the population, with one subpopulation producing sPNAG while another is responding to it. The producer subpopulation may even be defective in the initiation of the biofilm formation process, but could be incorporated into preformed structures. Examples of population heterogeneity in biofilm communities has been reported before [1]. Furthermore, it is likely that different cells produce different versions of the PNAG polymer, with different degrees of acetylation [39] or different polymer length distributions, which could make response to sPNAG more strain-specific. A better understanding of sPNAG-based biofilms acquired from studies like this could be useful for development of new therapeutic strategies against pathogens that use this polysaccharide as a virulence factor. The systematic framework presented in this study, along with the acquired insights, should also benefit the study of microbial biofilms formed by other species.
All strains used in this study are listed in Table S7, bacteriophages and plasmids are mentioned in Table S8. All the experiments were performed in LB (1% tryptone, 0.5% yeast extract, 1% NaCl), supplemented as required with the following antibiotics: ampicillin, 50 µg/ml; tetracycline, 25 µg/ml; spectinomycin, 100 µg/ml; chloramphenicol, 30 µg/ml and kanamycin, 100 µg/ml, unless otherwise mentioned. β-galactosidase measurements were performed in triplicate as described before [40].
Transposon mutagenesis and microarray based genetic footprinting were carried out as described before [22]. Chromosomal deletions were created using the previously described method [23] and transferred by generalized transduction with P1 phage as required. lacZ reporter strain was generated using the plasmid pCE37 [41]. The over-expression plasmid library construct was a kind gift of Joseph Sklar. Each plasmid contained a 1–3 kb long fragment of E. coli MC4100 genome (average fragment size of 2 kb) cloned into the β-lactamase coding sequence of the pBR322 vector. The plasmid pool was electroporated into E. coli MG1655 ΔrfaY cells, leading to ∼2.5×105 independent over-expressing mutants. For PCR amplification of DNA constructs for cloning, Pfu Ultra polymerase (http://www.stratagene.com) was used, whereas ExTaq DNA polymerase (http://www.takara-bio.com) was used for all other PCR reactions. Restriction endonucleases and T4 ligase were obtained from New England Biolabs (http://www.neb.com). DNA purification kits were obtained from QIAGEN (http://www1.qiagen.com). Primer sequences are available upon request.
Structural analysis of LPS samples was performed as described before [42]. LPS was purified from cell lysates after phenol-ether extraction, separated on 14% tricine-SDS-PAGE gel and visualized after silver staining.
All fluorescence and light microscopy experiments were performed using Zeiss AxioVision 4.5. Time-lapse microscopy was performed in the FC81 flow cell apparatus from Biosurface Technologies Corp (in the absence of flow). SEM analysis was performed using a Philips XL30 Field Emission Scanning Electron Microscope.
For the sake of consistency, a large batch of ΔcsrA spent media, grown in LB, was prepared and used as the source of sPNAG for all the experiments throughout this study. The amount of sPNAG present in this stock solution, was estimated by measuring its hexosamine content using 3-methyl-2-benzothiazolone hydrazone hydrochloride (MBTH) method [43] and found to be ∼0.175 mg equivalent of hexosamine per ml of saturated ΔcsrA culture supernatant. The quantity of sPNAG in each milliliter of this stock solution was defined to be 1 arbitrary unit (U). Control experiments were also carried out using sPNAG from different preparations of ΔcsrA spent media and the results were reproducible.
Spent media from 1 liter of saturated ΔcsrA culture in LB was passed through a 0.22 µ filter (http://www.nalgene.com/). The cell-free supernatant was concentrated 200 fold by Centriplus YM-100 columns (http://www.millipore.com). The concentrated sample was treated with DNase I (2 mg), RNase A (10 mg), and α-amylase (20 mg) and incubated for 2 hours at room temperature followed by 2 hours at 37°C. Next, sample was treated with Proteinase K (20 mg) for 1 hour at 37°C and 1 hour at 55°C. Enzymes and other proteins were removed by pre-warmed (55°C) phenol: ether extraction, followed by ethanol precipitation. The precipitate was re-suspended in water and fractionated on a S-300 Sephacryl column. The fractions with biofilm-inducing activity were pooled together and concentrated by Centriplus YM-100 columns. The sample was treated with Dispersin B (20 µg) for 4 hours at 37°C. After removing the enzyme by phenol: chloroform extraction, the sample was dialyzed with Spectra/Por cellulose ester membranes with MWCO = 500 (http://www.SPECTRUMLABS.com/) to remove the salt present in Dispersin buffer. Finally, the sample was analyzed by an ESI-LTQ Orbitrap Hybrid mass spectrometer from Thermo Fisher Scientific.
For all biofilm formation assays, cells and sPNAG were added to fresh LB in such a way as to obtain ∼5×108 cells and 0.1 U of sPNAG per milliliter of the mixture. Biofilm structures were studied or transferred (in case of serial enrichments) 12 hours after exposure to sPNAG at 25°C.
Plasmids from ∼109 cells of the enriched population (or in case of references, from the maximally diverse unselected library) were extracted using QIAGEN plasmid miniprep kit. The extracted plasmid pool was amplified with two separate PCR reactions, using primer pair pBR_Lib_T7_L and pBR_Lib_R or pBR_Lib_T7_R and pBR_Lib_L. The sequences of these primers are as follows:
The incorporated T7 promoter in pBR_Lib_T7_L and pBR_Lib_T7_R is underlined. Cycling conditions for PCR were 94°C for 2 min; 30 cycles of 94°C for 30 s, 55°C for 30 s, and 72°C for 5 min; and 72°C for 10 min, using ExTaq DNA polymerase. The PCR products from these two reactions were pooled together. The T7 promoter incorporated into the pBR_Lib_T7_L and pBR_Lib_T7_R primer sequences was used to generate RNA from the pooled PCR product in an in vitro transcription reaction using Ambion MEGAscript T7 Kit (http://www.ambion.com). Finally the RNA from the previous reaction was reverse transcribed to cDNA, using Cy3-labeled nucleotides with Invitorgen SupreScript II Reverse Transcriptase (http://www.invitrogen.com). The fluorescently labeled cDNA was used for microarray hybridization versus Cy5-labeled fragmented (nebulized) MG1655 genomic DNA. The hybridization data of the two maximally diverse unselected libraries were used as the reference.
Production of PNAG in S. epidermidis could be discriminated on congo red indicator plates [44]. In order to extend this to E. coli, congo red indicator plates (3% tryptic soy broth, 1% glucose, 0.08% congo red, and 1.5% agar) were supplemented with 200 mM NaCl which was reported to enhance pga locus transcription in E. coli [16]. Different dilutions of stationary phase cultures of ΔcsrA mutants were plated on these plates and incubated for two days at 37°C. PNAG-producing ΔcsrA cells formed dark brown colonies whereas ΔcsrA mutants defective in producing PNAG were red. To map the location of the mutation which abolished PNAG production, red colonies were picked individually and transduced with a P1 phage lysate obtained from the maximally diverse transposon insertion library. Kanamycin resistant transductants which recovered their PNAG production were identified by replica plating on congo red plates. Dark colonies were picked, and the location of transposon insertion, which should have been linked to the mutation site, was mapped using the same footprinting strategy. The exact location of the mutation was determined after PCR amplification of the candidate genomic locations and subsequent sequencing of the PCR product.
|
10.1371/journal.pgen.1002537 | Parkinson's Disease–Associated Kinase PINK1 Regulates Miro Protein Level and Axonal Transport of Mitochondria | Mutations in Pten-induced kinase 1 (PINK1) are linked to early-onset familial Parkinson's disease (FPD). PINK1 has previously been implicated in mitochondrial fission/fusion dynamics, quality control, and electron transport chain function. However, it is not clear how these processes are interconnected and whether they are sufficient to explain all aspects of PINK1 pathogenesis. Here we show that PINK1 also controls mitochondrial motility. In Drosophila, downregulation of dMiro or other components of the mitochondrial transport machinery rescued dPINK1 mutant phenotypes in the muscle and dopaminergic (DA) neurons, whereas dMiro overexpression alone caused DA neuron loss. dMiro protein level was increased in dPINK1 mutant but decreased in dPINK1 or dParkin overexpression conditions. In Drosophila larval motor neurons, overexpression of dPINK1 inhibited axonal mitochondria transport in both anterograde and retrograde directions, whereas dPINK1 knockdown promoted anterograde transport. In HeLa cells, overexpressed hPINK1 worked together with hParkin, another FPD gene, to regulate the ubiquitination and degradation of hMiro1 and hMiro2, apparently in a Ser-156 phosphorylation-independent manner. Also in HeLa cells, loss of hMiro promoted the perinuclear clustering of mitochondria and facilitated autophagy of damaged mitochondria, effects previously associated with activation of the PINK1/Parkin pathway. These newly identified functions of PINK1/Parkin and Miro in mitochondrial transport and mitophagy contribute to our understanding of the complex interplays in mitochondrial quality control that are critically involved in PD pathogenesis, and they may explain the peripheral neuropathy symptoms seen in some PD patients carrying particular PINK1 or Parkin mutations. Moreover, the different effects of loss of PINK1 function on Miro protein level in Drosophila and mouse cells may offer one explanation of the distinct phenotypic manifestations of PINK1 mutants in these two species.
| Parkinson's disease (PD) is the second most common neurodegenerative disease. It mainly affects movement in elderly people and was traditionally considered a sporadic disease with no known cause. Discoveries of genes associated with familial PD (FPD) have demonstrated that PD pathogenesis can be significantly influenced by an individual's genetic makeup. Understanding the functions of these FPD genes will allow better understanding of the sporadic PD cases. PINK1 and Parkin are genes associated with FPD that affect patients at an early age. Mutations in PINK1 and Parkin lead to the accumulation of damaged mitochondria, the powerhouse of the cell, as a result of impairments of the mitochondrial quality control system. However, the mechanism of PINK1/Parkin action remains poorly understood. Here we show that PINK1 and Parkin act together to regulate Miro, a key component of the mitochondrial transport machinery, and that altered activities of PINK1 cause aberrant mitochondrial transport. Regulation of mitochondrial transport may be a critical aspect of the mechanisms by which the PINK1/Parkin pathway governs mitochondrial quality control. Dysfunction of this process could contribute to the loss of DA neurons, the cardinal feature of PD, as well as the peripheral neuropathy symptom associated with particular PINK1 or Parkin mutations.
| PD is a neurodegenerative disorder characterized by the dysfunction and loss of dopaminergic (DA) neurons in the substantia nigra, although neurons in other brain regions are affected as well. Mutations in PINK1 and Parkin are linked to familial forms of early-onset PD [1], [2]. PINK1 encodes a Ser/Thr kinase with a mitochondrial targeting sequence, whereas Parkin encodes an E3 ubiquitin ligase. Studies in Drosophila first revealed that PINK1 and Parkin act in a common pathway to impact mitochondrial function and DA neuron maintenance [3]–[6], in part through the regulation of mitochondrial fission/fusion dynamics [7]–[11]. At least in primary cultured mammalian hippocampal neurons and DA neurons, PINK1 and Parkin have been shown to exert similar effects on mitochondrial dynamics as seen in Drosophila DA neurons [7], [12]. PINK1 and Parkin are also implicated in mitochondrial quality control [13]. Decreased mitochondrial membrane potential stabilizes the normally labile PINK1, which recruits Parkin to damaged mitochondria, leading to ubiquitination of mitochondrial proteins and marking damaged mitochondria for removal by autophagy [14]. Both mitochondrial fission/fusion dynamics and autophagy are considered important aspects of the mitochondrial quality control mechanism that mediates PINK1/Parkin function in DA neuron maintenance [11], [15]–[17].
In some PINK1- or Parkin-linked PD patients, symptoms of peripheral neuropathy were also reported [18]–[20]. It is not clear whether this is caused by defects in the aforementioned functions or some other unknown function of PINK1/Parkin. Peripheral neuropathy is a clinical term used to describe various forms of damages to nerves of the peripheral nervous system by distinct mechanisms [21]. Many types of peripheral neuropathy are dependent on the length of neuronal axon, with neurons carrying long axons frequently affected. It is hypothesized that this is caused by defects in the axonal transport of key proteins and/or organelles such as mitochondria, which are critical for maintaining the axonal and synaptic physiology of those extremely polarized neurons [22]. This notion has gained significant support from recent studies of the inherited forms of peripheral neuropathies [22], [23]. Defective mitochondrial transport has also been considered a pathogenic event in other neurodegenerative diseases [22], [24], [25], including rodent models of PD [26], [27]. In primary cultured rat hippocampal neurons, overexpression of PINK1 has been shown to inhibit the lateral movement of photoactivated, mito-Dendra2-labelled mitochondria [12], raising the possibility that defects in the axonal transport of mitochondria may actively participate in PINK1-related PD pathogenesis.
A major aspect of axonal transport is mediated by motor proteins that travel on axonal microtubules, which are polarized and uniformly orientated, with their plus-ends pointing towards nerve terminals. The kinesin and dynein motors are involved in microtubule plus-end (anterograde) and minus-end directed (retrograde) transport, respectively [28]. Mitochondria are mainly produced in neuronal cell body and delivered to sites where metabolic demand is high, such as the synapses and nodes of Ranvier [29]. The functions of Mitochondrial Rho (Miro), a mitochondrial outer membrane GTPase [30], and the cytosolic protein Milton are critical for mitochondrial transport, as they serve to link mitochondria with kinesin motors and the microtubule cytoskeleton [31], [32]. In Drosophila Miro or Milton mutants, mitochondria accumulate in neuronal soma and fail to move into the axons [31], [32]. In cultured mammalian cells, overexpression of a constitutively active mutant of Miro was shown to induce cell death, suggesting that mitochondrial transport or some other aspect of Miro function is important for cell survival [30]. Whether this is relevant to in vivo conditions such as neurodegenerative disease settings is not known.
The fruit fly Drosophila melanogaster has served as an excellent model for studying neurodegenerative diseases [10]. It was in Drosophila that the in vivo function of PINK1 was first revealed [3]–[6]. PINK1 mutant flies exhibit abnormal wing postures, reduced flight ability and thoracic ATP level, degeneration of indirect flight muscle and DA neurons, and male sterility, which are caused by the accumulation of dysfunctional mitochondria, thus suggesting a role of PINK1 in mitochondrial function and/or quality control [3]–[6]. Further genetic studies in Drosophila have also uncovered important functions of PINK1 in regulating mitochondrial morphology and electron transport chain activity [7]–[9], [33].
The power of the Drosophila neurodegenerative disease models lies in the ability to facilitate unbiased genetic modifier screens to identify new players involved in the disease process. Using this approach, we show in this study that PINK1 genetically interacts with the mitochondrial transport machinery. Reduction of function in Miro, Milton, or kinesin heavy chain effectively rescued the PINK1 mutant phenotypes. On the other hand, overexpression (OE) of Miro led to the formation of enlarged mitochondria and resulted in DA neuron loss, thus phenocopying PINK1 mutants. By monitoring mitochondrial movement in live Drosophila larval motor neurons, which possesses long axons and could serve as a model system for studying peripheral neuropathy, we provide evidence that PINK1 directly regulates mitochondrial transport. The function of PINK1 in mitochondrial transport may contribute to PD pathogenesis in DA neurons and underlie the peripheral neuropathy symptoms associated with certain PINK1 mutations in some PD patients.
Our biochemical analysis demonstrated that overexpressed PINK1 in cooperation with Parkin could regulate Miro protein ubiquitination and stability, which might contribute to the regulatory effect of PINK1 on mitochondrial motility. While our paper was under review, it was suggested that PINK1 phosphorylates Miro at a conserved S156 residue, and that this phosphorylation event is required to activate proteasomal degradation of Miro in a Parkin-dependent manner [34]. However, our in vitro kinase assay using active recombinant PINK1 failed to show direct phosphorylation of Miro by PINK1. Moreover, a mutant form of hMiro1 with the S156 site mutated to Ala was equally susceptible to PINK1/Parkin-mediated degradation in HeLa cells. Thus, the exact molecular mechanism by which the PINK1/Parkin pathway regulates Miro protein level will require further investigation.
By taking advantage of the easily identifiable phenotype of abnormal wing posture induced by dPINK1 inactivation, we performed a genetic screen for modifiers of PINK1. The scheme was similar as described before [35]. In this screen, we identified components of the mitochondrial transport machinery as genetic modifier of PINK1. Knockdown of Miro, Milton or Kinesin heavy chain (Khc) each rescued the muscle phenotypes in PINK1B9 null mutant, including abnormal wing posture, decreased fly ability and ATP depletion (Figure 1A–1C). Conversely, overexpression (OE) of Miro and Khc enhanced such phenotypes (Figure 1A–1C). These results demonstrated strong genetic interaction between PINK1 and the mitochondrial transport machinery as a whole, supporting that mitochondrial transport is the underlying mechanism mediating their genetic interaction. In this study, we will focus our analysis on Miro, as a previous study in cultured cells suggested that mammalian Miro might physically interact with PINK1 [36].
To test the relevance of the functional interaction between PINK1 and the mitochondrial transport machinery to PD pathogenesis, we examined their interaction in DA neurons, the disease-relevant cell type. As in the muscle, Miro-RNAi effectively rescued PINK1 mutant phenotypes in DA neurons, both in terms of mitochondrial aggregation (Figure 2A–2D) and DA neuron loss (Figure 2F). Moreover, Miro-OE alone, driven by the TH-Gal4 driver, caused aberrant mitochondrial aggregation (2E, 2G) and DA neuron loss (Figure 2F), thus phenocopying PINK1 loss-of-function effects (Figure 2B, 2F, 2G), although the Miro-OE effect was noticeably stronger than PINK1 mutant. It is worth noting that TH-Gal4-driven Miro-OE in PINK1 mutant background resulted in dramatically reduced viability (data not shown), although the surviving adults did not show further DA neuron loss than that induced by Miro-OE alone (Figure 2F). Together, these results demonstrate that PINK1 and Miro also exhibit strong genetic interaction in DA neurons, with decreased dMiro level/activity ameliorating the detrimental effects caused by the loss of dPINK1, whereas increased dMiro level or activity phenocopying dPINK1 mutants.
The strong genetic interaction between PINK1 and Miro raised the interesting possibility that PINK1 might directly regulate mitochondrial transport, the impairment of which might contribute to PINK1-related parkinsonism. This was further supported by the DA neuron loss induced by Miro-OE alone, which presumably acted by altering mitochondrial transport. To test this idea, we examined the effect of PINK1 on mitochondrial movement in Drosophila larval motor neurons, a system amenable to live imaging of mitochondrial transport. Mitochondrially-targeted GFP (mitoGFP) expressed specifically in motor neurons was used to track mitochondrial movement via live imaging in anesthetized third instar larvae (Figure 3A), using well-established procedures [37], [38]. To highlight the mitochondria undergoing active transport, a 61.5 µm-long segment of motor neuron was photobleached and the movement of fluorescently labeled mitochondria moving into the bleached area from both directions was recorded at 1 frame/2 s for 300 s (Figure 3B, Videos S1, S2, S3, S4, S5, S6, S7). From these videos, mitochondrial flux (the normalized number of mitochondria that passes certain point over time), mitochondrial net velocity (the normalized mitochondrial net displacement over time), and mitochondrial morphology (e.g. mitochondrial length) in different genetic backgrounds were analyzed. In general, mitochondrial net velocity is controlled primarily by the intrinsic properties and quantities of motor proteins associated with the mitochondria [38], while mitochondrial flux can also be significantly affected by mitochondrial morphology, as changes in mitochondrial morphological features such as length can increase or decrease the number of motile mitochondria.
We found that PINK1-OE decreased mitochondrial flux as well as net velocity in both anterograde and retrograde directions, similar to the effect of Miro-RNAi, although the PINK1-OE effect appeared to be slightly weaker (Figure 3B–3D; Videos S1, S2, S4). In contrast, PINK1-RNAi and Miro-OE both increased the net velocity of anterograde mitochondrial transport, with retrograde transport largely unaffected (Figure 3B, 3C; Videos S3, S5). PINK1-RNAi also increased anterograde mitochondrial flux (Figure 3C), while mitochondrial flux in Miro-OE background was reduced in both anterograde and retrograde directions (Figure 3C). The reduction of mitochondrial flux by Miro-OE could be partially explained by the formation of very long mitochondria in Miro-OE motor neurons (Figure 3E, 3F), as previously observed [38].
In addition to mitochondrial motility, PINK1 also affected mitochondrial length in motor neurons. PINK1-RNAi increased mitochondrial length in the axons of larval motor neurons as in Miro-OE case, although the effect of Miro-OE was much stronger. Conversely, PINK1-OE and Miro-RNAi both decreased mitochondrial length (Figure 3E, 3F). To address whether PINK1-induced mitochondrial motility change was due to its effect on mitochondrial length, we examined mitochondrial transport in genetic backgrounds where mitochondrial fusion/fission machinery was directly manipulated to alter mitochondrial length. Increasing mitochondrial fission by overexpression of the fission protein Fis1 or knockdown of the fusion protein Marf led to decreased mitochondrial length (Figure 3E, 3F), similar to the effects of Miro-RNAi or PINK1-OE. However, in contrast to the decreased mitochondrial flux and net velocity as observed in the Miro-RNAi or PINK1-OE backgrounds, Fis1-OE and Marf-RNAi both increased mitochondrial flux and net velocity in anterograde and retrograde directions (Figure 3B–3D; Videos S6, S7), suggesting that mitochondrial length and transport kinetics are not always directly correlated, which is consistent with a previous report [38]. Collectively, these results support the notion that PINK1 regulates mitochondrial transport and that its effect on mitochondrial motility is direct, rather than a secondary effect of mitochondrial length change.
In addition to mitochondrial motility, we also examined mitochondrial distribution at motor neuron nerve terminals at the larval neuromuscular junction (NMJ), which can be used as an indirect measure of mitochondrial motility. Consistent with a previous report [31], Miro-OE led to the accumulation of mitochondria in the most distal boutons, which is likely the consequence of net anterograde transport (Figure 4A, 4B). PINK1 knockdown showed similar effect (Figure 4A, 4B). Thus, the dysfunctional mitochondria in PINK1 mutant might gain longer retention time in the distal segment of motor neuron axons where synapses are formed. This finding may have clinical implications for PINK1 pathogenesis. In contrast, Miro-RNAi and PINK1-OE both led to decreased accumulation of mitochondria in the most distal boutons (Figure 4A, 4B). We conclude that PINK1 regulates mitochondrial distribution in motor neuron nerve terminals, likely through its effect on mitochondrial transport.
Our results so far showed that PINK1 and Miro exert opposite effects on mitochondrial morphology, motility and distribution. We next explored the biochemical mechanisms underlying their negative genetic relationship. We first used HeLa cells to test whether Miro protein level might be regulated by PINK1 and possibly Parkin, which tends to work together with PINK1 in a common pathway [3]–[5], [39]. There are two Miro homologues in human cells, hMiro1 and hMiro2, that are ∼60% identical [30]. Overexpression of either hPINK1 or hParkin did not lead to obvious change of exogenous hMiro1 protein level under normal conditions, but a modest reduction of hMiro1 level was observed when hPINK1 and hParkin were co-expressed (Figure 5A, lane 5). A decline in mitochondrial membrane-potential induced by the mitochondrial uncoupler carbonyl cyanide m-chlorophenylhydrazone (CCCP) was reported to activate the PINK1/Parkin pathway [13], [14], [39]. Under CCCP treatment condition, hPINK1 or hParkin each significantly stimulated hMiro1 ubiquitination (Figure 5A). Since HeLa cells express very little endogenous Parkin [13], the effect of hPINK1 alone on hMiro1 ubiquitination (Figure 5A, lane 8) suggested that other E3 ligase(s) might be recruited by hPINK1 to ubiquitinate hMiro1. However, this ubiquitination event did not appear to lead to destabilization of hMiro1 (Figure 5A, IB: Myc). In contrast, coexpression of hPINK1 and hParkin dramatically reduced hMiro1 level in the presence of CCCP (Figure 5A, lanes 9). Importantly, pathogenic mutations in hPINK1 or hParkin abolished this effect (Figure 5B, lanes 4 and 5 compared with lane 3, and lanes 9–11 compared with lane 8), indicating that functional hPINK1 and hParkin are both required in the destabilization of hMiro1. Previously, many outer mitochondrial membrane (OMM) proteins were shown to be degraded by the ubiquitin proteasome system (UPS) pathway in a PINK1/Parkin-dependent reaction at an early step of mitophagy, while other OMM proteins might be eliminated by subsequent autophagosome-dependent events [40], [41]. Thus, direct or indirect substrates of PINK1/Parkin could be distinguished by their degradation kinetics [40]. In our experiments, hMiro1 was more rapidly degraded than another OMM protein VDAC1 (Figure 5B, VDAC1), a reported Parkin substrate involved in mitophagy [14], supporting that hMiro1 is a direct substrate of Parkin.
Similar to hMiro1, hMiro2 could also be ubiquitinated by PINK1 and Parkin co-expression or after CCCP treatment. However, the degradation of hMiro2 was at a much slower rate compared to hMiro1 (Figure S1), consistent with a previous result [40]. Furthermore, like exogenous hMiro1, endogenous hMiro1 was also rapidly degraded by PINK1/Parkin overexpression in HeLa cells and its level was dramatically reduced within 15 min of CCCP treatment. The degradation of endogenous hMiro2 was again at a much slower rate than that of hMiro1 (Figure 5C, 5D).
We next examined the effect of the PINK1/Parkin pathway on Miro protein level in an in vivo setting. Similar to the results in HeLa cells, Drosophila dMiro protein level was decreased in the brain extracts of PINK1 or Parkin overexpression adult flies (Figure 5E). Conversely, dMiro level was increased in PINK1B9 mutant brain extracts (Figure 5E). These results are consistent with dPINK1 negatively regulating dMiro protein level in vivo. It is worth noting that different from the effects seen in HeLa cells, overexpression of PINK1 or Parkin alone was sufficient to reduce dMiro level in adult Drosophila brain, and the co-expression of PINK1 and Parkin did not lead to much further reduction of dMiro level than PINK1-OE alone, suggesting that the endogenous levels or activities of PINK1 and Parkin are already sufficient to support each other's action in the Drosophila brain.
Removal of damaged or dysfunctional mitochondria through mitophagy could be one mechanism by which the PINK1/Parkin pathway maintains mitochondrial health, at least under some conditions, and the accumulation of those abnormal mitochondria in PINK1 mutants could be the underlying cause of disease pathogenesis. Consistent with this notion, it was previously shown that enhancing autophagy could efficiently rescue dPINK1 mutant phenotypes [35]. To better understand the rescuing effect of Miro-RNAi in PINK1 mutant background, we examined the effect of Miro knockdown on mitophagy. We used CCCP treatment to induce mitochondrial damage in HeLa cells stably transfected with venus-Parkin, and subsequently monitored the removal of damaged mitochondria over time by examining the protein levels of mitochondrial markers on the inner/outer membrane or in the matrix and inter-membrane space. Simultaneous knockdown of hMiro1 and hMiro2 significantly accelerated the mitochondrial removal process, with all the mitochondrial markers disappearing faster in hMiro knockdown cells than in the control siRNA-treated cells (Figure 6A). This suggested that there was more active mitophagy after hMiro knockdown. To confirm this result, we monitored the mitochondrial network by immunofluorescence staining. Compared to the control siRNA-treated cells, knockdown of either hMiro1 or hMiro2 led to the accumulation of mitochondria in the perinuclear region, and knockdown of both hMiro1 and hMiro2 further enhanced this effect (Figure 6B). Fluorescence from the immunostaining of Tom20 (an OMM marker) but not HSP60 (a matrix marker) in hMiro1 and hMiro2 double knockdown cells was noticeably weaker than that in control siRNA treated cells at 3 h after CCCP treatment (Figure 6C), supporting the notion that hMiro knockdown facilitated an early event in Parkin-mediated mitophagy.
We further investigated the molecular mechanisms by which the PINK1/Parkin pathway regulates Miro protein level or stability. While our paper was under review, a report showed that PINK1 phosphorylates Miro at a conserved S156 residue, and that this phosphorylation activates proteasomal degradation of Miro in a Parkin-dependent manner [34]. However, repeated in vitro kinase assays using an active GST-dPINK1 recombinant protein capable of efficient autophosphorylation [42] failed to show phosphorylation of GST-dMiroΔTM, a GST fusion protein of full-length dMiro with the transmembrane domain deleted (Figure 7A). Drosophila or mammalian PINK1 protein affinity purified from HEK293 cells by immunoprecipitation also failed to phosphorylate GST-dMiroΔTM in our assays (data not shown).
To further probe the significance of S156 phosphorylation in facilitating the proteasomal degradation of Miro promoted by the PINK1/Parkin pathway, we introduced S156A mutations into hMiro1 or hMiro2 and examined the stability of the mutant proteins in HeLa cells co-transfected with PINK1 and Parkin, under normal or CCCP treatment conditions. As shown in Figure 7B, the wild type and S156A mutant forms of hMiro1 were equally susceptible to PINK1/Parkin- mediated degradation under both conditions. The wild type and S156A mutant forms of hMiro2 behaved similarly as well (data not shown). Thus, the PINK1/Parkin pathway may regulate Miro protein level independent of S156 site phosphorylation under these experimental conditions in HeLa cells.
We also examined the effect of loss of PINK1 function on Miro protein level in mammalian cells. For this purpose, we used both HeLa cells with PINK1 knockdown and MEF cells derived from PINK1 (−/−) knockout mice. Surprisingly, unlike the situation in Drosophila, endogenous Miro1 or Miro2 protein levels were significantly reduced in PINK1 RNAi cells under normal or CCCP treatment conditions (Figure 7C, PINK1 RNAi in HeLa cells or Venus-Parkin stably transfected HeLa cells) and in PINK1 (−/−) MEF cells (Figure 7D). The introduction of Venus-Parkin resulted in CCCP/PINK1-dependent degradation of Miro1 (Figure 7C, Control RNAi in Venus-Parkin transfected HeLa cells). Thus, loss of PINK1 function in mammalian cells can lead to reduced expression of Miro1 and Miro2 proteins, presumably through mechanisms distinct from that operating under PINK1/Parkin co-overexpression condition.
Mitochondrial dysfunction has long been implicated in the pathogenesis of PD. However, the exact mechanisms by which mitochondrial dysfunction arises in the disease process and how cells, particularly neurons, handle dysfunctional mitochondrial are not well understood. The identification of a mitochondrial quality control system involving two FPD genes, PINK1 and Parkin, has provided a much-needed point of entry to elucidate the role of mitochondria in the pathogenesis of PD. Here we showed that PINK1 directly regulates mitochondrial transport and that it affects the stability and/or activity of Miro, a mitochondrial Rho GTPase with a well-establish function in mitochondrial transport. Our conclusion is supported by the following evidence: 1) dMiro protein level is negatively regulated by PINK1 and Parkin in vivo in Drosophila; 2) Overexpressed PINK1 and Parkin act together to promote the ubiquitination and degradation of hMiro1 in HeLa cells; 3) Reduction of the activities of Miro or other components of the mitochondrial transport machinery effectively rescued dPINK1 mutant phenotypes. 4) Overexpression of dMiro in DA neurons phenocopied dPINK1 loss-of-function effects; 5) Manipulation of dPINK1 activity produced clear mitochondrial motility phenotypes opposite to that observed for dMiro manipulation in Drosophila larval motor neurons. Together, these results support that the mitochondrial transport defects caused by PINK1 inactivation represent one of the key pathogenic events that contribute to PD pathogenesis in the Drosophila model.
Neurons are highly polarized cells that rely heavily on axonal transport to distribute to axons and synapses critical proteins and organelles synthesized in the cell body, thereby maintaining neuronal function and health. Defects in axonal transport are often linked to diseases affecting peripheral neurons that tend to extend very long axons [21], [22], [29]. Although the symptoms of PD patients mainly arise from the loss of DA neurons, some PD patients carrying particular PINK1 and Parkin mutations developed peripheral neuropathy with unknown cause [18]–[20]. Our results showing the PINK1/Parkin pathway playing a critical role in regulating mitochondrial transport offers one potential explanation of the peripheral neuropathy symptoms observed in these PINK1 or Parkin-linked PD cases. It would be interesting to examine whether Miro protein level or activity is affected by these particular mutations in human cells. Moreover, we propose that defects in PINK1/Parkin-regulated mitochondrial transport may offer one explanation of the selective vulnerability of DA neurons observed in PD patients and animal models. DA neurons that make elaborate and long projections may be particularly vulnerable to impairment of the mitochondrial transport system.
Our results offer new insights into the mode of action of the PINK1/Parkin pathway in mitochondria quality control. We showed that, in Drosophila models, PINK1 OE led to decreased mitochondrial flux and net velocity, as observed in Miro knockdown background. In addition, we found that Miro knockdown could facilitate an early step of mitophagy in mammalian cells. These observations, together with the finding that the normally labile PINK1 protein is stabilized on damaged mitochondria [39], suggest a scenario whereby the accumulation of PINK1 on damaged mitochondria and the subsequent turnover of Miro could exert neuroprotection by (1) preventing damaged mitochondria from being anterogradely transported along the axons, thus increasing their chance of getting eliminated in the soma; and (2) promoting elimination of damaged mitochondria through mitophagy. This potentially explains the normal protective function of PINK1. When PINK1 function is impaired, however, on one hand mitochondria become dysfunctional as evidenced by morphology changes and impaired electron transport chain function [3]–[5], [33], [43]–[45], on the other hand, the anterograde mitochondrial transport is enhanced as shown in this study in Drosophila models. As a result, the dysfunctional mitochondria would have increased retention in the axons and synapses, resulting in increased reactive oxygen species (ROS) production, oxidative damage, and subsequent synaptic and axonal degeneration and eventual neuronal loss, at least in the Drosophila models. Many details of this model await further experimental validation. For example, it has been suggested that the reported effect of PINK1/Parkin on mitochondrial autophagy may not operate in the same manner in primary neurons as compared to cultured non-neuronal cells [46]. It also remains to be determined whether the effects of Miro on mitochondrial transport and mitophagy reflect a functional antagonism between these two processes, or two distinct functions of Miro in neuronal maintenance. In this respect, it is worth noting that the effect of Miro overexpression on cell survival in Drosophila is cell type-dependent: it causes DA neuron loss but has no obvious effect on muscle integrity (data not shown). It is possible that different tissues may have different sensitivities to impairments of Miro function. For example, muscle cells may be less susceptible to mitochondrial transport defects than neurons.
One interesting difference we observed between Drosophila and mouse systems was that although activation of the PINK1/Parkin pathway led to reduced Miro protein level in both systems, the loss of PINK1 in Drosophila resulted in increased steady-state Miro protein level, whereas its loss in mammalian cells as in PINK1 (−/−) MEF cells or PINK1 RNAi HeLa cells had the opposite effect. The mechanism of Miro downregulation in PINK1 loss-of-function mammalian cells is currently unknown, but it is presumably different from that used by activation of the PINK1/Parkin pathway. Since the upregulation of dMiro in dPINK1 mutant background is likely causal to DA neuron degeneration, as indicated by the rescue of DA neuron loss in dPINK1 mutant by dMiro-RNAi and the induction of DA neuron loss by dMiro-OE alone, it is tempting to speculate that the downregulation of Miro levels in PINK1 (−/−) mouse, as opposed to the dMiro upregulation in Drosophila PINK1 mutant, might contribute to the lack of DA neuron degeneration phenotype in the mouse PINK1 models [47]–[50]. Testing this hypothesis will require in vivo studies boosting Miro expression levels in wild type and PINK1 (−/−) mouse.
Our results also provide new insights into the process by which the PINK1/Parkin pathway promotes mitophagy. Previous studies suggested that upon recruitment to damaged mitochondria, Parkin activates the ubiquitin proteasome system to effect wide-spread degradation of OMM proteins in an autophagy-independent manner, and it was further proposed that this remodeling of OMM is important for a subsequent step of mitophagy [40]. The previously identified Parkin substrates, Mfn1 and Mfn2, although important for the effect of Parkin on mitochondrial fission/fusion dynamics, are not necessary for Parkin-induced mitophagy [40], [51]. Here we show that removal of Miro by the PINK1/Parkin pathway, in a presumably autophagy-independent but ubiquitination-dependent manner, facilitated mitophagy. Interestingly, knockdown of mammalian Miro itself promotes the formation of ring-like or round-shaped mitochondrial morphology, which is often observed in depolarized, mitophagy-ready mitochondria (Figure 6B; [13]). It is possible that the removal of Miro from OMM exposes certain recognition signals for the autophagy machinery, or that Miro/Milton/Kinesin-mediated mitochondrial transport may normally antagonize the mitophagy process. Supporting the latter scenario, an interaction between the BECLIN 1-interacting protein AMBRA1 and the dynein motor complex has been implicated in mammalian autophagy [52]. It also remains to be understood at the mechanistic level how PINK1 cooperates with Parkin to promote the ubiquitination and degradation of Miro. One attractive hypothesis is that PINK1 may directly phosphorylate Miro to promote its subsequent ubiquitination and degradation by Parkin, as suggested by a recent study [34]. However, our biochemical data have so far failed to support this hypothesis. It is possible that the divergent results are due to the different cell lines used or other experimental conditions. Alternatively, PINK1 may directly act on Parkin to promote Parkin's mitochondrial recruitment or activity in activating ubiquitin proteasome system-mediated ubiquitination and degradation of Miro. Further studies are needed to elucidate the molecular mechanisms of PINK1/Parkin action.
Finally, it is worth mentioning that studies in Drosophila models have identified a number of genetic modifiers of PINK1/Parkin [35], [53], [54]. While some of these genetic modifier genes may be directly related to the seemingly diverse biological activities of the PINK1/Parkin pathway, possibly mediated by distinct PINK1/Parkin substrates, others may reflect cellular compensatory responses to cope with the mitochondrial dysfunction caused by PINK1/Parkin inactivation. The fact that manipulations of each of these different cellular processes exert clear functional rescue of PINK1/Parkin mutant phenotypes suggests that there exists a signaling network linking the diverse activities of PINK1/Parkin in mitochondria biology with the nuclear-encoded cellular responses to mitochondrial dysfunction, and that many key players in this network represent novel and rational therapeutic targets.
Flies were raised according to standard procedures at indicated temperatures. Sources of fly strains and other reagents are as follows: dPINK1B9: Dr. J. Chung [3]; UAS-dMiro and anti-dMiro antibody: Dr. K. Zinsmaier [31]; TH-GAL4, UAS-PINK1, UAS-PINK1 RNAi and rabbit anti-Drosophila TH antibody: described before [5]; UAS-Miro RNAi106683: Vienna Drosophila RNAi Center; UAS-Miro RNAi27695, UAS-Milton RNAi28385 and UAS-Khc RNAi25898: Harvard Transgenic RNAi Project (TRiP) and Bloomington Drosophila Stock Center; all other fly lines: Bloomington Drosophila Stock Center; FLAG-hParkin mutants and HA-ubiquitin: Drs. N. Matsuda, K. Tanaka and S. Hatakeyama; Myc-hMiro1 and Myc-hMiro2 plasmids: Dr. P. Aspenström [30]; hPINK1 cDNAs were cloned into pcDNA3-FLAG vector.
Antibodies used in this study are as follows: anti-RHOT1/Miro1 (4H4, Abnova), anti-RHOT2/Miro2 (Protein technology Group), anti-PINK1 (Novus), anti-Parkin (PRK8, Santa Cruz Biotechnology), anti-Tom20 (FL-145, Santa Cruz Biotechnology), anti-VDAC1 (Abcam), anti-OXPHOS Complex IV subunit I/COX I (Invitrogen), anti-Tim23 (BD), anti-NDUFA9 (Invitrogen), anti-Cytochrome c (BD), anti-HtrA2/Omi (as described in [55]), anti-Hsp60 (BD), anti-PDHA1 (Abcam), anti-HA (3F10, Roche), anti-Myc (4A6, Millipore; #2272, Cell Signaling Technology), anti-β-actin (AC-15, Sigma-Aldrich), anti-α-tubulin (DM1A, Millipore), Peroxidase anti-Guinea Pig IgG antibody (Jackson ImmunoResearch), Texas Red-conjugated anti-HRP (Jackson ImmunoResearch), Alexa Fluor 488 nm-conjugated goat anti-chicken IgG (Invitrogen) and Alexa Fluor 594 nm-conjugated goat anti-rabbit IgG (Invitrogen).
These assays were carried out essentially as described before [35]. The thoracic ATP level was measured using a luciferase based bioluminescence assay (ATP Bioluminescence Assay Kit HS II, Roche applied science) as described [35].
Whole-mount brain immunohistochemistry for TH and mitoGFP was performed as described previously [35]. For DA neuron mitochondrial morphology analysis, mitoGFP was expressed in Drosophila DA neurons using the TH-Gal4 driver. Brains from 3-day-old adult flies of the indicated genotypes were immunostained with the anti-TH antibody to label DA neuron and anti-GFP antibody to label mitochondria. For measurement of mitochondrial size distribution, the size of each mitochondrial aggregate was represented by the length of its longest axis. The percentage of DA neurons in the PPL1 cluster that have one or more mitochondria exceeding the indicated size was shown. Six flies of each genotype were used for the analysis.
The motor neuron-specific OK6-Gal4 driver was used to express UAS-mitoGFP in the larval segment neurons, and 3rd instar male larvae raised at 29°C were used for live imaging. Larvae were briefly washed with water and anesthetized for 4 min in 1.5 ml eppendorf tubes containing 4 µl suprane (Baxter International Inc.), before being placed ventral side up into a small chamber. The chamber was created on glass slide with double-sided tape and cover glass. Additional suprane (4 µl) was introduced into the chamber before it was sealed with Valap (1∶1∶1 amount of vaseline, lanolin, parafin wax). Mitochondria were viewed with an upright Leica DM6000 B microscope equipped with a laser scanner and a 63× oil-immersion objective. Larvae were positioned to have their ventral ganglion (VG) appearing on the right of the acquired image and segmental nerves aligned horizontally across the image. Before recording mitochondrial movement, a centered region of 1024×200 pixel (61.5 µm×12.0 µm, 4× digital zoom) close to the VG was photobleached for 30 s with 488 nm excitation argon laser set at 80% output power. The viewing field was then zoomed out (2.5× digital zoom) and mitochondrial movement was immediately recorded by time-lapse video (2 s/frame, 300 s total) in a region of 1024×150 pixel (98.4 µm×14.4 µm) with the laser power reduced to 10% of the maximum output. The pinhole was set at 200 µm for all the experiments. Time-lapse images were acquired within 30 min of anesthetization. Mitochondrial flux was calculated by normalizing the number of mitochondria that passes certain point within the time frame examined. Mitochondrial net velocity was calculated by normalizing mitochondrial net displacement with time. At least 5 larvae of each genotype were analyzed.
HeLa cells were maintained at 37°C with 5% CO2 atmosphere in DMEM (Wako) supplemented with 10% FCS (GIBCO) and non-essential amino acids (Invitrogen). Plasmids and siRNA duplexes (Invitrogen) were transfected using Lipofectamine 2000 (Invitrogen) and Lipofectamine RNAiMAX (Invitrogen), respectively, according to manufacturer's instructions. To depolarize the mitochondria, HeLa cells were treated with 10 µM CCCP (Sigma-Aldrich) at 36 hr (for plasmids) or 72 hr (for siRNA) post-transfection.
The S156A mutations were introduced into human Miro1 and Miro2 by PCR-based mutagenesis using the following primers: For hMiro1, Forward primer: 5′- GCA gag ctcttttatt acgcac -3′; Reverse primer: 5′- tatg ttcttcaggt ttttcgc -3′. For hMiro2, Forward primer: 5′- GCA gagct gttctactac gc -3′; Reverse primer: 5′- ga tgttcctcag gttcttggc -3′. Full-length sequences of hMiro1 S156A and hMiro2 S156A were confirmed by sequencing.
For brain extract preparation, 6 fly heads were quickly dissected and homogenized in 60 µl SDS-PAGE sample buffer. 5 µl brain extracts from each genotype were loaded onto SDS-PAGE gel. Guinea pig anti-dMiro antibody (1∶20000, from Dr. K. Zinsmaier) and Peroxidase Anti-Guinea Pig IgG antibody (1∶10000, Jackson ImmunoResearch Labs) were used for Western Blot. For HeLa cell-based biochemical analysis, cells were lysed in 1% Triton X-100 -based lysis buffer (10 mM Tris-HCl [pH 7.6], 120 mM NaCl, 5 mM EDTA, 1% Triton X-100 and protease inhibitor [Nacalai Tesuque]). Immunoprecipitation was performed using Immunoprecipitation Kit-Dynabeads Protein G (Invitrogen) according to manufacturer's instructions.
In vitro kinase assay was performed essentially as described [42], using a 2× GST-dPINK1 fusion protein with GST fused at both the N- and C-terminus of dPINK1 as the kinase source and a GST-dMiroΔTM fusion protein as the substrate. GST-dMiroΔTM covers amino acids 1–634 of the full-length dMiro protein. The GST-dMiro-ΔTM plasmid was constructed by amplifying a Myc-tagged dMiro fragment without the transmembrane domain from a pUAST-Myc-dMiro plasmid (obtained from Dr. K. Zinsmaier) using 5′-CGCCCG-GGTGAGCAGAAACTCATCTCTGAAGAAG-3′ and 5′-ATGCGGCCGCTACTTGGG-GTCCTCCGTC-ATC-3′ as primers. The amplified fragment was inserted into the SmaI and NotI cloning sites of the pGEX-6P-1 vector. Recombinant GST fusion proteins were purified from bacteria according to standard protocols.
Cells were fixed with 4% paraformaldehyde in PBS and permeabilized with 0.2% Triton X-100 (for mitophagy in Figure 6C) or 0.5% Triton X-100 (for mitochondrial morphology in Figure 6B) in PBS. Cells stained with the appropriated antibodies and counterstained with DAPI were imaged using a laser-scanning microscope (LMS510 META; Carl Zeiss, Inc.) with a Plan-Apochromat 63×NA1.4 or 100×/1.4 Oil differential interference contrast objective lens. Image contrast and brightness were adjusted in Image Browser (Carl Zeiss, Inc.)
Two-tailed Student's t tests were used for statistical analysis. p values of <0.05, <0.01, and <0.005 were indicated with one, two, and three asterisks (*), respectively.
|
10.1371/journal.pcbi.1003200 | Predicting Disease Risk Using Bootstrap Ranking and Classification Algorithms | Genome-wide association studies (GWAS) are widely used to search for genetic loci that underlie human disease. Another goal is to predict disease risk for different individuals given their genetic sequence. Such predictions could either be used as a “black box” in order to promote changes in life-style and screening for early diagnosis, or as a model that can be studied to better understand the mechanism of the disease. Current methods for risk prediction typically rank single nucleotide polymorphisms (SNPs) by the p-value of their association with the disease, and use the top-associated SNPs as input to a classification algorithm. However, the predictive power of such methods is relatively poor. To improve the predictive power, we devised BootRank, which uses bootstrapping in order to obtain a robust prioritization of SNPs for use in predictive models. We show that BootRank improves the ability to predict disease risk of unseen individuals in the Wellcome Trust Case Control Consortium (WTCCC) data and results in a more robust set of SNPs and a larger number of enriched pathways being associated with the different diseases. Finally, we show that combining BootRank with seven different classification algorithms improves performance compared to previous studies that used the WTCCC data. Notably, diseases for which BootRank results in the largest improvements were recently shown to have more heritability than previously thought, likely due to contributions from variants with low minimum allele frequency (MAF), suggesting that BootRank can be beneficial in cases where SNPs affecting the disease are poorly tagged or have low MAF. Overall, our results show that improving disease risk prediction from genotypic information may be a tangible goal, with potential implications for personalized disease screening and treatment.
| Genome-wide association studies are widely used to search for genetic loci that underlie human disease. Another goal is to predict disease risk for different individuals given their genetic sequence. Such predictions could either be used as a “black box” in order to promote changes in life-style and screening for early diagnosis, or as a model that can be studied to better understand the mechanism of the disease. Current methods for risk prediction have relatively poor performance, with one possible explanation being the fact they rely on a noisy ranking of genetic variants given to them as input. To improve the predictive power, we devised BootRank, a ranking method less sensitive to noise. We show that BootRank improves the ability to predict disease risk of unseen individuals in the Wellcome Trust Case Control Consortium (WTCCC) data, and that combining BootRank with different classification algorithms improves performance compared to previous studies that used these data. Overall, our results show that improving disease risk prediction from genotypic information may be a tangible goal, with potential implications for personalized disease screening and treatment.
| Genome-wide association studies (GWAS) have recently become the dominant method for searching for the genetic basis that underlies human diseases [1]–[3]. A typical GWAS consists of a collection of genotypes from affected (cases) and healthy (controls) individuals, allowing researches to search for single nucleotide polymorphisms (SNPs) that significantly differ in frequencies between the two groups [4], [5]. Such studies have identified over 1,000 loci associated with more than 165 diseases and traits [6], [7], such as diabetes [8]–[10], cancer [11]–[13], and rheumatoid arthritis [14].
In an effort to move beyond single SNP associations, multiple studies tried to predict disease risk in individuals based on their SNP profiles [15]–[19]. In such risk prediction studies, the entire set of SNPs is potentially used to estimate the risk of every individual to suffer from the disease, and this risk is then compared with the actual disease status of the individual (e.g., case/control). The quality of the prediction is assessed in several ways, with the AUC value (area under the receiver operating curve) being a popular choice, albeit not a perfect one [18], [20]. Intuitively, the AUC can be thought of as the probability that a predictor will correctly classify a pair of samples, one positive and one negative, with a perfect predictor having an AUC of 1, and a random predictor having an AUC of 0.5. Currently, GWAS-based predictors achieve a broad range of AUCs, ranging from relatively high AUC values such as ∼0.9 for Type 1 diabetes, and near random AUC values for other diseases such as mood disorders [18].
Here, we set out to improve our ability to predict disease risk of individuals based only on their SNP genotypes. In risk prediction algorithms, a large set of SNPs is used to perform the predictions, with the identity of the selected SNPs varying due to noise in the choice of data. We thus hypothesized that improvements to risk prediction may be achieved by selecting SNPs that are less sensitive to noise and to the exact choice of data. To test this idea, we devised BootRank, a method that uses bootstrapping in order to rank SNPs for use within predictive models. In BootRank, the data is re-sampled multiple times and a SNP ranking is produced for each such sample, with the final SNP ranking being an aggregate of all the sample rankings.
We tested BootRank on the Wellcome Trust Case Control Consortium (WTCCC) data [21] and found that it increases the robustness of the top-ranked SNPs across different cross-validation sets. In order to validate that the SNP ranking produced by BootRank is also more biologically relevant, we compared its ranking to that based on GWAS association p-values (termed GWASRank). We found that BootRank results in a larger number of enriched pathways associated with the different diseases, and that the pathways detected have substantial support in the literature. Finally, we used the SNP rankings of either BootRank or GWASRank as inputs to seven different classification algorithms and found that using BootRank significantly improves the predictive power for held-out test individuals. Notably, the diseases where using BootRank improves performance the most were recently found to have an underestimated value of heritability, likely because they are predominately affected by variants that have low minimum allele frequency (MAF) [22]. This unexpected finding suggests that BootRank is especially beneficial in cases were the underlying SNPs that affect the disease are poorly tagged or have low MAF.
In summary, our results highlight the importance of robust SNP ranking in the task of disease risk prediction, and offer a concrete method to improve ranking robustness, and consequently the power to predict disease risk and identify biological pathways that may play a role in the different diseases.
Since current disease risk predictors are highly dependent on the initial SNP ranking they take as input, we hypothesized that a common limitation in these predictors may be the sensitivity of the ranking to the exact choice of data. We thus wished to test whether we can improve the ability to predict risk by selecting SNPs that are less sensitive to noise and the exact choice of data. To achieve such robust ranking of SNPs, we used bootstrapping [23], which uses resampling of the data in order to overcome noise. Bootstrapping makes the assumption that individuals are independent and identically distributed, which could be problematic for scenarios such as familial datasets. In the WTCCC dataset, however, individuals have no known dependencies [21]. In our method, we resample the data multiple times, producing a SNP ranking for each such sampling based on GWAS p-values, and then aggregate rankings from all samples to produce a final SNP ranking. We compared our bootstrapping SNP ranking method (termed BootRank), with the commonly used GWAS p-value SNP ranking (termed GWASRank) in a strict cross-validation analysis.
When using cross-validation (CV) to generate training and test sets, there are multiple training sets formed in the process (e.g., in a 5-fold CV, there are 5 different training sets, with a 75% overlap of individuals between every pair of training sets). The fraction of top SNPs that overlap between different training sets when ranking is performed either by GWASRank (as commonly done), or by BootRank, is indicative of the robustness of the SNP ranking method.
To test the robustness of both ranking methods, we used the Wellcome Trust Case Control Consortium (WTCCC) data consisting of ∼2000 cases and ∼1500 control genotypes for 7 different diseases (T1D, Type 1 diabetes; T2D, Type 2 diabetes; CD, Crohn's disease; CAD, coronary artery disease; BD, bipolar disorder; RA, rheumatoid arthritis; HT, hypertension). To ensure minimal bias in the genotypes, we removed all SNPs that were excluded in the original WTCCC paper [21], [24], including SNPs with deviation from Hardy-Weinberg equilibrium or bad clustering (see methods). For each disease, we randomly split the data into training and test sets using a 5-fold CV partition scheme. Next, in each training set, we computed for each SNP the minimum (best) p-value it obtains in one of the genetic association tests (i.e., general, dominant, recessive and additive, see Methods) and ranked SNPs accordingly, as usually done in GWAS studies (i.e., by GWASRank). In addition, we employed a bootstrapping approach, where we re-sample our data from the training set and produce a p-value based ranking for each such sample, and aggregate all rankings to a final SNP ranking based on the median rank SNPs obtained in all bootstrap samples (i.e., by BootRank). To ensure that our results are not due to a specific random partition, we repeated this 5-fold CV analysis 10 times and reported the overall averages.
We found that across all 7 diseases examined, a significantly higher fraction of the top SNPs are shared between different CV training sets when using BootRank as compared to GWASRank (Figure 1). In addition, the fraction of overlapping SNPs for GWASRank decreases as more SNPs are employed (e.g., from 50% for 100 SNPs to 30% for 2000 SNPs in Bipolar disorder (BD)), whereas the fraction of overlapping SNPs in BootRank remains relatively constant when the number of SNPs increases (e.g., ∼70% in BD regardless of the number of input SNPs). We also found that the advantage of BootRank becomes larger if a smaller sample size is used to rank the SNPs (e.g., if only 25% of data is used to rank SNPs for T2D, GWASRank shows ∼0% overlap while BootRank shows ∼75% overlap, Figure S8), further attesting to the robustness of BootRank.
These results show that the fraction of top SNPs that overlap between different training sets using GWASRank is rather low, and that using BootRank is beneficial for increasing the robustness of SNP ranking across multiple CVs in 7 different diseases. However, since ranking robustness by itself is meaningless (e.g., as in ranking SNPs lexicographically by their ID), we next sought to test whether this more robust ranking also has merit in the biological sense.
To independently validate the biological relevance of our SNP ranking, we searched for enriched biological pathways in the different diseases. A popular method to detect pathways that are important to a specific disease is to rank genes according to their expected association, and then use this ranking to compute an enrichment p-value for different pathways (e.g., KEGG pathways). The gene rankings tend to be based on the p-values obtained by GWAS, either by assigning to each gene the best p-value obtained by one of its nearby SNPs [25]–[28], or by pooling all p-values of SNPs inside a gene to one combined p-value [29], [30]. However, in both methods, the initial GWAS p-values are critical, and we thus tested whether the more robust ranking of BootRank also improves the enrichment of pathways in the different diseases.
For each disease and CV training set, we first ranked all the SNPs using either GWASRank or BootRank. Next, we assigned to each gene the best p-value or bootstrap rank obtained by one of the SNPs that resides within it or within its flanking 5 kb, resulting in a ranked list of genes (a total of 172,854 SNPs were mapped to 13,294 genes). We then computed an enrichment p-value for each KEGG [31], [32] pathway using the Wilcoxon rank-sum test, and defined a pathway as enriched if it passed the p-value threshold of P<0.01 in at least 90% of all CV training sets (Figure 2). We found that ranking by BootRank reduced the noise of computed p-values across different CV training sets (as measured by the p-value's coefficient of variation) for 155/163 (95%) of enriched pathways. In addition, BootRank increased the overall number of enriched KEGG pathways in 6 of 7 diseases (Figure 2, a total of 52 uniquely enriched pathways in BootRank compared with 21 in GWASRank).
Next, we examined whether the enriched pathways are known to affect the corresponding disease. In Type 1 diabetes (T1D, Table S1), three interesting pathways that were enriched only in BootRank were “Valine, leucine and isoleucine biosynthesis”, “Chemokine signaling pathway”, and “MAPK signaling pathway”. Notably, a paper that studied longitudinal changes in the amino acid profile in T1D mice, found that the plasma concentrations of valine, leucine, and isoleucine were significantly higher in the diabetic mice [33]. In addition, a few works showed that people with high risk of developing T1D have abnormal level of chemokines [34], and that higher expression levels of the chemokine receptor accelerated disease progress in mice [35]. Another paper also found an enrichment of the MAPK pathway in T1D [36].
In Type 2 diabetes mellitus (T2DM, Table S2), pathways that were enriched only in BootRank included “Type II diabetes mellitus”, “Caffeine metabolism”, “Complement and coagulation cascades” and “alpha-Linolenic acid metabolism”. The “Type II diabetes mellitus” pathway is connected to the disease since it is in fact based on multiple studies of T2DM [37]. In addition, several works showed that caffeine intake affects glucose levels and T2DM risk [38]–[40]. Genes involved in coagulation were found to be upregulated in T2DM patients [41], [42], and consumption of alpha-Linolenic acids was found to ameliorate features of T2DM [43].
In Bipolar disorder (BD, Table S3) the “Neuroactive ligand-receptor interaction” and “Propanoate metabolism” pathways were enriched only in BootRank. The Neuroactive ligand-receptor interaction pathway was indeed found to be enriched in a GWAS replication study [44], and in a study listing potential targets for novel therapeutics for BD, one of the suggested targets was a glutamate propionic acid receptor [45], which is part of the Propanoate metabolism pathway.
In Coronary artery disease (CAD, Table S4) several pathways had enrichment only in BootRank including “Type II diabetes mellitus”, “Colorectal cancer” and “Endometrial cancer”. Notably, there is evidence that all three diseases are also associated with increased levels of vascular diseases [46]–[49].
In Hypertension (HT, Table S5), “Amyotrophic lateral sclerosis (ALS)” and “Ether lipid metabolism” pathways were enriched only in BootRank. A recent paper showed that ALS patients have higher frequency of HT [50], and another paper found an Ether lipid deficiency in the blood plasma of HT patients [51].
In Crohn's disease (CD, Table S6), pathways unique to BootRank, included “Starch and sucrose metabolism”, “Pantothenate and CoA biosynthesis”, “MAPK signaling pathway”, and “Neuroactive ligand-receptor interaction”. Several studies showed that patients with CD have higher intake of starch and sugar [52], and that a low-starch diet can be beneficial for them [53]. In addition, the mRNA and protein levels of Acyl-CoA-synthetase-5 were found to be substantially reduced in CD patients [54], and MAPKs were found to be critically involved in the pathogenesis of Crohn's disease [55]. Interestingly, a new drug intended to treat patients with irritable bowel syndrome targets the GABAA-receptor gene [56], which is part of the Neuroactive ligand-receptor pathway.
In rheumatoid arthritis (RA, Table S7), among many pathways that were enriched only in BootRank were also “Melanogenesis”, “Wnt signaling pathway” and “Hedgehog signaling pathway”. In a study of the regulation of melanin pigmentation, it was shown that patients with RA show localized increased levels of β-Endorphin, which in turn had been implicated in skin pathogenesis [57]. In addition, Wnt signaling and the hedgehog pathway were found to be implicated in RA in mice and humans [58], [59].
Thus, these pathway enrichments show that using BootRank can reduce the noise in p-value computations, allowing more enriched pathways to be detected. Moreover, the enriched pathways have considerable support in the literature as being involved in the various diseases. These results show that BootRank SNP ranking is not only more robust, but also more biologically relevant than that of GWASRank.
Next, we tested whether BootRank's more robust ranking of SNPs can also improve risk prediction of held-out test individuals. To this end, for each disease we used the SNP ranking based only on the training data to filter the top SNPs at some given threshold (e.g., top 1000 SNPs), and then used these SNPs to learn a predictive discriminative model on training individuals using seven different classification algorithms: (1) Random forest (RF) [60], an ensemble classifier that consists of many decision trees; (2) Regularized logistic regression (RLR) [61], where the solution is a sparse vector of weights over the features; (3) A support vector machine (SVM) [62] that uses weights on the training examples to classify test cases; (4) Naïve Bayes (NB), a probabilistic classifier based on applying Bayes' rule with independence assumptions; (5) Robust adaboost (RAB), an adaptive ensemble algorithm where new classifiers are tweaked in favor of those instances misclassified by previous classifiers; (6) Allele count (AC), where classification is done based on counting risk alleles in each individual; and (7) Log odds (LO), where the classification is also based on the frequencies of alleles in cases and controls.
We ranked SNPs either by BootRank or by GWASRank in each of the 7 diseases, and used the different algorithms to predict the disease status of the held-out test individuals, as well as the algorithms' majority vote. We summarize our predictive power using the area under the receiver operating curve (AUC) value, as it is widely used to assess predictive performance in a case-control scenario. Intuitively, the AUC can be thought of as the probability that a predictor will correctly classify a pair of samples, one positive and one negative. A perfect predictor therefore has an AUC of 1, whereas a random predictor has an AUC of 0.5.
Notably, we found that in three diseases, T2D, BD, and HT, using BootRank compared to GWASRank significantly increased the test AUC from 0.69, 0.68 and 0.65 to 0.82, 0.83 and 0.68, respectively, and classification accuracy by ∼7% on average (Tables 1 and S10, 11, Figures 3 and S1, S2, S3, S4, S5, S6, S7). In the remaining 4 diseases (T1D, CD, CAD, and RA), using BootRank was not significantly beneficial to using GWASRank. Moreover, these three diseases for which BootRank improved the performance the most (i.e., T2D, BD and HT) were recently found to have an underestimated value of heritability (h2), probably because they are mainly affected by variants that have a low minimum allele frequency (MAF) [22]. This result suggests that BootRank is especially beneficial in cases were the underlying SNPs that affect the disease are poorly tagged or have low MAF.
We also found that in all cases, predictors based on multiple SNPs ranked by either GWASRank or BootRank had better test AUCs than a predictor based on the best single SNP in the training set, and that in all cases BootRank significantly decreased the over-fitting of the model, as seen by the lower difference between training and test results (Figure 3b).
Next, we compared the performance of the 7 different classification algorithms as well as their majority vote in terms of their AUC values, precision-recall, and accuracy (Tables 2, S8, S9, S10, S11). We found that for each disease, there is one algorithm that performs best, but that overall, the combined majority vote either outperforms individual algorithms or is very close in performance to the best one, encouraging the use of such multi-algorithm approaches to prediction of disease risk.
Finally, we sought to compare our predictive power with that reported in other studies. Since different datasets have different properties, such as number of cases and controls, genotyping density and sampling biases, we only considered for this comparison predictions made on the WTCCC data using a strict cross-validation (CV) training and test scheme (i.e., where as we did here, the test data was not used during the training process) [15], [16], [24], [63]–[67]. We found that in all of these cases, our approach achieved better predictions (Table 1). We note that we excluded methods such as [36], in which the ranking of SNPs was partly done on the entire dataset before splitting into training and test, because such a setting makes use of the held-out test data during the training phase.
Taken together, we show that multi-SNP models can significantly outperform single-SNP models in disease risk prediction, and that BootRank improves the performance and robustness of the predictions over GWASRank. In addition, we show that using BootRank with a majority vote of several algorithms achieves higher AUC on test data compared to previous reports in the literature.
Predicting the risk of individuals to develop a disease given their genetic sequence is a desirable goal, yet the current ability to make such predictions is relatively poor. Since they highly depend on the initial ranking of SNPs that is given to them as input, one common limitation of current methods is the dependence of this input set of SNPs on noise or on the specific choice of data.
Here, we presented BootRank, a SNP ranking method based on bootstrapping, and applied it to the Wellcome Trust Case Control Consortium (WTCCC) data [21] consisting of thousands of individuals suffering from seven different diseases. We first showed that using BootRank results in a more robust SNP ranking compared to using the GWAS p-value (GWASRank), and in more significantly enriched pathways for the different diseases. Moreover, the pathways detected have considerable support in the literature as being involved in the different diseases, validating the biological merit of ranking SNPs by BootRank.
Next, we showed that BootRank could improve disease status prediction in held-out test individuals, by comparing the performance of an approach that uses a single SNP, to approaches that use multiple SNPs ranked by BootRank or by GWASRank. It is important to note, that we are not trying to identify the list of “truly” associated SNPs. In a sense, we do not believe that such a list exists, since many SNP can have a very small effect, or be context specific (e.g., affect only if another SNP is present), or be environment-dependent (e.g., affect only if a certain virus is attacking the cells). Therefore, BootRank is not a feature-selection algorithm per se, but rather its goal is to improve the ranking of SNPs with respect to a disease, and let the classification algorithm use the data at hand to do the final feature selection and model building. For the multi-SNP models, we compared seven different classification algorithms as well as their majority vote. We found that using both multi-SNP models outperform the single-SNP model, and that BootRank significantly improves the test AUC in 3/7 of diseases (T2D, BD and HT, by 0.03–0.15) compared to GWASRank. In addition, we showed that using BootRank with the majority vote of all seven algorithms outperforms previous disease status prediction values of test individuals in the WTCCC data.
Although the improvement over existing publications is large in some cases (e.g., Type 2 diabetes improves from 0.6 to 0.82 AUC), in Type 1 diabetes (T1D) we were not able to improve significantly over existing work and the test AUC remained ∼0.9, suggesting that in T1D we are perhaps approaching the limit of predictive power solely from SNPs. The rest of the variance among individuals could be attributed either to other genetic alterations such as copy number or structural variations (CNVs and SVs, respectively), epigenetic differences such as methylation, or environmental factors such as nutrition or maternal effects.
In the other diseases where our test AUCs ranged from ∼0.7 to ∼0.8, the results are not yet clinically relevant, especially since the propensity of disease cases in the population is low, and thus having low specificity would generate many false positives. Clearly, for these diseases, more could be done in order to improve performance. Adding other genetic variants such as CNVs and SVs could boost performance, as can integrating existing biological knowledge such as pathway information, or protein-protein interaction networks. We note that testing our method on an external dataset to see how well it generalizes would be very beneficial, but this is currently beyond the scope of this work.
The ability to predict disease risk across individuals could transform human health by directing life-style changes among high-risk individuals or by helping early diagnosis by promoting periodical screenings. The ideal risk predictor would make use of both genetic and epigenetic variants, as well as take into account known biological mechanisms and environmental factors, and will enable individuals to appreciate their different risks, aiding them to make the right decisions regarding their health.
All genotype and phenotype (disease state) data was obtained from the Wellcome Trust Case Control Consortium (WTCCC) [21], consisting of 7 different disease sets, with ∼2000 cases for each disease, and a shared set of ∼1500 control individuals. All SNPs that were removed by the original publication due to bad quality, deviation from Hardy-Weinberg equilibrium or bad clustering were removed in this study as well. No correction for family structure was applied to the data.
For each set of cases and controls of a certain disease, 4 different p-values were calculated for each SNP corresponding to the 4 possible genetic models:
The best (lowest) p-value out of the 4 was assigned to each SNP as the GWAS p-value.
For a given training set of N individuals, 100 bootstrap samples were generated by randomly selecting N individuals with replacement. For each such sample, GWAS p-values were calculated for all SNPs, and the corresponding ranking was recorded (GWASRank). The final bootstrap ranking (BootRank) is based on the median rank each SNP achieved across the 100 bootstrap samples. The BootRank code was written in-house and has been made freely available for academic use in the following website: http://genie.weizmann.ac.il/pubs/BootRank/.
For a given 5-fold cross-validation (CV) partition into training and test, a SNP ranking was calculated (by either GWASRank or BootRank), and a number of top-SNPs were selected (e.g., top 1000) as the top-SNPs list for this partition. As a 5-fold CV has 5 such partitions, 5 lists of top-SNPs were calculated. Next, for each pair of partitions, the fraction of overlapping top-SNPs was calculated (i.e., how many top-SNPs they share out of a 1000), and the average across the 10 possible pairs was recorded. This procedure was repeated 10 times to remove possible biases from a specific CV partition, and the overall average of the 10 repeats was reported.
For each disease, we created 10 repeats of 5-fold cross-validation sets, resulting overall in 50 training sets. For each training set, we first computed a SNP ranking (by either GWASRank or BootRank). Next, we converted the SNP ranking to a gene ranking, by assigning each gene the best rank obtained by a SNP that resides within it or within its flanking 5 kb region. For a given gene ranking, we computed enrichment p-values for all KEGG [31], [32] pathways by using the Wilcoxon rank-sum test using an in-house script. Intuitively, testing whether genes belonging to a pathway appear more at the top of the ranked list than expected by chance, by comparing the sum of their ranks to the sum of ranks of a random set of genes of equal size. A pathway was defined as significantly enriched in the disease if it passed an enrichment p-value of P<0.05 in at least 45/50 (90%) of the CV training sets.
For each disease cross-validation partition, we used the training set to produce a SNP ranking (by either GWASRank or BootRank). Next, we selected some top number of SNPs (e.g., 1000), and used these SNPs as input for one of seven classification algorithms: (1) Random forest (RF) [60]; (2) Regularized logistic regression (RLR) [61]; (3) A support vector machine (SVM) [62]; (4) Naïve Bayes (NB); (5) Robust adaboost (RAB); (6) Allele count (AC); and (7) Log odds (LO). Running the algorithms was done by either using built-in functions in MATLAB (e.g., Adaboost), or by coding the algorithm in MATLAB ourselves. After learning the discriminative model from training examples, disease risk was predicted for the held-out test individuals (i.e., not a binary prediction but rather a continuous one), and this ranking of test individuals (i.e., from most likely to have the disease to least likely) was used to compute the area under the receiver operating characteristic curve (AUC value).
In all of the classification algorithms we set the different parameters based only on the training data.
In Random forest, we use an in-house implementation, and set the number of weak predictors to be 500, as we found that the AUC appears to converge around that point. For the number of random features selected at each node in the tree, we used the original recommendation of Breiman [60] as the square root of the total number of features.
In the Regularized logistic regression we use the GLMNET implementation [61], and set the penalty parameter using an internal cross-validation on the training data, where we a subset of the training is set aside as validation and the best penalty is chosen by the validation set. Once the penalty is chosen, we use it to learn the feature weights on the whole training set.
In the support vector machine we use the Radial basis function (RBF) kernel in all cases, and set the penalty parameter C to 1 and the gamma parameter to 1 over the number of features in the model, which are all the default values in the implementation we used (LIBSVM [62]).
In Naïve Bayes we use the Matlab implementation, and set the distribution type as Multivariate multinomial (since the data is discrete) and the prior to be empirical.
In Robust adaboost we use the Matlab implementation, with 500 learning cycles and trees as weak learners.
In Allele count and Log odds we use an in-house implementation and there are no fitted parameters.
For a given cross-validation partition, the predicted rankings of held-out test individuals (i.e., from most likely to have the disease to least likely) were calculated for all algorithms. Next, the disease-risk rankings were combined by assigning each test individual the median ranking obtained across the different algorithms.
|
10.1371/journal.pntd.0000130 | Transcriptional Changes in the Hookworm, Ancylostoma caninum, during the Transition from a Free-Living to a Parasitic Larva | Third-stage larvae (L3) of the canine hookworm, Ancylostoma caninum, undergo arrested development preceding transmission to a host. Many of the mRNAs up-regulated at this stage are likely to encode proteins that facilitate the transition from a free-living to a parasitic larva. The initial phase of mammalian host invasion by A. caninum L3 (herein termed “activation”) can be mimicked in vitro by culturing L3 in serum-containing medium.
The mRNAs differentially transcribed between activated and non-activated L3 were identified by suppression subtractive hybridisation (SSH). The analysis of these mRNAs on a custom oligonucleotide microarray printed with the SSH expressed sequence tags (ESTs) and publicly available A. caninum ESTs (non-subtracted) yielded 602 differentially expressed mRNAs, of which the most highly represented sequences encoded members of the pathogenesis-related protein (PRP) superfamily and proteases. Comparison of these A. caninum mRNAs with those of Caenorhabditis elegans larvae exiting from developmental (dauer) arrest demonstrated unexpectedly large differences in gene ontology profiles. C. elegans dauer exiting L3 up-regulated expression of mostly intracellular molecules involved in growth and development. Such mRNAs are virtually absent from activated hookworm larvae, and instead are over-represented by mRNAs encoding extracellular proteins with putative roles in host-parasite interactions.
Although this should not invalidate C. elegans dauer exit as a model for hookworm activation, it highlights the limitations of this free-living nematode as a model organism for the transition of nematode larvae from a free-living to a parasitic state.
| Hookworms are soil-transmitted nematodes that parasitize hundreds of millions of people in developing countries. Here we describe the genes expressed when hookworm larvae make the transition from a developmentally arrested free-living form to a tissue-penetrating parasitic stage. Ancylostoma caninum can be “tricked” into thinking it has penetrated host skin by incubating free-living larvae in host serum – this is called “activation”. To comprehensively identify genes involved in activation, we used suppressive subtractive hybridization to clone genes that were up- or down-regulated in activated larvae, with a particular focus on up-regulated genes. The subtracted genes, as well as randomly sequenced (non-subtracted) genes from public databases were then printed on a microarray to further explore differential expression. We compared predicted gene functions between activated hookworms and the free-living nematode, Caenorhabditis elegans, exiting developmental arrest (dauer), and found enormous differences in the types of genes expressed. Genes encoding secreted proteins involved in parasitism were over-represented in activated hookworms whereas genes involved in growth and development dominated in C. elegans exiting dauer. Our data implies that C. elegans dauer exit is not a reliable model for exit from developmental arrest of hookworm larvae. Many of these genes likely play critical roles in host-parasite interactions, and are therefore worthy of pursuit for vaccine and drug development.
| Parasitic nematodes are of considerable medical, veterinary and agricultural importance. For example, it is estimated that the morbidity attributable to hookworms, Trichuris and Ascaris, the three most prevalent parasitic nematodes in humans globally, could be as high as 39 disability adjusted life years (DALY) [1]. This assessment takes into account the long-term impact of infection on cognitive and physical development and the overall health of the host. World-wide, ∼1.3 billion people are infected with at least one of these geohelminths [2]. The prevalence of the human hookworms, Ancylostoma duodenale and Necator americanus, alone approaches 740 million, with the foci predominantly within Asia, sub-Saharan Africa, and Latin America [3].
Facultative developmental arrest in the free-living nematode, Caenorhabditis elegans, can occur transiently in the first larval stage (L1) as well as for prolonged periods at the L3 stage. Developmental arrest (often referred to as the dauer stage) in the L3 is triggered in response to conditions, such as crowding, scarcity of food and elevated temperature [4]. When the environment improves, worms exit the arrest to resume development. However, under permissive conditions, arrest is bypassed and adult and reproductive development is favoured. For many parasitic nematodes, arrest at the L3 facilitates survival in the environment. The exit from arrest marks the return to growth and development as well as the transmission of the parasite to its host. Larvae invade a suitable host and undergo a migration through particular tissues to then establish in a target organ and complete the life cycle or arrest in specific tissues. The infective L3 of many parasitic nematodes produce mRNAs which are thought to relate to invasion, migration, and/or survival [5]–[10]. Therefore, the characterization of mRNAs transcribed in the L3 during its transition from the free-living to the parasitic stage may aid in the identification of genes associated with these processes. An attractive parasite model in which to experimentally study this transition is the dog hookworm, Ancylostoma caninum, for which an in vitro serum-stimulation assay exists [11].
Several molecular aspects associated with serum stimulation have been investigated previously in A. caninum. Some researchers have focused on the release of activation-associated proteins; these molecules include the pathogenesis related protein (PRP) superfamily members Ac-ASP-1 [12] and Ac-ASP-2 [13], and the metalloprotease Ac-MTP-1 [14], all of which represent the most abundant excreted/secreted proteins released by serum-stimulated (activated) L3. Other workers have studied activation-associated genes of hookworms using a transcriptomic approach. For instance, Mitreva et al. [8] generated expressed sequence tags (ESTs) for A. caninum (serum stimulated, unstimulated and tissue-arrested L3) and A. ceylanicum (unstimulated L3 and adults), being the first systematic study of genes associated with the host invasion process. However, this study had some limitations in that (1) comparative analyses made between larval stages were qualitative rather than quantitative; (2) some of the observed differences in the abundance of ESTs between activated and non-activated A. caninum L3 seemed to be attributable to differences in the procedures employed for the construction of the cDNA libraries from these life-cycle stages; and (3) the study included a relatively small number of randomly generated sequences available at the time for A. caninum (n = 3840) and A. ceylanicum (n = 3149). Moser et al. [7] addressed the first two points by conducting a quantitative microarray analysis of A. caninum genes associated with the transition to parasitism, focusing on decreased transcription after serum stimulation (i.e., those mRNAs which are “switched off” or reduced in transcription upon host entry). However, this study was also limited to known ESTs available in the public databases.
To infer the mRNAs involved in the infective process of A. caninum, we conducted herein a quantitative study of all known A. caninum sequences as well as newly identified genes discovered through suppressive-subtractive hybridisation (SSH) of activated versus non-activated L3 of A. caninum. The method of SSH was employed to selectively enrich differentially transcribed genes [15]. In summary, 242 potentially up-regulated and 109 potentially down-regulated mRNAs were identified by SSH. There were many mRNAs that were differentially expressed but not identified by SSH, although this might be a function of the number of clones randomly sequenced from our subtracted libraries. The final repertoire of activation-associated genes consisted of 240 up-regulated and 362 down-regulated mRNAs. Among these nearly 600 activation-associated genes were numerous (often substantially up-regulated) mRNAs encoding PRPs and three of the major catalytic classes of proteases (metallo-, cysteine, and aspartic). Several mRNAs encoding novel secreted proteins without any known homologues were also identified. These mRNAs, if demonstrated to be integral to the parasitic process, could represent a new generation of potential vaccine antigens and drug targets against hookworms.
A. caninum L3 were isolated from the faeces of stray dogs in the greater Brisbane area and surrounding towns in Queensland, Australia, using a standard charcoal coproculture method. Cultures were incubated at ∼23°C in a humidified chamber for one week, after which L3 were concentrated using a modified Baermann technique and purified through a nylon filter (20 µm). Larvae were stored for up to four weeks in 50 mM Na2HPO4, 22 mM KH2 PO4, 70 mM NaCl, pH 6.8 [16] in 12.5 cm2 vented tissue culture flasks in the dark at room temperature until use. In total, four separate groups of L3 representing four separate infections from different geographical locations were obtained. The first group was used for SSH and time course studies, whereas the others were employed as biological replicates in microarray validation and real-time PCR analyses (Figure 1).
The specific identity of the parasite material was confirmed by PCR amplification of the first and second internal transcribed spacers (ITS-1 and ITS-2) of nuclear ribosomal DNA (as described by [17]) and automated sequencing (using BigDye chemistry, ABI). The sequences determined were required to be identical to those with GenBank accession numbers Y19181 (ITS-1) and AJ001591 (ITS-2).
Prior to in vitro activation (serum-stimulation), ensheathed L3 were incubated in 1% HCl for 30 min at ∼23°C and then resuspended in RPMI-C (RPMI-1640 tissue culture medium supplemented with 25 mM HEPES (pH 7.0), 100 IU/ml of penicillin, 100 µg/ml of streptomycin, and 40 µg/ml of gentamycin) [12]. To each well of a 24-well tissue culture plate, 5,000 L3 were added. For the SSH, a total of 40,000 L3 were activated in 15% serum and 25 mM S-methylglutathione in RPMI-C, whereas 25,000 L3 were incubated in RPMI-C alone (non-activated control). Five thousand L3 were sampled at each of four time points during the activation (1, 6, 13, and 24 h), in order to perform a time-course analysis of transcripts using real-time PCR (described in “Validation of transcription via real-time PCR”). The same number of non-activated control L3 were separately prepared for this analysis, leaving 20,000 activated and 20,000 non-activated worms for SSH.
For the microarray analysis, activated and non-activated L3 (50,000 of each) were prepared from each of two separate populations of A. caninum. Also, activated and non-activated L3 (5,000 of each) were prepared for real-time PCR. For the in vitro activation, L3 were incubated overnight at 37°C in 5% CO2; pharyngeal pumping in activated L3 was verified by feeding ∼100 of them with FITC-BSA (10 mg/ml) for 3 h and fluorescence was detected using a Leica DM IRB inverted microscope with a Leica DC 500 high-resolution digital camera [11]. Activated and non-activated L3 were each washed twice in phosphate-buffered saline (PBS, pH 7.4; 23°C) and immediately frozen at −80°C. For RNA isolation, larvae were resuspended in 100 µl of Trizol reagent and homogenized in a 1.5 ml tube using an RNase-free, disposable, in-tube pestle and subjected to three rapid (1 min) freeze/thaw cycles. Trizol was added to a final volume of 500 µl, before snap freezing in liquid nitrogen. These samples were stored for ≤1 month at −80°C before RNA was isolated.
Frozen samples of L3 in Trizol were brought to 4°C and centrifuged (16,000×g at the same temperature) for 10 min to remove insoluble debris and residual genomic DNA. RNA was then extracted with chloroform, precipitated with isopropanol, washed with absolute ethanol and resuspended in 50 µl of RNAse-free water. Each RNA sample was treated with 2 U of DNase I (Promega) prior to heat denaturation of the enzyme (75°C for 5 min) and frozen immediately at −80°C. The integrity of RNA was verified to have an RNA Integrity Number >8.0 using an Agilent 2100 Bioanalyzer and RNA 6000 LabChip Kit (Agilent Technologies). RNA used for microarray analysis was stored as an ethanol precipitate in 75% ethanol at −80°C.
First strand cDNA was synthesized from 1 µg of total RNA using the SuperSmart cDNA synthesis kit (Clontech), according to the manufacturer's protocol. Subsequently, double stranded cDNA was produced through 17 rounds of PCR amplification and purified by phenol:chloroform:isoamyl alcohol (25:24:1) extraction, followed by sodium acetate precipitation. SSH was carried out using the PCR Select cDNA subtraction kit (Clontech) according to the manufacturer's protocol. Briefly, cDNA from activated or non-activated A. caninum L3 was digested with the endonuclease Rsa I and ligated to adapters, yielding tester cDNAs for each treatment. Activated tester cDNA was denatured and allowed to re-hybridize in an excess of non-activated “driver” cDNA. This hybridization was termed the forward subtraction and enriched for cDNAs with a higher abundance in activated worms. A reverse subtraction was also performed which enriched for cDNAs which were more abundant in the non-activated worms. Also, an unsubtracted control was prepared according to the standard protocol. Hybridized cDNAs were amplified via two rounds of PCR (according to the recommended protocol), purified by spin-column (QIAGEN) and then cloned into the plasmid vector pGEM-T (Promega). Chemically competent Escherichia coli (TOP 10) were transformed and grown for 8 h at 37°C in Luria Bertani medium (LB) with 100 µg/ml ampicillin. Stocks were stored in glycerol (20%) at –80°C. Immediately prior to sequencing, 50 µl of LB (22°C) was added to each E. coli stock and grown overnight on LB agar plates containing 100 µg/ml of ampicillin. Recombinant colonies were isolated by blue/white selection and then arrayed on a grided LB plate containing 100 µg/ml of ampicillin. Inserts amplified using the TempliPhi DNA Sequencing Template Amplification kit (GE Healthcare Life Sciences) were sequenced unidirectionally using the T7 vector primer in an ABI 3730X1 DNA analyser.
The chromatograms for all raw ESTs were inspected and processed to remove poor quality sequence, with subsequent removal of contaminating vector sequences using BioEdit software v.7.0.1. Following this pre-processing, ESTs were organized into contigs and clusters through an iterative approach using the Cap contig assembly facility in BioEdit under strict conditions, requiring at least a 100 bp overlap and 95% identity among sequences. The resultant contigs and singletons were named according to a simple convention. A “C” in the sequence name identifies sequences composed of multiple ESTs while singletons are indicated with an “S”. Sequences from the forward-subtracted library have four digit identifiers, whereas those from the reverse-subtracted library have three digits. For example, Ac_SSH_C_0056 indicates that the SSH sequence 0056 is composed of multiple ESTs from the forward subtracted library. SSH sequences were compared with existing sequences in GenBank and Wormbase (www.wormbase.org) via BLASTx through NCBI (www.ncbi.nlm.nih.gov/BLAST/) and WU-BLAST (www.ebi.ac.uk/blast2). Alignments were considered statistically significant if an E- or P-value was ≤1×10−5. Neural networks and hidden Markov models were used to predict signal peptides and transmembrane domains by way of the SignalP 3.0 (www.cbs.dtu.dk/services/SignalP/) and TMPred (www.ch.embnet.org/software/TMPRED_form.html) interfaces, respectively. Conserved protein motifs of activation-associated ORFs were identified using the InterProScan website (www.ebi.ac.uk/InterProScan). Potential proteases were classified using the MEROPS protease database (http://merops.sanger.ac.uk/index.htm). Contigs were also mapped to gene ontology (GO) terms based on sequence similarity using the BLAST2GO platform (www.blast2go.de) which compares all contigs with sequences available in several databases, including Wormbase and Uniprot [18]. Only BLASTx hits with a maximum E-value ≤1×10−10 and a minimum of 50% similarity (default software settings) were selected for annotation. A modified one-tailed Fisher exact test based on a hypergeometric distribution was employed in the identification of GO terms for differentially transcribed genes, which were significantly over-represented [19]. This assessment was made relative to the total number of A. caninum genes which had been GO-annotated. Setting the “false discovery” rate limit to 0.5 aided in controlling for multiple testing errors [18].
Sequence data for 9,618 A. caninum ESTs were obtained from the Washington University Genomics Department via the NCBI sequence database (http://www.ncbi.nlm.nih.gov/Genbank/index.html). Chromatograms were pre-processed with the Phred software [20],[21] and organized into contigs and clusters with the Cap3 contig assembly program [22], employing a minimum sequence overlap length of 30 bases and an identity threshold of 95%. Contigs (n = 1311) were assembled from the ESTs and are hereafter designated with “Contig”, followed by a number between 1 and 1311. The remaining singletons were filtered by BLAST E-values (<0.001) to remove potentially spurious sequences and are henceforth referred to by their GenBank accession number. In total, 2,889 individual sequences were identified from the total EST dataset for A. caninum. Sequences representing individual clusters assembled from the sequence data from the forward and reverse subtracted cDNAs as well as the publicly available repository were combined. The combined dataset (a total of 3,100 representative sequences) were submitted for the design of 60-mer oligonucleotides using eArray (Agilent). A total of 9,288 oligonucleotides (3 per target) were proposed for 3,096 contigs. Of these oligonucleotides, 3,443 possessed a non-self perfect match, resulting in 5,845 representing 1,967 genes suitable for microarray analysis. These 5,845 oligonucleotide probe sequences were electronically submitted using eArray for ink-jet in-situ synthesis onto glass slides by Agilent Technologies.
To generate cRNA, 200 ng of total RNA extracted from each activated and non-activated L3 population of A. caninum was reverse transcribed and simultaneously labelled with Cy3 or Cy5 (Agilent). Immediately prior to hybridisation, 500 ng of labelled cRNAs from each of activated and non-activated worms were quantified using a NanoDrop ND-1000 UV-VIS spectrophotometer (NanoDrop), assessed for size distribution and Cy5-dye incorporation using an Agilent 2100 Bioanalyzer and RNA 6000 LabChip Kit (Agilent), mixed together and fragmented. The cRNA from the combined treatments for each population was hybridised to the array in duplicate, with the second hybridisation representing a dye swap to control for any bias in signal intensity between the two dyes. Hybridisations and washes were conducted as per Agilent's Two-colour Gene Expression Hybridisation protocol version 5.0.1. Briefly, 250 µl of hybridisation solution was applied and the microarrays were hybridised for 17 hours at 65°C, 10 rpm. Slides were then washed for 1 minute in Wash Buffer 1 (RT), 1 minute in Wash Buffer 2 (37°C), 1 minute in Acetonitrile (RT) and 30s in Stabilisation and Drying Solution (RT). Slides were scanned using a DNA Microarray Scanner (Agilent). Scanning and feature extraction were performed using Feature Extraction software version 9.1 (extraction protocol GE2-v4_91; Agilent). During extraction, signal intensities were Linear and Lowess-normalized, dye-corrected, and adjusted for local background. Data handling and analysis were carried out using the program SAS v.8.0 (SAS Institute). Processed signal intensities for each probe were averaged across genes, replicates and populations for comparison between treatments by a two-sided t-test with a Type I error rate of 0.01. Only signals differing by at least 1.5 fold (P≤0.01) for each population were considered to represent molecules differentially transcribed in A. caninum as a consequence of serum stimulation in vitro. The effects of dye and probe on the mean signal were assessed graphically. Fold changes in hybridisation were expressed as log2-transformed ratios. The absolute log2 ratios within each level-three GO category were averaged and divided by the mean absolute log2 ratio of all spots on the chip to derive an expression quotient (EQ). The EQ provides an indication of the degree of differential expression associated with a specific GO term.
Reverse transcription real-time PCR was used for the validation of microarray data and for studying levels of transcription in L3 at different time points during the course of serum stimulation in vitro. Ten target sequences were chosen at random and seven others were selected to represent contigs with high, medium and low levels of hybridisation in the microarray. The sequences of all of the primers used in the real-time PCR are listed in Table S2. The single-stranded cDNA template was quantified spectrophotometrically and diluted to an appropriate concentration (2 ng/µl). Two ng of cDNA from each activated and non-activated A. caninum L3 population were subjected to PCR in the presence of 100 nM of the forward and reverse primers in 1× Platinum SYBR Green qPCR SuperMix-UDG (Invitrogen). All experiments were repeated three times with two replicates in each using a Rotor-Gene 6000 Series 2-Plex real-time PCR thermal cycler (Corbett Life Science) employing the following cycling parameters: 50°C for 2 min, 95°C for 2 min, and 40 cycles of 95°C for 15 sec and 60°C for 30 sec. A melt curve analysis was performed from 60°C to 95°C in 1°C intervals to demonstrate the specificity of each amplicon and to identify the formation of primer dimers. Amplicons were also inspected on a 1.2% agarose gel and subjected to automated sequencing to prove their identity. Fold changes in transcripts between activated and non-activated L3 were normalized to the 60S acidic ribosomal protein gene (accession number BF250585) [23] according to an established method [24],[25]. The standard error of the log2 ratios was calculated from the error of the crossing points and observed reaction efficiencies propagated through the calculation of the ratio. Non-parametric statistical inference testing of log2-transformed ratios was performed using a pairwise fixed reallocation randomisation approach with 10,000 simulations which calculated the probability of observing ratios of randomly assigned control and treatment pairs greater than or equal to the treatment effect observed.
Approximately 120,000 activated and 120,000 non-activated L3 of A. caninum were prepared for RNA extraction. As evidenced by the ingestion of FITC-BSA, >95% of all activated L3 resumed feeding, whereas <4% of the non-activated L3 fed (Figure 2). L3 that failed to feed in the presence of the serum stimulus or those that fed in the absence of the stimulus could not be separated from each other.
The RNAs from activated and non-activated L3 were extracted, processed, subjected to forward and reverse SSH and cloned into a plasmid vector. A total of 958 sequencing reactions from the forward library and 171 from the reverse library yielded high quality ESTs. The sequences were deposited in GenBank (accession numbers ES671894–ES672870) and dbEST (accession numbers 46880363 – 46881339) databases. Clustering of these SSH ESTs yielded totals of 242 forward subtracted and 109 reverse-subtracted contigs (Table 1). Approximately half of all forward subtracted sequences were represented by a single EST, although this figure was ∼80% for reverse subtracted sequences. The minimum and maximum lengths of the ESTs were 100 bp and 1500 bp, respectively, with the forward subtracted contigs being slightly larger (571 bp) than those from the reverse subtracted contigs (499 bp). Contigs assembled from the publicly available ESTs for A. caninum had a similar size distribution compared with those assembled from the subtracted ESTs.
From the SSH-derived ESTs, almost 30% of all contigs from both libraries lacked significant sequence similarity to any of the ESTs generated previously for this species [8]. Furthermore, 64–70% of the ESTs from the forward and reverse subtracted libraries respectively did not exhibit significant similarity (at both the nucleotide and protein levels) to sequences within the databases queried (GenBank, EMBL and WormBase) (Table 2). Most mRNAs identified by SSH (63.8%) had a predicted ORF of >50 amino acids. Of these, 19% from the forward subtracted contigs had ORFs with a predicted signal sequence as compared with 6% from the reverse subtracted contigs.
The differential hybridisation of forward and reverse subtracted contigs identified by SSH was verified using a custom designed oligonucleotide microarray. In order to assess the sensitivity and specificity of SSH, we clustered the entire A. caninum EST dataset (9,618 ESTs) and submitted the union of the public ESTs and the SSH contigs for oligonucleotide design. Slides were hybridised with cRNA derived from two separate populations of hookworms (Groups III and IV; Figure 1). The number of L3 obtained from Group III was sufficient to produce two separate pools, serving as a technical replicate for RNA extraction. In total, eight separate hybridisations were performed, one for each of the three RNA samples, plus a dye swap as well as two self-hybridisations, in which Cy3 and Cy5 probes generated from the same RNA stock were used together to hybridise to the slide. For the two populations of A. caninum L3 used, the response to serum stimulation was very similar, as can be seen from the Magnitude (M) versus Amplitude (A) plots generated for each population (Figure 3). The three different oligonucleotides designed for each target yielded consistent log2 ratios among the 50 A. caninum control genes (data not shown). Similar log2 ratios were also observed between arrays and dye-swaps (Figure S1). Furthermore, real-time PCR analysis, performed on 17 randomly chosen SSH-derived sequences, demonstrated the validity of the microarray data. However, the microarray consistently under-estimated the log2 ratio for highly abundant mRNAs, most likely attributable to probe saturation at both the 100% and 50% scans (data not shown). The data discussed in this publication have been deposited in NCBIs Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/) and are accessible through GEO Series accession number GSE8155
In total, 602 mRNAs were associated with log2 transcription ratios significantly greater than zero (P≤0.01). A total of 103 mRNAs new to A. caninum (not in the public databases prior to this study) were identified by SSH, of which 79 had microarray data available. More than 60% of these mRNAs exhibited no significant similarity to any sequences other than A. caninum nor did they have any homologues/orthologues in the publicly available sequences for the congeneric hookworm, A. ceylanicum, or other strongylid nematodes. Sixty-three percent of unique genes, which were up-regulated upon serum stimulation, possessed a signal sequence, in comparison with 33% for down-regulated mRNAs (Table 3). The ten most abundant mRNAs in activated and non-activated L3 are listed in Table 4. Cytochrome c oxidase large subunit of nuclear ribosomal RNA and two novel mRNAs were amongst the most highly expressed mRNAs in both activated and non-activated larvae, with all but the two novel mRNAs being slightly, albeit significantly (P≤0.01) up-regulated upon stimulation. One of the most abundant mRNAs in non-activated larvae was Ac-mtp-1, encoding a metalloprotease involved in skin penetration [26], and this was the only molecule to also exhibit dramatic differential transcription upon stimulation (Table 4).
The types of proteins encoded by mRNAs that were up-regulated upon serum stimulation are very different from those that were down regulated (Tables 5 and 6). Among the 30 most highly up-regulated mRNAs, 17 encoded members of the PRP superfamily [27]. In addition >30% of these mRNAs were predicted to encode secreted proteins. In contrast, most mRNAs that were highly down regulated upon serum stimulation did not possess signal sequences and represented a more diverse group of molecules, including two heat-shock proteins, three PRPs, several novel sequences and a cytochrome P450. Thirteen mRNAs encoding proteases representing all four mechanistic classes, as determined from the MEROPS database [28], were also differentially expressed following serum stimulation (Table 7). In general, the most abundantly represented group of mRNAs associated with activation was the PRPs. Sixty-one different PRP transcripts were identified among the publicly available ESTs and the SSH dataset herein. Thirty-two of these PRPs were associated with activation, 21 of which shared greater than 50% amino acid identity with at least one of the other activation-associated PRPs. All but three of the activation-associated PRPs, Ac-asp-2 (AW626807), an mRNA similar to Ac-asp-1 (SSH Contig 017), and un-clustered A. caninum EST (BQ667555), were up-regulated upon serum stimulation.
Selected mRNAs were examined at different time points throughout the course of serum stimulation (Table 8). Four mRNA's were up-regulated and 4 were down-regulated within 1 hour of stimulation, then a further 80 responded to stimulation after 6–13 hours.The mRNAs that were rapidly up- or down-regulated following activation included the heat shock protein gene, hsp 12.6, the cysteine protease Ac-cp1, a metalloprotease and an mRNA similar to the Onchocerca related antigen (ora-1) from C. elegans. Several PRPs, as well as the metalloprotease Ac-mtp-1 and a neuropeptide-like protein exhibited obvious differential transcription after at most 6 h. Interestingly, Ac_SSH_0042_A, one of the most highly up-regulated activation-associated PRPs, did not appear to increase in transcription until after 13 h of incubation.
All activation-associated mRNAs were annotated with GO terms based on sequence similarity using the Blast2GO platform, and 3-level summaries were prepared for each aspect of GO, molecular function, biological process and cellular component (Figure 4). It is important to note that these classifications provide an estimation only of gene function, because the sequence data used are mRNAs, and often only partial sequences. Nonetheless, we identified a number of gene families that were highly upregulated in activated L3s. The GO category for catalytic activity was significantly over-represented in the mRNAs which were up-regulated upon serum stimulation. Inspection of higher-level terms in the GO tree showed that this significance was likely accounted for by the many proteases and other hydrolytic enzymes which were up-regulated during serum stimulation (Table 7). Based on the frequency of proteases encoded by the current gene/cDNA entries (n = 48) for A. caninum in the NCBI databases, only three such molecules would be expected among the 66 annotated, up-regulated mRNAs. Instead, a total of seven were observed. Additionally, three of the 13 genes predicted to encode metallopeptidases were up-regulated in activated L3. The importance of proteases in activation seems evident from Figure 4, which shows that a majority of the up-regulated genes encode ‘protein catabolism’ functions, a sub-category of the ‘biological process’ GO category.
The GO data were also analysed in the context of mRNA expression data [29]. This analysis focused on the degree to which mRNA expression within specified GO categories was greater than the global array average (Table 9). The absolute log2 ratios of genes associated with defence and the response to external stimuli were 2.4 to 2.9 times greater than the average. The log2 ratio of genes encoding proteins with a predicted extracellular localization was greater than three times the average, with this trend being reflected primarily by the PRPs and proteases. Interestingly, mRNAs associated with carbohydrate binding were more highly represented than average (EQ = 2.7). The largest group of annotated genes was associated with catalytic activity. Collectively, these genes had log2 ratios, which were 1.3 times greater than the average. However, hydrolases and lyases were largely responsible for the trend. Hydrolases (n = 93) and lyases (n = 14) were generally up-regulated 1.5 to 1.7 times more than the average, which was considerable given the size of those GO categories.
The GO classifications can be useful for functional comparisons among species. A comparison was made between activated L3 in A. caninum and C. elegans larvae exiting from dauer. The 3-level charts (Figure 5) display the distribution of GO terms specific to the category of biological process for the 30–36% of hookworm ESTs where a GO function could be assigned. While the “post-stimulation” transcriptome of both organisms was dominated by genes associated with cellular and physiological processes, it was evident that dauer exit in C. elegans was associated with a substantial increase in the proportion of genes involved in growth, development and reproduction. By comparison, serum stimulation in A. caninum did not result in an increased representation of these mRNAs. Interestingly, even the pool of annotated sequences from C. elegans dauer larvae included (15%) of mRNAs associated with development. This was not the case for the ensheathed, non-activated L3 of A. caninum.
For many parasitic nematodes, developmental arrest at the L3 stage is critical for their survival in the environment. In the hookworm model, the exit from developmental arrest is associated with the invasion of the host, assessment of host suitability (larval hypobiosis vs. development to adulthood) and evasion of the host immune response. Therefore, it is plausible that genes associated with this exit facilitate these processes. Using a SSH-based EST approach, we explored transcription in the L3 stage of A. caninum during the transition from a free-living to a parasitic larva, by simulating the earliest stage of parasitism by hookworms in the mammalian (canine) host via serum stimulation. The ESTs produced using this approach were then incorporated into a customised oligonucleotide microarray together with a set of known A. caninum sequences, in order to carry out a large-scale analysis of transcripts during this transition to parasitism.
High-level GO summaries demonstrated that a large proportion (12%) of differentially expressed mRNAs appear to be involved in extracellular localization, of which the majority (27) encode a group of proteins belonging to the PRP superfamily. The greater than average transcription of genes encoding PRPs highlights their importance in the transition of A. caninum from a free-living to a parasitic larva. Although the functions of these molecules are largely unknown, the identification of eight A. caninum PRP superfamily members from the excretory/secretory products of larvae [13],[30] and adults [31]–[33] suggests involvement in host-parasite interactions. Consistent with this hypothesis is the observation that the N. americanus orthologue of Ac-ASP-2, a major vaccine antigen from N. americanus [34], possesses a crystal structure similar to a chemokine, suggesting that it may serve as an extracellular ligand for an unknown host receptor involved in inflammation [35]. Furthermore, Ac-NIF (neutrophil inhibitory factor) [33] and Ac-HPI (platelet inhibitor) [32], the only two A. caninum PRPs for which in vivo functions have been proposed, both exhibit in vitro activities in the mediation of the inflammatory response. All eight of the A. caninum PRPs characterized to date are secreted proteins. Similarly, the presence of predicted signal peptides in most of the activation-associated PRPs suggests that they too are secreted.
Although most of the mRNAs encoding extracellular proteins were PRPs, others encoded proteases. The importance of the activation-associated proteases was evident in the high-level GO summary of differentially expressed mRNAs (Table 9). In broad terms, catalytic activity was over-represented in mRNAs from activated L3, with proteases largely reflecting this trend. The activation-associated proteases represented four of the major catalytic families, namely the metallo-, aspartic, cysteine and serine proteases. These proteases may serve roles in host tissue degradation, digestion and/or development. For example, the activation associated mRNA Ac_SSH_C_0180 is a likely homologue of Parelaphostrongylus tenuis cpl-1, which encodes a cysteine protease implicated in the digestion of host tissue during the escape of the L3 from the intermediate snail host [36]. While many mRNAs encoding proteases were upregulated upon activation, a few such as those encoding the astacin-like metalloprotease, Ac-MTP-1, were down-regulated. MTP-1 is thought to play a critical role in skin penetration in vivo [26], thus supporting its candidacy as a vaccine antigen [37]. A similar scenario exists in larval schistosomes, where the major protease involved in tissue penetration is pre-synthesized and its mRNA is down-regulated before the cercarial stage infects the mammalian host (reviewed in [38]).
Proteases also serve roles in nematode development. For example, O. volvulus cathepsin Z (Ov-cpz-1) is expressed in the cuticle of O. volvulus and is essential for the moult of the L3 to L4 stage [39]. Evidence suggests that the cpz-1 orthologue in C. elegans is also necessary for normal moulting and development [40]. The activation-associated A. caninum mRNA, BQ125325, is cathepsin Z-like and could therefore fulfil a similar role in the moult of hookworm L3 to L4. In addition to moulting, the beginning of the exit from dauer in C. elegans involves a major neurological restructuring [41]. The aspartyl protease Ce-ASP-2 plays an important role in neurodegeneration in this species [42], and the association of a likely hookworm orthologue (Ac_SSH_C_0068) with serum stimulation may also indicate a role in neurological development.
Lastly, other activation-associated proteases may be involved in the digestion of host proteins for nourishment. Ac_SSH_C_0046 is 60% identical to H. contortus pepsinogen (CAA96571) and necepsin I (also referred to as Na-APR-2) from N. americanus (CAC00542.1) [43]. The pepsinogen of H. contortus is expressed in the gut of the adult stage, and mRNAs have been detected in the L4 and adult stages but not in the L3 [44]. Furthermore, its ability to degrade haemoglobin indicates that it could be involved in feeding [44]. The N. americanus aspartyl proteases Na-APR-2, Na-APR-1 and the A. caninum orthologue, Ac-APR-1, are all expressed in the gut of the adult stage where they digest haemoglobin [45].
Mitreva et al. [8] observed that nearly 80% of the A. caninum clusters that were publicly available (before our study here) shared some degree of significant sequence similarity with C. elegans sequences. Moser et al. [7] compared the “serum-stimulated expression data” from many of these clusters to the 1,984 mRNAs associated with dauer exit in C. elegans. This was done on a gene-by-gene basis, and it was observed that cytochrome P450, two neuropeptides, phospholipase and alcohol dehydrogenase were enriched in the dauer form of C. elegans and in non-activated L3 of A. caninum [7]. Conversely, these authors identified mRNAs representing cytochrome c oxidases, an arginine kinase, a heat shock protein, a glycerol hydrolase and glyceraldehyde 3-phosphate dehydrogenase, which were up-regulated in both nematode species following stimulation. Our findings are in accordance with these reports. Neither our study nor that of Moser et al. [7] identified major similarities between the “activated” states of C. elegans and A. caninum. This was attributed to the fact that many of the C. elegans genes which were up-regulated were under-represented in the A. caninum dataset [7]. However, even after enriching for mRNAs that are differentially expressed between free-living and activated L3, the lack of similarity between recovered dauers and activated hookworm L3 persisted.
In lieu of a gene-by-gene approach, we used species-independent GOs to assess the similarity of the relevant A. caninum and C. elegans transcriptomes. This analysis made use of the microarray data generated by [46]. This comparison demonstrated that GO annotations specific to growth, development and reproduction were highly represented in the recovered dauers of C. elegans (Figure 5). This was not the case in A. caninum and is consistent with the observation that serum-stimulation does not invoke moulting of the L3 stage [47]. Conversely, mRNAs from C. elegans dauers or larvae recovered 12 h after beginning of the exit from dauer did not exhibit the significant over-representation of extracellular products as was observed for A. caninum. This finding supports the hypothesis that many of the highly up-regulated mRNAs encoding putatively secreted products are involved in parasitism.
Another major difference between “activation” in C. elegans and A. caninum is the down-regulation of several mRNAs encoding genes involved in G-protein coupled signal transduction during dauer exit in C. elegans. Such a down-regulation was not observed for A. caninum in the present study or that of [7]. As opposed to the PRPs of activated A. caninum L3, the predominant transcripts in C. elegans 12 h after beginning the exit from dauer included a plethora of collagens, many of which were up-regulated ≥32-fold [46]. Even after enrichment for activation-associated mRNAs, only three potential collagens were identified from A. caninum and only one of these (cuticulin) was significantly up-regulated. Based on this information, the activation of A. caninum larvae and the exit from dauer involve considerably different mRNAs. However, the mechanisms by which these mRNAs are regulated may be similar. Ce-hsp-12.6 is a well-known direct target of the fork head transcription factor (designated DAF-16) in C. elegans [48]. Transcripts for this gene are down regulated during dauer exit. Interestingly, it was observed that the A. caninum orthologue of hsp-12.6 (contig 313 from the publicly available ESTs) was also significantly down-regulated during serum stimulation. Assuming that the transcription of this gene is also under the direct regulation of a DAF-16 homologue, the earliest transcriptional events in the transition of A. caninum L3 to parasitism may also be regulated by DAF-16. Real-time PCR conducted on several activation-associated mRNAs at various time points throughout the serum-activation process showed that the levels of many of these mRNAs changed rapidly, as half of those assessed achieved log2 ratios of noticeably more than zero in less than 1 h. The hsp-12.6 transcript was represented in this group of “early responder” molecules. Other mRNAs with similar expression profiles may also be directly regulated by DAF-16. Interestingly, one of the most highly up-regulated PRP mRNAs did not increase substantially in transcription until 13 h after serum-stimulation, which suggests that it may be under the indirect control of DAF-16.
The SSH enrichment of activation-associated mRNAs identified 17 sequences which were up-regulated ≥4-fold and appeared to be unique to parasitic nematodes (Table S1). For example, the EST Ac_SSH_C_0056 was similar in sequence to an uncharacterised gene, ora-1, from C. elegans which is related to an O. volvulus antigen (Ov39) and is thought to play a role in the ocular pathogenesis caused by this parasite [49]. The mRNA representing Ac_SSH_S_0199 was up-regulated nearly 9-fold with the EST showing sequence similarity to the genes Hc-nim-1 and Hc-nim-2 from H. contortus. The mRNAs encoding these genes are abundant in adult H. contortus and represent almost 10% of total mRNA [50]. Hc-NIM-1 is expressed in the hypodermis of the pharyngeal region of the adult worm [50]. Lastly, SSH contigs 0099 and 0032 were among the most highly up-regulated mRNAs in activated L3 of A. caninum, with log ratios of ∼4.7 and 5.9, respectively. Both were predicted to possess a signal peptide and appeared to be specific to A. caninum. Their apparent novelty and stage specificity suggest that they are parasite-specific molecules which might be involved in interactions with host tissues. Functional characterization of these and other novel activation-associated mRNAs may provide insights into the roles that such molecules play in the transition to parasitism. Furthermore, this information may warrant investigating their potential as targets for novel therapeutics.
Having identified a suite of mRNAs associated with serum stimulation, future efforts should be focused on gaining an understanding of the biological function/s of selected members of these parasitism-associated genes. Of particular interest is the large group of PRPs that are up-regulated upon serum stimulation. In combination with their considerable stage specificity and diversity, many of these PRPs may have evolved to perform several coordinated yet distinct functions involved in the parasitic process. Given that PRP-like proteins occur in a wide range of taxa, delineating their function could potentially provide a deeper insight into their roles in parasitism as well as their broader biological significance. It is also of interest that the two most efficacious hookworm vaccine antigens, ASP-2 and APR-1, are members of the two most represented families/groups of proteins associated with this transition to parasitism, PRPs and proteases. We believe that this bodes well for the pursuit of these new molecules identified by SSH as targets for novel vaccines and drugs.
The near absence of mRNAs associated with reproduction, growth and development among activated hookworm L3 probably reflects their ability to further arrest in tissues of non-permissive hosts or in the external environment when conditions for transmission are unfavourable. Although this should not invalidate C. elegans dauer exit as a model for hookworm activation, it highlights the limitations of this free-living nematode as a model organism for the transition of nematode larvae from a free-living to a parasitic state. |
10.1371/journal.pmed.1002183 | Patient-Reported Barriers to Adherence to Antiretroviral Therapy: A Systematic Review and Meta-Analysis | Maintaining high levels of adherence to antiretroviral therapy (ART) is a challenge across settings and populations. Understanding the relative importance of different barriers to adherence will help inform the targeting of different interventions and future research priorities.
We searched MEDLINE via PubMed, Embase, Web of Science, and PsychINFO from 01 January 1997 to 31 March 2016 for studies reporting barriers to adherence to ART. We calculated pooled proportions of reported barriers to adherence per age group (adults, adolescents, and children). We included data from 125 studies that provided information about adherence barriers for 17,061 adults, 1,099 children, and 856 adolescents. We assessed differences according to geographical location and level of economic development. The most frequently reported individual barriers included forgetting (adults 41.4%, 95% CI 37.3%–45.4%; adolescents 63.1%, 95% CI 46.3%–80.0%; children/caregivers 29.2%, 95% CI 20.1%–38.4%), being away from home (adults 30.4%, 95% CI 25.5%–35.2%; adolescents 40.7%, 95% CI 25.7%–55.6%; children/caregivers 18.5%, 95% CI 10.3%–26.8%), and a change to daily routine (adults 28.0%, 95% CI 20.9%–35.0%; adolescents 32.4%, 95% CI 0%–75.0%; children/caregivers 26.3%, 95% CI 15.3%–37.4%). Depression was reported as a barrier to adherence by more than 15% of patients across all age categories (adults 15.5%, 95% CI 12.8%–18.3%; adolescents 25.7%, 95% CI 17.7%–33.6%; children 15.1%, 95% CI 3.9%–26.3%), while alcohol/substance misuse was commonly reported by adults (12.9%, 95% CI 9.7%–16.1%) and adolescents (28.8%, 95% CI 11.8%–45.8%). Secrecy/stigma was a commonly cited barrier to adherence, reported by more than 10% of adults and children across all regions (adults 13.6%, 95% CI 11.9%–15.3%; children/caregivers 22.3%, 95% CI 10.2%–34.5%). Among adults, feeling sick (15.9%, 95% CI 13.0%–18.8%) was a more commonly cited barrier to adherence than feeling well (9.3%, 95% CI 7.2%–11.4%). Health service–related barriers, including distance to clinic (adults 17.5%, 95% CI 13.0%–21.9%) and stock outs (adults 16.1%, 95% CI 11.7%–20.4%), were also frequently reported. Limitations of this review relate to the fact that included studies differed in approaches to assessing adherence barriers and included variable durations of follow up. Studies that report self-reported adherence will likely underestimate the frequency of non-adherence. For children, barriers were mainly reported by caregivers, which may not correspond to the most important barriers faced by children.
Patients on ART face multiple barriers to adherence, and no single intervention will be sufficient to ensure that high levels of adherence to treatment and virological suppression are sustained. For maximum efficacy, health providers should consider a more triaged approach that first identifies patients at risk of poor adherence and then seeks to establish the support that is needed to overcome the most important barriers to adherence.
| Despite more than two decades of research on adherence to antiretroviral therapy (ART) and more than 17 million HIV-positive individuals on treatment, adherence to ART remains a major challenge.
This review aimed to assess the most frequently reported barriers to adherence by patients experiencing adherence challenges.
Published data from 125 studies on patient-reported barriers to adherence were systematically reviewed and analyzed by age group.
The most frequently reported individual barriers across all age groups included forgetting, being away from home, and a change to daily routine. Depression was reported as a barrier to adherence by more than 15% of patients across all age categories, while alcohol/substance misuse was commonly reported as a barrier by adults and adolescents.
With respect to contextual barriers, secrecy/stigma was a commonly cited barrier to adherence, reported by more than 10% of patients across all regions.
Health service–related barriers were frequently reported, including distance to clinic and stock outs.
Evidence-based interventions exist that may address many of the most common patient-reported barriers to adherence.
However, no single intervention will be sufficient to ensure that high levels of adherence to treatment and virological suppression are sustained.
Health providers should consider a more triaged approach that first identifies patients at risk of poor adherence and then seeks to establish the support that is needed to overcome the most important barriers to adherence.
Several key health service improvements are also required to ensure that patients are able to consistently access ART.
| Global targets for scaling up antiretroviral therapy (ART) include ensuring that 90% of patients on ART achieve viral suppression. This gives a renewed emphasis to ensuring optimal levels of adherence. Negative outcomes of longer-term suboptimal adherence include increased risk of disease progression [1], drug resistance [2], high viral load and consequent risk of transmission [3,4], and death [5,6].
Maintaining high levels of adherence is a challenge across settings. Suboptimal adherence to antiretroviral medication has been reported for specific patient groups such as adolescents [7], pregnant women [8], and others in high-, middle-, and low-income countries. A broad range of context-specific barriers to adherence have been reported, including forgetfulness, stigma, adverse drug reactions, and competing responsibilities [9,10]. These challenges have been categorized as individual, interpersonal, community, and structural factors [11].
Several interventions have been found to improve adherence in randomized trials, including adherence counselling, text messaging, and reminder devices [12], and these are recommended by the WHO [13]. However, there remains a need to understand the relative importance of different barriers to adherence in order to inform the targeting of different interventions and inform future research.
We conducted this systematic review to assess patient reported barriers to adherence among HIV-infected adults, adolescents and children in high-, middle-, and low-income countries.
This study follows the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) statement [14]. The study protocol and PRISMA statement are available in the Supporting Information (S1 Text and S2 Text) [14].
To be included, studies had to provide information about barriers to adherence reported by at least 50 adult patients or 20 children or their caregivers who were non-adherent to ART according to study definitions. These cut offs were chosen in practical consideration of the high number of studies identified through preliminary searches and the limited and sometimes unreliable information to be gained from small studies [15]. Using a search strategy that combined terms for ART, adherence, and commonly reported reasons for non-adherence (See S1 Text), two investigators (ZS, NF) screened MEDLINE via PubMed, Embase, Web of Science, and PsychINFO from 01 January 1997 to 31 March 2016. We also screened abstracts from all International AIDS Society conferences, all conferences on HIV Treatment and Prevention Adherence, and all conferences of the European Society for Patient Adherence, Compliance, and Persistence from 2012 to 2015, and electronically available abstracts of the Conference on Retroviruses and Opportunistic Infections (2014–2016) to identify recent studies that may not yet have been published in full. We supplemented database searches by screening bibliographies of review articles and all included full-text articles. Studies that included patients mainly receiving dual- or mono-therapy (>20% of cohort), antiretroviral interventions other than treatment (PEP, PrEP, or PMTCT), or reported adherence to medication other than ART were excluded from review. We only extracted data on non-adherent patients, following the definition of non-adherence provided by the studies; if studies included both adherent and non-adherent patients and data could not be disaggregated, the study was excluded. Studies assessing adherence interventions were also excluded unless baseline (i.e., pre-intervention) information was relevant to our analysis. Qualitative studies were excluded unless relevant quantitative data were also provided. No language restriction was applied.
We extracted data independently in duplicate (ZS, NF) using a standardized, piloted data extraction form and following the predefined protocol. Barriers to adherence were initially categorized according to AIDS Clinical Trials Group adherence instrument items, with additional categories developed to capture challenges common in resource-limited settings [16]. Using categories adapted from a qualitative review of adherence barriers [11], responses were grouped into individual, contextual, and health service–related barriers; where there was uncertainty, investigators discussed to achieve consensus. Information about study design and setting, age of patients, and adherence measure was also extracted. The following variables were extracted to assess study quality: use of a previously piloted and/or validated questionnaire to assess barriers to adherence; random sampling; and use of objective adherence measures. These indicators were identified following a review of the first ten eligible studies.
We calculated proportions and corresponding 95% CIs for all reported barriers to adherence and pooled data following transformation [17,18], using random effects models stratified by age (adults, adolescents, and children, as defined by the studies) [19]. Because statistical tests for heterogeneity are not reliable for pooled proportions [20], we assessed heterogeneity through visual inspection of forest plots. We ran prespecified subgroup analyses to assess the potential influence of study quality indicators and explored changes over time (assessed using date of study completion) using meta-regression. All analyses were conducted in Stata version 13.0.
From an initial screen of 5,560 abstracts, 125 studies met our inclusion criteria (Fig 1). These studies provided information about barriers to adherence for 19,016 patients—17,061 adults, 1,099 children, and 856 adolescents—with documented non-adherence to ART. Studies were carried out across 38 countries, with the majority carried out in the Africa region (58 studies, 16 countries), the European region (14 studies, 10 countries), and the Western Pacific region (8 studies, 6 countries). Study quality was rated to be moderate overall. The majority of studies (78/125, 62%) used a validated questionnaire to assess barriers to adherence and piloted the questionnaire (89/125, 71%); however, less than half of studies (38/125, 30%) used random sampling, and objective adherence measures (pill count, pharmacy refill, and viral load) were only used in the minority (18/125, 14%) of studies; these limitations are important potential sources of selection and information bias. The most common definitions of adherence were no missed doses (55 studies) and >95% adherence (37 studies). Average duration on ART ranged from 4 wk to 239 wk (median 78 wk). Characteristics of included studies are summarized in S1 Table.
The most frequently reported barriers to adherence for adults, adolescents, and children are summarized in Figs 2–4. The most frequently reported individual barriers across all age groups included forgetting, being away from home, and a change to daily routine. Depression was reported as a barrier to adherence by more than 15% of patients across all age categories, while alcohol/substance misuse was commonly reported as a barrier by adults and adolescents. Among adults, feeling sick was a more commonly cited barrier to adherence than feeling well (relative risk 1.68, 95% CI 1.23–2.30). The proportion of adolescents reporting barriers to adherence was higher per barrier compared to adults, but data are limited and confidence intervals are wide for most estimates.
With respect to contextual barriers, secrecy/stigma was a commonly cited barrier to adherence, reported by more than 10% of patients across all regions (S2 Table). Notably, secrecy/stigma was more commonly reported as a barrier to adherence among children/caregivers compared to adults.
Health service–related barriers were frequently reported, including distance to clinic and stock outs. Distance to clinic was reported as a barrier for all age groups across 11 low- and middle-income countries in the Africa, South East Asia, and Western Pacific regions. Stock outs were reported across 13 countries for adults, all in low- and middle-income countries. For children, two studies reported stock outs in the United States. Barriers related to drug toxicity were frequently reported among adults and adolescents, while for children and adolescents, palatability was an important concern. In meta-regression, there was evidence that the frequency of reporting toxicity, pill burden, and being sick as barriers to adherence have reduced over time (Table 1).
Our review highlights the diversity of patient-reported barriers to adherence across age groups. The most frequently reported individual barriers across all age groups included forgetting, being away from home, depression, and a change to daily routine; alcohol and/or substance misuse was commonly reported by adults and adolescents. Health service–related barriers, including distance to clinic and stock outs, were also frequently reported.
Most barriers to adherence are amenable to interventions that have been evaluated in randomized trials (Table 2) [12,21–30]. Notably, forgetting was the most frequently cited barrier to adherence across all age groups. Challenges relating to timing of medication, including being asleep, could be overcome through text messaging, reminder devices, and individual counseling that seeks to routinize medication taking in a way that fits in with other daily activities [31].
We found that health service barriers played an important role in frustrating efforts to maintain high levels of adherence to treatment. Long distance to clinics is a risk factor for loss to care [32], and decentralization of HIV services is associated with better retention [23]. Recent stock outs of antiretroviral medication have been recorded in several countries in Africa [33], and there is a pressing need for increased vigilance as countries move to adopt the policy of treating all HIV-positive individuals and consider transitioning from established first-line medications to newer regimens.
Previously, concern has been expressed that people who receive ART early in their disease progression may be less adherent to treatment [34,35]. The finding that feeling sick was a more commonly reported barrier to adherence than feeling well suggests that this may not be the case and supports the recent recommendation by WHO to treat all HIV-positive individuals regardless of immune status [36]. As HIV programs start to provide early ART to people earlier in their HIV infection, it will be important to prospectively collect data to further evaluate this concern.
Toxicity and pill burden have both been found to be associated with poor adherence in other reviews [37,38]. This review found that the frequency of reporting these factors as barriers to adherence has reduced over time, which is consistent with efforts by WHO and other agencies to promote fixed-dose combinations and rationalize treatment guidelines towards the use of antiretroviral drugs associated with a better safety profile [39].
The main adherence barriers identified by our review are consistent with a recent review by Langenbeek et al that found substance use, concerns about ART, satisfaction with care providers, stigma, social support, and self-efficacy to be strongly associated with adherence [40]. In contrast to the Langenbeek review, which assessed the influence between baseline patient characteristics and adherence, our review assesses adherence barriers that are reported by patients and, as such, was able to identify a number of additional frequently reported barriers that could not be gleaned from clinic records; because of this difference in approach, the number of studies in our review that were included in the previous review is small (19%). This approach builds on a previous review of patient-reported barriers to adherence that was published in 2006 [10]. There have been considerable changes in ART delivery over the past decade: the number of patients on ART globally has increased; drug regimens have improved with respect to tolerability and simplicity; and service provision has been decentralized. In updating this review, we have been able to include a larger sample size that allowed for a ranking in the frequency of reporting of barriers, disaggregated by age, and an understanding of how these barriers differ by geographical region and over time. This can allow for a better understanding about where resources need to be focused in order to improve adherence among different patient populations.
Several recent studies have indicated that adolescents face challenges across the continuum of HIV care, and outcomes of ART are worse for adolescents compared to adults [41]. HIV programs should pay particular attention to the adherence challenges faced by this vulnerable population and target adherence interventions accordingly; such an approach would be facilitated by the development of better ways to measure adherence. To date, few adolescents have been enrolled into trials of interventions to improve adherence, and this is an important area of future research [42]. A pilot feasibility study found that personalized, interactive, daily text message reminders were feasible and acceptable, and significantly improved self-reported adherence. However, larger controlled studies are needed to determine the impact of this intervention on ART adherence and other related health outcomes for youth living with HIV/AIDS globally [43,44].
Pregnant and postpartum women are another group who face challenges in maintaining high levels of adherence to medication [8]. Despite the critical need for ART during pregnancy and the postpartum period, evidence-based interventions to promote ART adherence during this period are lacking. A recent exploratory study of 109 HIV-positive pregnant South African women found that mobile phone access (>90%) and interest in text messaging for adherence support (88.1%) was high, and the majority (95%) of women were willing to disclose their status to a treatment buddy/supporter [45].
More generally, the fact that most adherence intervention studies are only able to show a modest effect in randomized trials is likely in part a consequence of the multiple challenges patients face in adhering to treatment as indicated by the findings of our review (i.e., within studies the percentages of reported barriers added up to more than 100%). Future research is encouraged to evaluate the effectiveness (effect size and interaction effects) of more than one intervention on virological suppression, using a factorial or adaptive clinical trial design to precisely determine the specific interventions and components of interventions that work best. [46].
Our review has several strengths and limitations. Strengths include our broad search strategy and inclusion criteria that allowed for the identification of a substantial number of studies and synthesis of a large dataset. Limitations are mainly related to study quality and include the variable definitions of adherence used by the different studies, different approaches to assessing adherence barriers and time on ART and it is possible that these and other unreported factors may have influenced outcomes. Information about drug toxicity is limited by the possibility that not all experienced adverse events are related to ART, even if they were perceived as a reason to stop taking the medication. Caution is also needed in the interpretation of results as some reasons for poor adherence (e.g., forgetting) may be put forward because they are perceived to be more socially acceptable than others (e.g., chaotic lifestyle or substance misuse). Although several analyses were undertaken to identify potential explanations for variance in findings, we could not thoroughly explore all possible differences in covariates (e.g., geographic region, income) due to the need to avoid spurious associations that may arise from large numbers of outcomes and covariates. We searched multiple databases and conferences, which allowed us to include data from over 100 studies for analysis; however, we did not include regional databases, and this may have limited identification of potentially eligible studies. An additional limitation to note with respect to children is that barriers were mainly reported by caregivers, and these may not represent the most important barriers faced by children themselves [47]. Finally, any study that looks at self-reported adherence will likely underestimate the frequency of non-adherence, and studies that assessed objective measures of adherence are more likely to be accurate in terms of reflecting true adherence rates.
In conclusion, this review highlights that patients on ART face multiple barriers to adherence and no single intervention will be sufficient to ensure that high levels of adherence to treatment and virological suppression are sustained. Rather than introducing single interventions into HIV programs, health providers should consider a more triaged approach that first identifies patients at risk of poor adherence and then seeks to establish the support that is needed to overcome the most important barriers to adherence. For maximum efficacy, adherence support strategies should be targeted to those individuals who require support. Finally, although the majority of the most commonly reported barriers are amenable to intervention at the individual level, several key health service improvements are also required to ensure that patients are able to access ART.
|
10.1371/journal.pntd.0006527 | The initial effectiveness of liposomal amphotericin B (AmBisome) and miltefosine combination for treatment of visceral leishmaniasis in HIV co-infected patients in Ethiopia: A retrospective cohort study | North-west Ethiopia faces the highest burden world-wide of visceral leishmaniasis (VL) and HIV co-infection. VL-HIV co-infected patients have higher (initial) parasitological failure and relapse rates than HIV-negative VL patients. Whereas secondary prophylaxis reduces the relapse rate, parasitological failure rates remain high with the available antileishmanial drugs, especially when administered as monotherapy. We aimed to determine the initial effectiveness (parasitologically-confirmed cure) of a combination of liposomal amphotericin B (AmBisome) and miltefosine for treatment of VL in HIV co-infected patients.
We conducted a retrospective cohort study at a Médecins Sans Frontières—supported health center in north-west Ethiopia. We included VL-HIV co-infected adults, treated for VL between January 2011 and August 2014, with AmBisome infusion (30 mg/kg total dose) and miltefosine orally for 28 days (100 mg/day). Proportions of initial treatment outcome categories were calculated. Predictors of initial parasitological failure and of death were determined using multivariable logistic regression. Of the 173 patients included, 170 (98.3%) were male and the median age was 32 years. The proportion of patients with primary VL (48.0%) and relapse VL (52.0%) were similar. The majority had advanced HIV disease (n = 111; 73.5%) and were on antiretroviral therapy prior to VL diagnosis (n = 106; 64.2%). Initial cure rate was 83.8% (95% confidence interval [CI], 77.6–88.6); death rate 12.7% (95% CI, 8.5–18.5) and parasitological failure rate 3.5% (95% CI, 1.6–7.4). Tuberculosis co-infection at VL diagnosis was predictive of parasitological failure (adjusted odds ratio (aOR), 8.14; p = 0.02). Predictors of death were age >40 years (aOR, 5.10; p = 0.009), hemoglobin ≤6.5 g/dL (aOR, 5.20; p = 0.002) and primary VL (aOR, 8.33; p = 0.001).
Initial parasitological failure rates were very low with AmBisome and miltefosine combination therapy. This regimen seems a suitable treatment option. Knowledge of predictors of poor outcome may facilitate better management. These findings remain to be confirmed in clinical trials.
| North-west Ethiopia faces the highest burden world-wide of visceral leishmaniasis (VL) and HIV co-infection. VL treatment outcomes in HIV co-infected patients are associated with high initial treatment (parasitological) failure and recurrence rates after cure (relapse). With secondary chemoprophylaxis, the risk of relapse can be reduced. However, with the current VL treatment regimens, the initial parasitological failure rates remain high. In this study, we aimed to determine the initial effectiveness of a combination of liposomal amphotericin B (AmBisome) and miltefosine for treatment of VL in HIV patients in Ethiopia. We conducted a retrospective study using routine program data from a Médecins Sans Frontières—supported health center in north-west Ethiopia. We included 173 adult VL-HIV co-infected patients treated for VL with a combination of AmBisome and miltefosine. Initial cure rate was 83.8%, death rate 12.7% and parasitological failure rate 3.5%. Tuberculosis co-infection at VL diagnosis was predictive of initial parasitological failure. Predictors of death were age >40 years, hemoglobin ≤6.5 g/dL and primary VL. Initial parasitological failure rates were very low with AmBisome and miltefosine combination therapy. This regimen seems a suitable treatment option. Knowledge of predictors of poor outcome may facilitate better management. These findings remain to be confirmed in other studies.
| Visceral leishmaniasis (VL) is a protozoan infection caused by the Leishmania donovani species complex [1]. In East Africa and the Indian subcontinent, it is caused by L. donovani, whereas in the Mediterranean region and South America, by L. infantum [2]. Ethiopia is among the top six high burden countries, with approximately 3.2 million people at risk and 3400–5000 VL cases occurring annually [3–5]. North-west Ethiopia faces the highest burden world-wide of VL-HIV co-infection, an estimated 20% of VL patients are HIV co-infected [6]. HIV infection influences the clinical course of VL: it reactivates latent Leishmania infection, increases VL severity, and negatively affects treatment outcomes [7]. VL in turn promotes the progression of HIV infection [7].
As to VL treatment outcomes, both higher (initial) parasitological failure rates and higher relapse rates have been described [7,8]. To reduce the relapse rate, secondary prophylaxis is the way to go [7,9–11]. We have recently documented the effectiveness, safety and feasibility of pentamidine secondary prophylaxis—started after parasitological cure was achieved—in Ethiopian VL-HIV co-infected adults [9,11]. However, achieving parasitological cure remains challenging, as co-infected patients have shown poor treatment response to all available antileishmanial drugs especially when administered as monotherapy [7].
Several studies have shown that pentavalent antimonials cause severe adverse events (cardiotoxicity, nephrotoxicity, hepatotoxicity, pancreatitis) resulting in high case fatality rates [7,12–18]. Antimonials have also been shown to stimulate HIV-1 replication in vitro [19]. In East Africa, high case fatality rates of 6.8% to 33.3% have been reported [14–16]. Due to the high case fatality rates, the World Health Organization (WHO) recommends that pentavalent antimonials should ideally not be used as a first line treatment for VL in HIV co-infected patients [20].
In comparison to antimonials, there is relatively limited clinical experience with miltefosine—a newer antileishmanial agent [7,15,20,21]. An Ethiopian study comparing miltefosine and the antimonial—sodium stibogluconate (SSG), showed that miltefosine was safer (lower death rates: 1.6% vs. 6.8%) but had lower initial effectiveness (higher initial parasitological failure rates: 17.5% vs. 2.3%) [15]. These high parasitological failure rates increase the potential for the emergence of resistant parasites [21]. Because patients with parasitological failure are potential reservoirs of resistant parasites, they are a major public health concern, especially in East Africa where the main mode of transmission of Leishmania parasites is anthroponotic [21]. Furthermore, miltefosine has a long half-life of approximately one week, and can develop resistance with a single point mutation [22]. The most optimal way to use miltefosine would be in combination with another antilesihmanial agent [21,23].
Several studies from the L. infantum areas of the Mediterranean region, in small numbers of co-infected patients, showed that liposomal amphotericin B was safe and effective [7,20]. Based on these findings and absence of similar studies from other VL endemic areas, the WHO recommended liposomal amphotericin B as the first line treatment for VL in HIV co-infected patients [20]. However, liposomal amphotericin B (AmBisome) monotherapy at a total dose of 30 mg/kg also had limited effectiveness in Ethiopia, with initial parasitological failure rates of 32.8% [24].
Combination treatment has been used in tuberculosis, HIV and malaria with good outcome and is increasingly being used for VL [23,25]. We reasoned that, as both AmBisome and miltefosine had been found to be safe, but with high initial parasitological failures as monotherapy, the combination of the two drugs with different modes of action and non-overlapping toxicity might yield a safe regimen able to decrease the high initial parasitological failure rates [7,15,20,21,24,26]. In 2011, Médecins Sans Frontières (MSF) introduced a compassionate treatment regimen of AmBisome and miltefosine combination as first line treatment for VL in HIV co-infected patients. In this study, we aimed to determine the initial effectiveness (cure, death and parasitological failure rates) of this regimen for treatment of VL in HIV co-infected patients in Ethiopia.
The study was conducted at Abdurafi health center—an MSF supported health facility located in a remote town in Amhara region, in northwestern Ethiopia. The MSF support focuses on clinical management of VL, HIV and concomitant infections. It is a major VL treatment site in Ethiopia and medical services are free of charge. The majority (>95%) of VL patients treated at the health center, are young adult males working on the large-scale agricultural schemes in the northwestern lowlands.
We conducted a retrospective cohort study using routine program data. In the main (per-protocol) analysis, we included all VL-HIV co-infected patients diagnosed between January 2011 and August 2014, aged ≥18 years, treated with an initial VL treatment regimen composed of a combination of AmBisome and miltefosine. Patients that discontinued treatment, defaulted, were transferred-out or had a missing VL treatment outcome were excluded.
We also conducted a sensitivity analysis that is similar to an intention to treat analysis. In this analysis, we also included patients that defaulted or were transferred-out, and considering that they probably had Leishmania parasites at exit, they were all classified as having parasitological failure.
Patients with prolonged fever, splenomegaly and wasting were considered VL suspects and underwent further diagnostic evaluations [20]. Patients without prior VL treatment history (primary VL) were first screened using the rK39 rapid diagnostic test (IT-Leish, Bio-Rad laboratories, USA)[27] and a positive result confirmed VL. Those testing negative were screened with the leishmania direct agglutination test (DAT, Royal Tropical Institute, Amsterdam, The Netherlands)[28] and a high titer (≥1:3200) confirmed VL. Those with an intermediate DAT titer (1:800–1:1600) underwent tissue aspiration (spleen, bone marrow or lymph node) and a positive result confirmed VL. Patients with prior VL treatment history (relapse VL) underwent tissue aspiration and a positive result confirmed VL. A clinical diagnosis was made in patients [primary VL (with negative rK39 test results and intermediate DAT results) and relapse VL] who were contra-indicated for spleen aspirate (i.e. spleen size ≤2 cm, bleeding tendency, pregnant, severely anemic, jaundiced or in a state of collapse) or who declined a bone marrow aspirate and didn’t have palpable lymph nodes. Furthermore, a clinical diagnosis was also made in patients [primary VL (with negative rK39 test results and intermediate DAT results) and relapse VL] with negative bone marrow aspirate results but persistent strong VL clinical suspicion in the absence of differential diagnoses [24,29].
HIV positive status was defined by two positive results of serological tests performed in parallel {KHB (Shanghai Kehua Bio-engineering Co-Ltd, Shanghai, China) and STAT-PAK (Chembio HIV1/2, Medford, New York, USA)} and confirmed by the ELISA test {ImmunoComb (Orgenics ImmunoComb II, HIV 1&2 Combfirm)}. Antiretroviral therapy (ART) prescription was according to national guidelines and tenofovir, lamivudine and efavirenz combination was the most common first line regimen [30].
The initial treatment regimen was a combination of liposomal amphotericin B (AmBisome, Gilead Sciences) at a total dose of 30 mg/kg, divided into 6 infusions of 5 mg/kg on alternate days and miltefosine (Impavido, Paladin Labs, Montreal, Canada) administered orally for 28 days (100 mg/day). Patients that showed a slow treatment response received a second course of the combination regimen at the same dosage (treatment extension). Slow treatment response was defined as a substantial parasite reduction after finishing the treatment course (day 29) as compared to baseline (parasite decrease ≥2 log-grades but parasitology result was still positive) [29,31]. Patients that showed no parasitological response to the combination treatment received sodium stibogluconate (Albert David Ltd., Kolkata) at a dose of 20 mg/kg/day by intramuscular injection for a minimum of 30 days (rescue therapy). No parasitological response was defined as no substantial parasite reduction at the end-of-treatment as compared to baseline (parasite decrease ≤1 log-grade and parasitology result was still positive) [29,31].
Our main outcome of interest was the initial treatment outcome which was defined as the treatment outcome after completion of the first VL treatment course. The categories of initial treatment outcome included: cure, death, parasitological failure, defaulter and transfer-out. Parasitological tests were performed at the end-of-treatment in all co-infected patients except for those without palpable spleen or lymph nodes and who refused bone marrow aspirate, or for those with a contraindication for spleen aspirate. For this category of patients, cure was assessed clinically. Patients with parasitological failure received additional treatment (retreatment) and the subsequent treatment outcome was classified as retreatment outcome. Treatment outcome at discharge was either the initial treatment outcome or where applicable the retreatment outcome. The categories of treatment outcome at discharge were as for the initial treatment outcome.
Cure was defined as improvement in symptoms and signs of VL after treatment initiation (i.e. absence of fever, decrease in spleen size, increase in hemoglobin, weight gain) and a negative parasitological test at the end-of-treatment. Parasitological failure was defined as a positive parasitological test at the end-of-treatment. Death from all causes during VL treatment at the health center were documented. Defaulting was defined as absconding from treatment. Transfer-out was defined as referral to another hospital facility. Treatment discontinuation was defined as discontinuation of a VL treatment regimen prior to using less than 90% of the total recommended dosage.
Since the program onset, clinical data were collected using standardized data collection tools and stored in electronic databases. The databases were updated on a daily basis by data managers. The data were collected at admission through history taking, clinical examination, laboratory and/or radiological investigations, and treatment prescriptions (VL and ART regimens). The following variables were assessed from patient history: age (years), sex, residential status [migrant worker (an individual who seasonally relocates to another area in search of work); settler (an individual who has been relocated to another area by the state) and resident (an individual who has permanently lived within a specific area for a duration of 2 or more years)], duration of illness (months) and VL treatment history (primary, relapse).
The following variables were assessed by clinical examination: weight (kilograms), height (meters)/length (centimeters), body mass index [BMI; weight in kilograms ÷ (height in meters)2], spleen size (centimeters), the level of weakness, ascites, peripheral edema, bleeding and jaundice. The spleen size (centimeters) was measured from the junction of the anterior axillary line and the left coastal margin to the tip of the spleen. Weakness severity was defined according to MSF guidelines [29] as follows: [State of collapse: unable to sit up unaided and cannot drink unaided; severely weak: cannot walk 5 meters without assistance; other types of weakness were classified as “other”].
The following variables were assessed by laboratory and/or radiological investigations. The mode of diagnosis of HIV was defined above (see HIV diagnosis). Using a microscope with a 10X eyepiece and 100X oil objective, tissue parasite grading were as follows: [0 (0 parasites/1000 fields); 1+ (1–10 parasites/1000 fields); 2+ (1–10 parasites/100 fields); 3+ (1–10 parasites/10 fields); 4+ (1–10 parasites/field); 5+ (10–100 parasites/field); 6+ (>100 parasites/field)]) [31]. Hemoglobin level was measured using a hematology analyzer—Beckman Coulter AcT diff, Beckman Coulter Inc., 2003, USA. CD4 count was measured at baseline and every six months after ART initiation using the FACS counter (BD FACS Calibur flow cytometer, 2009, USA). Tuberculosis diagnosis and WHO clinical staging were according to WHO guidelines [32,33].
The primary outcome was the initial treatment outcome (cure, death or parasitological failure). The proportion of individuals with the different outcome categories (excluding patients that defaulted or were transferred-out), were calculated with 95% Wilson confidence intervals (CI). In secondary analysis, the association of initial treatment outcome with VL treatment history was assessed by the Chi-squared or Fisher’s exact test.
Predictors of initial parasitological failure and predictors of death were determined. Predictors of parasitological failure were analyzed among patients with parasitological failure or cure, whereas predictors of death were analyzed among patients who died or stayed alive (cured or parasitological failure). The choice of variables analyzed as predictors was based on literature review and consideration of variables available in the dataset. To overcome the problem of substantial missing baseline CD4 count results, we created a composite marker for advanced HIV disease, defined as having either additional WHO stage IV disease [33] or a CD4 count <50 cells/μL at VL diagnosis [34]. Continuous variables were categorized based on information from literature and a recent study on predictors of death [35]. The association between predictors and parasitological failure or death were first assessed with Chi-squared or Fisher’s exact test. When the p-value was <0.1, the predictor was included in a multivariable logistic regression model. Non-significant variables (p-value ≥0.05) were removed step by step until no more variables could be dropped.
Lastly, a sensitivity analysis similar to an intention to treat analysis was performed. Defaulters and transfer-outs were included in this sensitivity analysis; and considering that they probably had Leishmania parasites at exit, they were all classified as having parasitological failure. All the statistical methods described above were then repeated. All statistical analyses were performed with Stata version 14.
Ethics approval was received from the Institutional Review Board of the Institute of Tropical Medicine, Antwerp, Belgium, and the Ethical Review Committee of the Institute of Public Health, Gondar University, Ethiopia. This research fulfilled the exemption criteria set by the MSF Ethical Review Board (ERB) for a posteriori analyses of routinely collected clinical data, and thus did not require MSF ERB review. It was conducted with permission from the Medical Director of the MSF Operational Centre Amsterdam.
Between January 2011 and August 2014, 227 patients were diagnosed with VL-HIV co-infection and treated at the Abdurafi health center. Forty patients were treated with other VL treatment regimens (AmBisome alone, n = 29; SSG based, n = 11). Two patients were started on AmBisome and miltefosine combination treatment, however miltefosine was later discontinued. The reason for discontinuing miltefosine was because of miltefosine stock-out rather than adverse event. Both of these patients completed AmBisome monotherapy and one was cured and the other died. Four patients defaulted, 5 were transferred out and 3 had unknown treatment outcome. These 54 (23.8%) patients were excluded from the main (per-protocol) analysis. A total of 173 patients were included in the main analysis (Fig 1).
A comparison of characteristics of patients excluded or included in the study is shown in S1 Table.
Most patients were male (n = 170; 98.3%), residents (n = 101; 59.1%) and young (median age of 32 years; interquartile range [IQR] 28–39). The proportion of patients with primary VL (n = 83; 48.0%) and relapse VL (n = 90; 52.0%) were similar. Most patients had advanced HIV disease (n = 111; 73.5%) and were on ART prior to VL diagnosis (n = 106; 64.2%) (Table 1).
Compared to relapse VL patients, a higher proportion of primary VL patients were aged 18–40 years (88.0% vs. 75.6%), had been ill for ≥2 months (46.2% vs. 27.9%) and had not started ART prior to the VL episode (57.7% vs. 16.1%). Lastly, primary VL patients had a lower parasite load at admission (median +4 vs. +5) and a lower proportion had cure confirmed by a parasitological test (43.1% vs. 65.0%) (Table 1).
The outcomes were: cured, 145/173 (83.8%; 95% CI, 77.6–88.6); died, 22/173 (12.7%; 95% CI, 8.5–18.5) and parasitological failure, 6/173 (3.5%; 95% CI, 1.6–7.4). The outcome by VL treatment history was significantly different as shown in Table 2.
Of the 6 patients with initial parasitological failure (Table 2), 1 was retreated with AmBisome and miltefosine combination, 2 with AmBisome alone and 3 with SSG based regimen. One of the patients retreated with SSG based regimen died, all the rest were cured. The treatment outcomes at discharge were: cured, 150/173 (86.7%); died, 23/173 (13.3%) and no parasitological failure.
Tuberculosis co-infection at VL diagnosis was predictive of initial parasitological failure (adjusted odds ratio (aOR), 8.14; 95% CI, 1.42–46.72; p = 0.02). There was a statistically non-significant association between high tissue parasite load (parasite grade 6+) at VL diagnosis and initial parasitological failure. In multivariable analysis, VL treatment history was not significantly associated with initial parasitological failure (Table 3).
Independent predictors of death were age >40 years (aOR, 5.10; 95% CI, 1.50–17.44; p = 0.009), hemoglobin ≤6.5 g/dL (aOR, 5.20; 95% CI, 1.83–14.79; p = 0.002) and primary VL (aOR, 8.33; 95% CI, 2.27–30.63; p = 0.001) as shown in Table 4.
The initial treatment outcomes were similar to those from the main analysis as shown in S2 Table. As also reported in the main analysis, tuberculosis co-infection at VL diagnosis was predictive of initial parasitological failure. Additionally, BMI<16 kg/m2 (severe malnutrition) was also predictive of initial parasitological failure as shown in S3 Table. The independent predictors of death were similar to those from the main analysis as shown in S4 Table.
VL-HIV co-infected patients are confronted with high (initial) parasitological failure rates and high relapse rates [7,8]. While we recently identified pentamidine secondary prophylaxis as a promising option to reduce the relapse rates [9,11], achieving parasitological cure has been challenging. In this study, we determined the initial effectiveness (cure, death and parasitological failure rates) of a combination regimen of AmBisome and miltefosine for treatment of VL in HIV co-infected patients in Ethiopia. The initial cure rate was 83.8%, death rate 12.7% and parasitological failure rate 3.5%. Tuberculosis co-infection at VL diagnosis was predictive of initial parasitological failure. Age >40 years, hemoglobin level ≤6.5 g/dL and primary VL were predictive of death.
Although it remains difficult to compare historical cohorts, the initial treatment outcomes with combination therapy compared with those for AmBisome monotherapy—the previous first line treatment at the MSF treatment site [24], are as follows: parasitological failure rates were significantly lower (3.5% vs. 32.8%; p<0.001), cure rates were significantly higher (83.8% vs. 60.4%; p<0.001), and death rates were non-significantly higher (12.7% vs. 6.8%; p = 0.05). None of the deaths are considered treatment-related. In the present study, we had a higher admission rate of late stage VL patients that were referred from other hospitals as compared to the previous AmBisome monotherapy study. We have recently shown that besides HIV serostatus, other important predictors of death were: age >40 years, hemoglobin ≤6.5 g/dL, bleeding, jaundice, edema, ascites and tuberculosis [35]. We also found that in the presence of major predictors of death, the predictive effect of treatment on outcome may be minimal [35,36]. If we are to consider initial parasitological failure and death as overall initial failure (assuming that patients who died also had parasitological failure), then the overall initial failure rate with the combination regimen were also significantly lower than with AmBisome monotherapy (16.2% vs. 39.6%; p<0.001) [24].
In contrast to pentavalent antimonials that cause severe adverse events resulting in high case fatality rates [7,12–18], AmBisome and miltefosine have been shown to be safe [7,15,20,21,24,26]. Therefore the higher case fatality rates reported here are more likely related to the patients clinical conditions (late stage VL patients), than due to AmBisome and miltefosine toxicity [7,15,20,21,24,26]. Combination treatment may increase treatment efficacy and tolerance, reduce treatment duration and cost, and limit the emergence of drug resistance [23,25]. VL combination therapies have been successfully used and implemented in HIV-negative patients [37,38]. In vitro studies have demonstrated synergy between liposomal amphotericin B and miltefosine [39]. A combination of synergistic treatment regimens with different modes of action and mechanisms to develop resistance can also delay the emergence of drug-resistance [23]. AmBisome and miltefosine combination therapy was safe and effective in HIV-negative patients [38]. For co-infected patients, an Indian retrospective study on AmBisome and miltefosine combination therapy showed it was safe and effective, however, initial treatment outcomes were not reported [40].
Thanks to an agreement between Gilead and WHO, on a donation programme, WHO is providing AmBisome for treatment of visceral leishmaniasis for free to low income countries in East Africa and South Asia. This donation programme started in 2012, and has been extended in 2017 for another five years, including middle income countries. This access to free AmBisome will enhance affordability of implementing AmBisome-based treatment regimens for VL-HIV [41–44]. Combining AmBisome and miltefosine may be crucial: with lower initial parasitological failure rates, fewer patients required retreatment and therefore treatment duration was shortened. This promotes patient compliance, reduces risk of adverse events, and patient and health facility costs [23]. AmBisome must be transported and stored at temperatures below 25° centigrade [26]. Miltefosine may be teratogenic, therefore it is contraindicated during pregnancy, women of reproductive age must use effective contraception during and for 3 months after treatment [21]. Miltefosine may also cause gastrointestinal symptoms (nausea and vomiting) [21]. In this study, there was no treatment discontinuation secondary to gastrointestinal symptoms. Good tolerance to miltefosine has also been reported in other studies [15,40]. Overall, similarly as in India, our findings are encouraging in terms of efficacy of combination therapy. Data from clinical trials are now needed to enhance the evidence base. Importantly, studies evaluating AmBisome and miltefosine combination therapy in HIV patients have been conducted in Ethiopia [45] and started in India [46]. The trial findings are to be published soon.
We found that tuberculosis co-infection at VL diagnosis was predictive of initial parasitological failure. This is similar to findings from a study conducted in Sudan [47]. Tuberculosis causes immunosuppression which may inhibit parasite clearance, resulting in parasitological failure [47]. In another Ethiopian study, high tissue parasite load at VL diagnosis was shown to predict parasitological failure [16]. Possibly in the presence of underlying immunosuppression, a high parasite load on admission might be more difficult to clear. In our study, we also found an association between high tissue parasite load at VL diagnosis and initial parasitological failure, however, it was not statistically significant. This finding could be explained by the few outcomes observed in this study—only 6 patients with initial parasitological failure. In sensitivity analysis, BMI<16 kg/m2 (severe malnutrition) was predictive of initial parasitological failure. As with tuberculosis, severe malnutrition causes immunosuppression which may inhibit parasite clearance, resulting in parasitological failure.
The predictors of death identified—age >40 years and hemoglobin level ≤6.5 g/dL—are similar to those reported from other studies [35,48,49]. Patients aged >40 years, may have underlying co-morbidities (e.g. cardiovascular diseases), lower immunity and/or severe VL disease [48,50–53], which increases their risk to die. Severe anemia may cause congestive heart failure [54]. In comparison with relapse VL patients, we found that primary VL patients were more likely to die. The exact reason for this is unknown. However, since VL-HIV coinfection is a severe illness [7], and the risk of VL relapse is high (26% at one year) [8], it is probable that relapse VL patients may be more likely to be aware of the dangers of VL than primary VL patients, they may present to the health center with an earlier stage of illness in comparison with primary VL patients that may arrive with more end stage illness.
Most of our patients are young adult males who get infected with leishmania while working in the agricultural fields within the VL endemic region. HIV infection is also more common in young adults than children. As shown in Table 1, the lowest age in this VL-HIV cohort was 18 years old. However, if we were to treat a younger (<18 years old) VL-HIV co-infected patient cohort with this combination regimen, probably those <5 years old would have high case fatality rates, that would be comparable to those of patients aged >40 years in this cohort. This is because several studies have shown that younger HIV negative VL patients have higher risk of death [49,51,55]. AmBisome and miltefosine would still be the treatment of choice, however, an allometric dosing table for miltefosine should be used in children, as it might improve treatment outcomes [56].
There are some limitations to this study. Diagnosis and cure were not systematically confirmed by parasitological tests. Fifty-seven (88.0%) of the patients whose cure was assessed clinically had non or barely palpable spleen at the end-of-treatment, inhibiting performing a spleen aspirate, they declined having a bone marrow aspirate because the procedure is painful and they didn’t have palpable lymph nodes. This occurs commonly in settings without non-invasive investigations to assess VL cure. We acknowledge that this could likely lead to some degree of underestimation of the failure rates. However, it is important to note that in this cohort of patients, the more ill patients at admission (for instance with tuberculosis co-infection) and those more likely to fail (e.g. patients with a history of VL) were more likely to get a parasitological test for confirmation of cure at the end of treatment. Furthermore, in a recent study from this setting, we found no difference in long term outcomes (relapse or death) among patients with treatment outcome at discharge of parasitological cure versus those with clinical cure [8]. In a worst-case scenario, assuming similar failure rates for those with clinical cure compared to those undergoing tissue aspiration, the overall failure rates would still only be 6.4%, clearly better as what has been reported with miltefosine and AmBisome monotherapy. While longer patient follow-up to report on the relapse rates would have been of interest, this was not done as some patients were included in the pentamidine secondary prophylaxis trial, which has been published recently [11]. Indeed, to prevent relapse, secondary prophylaxis is likely the most important intervention, and not the initial treatment [10]. Consequently, the focus of this paper was on the initial effectiveness of the combination regimen in achieving parasitological cure, which is a prerequisite before starting secondary prophylaxis. In this study, CD4 counts were missing for a significant proportion of patients. Working in a remote area with relatively limited capacity, we did not have the capacity to perform autopsies. However, basing on clinical experience, some of the underlying causes of death include: severe anemia, severe pneumonia, tuberculosis, hepatic failure and sepsis [35,57,58]. Also, as a retrospective study, we could only study predictors from the collected variables.
In conclusion, we determined the initial effectiveness of a combination regimen of AmBisome and miltefosine for treatment of VL in HIV co-infected patients in Ethiopia. Initial parasitological failure rates were very low with AmBisome and miltefosine combination therapy when compared with the initial parasitological failure rates with either drug administered as monotherapy [15,24]. Therefore, this combination regimen seems a suitable VL treatment option in HIV patients. These findings remain to be confirmed in clinical trials. After achieving initial cure, those at high risk of VL relapse should be initiated on secondary prophylaxis.
|
10.1371/journal.pntd.0003188 | Aedes hensilli as a Potential Vector of Chikungunya and Zika Viruses | An epidemic of Zika virus (ZIKV) illness that occurred in July 2007 on Yap Island in the Federated States of Micronesia prompted entomological studies to identify both the primary vector(s) involved in transmission and the ecological parameters contributing to the outbreak. Larval and pupal surveys were performed to identify the major containers serving as oviposition habitat for the likely vector(s). Adult mosquitoes were also collected by backpack aspiration, light trap, and gravid traps at select sites around the capital city. The predominant species found on the island was Aedes (Stegomyia) hensilli. No virus isolates were obtained from the adult field material collected, nor did any of the immature mosquitoes that were allowed to emerge to adulthood contain viable virus or nucleic acid. Therefore, laboratory studies of the probable vector, Ae. hensilli, were undertaken to determine the likelihood of this species serving as a vector for Zika virus and other arboviruses. Infection rates of up to 86%, 62%, and 20% and dissemination rates of 23%, 80%, and 17% for Zika, chikungunya, and dengue-2 viruses respectively, were found supporting the possibility that this species served as a vector during the Zika outbreak and that it could play a role in transmitting other medically important arboviruses.
| Arthropod-borne viruses (arboviruses) cause significant human morbidity and mortality throughout the world. Zika virus, which is reported to be transmitted by Aedes (Stegomyia) species mosquitoes, caused an outbreak on the island of Yap, in the Federated States of Micronesia in 2007. This was the first described outbreak of Zika in Oceania, which has had several arbovirus outbreaks in the past. Diagnosing the outbreak was difficult due to the similarity in clinical symptoms between disease caused by Zika virus and other viruses. This work describes the efforts to identify the mosquito species that were responsible for transmission of the virus. While no virus was isolated from any species of mosquito collected during the current study, the predominant species found was Aedes hensilli and through the complementary laboratory studies, this mosquito was implicated as a probable vector for Zika virus. In addition, this species was found to be susceptible to both the medically important dengue-2 and chikungunya viruses.
| Outbreaks of arboviral disease have been documented in islands of the western Pacific including The Federated States of Micronesia (FSM) and Palau. Multiple dengue outbreaks have been reported in the western pacific [1]–[4] with an outbreak of dengue 4 virus occurring in Palau in 1995 after a 7 year absence of dengue on this island [5]. This first outbreak of dengue 4 in the Western Pacific also affected FSM the same year [6]. Additional dengue outbreaks occurred more recently in FSM during 2004 and 2012–13 [7], [8]. In 2007, an outbreak of acute febrile illness characterized by rash, conjunctivitis, fever, and arthralgia was reported on the island of Yap in the Federated States of Micronesia. While dengue was originally suspected, clinicians noted differences from classical dengue fever and collected serum from acutely-ill individuals for diagnosis. Chikungunya virus (CHIKV) was also considered as the clinical presentation was representative of CHIKV infection and an ongoing epidemic of CHIKV was occurring in Southeast Asia. However, Zika virus (ZIKV) nucleic acid was detected in 14% of the samples tested and no evidence of alternate etiologies was identified [9].
Zika virus (ZIKV) is a member of the family Flaviviridae. Presence of the virus in human specimens has been demonstrated by virus isolation (samples from Africa and Asia) and antibody presence (Asia) [10]–[15]; however, only a handful of clinical disease cases were described in the literature prior to this 2007 outbreak [16]–[19]. Since the outbreak in Yap, additional ZIKV outbreaks have been documented in Gabon in 2007 and in French Polynesia in 2013 [20], [21]. Mosquito vectors from which virus has been identified include (among others) Aedes africanus, Aedes luteocephalus, Aedes aegypti, and Aedes albopictus (all belonging to the subgenus Stegomyia) [15], [21]–[25]. However, little else is known regarding the natural ecology of the virus.
Because this was the first documentation of the virus in Oceania, understanding the biological transmission of the virus was a public health priority. A team including epidemiologists, clinicians, entomologists, and public health personnel investigated the outbreak with the objectives of characterizing the epidemiology, course of clinical illness, and ecological factors contributing to the epidemic and transmission of the virus. Household surveys were performed to obtain serum specimens, to obtain clinical and epidemiological data, to identify risk factors for infection, and to collect entomological specimens for the purpose of determining the most probable epidemic vector [9]. The entomological studies included both immature (larval and pupal) and adult surveys to determine the species present on the island, contributions of distinct container types in mosquito maintenance, and to perform virus isolation. This report describes the entomologic findings from the field collected material as well as subsequent laboratory studies assessing the vector capacity of the likely outbreak vector.
The Federated States of Micronesia are located in the Western Pacific Ocean northeast of Papua New Guinea (Figure 1). Yap State is the westernmost state and is comprised of a main island group consisting of four closely associated islands situated at 9° North and 138° East. It is approximately 6 km wide by 15 km long with a population of 7,391 persons during the 2000 census. The climate is tropical with warm temperatures and rainfall reported throughout the year. Mosquitoes were collected and household surveys performed between July 4, 2007 and July 16, 2007 at 170 randomly selected homes in 9 out of the 10 municipalities, representing 16% of the total households on the island. The outbreak was estimated to have begun in April, 2007 and continued through July of 2007 [9].
Adult mosquito sampling was carried out using three collection methods. Host seeking mosquitoes were collected using light traps, resting mosquitoes were collected using vacuum aspiration, and mosquitoes looking for oviposition habitat were collected with gravid traps. Gravid and light traps (light only) were set in the evening at three sites in the state capital city of Colonia from July 4–9, and July 12–16. Collection bags from the light and gravid traps were recovered daily for 9 days in the early morning. A battery operated backpack or handheld mechanical aspirators were used to collect mosquitoes resting in and around random houses where serosurveys were being performed during daytime hours, July 9–15.
All collected mosquitoes were identified morphologically using keys from Bohart [26], and Rueda [27]. They were then sorted by sex, species, collection method, and collection period and placed into cryovials at a temporary laboratory set up in Colonia. Specimens were frozen at −20°C on-site; they were later transported to the CDC at Fort Collins CO, USA where storage was at −70°C until processing.
All indoor and outdoor water containing receptacles at the randomly selected households [9] were inspected for mosquito larvae and pupae. Live larvae observed in receptacles were collected and identified to species and allowed to emerge to confirm identification. All pupae found were collected and reared to adulthood. Key habitat information was recorded and larval indices (Breteau and household [28]) were calculated from the collected data for each of the species observed.
Each pool of mosquitoes (not exceeding 40 individuals) were placed into a 1.7 mL polypropylene tube (Eppendorf, Hauppauge, NY) and ground with a pestle (Kontes) and 500 µl of Dulbecco's minimal essential medium (DMEM) (Gibco) supplemented with 10% fetal bovine serum (FBS), 100 U/mL of penicillin and streptomycin, 1 U/mL of fungizone and gentamycin. The homogenized mosquitoes were then centrifuged at 15,000 g for 1 min. Triturate was then transferred to a new tube and frozen at −70°C. 100 µl of thawed triturate was then plated onto a 96-well cell culture plate (Corning). 50 µl of a Vero cell suspension was then added to the same well and placed into an incubator at 37°C and 5% carbon dioxide. The cell and homogenized mosquito mixture was then monitored daily for cytopathic effect (CPE) for 10 days. Medium from those wells presenting signs of CPE (presumptive positives) were removed and placed at −70°C until use [29].
The viruses used for laboratory mosquito infections were obtained from the Arbovirus Reference Collection at CDC, Fort Collins, CO (Table 1). As no ZIKV field strain was obtained, we used the prototype strain for ZIKV laboratory infections.
Viral RNA was isolated using the QiaAmp viral RNA protocol (Qiagen). Total RNA was extracted from 50 µl of cell supernatant (CPE positive pools) or 100 µl of mosquito homogenate (artificial infections) and eluted from the kit columns using 60 µl of elution buffer. The RNA was stored at −70°C until use.
Both reverse-transcription PCR (RT-PCR) and real-time RT-PCR assays were utilized to detect viral nucleic acid. The Titan one-step RT-PCR (Roche) kit was paired with the primers FU1 and cFD3 for detection of Zika [30], [31]. Briefly, 5 µl of sample RNA was added to the kit components and 400 nM of primers. The manufacturer's protocol was followed with no modifications. The reactions were analyzed by gel electrophoresis. The real-time PCR assay was used on both the presumptive positive pools and the experimentally infected mosquitoes. The previously described Zika (800 series set) and chikungunya virus specific primer and probe sets were used [32], [33]. The DENV-2 oligonucleotide set were designed with the Primer Select software program (DNASTAR) (1085 CCAAACAACCCGCCACTCTAAG, 1244c TTTCCCCATCCTCTGTCTACCATA, and TaqMan probe 1145 FAM-AACAGACTCGCGCTGCCCAACACA-BHQ1) and were based on the published GenBank full-length sequences. All real-time assays were performed by using the QuantiTect probe RT-PCR reagent kit (Qiagen). Briefly, a 50µl total reaction volume consisted of kit components, 10 µl of RNA, 400 nM of each primer, and 150 nM of probe. The reactions were subjected to 45 cycles of amplification in an iQ5 Real-Time PCR detection system (BioRad) following the manufacturer's protocol. The limits of detection for DENV, ZIKV, and CHIKV assays were found using the previously described techniques [34] and were cycle threshold (Ct) values of 37.7, 36.1, and 38.0 respectively, which is equivalent to approximately 1.0 plaque forming unit/mL. In addition, each run included a standard RNA curve. The standard curve was completed by serially diluting the virus stock and extracting the RNA from each dilution, according to the previously mentioned RNA extraction protocol, while simultaneously titrating each dilution in a standard plaque assay. A curve correlation coefficient of ≥0.950 and a 90–100% PCR efficiency was used to validate each detection assay.
Mosquito eggs were collected at selected houses in Yap using oviposition cups. Briefly, black, plastic cups were lined with seed germination paper [35] and filled approximately half full with water. Cups were placed under foliage near selected homes (2–4 feet above the ground) and collected after 3–5 days. Field collected egg liners were wrapped in moist paper towels, sealed in Ziploc-style bags, and transported to the insectary at the Center for Disease Control and Prevention (CDC), Fort Collins Colorado for colonization. The eggs were washed with a 10% bleach solution prior to hatching in a pan of tap water to eliminate surface fungal and bacterial contaminants.
Larvae were supplied with either a liver powder solution or mouse pellets as appropriate for the developmental stage and identified to species as 4th instar. All larvae collected were identified as Aedes (Stegomyia) hensilli. Pupation occurred between days 5–7 post hatching. Pupae were removed from the larval pans and allowed to emerge into 1 ft3 adult mosquito cage (BioQuip). In order to produce the next generation, adults were provided an anesthetized mouse as a blood meal source and the engorged females were provided with an oviposition site (seed germination paper) to deposit their eggs. The process was repeated in order to get sufficient numbers of experimental mosquitoes. In addition, species verification was performed on F2 adult mosquitoes.
Three to four day-old adult Ae. hensilli mosquitoes (F12–15) were fed on blood meals containing ZIKV, CHIKV, or DENV-2. The blood meals contained equal parts of virus, FBS with 10% sucrose, and sheep blood (Colorado Serum CO) washed with phosphate-buffered saline and packed by centrifugation. A Hemotek feeding system (Discovery Workshops) was used to deliver the blood meal to the mosquitoes for 1 hour at 37°C. The fully engorged females were separated and placed into a humidified environmental chamber (Thermo Scientific) and held at 28°C for 8 days until processing. Blood meal titer was determined by plaque assay to determine input titer.
After the 8 day holding period, mosquitoes were cold anesthetized and decapitated with the heads and bodies placed into separate 1.7 mL tubes (Eppendorf). A 400 µl aliquot of Dulbecco's minimal essential medium (DMEM) (Gibco) supplemented with 10% fetal bovine serum (FBS), 100 U/mL of penicillin and streptomycin, 1 U/mL of fungizone and gentamycin was added to each tube and the sample was homogenized using a micropestle (Kontes). The supernatant was clarified by filtration through a 0.2 µM syringe filter (Pall) and stored at −70°C until use [36].
Virus presence was determined using the virus isolation method as described above. An infected mosquito exhibited a virus positive body [percent infected = (number positive bodies/total number of mosquitoes processed) X 100] while those with disseminated infections were the infected individuals with virus in the head [percent disseminated = (number of positive heads/number of positive bodies) X 100]. Quantities of viral RNA were determined using real-time RT-PCR (above) and correlated with viral titer.
Adult mosquitoes were captured using three different collection methods (light trap, gravid trap, and vacuum aspirations). A total of 879 mosquitoes were collected in 84 trap nights. Additionally, 475 individuals collected as larvae and/or pupae were reared to adults for confirmatory identification and processing. Nine species were identified in these collections (Table 2). The most abundant adult species collected was Aedes hensilli (41.2%) followed by Culex quinquefasciatus (28.1%). All other species each comprised less than 10% of the total collection. All adult mosquitoes (field collected adults and those reared from immatures) were processed and subjected to virus isolation efforts. No viable virus was recovered from any of these mosquitoes.
From 170 randomly surveyed households (July 4–16, 2007), 1366 water holding habitats were identified. Larvae and/or pupae were collected from 586 of these containers and 85% of surveyed households had at least one infested habitat; individual habitats sometimes contained more than one species (Figure 2). The most prevalent containers with larvae or pupae were discarded cans followed by coconut shells (Table 3). Proportionally, containers including tires, tarps, floats, and bamboo had high percentages of immatures but several of these container types were found only infrequently (Table 4). Containers such as water barrels, used to collect rainwater, while proportionally fewer in number than other containers, were actually major contributors to mosquito production due to the sheer number of larvae and pupae present (e.g. thousands of immature mosquitoes per water barrel in comparison with cans or shells which typically contained fewer than 10 individuals each). In total, ten different species were identified from the larval collections. Ae. hensilli was both the most abundant and most prevalently identified immature species being found in 83% of the infested containers distributed all over the island (household index of 81.2 and Breteau index of 282.9).
Because no virus was found in any of the field collected material, laboratory infections were performed on the most common mosquito collected, Ae. hensilli, to determine if this species could have served as the epidemic vector. Cohorts of Ae. hensilli were infected with three different viruses during these studies: 1) ZIKV - to determine if this was the likely vector during the outbreak; 2) CHIKV- to ascertain whether Ae. hensilli could serve as a vector for this virus which was expanding through SE Asia and was considered as a possible etiology of the outbreak prior to ZIKV diagnosis; 3) DENV - as Ae. hensilli was previously postulated as the vector of the 1996 dengue outbreak in Yap [5].
Cohorts of 3–4 day old adults were provided infectious blood meals with titers of at least 4.9 log10 pfu/mL. Mosquitoes provided the lowest dose of ZIKV were resistant to infection with only 7% becoming infected (Table 5). However, at least 80% of those receiving a slightly higher dose became infected. Curiously, only 13–23% of those developed disseminated infections. Only a small percentage of mosquitoes exposed to DENV-2 became infected (0–21%) and few of these had virus dissemination. In contrast, Ae. hensilli was found to be exceptionally sensitive to CHIKV with infection and dissemination rates greater than 60% and 80% respectively.
The 2007 outbreak of ZIKV in Yap prompted the investigation of vectorial capacity of the predominant local mosquito to transmit this virus and other related viruses that are present or threaten to affect FSM and other Western Pacific island countries. Yap State, the western-most part of FSM, has previously been affected by arboviral outbreaks [6] but the discovery of ZIKV on the island highlighted the risk of epidemics due to agents previously unknown to the area. During the entomological investigations, collection of mosquito larvae and pupae from over 15 distinct container types revealed a wide range of habitats, both natural and artificial, that could support development of a variety of mosquito species. Because the island extensively imports products via cargo ships, introduction of exotic species that could utilize the variety of habitats is a strong possibility. This could allow further novel arboviral introduction events on the island. For example, Ae. albopictus could easily be or have been introduced to the island due to the proximity and intense air and sea traffic with Guam and Mariana Islands where this species is widespread [37]. None were found during this study.
The overwhelmingly predominant mosquito species found on the island was Ae. hensilli. This mosquito was previously speculated to be the vector of DENV during the 1995 outbreak in Yap State as it was the only Aedes (Stegomyia) present on some affected islands [6]. However, like in this outbreak, no isolations were made from field-collected mosquitoes and no arboviruses have ever been reported from this species so incrimination as a vector could not be biologically confirmed. The collection of additional mosquitoes may have allowed virus isolation from field material but repeated strong rainstorms limited the number of adults collected. As in the previous dengue outbreak, Ae. hensilli is the most probable outbreak vector due to its high density, widespread distribution on the island, and its tendency to bite humans. Although transmission studies may have helped clarify vector status, laboratory infection studies reported here further suggest that this is a probable vector due to the high infection rates with ZIKV. While there is an admittedly suboptimal dissemination rate to indicate vector status for ZIKV, there has been documentation of other Aedes (Stegomyia) mosquitoes serving as outbreak vectors even with low susceptibility to infection or dissemination. For example, Ae. aegypti, which has been reported to be relatively resistant to infection to yellow fever virus, has nevertheless been implicated in outbreaks of yellow fever [38]. Vector status of Ae. hensilli for DENV-2 is more difficult to assert based on the laboratory data indicating less than 20% infection rates with virtually no dissemination. However, susceptibility to viruses in at least 2 distinct arboviral genera (flavivirus and alphavirus) suggests that this species could possibly serve as a vector of other medically important arboviruses typically transmitted by Aedes (Stegomyia) species (e.g. yellow fever and chikungunya viruses). It could also serve as a vector of arboviruses in large population centers where the mosquito is found [39]. Aedes hensilli has a limited known distribution consisting of FSM, Palau, and Singapore [39] suggesting that these additional areas might also be potentially at risk due to arboviral pathogens vectored by this species.
Since little is known of the biology or zoonotic transmission of ZIKV, it is also possible that other Scutellaris group species (among others) could be possible vectors of the virus. This is supported by the findings that ZIKV has previously been associated with Ae. africanus [23], [40], [41], Ae. luteocephalus [42], and Ae. aegypti [15] mosquitoes. There are numerous Scutellaris group mosquitoes from island ecologies including Aedes cooki, Aedes polynesiensis, Aedes palauensis, Aedes rotumae, and Aedes scutellaris, and others, some of which have been implicated in arboviral transmission [42]–[48]. The range of the Scutellaris group mosquitoes should be considered as possible vectors of ZIKV in islands of the Pacific and elsewhere.
Aedes hensilli was found to be very susceptible to infection by CHIKV. This finding was interesting as the strain of CHIKV selected was a Central/East African genotype strain associated with the Indian Ocean lineage but not possessing the valine residue at E1that has been linked to increased infectivity in Ae. albopictus [49]. A strain without this mutation was specifically selected to evaluate the susceptibility of Ae. hensilli to a virus that may not have been adapted to alternate Scutellaris group mosquitoes. However, the high degree of susceptibility to CHIKV even without the valine reside at position 226 is not completely unexpected as distinct populations of Ae. albopictus have historically shown significant susceptibility to CHIKV [50]. The ability of Ae. hensilli to be infected with CHIKV again, like with ZIKV, indicates that geographic areas with less well characterized Scutellaris group mosquitoes should consider alternate species to be potential vectors of introduced arboviral diseases.
|
10.1371/journal.pntd.0004663 | Impact of Helminth Infection on the Clinical and Microbiological Presentation of Chagas Diseases in Chronically Infected Patients | Helminth infections are highly prevalent in tropical and subtropical countries, coexisting in Chagas disease endemic areas. Helminth infections in humans may modulate the host immune system, changing the Th1/Th2 polarization. This immunological disturbance could modify the immune response to other infections. The aim of this study is to evaluate the relationship between clinical, microbiological and epidemiological characteristics of Chagas disease patients, with the presence of helminth infection.
A prospective observational study was conducted at Vall d’Hebron University Hospital (Barcelona, Spain). Inclusion criteria were: age over 18 years, diagnosis of Chagas disease, and not having received specific treatment for Chagas disease previously to the inclusion. The study protocol included Chagas disease assessment (cardiac and digestive evaluation, detection of T. cruzi DNA measured by PCR in peripheral blood), and helminth infection diagnosis (detection of IgG anti-Strongyloides stercoralis by ELISA, microscopic examination of stool samples from three different days, and specific faecal culture for S. stercoralis larvae).
Overall, 65 patients were included, median age was 38 years, 75.4% were women and most of them came from Bolivia. Cardiac and digestive involvement was present in 18.5% and 27.7% of patients respectively. T. cruzi PCR was positive in 28 (43.1%) patients. Helminth infection was diagnosed in 12 (18.5%) patients. No differences were observed in clinical and epidemiological characteristics between patients with and without helminth infection. Nevertheless, the proportion of patients with positive T. cruzi PCR was higher among patients with helminth infection compared with patients without helminth infection (75% vs 35.8%, p = 0.021).
We observed a high prevalence of S. stercoralis infection among chronic Chagas disease patients attended in our tropical medicine unit. Strongyloidiasis was associated with significantly higher proportion of positive T. cruzi RT-PCR determined in peripheral blood.
| Helminth infections (viz. Strongyloides stercoralis, hookworms, Ascaris lumbricoides, Trichuris trichiura) are highly prevalent in tropical and subtropical areas, and some of these infections may persist in the human host for many years after leaving the endemic area. It is known that helminth infection in humans may modulate the host immune system. This immunological disturbance could modify the immune response to other infections or the antibody production after vaccination. We prospectively studied a group of patients with chronic Chagas disease, with the aim of evaluate the impact of helminth co-infection in the clinical manifestations and microbiological features of Chagas disease. We observed a high prevalence of helminth infection (mostly due to S. stercoralis infection) among chronic Chagas disease patients attended in our tropical medicine unit. Strongyloidiasis was associated with significantly higher proportion of positive T. cruzi RT-PCR determined in peripheral blood. These data increase the scarce available information to understand the role of PCR techniques in the management of Chagas disease patients. Further studies are needed to deepen and confirm this interesting relationship.
| Chagas disease is a parasitic infection caused by the hemoflagellated protozoan Trypanosoma cruzi. Chagas disease is an endemic disease of Latin America affecting rural and poor population; nevertheless, progressive urbanization and the increase of population mobility during last decades, have made Chagas disease an urban and global disease outside endemic countries: mainly in United States and Spain [1, 2].
After the acute phase of the infection, a subsequent usually asymptomatic chronic stage (or indeterminate phase) takes place during years; after 20–30 years, up to a 30–40% of patients will develop the symptomatic chronic phase, with cardiac and/or digestive involvement [2]. Chagas disease diagnosis in the chronic phase is based on serological tests. Due to the low parasitaemia in this phase, classical direct parasitological tests (microhematocrit, hemoculture, xenodiagnose) are usually negative [3]. Nevertheless, more sensitive tests such as the polymerase chain reaction (PCR) have being developed [4]. The percentage of positive T. cruzi PCR in peripheral blood in patients with Chagas disease in the chronic phase highly varies depending on the study: it ranges from 80% to 90% in studies performed in endemic countries, and is lower in non-endemic countries, ranging from 28% to 66% [5–10]. T. cruzi PCR is not routinely performed in the management of chronic Chagas disease patients, but is becoming very useful in specific situations, such as follow-up in immunosuppressed patients in order to detect reactivation or in clinical trials to detect treatment failures [11–13]. However the role of the T cruzi PCR in the chronic phase of Chagas disease needs to be defined.
Helminth infections (viz. Strongyloides stercoralis, Necator americanus, Ancylostoma duodenale, Ascaris lumbricoides, Trichuris trichiura) are highly prevalent in tropical and subtropical areas, coexisting in Chagas disease endemic areas, and some of these infections may persist in the human host for many years after leaving the endemic area [14]. It is known that helminth infection in humans may modulate the host immune system, changing the Th1/Th2 polarization. This immunological disturbance could modify the immune response to other infections or the antibody production after vaccination [15, 16].
The aim of the present study is to evaluate the relationship between clinical and epidemiological characteristics of chronic Chagas disease patients, with the presence of helminth infection.
The study protocol was approved by the Ethical Review Board of the Vall d’Hebron University Hospital (Barcelona, Spain), and written informed consent was obtained from all patients. Procedures were performed in accordance with the ethical standards laid down in the Declaration of Helsinki as revised in 2000.
This is a prospective observational study performed at the Infectious Diseases Department of the Vall d’Hebron University Hospital, a tertiary hospital included in the International Health Program of the Catalan Health Institute (PROSICS Barcelona, Spain), from March 2014 to February 2015. All adults (over 18 years old) with recently diagnosis of Chagas disease in the chronic or indeterminate form attended during the study period were offered to participate. Exclusion criteria included: previous treatment for Chagas disease or helminth infections, pregnancy or immunosuppression.
Diagnosis of Chagas disease was performed through two positive different serological tests according to WHO recommendations [17]: an enzyme-linked immunosorbent assay (ELISA) with recombinant antigen (Bioelisa Chagas, Biokit, Lliçà d’Amunt, Spain), and an ELISA with crude antigen (Ortho T.cruzi ELISA, Johnson & Johnson, High Wycombe, United Kingdom). Cardiac and digestive involvement was assessed through a clinical symptoms questionnaire, physical examination, 12-lead electrocardiography, chest radiography, and barium enema. Patients were stratified according to the clinical Kuschnir classification for cardiac involvement assessment [18]. Pathologic barium enema was defined by dolichocolon or sigmoid diameter > 6cm (megacolon) [19]. A real time PCR (RT-PCR) to detect T. cruzi DNA in peripheral blood was performed in all patients according to the method described by Piron et al [20].
For helminth infection diagnosis, microscopic examination of stool samples from three different days after concentration techniques using Ritchie’s formalin-ether technique were performed in all patients. A faecal culture for S. stercoralis larvae detection (charcoal culture) was also performed. Moreover, blood cell count to detect presence of eosinophilia (defined as ≥500 cells/mm3 and/or ≥7%), and detection of serum IgG anti-S. stercoralis by ELISA (SciMedx Corporation, Denville, NJ, United States) were conducted.
Definition of helminth infection included: confirmed infections through direct observation, and probable infection (presence of eosinophilia and positive S. stercoralis serology in the absence of other causes of eosinophilia).
Categorical data are presented as absolute numbers and proportions, and continuous variables are expressed as medians and ranges. The χ2 test or Fisher exact test, when appropriate, was used to compare the distribution of categorical variables, and the Mann-Whitney U test for continuous variables. Results were considered statistically significant if the 2-tailed P value was <0.05. SPSS software for Windows (Version 19.0; SPSS Inc, Chicago, IL, USA) was used for statistical analyses.
Overall, 72 patients were included during the study period. Six patients were excluded because they did not complete the study protocol; therefore, 66 patients were analyzed. The median age of patients was 38 (18–67) years, and 50 (75.8%) were women. The vast majority came from Bolivia (64 patients, 97%) and, at the time of the first visit, the median duration of residence in our country was 9 (1–14) years, and 41 (62.1%) patients had traveled again to their countries after arriving to Spain (most of them spent less than 2 months in their countries, and stayed in an urban setting). Cardiac involvement was diagnosed in 12 (18.2%) patients (nine patients in the stage I, and three patients in the stage II of the Kushnir classification respectively). Eighteen (27.3%) patients presented abnormalities in the barium enema: 16 patients with dolichocolon, and 2 patients with megacolon. At the time of Chagas disease diagnosis, T. cruzi RT-PCR in peripheral blood was positive in 28 (42.4%) patients.
Helminth infection was diagnosed in 12 (18.2%) patients: two patients with confirmed infection (one patient with S. stercoralis and another patient with Hymenolepis nana), and 10 patients with probable infection. Patients with helminth infection had a median eosinophil cell count of 500 (100–1200) cells/mm3. Table 1 shows other protozoan parasites observed in the microscopic examination of stool samples. When comparing main clinical and epidemiological characteristics between patients with and without helminth infection, no differences were observed (Table 2). Nevertheless, the percentage of patients with positive T. cruzi RT-PCR was higher in patients with helminth infection compared with those without helminth infection (75% versus 35.2%, p = 0.021).
We prospectively studied 66 adult patients with Chagas disease to evaluate the relationship between microbiological, clinical and epidemiological characteristics with the presence of helminth infection. Positive T. cruzi RT-PCR was more frequent in patients with helminth infection compared with those without helminth infection.
Although the study was carried out in a limited group of Chagas disease patients, clinical and epidemiological characteristics found in the study population were similar to those found in larger studies performed in non-endemic countries: most of them coming from Bolivia, young people, majority of women, and low prevalence of cardiac and digestive involvement [2, 8–10]. Therefore, our study population is representative of the Chagas disease population diagnosed and treated in non-endemic areas.
Helminth infection has been diagnosed in 18.2% of the study population, being strongyloidiasis the most frequent infection (all except from one). The S. stercoralis infection predominance was rather expected, since the median time of residence in Spain in our population was 9 years, thus decreasing the probability of other helminth infections such as Ascaris lumbricoides, hookworms or Trichuris trichiura. S. stercoralis is distributed worldwide, being more frequent in tropical and subtropical areas. High prevalence has been found in Latin American countries where Chagas disease is also endemic, hence co-infection is supposed to be high in this area [21]. Scarce information about the prevalence of strongyloidiasis in Bolivia is available, and it is centered in at risk groups [22]. A study published by Ramos et al showed a S. stercoralis seroprevalence of 44.4% among Bolivian immigrants living in Spain [23]. Although our study was not focused on intestinal protozoa, it is important to note the high prevalence of Blastocystis hominis and Dientamoeba fragilis infections (34.8% and 10.6% respectively) observed in our study population; despite their pathogenicity remains uncertain and controversial, the presence of these parasites may be used as a marker of potential exposure to other pathogenic parasites.
When comparing epidemiological, clinical and microbiological characteristics between patients with and without helminth infection, the first group had statistically significant higher proportion of positive T. cruzi RT-PCR in peripheral blood than the second group (75% and 35.2% respectively). To our knowledge, no previous study has addressed the possible implications of Chagas disease and helminth co-infection in humans. Nevertheless, some interesting studies in animal model have been published with similar findings. Monteiro et al described higher prevalence of T.cruzi-positive blood cultures in golden lion tamarins infected with T. cruzi when they were co-infected with intestinal helminths of the Trichostrongylidae family, which is coherent with the results obtained in our study [24]. Another study performed with T. cruzi infected mice went in depth in this relationship between helminth infection and T. cruzi parasitaemia: no differences in the parasitaemia were found between non co-infected and early co-infected mice (the T. cruzi infection took place 2–4 weeks after Taenia crassiceps infection), however, late co-infected mice (the T. cruzi infection took place 8–12 weeks after Taenia crassiceps infection, when a predominant Th2-type cytokine response is expected) showed significantly higher parasitaemia compared with non co-infected and early co-infected mice [25].
Strongyloides spp infection in the murine model induces a Th2 response and regulatory cytokine induction (IL-10), leading to a suppression of pro-inflammatory cytokines and diminishing Th1 response [26]. These pro-inflammatory cytokines (Th1 response) are present in the acute phase of Chagas disease [27]. Thus, co-infection with different parasites may result in complex interactions, which may lead to altered immunological responses of the host.
The relationship between positive T.cruzi RT-PCR in peripheral blood and helminth infection (mostly strongyloidiasis) found in this study provides highly relevant data to better understand the role of the PCR in the management of Chagas disease patients. Helminth infection could increase the probability of having a positive T. cruzi PCR. Given that current clinical trials that evaluate treatment efficacy in Chagas disease are based on the positivity of T. cruzi PCR, this fact may be relevant. [13, 28]. Further studies are needed to evaluate the impact of treating the helminth infection on the positivity of T. cruzi PCR.
Another issue that has to be taken into account is that almost all patients in our study came from Bolivia. The geographical distribution of the different T. cruzi discrete typing units (DTUs) differs from country to country, which may have impact in the clinical presentation or in the proportion of patients with positive T. cruzi PCR in peripheral blood [29].
This study has some limitations. First of all, as we have mentioned previously, the study has been performed with a relatively small number of patients; nevertheless, the study population is representative of Chagas disease patients attended in Spanish tropical medicine units. Secondly, the diagnosis of strongyloidiasis has relied in serological tests in most of the cases. Although serology is not the gold standard for the S.stercoralis infection diagnosis, previous studies have demonstrated its usefulness [30]. New tests based on molecular biology such as PCR could increase the accuracy of helminth infection diagnosis. Finally, T. cruzi PCR was determined only at one point, which may underestimate the kinetics of the parasite.
In summary, we observed a high prevalence of S. stercoralis infection among chronic Chagas disease patients attended in our tropical medicine unit. Strongyloidiasis was associated with significantly higher proportion of positive T. cruzi RT-PCR determined in peripheral blood. These data increase the scarce available information to understand the role of PCR techniques in the management of Chagas disease patients. Further studies are needed to deepen and confirm this interesting relationship.
|
10.1371/journal.ppat.1003956 | Male-Killing Spiroplasma Induces Sex-Specific Cell Death via Host Apoptotic Pathway | Some symbiotic bacteria cause remarkable reproductive phenotypes like cytoplasmic incompatibility and male-killing in their host insects. Molecular and cellular mechanisms underlying these symbiont-induced reproductive pathologies are of great interest but poorly understood. In this study, Drosophila melanogaster and its native Spiroplasma symbiont strain MSRO were investigated as to how the host's molecular, cellular and morphogenetic pathways are involved in the symbiont-induced male-killing during embryogenesis. TUNEL (terminal deoxynucleotidyl transferase dUTP nick end labeling) staining, anti-cleaved-Caspase-3 antibody staining, and apoptosis-deficient mutant analysis unequivocally demonstrated that the host's apoptotic pathway is involved in Spiroplasma-induced male-specific embryonic cell death. Double-staining with TUNEL and an antibody recognizing epidermal marker showed that embryonic epithelium is the main target of Spiroplasma-induced male-specific apoptosis. Immunostaining with antibodies against markers of differentiated and precursor neural cells visualized severe neural defects specifically in Spiroplasma-infected male embryos as reported in previous studies. However, few TUNEL signals were detected in the degenerate nervous tissues of male embryos, and the Spiroplasma-induced neural defects in male embryos were not suppressed in an apoptosis-deficient host mutant. These results suggest the possibility that the apoptosis-dependent epidermal cell death and the apoptosis-independent neural malformation may represent different mechanisms underlying the Spiroplasma-induced male-killing. Despite the male-specific progressive embryonic abnormality, Spiroplasma titers remained almost constant throughout the observed stages of embryonic development and across male and female embryos. Strikingly, a few Spiroplasma-infected embryos exhibited gynandromorphism, wherein apoptotic cell death was restricted to male cells. These observations suggest that neither quantity nor proliferation of Spiroplasma cells but some Spiroplasma-derived factor(s) may be responsible for the expression of the male-killing phenotype.
| Symbiotic bacteria are ubiquitously associated with diverse insects, and affect their host biology in a variety of ways. In Drosophila fruit flies, infection with Spiroplasma symbionts often causes male-specific embryonic mortality, resulting in the production of all-female offspring. This striking phenotype is called “male-killing”, whose underlying mechanisms are of great interest. Here we investigated Drosophila melanogaster and its native Spiroplasma symbiont strain to understand how the host's molecular, cellular and morphogenetic pathways are involved in the symbiont-induced male-killing. Specifically in Spiroplasma-infected male embryos, pathogenic phenotypes including massive cell death throughout the body and neural malformation were observed. We unequivocally identified that the male-specific cell death preferentially occurs in the embryonic epithelium via the host's apoptotic pathway. Meanwhile, we found that, unexpectedly, the male-specific neural defects occur independently of host's apoptosis, suggesting that at least two different mechanisms may be involved in the Spiroplasma-induced male-killing. Also unexpected was the finding that Spiroplasma titers are almost constant throughout embryogenesis irrespective of sex despite the male-specific severe apoptosis. We serendipitously found Spiroplasma-infected sexual mosaic embryos, wherein apoptosis was associated with male cells, which suggests that some Spiroplasma-derived factor(s) may selectively act on male cells and cause male-killing.
| Symbiotic microorganisms are ubiquitously associated with diverse insects, and affect their host biology in a variety of ways [1], [2]. Some symbionts play important biological roles such as provisioning of essential nutrients to their hosts [3], helping food digestion for their hosts [4], or improving the fitness of their hosts under specific ecological conditions [5]. Other symbionts like Wolbachia, Cardinium and Spiroplasma are generally parasitic rather than beneficial to their hosts, often causing negative fitness effects and also inducing reproductive phenotypes like cytoplasmic incompatibility, male-killing, parthenogenesis or feminization, by which these symbionts are able to spread their own infections into the host populations in a selfish manner [6]–[8].
Members of the genus Spiroplasma, belonging to the class Mollicutes, are wall-less bacteria associated with diverse arthropods and plants [9]. Some Spiroplasma species and strains are known to cause male-killing phenotypes in fruit flies, ladybird beetles and butterflies, wherein infected females produce all-female or female-biased offspring due to male-specific mortality during embryogenesis and/or larval development [10], [11]. Male-killing symbiotic bacteria belonging to Spiroplasma poulsonii [12] have been identified from fruit flies of the genus Drosophila, which are represented by the strains WSRO from D. willistoni, NSRO from D. nebulosa, MSRO from D. melanogaster and others [13], [14].
While the Drosophila-Wolbachia symbiosis represents one of the best-studied model symbiotic systems [8], [15], the Drosophila-Spiroplasma symbiosis has also been well-studied as another model system of infection dynamics [16]–[18], immune regulation [19]–[21], vertical transmission [22], [23] and male-killing expression [24]–[27]. However, molecular and cellular mechanisms underlying the Spiroplasma-induced male-specific embryonic pathology are still not well understood. Histological observations, mosaic analysis and in vitro culturing have suggested that nervous system is among the major target sites of Spiroplasma-induced male-killing in Drosophila embryos [28]–[32]. In D. melanogaster, Spiroplasma-infected mutants deficient in dosage compensation complex genes fail to show male-killing phenotype, indicating that a functional dosage compensation complex is required for expression of the Spiroplasma-induced make-killing [24]. In D. nebulosa infected with its native Spiroplasma strain NSRO, dying male embryos exhibit widespread TUNEL (terminal deoxynucleotidyl transferase dUTP nick end labeling) signals, suggesting possible involvement of host's pathway of programmed cell death or apoptosis [25].
In this study, we performed detailed investigation of the male-killing process during embryogenesis of D. melanogaster infected with its native Spiroplasma strain MSRO. In particular, we focused on host's molecular, cellular and morphogenetic pathways that may potentially be involved in the male-killing phenotype by utilizing the wealth of genetic resources available in D. melanogaster. Our observations unveiled several previously unknown aspects of Spiroplasma-induced male-killing, which include unequivocal demonstration of male-specific up-regulation of apoptotic pathway, identification of embryonic epithelium as the main target of male-specific apoptosis, male-specific malformation of embryonic nervous system independent of apoptosis, and specific killing of male cells in gynandromorphic embryos.
In Spiroplasma-infected female embryos (sexed according to Sex-lethal [Sxl] expression, see Materials and Methods and Fig. S1A and B), TUNEL-positive cells were scarcely found by stage 9 (Fig. 1A), and first appeared at stage 10 in the cephalic region (Fig. 1B, arrowhead). Subsequently, TUNEL-labeled cells spread to the other regions (Fig. 1C and D) and reached a peak level at stages 12–13 (Fig. 1K). These patterns are typical of normal programmed cell death in Drosophila development [33]. Actually, the Spiroplasma-infected female embryos exhibited no substantial differences in the spatiotemporal appearance of TUNEL-positive cells in comparison with uninfected male and female embryos (Fig. 1C; Fig. S1C–E). The Spiroplasma-infected female embryos developed normally (Fig. 1E), and finally emerged as first instar larvae. By contrast, in Spiroplasma-infected male embryos, ectopic TUNEL signals were observed at stage 10: in addition to the dense signals at the cephalic region (Fig. 1G, arrowhead), TUNEL-positive cells were detected throughout the embryonic body (Fig. 1G). Subsequently, the excessive TUNEL signals became more prominent and progressively increased during germ band retraction (Fig. 1H, I and K). From stage 13 and on, the Spiroplasma-infected male embryos started to disintegrate with massive cell death, wherein segmentation and other morphological traits of the embryos became difficult to recognize (Fig. 1J), and finally died. These results indicate that Spiroplasma-infected Drosophila males exhibit ectopic programmed cell death from the early stage of embryonic development. A previous study reported that, in D. nebulosa and its natural Spiroplasma strain NSRO, Spiroplasma-infected male embryos exhibit developmental arrest between stages 12 and 13 with segmentation failure, disintegrated embryonic morphology, and widespread apoptosis as testified by TUNEL staining [25]. Our observations with D. melanogaster and its natural Spiroplasma strain MSRO are highly concordant with the previous observations, suggesting that the same molecular and cellular processes are operating under the symbiont-induced male-specific cell death in the different host species.
Previous studies have demonstrated that normal programmed cell death in Drosophila development requires the activity of Caspase-9-like initiator caspase Dronc (Drosophila Nedd2-like caspase) [34]–[38], and an antibody against cleaved-Caspase-3 can detect the Dronc activity [39]. When probed with the anti-cleaved-Caspase-3 antibody, Spiroplasma-infected male embryos exhibited more immunopositive signals than Spiroplasma-infected female embryos as well as uninfected male and female embryos, and the spatiotemporal patterns of the signals looked similar to those of the TUNEL signals (Fig. 1L). These results suggest that Spiroplasma-infected Drosophila males exhibit ectopic programmed cell death during embryonic development, at least in part by activating the caspase-dependent apoptotic pathway.
Spiroplasma-infected male embryos and female embryos were individually subjected to quantitative PCR targeting Spiroplasma dnaA gene copies. Throughout the embryonic stages examined (from 10 to 13), Spiroplasma titers per embryo remained almost constant, exhibiting no significant differences between male embryos and female embryos (Fig. 1M). By contrast, Spiroplasma titers per host elongation factor 1α 100E (EF1a) gene copy exhibited higher values in male embryos than in female embryos (Fig. S1F), which was attributable to lower EF1α gene titers in male embryos presumably because of male-killing phenotype (Fig. S1G). These results strongly suggest that the ectopic programmed cell death specific to male embryos entails no Spiroplasma proliferation during embryogenesis. Meanwhile, it should be noted that, since quantitative PCR detects not live bacterial cells but DNA molecules, the possibility cannot be ruled out that titers of live Spiroplasma cells may change during embryogenesis and/or between male embryos and female embryos.
Previous studies have demonstrated that programmed cell death in normal Drosophila development requires proapoptotic genes reaper (rpr), head involution defective (hid) and grim, which are collectively termed RHG genes [40]–[42]. In Drosophila mutant H99 deficient in all these genes, apoptotic cell death is almost completely blocked during embryogenesis [40]. RHG proteins bind to Drosophila inhibitor of apoptosis protein 1 (DIAP1) and disrupt its ability to inhibit caspase activity, by which apoptosis is triggered [43]–[46]. When Spiroplasma-infected H99 mutant embryos were subjected to the TUNEL assay, TUNEL-positive cells were observed neither in female embryos nor in male embryos at stages 11 and 12 (Fig. 2A and C–F). These results strongly suggest that the majority of the ectopic programmed cell death observed in Spiroplasma-infected male embryos is induced via host's apoptotic pathway.
In later stages of embryogenesis (stages 13 and 14), some TUNEL-positive cells were detected specifically in Spiroplasma-infected H99 male embryos, although the level of the signals was significantly lower than that in Spiroplasma-infected control male embryos (Fig. 2B and G–J). The residual TUNEL-positive cells may implicate the presence of a minor pathway of Spiroplasma-induced male-specific cell death independent of RHG proteins. Alternatively, the residual TUNEL-positive cells may be due to attenuated clearance of dead cells in the infected embryos.
One of the major mechanisms of apoptosis regulation in Drosophila development is the expression control of RHG genes [44], [47]. Notably, it was reported that several Hox proteins, such as Deformed (Dfd) and Abdominal B (Abd-B), may directly activate the expression of rpr by binding to its enhancer elements [48]. It was also reported that some segment polarity genes may regulate cell survival and death to establish morphological patterns during embryogenesis [49]. However, when we performed immunohistochemical visualization of Hox proteins Antennapedia (Antp), Ultrabithorax (Ubx) and Abd-B, and several segment polarity proteins Wingless (Wg) and Engrailed (En) in Spiroplasma-infected male embryos and female embryos, no sex-related differences were observed in their localization patterns (Fig. S2). These results refute the possibility that Spiroplasma infection may induce ectopic cell death by affecting such developmental signals as Hox genes and segment polarity genes in male embryos.
In Drosophila, atypical protein kinase C (aPKC) localizes to the subapical region (SAR) in the epithelial junctions, thereby establishing apical-basal cell polarity (Fig. 3A) [50], [51]. We visualized embryonic epithelial cells by immunostaining with anti-aPKC antibody. In Spiroplasma-infected female embryos, aPKC showed normal junctional localization and subapical localization in epithelial cells (Fig. 3B–D). In Spiroplasma-infected male embryos, by contrast, junctional localization of aPKC was significantly impaired (Fig. 3E and F), and concentrated TUNEL signals were observed at the level of the subapical region of epithelial cells (Fig. 3G). Notably, intersegmental furrows, which were normally formed in Spiroplasma-infected female embryos (Fig. 3B–D, brackets), became obscure in Spiroplasma-infected male embryos (Fig. 3E–G). These results indicate that embryonic epithelium is among the target tissues wherein Spiroplasma-associated male-specific programmed cell death is induced. Spiroplasma-infected male embryos show ambiguous segments after germ band retraction and die (Fig. 1I) [25]. The loss of segmentation may be relevant to the epithelial damage due to the male-specific apoptosis.
Previous histological observations, mosaic analysis and in vitro culturing have suggested that nervous system is among the major target sites of Spiroplasma-induced male-killing in Drosophila embryos [28]–[32]. In Drosophila embryos, Elav (embryonic lethal abnormal vision) protein is specifically expressed in differentiated neural cells [52], [53]. We performed immunostaining of developing embryos with anti-Elav antibody, which clearly visualized neurons in central nervous system (CNS) and also neurons in peripheral nervous system (PNS) (Fig. 4A). In Spiroplasma-infected male embryos, both CNS and PNS were disorganized (Fig. 4E and F), whereas these neural structures were intact in Spiroplasma-infected female embryos (Fig. 4C and D). It is noteworthy that, despite the remarkable structural disorder, the regions of concentrated TUNEL-positive cells in Spiroplasma-infected male embryos did not agree with the locations of CNS and PNS (Fig. 4E and F). These results suggest that Spiroplasma infection certainly disrupts the formation of normal nervous system specifically in male embryos, but the neural defects are not due to apoptosis of already differentiated neural cells.
In Drosophila embryos, precursor neural cells, or neuroblasts, delaminate basally from neuroectoderm and undergo asymmetric cell division to generate neuroblast itself and ganglion mother cells (GMC). GMCs divide once to give rise to two neurons (Fig. 4B). Neuroblasts express several transcription factors sequentially to generate diverse populations of neurons, one of which is the transcription factor Krüppel (Kr) [54]. Spiroplasma-infected embryos at stage 10 were immunostained with anti-Kr antibody to observe neuroblasts at the beginning of male-specific ectopic cell death. Numerous Kr signals were identified in the neuroblast layer of both Spiroplasma-infected female embryos and male embryos, but they did not co-localize with the TUNEL signals (Fig. 4G and H). These results suggest that Spiroplasma infection disrupts the formation of normal nervous system specifically in male embryos, but the neural defects are unlikely due to direct killing of precursor neural cells via apoptosis.
In the light of these results, it is of focal interest how the epithelial apoptosis and the neural malformation in the Spiroplasma-infected male embryos are interconnected to each other. Are there any causal relationships between them, or do they represent independent processes leading to the embryonic male lethality? To address this question, we observed the development of nervous system in Spiroplasma-infected H99 mutant embryos, in which apoptotic cell death is almost completely blocked [40], by immunostaining with anti-Elav antibody. In Spiroplasma-infected H99 female embryos, CNS and PNS developed normally (Fig. 5A and C), while in Spiroplasma-infected H99 male embryos, remarkable neural malformation was observed (Fig. 5B and D). These results strongly suggest that host's apoptotic pathway is not required for expression of the neural malformation in Spiroplasma-infected male embryos. Plausibly, the male-specific neural malformation may occur independently of the male-specific epithelial apoptosis in the Spiroplasma-infected embryos. If so, the Spiroplasma-induced male-killing entails at least two independent mechanisms: one targets male epithelial cells via host's apoptotic pathway and another targets male nervous system via unknown pathway(s). Alternatively, the defects in neural tissues may somehow influence the organization of adjacent epithelial cells, thereby causing the male-specific epithelial apoptosis secondarily, or vise versa.
In the survey of Spiroplasma-infected embryos, we occasionally identified gynandromorphic embryos with mosaic expression of Sxl (Fig. 6). Sxl-mosaic embryos were observed at stage 12 and later (3/162; 1.9%), but not found at earlier stages (stage 9 to 11; 0/142). In D. melanogaster, spontaneous gynandromorphism has been reported to occur at frequencies between 0.02 to 0.1% in XX zygotes [55]. While symbiont-induced gynandromorphism has been reported from Wolbachia-infected moth, butterfly, planthopper, wasp and wood louse [56]–[60], it requires further verification whether or not the infrequent occurrence of gynandromorphism in D. melanogaster is induced by Spiroplasma infection. Interestingly, apoptotic cells labeled with anti-cleaved-Caspase-3 antibody were restricted to Sxl-negative, presumable male areas in the gynandromorphic embryos (Fig. 6B and C). These results provide strong evidence that Spiroplasma infection selectively acts on male cells but not on female cells, thereby causing male-killing phenotype.
In conclusion, our study unveiled previously unknown molecular and cellular aspects underlying the Spiroplasma-induced male-killing in D. melanogaster. We demonstrated that in Spiroplasma-infected Drosophila embryos (i) host's apoptotic pathway is up-regulated in a male-specific manner, (ii) the male-specific apoptosis mainly targets embryonic epithelial cells, (iii) as previously reported, remarkable neural malformation is observed in male embryos, (iv) however, neither differentiated neural cells nor precursor neural cells exhibit apoptosis in male embryos, (v) the male-specific neural malformation occurs even when host's apoptotic pathway is disrupted, and (vi) therefore, the apoptosis-dependent epidermal cell death and the apoptosis-independent neural malformation may represent different mechanisms underlying the Spiroplasma-induced lethality in male embryos. We also found that (vii) Spiroplasma titers remain almost constant throughout the embryonic development and across male and female embryos, (viii) although at a low frequency (∼2%), gynandromorphic embryos are found in the Spiroplasma-infected embryos, (ix) in these embryos, apoptotic cell death is preferentially observed in male cells, and (x) therefore, neither quantity nor proliferation of Spiroplasma but some Spiroplasma-derived factor(s) selectively acting on host's male cells may be responsible for the expression of male-killing phenotype. These findings highlight complex molecular and cellular interactions in the Spiroplasma-Drosophila symbiosis, and provide invaluable clues to our deeper understanding of the symbiont-induced manipulation of host's development and reproduction.
The following laboratory strains of D. melanogaster were raised at 25°C on a standard cornmeal diet in plastic tubes unless otherwise indicated. Oregon-R (wild-type strain) was provided by Takehide Murata (the Institute of Physical and Chemical Research, RIKEN). Sxl-Pe-EGFP G78b [61] and Df(3L)H99, kniri-1, pp/TM3, Sb1 [40] were obtained from the Bloomington Stock Center, USA, and the Drosophila Genetic Resource Center (DGRC) at Kyoto Institute of Technology, Japan, respectively. After tetracycline treatment for curing bacterial infections as described [62], the fly strains were infected with the Spiroplasma strain MSRO by hemolymph injection as described [16]. The MSRO-containing hemolymph was collected from naturally infected D. melanogaster strain Ug-SR derived from Uganda [63], which was gifted by John Jaenike (University of Rochester, USA). Since the MSRO-infected fly strains produce all-female offspring, these strains were maintained by supplying males from corresponding uninfected fly stocks. H99 mutant strain was re-balanced with GFP-tagged balancer (TM3, ActGFP, Ser1) and homozygous mutant individuals were identified by immunostaining with anti-GFP antibody.
Spiroplasma-infected female flies within three days after eclosion were allowed to mate with male flies for three days in plastic tubes. These insects were kept with grape juice agar plates for embryo collection. Embryos at different developmental stages were dechorionated, fixed in 4% formaldehyde and heptane for 20 min, and devitellinized by vigorously shaking in heptane and methanol. In this study, female-specific expression of Sxl, the master regulator of the sex determination system in Drosophila [64], [65], was used for sexing of embryos (Fig. S1A and B). The following primary antibodies were used for immunohistochemical staining: mouse anti-Sex-lethal (M18; 1∶20 dilution; Developmental Studies Hybridoma Bank [DSHB]), rabbit anti-cleaved-Caspase-3 (1∶100; Cell signaling Technology, Inc.), rat anti-Dα-Catenin (DCAT-1; 1∶10; DSHB), rabbit anti-PKCζ (C-20; 1∶100; Santa Cruz Biotechnology, Inc.), mouse anti-Elav (9F8A9; 1∶20; DSHB), rat anti-Elav (7E8A10; 1∶20; DSHB), chicken anti-GFP (GFP-1020; 1∶400; Aves Labs, Inc.), mouse anti-Wingless (4D4; 1∶20; DSHB), mouse anti-Engrailed/Invected (4D9; 1∶20; DSHB), mouse anti-Antp (8C11; 1∶20; DSHB), mouse anti-Ubx (FP3.38; 1∶20; DSHB), mouse anti-Abd-B (1A2E9; 1∶20; DSHB), and guinea pig anti-Krüppel (#573; 1∶300; Asian Distribution Center for Segmentation Antibodies) [66]. Fluorochrome-labeled secondary antibodies were purchased from Jackson ImmunoResearch Laboratories, Inc. and Molecular Probes. Nuclear DNA was stained with SYTOX Orange Nucleic Acid Stain (S-11368; 1∶20,000; Molecular Probes). TUNEL staining was performed to detect DNA fragmentation associated with programmed cell death or apoptosis [67], [68] by In Situ Cell Death Detection Kit, TMR red (Roche Applied Science) as described [69]. Images were taken on a confocal microscope (Zeiss LSM 5 Pascal and LSM 510 META). Serial Z-sections of confocal images were compiled to create projection images (maximum intensity projection) unless otherwise described, using a custom macro in ImageJ software (National Institutes of Health, USA).
After dechorionization, developmental stages and sexes of Sxl-Pe-EGFP embryos were determined visually under a stereoscopic fluorescent microscope (Leica M165 FC). Each of 12 embryos of both sexes, which were collected at stage 10, 11, 12 or 13, was individually subjected to DNA extraction using QIAamp DNA mini kit (Qiagen). The DNA samples were subjected to real-time quantitative PCR using SYBR Green (Takara) and Mx3000P qPCR system (Stratagene) essentially as described [16]. Spiroplasma titers in terms of dnaA gene copies were quantified using the primers 5′-TGA AAA AAA CAA ACA AAT TGT TAT TAC TTC-3′ and 5′-TTA AGA GCA GTT TCA AAA TCG GG-3′. The copy numbers of the host EF1α gene were also quantified using the primers 5′-TTA ACA TTG TGG TCA TTG GCC A-3′ and 5′-CTT CTC AAT CGT ACG CTT GTC G-3′. The reaction mixture consisted of 1× AmpliTaq Gold buffer, 1.5 mM MgCl2, 0.2 mM each of dATP, dGTP, dCTP and dUTP, 0.3 µM each of the forward and reverse primers, 1/100,000 SYBR green, and 0.02 U/µl AmpliTaq Gold DNA polymerase (Applied Biosystems). PCR was performed under a temperature profile of 95°C for 10 min followed by 38 cycles of 95°C for 1 min, 60°C for 1 min and 72°C for 1 min. The data were statistically analyzed using the software R version 2.15.0 (R Foundation for Statistical Computing). Multiple comparison was performed using non-parametric Kruskal-Wallis test followed by Scheffe test.
Quantitative analyses of TUNEL signals were performed by custom R scripts with EBImage package for image processing [70]. Briefly, maximum projections of confocal slices stained with TUNEL and Sxl antibody were obtained. Images showing lateral view of the embryos were selected for further processing. Embryonic regions were determined by binarization of projected images of Sxl. Processed images were visually checked and signals derived from other objects (e.g. flanking embryos, backgrounds etc.) were manually blacked out to obtain the area of each embryo precisely (mask image). TUNEL signals were also binarized and signals inside the mask image were calculated by image integration. For normalization, TUNEL signals in each embryo were divided by the embryonic area calculated by mask image. Statistical test was performed using Wilcoxon rank sum test.
|
10.1371/journal.pntd.0002658 | Parasite-Antigen Driven Expansion of IL-5− and IL-5+ Th2 Human Subpopulations in Lymphatic Filariasis and Their Differential Dependence on IL-10 and TGFβ | Two different Th2 subsets have been defined recently on the basis of IL-5 expression – an IL-5+Th2 subset and an IL-5−Th2 subset in the setting of allergy. However, the role of these newly described CD4+ T cells subpopulations has not been explored in other contexts.
To study the role of the Th2 subpopulation in a chronic, tissue invasive parasitic infection (lymphatic filariasis), we examined the frequency of IL-5+IL-4+IL-13+ CD4+ T cells and IL-5−IL-4 IL-13+ CD4+ T cells in asymptomatic, infected individuals (INF) and compared them to frequencies (Fo) in filarial-uninfected (UN) individuals and to those with filarial lymphedema (CP).
INF individuals exhibited a significant increase in the spontaneously expressed and antigen-induced Fo of both Th2 subpopulations compared to the UN and CP. Interestingly, there was a positive correlation between the Fo of IL-5+Th2 cells and the absolute eosinophil and neutrophil counts; in addition there was a positive correlation between the frequency of the CD4+IL-5−Th2 subpopulation and the levels of parasite antigen – specific IgE and IgG4 in INF individuals. Moreover, blockade of IL-10 and/or TGFβ demonstrated that each of these 2 regulatory cytokines exert opposite effects on the different Th2 subsets. Finally, in those INF individuals cured of infection by anti-filarial therapy, there was a significantly decreased Fo of both Th2 subsets.
Our findings suggest that both IL-5+ and IL-5−Th2 cells play an important role in the regulation of immune responses in filarial infection and that these two Th2 subpopulations may be regulated by different cytokine-receptor mediated processes.
| Th2 cells are CD4+ T cells that produce a unique set of cytokines - IL-4, IL-5 and IL-13. Th2 cells are commonly associated with allergies, asthma and helminth infections. A common helminth infection that infects over 120 million people worldwide is lymphatic filarial infection caused by filarial parasites. We show here data that filarial infection is associated with the expansion of two types of Th2 cells, one which produces IL-4 and IL-13 alone without IL-5 and the other which produces all three cytokines. Interestingly, while the former subset is associated with the levels of antibodies - IgG4 and IgE; the latter is associated with the presence of eosinophilia in filarial infected individuals. In addition, these subsets appear to be modulated differently by the immunoregulatory cytokines - IL-10 and TGFβ. Therefore, our study highlights a novel regulation of Th2 cells and suggests that the Th2 compartment is quite heterogeneous in phenotype with possible functional consequences.
| Th2 cells were initially characterized as expressing the cytokines – IL-4, IL-5 and IL-13 [1]. Although Th2 cells can express a variety of other cytokines, these three cytokines remain the hallmark Th2 cytokines. Each Th2 cytokine has a well-defined and relatively specific function. While IL-4 is the major driving force behind Th2 differentiation, IgE class switching and alternative macrophage activation, IL-13 is an important mediator of goblet cell hyperplasia, mucus secretion and airway hyper reactivity [2]. IL-5, in contrast, acts primarily on eosinophils and their precursors in the bone marrow to induce enhanced production, survival and activation of these cells [3]. While Th2 cells have generally been considered a homogenous population, recent reports provide evidence for subpopulations within the Th2 lineage [4]. Two of the main subsets identified recently are the IL-5 expressing Th2 subset (co-expressing IL-4, IL-5 and IL-13) and the non IL-5 expressing Th2 subset (co-expressing only IL-4 and IL-13) [4]. Since the three established Th2 cytokines each play a non-redundant role in allergic disease pathology, it was postulated that these Th2 subsets might play an important role in allergic diseases. Indeed, IL-5+Th2 cells have been found in greater frequencies (Fo) in patients with eosinophilic gastrointestinal disease, while peanut allergy was found to be associated with higher Fo of IL-5−Th2 cells [5].
The canonical host immune response seen in human filarial infections is of the Th2 type and involves the production of cytokines – IL-4, IL-5, IL-9, IL-10 and IL-13, the antibody isotypes – IgG1, IgG4 and IgE, and expanded populations of eosinophils and immunoregulatory monocyte [6]. Human filarial infection is known to be associated with down regulation of parasite-specific Th1 responses and T cell proliferation and but with augmented Th2 responses [7]. Thus, in human lymphatic filariasis (LF) patent filarial infection is associated with an antigen – specific expansion of Th2 cells (mostly defined by IL-4 expression) and enhanced production of IL-4 and IL-13 [7]. However, antigen – driven IL-5 production has been shown to be diminished in patently infected individuals [8], [9] in some studies. Similarly, although protective immunity to filarial infections in mice is dependent primarily on IL-4, IL-5 does not appear to play a role in resistance to infection [10], [11]. Hence, filarial infections provide a natural setting in which to explore the differential role (if any) of Th2 subsets. We wanted to explore the hypothesis that Th2 subsets would be differentially regulated in asymptomatic infection compared to uninfected or individuals with chronic pathology.
We, therefore, examined the Th2 cytokine expression patterns of CD4+ T cells in clinically asymptomatic patently infected (INF) individuals, filarial-uninfected endemic normal (UN) individuals, and in previously infected patients with filarial lymphedema (CP) both at homeostasis and following stimulation with parasite and control antigens. We show that active human LF is characterized by a significant enhancement in the Fo of both spontaneously-expressed and parasite antigen – driven IL-5− and IL-5+Th2 cells. We show that the Fo of IL-5+Th2 subpopulation is positively correlated with peripheral eosinophil and neutrophil numbers in filarial infections, while the IL-5−Th2 cells are strongly positively related to the levels of parasite specific IgE and IgG4. We also show that these Th2 subpopulations appear to have differing programs of regulation by both IL-10 and TGFβ in filarial-infected individuals and that definitive treatment (and subsequent cure) of this infection causes reversion of the CD4+ Th2 subpopulation expansion to normal levels.
All individuals were examined as part of natural history studies approved by Institutional Review Boards of both the National Institutes of Allergy and Infectious Diseases and the National Institute for Research in Tuberculosis (NCT00375583 and NCT00001230), and informed written consent was obtained from all participants.
The study took place in Tamil Nadu, South India where from 1000 subjects in the area, 70 subjects who were willing to provide blood were included. These subjects originate from hospital and surroundings and therefore may represent very different populations. The 70 individuals comprised of 32 clinically asymptomatic, infected (hereafter INF) individuals, 23 individuals with filarial lymphedema or elephantiasis (hereafter CP) and 15 uninfected, endemic normal (hereafter UN) individuals (Table 1). This was primarily a retrospective study using previously collected samples that had been fixed and cryopreserved. The samples obtained post-treatment were collected prospectively.
The study individuals were recruited from individuals attending the Filariasis Clinic of the Government General Hospital, Chennai and from community screening of the areas where the individuals reside (Puliayonthope and Ponneri areas). All CP individuals were circulating filarial antigen negative by both the ICT filarial antigen test (Binax, Portland, ME) and the TropBio Og4C3 enzyme-linked immunosorbent assay (ELISA) (Trop Bio Pty. Ltd, Townsville, Queensland, Australia), indicating a lack of current active infection. The diagnosis of prior filarial infection was made by history and clinical examination as well as by positive Brugia malayi antigen (BmA) -specific IgG4. BmA-specific IgE, IgG4 and IgG ELISA were performed exactly as described previously [12]. All INF individuals tested positive for active infection by both the ICT filarial antigen test and the TropBio Og4C3 ELISA and had not received any anti-filarial treatment prior to this study. All INF individuals were treated with a standard dose of diethylcarbamazine (DEC) and albendazole and follow – up blood draws were obtained one year later from 16 individuals. Among the 32 INF individuals, 25 were used for whole blood culture with parasite antigens and 7 were used for cytokine blocking studies. Among the 32 treated individuals, we were able to follow up only 16 individuals out of which 9 were circulating antigen negative (Cured) and 7 remained circulating antigen positive (Not-cured). These individuals were used for post-treatment analysis. We also used 7 of the 32 INF individuals exclusively for performing cytokine blocking studies. All UN individuals were circulating filarial antigen negative and without any signs or symptoms of infection or disease. We have selected individuals from the same community and socio-economic backgrounds to account for exposure and socio-economic status.
Saline extracts of B. malayi adult worms (BmA) and microfilariae (Mf) were used for parasite antigens and mycobacterial PPD (Serum Statens Institute, Copenhagen, Denmark) was used as the control antigen. Final concentrations were 10 µg/ml for BmA, Mf and PPD. Endotoxin levels in the BmA was <0.1 EU/ml using the QCL-1000 Chromogenic LAL test kit (BioWhittaker). Phorbol myristoyl acetate (PMA) and ionomycin at concentrations of 12.5 ng/ml and 125 ng/ml (respectively), were used as the positive control stimuli.
Leukocyte counts and differentials were performed on all individuals using the Act-5 Diff hematology analyzer (Beckman Coulter). The INF individuals had higher eosinophil counts but did not differ from the other two groups in any of the other hematological parameters (data not shown).
Whole blood cell cultures were performed to determine the Fo of intracellular cytokine-producing cells.. Briefly, whole blood was diluted 1∶1 with RPMI-1640 medium, supplemented with penicillin/streptomycin (100 U/100 mg/ml), L-glutamine (2 mM), and HEPES (10 mM) (all from Invitrogen, San Diego, CA) and placed in 12-well tissue culture plates (Costar, Corning Inc., NY, USA). The cultures were then stimulated with BmA, Mf, PPD, PMA/ionomycin (P/I) or media alone in the presence of the co-stimulatory reagent, CD49d/CD28 (BD Biosciences) at 37°C for 6 hrs. FastImmune Brefeldin A Solution (10 µg/ml) (BD Biosciences) was added after 2 hours. After 6 hours, whole blood was centrifuged, washed using cold PBS, and then 1× FACS lysing solution (BD Biosciences, San Diego, CA, USA) was added. The cells were fixed using cytofix/cytoperm buffer (BD Biosciences, San Diego, CA, USA), cryopreserved, and stored at −80°C until use. The same procedure was used for both prospectively collected as well as for retrospectively stored samples. For cytokine neutralization experiments (n = 7), whole blood from INF individuals was cultured in the presence of anti-IL-10 (5 µg/ml) or anti-TGFβ (5 µg/ml) or isotype control antibody (5 µg/ml) (R& D Sytems) for 1 h following which BmA and brefeldin A was added and cultured for a further 5 h.
The cells were thawed and washed with PBS first and PBS/1% BSA later and then stained with surface antibodies for 30–60 minutes. Surface antibodies used were CD3 - Amcyan, CD4 - APC-H7 and CD8 - PE-Cy7 (all from BD Biosciences). The cells were washed and permeabilized with BD Perm/Wash™ buffer (BD Biosciences) and stained with intracellular cytokines for an additional 30 min before washing and acquisition. Cytokine antibodies used were IL-4 FITC, IL-5 APC and IL-13 PE (all from BD Pharmingen. Flow cytometry was performed on a FACS Canto II flow cytometer with FACSDiva software v.6 (Becton Dickinson). The lymphocyte gating was set by forward and side scatter and 100,000 gated lymphocyte events were acquired. Data were collected and analyzed using Flow Jo software. All data are depicted as Fo of CD4+ T cells expressing cytokine(s). Frequencies following media stimulation are depicted as baseline Fo while Fo following stimulation with antigens or PMA/ionomycin are depicted as net Fo (with baseline Fo subtracted).
Data analyses were performed using GraphPad PRISM v6.0 (GraphPad Software, Inc., San Diego, CA, USA). Geometric means (GM) were used for measurements of central tendency. Comparisons were made using either the Kruskal-Wallis test with Dunn's multiple comparisons (unpaired comparisons) or the Wilcoxon signed rank test (paired comparisons). Correlations were calculated using the Spearman rank test.
We first measured the spontaneously expressed and antigen - stimulated Fo of CD4+ T cells expressing IL-4 and IL-13 but not IL-5 (CD4+IL4+IL13+IL5−) as well as those expressing IL-4, IL-13, and IL-5 (CD4+IL4+IL13+IL5+) in INF (for representative flow plot, see Fig. S1) and contrasted these with the Fo of these subpopulations in UN and CP. As shown in Figure 1A, INF individuals exhibit significantly higher Fo of IL-5−Th2 cells in response to the parasite antigens BmA (2.2 fold) and Mf (2.1 fold) but not at baseline nor following PPD or PMA/ionomycin stimulation in comparison to UN individuals, INF individuals also had significantly higher Fo of IL-5+Th2 cells at baseline (1.2 fold) and following BmA (2.6 fold) and Mf (1.9 fold) (but not PPD or PMA/ionomcyin) stimulation in comparison to UN (Figure 1B). Similarly, INF individuals exhibit significantly higher Fo of IL-5−Th2 cells at baseline (1.8 fold) and in response to BmA (3 fold) and Mf (2.2 fold) but not following PPD or PMA/ionomycin stimulation in comparison to CP individuals (Figure 1A). Finally, INF individuals also had significantly higher Fo of IL-5+Th2 cells both at baseline (1.2 fold) and following stimulation with BmA (2.1 fold) and Mf (1.7 fold) (but not with PPD or PMA/ionomcyin) in comparison to CP (Figure 1B). Thus, filarial infection in this population was associated with expanded Fo of antigen – stimulated Th2 subpopulations, perhaps indicating a role for these cells in filarial infection and in the prevention of overt pathology.
To determine whether the IL-5− and IL-5+Th2 subsets were associated with differential functions in filarial infections, we examined the relationship between the baseline Fo of IL-5− and IL-5+Th2 cells and peripheral eosinophil, neutrophil and basophil numbers in INF individuals (n = 32). As shown in Figure 2A, IL-5+ (but not IL-5−) Th2 cells had a significantly positive correlation between their Fo ex vivo and the absolute eosinophil count (r = 0.4667, p = 0.0071) as determined by Spearman rank correlation. Similarly, as shown in Figure 2B, IL-5+ (but not IL-5−) Th2 cells exhibited a significant positive association with the absolute neutrophil count (r = 0.4115, p = 0.0193). In contrast, both IL-5− and IL-5+Th2 cells showed no correlation with the absolute basophil numbers in INF individuals (Figure 2C).
We next examined the relationship between the Fo of IL-5− and IL-5+Th2 cells ex vivo and IgE and IgG4 levels in INF individuals (n = 32). As shown in Figure 3A, IL-5− (but not IL-5+) Th2 cells exhibited a significantly positive correlation between baseline Fo and the BmA – specific IgE levels (r = 0.7717, p<0.0001) as determined by Spearman rank correlation. Similarly, as shown in Figure 3B, IL-5− (but not IL-5+) Th2 cells exhibited a significant positive association with the BmA – specific IgG4 levels (r = 0.4115, p = 0.0193).
To determine the role of IL-10 and TGFβ in the modulation of Th2 subpopulations in INF, we measured the frequency of IL-5−Th2 cells and IL-5+Th2 cells following short-term stimulation with the parasite antigen BmA in the presence or absence of anti-IL-10 or anti-TGFβ neutralizing antibody in INF individuals (n = 7). As shown in Figure 4A, both IL-10 and TGFβ neutralization resulted in significantly decreased Fo of IL-5−Th2 cells in INF individuals (2. 3 and 1.9 fold respectively). In marked contrast, as shown in Figure 4B, both IL-10 and TGFβ blockade resulted in significantly increased Fo of IL-5+Th2 cells following BmA stimulation (1.5 and 1.3 fold respectively). Thus, both IL-10 and TGFβ play an important role in the modulation of Th2 subset Fo in filarial infections.
To determine the role of antigen – persistence in the regulation of Th2 subsets in filarial infections, we measured the frequency of IL-5−Th2 and IL-5+Th2 cells in a subset of INF individuals (n = 9), who had been treated with anti-filarial chemotherapy and as a result had eliminated infection (as demonstrated by the absence of circulating filarial antigen) or those who had been treated but continued to harbor infection (n = 7). As shown in Figure 5A, treatment of filarial infection and consequent cure resulted in significantly decreased Fo of IL-5−Th2 cells upon filarial antigen stimulation (1.4 fold for BmA and 1.3 fold for Mf) when compared to their pre-treatment levels. Similarly, treatment also resulted in significantly decreased Fo of IL-5+Th2 cells following filarial antigen stimulation (1.7 fold for BmA and 1.5 fold Mf) (Figure 5B). However, individuals who were treated but did not eliminate infection continued to exhibit significantly increased Fo of IL-5− and IL-5+Th2 cells in response to filarial antigens compared to pre-treatment Fo (Figure 5). Interestingly, this effect was specific to the filarial – antigen stimulated Fo of CD4+ Th2 subpopulations as neither responses to PPD nor to PMA/ionomycin were significantly among the groups. Thus, the antigen – driven expansion of Th2 subpopulations appears to be dependent on the continued presence of circulating filarial antigen (an indicator of active infection).
Th2 responses are considered to be the hallmark of helminth infection and are indeed required for host resistance to a variety of helminths in animals [6]. The three prototypical Th2 cytokines – IL-4, IL-5 and IL-13 have all been shown to play important but non – redundant roles in helminth immunity [13]. In addition, the recent explosion of interest in CD4+ T cell subpopulations and the availability of polychromatic cytokine staining has helped define heterogeneity within the Th2 compartment [4]. Thus, two major subsets of Th2 cells were recently described – an IL-5 expressing Th2 population, which is thought to play an important role in eosinophilic inflammation and an IL-5 non expressing Th2 population, which is thought to play an important role in other forms of allergic inflammation [4]. Moreover, it has been demonstrated that IL-5+ and IL-5−Th2 cells represent more and less highly differentiated Th2 subpopulation, respectively [14]. However, the role of these subsets in helminth infections is not known.
The induction of prototypical Th2 response with high IL-4, IL-5 and IL-13 secretion has long been considered to be the hallmark of active infection in human LF [7]. However, not all studies have consistently shown a predominant prototypical Th2 response in filarial infections. A recent study in Mali suggested that patent, long-standing filarial infection is associated with expanded adaptive regulatory T cell cells rather than an expansion of classical Th2 cells environment [15]. Previous studies have reported a down regulation of IL-5 upon parasite stimulation [8], [9], [16]. In addition, the role of Th2 responses in protection from or in the pathogenesis underlying the disease associated with LF has not been well characterized either. We therefore utilized two sets of comparisons to help elucidate the role of Th2 subsets in human filarial infections – (1) comparisons of Th2 subsets in INF and UN individuals, (2) comparisons of these subsets in INF and CP individuals. We were able to demonstrate that both Th2 subsets are expanded preferentially in active, subclinical infection but not in filarial disease (without active infection). Our data on Ag – induced expression of CD4+ Th2 cell subpopulations also reveal interesting facets of T cell driven immune regulation in filarial infection and disease. First, the alterations in the CD4+ Th2 cell cytokine repertoire is filarial – antigen specific since the since these alterations in Th2 subpopulation Fo were primarily observed only in response to the filarial-derived BmA and Mf antigens but not to PPD nor in response to polyclonal stimulation. Second, the importance of antigen – persistence is clearly illustrated by our data on a small subset of individuals who cleared infection following treatment and were therefore proven to be filarial antigen negative. The follow up data on these individuals indicate a clear reversion to the normal/homeostatic levels of Th2 subset populations. On the other hand, the different Th2 subpopulations continue to expand in a control group of individuals, who also received treatment but failed to clear infection. Therefore, the expansion of antigen - specific Th2 subsets is closely associated with the presence of parasite antigen in vivo.
Th2 cells are thought to play a counter-regulatory role in a variety of infectious and inflammatory conditions to offset pathology and promote tissue repair and wound healing mechanisms [17]. Th2 responses are considered to be fundamentally important in protection against the development of pathology both because of their ability to ameliorate Th1 induced inflammatory responses and because of their propensity to promote wound healing and tissue repair [18]. For example, IL-5 and IL-13 have pro-fibrotic activity and, in addition, IL-4 and IL-13 are critical mediators of alternative activation of macrophages. Our study of the Ag – stimulated expansion of Th2 subpopulations reveals a significant association of these cells with asymptomatic infection, confirming a previously reported association [19]. By contrasting these Th2 subpopulations in clinically asymptomatic patients to those with filarial lymphedema (CP) we may be able to infer a role for these expanded Th2 subsets in protection from the development of clinical pathology. Moreover, our data on the lower levels of Th2 subpopulations in CP supports the suggestion that unchecked parasite-specific Th1/Th17 cells may contribute to the pathological process in LF.
Our data reveal clear distinctions in the relationship between IL-5− and IL-5+Th2 cells and expansion of innate leukocyte populations in filarial infections. Eosinophils are considered to be important innate effectors in immunity to helminth infections and have been shown to play a role in protection against S. mansoni and other helminths [20], [21]. Similarly, basophils are known to act as effectors to promote parasite killing during challenge infections of immunized animals, perhaps through antibody dependent mechanisms [22], [23], while neutrophils have been demonstrated to attack helminth larvae in response to IL-4 and IL-5 [24], [25]. Our study implicates the IL-5+Th2 subpopulation in this innate defense mechanism by promoting the recruitment and/or activation of eosinophils and neutrophils. Our study also demonstrates an important association of IL-5−Th2 cells in promoting Th2 associated (IgE and IgG4) antibody responses in filarial infection. All helminth infections are characterized by the induction of antibody isotypes of the class – IgE and IgG4 (IgG1 in mice), that are largely dependent on the IL-4 [26].
Not only did we assess the expansion of these Th2 subpopulations, we also examined the mechanisms regulating the expression of these cytokines in these two subpopulations. Since IL-10 and the TGFβ in chronic infections are known to play a role in modulating T cell expression of cytokines in filarial infections [19], we examined the Fo of IL-5+ and IL-5−Th2 cells following either IL-10 or TGFβ blockade during in vitro stimulation with filarial antigen. Our data, through preliminary due to the small number of samples able to obtained, show clear differences in the modulation of the Th2 subsets. We demonstrate that the expansion of IL-5−Th2 cells is dependent on both IL-10 and TGFβ since blockade of these cytokines significantly reduces the frequency of IL-5−Th2 cells. Conversely, both IL-10 and TGFβ appear to impair the induction of the IL-5+Th2 subset. While it has been previously shown that IL-5 expression in Th2 cells is limited to the effector memory subset whereas IL-4 is expressed in both central and effector memory subsets [4], this finding that IL-10 and TGFβ signaling may be critical to Th2 subpopulation expansion provide new insight into the interrelationship between the IL-5+ and IL-5−Th2 subpopulations and provides new avenues for the study of filarial-specific immune regulation and protection from immune-mediated pathology in LF.
In summary, our study examines in depth the CD4+ Th2 cell subset repertoire in a chronic parasitic infection and sheds light on the role of these subsets in both the regulation of immune responses in active infection and in the pathogenesis of filarial lymphedema. While we have not performed longitudinal studies to define the development of pathology in filarial infection and this was a study using a combination of previously (pre-treatment) and prospectively (post-treatment) collected samples with inherent potential limitations with respect to bias, our strategy of contrasting immune responses in individuals with early or subclinical disease and those with late or clinical disease yields important information on the association of Th2 subsets in pathogenesis. However, the potential drawbacks in the study include potential bias in using both prospectively and retrospectively collected samples and lack of rigorous controls in eliminating potential confounders including socio-economic status of individuals in the study. The lack of proper information in the study area regarding the prevalence of the different clinical groups, adds to the problem of potential bias which means that our conclusions cannot be generalized. In addition, while we have demonstrated the presence of Th2 subsets in filarial infections, disease association is not formal proof of function and the elucidation of function needs to be explored in the future. Nevertheless, our study clearly defines and important association of filarial infection with heightened expansion of Th2 cells suggesting that these subsets play an important role in infection.
|
10.1371/journal.ppat.1007328 | Function of BriC peptide in the pneumococcal competence and virulence portfolio | Streptococcus pneumoniae (pneumococcus) is an opportunistic pathogen that causes otitis media, sinusitis, pneumonia, meningitis and sepsis. The progression to this pathogenic lifestyle is preceded by asymptomatic colonization of the nasopharynx. This colonization is associated with biofilm formation; the competence pathway influences the structure and stability of biofilms. However, the molecules that link the competence pathway to biofilm formation are unknown. Here, we describe a new competence-induced gene, called briC, and demonstrate that its product promotes biofilm development and stimulates colonization in a murine model. We show that expression of briC is induced by the master regulator of competence, ComE. Whereas briC does not substantially influence early biofilm development on abiotic surfaces, it significantly impacts later stages of biofilm development. Specifically, briC expression leads to increases in biofilm biomass and thickness at 72h. Consistent with the role of biofilms in colonization, briC promotes nasopharyngeal colonization in the murine model. The function of BriC appears to be conserved across pneumococci, as comparative genomics reveal that briC is widespread across isolates. Surprisingly, many isolates, including strains from clinically important PMEN1 and PMEN14 lineages, which are widely associated with colonization, encode a long briC promoter. This long form captures an instance of genomic plasticity and functions as a competence-independent expression enhancer that may serve as a precocious point of entry into this otherwise competence-regulated pathway. Moreover, overexpression of briC by the long promoter fully rescues the comE-deletion induced biofilm defect in vitro, and partially in vivo. These findings indicate that BriC may bypass the influence of competence in biofilm development and that such a pathway may be active in a subset of pneumococcal lineages. In conclusion, BriC is a part of the complex molecular network that connects signaling of the competence pathway to biofilm development and colonization.
| Pneumococcal biofilms occur in chronic otitis media, chronic rhinosinusitis, and nasopharyngeal colonization. These biofilms are an important component of pneumococcal epidemiology, particularly in influencing transmission, maintenance of asymptomatic colonization, and development of disease. The transcriptional program initiated via signaling of the competence pathway is critical for productive biofilm formation and is a strong contributor of pneumococcal infection and adaptation. In this study, we have identified BriC, a previously uncharacterized peptide that serves as a bridge between the competence pathway and biofilm development. We show that briC is induced by ComE, the master regulator of competence, and promotes biofilm development. Moreover, our studies in the murine model demonstrate that BriC is a novel colonization enhancer. Our studies of briC regulation capture an instance of genomic plasticity, where natural variation in the briC promoter sequence reveals the existence of an additional competence-independent regulatory unit. This natural variation may be able to modify the extent to which competence contributes to biofilm development and to nasopharyngeal colonization across different pneumococcal lineages. In summary, this study introduces a colonization factor and reveals a molecular link between competence and biofilm development.
| Bacteria form sessile communities termed biofilms, where they interact with each other to engage in collaborative and/or competitive behaviors [1]. In Streptococcus pneumoniae (pneumococcus), these cell-cell interactions are commonly mediated by secreted peptides that interact with both producing and neighboring cells of the same species, and induce changes in gene regulation that result in altered phenotypes [2]. These dynamic pneumococcal biofilms occur in chronic otitis media, chronic rhinosinusitis and nasopharyngeal colonization [3–8].
The ability to form biofilms is a critical component of pneumococcal disease [9]. Biofilms serve as reservoirs for acute infections [10]. In the middle ear, cells released from a biofilm are thought to be responsible for recurrent episodes of infection [4]. Bacterial cells released from nasopharyngeal biofilms can seed pneumococcal transmission between individuals by being incorporated into nasal shedding. Alternatively, these cells can disseminate to tissues causing mild to severe diseases, such as otitis media, pneumonia, and sepsis [10]. Pneumococcal cells released from biofilms display increased virulence relative to their planktonic or biofilm counterparts, suggesting that chronic biofilms set the stage for the stimulation of a virulence program activated upon the dispersal of cells [11]. Moreover, pneumococci in a biofilm display decreased susceptibility to antibiotics, and are recalcitrant to treatment [6]. Thus, biofilms are an important component of pneumococcal epidemiology in transmission, maintenance of asymptomatic colonization, and development of disease.
The transcriptional program required for the initiation and the growth of pneumococcal biofilms has been the subject of numerous investigations. It is clear that at least two quorum sensing (QS) signal transduction pathways are critical for biofilm development: competence and Lux [7], [12–15]. The competence pathway has been the subject of intense investigation for decades [16–28]. Competence is activated by a classic two-component system (TCS) where the extracellular competence stimulating peptide (CSP, encoded by comC) binds to the surface exposed ComD histidine kinase receptor, inducing its autophosphorylation and the subsequent transfer of the phosphate group to its cognate regulator, ComE [23], [29]. Activation of the competence pathway leads to increased expression of 5–10% of the pneumococcal genome in two main waves of gene expression [21], [26]. The first wave of induction is carried out directly by ComE; it upregulates a subset of competence genes (early genes) that include comAB, comCDE, as well as the alternative sigma factor, comX. The second wave of competence induction is regulated by ComX; it leads to an increase in the levels of at least 80 genes (late genes), that subsequently modulate important phenotypes such as transformation, metabolism, fratricide and biofilm formation [16], [26], [30]. This competence program is upregulated during biofilm mode of growth in vitro, during interactions with human epithelial cells, and in lungs and brain after intranasal and intracranial challenges respectively in murine infection models [7], [12], [31]. Importantly, in cell culture models, comC is required for biofilm development [12], [14]. Thus, activation of the competence pathway is important for productive biofilm formation and critical for pneumococcal infection and adaptation.
The Lux QS system also plays a role in biofilm formation. In this system, Lux QS is controlled by the AI-2 autoinducer, which is secreted and sensed by both Gram-positive and Gram-negative species. LuxS is a node in the regulation of competence, fratricide, and biofilm development [15], [32]. Lux upregulates competence via ComD and ComX [13]. It contributes to bactericidal activity via upregulation of the choline binding murein hydrolase (CbpD). Through lysis, this bactericidal activity increases the levels of extracellular DNA, which is a key ingredient in the extracellular polymeric substance (EPS) that makes up the biofilm. Thus, the competence and Lux systems provide the molecular framework to coordinate multi-cellular bacterial communities to form and develop robust biofilms during infection.
Whereas the role of competence signaling in biofilm development is well established, the molecules that connect competence to biofilms are poorly understood [3], [7], [15], [33]. In this study, we identify one such molecule that links competence and biofilms. We characterize the gene encoding BriC (Biofilm regulating peptide induced by Competence), a novel colonization factor in the competence pathway. Levels of briC are regulated by ComE, and briC promotes biofilm development and nasopharyngeal colonization. Further, genomic analysis of briC reveals polymorphisms in its promoter, where a subset of strains encode a RUP (for repeat unit of pneumococcus) sequence, which leads to additional and CSP-independent expression of briC.
We have identified the gene encoding a putative secreted peptide that is co-regulated with competence (spd_0391 (D39); spr_0388 (R6); sp_0429 (TIGR4)). Based on the results presented in this manuscript, we have termed it Biofilm-regulating peptide induced by Competence (BriC). BriC was identified in our previously described in silico screen designed to capture cell-cell communication peptides in the pneumococcal genome [34]. The known double glycine (Gly-Gly) streptococcal peptides are exported and proteolytically processed by dedicated ABC transporters that recognize N-terminal sequences with the Gly-Gly leader peptide (LSXXELXXIXGG)[23]. In our previous work, we identified novel secreted pneumococcal peptides using computational analysis to search for proteins with N-termini that contain a Gly-Gly leader. Our input set consisted of the alleles of two exported Gly-Gly peptides, the signaling molecule CSP and the bacteriocin BIP [23], [35]. Our output consisted of a position dependent probability matrix that captures the length and positional variability at each residue of the Gly-Gly motif. When we searched for this motif in a database of sixty streptococcal genomes, we defined a predicted secretome consisting of twenty-five sequence clusters, one of which corresponds to BriC [34].
To identify genes co-regulated with briC, we performed transcriptional studies using a NanoString probe set that reports on the abundance of the briC transcript as well as transcripts encoding a subset of pneumococcal regulators and cell wall proteins. We assessed the levels of briC transcript in vitro and in vivo. In vitro expression was measured by screening RNA extracted from mid-log planktonic cultures of a laboratory strain (R6-derivative (R6D)). In vivo expression was evaluated by analysis of middle-ear effusions recovered from chinchillas infected with a clinical PMEN1 strain. The mRNA levels of briC were positively associated with comC and comE in vitro (strain R6D: R2 = 0.61 and 0.79, respectively) and in vivo (strain PN4595-T23: R2 = 0.92 and 0.88, respectively). It is noteworthy that changes in the expression of genes in this locus were observed in the studies that first documented the competence regulon [21], [26]. Specifically, Peterson and colleagues observed changes in briC levels, however the association between briC and CSP was below the statistical threshold [26]. Further, Dagkessamanskaia and colleagues observed an upregulation in the gene downstream of briC, predicted to be in the same operon [21]. Given that in strains R6, R6D, and D39, this downstream gene is truncated, this study does not explore the function of the downstream gene. In summary, our gene expression analysis suggests that briC is induced by competence.
To directly test whether briC is a competence-regulated peptide, we employed a fusion of the briC promoter to the lacZ reporter (R6 PbriC-lacZ). Stimulation of the signal transduction system that initiates competence by addition of CSP1 led to an induction of the β-galactosidase activity by about twenty-five-fold (Fig 1). Induction of the briC promoter was specific to the CSP pherotype encoded by strain R6. The β-galactosidase activity was observed upon addition of CSP1, the CSP pherotype from strain R6, but not upon addition of the non-cognate CSP2 pherotype (Fig 1). Thus, we conclude that briC is a competence-responsive gene.
Our in silico analysis of the briC promoter in strains R6 and R6D revealed the presence of a ComE-binding site. ComE binds a well-defined sequence consisting of two imperfect direct repeats of nine nucleotides separated by a gap of twelve or thirteen base pairs [36]. Our analysis of the putative briC promoter across pneumococcal strains revealed an excellent match to the ComE-binding box (Fig 2A). To further investigate the association between ComE and briC, we tested whether CSP-induction of briC requires ComE. We compared the CSP-induction of briC in a wild-type (R6D WT) strain to that of an isogenic comE-deletion mutant (R6DΔcomE), using qRT-PCR analysis. In WT cells, the addition of CSP triggered a significant increase in levels of briC at 10 minutes post-addition, with levels slowly decreasing by 15 minutes (Fig 2B). This trend follows the temporal pattern observed for the levels of comE that has been associated with genes under direct controls of ComE [21], [26]. In contrast, the transcript levels of briC were unaffected by CSP addition in the ΔcomE strain, indicating that the expression of briC requires ComE (Fig 2B). These results strongly suggest that briC is directly regulated by ComE. In addition, ComE plays a critical role in controlling transformation, thus we investigated whether briC impacts transformation efficiency (S1 Result and S1 Fig). We find that absence of briC leads to only a minor decrease in transformation efficiency.
To investigate whether expression of briC plays a role in biofilm development, we compared biofilm formation across WT (R6D WT), briC deletion mutant (R6DΔbriC), and briC complemented (R6DΔbriC::briC) strains grown on an abiotic surface at 24h and 72h post-seeding. No difference was observed in biofilm biomass and thickness at 24h post-seeding, suggesting that expression of briC does not contribute to early stages of biofilm formation (Fig 3A and 3B). In contrast, at 72h post-seeding, ΔbriC biofilms displayed significantly reduced biomass and thickness when compared to WT (Fig 4A and 4B). Further, biofilms with ΔbriC::briC cells restored the WT phenotype at this time-point (Fig 4A and 4B). The indistinguishable biofilm parameters of WT and ΔbriC cells at 24h post-seeding suggests that there is no fitness-related growth difference between the strains and indicates that the biofilm defect is biologically relevant. These findings suggest that briC contributes to late biofilm development.
To investigate the prevalence of briC, we investigated its distribution across the genomes of pneumococcus and related streptococci. To place the distribution in the context of phylogeny, we used a published species tree generated from a set of fifty-five genomes [34], [38] (S1 Table). The genomes encompass thirty-five pneumococcal genomes that span twenty-nine multi-locus sequence types as well as eighteen serotypes and nontypeable strains; eighteen genomes from related streptococcal species that also colonize the human upper respiratory tract, namely S. pseudopneumoniae, S. mitis, and S. oralis; and finally, two distantly related S. infantis strains as an outgroup. In this set, all the pneumococcal genomes encode briC, and there are two highly similar alleles (labeled allele 1A and 1B, Figs 5 and 6). Further, we identified four additional alleles in the related streptococci (Fig 5, S1 File). Next, we extended this analysis to a set of 4,034 pneumococcal genomes available in the pubMLST database (these correspond to all the genomes with at least 2Mb of sequence) [39]. In total, 98.5% (3,976 out of 4,034) of these genomes encode a briC allele, suggesting briC is highly prevalent across pneumococcal strains. We find that alleles 1A and 1B are prominent in this larger set, with 1,824 and 1,187 representatives respectively. After manual curation, we retrieved nineteen distinct briC alleles across these pneumococcal genomes (Fig 6). Six of the polymorphic residues are located in the putative leader sequence. The conserved region of the leader sequences corresponds to the amino acids preceding the Gly-Gly [23], [34], thus the polymorphisms at the N-terminal end of the BriC sequence are not predicted to influence export. One polymorphic residue replaces the glycine in the Gly-Gly motif with a serine; it seems probable that this variation may influence processing and/or export. In addition, position -2 from the C-terminus encodes either an alanine or a threonine. This variation is at the C-terminus, predicted to be the functional end of the molecule, such that the difference in size and polarity at this residue is likely to impact structure or binding of BriC to its targets. The briC gene is induced by competence, so we investigated whether there is a correlation between CSP pherotypes and alleles encoding BriC. However, we did not find these to be associated (S2 Table). Finally, to investigate the distribution of briC in other species and genera, we used BLASTp to search the non-redundant database [40]. We find that BriC homologs are present in strains of related streptococci, S. pseudopneumoniae, S. mitis and S. oralis, but we did not identify homologues in more distant species. The phylogenetic distribution of briC supports a conserved role across pneumococci and a subset of related streptococcal species.
Our analysis of the promoter region of briC in our curated set of genomes reveals that a subset of strains encode for a 107bp insertion within the region upstream of briC (Fig 5). The additional nucleotides are located after the ComE-binding site and before the transcriptional start site, and correspond to a repeat unit in pneumococcus (RUP) sequence [41], [42]. RUP is an insertion sequence derivative with two clear variants, which may still be mobile [42]. The RUP sequence upstream of briC corresponds to RUPB1.
In our curated genomes, the long RUP-encoding promoter is present in multiple strains, including those from the clinically important PMEN1 and PMEN14 lineages (Fig 5). Using our expanded database of 4,034 pneumococcal genomes, we determined that the vast majority of the PMEN1 and PMEN14 genomes encode a long promoter. Specifically, of the manually curated subset, 99.4% of PMEN1 strains and 100% of PMEN14 strains encode the long promoter. The high prevalence of the long promoter in these lineages suggests that this form was present in the ancestral strains from these lineages and/or provides a fitness advantage in these genomic backgrounds.
To investigate how this genomic difference influences briC expression, we generated a LacZ reporter strain. The 263bp upstream of briC from the PMEN1 strain, PN4595-T23, were fused to lacZ to produce the PbriClong-lacZ reporter, and its reporter activity was compared to that of the PbriC-lacZ generated with the fusion of 159bp upstream of briC obtained from strain R6. To control for the possibility that the influence of the RUP sequence might be strain-dependent, we tested these reporter constructs in the R6 and the PMEN1 backgrounds, in both the absence and presence of CSP treatment (Fig 7A and 7B). The presence of RUP dramatically increased the basal levels of briC in the absence of CSP, and this increase was observed in both R6 and PMEN1. Furthermore, both constructs were induced upon addition of the cognate CSP. These findings suggest that the RUP sequence serves as an expression enhancer; it increases the levels of briC transcripts and this increase is CSP-independent. Thus, in some lineages, briC appears to be under the control of both CSP-dependent and CSP-independent regulation.
Next, we investigated the biological impact of the natural variations in the briC promoter on biofilm development. It has been well established that competence promotes biofilm development. Specifically, deletion of the comC (encodes CSP) and comD (encodes histidine kinase of competence TCS) genes lead to a reduction in in vitro biofilms in strains D39 and TIGR4 [7], [12]. In this study, we have established that briC also promotes biofilm development (Fig 4A and 4B), and that the RUP-containing long promoter serves as an expression enhancer, increasing basal expression of briC to levels higher than the CSP-dependent induction observed with the short promoter (Fig 7). Thus, we hypothesized that expression of briC from the long promoter may bypass the impact of competence in biofilm development.
In concurrence with previous work, we observed that a strain with a comE deletion (R6DΔcomE) displays a reduction in biofilm biomass and thickness relative to the WT strain (Fig 8A and 8B). ComE influences the expression of numerous genes. To determine whether the biofilm defects were primarily due to its impact on briC induction, we tested a construct with a disruption of the ComE-binding box in the briC promoter (R6DΔbriC::PbriCShuffled ComE-box-briC). This strain displays a significant reduction in biofilm biomass and thickness relative to the WT strain (Fig 8A and 8B). Moreover, no difference was observed in the biofilm parameters for both of these mutants, suggesting that the absence of briC expression is a contributor to the in vitro biofilm defect in the comE deletion mutant. Next, we determined that a strain with increased basal levels of briC driven by the RUP-containing long promoter (R6DΔcomE::PbriClong-briC) fully rescued the biofilm defect observed in R6DΔcomE (Fig 8A and 8B). In addition, increased expression of briC in the wild type background (R6D WT::PbriClong-briC) did not lead to a significant increase in biofilm biomass and thickness relative to the wild type (Fig 8A and 8B). Together, these data strongly suggest that briC is a critical molecular link between competence and biofilm formation, and that natural variations in the briC promoter are physiologically relevant.
Since BriC is associated with the competence pathway and is able to rescue the biofilm defects associated with competence signaling, we investigated whether competence associated transporters play a role in exporting BriC. In pneumococcus, the ComAB and BlpAB C39-peptidase transporters export peptides with a Gly-Gly leader [43–45]. These transporters recognize the N-terminal leader of target sequences, and cleave these sequences at the Gly-Gly motif [45], [46]. In strains R6 and R6D, BlpAB is not functional due to a frameshift mutation that leads to an early stop codon [47]. Thus, we hypothesized that as a Gly-Gly peptide co-expressed with genes of the competence pathway, BriC may be exported via the ComAB transporter. We tested this hypothesis in two ways. First, we measured whether deletion of comAB influences the ability of a strain with competence-independent expression of briC to rescue the ΔcomE-biofilm defect. Second, we compared secretion of a BriC reporter construct in a WT strain with that in a comAB deletion mutant.
Our biofilm data suggests that ComAB plays a role in transporting BriC. At 72h post-seeding, a comE/comAB-double deletion mutant strain expressing briC from the RUP-encoding long promoter (R6DΔcomEΔcomAB::PbriClong-briC) displayed a biofilm with reduced biomass and thickness when compared to the equivalent construct in only a comE-deletion background (R6DΔcomE::PbriClong-briC) (Fig 9A and 9B). However, the biofilm levels were not reduced to the levels observed in the ΔcomE strain. These results suggest that under these conditions, ComAB may not be the only transporter that contributes to the export of BriC.
To further elucidate the role of ComAB in the export of BriC, we employed the HiBiT tag detection system, which was recently used to detect secretion of BlpI [44]. The HiBiT tag corresponds to an 11-residue peptide. The assay works by addition of an inactive form of luciferase (LgBit) to the extracellular milieu. When both LgBit and HiBiT combine, they generate bioluminescence [48], [49]. To study BriC transport, we fused the putative BriC leader sequence to the HiBiT tag and expressed this reporter under the control of the native (short) briC promoter in WT and ΔcomAB R6-strains. We measured the extracellular bioluminescence produced by this reporter both in the presence and absence of CSP (Fig 9C, S3 Table). In the absence of CSP, the levels of secreted HiBiT resembled that of background (WT cells without HiBiT), consistent with very low expression of N-terminal BriC-HiBiT as well as low expression of the ComAB transporter. In the WT background, upon addition of CSP, N-terminal BriC-HiBiT is induced and the extracellular level of HiBiT is significantly increased, consistent with HiBiT export. In the ΔcomAB background, upon addition of CSP, N-terminal BriC-HiBiT is induced and the extracellular levels of HiBiT also increase. However, the increase in the level of extracellular HiBiT observed in ΔcomAB strain is significantly less than that in the WT strain, consistent with a severe reduction in HiBiT export in the absence of comAB. Combined, these results strongly suggest that ComAB serves as a major transporter for BriC.
During nasopharyngeal colonization, pneumococci form biofilms and upregulate the competence pathway. Thus, we investigated the role of briC in nasopharyngeal colonization using an experimental murine colonization model. Our in vitro investigations have been performed using strain R6D, which is defective in colonization due to the absence of a capsule [50]. Thus, we performed colonization experiments with the serotype 2 strain, D39, which is the ancestor of strain R6 [51]. Mice were colonized with D39 WT, the briC-deletion mutant (D39ΔbriC) or the briC-complemented (D39ΔbriC::briC) strains. Comparison of the number of bacteria in nasal lavages immediately after inoculation revealed that mice in the three cohorts received the same number of bacteria. In contrast, nasal lavages at three and seven days post-inoculation revealed decreased levels of D39ΔbriC relative to WT in the nasal wash (Fig 10A). Furthermore, the WT levels were restored in the complemented strain (Fig 10A). These findings indicate that briC encodes a novel colonization factor.
In in vitro biofilms, briC links competence to biofilms. First, disruption of the ComE- binding box in the briC promoter led to a biofilm defect similar to that observed in a ΔcomE strain. Second, overexpression of briC driven by the long version of the promoter was found to restore the competence-dependent defect in in vitro biofilm development. Thus, we investigated the behavior of these strains in the pneumococcus colonization model. We found that the strain with a disruption of the ComE-binding box (D39ΔbriC::PbriCShuffled ComE-box-briC) within the briC promoter was defective for colonization, the decreased bacterial counts resembled those in the D39ΔcomE strain (Fig 10B). These findings suggest that briC is a substantial mediator of the role of ComE on colonization. Further, addition of this long briC promoter to ΔcomE cells (D39ΔcomE::PbriClong-briC) partially rescues the colonization defect of the D39ΔcomE strain. That is, the numbers of bacterial cells of strain D39ΔcomE::PbriClong-briC recovered from the nasal lavages at both three and seven days post-inoculation were significantly higher than the numbers of D39ΔcomE cells recovered, but less than that of the D39 WT (Fig 10B). Finally, the overexpression of briC in the WT background (D39 WT::PbriClong-briC) does not impact colonization. Thus, we conclude that BriC is a contributor to the competence-induced stimulation of nasopharyngeal colonization observed in strain D39. Further, natural variations leading to a long briC promoter appear to dampen the impact of competence in colonization.
An important component of pneumococcal pathogenesis is its ability to form complex biofilm structures. Pneumococci in a biofilm mode of growth display decreased sensitivity to antibiotics and increased resistance to host immune responses [6]. These properties make the bacteria recalcitrant to treatment and highlight the need to better understand the molecular mechanisms that drive biofilm development. Activation of the competence pathway is critical for biofilm development. Previous in vitro studies have demonstrated that while cell-adherence and early biofilm formation are competence-independent, an intact competence system is required in the later stages of biofilm development. It was shown that the competence pathway positively influences structure and stability of late stage biofilms [12]. However, the molecules downstream of competence activation by ComDE that regulate biofilm development remain poorly understood. In this study, we present BriC, a previously uncharacterized peptide, that we show is regulated by competence and plays a role in promoting biofilm development and nasopharyngeal colonization.
We have presented extensive evidence that briC is a competence regulated gene. We have shown that induction of briC is triggered by addition of CSP and requires ComE. Further, we have also shown that the briC promoter encodes the consensus ComE-binding box, and that briC expression follows the temporal pattern described for genes directly regulated by ComE. Previous studies have used microarray analysis to identify pneumococcal genes differentially regulated upon CSP stimulation [21], [26] and have categorized these genes into two main groups—early genes regulated by ComE or late genes regulated by the alternative sigma factor, ComX. In the study by Peterson and colleagues, briC was found to be upregulated in a pattern consistent with early genes [26]. However, the upregulation was not found to be statistically significant, and this study is the first validation of briC as a competence-regulated peptide.
We have provided evidence that briC stimulates biofilm development on abiotic surfaces and promotes nasopharyngeal colonization in a murine model. These findings are consistent with studies that show that pneumococcal biofilms contribute to colonization. Colonization of the upper respiratory tract is a requisite for pneumococcal dissemination to distant anatomical sites and subsequent disease [10]. These sessile communities serve as a source of pneumococcal cells with an activated virulence-associated transcription program. That is, when compared to cells originating from a planktonic mode of growth, those originating from a biofilm mode of growth are more likely to cause disease upon infecting other tissues [11]. In this manner, increased biofilm development likely heightens the risk of disease. Biofilms and competence are also associated with transformation efficiency. We have observed a mild but significant decrease in the transformation efficiency of briC-deletion mutants relative to WT R6D cells (S1 Fig). Finally, colonization of the upper respiratory tract is also a reservoir for pneumococcal transmission. Transmission occurs when cells migrate from the nasopharynx of one host to that of another. Thus, BriC’s contribution to colonization may influence both disease severity and transmission.
While it has been established that CSP contributes to biofilm development, the competence-dependent genes that regulate biofilm development are not well understood [7], [12]. Our finding that increased levels of briC can fully rescue biofilm defects from a comE deletion mutant in vitro, and partially rescue its colonization defects in vivo suggests that briC expression may bypass the requirement for competence in biofilm development. ComE is a key regulator of competence whose activity is required to regulate approximately 5–10% of the genome, and as such deletion of comE is expected to have substantial global consequences [21], [26]. In this context, it is noteworthy that overexpression of one gene (briC) in the comE-deletion mutant was able to improve colonization in the murine model. These findings strongly suggested that BriC is a molecular link between competence, biofilm development, and colonization.
Our data suggests that many strains have multiple inputs to the regulation of briC. Shared across all strains is the regulation by ComE, the key regulator of the competence pathway. Competence is responsive to environmental cues, such as changes in cell density, pH, mutational burden in cells, and exposure to antibiotics [16], [52–54]. Conversely, competence is inhibited by the degradation of CSP via the activity of the CiaHR TCS and the serine protease, HtrA [55], [56]. Factors altering competence will also alter briC levels due to its competence-dependent induction. Our comparative genomics suggest that a subset of pneumococcal lineages may encode an additional briC-regulatory element. Specifically, the briC promoter differs across strains, in that a subset of lineages encodes a long promoter with a RUP sequence (PbriClong) and has higher basal levels of briC expression. This long promoter is constitutively active, even when competence is off.
The long promoter is encoded in the vast majority of strains from the PMEN1 lineage (Spanish-USA) and the PMEN14 (Taiwan-19F) lineages. These lineages are prominent in the clinical setting; they are multi-drug resistant and pandemic [57–59]. This additional competence-independent regulation of the long promoter may provide promoter-binding sites for additional regulators or reflect consequences of positional differences for the existing promoter binding sites. Our biofilm and colonization experiments suggest that encoding the long briC promoter has functional consequences. We conclude that the response of briC to competence is ubiquitous, but that additional lineage-specific factors influence briC regulation and downstream phenotypic consequences.
We propose a model where briC encodes a signaling molecule with a role in biofilm development and colonization. First, the transcription of briC is induced by ComE through competence signal transduction pathway in all lineages, and possibly by additional regulator(s) in a subset of lineages. Once this Gly-Gly peptide is produced, we propose that it is exported through ABC transporters, a process in which ComAB plays a prominent role. Based on a bioinformatic comparison with other Gly-Gly peptides we suggest that BriC is cleaved into its active form (BRIC) during export. It is tempting to speculate that BRIC is a new member of the expanding set of pneumococcal secreted peptides that signal to neighboring cells promoting population-level behaviors. In this era of emerging antibiotic resistance, it is imperative that we test the potential of alternative strategies to inhibit bacterial carriage and disease. One such strategy is to specifically target bacterial communities and population-level behaviors. In that regard, molecules such as BriC present promising alternatives to be used as targets for discovery of novel drugs and therapeutic interventions.
Four wild-type (WT) Streptococcus pneumoniae strains were used for this experimental work. The majority of studies were performed on a penicillin-resistant derivative of R6 (R6D); this strain was generated from a cross where parental strain R6 was recombined with Hungary19A and the recombinant was selected for penicillin resistance [60]. The briC allele in R6D is identical to the allele present in the parental R6. This laboratory strain is non-encapsulated and does not colonize mice, thus mice colonization experiments were performed with the serotype 2 D39 strains (GenBank CP000410)[61]. The D39 strain contains the same briC allele as is present in the R6D strain, which has been used for most of the work in this study. Finally, for a representative of PMEN1, we used the carriage isolate, PN4595-T23 (GenBank ABXO01) graciously provided by Drs. Alexander Tomasz and Herminia deLancastre [62].
Colonies were grown from frozen stocks by streaking on TSA-II agar plates supplemented with 5% sheep blood (BD BBL, New Jersey, USA). Colonies were then used to inoculate fresh Columbia broth (Remel Microbiology Products, Thermo Fisher Scientific, USA) and incubated at 37°C and 5% CO2 without shaking. When noted, colonies were inoculated into acidic Columbia broth prepared by adjusting the pH of Columbia broth to 6.6 using 1M HCl. Acidic pH was used to inhibit the endogenous activation of competence.
The mutant strains (R6DΔbriC and PN4595ΔbriC) were constructed by using site-directed homologous recombination to replace the region of interest with erythromycin-resistance gene (ermB) or kanamycin-resistance gene (kan), respectively (S4 Table). The kan and spectinomycin-resistance gene (aad9) were used to construct ΔcomE strains in R6D and PN4595-T23 respectively. Briefly, the transformation construct was generated by assembling the amplified flanking regions and antibiotic resistance cassettes. ~2kb of flanking regions upstream and downstream of the gene of interest was amplified from parental strains by PCR using Q5 2x Master Mix (New England Biolabs, USA). The antibiotic resistance genes, kan and aad9 were amplified from kan-rpsL Janus Cassette and pR412, respectively (provided by Dr. Donald A. Morrison), and ermB was amplified from S. pneumoniae SV35-T23. SV35-T23 is resistant to erythromycin because of the insertion of a mobile element containing ermB [62]. These PCR fragments were then assembled together by sticky-end ligation of restriction enzyme-cut PCR products. The deletion mutant in R6D is an overexpressor of the downstream peptide (spr_0389).
The briC complement and overexpressor strains were generated using constructs containing the CDS of briC along with either its entire native promoter region or overexpressing promoter respectively, ligated at its 3’ end with a kanamycin resistance cassette. The promoters used to overexpress briC included either the constitutive amiA promoter, or PbriClong. These were assembled with the amplified flanking regions by Gibson Assembly using NEBuilder HiFi DNA Assembly Cloning Kit. The construct was introduced in the genome of R6D downstream of the bga region (without modifying bga), a commonly employed site [63]. Primers used to generate the constructs are listed in S5 Table. Like R6DΔbriC, R6DΔbriC::briC is also an overexpressor of the downstream peptide (spr_0389), which is annotated as a pseudogene in strains R6D, R6 and D39 (Fig 2A). The expression of spr_0389 remains unchanged in the mutant and the complement (expression of spr_0389 was induced by 5-6-fold in both the mutant and the complement relative to WT).
The R6DΔcomE::PbriClong-briC strain was constructed by replacing comE with spectinomycin resistant cassette in the R6D PbriClong-briC strain. The comAB-deletion mutant in a briC overexpressor R6D genomic background strain (R6DΔcomAB::PbriClong-briC) was constructed by transforming the R6D::PbriClong-briC strain with the genomic DNA of ADP226. ADP226 is a strain from the D39 genomic background with comAB replaced by erythromycin resistance cassette. To make the construct, the flanking regions and erythromycin resistance cassette were amplified, and then assembled together by sticky-end ligation of restriction enzyme-cut PCR products. The construct was then transformed into D39 ADP225 and selected on Columbia blood agar supplemented with 0.25 μg mL−1 erythromycin (S4 Table).
The briC promoter region was modified by shuffling the ComE-binding box (R6DΔbriC::PbriCShuffled ComE-box-briC). The ComE-binding box was shuffled using PCR by amplifying from R6DΔbriC::briC and introducing the shuffled sequence (CAGACCAGTTAGTCTAGGATAGAGCTTAAG) into the primers. The resulting amplicons were assembled using Gibson Assembly. The modified construct was transformed into R6DΔbriC strain in the region downstream of the bgaA gene.
The D39 briC deletion mutant (D39ΔbriC), briC complemented (D39ΔbriC::briC), comE deletion mutant (D39ΔcomE), briC overexpressor in comE deletion background (D39ΔcomE::PbriClong-briC), and briC expressed from a promoter with a shuffled ComE-binding box (R6DΔbriC::PbriCShuffled ComE-box-briC) strains were generated by transformation with the corresponding constructs amplified from R6D into strain D39.
Chromosomal transcriptional lacZ-fusions to the target promoters were generated to assay promoter activity. These lacZ-fusions were generated via double crossover homologous recombination event in the bgaA gene using modified integration plasmid pPP2. pPP2 was modified by introducing kan in the multiple cloning site, in a direction opposite to lacZ. The modified pPP2 was transformed into E. coli TOP10. The putative briC promoter regions were amplified from R6 and PN4595-T23 strains, and modified to contain KpnI and XbaI restriction sites, which were then assembled in the modified pPP2 plasmid by sticky-end ligation of the enzyme digested products. These plasmids were transformed into E. coli TOP10 strain, and selected on LB (Miller’s modification, Alfa Aesar, USA) plates, supplemented with ampicillin (100μg/ml). These plasmids were then purified by using E.Z.N.A. Plasmid DNA Mini Kit II (OMEGA bio-tek, USA), and transformed into pneumococcal strains R6 and PN4595-T23 and selected on Columbia agar plates supplemented with kanamycin (150μg/ml).
For all bacterial transformations to generate mutants, target strains (R6D or D39) were grown in acidic Columbia broth, and 1μg of transforming DNA along with 125μg/mL of CSP1 (sequence: EMRLSKFFRDFILQRKK; purchased from GenScript, NJ, USA) was added to them when the cultures reached an OD600 of 0.05, followed by incubation at 37°C. After 2 hours, the treated cultures were plated on Columbia agar plates containing the appropriate antibiotic; erythromycin (2μg/ml), or kanamycin (150μg/ml). Resistant colonies were cultured in selective media, and the colonies confirmed using PCR. Bacterial strains generated in this study are listed in S4 Table.
For transformation efficiency experiments, R6D strain was grown in acidic Columbia broth until it reached an OD600 of 0.05. At this point, number of viable cells was counted by plating serial dilutions on TSA-blood agar plates. Transformations were carried out by adding either 100ng or 500ng of transforming DNA in the media supplemented with 125μg/mL of CSP1 and incubated at 37°C for 30mins. For transforming DNA, we used either genomic DNA or PCR products. The donor DNA contained spectinomycin-resistance gene (aad9) in the inert genomic region between spr_0515 and spr_0516. This construct was generated in PN4595-T23, specR, followed by its amplification and transformation into R6D and Taiwan-19F strains (Sp3063-00). The genomic DNA was extracted from Taiwan-19F, specR strain. The purified linear DNA was an amplimer of the region from R6D. After 30 minutes, the cultures were plated on Columbia agar plates containing spectinomycin (100μg/ml), incubated overnight, and colonies were counted the next day.
RNA extraction consists of sample collection, pneumococcal cell lysis, and purification of RNA. For qRT-PCR analysis, the strains (R6D and R6DΔcomE) were grown to an OD600 of 0.3 in acidic Columbia broth, followed by CSP1 treatment for 0, 10, or 15 minutes. For in vitro transcriptomic analysis using NanoString Technology, the R6D strain was grown to an OD600 of 0.1 in Columbia broth (in one experimental set, the samples were grown in sub-lethal concentration of penicillin (0.8μg/ml) for an hour). RNA was collected in RNALater (Thermo Fisher Scientific, USA) to preserve RNA quality and pelleted. For the in vivo experiments, the RNA was extracted from middle-ear chinchilla effusions infected with PN4595-T23 and PN4595-T23ΔcomE strains and preserved by flash freezing the effusion. In all the samples, the pneumococcal cell lysis was performed by re-suspending the cell pellet in an enzyme cocktail (2mg/ml proteinase K, 10mg/ml lysozyme, and 20μg/ml mutanolysin), followed by bead beating with glass beads (0.5mm Zirconia/Silica) in FastPrep-24 Instrument (MP Biomedicals, USA). Finally, RNA was isolated using the RNeasy kit (Genesee Scientific, USA) following manufacturer’s instructions. For analysis with the NanoString, which does not require pure DNA, the output from the RNeasy kit was loaded on the machine without further processing. For analysis using qRT-PCR, contaminant DNA was removed by treating with DNase (2U/μL) at 37°C for at least 45 mins. The RNA concentration was measured by NanoDrop 2000c spectrophotometer (Thermo Fisher Scientific, USA) and its integrity was confirmed on gel electrophoresis. The purity of the RNA samples was confirmed by the absence of a DNA band on an agarose gel obtained upon running the PCR products for the samples amplified for gapdh.
nCounter Analysis System from NanoString Technology provides a highly sensitive platform to measure gene expression both in vitro and in vivo, as previously described [64]. Probes used in this study were custom-designed by NanoString Technology, and included housekeeping genes gyrB and metG, which were used as normalization controls. 5μL of extracted RNA samples were hybridized onto the nCounter chip following manufacturer’s instructions. RNA concentration ranged from 80-200ng/μL for in vivo samples, and between 60-70ng total RNA for in vitro samples. A freely available software from manufacturers, nSolver, was used for quality assessment of the data, and normalization. The RNA counts were normalized against the geometric mean of gyrB and metG [65], [66]. The 16S rRNA gene is not optimal for normalization in the NanoString, as the high abundance of this transcript packs the field of view. Pearson’s Correlation Coefficient was used to estimate correlation in the expression levels of different genes.
High quality RNA was used as a template for first-strand cDNA synthesis SuperScript VILO synthesis kit (Invitrogen). The resulting product was then directly used for qRT-PCR using PerfeCTa SYBR Green SuperMix (Quantabio, USA) in an Applied Biosystems 7300 Instrument (Applied Biosystems, USA).16S rRNA counts were used for normalization. The raw data was then run through LinregPCR for expression data analysis, where the output expression data is displayed in arbitrary fluorescence units (N0) that represent the starting RNA amount for the test gene in that sample [67], [68]. Fold-change relative to WT was then calculated for each individual experiment.
β-galactosidase assays were performed as previously described [69] using cells that were grown in acidic Columbia broth to exponential phase. Cells were either left untreated, or independently treated with CSP1 (EMRLSKFFRDFILQRKK) or CSP2 (EMRISRIILDFLFLRKK) (Genscript, USA) for 30 minutes and processed for analysis.
Pneumococcal cultures grown in Columbia broth were used to seed biofilms on abiotic surfaces. When the cultures reached an OD600 of 0.05, each bacterial strain was seeded on 35MM glass bottom culture dishes (MatTek Corporation, USA). To promote biofilm growth, the plates were incubated at 37°C and 5% CO2. Every 24 hours, the supernatant was carefully aspirated, followed by addition of the same volume of pre-warmed Columbia broth at one-fifth concentration. The biofilm samples were fixed at two time-points: 24 and 72h. For fixing, the supernatants were carefully aspirated, and biofilms were washed thrice with PBS to remove non-adherent and/or weakly adherent bacteria. Subsequently, biofilms were fixed with 4% PFA (Electron Microscopy Sciences), washed three times with PBS, and prepared for confocal microscopy.
Fixed biofilms were stained with SYTO59 Nucleic Acid Stain (Life Technologies, USA) for 30 minutes, washed three times, and preserved in PBS buffer for imaging. Confocal microscopy was performed on the stage of Carl Zeiss LSM-880 META FCS, using 561nm laser line for SYTO59 dye. Stack were captured every 0.46 μm, imaged from the bottom to the top of the stack until cells were visible, and reconstructed in Carl Zeiss black edition and ImageJ. The different biofilm parameters (biomass, maximum thickness, and average thickness over biomass) were quantified using COMSTAT2 plug-in available for ImageJ [70]. For depiction of representative reconstructed Z-stacks, empty slices were added to the images so the total number of slices across all the samples were the same. These reconstructed stacks were pseudo-colored according to depth using Carl Zeiss black edition. The color levels of the images being used for representation purposes were adjusted using GNU Image Manipulation Program (GIMP).
HiBiT constructs were designed by fusing the C-terminus of the region of interest with the 11-amino acid HiBiT peptide using a 10-amino acid linker. The region of interest was the putative secretion signal (until the double glycine) of the briC gene. The expression of these constructs was designed to be controlled by the briC promoter region. The construct was introduced in the genome of R6D and R6DΔcomAB strains downstream of the bgaA gene (without modifying bgaA).
R6D strains containing HiBiT constructs were started from overnight blood agar plates into acidic Columbia broth (pH 6.6) and incubated at 37°C and 5% CO2 without shaking. Cultures were grown to an OD600 of ~0.2. Cultures were either left untreated or treated with 125μg/mL of CSP1 for 30 minutes, followed by measuring optical density at 600nm. Cells were pelleted by centrifuging the cultures for 5 minutes at 3700×g. The resulting supernatants were removed and filtered through 0.2μm syringe filters. The cell pellets were resuspended in equal volume of PBS. To obtain cell lysate, triton X-100 was added to 1ml of the resuspended cells to a final concentration of 1%. Additionally, to minimize non-specific binding, triton X-100 was also added to 1ml of the filtered supernatant to a final concentration of 1%. 75μl of the supernatant, whole cells, lysates, were added to a Costar96 well flat white tissue culture treated plates and mixed with an equal volume of the Nano-Glo Extracellular Detection System reagent as specified in the manufacturer’s instructions. Additionally, media and PBS samples were used as controls. Reactions were incubated at room temperature for 10 minutes followed by measuring luminescence on a Tecan Spark with an integration time of 2000 milliseconds.
All chinchilla experiments were conducted with the approval of Allegheny-Singer Research Institute (ASRI) Institutional Animal Care and Use Committee (IACUC) A3693-01/1000. Research grade young adult chinchillas (Chinchilla lanigera) weighing 400-600g were acquired from R and R Chinchilla Inc., Ohio. Chinchillas were maintained in BSL2 facilities and experiments were done under subcutaneously injected ketamine-xylazine anesthesia (1.7mg/kg animal weight for each). Chinchillas were infected with 100 CFUs in 100μL of S. pneumoniae PN4595-T23 by transbullar inoculation within each middle ear. For RNA extraction, chinchillas were euthanized 48h post-inoculation of pneumococcus, and a small opening was generated through the bulla to access the middle ear cavity. Effusions were siphoned out from the middle ear and flash frozen in liquid nitrogen to preserve the bacterial RNA. Animals were euthanized by administering an intra-cardiac injection of 1mL potassium chloride after regular sedation.
The role of briC in experimental pneumococcal colonization was assessed as previously described [71], [72]. For this, 10 weeks old female CD1 mice (Charles River), weighing approximately 30-35g were anesthetized with 2.5% isoflurane over oxygen (1.5 to 2 liter/min), and administered intranasally with approximately 1X105 CFU/mouse in 20μl PBS. At predetermined time intervals, a group of 5 mice were euthanized by cervical dislocation, and the nasopharyngeal lavage of each animal was obtained using 500μl PBS. The pneumococci in nasopharyngeal wash were enumerated by plating the serial dilutions onto blood agar plates.
The statistical differences among different groups were calculated by performing ANOVA followed by Tukey’s post-test, unless stated otherwise. p-values of less than 0.05 were considered to be statistically significant.
To identify briC homologs we used tBLASTn with default parameters on the RAST database to search the genome sequences of all fifty-five strains. Predicted protein sequences were downloaded as well as nucleotide sequences for the briC homolog and 1500-bp flanking regions surrounding the briC homolog. Predicted protein sequences for BriC were aligned using NCBI Cobalt [73] and visualized using Jalview [74]. One sequences (CDC3059-06) appeared to have a frame-shift after a string of guanines. Given that sequencing technologies are often inaccurate after a string of identical bases, we curated this sequence in the dataset. The sequences were translated in Jalview and organized based on polymorphisms in the translated sequences.
The briC alleles were then organized in the context of the species tree. For this we used a published phylogenetic tree [34], [38]. As previously described, the whole genome sequence (WGS) for these strains were aligned using MAUVE [75], [76], the core region was extracted and aligned using MAFFT (FFT-NS-2) [77]. Model selection was performed using MODELTEST [78], and the phylogenetic tree was built with PhyML 3.0 [79], model GTR+I(0.63) using maximum likelihood and 100 bootstrap replicates. On the visualization, each allelic type is shape-coded, and the visualization was generated using the Interactive Tree of Life (iTOL) [80].
Next, we expanded the search to a set of 4,034 genomes. These correspond to the genomes within pubMLST, with at least 2Mb of genomic data (Genome IDs are listed in S6 Table). We used BLASTn to search for genomes that encode sequences that are at least 70% identical over 70% of their length to briC alleles 1A or 1B. The 3,976 hits were organized to parse out and enumerate the unique sequences using Python. Next, the hits were visualized and further annotated using Jalview [74]. As in the smaller genomic set, one allele representative appeared to have a frame-shift after a string of guanines and was curated in the dataset. Next the DNA sequences were translated, and the predicted protein sequences were organized to display the unique alleles. The resulting 19 coding sequence were colored in Jalview based on percent identity to highlight the variability (Fig 6). To search for briC in related species, we performed a BLASTp analysis in NCBI. We used alleles 1A and 1B as query sequences, default parameters, and the non-redundant database excluding Streptococcus pneumoniae (taxid: 1313).
In order to examine the structure of the promoter region upstream of the briC gene, a 1500-bp flanking region on both sides of the briC gene was pulled from the RAST database [81]. Sequences were aligned using Kalign [82] and then visualized with Jalview [74]. The alignment revealed two clear groups within the dataset: those with the RUP insertion and those without. We also noted that CDC1087-00 may have an additional mobile element inserted within the RUP. However, given that the RUP and this mobile element exist in multiple places in the genome, we cannot determine whether this is real or an artifact of assembly without the isolate. Thus, we opted not to use the promoter sequence for the consensus in Fig 2A, and we did not mark this genome as having a long promoter in Fig 5. We marked the species tree with allelic variants that contain the RUP insertion. We observed that RUP was present in the representative isolates from two clinically important lineages PMEN1 and PMEN14. To check the distribution of the long promoter in a larger set strains, we used PubMLST [39] to inspect 4,034 sequences with complete genomes (sequence IDs for these 4,034 sequences are listed in S6 Table). This set includes 198 ST81 (PMEN1), as well as 104 ST236 (PMEN14) and 15 ST320 (PMEN14) strains. In the PMEN1 strains (ST81), one genome encodes the short promoter and 178 genomes encode the long promoter. In the PMEN14 strains, none of the strains encode the short promoter, and 71 out of 104 ST236 strains and 14 out of 15 ST320 strains encode the long promoter. Manual curation of selected genomes from the remaining set did not reveal a third promoter. Instead, we captured contig breaks and likely issues with assembly that we deduce are linked to repetitive nature of the RUP. Thus, we conclude that the long promoter is present in the vast majority of the PMEN1 and PMEN14 isolates. For analysis of the ComE-binding box, the ComE consensus sequence was extracted from the promoter regions of the pneumococcal strains and aligned with Jalview. The logo was generated using WebLogo [83].
Mouse experiments were performed at the University of Leicester under appropriate project (permit no. P7B01C07A) and personal licenses according to the United Kingdom Home Office guidelines under the Animals Scientific Procedures Act 1986, and the University of Leicester ethics committee approval. The protocol was agreed by both the U.K. Home Office and the University of Leicester ethics committee. Where specified, the procedures were carried out under anesthetic with isoflurane. Animals were housed in individually ventilated cages in a controlled environment and were frequently monitored after infection to minimize suffering. Chinchilla experiments were performed at the Allegheny-Singer Research Institute (ASRI) under the Institutional Animal Care and Use Committee (IACUC) permit A3693-01/1000. Chinchillas were maintained in BSL2 facilities, and all experiments with chinchillas were done under subcutaneously injected ketamine-xylazine anesthesia (1.7mg/kg animal weight for each). All chinchillas were maintained in accordance with the applicable portions of the Animal Welfare Act, and the guidelines published in the DHHS publication, Guide for the Care and Use of Laboratory Animals.
|
10.1371/journal.ppat.1006011 | Cytoadhesion to gC1qR through Plasmodium falciparum Erythrocyte Membrane Protein 1 in Severe Malaria | Cytoadhesion of Plasmodium falciparum infected erythrocytes to gC1qR has been associated with severe malaria, but the parasite ligand involved is currently unknown. To assess if binding to gC1qR is mediated through the P. falciparum erythrocyte membrane protein 1 (PfEMP1) family, we analyzed by static binding assays and qPCR the cytoadhesion and var gene transcriptional profile of 86 P. falciparum isolates from Mozambican children with severe and uncomplicated malaria, as well as of a P. falciparum 3D7 line selected for binding to gC1qR (Pf3D7gC1qR). Transcript levels of DC8 correlated positively with cytoadhesion to gC1qR (rho = 0.287, P = 0.007), were higher in isolates from children with severe anemia than with uncomplicated malaria, as well as in isolates from Europeans presenting a first episode of malaria (n = 21) than Mozambican adults (n = 25), and were associated with an increased IgG recognition of infected erythrocytes by flow cytometry. Pf3D7gC1qR overexpressed the DC8 type PFD0020c (5.3-fold transcript levels relative to Seryl-tRNA-synthetase gene) compared to the unselected line (0.001-fold). DBLβ12 from PFD0020c bound to gC1qR in ELISA-based binding assays and polyclonal antibodies against this domain were able to inhibit binding to gC1qR of Pf3D7gC1qR and four Mozambican P. falciparum isolates by 50%. Our results show that DC8-type PfEMP1s mediate binding to gC1qR through conserved surface epitopes in DBLβ12 domain which can be inhibited by strain-transcending functional antibodies. This study supports a key role for gC1qR in malaria-associated endovascular pathogenesis and suggests the feasibility of designing interventions against severe malaria targeting this specific interaction.
| Plasmodium falciparum sequesters in vital organs. This phenomenon mediated by cytoadhesion of infected-erythrocytes to host receptors in the microvasculature, contributes to the development of severe malaria. Although cytoadhesion to Endothelial Protein-C Receptor has a central role in severe malaria, other host receptors are also likely to be involved. Our results generated by the analysis of P. falciparum isolates from Mozambican patients and laboratory parasite lines indicate that a specific domain (DBLβ12) from DC8-type PfEMP1s can bind to the human receptor gC1qR, previously associated with severe malaria. Our findings revealed that antibodies against PfEMP1 could provide strain-transcending inhibition of gC1qR-binding. Overall, these results support a key role for the adhesion to gC1qR in malaria-associated endovascular pathogenesis and the feasibility of new interventions targeting this specific interaction.
| Case fatality rates for severe malaria (SM) remain unacceptably high even after administration of effective anti-malarial drugs [1]. There is an urgent need to develop novel interventions against life-threatening malaria. However, the mechanisms underlying the clinical heterogeneity and spectrum of malaria [2] remain largely unknown. The general state of health and physiological condition of the host, in particular variations in host immunity, together with genetic predisposition and parasite factors involved in the virulence of the infection, might influence the progression of malaria towards a life-threatening outcome. Sequestration of infected erythrocytes (IE) in vital organs is believed to constitute a key pathogenic event in P. falciparum SM [3], eventually leading to hemorrhages, thrombi formation and pathological inflammation [4], all at the basis of microvascular obstruction [4–6]. Strategies to inhibit or prevent parasite sequestration thus have the potential to reduce the high fatality rate in SM.
Surface proteins at the interface of malaria parasites and the human host contribute to sequestration through the cytoadhesion of IEs to the vascular endothelium, to uninfected erythrocytes to form rosettes [7] and to IEs through platelet binding to form agglutinates (Platelet-mediated [PM]-agglutination) [8]. Cytoadhesion is primarily mediated by interactions between Plasmodium falciparum erythrocyte membrane protein 1 (PfEMP1) [9] and host receptors such as CD36 [10], ICAM-1 [11], CSA [12], heparin [7], EPCR [13] and gC1qR [8,14]. PfEMP1 is a family of highly diverse antigens located on the surface of mature stage IEs that contain 2–9 adhesion domains termed DBL (Duffy binding-like) and CIDR (cysteine-rich interdomain region). Each parasite contains ∼60 different var genes per haploid genome that encode PfEMP1s, which subvert acquisition of protective immunity [15] through constant transcriptional switching [16] and mutually exclusive expression [17]. Antibodies to PfEMP1 that occur after natural infections or after immunization with recombinant PfEMP1 domains are predominantly variant- and strain-specific, as expected for highly variable parasite antigens [18–20]. However, epidemiological observations that children acquire immunity to non-cerebral severe malaria after a small number of infections [21] suggest that strain-transcending antibodies recognizing conserved epitopes on PfEMP1 may occur [19,22], or that the parasites that cause severe malaria are of restricted antigenic types [23,24].
PfEMP1s can be classified into three major groups (A, B and C) and two intermediate groups (B/A and B/C), based on motifs in non-coding sequences and locus position [25]. Whereas most group B and C PfEMP1 proteins appear to be under selection to bind CD36 [26] and tend to be associated with uncomplicated and asymptomatic malaria [27,28], groups A and B/A are often expressed in young children with limited malaria immunity [23] and in those with SM [28–31]. A subset of these A and B/A PfEMP1 variants that contain a combination of adhesion domains, termed domain cassettes 8 and 13 (DC8 and DC13) [32–34], can bind through their CIDRα1.1/4/5/7 domains to Endothelial Protein C Receptor (EPCR) [13]. It has been suggested that EPCR-mediated parasite cytoadhesion could interfere with activation of cytoprotective and anti-inflammatory pathways, which in turn may contribute to severe malaria pathology [13]. However, adhesion to human cell lines is likely to be mediated by interaction with several receptors [35]. Indeed other domains of DC8 and DC13 PfEMP1 variants have been shown to bind avidly to endothelial cells from different tissues through unknown host receptors [36]. These data highlight the heterogeneity of receptors used by IEs in different vascular beds and the importance of identifying other receptors involved in host-parasite interactions.
P. falciparum IEs use gC1qR as a receptor for both cytoadhesion to human cells and platelet-mediated clumping [14], a cytoadhesion phenotype which has been associated with SM in Mozambican children [8]. Human gC1qR is a multi-functional cellular protein expressed on a wide range of tissues and cell types including endothelial cells, lymphocytes, dendritic cells and platelets [37]. In addition to modulating the activation of complement through binding to C1q [38], gC1qR can serve as a receptor for diverse pro-inflammatory ligands [39] and functional antigens of viral and bacterial origin to promote pathogen attachment and/or entry [40]. However, the protein used by malaria parasites to mediate cytoadhesion of IEs to gC1qR is currently unknown.
Selection of IT/FCR3 parasite lines for binding to human brain microvascular endothelial cells (HBMEC) was associated with an up-regulation of DC8- and DC13-PfEMP1 and an increase in binding to gC1qR [33]. Based on this observation, we hypothesized that PfEMP1s containing DC combinations associated with SM may mediate binding to gC1qR. To address this, we assessed the var expression patterns and gC1qR cytoadhesion profile of P. falciparum isolates collected from Mozambican children [8] and in a P. falciparum 3D7 line selected in vitro for binding to gC1qR. The relationship of var transcript levels with disease severity, previous malaria exposure and antibody-mediated recognition of IEs was also analyzed. Our results demonstrate that transcript abundance of DC8 in field isolates is associated with binding of IEs to gC1qR and that DBLβ12 from the DC8-type PFD0020c mediates such interaction in the P. falciparum 3D7 line selected for binding to gC1qR. The successful induction of strain-transcending antibodies against DBLβ12 with activity to inhibit binding to gC1qR by field isolates suggests shared surface epitopes amongst heterologous gC1qR-binding PfEMP1 variants and the feasibility to designing interventions to prevent severe malaria.
Blood samples from 132 malaria patients were used in the study, 111 from Manhiça, Mozambique (86 children and 25 adults) and 21 from European travelers (Table 1). Among the Mozambican children, 43 had uncomplicated malaria (UM) and 43 had SM, defined as severe anemia, acidosis or respiratory distress, multiple seizures, prostration, cerebral malaria or hypoglycemia (Table 1) [8]. Among the 43 cases of severe malaria, 19 (44%) had a single criteria of malaria severity and the rest overlapping symptoms (13 [30%] had two and 11 [26%] three or more). Prostration was observed in 34 (79%) of the children, acidosis/respiratory distress in 17 (39%), severe anemia in 13 (30%) and multiple seizures in 11 (26%), whereas cerebral malaria and hypoglycemia was observed only in 3 and 2 of the children, respectively (Table 1). European travelers were coming from Western Africa (Ghana, Republic of Côte d'Ivoire, The Gambia, Guinea, Equatorial Guinea, Togo, Senegal and Burkina Faso), Middle Africa (Cameroon, Congo and Central African Republic) and Eastern Africa (Mozambique and Madagascar), with none of them presenting SM at recruitment. Parasitemia, quantified by qPCR, was the highest in Mozambican adults, followed by SM and UM cases, with travelers showing the lowest levels of parasitemia (P = 0.022). No differences were observed in the multiplicity of infection (MOI) between groups (P = 0.106).
The relationship between cytoadhesion and var/DCs transcript levels was assessed among P. falciparum isolates collected from Mozambican children (n = 86; Fig 1). Adhesion to CD36 was the most frequent cytoadhesion phenotype (76/86 [88%]; median binding of 180 IEs/mm2, IQR[101–353]), followed by PM-agglutination (57/86 [66%]; median of 7%, IQR[2–22]), adhesion to gC1qR (38/86 [44%]; median binding of 60 IEs/mm2, IQR[45–155]), ICAM1 (37/86 [43%]; median binding of 55 IEs/mm2, IQR[39–105]) and rosetting (31/86 [36%]; median of 2%, IQR[1–5]; Table A in SI Text). The percentage of isolates expressing var/DCs ranged from 41% (35/86) for DC13-CIDRα1.4 to 100% for varA-exon2, varB-UpsB and DC11-CIDRβ2+DBLγ7 (Table B in SI Text). Adhesion to gC1qR correlated positively with DC8 transcript levels (targeted by DC8-CIDRα1.1, rho: 0.287, P = 0.007) and with DC11 (rho: 0.324, P = 0.002). Adhesion to ICAM1 showed a positive correlation with transcript levels of DC13 (rho: 0.273, P = 0.011). Adhesion to CD36 correlated positively with varB (rho: 0.259, P = 0.016) and negatively with varA (rho = -0.228, P = 0.035) and varA-notDC3 (rho: -0.256, P = 0.018). No association was found between var transcript levels, PM-agglutination, rosetting or binding to the negative control Duffy receptor (Fig 1).
DC11 transcript levels were lower in isolates from children with UM than in those from SM children (P = 0.043), severe anemia (P = 0.022), prostration (P = 0.050) and acidosis/respiratory distress (P = 0.044, Fig 2 and Table C in SI Text). Similarly, transcript levels of DC8 were higher in children with severe anemia compared to their UM pairs (P = 0.030, Fig 2). Both DCs were transcribed at similar levels by isolates from travelers and children, being the lowest in isolates from Mozambican adults (Fig 3).
Isolates transcribing DC8 at high levels (i.e., copy number ≥0.5-fold of Seryl-tRNA-synthetase copy number) were more often recognized by plasmas from Mozambican children (n = 100; mean breadth of recognition: 29%, Standard Deviation (SD) 16) than those transcribing DC8 at low levels (16%, SD 11; incidence rate ratio = 2.3, 95% CI [1.2–4.5], P = 0.019; Fig 4A). No differences were observed for other DCs. Breadth of IgG recognition of parasites transcribing DC8 at high levels was higher among the Mozambican adult population (76%, SD 18) than among children (p = 0.010 by Signrank test). However, no difference was observed in the breadth among children with SM (p = 0.969). Finally, recognition by plasma from Mozambican children was the highest for IEs from travelers (mean breadth of recognition: 28%; SD 12), followed by isolates from SM (24%, SD 14) and UM (16%; SD 9), being the lowest for parasites from Mozambican adults (3%, SD 1; test for tend, P≤0.001; Fig 4B).
After two rounds of in vitro selection for binding to gC1qR followed by a limiting dilution cloning, a P. falciparum 3D7 clone was obtained (Pf3D7gC1qR) that showed high binding to gC1qR (mean: 900 parasites/mm2, SD 101) and low to CD36 (71 IEs/mm2, SD 2. In contrast, the unselected 3D7 clone (Pf3D7CD36) showed high levels of adhesion to CD36 (1400 IEs/mm2, SD 159) but no adhesion to gC1qR (Fig 5A). The selection for binding to gC1qR was associated with a 1.3-fold increase in the levels of IgG recognition by plasmas from Mozambican children (Geometric Mean Fluorescence Intensity [GMFI) for Pf3D7gC1qR of 1679, SD: 765 vs GMFI for Pf3D7CD36 of 1266, SD: 576, P≤0.001; Fig 5C). P. falciparum 3D7gC1qR transcribed mostly the DC8-PFD0020c (4.7-fold seryl-tRNA synthetase gene), as well as PFD0625c (5.3-fold seryl-tRNA synthetase gene), whereas Pf3D7CD36 mostly expressed PFD0625c (3.7-fold seryl-tRNA synthetase gene, Fig 5B). The fold ratio of PFD0020c transcript levels in gC1qR-selected Pf3D7 line compared with that in unselected parasites was 4656, and 1.45 for PFD0625c.
To identify the domain(s) mediating binding to gC1qR, we assessed the ability of recombinant constructs representing extracellular domains of PFD0020c (Fig 5D) to bind to gC1qR by ELISA-binding assays. The DBLβ12 from PFD0020c, but not CIDRα1.1, DBLγ6, DBLγ11 and CIDRγ8 was shown to interact with gC1qR (Fig 6A). We further confirmed the gC1qR-binding specificity through a bead-suspension technology (Luminex) in which the beads were coupled with the seven PFD0020c domains and tested for binding to gC1qR, allowing us to confirm that only DBLβ12 was able to bind to gC1qR (Fig A in SI Text). In contrast, only CIDRα1.1 from PFD0020c reacted with rEPCR but not the other domains tested, DBLβ12, DBLγ6, DBLγ11 and CIDRγ8. Moreover, purified polyclonal IgG generated in rabbit against domain DBLβ12PFD0020c, at concentrations of 300 μg/mL, were able to inhibit binding of Pf3D7gC1qR to gC1qR by ∼40% (SD 8) compared with binding of Pf3D7gC1qR to gC1qR in absence of antibodies (P = 0.029). Antibodies against the other PFD0020c domains tested and antibodies against DBLβ12PF08_0140 did not inhibit the binding of Pf3D7gC1qR to gC1qR (Fig 6B). IgGs against DBLβ12-PFD0020c inhibited binding of Pf3D7gC1qR to HBMEC cells, known to express gC1qR on their surface [14], by 46% compared to the control antibody (α -DBLγ6-PFD0020c; Fig 6C). We also tested inhibition of gC1qR binding by four P. falciparum isolates collected from Mozambican children which transcribed DC8, as targeted by DC8-CIDRα1.1, or DBLβ12&DBLβ3/5 domains, at high levels (Fig B in SI Text). In all the four isolates, binding to gC1qR was also reduced by ∼50% in the presence of antibodies against DBLβ12PFD0020c (Fig 6D). IgG against domain DBLβ12PFD0020c did not affect binding to EPCR nor CD36 (Fig B in SI Text). Finally, we show that DBLβ12, together with DBLγ6, DBLγ11, DBLδ1 and CIDRγ8, exhibited the highest increase in IgG recognition among malaria-infected Mozambican children compared to never-exposed Spanish individuals, as well as the highest increase with age of Mozambican children (more than 2.5 years versus less than 2.5 years of age; Fig C in SI Text).
The combined analysis of P. falciparum isolates from malaria infected Mozambique patients and an in vitro selected P. falciparum 3D7 line shows a relationship between cytoadhesion to gC1qR and transcription of DC8-type var genes. The clinical relevance of such a phenotype has been suggested in a field study conducted in Mozambique which showed that prevalence of parasite isolates exhibiting adhesion to gC1qR was associated with multiple seizures [8], although binding levels only tended to be higher compared with isolates from children with severe malaria. In the present study, the use of primer sets targeting the most clinically-relevant DCs [28,34,41] allowed us firstly to correlate the cytoadhesion to gC1qR with abundance of DC8 var transcripts in Mozambican isolates. Secondly, selection of P. falciparum 3D7 line for binding to gC1qR showed the up-regulation of the DC8-PFD0020c. Recombinant DBLβ12PFD0020c bound to gC1qR in ELISA assays and antibodies against this domain were able to inhibit binding of Pf3D7gC1qR and P. falciparum Mozambican isolates to gC1qR by 50%. Overall, these results point to the DBLβ12 domain present in DC8-PfEMP1 variants as the domain that mediates cytoadhesion to gC1qR.
Cytoadhesion to gC1qR by Mozambican isolates correlated positively with their transcript levels of DC8 which, in line with previous studies [34,42,43], were higher in parasites collected from Mozambican children with severe anemia than in those with UM. Moreover, DC8 was transcribed at higher levels by isolates from individuals with limited antimalarial immunity (i.e., Mozambican children and first-time infected travelers) compared to isolates from Mozambican adults with life-long exposure to malaria. As adults and children included in this study come from the same region in Mozambique, it is unlikely that differences observed are due to spatial heterogeneities in the DC8-expressing profile of parasite populations, especially when all parasite genomes appear to have similar repertoires globally [44]. The results rather suggest an exhaustion of the var gene repertoire mediating cytoadhesion and severe malaria with increasing immunity. Alternatively, antigenic variants different to DC8 may increase through ectopic recombination in chronic infections [45] which are expected to be more frequent among semi-immune adults. Also, parasites transcribing DC8 at high levels were more often recognized by plasma from malaria-exposed children than parasites with low DC8 transcription. This is in line with the observation that malaria-exposed Tanzanian population acquires antibodies to EPCR-binding CIDR domains more rapidly than antibodies to other CIDR domains [46]. Although isolates highly transcribing DC8 were better recognized by plasmas from semi-immune Mozambican adults than by children, no difference was observed between plasmas from children with severe and uncomplicated malaria. This latter observation, in line with previous studies conducted in the same area [47], might be attributed to difficulties in disentangling the role of antibodies as markers of exposure and protection among infected population. Overall, these results point towards the contribution of DC8 to gC1qR binding and severe malaria, the antigenic conservation of these PfEMP1 variants, their preferential transcription by malaria parasites infecting individuals who have still not developed antimalarial immunity [34,48,49] and the need to perform longitudinal studies to assess the role of antibodies against DC8 in reducing the risk of severe malaria.
Selection of P. falciparum 3D7 line for binding to gC1qR [14] was accompanied by a marked increase in the expression of a single varA gene, PFD0020c, whose transcript levels were 4656-fold higher than in the unselected line, as well as by an increase in the IgG recognition of IEs by plasmas from exposed children. In contrast, the unselected line, which bound to CD36 in static assays, transcribed B and C var genes at relatively low levels. Similarly to previous in vitro studies with P. falciparum 3D7 [50,51], PFD0625c was also detected in the selected and unselected 3D7 line, which may be due to some degree of relaxed transcription in 3D7 [52]. The PFD0020c specifically up-regulated in Pf3D7gC1qR is a PfEMP1 variant characterized by having three of the four domains usually found in DC8 (CIDRα1.1, DBLβ12, DBLγ4/6), differing only in the first DBLα domain. We were not able to show binding of Pf3D7gC1qR to recombinant EPCR, as would have been expected by the expression of a DC8-containing var gene, although we did not assess this binding specificity to endothelial cells [13]. The up-regulation of a var gene containing DC8 after selection of P. falciparum 3D7 for gC1qR binding fits well with previous in vitro studies showing the transcription of PFD0020c’s orthologs, IT4var19/IT4var07 and HB3var03, after selecting parasites lines IT4 and HB3 for binding to HBMEC cells [32,33]. Importantly, up-regulation of the PFD0020c ortholog IT4var19 after selection of IT for binding to HBMEC was associated with an increased binding to gC1qR as observed in static assays [33].
Recombinant DBLβ12 from PFD0020c, but not other domains from PFD0020c (DBLα1.2, CIDRα1.1, DBLγ4/6, DBLδ1 and CIDRγ8) and from the var type 3 PFI1820w (consisting in domains DBLα1.3-DBLε8), showed binding to gC1qR in ELISA- and Luminex-based binding assays. Moreover, antibodies against DBLβ12 from PFD0020c were able to inhibit the binding to gC1qR by ∼40% at antibody concentrations of 300 μg/mL. This inhibition was not observed in the CD36-binder Pf3D7CD36, which suggest that DBLβ12 is the domain with the ability to bind gC1qR. DBLβ12, which consists in 149 aa and 19 homologous blocks, is present in 12 of the 399 PfEMP1s present in the genomes of seven P. falciparum laboratory strains [44], 9 of them belonging to DC8-PfEMP1 and sharing 56% of similarity at the amino acid level. DBLβ12 was shown to be among the PFD0020c domains most immunogenic in natural infections, as shown by the increase in IgG recognition by malaria-infected Mozambican children compared to never-exposed Spanish individuals, as well as the increase in IgG levels with age of Mozambican children. Importantly, antibodies against DBLβ12 from PFD0020c raised in animal models were able to inhibit binding of Pf3D7gC1qR to HBMEC cells by 46% compared to the control antibody (antibodies against DBLγ6-PFD0020c), demonstrating the gC1qR-dependent adhesion of IEs to endothelial cells through the DBLβ12 domain. Finally, polyclonal antibodies against DBLβ12 from PFD0020c showed cross-inhibitory activity against all the 4 Mozambique clinical isolates sharing the same gC1qR adhesion in vitro, reduced binding by 50%. Three of the four Mozambican field isolates analyzed transcribed DC8, as targeted by DC8-CIDRα1.1. However, one of the isolates (Pfmoz2) did not transcribe DC8, but transcribed DBLβ12&DBLβ3/5 at high levels, suggesting that DBLβ12-containing DC8-like PfEMP1s may share the ability to bind gC1qR. Overall, these data show that parasites with a virulence-associated adhesion phenotype such as gC1qR share PfEMP1 epitopes that can be targeted by strain-transcending functional antibodies to PfEMP1. The existence of shared surface epitopes amongst functionally similar disease-associated P. falciparum parasite isolates suggests the feasibility of developing therapeutic interventions against severe malaria
This study also shows that binding level of IEs to CD36 correlated positively with transcript levels of group B genes and negatively with varA levels, confirming the earlier findings that parasite ligands for CD36 are PfEMP1 variants encoded by var genes belonging to groups B and C [26,53,54]. In contrast to other studies showing up-regulation of DC13 in children with SM [34,42], this DC was not found associated with SM in children in our study, probably due to the low prevalence (7%) of cerebral malaria in the study population. However, transcript levels of DC13 were positively correlated with binding levels to ICAM1. Although DC13 does not have a conserved DBLβ domain with a proven ICAM-1 binding capability [55] most of DC13s are flanked by DBLβ domains, and thus this DC might be associated with a ICAM1-binding DBLβ domain type yet to be described. In fact, the DBLβ domain following DC13 in PF11_0521 [44] has been shown to bind ICAM1 [56]. Importantly, results of this study provide evidences of the potential involvement of DC11 in the pathophysiology of severe malaria. DC11 transcripts were found at higher levels in parasites collected from children with SM than UM as well as in isolates from Mozambican children and first exposed individuals (travelers) compared to isolates from Mozambican adults with life-long exposure to malaria. This DC11 has been involved in rosetting mediated by IgM [57], which has been suggested as the most clinically important rosetting phenotype [58]. However, transcript levels of DC11 were not associated with rosetting in the Mozambican isolates tested. Similarly, platelet-mediated agglutination, previously associated with binding to gC1qR [14], did not show correlation with any var DC. These results suggest that other receptors may be involved in IE rosetting and platelet-mediated agglutination and point towards the relevance of DC11 in the physiopathology of SM. However, further work will be necessary to elucidate the role of DC11 in the severity of the malaria disease.
This study has several limitations. First, more than half of the 43 children with severe malaria included in this study (n = 24; 56%) had two or more criteria for malaria severity [59]. Such a high degree of overlap in severe symptoms, which is otherwise common in endemic areas [60], together with the limited sample size of the cerebral malaria group, may have hampered the identification of molecular correlates that are particular to a clinical form of SM. Second, the degenerate primers used in qPCR assays have incomplete coverage of the global var gene repertoire. Moreover, the parasite populations obtained from peripheral blood may only partially represent the sequestering parasite population. Third, the conditions of the binding assay may not allow for 100% inhibition as has been shown for other receptors [32,61]. Alternatively, residual binding may be supported by other domains, for example those present in DC11, that may also mediate gC1qR adhesion. Fourth, given limited amounts of RNA and cryopreserved IEs available from P. falciparum field isolates included in the study, we focused the transcriptional analysis on domain cassettes previously associated with severe malaria [34,57] and the binding phenotypes on those receptors previously analyzed [8]. Fifth, gC1qR binding assays were performed on five of the 7 PFD0020c domains, but we did not test multiple domains constructs potentially involved in the binding phenotype. Finally, the fact that levels of transcripts encoding certain PfEMP1 domains types associates with the cognate parasites receptor binding capability does not necessarily mean that the particular domain mediates that receptor-binding, and other PfEMP1 domains or structures (i.e., DC11) could convey parasites this binding phenotype. More studies are needed to assess the relationship between expression of DC11, binding to gC1qR and malaria severity.
In summary, the positive correlation between gC1qR cytoadhesion by P. falciparum field isolates and their DC8 transcript levels, the overexpression of DC8-PFD0020c after selection of P. falciparum 3D7 for binding to gC1qR, and the inhibition of gC1qR binding by antibodies against DBLβ12PFD0020c, supports that DC8-PfEMP1s mediate binding to gC1qR through a conserved motif present in the DBLβ12 domain. Overall, our findings suggest that binding to gC1qR, mediated by interactions with DBLβ12, constitutes one of the three different host receptors suggested by protease-treatment assays of IT4 [36]. Moreover, the successful induction of strain-transcending antibodies against DBLβ12 domain from the PfEMP1 variant PFD0020c capable of inhibiting binding to gC1qR by field isolates suggests shared surface epitopes amongst heterologous gC1qR-binding PfEMP1 variants and the feasibility of designing interventions to prevent severe malaria. DC8 may thus facilitate binding to endothelial cells [32,33] via the interactions with gC1qR, known to be expressed in a wide range of human cells [14], in concert with binding to EPCR [13]. Further studies are needed to assess the relationship between DC8 expression, EPCR and gC1qR cytoadhesion, and their influence on malaria disease. Similarly to EPCR, gC1qR has been implicated in inflammatory processes such as the modulation of the complement cascade [40] and suggested to mediate bacterial cell adhesion to sites of vascular injury and thrombosis [62]. Moreover, up-regulation of gC1qR in bone marrow endothelial cells through inflammatory mediators [63] could contribute to sequestration of asexual late stages observed in ex vivo studies [64,65]. The results of this study support the possibility of a role for gC1qR in malaria-associated endovascular pathogenesis.
The study was conducted at the Manhiça District Hospital (MDH) in Southern Mozambique, a malaria endemic area where transmission of P. falciparum is perennial with some seasonality and moderate intensity [66], and at the Tropical Medicine Unit in Hospital Clinic of Barcelona (HCB), Spain. Between April and November 2006, 86 children 1 to 5 years of age [8] were recruited at MDH with P. falciparum clinical malaria, defined as the presence of fever (axillary temperature ≥37.5°C) and an asexual parasitemia of P. falciparum ≥500 parasites/μL on thin blood film examination [67]. Children with SM were those presenting with at least one of the following clinical definitions: cerebral malaria, severe anemia, acidosis or respiratory distress, prostration, hypoglycemia or multiple seizures [8]. Children with clinical malaria not showing any of the mentioned signs of severity and able to take oral medication (uncomplicated malaria; UM) were sex and age (+/-3 months) matched to SM cases. All cases and controls were reviewed by the study pediatrician to confirm that malaria was the sole or principal cause of the disease. Children with concomitant positive bacteremia were excluded from the study. Non-pregnant Mozambican adults (women and men) with life-long exposure to P. falciparum (n = 25) presenting clinical malaria at MDH were recruited between 2004 and 2005 [41]. European adults presenting a first episode of malaria after a travel to malaria endemic areas (n = 21), were recruited between 2005 and 2009 at HCB (Spain) [68]. Before treatment, peripheral blood was collected by venipuncture and 2 drops were spotted onto filter paper. Following centrifugation, plasma and 300 μL of the red blood cell pellet resuspended in 3 mL of Trizol reagent (Invitrogen) were stored at -80°C. The remaining red blood cell pellet was cryopreserved in liquid nitrogen [8].
The study protocol was approved by the National Mozambican Ethics Review Committee and the Hospital Clínic of Barcelona Ethics Review Committee. All patients were included into the study after written informed consent was given by them or their parents/guardians and were treated following national guidelines of Mozambique or Spain at the time of the study.
Total genomic DNA was extracted from filter papers using QIAmp DNA Mini Kit (Qiagen). Parasitemia was measured by real-time quantitative PCR (qPCR) targeting the P. falciparum 18S ribosomal RNA gene [69]. The number of concurrent infections (multiplicity of infection, MOI) was estimated as the highest number of msp-1 or msp-2 alleles detected in the sample by nested-PCR genotyping [70].
Total RNA prepared in Trizol reagent was extracted using PureLink Micro-to-Midi RNA purification kit (Ambion). RNA was treated with DNaseI (Invitrogen) for 1.5h at 37°C. After discarding the presence of gDNA by PCR-based amplification of P. falciparum tubulin (PF10_0084) [41] or Seryl-tRNA synthetase genes (PF07_0073) [71], reverse transcription was performed using the Super Script III First Strand synthesis system (Invitrogen) with random hexamers primers. Complementary DNA (cDNA) synthesis was confirmed by PCR-based amplification of P. falciparum tubulin or seryl-tRNA synthetase genes. Then, the transcript levels of var subgroups was determined by qPCR using degenerated primers targeting varA-exon2, varB group (varB-UTR region), varC group (varC-UTR region) [28], varA-DBLα1 (varA-notDC3) [34] and var2CSA (DBL3X domain) [41]. DC transcript levels were assessed by qPCR using a set of primers targeting semi-conserved domains belonging to DC8 (CIDRα1.1), DC9 (DBLγ), DC11 (CIDRβ2+DBLγ7; Forward: TTRGTHACAGCAAAATAYGAAGGTG and reverse: CTCTTACRATATCWCCTATATCKGCA), DC13 (CIDRα1.4), DC16 (CIDRδ) and DC19 (DBLα0.16) [34]. Seryl-tRNA-synthetase gene was used as the reference gene [71]. Individual 20 μL qPCR reactions were performed in duplicate using ABI Prism 7500 Real-Time system (Applied Biosystems) containing 10 μL of Power SYBR Green Master Mix (Applied Biosystems), 4 μL of cDNA and primer concentration of 1μM with cycling conditions of 50°C for 2 min, 95°C for 10 min followed by 40 cycles at 95°C for 15 s and 60°C for 1 min. Data were analyzed using the 7500 System SDS software v1.4. PCR efficiencies of each primer pair were calculated on a standard curve from 7 log dilutions of P. falciparum 3D7 or P. falciparum ItG gDNA by the formula (E = 10−1/m), where m is the slope. Specificity of amplification was assessed by melting-curve analysis of final products. Non-template controls were tested in every plate. Samples with fluorescence detected over the 40 cycles were considered positive. If Ct value felt out of the linear range of the standard curve, a Ct value of 41 was assigned. Ct values were converted to copy numbers [72] using the formula C/EΔCt, where E is the efficiency of the PCR, C is the number of copies of the gene in the P. falciparum 3D7 or ItG genome [34,72] and ΔCt is the difference in Ct values between a sample and P. falciparum 3D7 or ItG reference gDNA loaded in each plate [73]. Relative copy number of target genes was calculated by dividing the target gene copies by Seryl-tRNA synthetase gene copies. Transcript levels of var/DCs were considered as high if the copy number was ≥0.5-fold of Seryl-tRNA-synthetase copy numbers and low if copy number was <0.5-fold.
Adhesion of P. falciparum pediatric isolates to gC1qR (Creative BioMart), CD36, ICAM-1 (R&D Systems) and Duffy-Fc [74], as well as rosetting and PM-agglutination was assessed as previously described [8] and expressed as IEs/mm2, percentage of IEs forming rosettes and percentage of IEs in a clump, respectively. Adhesion to purified receptors was considered positive if the number of IEs bound per mm2 was higher than the mean binding plus 2 standard deviations to Duffy-Fc coated Petri dishes (19.5 IE/mm2) [8], rosetting if frequency of rosettes was higher than 2% [75] and PM-agglutination if frequency of clumps was higher in presence of platelets than in buffer-control [8].
Forty five P. falciparum isolates were tested for IgG recognition by plasma from 50 children with SM and 50 with UM, as well as 22 adults recruited in the same study area [8]. After thawing and washing erythrocytes in incomplete RPMI 1640 medium, parasites were matured at 37°C for 18–36 hours until late-stages. Fifty μL of plasmas at 1/10 dilution, previously depleted of antibodies reacting against uninfected A/B-erythrocytes, were mixed with 50 μL of erythrocyte suspension at 1% hematocrit and 0.5–2.2% parasitemia in PBS-1% BSA for 1 hour at room temperature. After sequential incubations with 100 μL of polyclonal rabbit anti-human IgG (DakoCytomation; 1/200 dilution) and 100 μL of Alexa Fluor 488-conjugated donkey anti-rabbit IgG (Invitrogen; 1/1,000) plus 10 μg/mL of ethidium bromide, data from 1,000 ethidium bromide positive events were acquired with a Becton Dickinson LSR Fortessa flow cytometer. Reactivity against IEs was expressed as the difference between the geometric mean fluorescence intensity (GMFI) of IEs and the GMFI of uninfected erythrocytes. A pool of plasma samples from immune Mozambican adults and six plasma samples from non-exposed European adults were included as positive and negative controls, respectively. To allow comparability between isolates, GMFI values from each parasite/plasma combination were scored in relation to the threshold of positivity for each isolate defined as the GMFI of negative controls plus two standard deviations (cut-off). A score of 0 was assigned if GMFI values were below the cut-off; 1 if the value was between one- and two-fold the cut-off; 2 if the value was between two- and three-fold the cut-off; and so on until a maximum score of 5. Breadth of IgG recognition (BoR) was calculated as the sum of scores obtained for each parasite and expressed as percentage of the maximum score possible.
To select for binding to gC1qR, a P. falciparum 3D7 culture synchronized in trophozoite/schizont stages was incubated for 1 h in bacteriological Petri plates coated with 2 mL of recombinant gC1qR diluted in PBS (50 μg/mL) [14]. Unbound parasites were collected using a pipette and separated from bound parasites. Both unbound and bound parasites were cultured, with the latter being subjected to a second round of selection for binding to gC1qR. After a limiting dilution cloning, a selected and unselected clone were expanded and tested for binding to gC1qR, CD36, ICAM-1, CSA (Chondroitin sulfate A sodium salt from bovine trachea Sigma-Aldrich) and BSA (Bovine Serum Albumin, Santa Cruz Biotechnology), following standard procedures [8]. The var genes transcription profile was determined for both clones by individual qPCR performed in duplicate using primers covering the P. falciparum 3D7 var gene repertoire [71,76].
Recombinant PFD0020c domains produced in insect or Escherichia coli cells [13] were screened for binding against recombinant human EPCR or gC1qR by ELISA (CIDRα1.1, DBLβ12, DBLγ6, DBLγ11, CIDRγ8) and Luminex (DBLα1.2, CIDRα1.1, DBLβ12, DBLγ6, DBLγ11, DBLδ1 and CIDRγ8) in duplicate. For the ELISA assays, MaxiSorp immunoplates (Nunc) were coated overnight at 4°C with 50 μL per well of recombinant human EPCR and gC1qR at 3 μg/mL in PBS pH 7.4. After blocking with PBS 3%-skimmed milk and washing three times with PBS-0.05% TweenR20, PFD0020c domains were added at a concentration of 5 μg/mL in PBS 1%-skimmed milk and incubated for 1 h at 37°C. Secondary anti-V5-HRP antibody diluted in PBS 1%-skimmed milk at 1:3000 was added to each well and incubated for 1 hour at room temperature with gentle shaking. Plates were developed using 100 μL per well of a phosphate solution with o-phenylenediamine. The colorimetric reaction was stopped with 100 μL of 3 M H2SO4 after 10 minutes and the optical density (OD) was measured at 490 nm. For the Luminex assays, gC1qR was coupled at 50 μg/107 beads to MagPlex-C magnetic carboxylated microspheres (Luminex Corporation) following manufacturer’s instructions. Two thousand coupled beads were incubated with the recombinant PFD0020c domains (DBLα1.2, CIDRα1.1, DBLβ12, DBLγ6, DBLγ11, DBLδ1 and CIDRγ8) at 1ug/ml in incubation buffer (IB; 1% Skim Milk in PBS), overnight at 4°C. After 3 washes with washing buffer (PBS + 0.5% Tween20 + 0.25% skim milk), the beads were incubated with anti-V5 from mouse (ThermoFisher, R960-25) at 1/2500 in IB at room temperature for 1 hour, followed by an incubation with anti-mouse biotin conjugated antibody (Sigma, B7401) at 1/10000 in IB for 1 hour at RT, and streptavidin-R-phycoerythrin (Sigma, 42280) at 1/1000 in IB for 30 minutes at RT, with 3 washes after each incubation. Median Fluorescence Intensity was obtained using the Luminex 100/200 System (Luminex Corp., Austin, Texas).
Anti-sera against domains belonging to PFD0020c (α-CIDRα1.1PFD0020c, α-DBLβ12PFD0020c, α-DBLγ6PFD0020c, α-CIDRγ8PFD0020c)), PF08_0140 (α-DBLβ12PF08_0140) and PFI1820w (α-PFI1820w) were produced in rabbit [13]. After depleting rabbit sera of antibodies against human erythrocytes, IgGs were purified by Affi-Gel Protein-A MAPS II Kit (Bio-Rad, Richmond, CA) and quantified using EPOCH spectrophotometer. To test their ability to inhibit binding of P. falciparum to recombinant gC1qR, 20 μL pellet of P. falciparum 3D7 pigmented trophozoite (≥2% parasitaemia, 1% hematocrit) were incubated in duplicate for 1.5 h at 37°C with 300 μg/mL rabbit IgGs diluted in PBS and used for a standard adhesion assay in Petri dishes [8]. Similar procedures were used to test inhibition of gC1R binding by 4 Mozambican P. falciparum isolates (Pfmoz 1–4).
Human Brain Microvascular Endothelial Cells (HBMEC; Innoprot) were seeded on flat-bottomed Nunclon Δ Surface (Nunc cat number: 150628) 12-well plates 3 to 4 days before assays and allowed to growth to 30–40% confluence in endothelial cell medium (Innoprot). Prior to the adhesion assay, HBMECs were washed once with PBS followed by addition of 20 μl 2% FCS in RPMI/well. For binding inhibition, IgG-purified anti-PfEMP1 rabbit antibodies and PBS alone were added to 2% parasitemia and 2% hematocrit late-stage IEs at a final concentration of 300 μg/ml incubated for 1.5 h at 37°C. 300 μl of the IE suspension were added to each well and co-incubated on a rocking table for 1 hour at room temperature. Unbound infected erythrocytes were removed by several gentle washes. Wells were then fixed in 2% glutaraldehyde over night at room temperature and stained with Giemsa for 10 min. Binding was quantified by determining the number of IEs adhering per endothelial cells nuclei in 50 random fields under 400× magnification. All binding assays were done in triplicate. The percentage of binding was expressed relative to binding in the absence of antibodies.
IgG reactivity against the recombinant PFD0020c domains (DBLα1.2, CIDRα1.1, DBLβ12, DBLγ6, DBLγ11, DBLδ1 and CIDRγ8) was assessed in 135 malaria-infected Mozambican children (67 with severe malaria and 68 with uncomplicated malaria) and 18 Spanish adults never exposed to malaria. PfEMP1 domains or BSA (Sigma, A7030, as background control) were coupled at 50 μg/107 beads to MagPlex-C magnetic carboxylated microspheres (Luminex Corporation) following manufacturer’s instructions. Multiplexed beads were incubated with plasma samples (1/50 dilution) and antibody levels were detected as described elsewhere [77]. Positive, negative and background controls were added to each plate. Median Fluorescence Intensity (MFI) was obtained from the InVitrogen Luminex platform (xPONENT Software, at least 100 counts/analyte) and normalized for inter-plate variability by multiplying MFIs by the median value of a positive control from all plates and dividing by each plate’ value.
Correlations between variables were assessed by Spearman’s rank coefficient, with Benjamini-Hochberg correction for multiple comparisons. Continuous data were compared between matched case/control pairs by Sign-test and between non-paired groups by Mann-Whitney test. BoR was compared between groups by a Test for trend across ordered groups and between isolates transcribing var/DCs at low- or high-levels by negative binomial regression models adjusted by age. Mean ratio of IgGs and 95% confidence intervals between Mozambican children and Spanish adults, as well as between Mozambican children older than 2.5 years of age and less than 2.5 years were calculated in linear regression models, with log-transformed MFIs. Statistical analysis was performed with Stata/SE software (version 12.0; StataCorp).
|
10.1371/journal.pgen.1003318 | HIV Infection Disrupts the Sympatric Host–Pathogen Relationship in Human Tuberculosis | The phylogeographic population structure of Mycobacterium tuberculosis suggests local adaptation to sympatric human populations. We hypothesized that HIV infection, which induces immunodeficiency, will alter the sympatric relationship between M. tuberculosis and its human host. To test this hypothesis, we performed a nine-year nation-wide molecular-epidemiological study of HIV–infected and HIV–negative patients with tuberculosis (TB) between 2000 and 2008 in Switzerland. We analyzed 518 TB patients of whom 112 (21.6%) were HIV–infected and 233 (45.0%) were born in Europe. We found that among European-born TB patients, recent transmission was more likely to occur in sympatric compared to allopatric host–pathogen combinations (adjusted odds ratio [OR] 7.5, 95% confidence interval [95% CI] 1.21–infinity, p = 0.03). HIV infection was significantly associated with TB caused by an allopatric (as opposed to sympatric) M. tuberculosis lineage (OR 7.0, 95% CI 2.5–19.1, p<0.0001). This association remained when adjusting for frequent travelling, contact with foreigners, age, sex, and country of birth (adjusted OR 5.6, 95% CI 1.5–20.8, p = 0.01). Moreover, it became stronger with greater immunosuppression as defined by CD4 T-cell depletion and was not the result of increased social mixing in HIV–infected patients. Our observation was replicated in a second independent panel of 440 M. tuberculosis strains collected during a population-based study in the Canton of Bern between 1991 and 2011. In summary, these findings support a model for TB in which the stable relationship between the human host and its locally adapted M. tuberculosis is disrupted by HIV infection.
| Human tuberculosis (TB) caused by Mycobacterium tuberculosis kills 1.5 million people each year. M. tuberculosis has been affecting humans for millennia, suggesting that different strain lineages may be adapted to specific human populations. The combination of a particular strain lineage and its corresponding patient population can be classified as sympatric (e.g. Euro-American lineage in Europeans) or allopatric (e.g. East-Asian lineage in Europeans). We hypothesized that infection with the human immunodeficiency virus (HIV), which impairs the human immune system, will interfere with this host–pathogen relationship. We performed a nation-wide molecular-epidemiological study of HIV–infected and HIV–negative TB patients between 2000 and 2008 in Switzerland. We found that HIV infection was associated with the less adapted allopatric lineages among patients born in Europe, and this was not explained by social or other patient factors such as increased social mixing in HIV–infected individuals. Strikingly, the association between HIV infection and less adapted M. tuberculosis lineages was stronger in patients with more pronounced immunodeficiency. Our observation was replicated in a second independent panel of M. tuberculosis strains collected during a population-based study in the Canton of Bern. In summary, our study provides evidence that the sympatric host–pathogen relationship in TB is disrupted by HIV infection.
| Host–pathogen co-evolution plays an important role in the biology of infectious diseases [1]. Coevolution between interacting host and pathogen species is difficult to demonstrate formally, but indirect evidence can be obtained by studying geographical patterns, which can indicate local adaptation of particular pathogen variants to geographically matched host variants [2]–[4]. Local adaptation is often studied using so-called reciprocal transplant experiments, in which the fitness of locally adapted (sympatric) pathogen variants is compared to the performance of allopatric pathogen variants [2]. Studies in several invertebrate systems have shown that sympatric pathogens (infection with a phylogeographically concordant strain) tend to outperform allopatric pathogens (infection with a phylogeographically discordant strain) in the corresponding host variants [1], [5]–[7].
Mycobacterium tuberculosis, the agent causing human tuberculosis (TB) is an obligate human pathogen, which has been affecting humankind for millennia [8]–[13]. Contrary to previous beliefs linking the origin of TB to animal domestication ∼10,000 years ago [14], more recent data suggest that M. tuberculosis evolved as a human pathogen in Africa, and might have co-existed with anatomically modern humans since their origins ∼200,000 years ago [8], [10], [12], [13], [15]. Analyses of multiple global strain collections have shown that M. tuberculosis exhibits a phylogeographic population structure consisting of six main phylogenetic lineages associated with different geographic regions and sympatric human populations [9], [11]–[13], [16]–[20]. Studies in San Francisco, London, and Montreal have shown that these sympatric host–pathogen associations persist in cosmopolitan settings, even under a presumed degree of host and pathogen intermingling [11], [18], [19]. Moreover, transmission of M. tuberculosis has been shown to occur more frequently in sympatric host–pathogen combinations compared to allopatric host–pathogen combinations [9]. Taken together, these observations are consistent with the notion that the different phylogeographic lineages of M. tuberculosis have adapted to specific sympatric human populations [21].
Based on the assumption that M. tuberculosis has been coevolving with humans, and that M. tuberculosis has locally adapted to sympatric human populations [9], we hypothesized that HIV co-infection will alter this relationship [22]. Specifically, we postulated that because HIV induces immune suppression in humans, and because variation in host immunity will likely play a role in local adaptation, M. tuberculosis strains will cause disease in HIV–infected patients, irrespective of their usual sympatric host–pathogen relationship. To test this hypothesis, we performed a population-based molecular-epidemiological study of HIV–infected and HIV–negative TB patients in Switzerland between 2000 and 2008, a country with a long history of immigration [23].
A total of 518 patients were included in the study, of whom 112 (21.6%) were HIV–infected. Of these 518 patients, 233 (45.0%) were born in Europe (117 in Switzerland), 131 (25.3%) were born in sub-Saharan Africa, 48 (9.3%) in South-East Asia, 36 (7.0%) in the Indian subcontinent, and 24 (4.6%) in Central and South America. Similar to previous studies [9], [18], [19], we found an association between the patient's place of birth and the particular M. tuberculosis lineages (Figure 1). Lineage 4 (Euro-American lineage) was present in all regions, but particularly common in patients born in Europe and South America. Lineages 5 and 6 (West-African lineages also known as M. africanum [24]) were exclusively found in patients originating from West Africa, whereas Lineage 2 (which includes Beijing) and Lineage 1 were mainly seen in patients originating from the Western Pacific and East Asian regions. Patient characteristics are summarized in Table 1.
Because in European-born patients the host–pathogen combinations defined as sympatric (i.e. Lineage 4/Euro-American lineage in European-born patients) or allopatric (i.e. all other lineages in European-born patients) have been well established [9], [18], [19], [25], we focused on this patient group (n = 233) for the remaining of our analyses.
M. tuberculosis transmission was more likely among patients in a sympatric host–pathogen relationship compared to patients in an allopatric host–pathogen relationship (adjusted odds ratio [OR] 7.5, 95% confidence interval [95% CI] 1.2-infinity, p = 0.03, Table 2). Of note, there was no molecular clustering among European-born TB patients infected with an allopatric M. tuberculosis strain. Moreover, we found that only the sympatric Lineage 4 (Euro-American lineage) was detected in European-born clusters as well as in mixed clusters (Table S1), suggesting that sympatric host–pathogen combinations in TB favor transmission.
Overall, we found that HIV infection was strongly associated with allopatric M. tuberculosis lineages among European-born TB patients (unadjusted OR 7.0, 95% CI 2.5–19.1, p<0.0001; Table 3). Among the allopatric lineages, we found that Lineages 1, 2 and 3 were more likely to be found in HIV–infected compared to HIV–negative patients when taking the sympatric Lineage 4 (Euro-American lineage) as the reference (Table S2). When investigating the ancestry of the nine HIV–infected patients with an allopatric M. tuberculosis strain, seven patients were confirmed to be of Swiss ancestry over the last three generations, and two patients had Swiss and Italian ancestors in the previous generations (Italian father in the previous generation, or emigrating from Italy in the previous generation).
Several factors could contribute to the association between HIV infection and allopatric lineages. We found that patients with an allopatric M. tuberculosis lineage were younger (median age 41.5 versus 50 years), and had more often a history of frequent travelling (38.9% versus 4.2%, p<0.0001). Therefore, we developed a model (Figure 2) to take these and other putative confounding variables into account. These variables included age, sex, country of birth, frequent travelling, contact with the foreign-born population, and non–HIV associated immunosuppression. We considered “TB with an allopatric strain” as the outcome because disease is the only measurable outcome with a sympatric or allopatric M. tuberculosis strain (only diseased individuals can yield a positive mycobacterial culture). Our multivariate analyses revealed that the association between HIV infection and allopatry remained statistically significant after adjustment for all social and patient factors included in our model (OR 5.5, 95% CI 1.5–20.6, p = 0.01, Table 3). Age, sex, being Swiss-born, and non–HIV associated immunosuppression had only a minor effect on the association between HIV infection and TB with an allopatric strain (Table 3). In contrast, a history of repeated travelling to low-income countries had a stronger effect, decreasing the OR to 4.50 (95% CI 1.5–13.6, p = 0.008, Table 3) when adjusting for this variable.
We also tested if the degree of immunodeficiency as measured by the nadir CD4 T cell count (defined as the lowest CD4 T cell count ever measured in a patient) would have an impact on the association between host population and M. tuberculosis lineage. Among Europeans, the strength of the association between HIV infection and allopatric lineages increased with a decreasing nadir CD4 T cell count in a dose-dependent manner: from an OR of 4.6 (95% CI 0.9–24.7) in patients with a nadir CD4 T cell count of >200 cells/µL to an OR of 12.5 (95% CI 2.6–60.8) in patients with nadir CD4 T cell counts <50 cells/µL (test for trend p<0.0001; HIV–negative patients as reference). This trend remained statistically significant when adjusting for age, sex, being born in Switzerland, frequent travelling, contact with the foreign-born population, and non–HIV associated immunosuppression (Table 4).
Increased contact with foreigners originating from high TB burden countries, who have a higher risk of TB [26] and are more likely to have TB caused by an allopatric M. tuberculosis strain, could also lead to an allopatric host–pathogen relationship in European-born patients. However, the association between HIV and allopatry remained statistically significant when adjusting for this variable (Table 3). Furthermore, we examined molecular clusters defined by standard bacterial genotyping [27], [28], to test the hypothesis that HIV–infected patients were more frequently seen among ethnically mixed clusters where transmission occurred between non-European and European-born cases [29]. We found that the prevalence of HIV infection was similar among TB cases in mixed clusters (5 HIV–infected cases out of 26 cases, 19.2%) and among cases in clusters involving only European-born cases (4 out of 26 cases, 15.4%, see Table S1).
When restricting the main analysis (n = 233) to European-born patients without a history of frequent travelling, we found that the association between HIV infection and allopatric TB remained statistically significant (adjusted OR 6.96, 95% CI 1.25–38.88, p = 0.027). Furthermore, we explored associations of socio-demographic and clinical factors with TB with an allopatric M. tuberculosis strain in a model focusing on HIV–infected European patients only (Figure S1, Table S3): frequent travelling was confirmed to be an important factor, and patients with a low nadir CD4 T cell count tended to be associated with an allopatric TB although the associations did not reach statistical significance (Table S3). Finally, we obtained very similar results for the association between HIV infection and allopatric TB (Table S4), and for the association between the degree of immunodeficiency and allopatric TB (Table S5) when repeating analyses using a Bayesian approach [30], which is more robust when numbers are small.
The birth origin of HIV–infected and non-infected patients is shown on a map in Figure S2. The main phylogenetic M. tuberculosis lineages stratified by place of birth and HIV status are presented in Table S6.
To replicate our main finding, we investigated a second panel of M. tuberculosis strains from an ongoing population-based TB study in the Canton of Bern, Switzerland, between 1991 and 2011. Of the 1,642 M. tuberculosis isolates analyzed, 1,350 (82.2%) belonged to Lineage 4 (Euro-American lineage), and 292 (17.8%) to non-Euro-American lineages (Lineages 1, 2, 3, 5 or 6). We compared all 40 European-born patients with an allopatric strain (non-Lineage 4) with 400 randomly selected European-born patients with a sympatric strain (Lineage 4). We found that the proportion of HIV infection was 4.5 (95% CI 1.6–11.9) times higher in patients with an allopatric strain compared to patients with a sympatric strain (12.5% versus 2.8%, p = 0.010, Table 5).
The phylogeographic distribution of M. tuberculosis lineages observed here suggests local adaptation to sympatric human populations. In contrast, we found that allopatric host–pathogen relationships in European-born TB patients were strongly associated with HIV co-infection. The association with HIV infection became stronger in a ‘dose-dependent’ manner in patients with a history of more pronounced immunodeficiency, and was not explained only by frequent travelling to high TB-incidence countries or increased social mixing with the foreign-born population. The association of M. tuberculosis lineages with sympatric patient populations reported here is in agreement with previous findings [9], [11]–[13], [16]–[19]. Similarly, our finding that recent TB transmission was more likely to occur in sympatric compared to allopatric host–pathogen combinations supports previous work [9]. Taken together, these data are consistent with local adaptation of M. tuberculosis to different human populations, which in turn can be viewed as indirect evidence for coevolution between M. tuberculosis and its human host [1]–[4], [9]–[13].
We found that TB allopatric host–pathogen combinations were strongly associated with HIV infection in a nation-wide study and a second panel of strains from one Canton of Switzerland. This supports the notion that M. tuberculosis lineages have evolved subtle differences in their interaction with different human immune systems. However, in the presence of HIV–induced immunodeficiency, any M. tuberculosis lineage seems to cause disease in a given human host. M. tuberculosis is an obligate human pathogen which lives in constant interaction with the host immune system [31]. Human populations, however, are known to differ genetically and immunologically [15]. The clinical disease reflects host-dependent immune-pathological processes [31]. In other words, while initially triggered by the pathogen, it is the host immune response which is ultimately responsible for the chronic inflammation and associated tissue destructions. These processes contribute to the successful transmission of M. tuberculosis [22], [32]. On the other hand, only 5–10% of the 2 billion individuals estimated to be latently infected with M. tuberculosis globally will develop active TB during their lifetime [33]–[35]. Hence most of the time, humans are able to control the infection. In our study, we chose culture-confirmed TB cases as the main endpoint which reflects successful transmission and progression from infection to active disease.
Our study on the association between allopatric TB and HIV was able to control for important cofactors [36], [37]. These cofactors included frequent travelling abroad and increased contact to foreign-born populations. A particularly important cofactor for allopatric TB was frequent travelling to high TB burden countries with potential exposure to foreign M. tuberculosis strains; HIV–infected individuals may be at a higher risk for travel-related infectious diseases [38]. However, the association between HIV infection and allopatric TB remained even when adjusting for these behavioral and other patient characteristics. A previous study reporting on allopatric TB and HIV was not able to control for these factors [9]. Furthermore, we found no evidence for increased social mixing among HIV–infected individuals, which argues against mere social factors leading to the association between allopatric TB and HIV.
A biological basis for this association is further supported by the striking dose-dependency we observed with increasing immunosuppression as defined by lower nadir CD4 T cell counts. Of note, this trend was also independent of other variables. Low nadir CD4 T cell counts are associated with incomplete immune recovery after starting combination antiretroviral therapy [39], [40] and impaired functional immune restoration despite normalization of CD4 T cells [41]. More generally, infection with HIV and M. tuberculosis interferes with the immune system in many ways [42], [43]. HIV infection disrupts the function of M. tuberculosis-infected macrophages [44], [45], but also seems to reduce the number and functionality of M. tuberculosis-specific T cells over time [46]. On the other hand, M. tuberculosis strains have been shown to induce variable immune responses [47]. Based on these observations, it is reasonable to hypothesize that HIV/TB co-infection might impact immune cell functions, intracellular signaling and immune regulation, perhaps leading to an immune response less capable of discriminating between M. tuberculosis variants.
Besides M. tuberculosis, several other human pathogenic bacteria exhibit phylogeographic population structures, possibly reflecting local adaptation to different human populations. These include Haemophilus influenzae [48], Streptococcus mutans [49], M. leprae [50] and Helicobacter pylori [51], [52]. Interestingly, like M. tuberculosis, all of these microbes are obligate human pathogens. In the case of H. pylori, functional studies have shown that strains associated with South America have adapted their adhesins to the human blood group O, which is particularly frequent in native populations of this region [53]. Similarly, a study of the bacterial genome evolution of an asymptomatic Escherichia coli bacteriuria strain showed adaptation at the genomic level in distinct human hosts [54]. No similar experimental work has yet been carried out in TB. However, several studies have reported associations between human genetic polymorphisms and particular M. tuberculosis lineages [55]–[59], indicating possible interaction between human and M. tuberculosis variation. Whether such variation in pathogen and host genetics can be attributed to co-evolution will be difficult to demonstrate conclusively, but the data presented here support this possibility.
The strength of our study was that we used a nation-wide sample to specifically look at the impact of HIV infection on the host–pathogen relationship in human TB. Yet, our study is limited by the relatively small sample size, and the difficulty to quantify the complex social context through which the host–pathogen relationship is influenced in human TB. In addition, we looked at European-born patients only, because sympatric and allopatric host–pathogen combinations are more easily defined for this patient population [9], [18], [19], [25]. Additional studies in large cosmopolitan cities of Asia and Africa would be required to test whether the association between allopatric TB and HIV holds true in these settings. Ultimately, detailed experimental work is needed to establish the biological basis of the host–pathogen association in human TB.
In conclusion, our data suggest that the phylogeographical host–pathogen relationship in TB influences transmission patterns. Among the studied European-born TB patients, we showed that HIV infection disrupts the sympatric host–pathogen relationship in human TB, and that this effect increased as a function of immunodeficiency. Various interactions between HIV and M. tuberculosis at the cellular level make an association biologically plausible [42], [43]. Further studies are needed to investigate the impact of HIV on the genetic population structure of M. tuberculosis with its consequences for transmission and clinical manifestations in high TB burden countries [36]. This will lead to a better understanding of biological factors that shape the current HIV/TB syndemic [60].
The Swiss Molecular Epidemiology of Tuberculosis (SMET) study is a collaborative project between the Swiss HIV Cohort Study (SHCS), the National Center for Mycobacteria, diagnostic microbiology laboratories, departments of respiratory medicine and public health, and the Federal Office of Public Health (FOPH) [29], [61], [62]. The overarching aims were to examine the genetic population structure of M. tuberculosis and the associations between strain variation, patient origin, and clinical characteristics in HIV–infected and HIV–negative TB patients in Switzerland. Further information on the SMET project is available at www.tb-network.ch. All participating sites are listed in the Acknowledgements.
The SHCS is a prospective observational study of HIV–infected individuals followed up in HIV outpatient clinics in Switzerland [63]. All HIV–infected patients diagnosed with TB between 2000 and 2008 whose M. tuberculosis complex (MTBC) isolate was available were included in the SMET study [29]. Furthermore, we randomly selected 288 from the 4,221 culture-confirmed TB cases reported to the National TB Surveillance Registry during the same period (approximately three cases for one HIV–infected TB case within the SHCS). Finally, all reported drug-resistant TB cases were included. Two M. bovis isolates were excluded from this study as they are animal-adapted species within the MTBC and therefore represent a different host–pathogen relationship.
We obtained clinical data by standardized questionnaires sent to the treating physicians and extracted relevant data from the SHCS database. We collected socio-demographic data (age, sex, origin of birth, citizenship, legal status, immunosuppressive therapy, risk factors for TB such as recent TB within family or immediate social surroundings in the last two years), laboratory parameters (CD4 cell count and plasma HIV RNA in HIV–infected cases) and clinical information (site of disease, radiography findings). Chest radiography parameters were consolidation, cavitations, enlarged intrathoracic lymph nodes and pleural thickening. Any drug resistance was defined as any resistance to isoniazid, rifampicin or ethambutol as reported to the FOPH. Most TB cases in Switzerland are treated under the guidance of experienced infectious and respiratory disease specialists, and the clinical data were of high quality.
Geographic origin of patients was defined as the country of birth, and countries were grouped in seven geographic regions (see Figure 1) according to the current understanding of the phylogeography of M. tuberculosis [25]. Birth country was used as a proxy of the ancestry of the study population. Immunosuppression due to other causes than HIV infection was defined as use of TNF-alpha inhibitors, malignancy, solid organ transplantation, use of steroids or methotrexate. Nadir CD4 T cell count was defined as the lowest CD4 T cell count (cells/µL) ever measured in a patient. Nadir CD4 T cell count is a predictor of poor immune recovery after ART [39]. Travel history was extracted from the free text field “Risk factor for TB” and defined as repeated travelling of longer duration (>30 days) to low-income countries with a high TB burden and a relevant exposure to M. tuberculosis according to the physician's judgment. Belonging to a molecular cluster involving Swiss-born and foreign-born TB cases was used as a proxy for contact with the foreign-born population.
Mycobacterial isolates were cultured and DNA extracted according to standard laboratory procedures. We used spacer oligonucleotide typing (spoligotyping) and 24-loci mycobacterial interspersed repetitive units (MIRU-VNTR) which are based on repetitive DNA sequences as genotyping tools with high discriminatory power to identify recent TB transmission [29], [64]–[66]. Data were analyzed with the MIRU-VNTRplus online tool (http://www.miru-vntrplus.org). Molecular clusters were defined as a group of completely identical isolates in the spoligotyping and MIRU-VNTR pattern indicating a chain of TB transmission. In addition, we used single nucleotide polymorphisms (SNPs) as stable genetic markers to define the main phylogenetic M. tuberculosis lineages [67]. Lineages were determined by SNPs using multiplex real-time PCR with fluorescence-labeled probes (Taqman, Applied Biosystems, USA) adopted from previous studies [9], [12], [67], [68]. The SNP used to define Lineage 4 was originally described by Sreevatsan et al. [69] and shown to be specific to this lineage [9].
Graphical models were built using the principles of directed acyclic graphs [70]. Our model considered infection and disease as a combined outcome (“TB with an allopatric strain”). Our hypothesis that HIV infection causes TB with an allopatric strain is shown as a potentially causal direct effect, and risk factors potentially influencing this effect are shown in the hypothetical direction. Mediators represent variables that are caused by the independent variable and, in turn, have a direct effect on the outcome variable. We included age and sex in our model as risk factors for infection and disease [37]. We also considered contact with the foreign-born population who have a higher risk for TB compared to the native Swiss population [26] and who have a higher risk of exposure to “foreign” M. tuberculosis strains. Finally, we included frequent traveling to countries with a high TB burden, which increases exposure risk and thus potentially infection risk with “foreign” M. tuberculosis strains (Figure 2).
We used χ2 tests or Fisher's exact tests to assess differences between groups in binary variables, and the Wilcoxon rank sum test for continuous variables (Table 1, Table 2). Univariate and multivariate exact logistic regression models were fitted to estimate the association between transmission as defined by molecular clustering and patients with sympatric M. tuberculosis lineages (patients with allopatric lineages were used as the reference, Table 2). Results were presented as ORs unadjusted and adjusted for age group, being born in Switzerland and recent TB in families or social surroundings. To assess the association of HIV infection with allopatric TB, we fitted univariate and multivariate logistic models (Table 3), and presented ORs unadjusted and adjusted for age, sex, Swiss-born, frequent travelling, contact with foreign born populations, and/or immunosuppression. We used univariate and multivariate logistic models to estimate the association between the degree of immunodeficiency and allopatric TB (Table 4), and presented ORs unadjusted and adjusted for age, sex, Swiss-born, frequent travelling, contact with foreign-born populations, and immunosuppression other than HIV infection. Finally, we determined statistical significance of HIV prevalence in patients with allopatric M. tuberculosis lineages compared to patients with sympatric lineages using Fisher's exact tests (Table 5). All analyses were performed in Stata version 11.1 (Stata Corporation, College Station, TX, USA).
In sensitivity analyses, we excluded patients with a history of frequent travelling to remove its influence on the association between HIV infection and allopatric lineages. In addition, we repeated the analyses using fully probabilistic Bayesian methods using weakly informative prior distributions [71]. The CIs reported from these analyses are 95% credible intervals and correspond to tail probabilities of the coefficient's posterior distributions. Bayesian statistics are less sensitive to errors when calculating estimators and CIs in small datasets.
We obtained 1,642 M. tuberculosis isolates from all TB cases (n = 1,940, 84.6%) notified in the Canton of Bern, Switzerland, between 1991 and 2011. For all patient isolates, we determined the main phylogenetic M. tuberculosis lineages. Of these, we included all patient isolates belonging to a non-Euro-American lineage (Lineage 1, 2, 3, 5 or 6) from European-born TB patients (40 of a total of 292 isolates belonging to lineages other than Lineage 4). Furthermore, we randomly selected control strains belonging to the Euro-American lineage (Lineage 4) from European-born TB patients (400 of a total of 1,350 isolates belonging to Lineage 4). European ancestry was confirmed in HIV–infected patients with an allopatric M. tuberculosis strain. Finally, we determined the HIV status in these patients using the same procedures as in the main sample.
The study was approved by the ethics committee of the Canton of Bern, Switzerland. Written informed consent was obtained from all patients enrolled in the SHCS. For patients outside the SHCS, written informed consent was obtained by the treating physicians. In some cases informed consent could not be obtained from the patient because he or she could not be located or was known to have died. For these cases we obtained permission from the Federal expert commission on confidentiality in medical research to use the data provided by the treating physician.
|
10.1371/journal.pcbi.1006925 | Colony entropy—Allocation of goods in ant colonies | Allocation of goods is a key feature in defining the connection between the individual and the collective scale in any society. Both the process by which goods are to be distributed, and the resulting allocation to the members of the society may affect the success of the population as a whole. One of the most striking natural examples of a highly successful cooperative society is the ant colony which often acts as a single superorganism. In particular, each individual within the ant colony has a “communal stomach” which is used to store and share food with the other colony members by mouth to mouth feeding. Sharing food between communal stomachs allows the colony as a whole to get its food requirements and, more so, allows each individual within the colony to reach its nutritional intake target. The vast majority of colony members do not forage independently but obtain their food through secondary interactions in which food is exchanged between individuals. The global effect of this exchange is not well understood. To gain better understanding into this process we used fluorescence imaging to measure how food from a single external source is distributed and mixed within a Camponotus sanctus ant colony. Using entropic measures to quantify food-blending, we show that while collected food flows into all parts of the colony it mixes only partly. We show that mixing is controlled by the ants’ interaction rule which implies that only a fraction of the maximal potential is actually transferred. This rule leads to a robust blending process: i.e., neither the exact food volume that is transferred, nor the interaction schedule are essential to generate the global outcome. Finally, we show how the ants’ interaction rules may optimize a trade-off between fast dissemination and efficient mixing. Our results regarding the distribution of a single food source provide a baseline for future studies on distributed regulation of multiple food sources in social insect colonies.
| We study how food is distributed in colonies of ants. Food collected by a small fraction of ants is distributed throughout the colony through a series of mouth-to-mouth interactions. An interesting interplay exists between food dissemination and food mixing within the colony. High levels of dissemination are important as they ensure that any food type is available to any ant. On the other hand, high dissemination induces homogenization which reduces the required variety of nutritional choices within the colony. Tracking fluorescent-labelled food and interpreting the results within the concepts of information theory, we show that food collected by each forager reaches almost every ant in the colony. Nonetheless, it is not homogenized across workers, resulting in a limited level of mixing. We suggest that the difference in food mixture held by each individuals can provide ants with the potential to control their nutritional intake by interacting with different partners.
| Food sharing in social insects is a compelling example of cooperation within a population [1–7]. Ants and bees can store a considerable amount of liquids in a pre-digestion storage organ called the ‘crop’ [8–10]. The stored food can later be regurgitated and passed on to others by mouth-to-mouth feeding (oral trophallaxis) [10–12]. Trophallaxis is a principal mechanism of food-transfer between individuals and therefore, the crop is often referred to as a “social stomach” [8].
When food is exchanged through trophallaxis, it is stored within the crop of the recipient workers and mixed with the rest of food in the crop [13–17]. Food blending is therefore an important factor in any process mediated by trophallaxis: from nutrient transfer and the maintenance of gestalt odor to hormonal regulation and information sharing [8, 13, 18, 19]. The extent to which food is blended in the colony has only been partially addressed before [3, 14, 20–22] and is still an open question.
Food blending is especially interesting in light of the fact that most colony members do not leave the nest [5, 14, 16, 23, 24], and all food is brought in by a a small fraction of workers called the foragers [16, 25]. The inter-relations between food-supplies brought in by different foragers can be expected to have an important role in the nutritional regulation of the colony. Social insect colonies have a documented ability to tightly regulate both the global nutritional intake [15, 21] and the dissemination of food to various sub-populations (such as nurses, larvae and brood) which may have different nutritional needs [5, 14, 16, 23, 26, 27]. The mechanisms that underlie this regulation are, however, not fully understood [28].
Trophallactic food exchange requires physical contact between ants. The dissemination process is therefore conveniently described by a time ordered network, in which ants are the nodes and the food transfers are the (directed) edges. The topology of this network provides the underlying infrastructure of the food-sharing process [17, 29–31]. In the study of social insects and other real-world networks, the topology of the network can frequently be traced while the details of particular interactions are concealed [32, 33]. Indeed, previous studies that traced individuals in a colony have mainly focused either on the network topology [29, 31] or on coarse grained descriptions of food dissemination [1, 16, 22, 26, 27, 34]. In this study we use single ant identification and fluorescently-labeled food (Fig 1) to measure not only the interaction network but also the flow of food over this network. For technical reasons, these experiments are conducted with a single food source. Characterization of this basic case is a first but necessary step towards more complex scenarios which include multiple sources.
The flow of food is limited by capacity: As the crop of ants is of finite size, this imposes a constraint on the amount of food that can be transferred in an interaction. This physical constraint limits the rate of mixing as ants become more and more full. Therefore, a potential trade-off between fast rate of food accumulation and well mixed outcome is expected.
The main objective of our study is using single ant measurement techniques to quantify how food brought in by different foragers blends as it is being disseminated across an ant colony. To this end, we use Shannon entropy to quantify the quality of mixing in an ant’s crop. The Shannon entropy provides a single quantity that reflects the relative abundances of multiple constituents [35] and therefore sets a common scale by which food homogenization can be evaluated from our empirical data. Using our detailed measurements we characterize the interaction network and the rules by which food flows across this network. We then use hybrid simulations to identify which of these characteristics function as regulators of food mixing, and which might play a lesser role. Finally, we employ a theoretical model to study the trade-offs between food dissemination and nutritional homogenization.
We studied food (sucrose solution [80g/l]) dissemination in Camponotus sanctus ant colonies residing in an artificial, single chamber nest and following famine relief (see Materials and methods, Experimental Setup). The dissemination process begins when the foragers, a small subgroup of the ants which we label F = { 1 , 2 , … , N foragers ≡ | F | }, return to the nest with liquid food loaded at the food source. Back in the nest, the foragers transfer the food to the non-forager population, A = { N foragers + 1 , N foragers + 2 , … , N ants }, via trophallactic interactions (Fig 1a). As food accumulates in the colony (Fig 1b–1d) it also flows between non-forager ants as they interact among themselves [17, 36].
The amount of food held in the crop of each ant as well as the amount of food passed per interaction were directly measured by combining single ant tracking with imaging of fluorescently labeled (Rhodamine B [0.08g/l]) food (Fig 1a, S1 Data) [17]. We designate the total amount of food in the crop of a non-forager ant a at time t by na(t) and the total amount of food held by all non-forager ants by Z ( t ) = ∑ a ∈ A n a ( t ). During the course of an experiment, the total amount of food held by the colony grows until it reaches saturation (S1e Fig) [1, 36]. The fraction of the total food held by ant a by Pa(t) = na(t)/Z(t) is not uniform across colony members (Fig 2a, S2 Fig) and is restricted by variable physiological properties such as crop capacity.
As a first step towards quantifying food mixing in the ant colony we took a forager-centric approach. The idea is to track how food brought in by each forager spreads across the colony (Figs 1(b)–1(d), 2(a) and 2(b) and S1c Fig) and the degree to which these food flows may overlap and mix. Since our experiments included a single food source we implemented this approach using a computational procedure in which we define the type of each ‘food droplet’ by the index of the forager, f ∈ F, that had initially collected it at the food source (see ‘Food tracking’, Methods). This entails that the number of ‘food types’ in the system is taken to be equal to the number of foragers. Using the assumption that mixing of food inside the crop of an individual ant is extremely rapid when compared to the rate at which food is transferred between ants, we then tracked the trajectories of labeled food droplets as they flow through the colony (see ‘Food tracking’, Methods). This procedure allowed us to define the empirically measured probability, Pa(t), described above (Pa(t) can be viewed as the probability that a randomly chosen ‘food-droplet’ is found within the crop of ant a), and consider the inferred joint probability Pf,a(t) = nf,a(t)/Z(t), which represents the probability that food, originally collected by forager f, is located in the crop of ant a at time t.
To quantify the degree to which different foragers contributed to the total foraging effort we calculate the total amount of food of type f that has accumulated in the colony up to time t as P f ( t ) = ∑ a ∈ A P a , f ( t ). This probability function may be associated with an entropy, which we refer to as the types-entropy (Htypes), and which quantifies the relative abundance of the different food types (for all entropy definitions refer to SI, ‘Mathematical Framework’ and ‘Table B, S1 Text’). It is defined by:
H types ≡ H ( F ) = - ∑ f ∈ F P f log ( P f )
(we suppress the explicit notation of time from here onward). Our measurements show that Htypes increases as a function of time (Fig 2c) and quickly approaches the upper bound of log ( | F | ). This upper bound can only be saturated if all foragers bring in equal amounts of food. As discussed below, Htypes sets a limit on the total level of mixing in the colony.
The degree to which food of a given type, i.e food brought in by a single forager—f, spreads across the colony can be quantified by the conditional distribution P(a|F = f) = Pf,a/Pf. We found that the food initially collected by each and every forager reaches, practically, all members of the colony (Fig 2b). This degree of dissemination dictates overlapping food flows such that the crops of non-forager ants must hold a mixture of food of several types (Fig 1b–1d).
Mixing was assessed by tracking the differently labeled food droplets as they flow, via the trophallactic network, from ant to ant. The conditional distribution P(f|A = a) = Pf,a/Pa signifies the mixture of food-types in the crop of a specific ant a (Fig 2). Since each non-forager ant receives its load from multiple interactions with both foragers and non-foragers [17, 31] the food composition in her crop, P(f|A = a), contains a mixture of differently labeled ‘droplets’.
The level of blending in the crop of each individual ant, a, can be defined by the crop entropy:
h mix a ≡ H ( F | A = a ) = - ∑ f ∈ F P ( f | A = a ) log [ P ( f | A = a ) ] .
The range of individual crop entropy, h mix a, is [ 0 , log ( | F | ) ] where zero crop entropy indicates that all food in the ants crop originates in a single forager while log ( | F | ) indicates that food in the crop is equally divided among all possible food types. We find that the (non-weighted) average mixing entropy (Fig 2d) takes an intermediate value of 0.79 of the maximal possible mixing. While the actual components that mix to create the crop of each ant vary greatly (Fig 2) we find that the degree of mixing is actually quite uniform across the colony (standard deviation of 0 . 2 · log ( | F | ), Fig 2d).
Mixing within the entire colony, as a whole, can be quantified by the conditional entropy, H(F|A). This global mixing entropy is defined as the weighted average over individual crop entropies, h mix a, where each ant is weighted by its relative load, Pa [35]:
H mix ≡ H ( F | A ) = ∑ a ∈ A P a · h mix a .
Mixing entropy is bounded from below by zero, a value which signifies no mixing (this can happen if the food in the crop of any ant originates from a single forager only). An upper bound on mixing is obtained by the general rule H(F|A) ≤ H(F) (conditioning reduces entropy [35]) which, in our notation, translates into the fact that the mixing entropy is smaller or equal to the types entropy (Hmix ≤ Htypes). Equality signifies perfect blending and occurs only when all ants have identical crop-load compositions that exactly match the concentration-distribution of food types across the entire colony.
We find that as the number of interactions grows so does the mixing entropy, Hmix (Fig 2c, S1 Fig). However, while the crop composition of a typical ant contains food that originated from each of the foragers, the relative proportions of these food types differ from ant to ant and do not match the proportions of food types flowing into the system (Fig 2c, S1 Fig). In other words, even though the types entropy (Htypes) designating the partition of the total food in the colony into types, does approach the maximal bound of log ( | F | ), the mixing entropy (Hmix) designating a similar partition on an individual level within each crop, is lower during the entire course of the experiment and reaches Hmix/Htypes = 0.8 ± 0.02 (mean ±std over three experiments) at the end of the experiments (Fig 2c). If the mixing entropy does eventually reach the upper bound of the types entropy the time for this to occur is very long.
To discern the causes of these intermediate mixing levels we next focus on the underlying dynamics of food exchange. In the following sections, we characterize the pairwise interactions via which food spreads through the colony and study their implications on mixing.
The flow of food across the colony can be described by focusing on two processes: 1) The interaction network which is the time-ordered depiction of the pairs of ants that engage in trophallaxis. 2) The interaction volume which depicts food exchange during an interaction in terms of both direction and volume. Next, we briefly characterize these two components.
Different aspects of the trophallactic interaction may limit food mixing in different ways. One way in which mixing levels may be reduced stems from the details of the interaction rule. As an extreme example: if the crop capacity of all ants was about equal and in any trophallactic event ants would transfer as much food as possible this would lead to pure food loads that are simply relayed between the ants and therefore minimal mixing. Decreased mixing may also be the result an interaction network which is topologically segregated into several disjoint communities with limited food flows between them (as reported for other ant species [4]). In this section, we describe hybrid simulations, which preserve some of the empirically measured data while replacing others by simulated values (for details see SI, ‘Simulations’), to separately examine the effects of the different aspects of the interaction details on overall mixing.
Finite crop size naturally impacts an ant’s ability to mix food. Mixture composition can significantly change only if an ant receives a large enough portion relative to her present load. Therefore, as ants become more satiated, their free storage space (i.e., the difference between her capacity and her current load) becomes smaller and the ability to mix (the potential mixing rate) declines. Consequentially, a fast accumulation rate might interfere with the mixing process.
As implied by the empirical interaction rule, in a receiving interaction, an ant is provided with a random volume of food that follows an exponential distribution, with an average that is proportional to her free storage space. This means that on average, an ant receives food in a series of decreasing volumes with a parameter δ. The parameter δ can thus be expected to have opposite effects on the accumulation and mixing of the food: the larger the value of δ the higher the accumulation rate and the lower the mixing rate (and vice versa).
We used a simple model to explore the possible trade-offs between the rate at which food accumulates within the colony and the extent to which it is mixed. For simplicity, the model assumes that all ants have the same capacity, that foragers and non-foragers use the same δ ˜ (in a deterministic version of the original food-transfer rule, see SI, ‘Simulations’) and that interactions occur randomly. Furthermore, for the purpose of the model, we defined the amount of food held by a forager at time t = 0 to equal the total amount of food she collects at the food source during the entire course of the experiment. This definition sets the amount of food summed over all colony members, M, as a quantity that is conserved over time. Considering the entire colony we now define the probability P ˜ a = n a ( t ) / M as the fraction of total amount of food held by any ant, forager or non-forager.
Using these definitions entails that at t = 0 all food is held by the foragers being, therefore, completely non-mixed while at later times, as food flows into the colony, it mixes within the crops of non-forager ants. This interplay between food accumulation and food mixing can be captured by considering the mixing entropy over all ants in the colony:
H mix overall = ∑ a ∈ A ∪ F P ˜ a h mix a
Note that since foragers receive almost no food from other workers (see above) we can approximate P(f′|a = f) ≈ 1 for f′ = f and zero otherwise. This means that h mix f = 0 for f ∈ F and leads to a second representation of H mix overall (see SI, ‘Trade-off model’):
H mix overall = P colony · H mix
where P colony = ∑ a ∈ A P ˜ a is the colony’s satiation level which starts off at 0 and saturates at 1 as food flows into the system [17] and Hmix is the mixing entropy over all non-forager ants, as defined above. This representation neatly separates the dissemination behavior into a component which quantifies the extent at which food is accumulated and a second component which quantifies the extent at which it is mixed.
We simulated an approximation to this model (see SI, ‘Simulations’) to study the relative effects of these terms as a function of the parameter δ ˜. The interactions of the simulations approximate the empirical data by keeping the average number of interactions per ant and the ratio between forager to non-forager and non-forager to non-forager interactions. As may be expected, larger values of the parameter δ ˜ lead to larger transfer of food into the colony (Pcolony indicated by the green line in Fig 5a). However, due to the finite capacity of an ant’s crop, larger values of δ ˜ also hamper mixing among non-foragers (Hmix indicated by the blue line in Fig 5a). The compromise between these two factors is captured by their product, the total mixing entropy H mix overall. This entropy exhibits a maximum for an intermediate value of δ ˜.
These results demonstrate a robust process: as long δ ˜ does not approach the extremes, both the mixing and the accumulation are comparable for a given number of interactions (Fig 5a). Surprisingly, even though higher δ ˜ will result in a higher accumulation rate, the ants seem to function at smaller δ ˜ values (red bars in Fig 5a). A potential benefit of smaller δ ˜ values is the maintenance of similar mixing levels across all ants in the colony (Fig 5b). This stands in agreement with our empirical evaluation of the variance in mixing levels across the colony (Fig 2d).
It is well known that social insects manage their nutrient resources on the collective level and also on finer scales because the colony channels foods with different nutritional composition to different sub-populations. In this paper, we put forward the idea that this intricate regulation relates to the interplay between food dissemination and food mixing within the colony. High levels of dissemination are important as they ensure that any food type is available to any ant. On the other hand, high dissemination induces mixing and this reduces the required variety of nutritional choices within the colony.
A main finding of this work is that, despite repeated trophallactic interactions between the ants, food in the colony does not become evenly mixed. Quantifying mixing using entropy measures we showed that, compared to what was theoretically possible, mixing is slow to rise and levels up at around 80% of the full mixing potential. The logarithms in the definition of entropy make the significance of this number difficult to assess. For intuition, in the case of only two food sources, the maximal mixing entropy (1 bit) corresponds to each crop holding equal parts of the food sources (1: 1) while 80% of this (0.8 bits) corresponds to, a far from perfect, 3: 1 partition of food sources. This imperfect mixing offers the possibility for receiving ants to choose from a wide spectrum of nutritional compositions when the donors provide different blends. Such choices can allow ants within the nest to reach their nutritional target using feeding schemes similar to those described by the geometrical framework for food foraging [42].
We further explored the mechanisms that allow for intermediate levels of food blending. Using hybrid simulations, we found that the interaction network over which food flows does not pose any limits on mixing levels. Rather, it is the interaction rule employed by the ants that regulates the extent to which food blends. This is reminiscent of several examples in which cellular pathways with identical architecture can achieve starkly different regulatory behaviors depending on actual rate coefficients [43, 44]. Regulation by interaction rules rather than by meeting patterns is an intriguing possibility for social insects in which different collective functions often reside over very similar interaction networks [29]. For example, while proximity is required for both food sharing and disease transmission [45] different interaction rules may ensure that one of these is enhanced while the other is suppressed.
Quantifying a large number of trophallactic interactions, we directly measured the food-transfer rule (see also [36]) used by the ants. We stress several important aspects of this rule. First, the rule respects the physical limits on crop size of the ants. Broadly speaking, this limit along with the fact that ants receive a substantial fraction of their free crop space per each interaction imply that an ant may become relatively full following her first few interactions. Thus, an ant’s mixing entropy is, to a large extent, determined by a small number of large events. Since these events are random both in order and in volume it is likely that mixing entropies will not saturate their maximal upper limit (see SI, ‘Entropy by largest events’, S5 Fig). Second, we show that the interaction rule is most likely stochastic in nature and, therefore, does not entail any strong requirements on ant cognition or communication. Finally, the fact that in trophallactic interactions the recipients fill only partially (Fig 3) is in agreement with a model in which, similar to animals foraging in the environment, ants in the nest regulate their nutritional income by feeding off of multiple partners each with a different mixture of the available ‘food types’.
We explored the interplay between food dissemination and mixing using a simple model of food flow that is based on our empirical observations. We find that the intermediate levels of mixing, as measured, can viewed as a compromise between the requirements to quickly unload incoming food and the requirement to disseminate different food types to all parts of the colony. We show that this process is robust over a wide range of δ values and that the actual measured parameter ensures that all ants in the colony are equally well mixed (although each holds a different particular mixture).
Finally, we wish to highlight the limitations of this study. Due to current technological availability, this work was performed using a single food source labeled by a single dye. The ants may behave differently in terms of both interaction network and food-transfer-rule when several food sources with different nutritional values are available [4]. For example, ants may modulate the amount of food they receive in a trophallactic interaction according to its nutritional value. Such modulation, which can be captured in an extension of our current model, can allow the ants to differentially regulate the flow of different nutritional types across the colony. Further, our artificial setup contained a single chamber nest. More realistic, multi-chambered, nest structure may induce interaction networks that are more clustered than the one measured here. This may hold important consequences for nutrition dissemination. Last, is our choice to measure mixing by labeling food types by foragers. While arbitrary, this is a reasonable choice since, as we have shown, foragers are responsible for a large part of the mixing (Fig 4b). Taking all these limitations into account we view our findings as a baseline to which future results, where multiple food sources are provided and tracked may be compared to.
Overall, our finding that the interaction rule takes precedence over the interaction schedule manifests both the robustness of collective processes within the ant colony and the large extent to which individual behaviors may modulate global outcomes.
For a more comprehensive methods section please refer to the SI and [17, 36]. Our experiments were conducted on lab colonies of Camponotus sanctus which included 50-100 workers, reared from single queens that were collected during nuptial flights in Neve Shalom and Rehovot, Israel. ‘Table A, S1 Text’ contains further details on each experimental colony.
The experimental setup consisted of an IR-sheltered artificial nest chamber (~100 cm2), neighboring an open area which served as a yard. The setup was recorded by two cameras (details in [17]): the top camera images were used to extract ant identities, coordinates and orientations using the BugTag software (Robiotec, Israel). The bottom camera images were used to detect fluorescent-labelled, using the openCV library in Python. Combining the information from both images, we associated between the identity of an ant and her appropriate fluorescent image. Thus, for each experiment a database was obtained, which included for every frame the coordinates, orientation, and measured fluorescence (in arbitrary units of pixel intensity) of each identified ant.
The experimental trophallactic network includes a time-ordered pairwise-interaction schedule, and the volume of liquids that one ant passed (received) to (from) the other (see S1 Data). Food is tracked from the moment it is acquired by a forger from the food source. We associate this volume (food ‘droplets’) with the forager’s barcode identity (‘type’), and continue tracing these droplets as they split between the ants according to the interaction schedule. To do this, we assume that in each interaction the receiver ant receives a fraction of the donor’s food in which the food type distribution is identical to that of the donor. In other words the number, n receiver ′ ( f ) of food droplets of type f in the receiver’s crop, following an interaction, is given by:
n receiver ′ ( f ) = n receiver ( f ) + v V donor · n donor ( f )
where nx(f) is the number of food droplets of type f in an ant’s crop before the interaction and v V donor is the fraction of the donors crop content that is transmitted during the interaction. The updated distribution of food types in the receivers’ crop following the interaction is given by:
P ( f | A = r e c e i v e r ) = n receiver ′ ( f ) ∑ ϕ ∈ F n receiver ′ ( ϕ )
Note that the number of food droplets per milliliter of food is arbitrary and cancels out in this calculation. The interaction networks for all three colonies including interacting ants’ identities, time, and interaction volumes can be found in the accompanying file ‘S1 Data’.
|
10.1371/journal.pcbi.1004886 | To Cooperate or Not to Cooperate: Why Behavioural Mechanisms Matter | Mutualistic cooperation often requires multiple individuals to behave in a coordinated fashion. Hence, while the evolutionary stability of mutualistic cooperation poses no particular theoretical difficulty, its evolutionary emergence faces a chicken and egg problem: an individual cannot benefit from cooperating unless other individuals already do so. Here, we use evolutionary robotic simulations to study the consequences of this problem for the evolution of cooperation. In contrast with standard game-theoretic results, we find that the transition from solitary to cooperative strategies is very unlikely, whether interacting individuals are genetically related (cooperation evolves in 20% of all simulations) or unrelated (only 3% of all simulations). We also observe that successful cooperation between individuals requires the evolution of a specific and rather complex behaviour. This behavioural complexity creates a large fitness valley between solitary and cooperative strategies, making the evolutionary transition difficult. These results reveal the need for research on biological mechanisms which may facilitate this transition.
| Mutualistic behaviours wherein several individuals act together for a common benefit, such as a collective hunt, are often deemed of minor interest by theoreticians in evolutionary biology. These behaviours benefit all the individuals involved, and therefore, so the argument goes, their evolution is straightforward. However, mutualistic behaviours do pose a specific kind of evolutionary problem: they require the coordination of several partners. Indeed, a single individual expressing a preference for cooperation cannot benefit if others wish to remain solitary. Here we use simulations in evolutionary robotics to study the consequences of this problem. We show that it constitutes a far more serious obstacle for the evolution of cooperation than was previously thought on the basis of game theoretical analyses. We find that the transition from solitary to cooperative strategies is very unlikely, and we also observe that successful cooperation requires the evolution of a specific and rather complex behaviour, necessary for individuals to coordinate with each other. This reveals the critical role of the practical mechanics of behaviour in evolution.
| It is well known that, in the absence of genetic relatedness, altruistic behaviours in which individuals pay a fitness cost for the benefit of others cannot evolve by natural selection [1, 2]. However, it is often assumed that mutualistic behaviours, wherein individuals collectively gain a common benefit [28, 29], do not pose such a problem, and are therefore of limited interest to evolutionists: they simply evolve because they benefit the individuals who express them.
However, mutualistic behaviours do often pose a different kind of evolutionary problem than altruism: they require coordination [3, 5–28]. Many collective traits are only mutually beneficial if several individuals express them together in a coordinated fashion. That is, it would not be beneficial for a single individual to express the cooperative trait if others did not express it as well. Consequently, whereas altruistic behaviours pose a problem of stability, which can only be solved by genetic relatedness, many forms of mutualistic behaviours pose a problem of evolution. These collective strategies are stable equilibria but their evolution is complex.
This problem has been formalized in game theory as the stag hunt game [6]. In the stag hunt, two hunters are confronted with the choice of either hunting a hare alone for a small but guaranteed benefit, or coordinating to hunt a stag cooperatively for a bigger reward, with the risk of not being rewarded at all if they hunt the stag alone. There are two evolutionarily stable Nash equilibria in this game: (1) simultaneous defection (i.e. both players hunt hares), which is risk-dominant as it maximizes the minimum payoff an individual can expect, and (2) simultaneous cooperation (i.e. both players hunt stags), which is payoff-dominant as it maximizes the total payoff at equilibrium. One of the aims of evolutionary analyses of the stag hunt is to characterize the mechanisms that facilitate the transition from the solitary equilibrium to the cooperative equilibrium. The difficulty is that cooperation can only be favoured by selection when a sufficient proportion of individuals in the population also cooperate. The transition from a population with a majority of solitary individuals to one with a majority of social individuals requires the rise of cooperation above an invasion threshold, which must occur for non-selective reasons.
In game-theoretic analyses, the hunting strategy of individuals is generally assumed to be encoded by a single genetic locus with two alleles: solitary or social [6]. In this case, random mutations and/or demographic stochasticity can lead to the appearance of a subpopulation of mutants playing the social strategy which is sufficient to overcome the invasion threshold. Moreover, Skyrms [6] showed that this cooperation can be further facilitated in a spatially structured population in which individuals tend to interact more with genetically related partners.
However, this approach makes a very strong assumption about the underlying mechanistic nature of behaviour: that a single mutation is sufficient to transform an individual playing a solitary strategy into an individual playing a perfectly efficient social strategy. In reality, hunting socially implies several novel behavioural abilities. In particular, it implies the ability to coordinate with others in order to focus on the same prey, which is unlikely to occur with only a single random mutation. In this paper, we postulate that critical aspects of coordinated cooperation have been neglected by game-theoretic analyses and investigate the mechanistic constraints which interfere with the evolution of coordination in a more realistic setting where the mapping between genotype and phenotype is not limited to a strict binary encoding.
Evolutionary robotics is a useful methodology for the simulation and study of this more realistic conception of behaviour and its genetic underpinnings [7, 8]. This approach allows to simulate the evolution of complex genotypes and observe the resulting behaviours in robotic agents. Such simulations also make it possible to investigate the complex mechanistic constraints at play in the translation from genotype to phenotype [24]. A considerable body of work has already been dedicated to modeling social evolution with robotic approaches [14]. These studies have been interested in a large diversity of issues: the evolution of swarms [11], the mechanics of division of labour in social insects [15, 16] or the evolution of communication [10, 17–19]. The evolution of cooperation in particular has been addressed in numerous papers. In the vast majority of this literature, however, social partners are genetically related [20], whether motivated by design [21, 22] or to study the evolution of altruism [9, 23]. Few articles, in comparison, have been interested in the evolution of mutualistic cooperation between genetically unrelated individuals [19]. Moreover the specific problem posed by the stag hunt game, where cooperation is not the only evolutionarily stable strategy and a non-collective solution acts as a stable attractor, has never been studied in evolutionary robotics.
In this paper, we use an experimental model where simulated robotic agents interact in a situation equivalent to the stag hunt and compare the results of our model to those of standard game-theoretic analyses. Our results shed new light on the influence of mechanistic constraints in the evolution of coordinated actions. We then use this model to explore realistic mechanisms that could drive the transition to collective behaviours.
We consider an environment with two hunters and several prey, both hares and stags. Hunters can choose to hunt either of these prey, earning different food rewards depending on whether they hunt alone or cooperate (see Fig 1).
Food rewards for killing a prey are shown in Table 1. A hare yields a reward of 50, regardless of whether it is hunted in a solitary or cooperative fashion. A stag yields a reward of 500 for each hunter only if it is hunted cooperatively. If a stag is killed by a single hunter, it is still removed from the arena but is considered a failed hunt and rewards nothing. None of the rewards are split between cooperators.
Simulated robotic agents are evaluated in an 800 by 800 unit square arena, which has four solid walls and is devoid of any obstacles aside from other agents. Each circular-shaped agent, with a diameter of 14 units, is equipped with two independent wheels and a collection of sensors. Hunters can use the information provided by 12 proximity sensors and a front camera. Proximity sensors have a range of approximately twice the diameter of the agent’s body, and provide the agent with the proximity of the nearest obstacle. They are evenly distributed around the agent’s body. The front camera consists of 12 rays with infinite range spread out in a 90 degree cone in front of the body. Each ray in the camera provides two different pieces of information about the first target it intersects with: the type of target (hunter, hare, or stag) and its proximity. This robot model facilitates the evolution of basic walls avoidance and agents recognition behaviours, which we consider not to be of interest here. Hence we separate obstacles recognition (by the proximity sensors) from agents’ recognition (by the camera).
Only the hunters are capable of movement; prey remain at their initial positions. (Complementary experiments with moving prey capable of avoidance behaviours did not produce significantly different results; not shown.) A prey is caught if any hunter remains close enough during a fixed amount of time steps (800 steps, in a simulation lasting 20.000 time steps). Cooperative hunting is defined as a prey with two hunters in catching distance at the time of its capture. Therefore, cooperation happens even if only one of the two hunters is in catching distance of the prey for most of the time, as long as the two hunters are there in the final step. The prey is then immediately replaced at a random position in the arena, thus keeping a fixed number of agents and prey during the whole simulation.
The hunters’ behaviour is computed by an artificial neural network which maps sensory inputs to motor outputs. The neural network is a fully connected multi-layer perceptron with a single hidden layer of 8 neurons. The inputs of this network are the perceptions of the agent, with 12 neurons for the proximity sensors and 48 for the camera (4 for each of the 12 rays) plus a bias neuron (whose value is always 1), for a total of 61 input neurons. The two outputs of the network control the speed of each of the agent’s wheels and the mapping function between inputs and outputs is a sigmoid function (see Fig 2). Changing the number of hidden neurons did not yield significantly different results (not shown).
To simulate evolution, we use an evolutionary algorithm to evolve the genome of the hunters. This genome is comprised of a collection of 410 real values in the range [0, 1], one for each of the neural network’s weights, and is initially randomized for each individual in the population. In order to obtain its fitness, each individual is successively paired five times with a partner randomly chosen each time (except itself) in the arena presented in the Experimental Setup subsection, for an evaluation round of 20.000 time steps. The payoff of the evaluated individual at the end of a round is given by the total amount of food it has managed to obtain by killing prey in this round. As this quantity depends heavily on the initial conditions (random initial positions of the prey), five simulations are performed for each pair of individuals. The individual’s fitness is then obtained by computing the sum of payoffs averaged over the total number of simulations for the individual. In this case the number of simulations is 25, with 5 partners and 5 simulations with each partner.
Experiments were conducted using a Wright-Fisher model [12] with constant population size (20 individuals), which is commonly known as a fitness-proportionate selection method in evolutionary robotics [13]. Using this model, the population of the next generation is formed by a random sampling of offspring from the previous generation, with the probability of sampling a particular parent proportional to the parent’s fitness. Each offspring is simply a mutated clone of its parent; recombination is not included in our simulation. Consequently, new genotypes appear only through mutation. These mutations are performed using a Gaussian function, with a standard deviation of 2 × 10−1 and a mutation probability of 5 × 10−3. Each experiment lasted 3000 generations. All simulation parameters are summarised in Table 2.
In order to explore the evolutionary transition between the risk-dominant equilibrium (hare hunting) and the payoff-dominant equilibrium (cooperative stag hunting), individuals first evolved in an environment composed solely of hares. This ensured that the populations initially reached the solitary equilibrium. Only then did we add stags and study the dynamics of evolution. Fig 3(a) shows the evolution of the mean percentage of stags hunted successfully (i.e., hunted cooperatively) out of the total number of prey hunted over time for 30 independent runs. Fig 3(b) shows the mean proportion of each type of prey hunted during the last generation of each run. Stag hunting evolved in only one run out of 30 and even in that run accounted for less than 30% of the total number of prey hunted. In the other 29 runs, the individuals hunted only hares as they had previously evolved to do. These simulations demonstrate that the evolution of collective hunting is very unlikely when the population is composed of individuals who are already efficient solitary hunters.
For comparison we simulated the same scenario using the standard game-theoretic version of the stag hunt, where the expression of the two types of behaviour was encoded by a single binary locus. Each individual in the population initially possessed the allele for hare hunting (Fig 4).
Here the transition to collective hunting occurred in each of the 30 independent runs and this strategy then remained stable. This result differs drastically from the results of our robotic simulations in which this transition never fully occurred (Mann-Whitney U test on the proportion of stags hunted successfully during the last generation, p-value <0.001).
In a second experiment, we wanted to investigate the evolution of hunting strategies “from scratch”, with the individuals’ genotypes initialized with random values, rather than evolved with a specific hunting strategy. Fig 5 shows the mean percentage of stags hunted over time and the mean number of prey hunted during the last generation. We observed the transition to a clearly cooperative strategy in a single run, while in two other runs, 50% of prey hunted were stags. In the 27 remaining runs the proportion of stags hunted was less than 25%. In comparison, in simulations using the standard game-theoretic version of the stag hunt where individuals are initially unable to hunt, stag hunting evolved and remained stable in every run (see supporting information S1 Fig).
The above experiments show that mechanistic constraints have a critical effect on the evolution of coordinated collective actions. In a simple game-theoretic analysis in which the hunting strategy is encoded by a single binary gene, collective behaviour systematically evolved. However, in a setting where the hunting strategy was determined by a more complex artificial neural network, cooperative behaviour evolved in fewer than 10% of cases. These results encourage further exploration into the evolutionary origin of coordinated collective actions and the mechanisms which may facilitate their evolution. In the following section, we explore two such mechanisms.
In the next experiment, food was also rewarded for hunting a stag in a solitary fashion so that cooperative behaviour did not entail a risk. We wanted to study whether hunting a stag alone could act as a transition towards the evolution of the collective strategy. Hunting a stag alone was given the same reward as hunting a hare (Table 3), differing from classical models of the stag hunt.
Fig 6 shows the results of robotic simulations where individuals initially evolved to hunt hares (as in Fig 3). As expected, the evolution of collective hunting was significantly facilitated when the risk of hunting stags alone was removed (Mann-Whitney, p-value <0.001). The populations completely switched to hunting stags in two runs out of 30, and in three other runs, more than 50% of the prey hunted were stags, with a large part of the prey hunted cooperatively in each of these runs. However, in most of the runs (25 out of 30), the evolved strategy was to hunt both types of prey in a solitary fashion. From these results, it entails that the individuals are still hindered by the evolution of a successful coordination strategy.
Genetic relatedness among social partners is known to influence the evolution of many types of social traits [1]. In particular, [6] showed how it can facilitate the evolution of cooperation in a stag hunt game [6, chapetr 3]. It can yield more frequent interaction between cooperators, which in turn increases their probability of benefiting from cooperative behaviour. In order to include this mechanism, we considered an extreme situation in which each individual is always paired with a clone of itself, known as “clonal selection” in robotics, ensuring a maximal genetic relatedness of 1.
These results show that genetic relatedness has a positive effect on the evolution of cooperation (Fig 7). In four out of 30 runs the population evolved the cooperative strategy. Moreover, in two other runs, stags accounted for more than 75% of prey hunted, as compared to less than 25% without relatedness (Mann-Whitney, p-value <0.005). When the initial population was random, rather than only hare hunters (see supporting information S2 Fig), the positive effect of genetic relatedness was also observed in 12 out of 30 runs, where more than 50% of prey hunted were stags.
There is a profound difference between evolutionary game-theoretic and robotic simulations of the stag hunt. Using identical model parameters, the transition from the solitary equilibrium to the social equilibrium always occurred in game-theoretic simulations, but was extremely unlikely in robotic simulations, occurring in 1 run out of 30. The complexity of the mapping between genotype and phenotype is responsible for much of this contrast. Individuals involved in a coordination game such as the stag hunt face a chicken & egg problem: the cooperative behaviour must be beneficial in order to evolve, but no individual can benefit from this behaviour unless the behaviour is already expressed by other individuals. When binary variation at a single genetic locus encodes the expression of the solitary or cooperative strategy, a single mutation is sufficient for a cooperative mutant to appear in a resident population of solitary individuals. In a finite population, demographic stochasticity can then lead to the rise of cooperators above the invasion threshold, at which point natural selection leads to their fixation, switching from a solitary equilibrium to a social one. In contrast, in our robotic simulations, the mapping between genotype and phenotype is more complex. Adopting the social strategy entails both a modification of the preferred hunting target and the ability to coordinate with others. Thus, several mutations are necessary for the appearance of full-fledged cooperative behaviour. As several individuals must carry these multiple mutations for the behaviour to become beneficial, the transition to the cooperative equilibrium is nearly impossible.
In particular, in our robotic simulations we were able to observe that coordination entails a specific and rather complex behaviour. Fig 8 (see also supporting information S1 Movie) shows the behaviours evolved by the best individuals in the cooperative run shown in Fig 5 (Run 9). The solution they evolved for coordination was to circle around one another, allowing each of them to constantly see their partner while both moving closer to a stag. This behaviour was replicated in every cooperative run. We thus observed the evolution of an ingenious (given the agents’ limited capabilities) and complex hunting strategy. These findings demonstrate that the practical mechanics of behaviour can have important evolutionary consequences, and that models which ignore these properties may lead to misleading predictions.
Moreover, the evolution of cooperation is also strongly impacted by ecological features. Social hunting poses a bootstrapping problem because it entails both a modification of the preferred hunting target and an ability to coordinate with others. Its evolution can be facilitated, therefore, if hunters have a reasonable probability of hunting the same prey as their partner, just by chance, with no need of active coordination. Biologically, this could occur if hunters live in a dense social environment (with many other hunters in the vicinity), and/or if the density of prey is low, such that the likelihood of ending up on the same prey is large. To test this possibility, we conducted additional experiments where the density of prey was varied. The number of prey was whether (1) decreased from 18 to 6 or (2) increased from 18 to 30. The population was initially constituted of hare hunters and we kept the same ratio of prey as in previous experiments (i.e. 50% of hares and 50% of stags). We show (see supporting information S3 Fig(a)) that when the number of prey is decreased (6) the transition to a cooperative strategy is facilitated (Mann-Whitney, p-value <0.05) as in 9 runs out of 30, more than 30% of the prey hunted are stags. In comparison, a higher density of prey (30) entails that it is impossible to evolve cooperation (see supporting information S3 Fig(b)). These results reinforce our claim that the practical mechanics of coordination are crucial in understanding the evolution of cooperation. In particular, here, the precise ecological situation faced by individuals plays a key role in the transition to the collective equilibrium.
Finally, the complexity of coordination suggests that the recycling of a previously evolved trait could be necessary for the transition to cooperation, i.e. individuals could coordinate thanks to behavioural features that may not have been selected for cooperation at first. Such features could include the evolution of communication, or a leader-follower strategy. The role of both of these behaviours has already been studied in real-life stag hunt type interactions in chimpanzees and human children [25, 26], and there is an already extensive literature in evolutionary robotics on their role in the evolution of collective actions [16, 19, 22, 27]. This offers some directions for future works on this problem.
|
10.1371/journal.pgen.1007790 | miR-3607-3p suppresses non-small cell lung cancer (NSCLC) by targeting TGFBR1 and CCNE2 | Accumulating evidence indicates that miRNAs can be promising diagnostic and/or prognostic markers for various cancers. In this study, we identified a novel miRNA, miR-3607-3p, and its targets in non-small cell lung cancer (NSCLC). The expression of miR-3607-3p was measured and its correlation with patient prognosis was determined. Ectopic expression in NSCLC cells, xenografts, and metastasis models was used to evaluate the effects of miR-3607-3p on proliferation and migration of NSCLC. Luciferase assay and western blotting were performed to validate the potential targets of miR-3607-3p after preliminary screening by microarray analysis and computer-aided algorithms. We demonstrated that miR-3607-3p was downregulated in NSCLC tissues and that miR-3607-3p might act as an independent predictor for overall survival in NSCLC. Moreover, serum miR-3607-3p may be a novel and stable marker for NSCLC. We found that overexpression of miR-3607-3p inhibited cell proliferation, colony formation, migration and invasion, and hampered the cell cycle of NSCLC cell lines in vitro. Our results suggested that miR-3607-3p directly targets TGFBR1 and CCNE2. In accordance with in vitro studies, we confirmed that miR-3607-3p functions as a potent suppressor miRNA of NSCLC. We showed that miR-3607-3p agomir could reduce tumor growth and inhibit TGFBR1 and CCNE2 protein expression. Taken together, our findings indicate that miR-3607-3p can inhibit NSCLC cell growth and metastasis by targeting TGFBR1 and CCNE2 protein expression, and provide new evidence of miR-3607-3p as a potential non-invasive biomarker and therapeutic target for NSCLC.
| We first showed downregulation of miR-3607-3p in NSCLC tissues and demonstrated that miR-3607-3p may act as an independent predictor of overall survival. Serum miR-3607-3p may be a novel marker of patients with NSCLC. We further found that miR-3607-3p possesses the potency to suppress NSCLC growth and metastasis and induce cell cycle arrest by regulating TGFBR1 and CCNE2. Importantly, agomiR-3607-3p could reduce tumor growth and lung and brain metastasis and inhibit TGFBR1 and CCNE2 protein expression. Our findings suggest that miR-3607-3p is a tumor suppressor in NSCLC and holds promise as a prognostic biomarker and potential therapeutic target for NSCLC.
| Lung cancer is one of the most frequently diagnosed cancers and the leading cause of cancer death in developed and developing countries [1]. It was estimated that there were 1,241,600 new cases of lung cancer in males and 583,100 in females in 2014 [1]. In China, it is estimated that approximately 733,300 new cases were diagnosed and 610,200 patients died of lung cancer in 2015 [2]. Approximately 85% of lung cancers are classified histopathologically as non-small cell lung carcinoma (NSCLC) [3]. In contrast to the steadily improving survival for most cancers, the 5-year survival of lung cancer is only 18% [4]. Poor outcomes and frequent relapses associated with lung cancer urgently demand the development of new screening and early biomarkers for the accurate and non-invasive detection of lung cancer metastasis and recurrence [5]. Since pathologically diagnosing all suspicious nodules is currently impossible, a noninvasive and easy sampling strategy that provides reliable information on the metastatic state of NSCLC is urgently required.
MicroRNAs (miRNAs) are small non-coding RNAs of 18–25 nt and inhibit mRNA translation and/or negatively regulate mRNA stability by binding to the 3’-untranslated region (3’-UTR) of target mRNAs [6]. The “seed” region of a miRNA is the most important known determinant of the miRNA’s ability to recognize its target mRNA [7]. Depending on the target mRNAs, miRNAs can act as either tumor oncogenes or tumor suppressors in cancers. Recent studies suggested that miRNAs are involved in a number of biological processes such as development, differentiation, proliferation, and apoptosis [8]. Dozens of miRNAs (such as miR-143/145, miR-21 and miR-34) play essential roles in lung tumorigenesis by regulating critical oncogenes or tumor suppressors [9–11]. Accumulating evidence has demonstrated that miRNAs play a critical role in either the progression or prognosis of NSCLC [12]. Several studies have shown that miRNAs could be used as diagnostic and prognostic biomarkers. For example, miR-195 expression was shown to be lower in NSCLC tissues and associated with poor survival [13]. In colorectal cancer, high expression of miR-135b and low expression of miR-590-5p are associated with clinical stage, progression and survival [14,15]. Nevertheless, most of these studies are sample and examined relatively limited numbers of miRNAs [12,16]. miR-3607-3p has been reported to be significantly attenuated and act as a crucial tumor-suppressive miRNA in prostate cancer [17]. On the other hand, Lin et al. showed that miR-3607-5p was upregulated in lung cancer tissues and cells and that miR-3607-5p overexpression promoted lung cancer A549 cells proliferation by inhibiting adenomatous polyposis coli (APC) [18], highlighting that different miRNAs from the same precursor may have different roles in different types of cancers. Since the mature miR-3607-3p and -5p species could origin from opposite sides of the miRNA hairpin and contain complex post-transcriptional miRNA regulation, it is possible that they play overlapping roles or possess different biological functions in NSCLC progression.
Therefore, further study is necessary to elucidate this point. Furthermore, the precise molecular mechanism through which miR-3607-3p influences NSCLC progression remains largely unknown and further investigations are required. In addition, we were curious to determine if miR-3607-3p was associated with NSCLC in the same way as miR-3607-5p does. In this study, we identified a novel miRNA, miR-3607-3p, and its role in signaling pathways involved in the pathogenesis of NSCLC. This miRNA could be a promising biomarker and therapeutic target for NSCLC.
To determine the potential functions of miR-3607-3p in NSCLC pathogenesis, we analyzed the expression of miR-3607-3p in 162 NSCLC tissue samples compared with their adjacent normal lung tissues. miR-3607-3p staining in NSCLC tissues was negative or weak relative to normal the adjacent normal lung tissues that exhibited light to dark staining (Fig 1). In contrast, miR-3607-3p signals were confined to scattered and positive images in lung tissues (Fig 1A and 1B). In situ hybridization showed that miR-3607-3p expression was significantly downregulated in tumor tissue samples compared with control samples, especially for NSCLC stages III and IV (Fig 1C and 1D, Table 1). In comparison, miR-3607-5p expression was significantly upregulated in 93 tumor tissue samples compared with control samples by in situ hybridization (S1 Fig). Downregulation of miR-3607-3p was associated with lymph node metastasis and NSCLC stage (Table 1). The Kaplan-Meier analysis indicated that patients with low miR-3607-3p expression had significantly shorter survival (commercial tissue microarray, n = 162; Fig 2A).
qRT-PCR showed that miR-3607-3p levels in NSCLC tissue samples were markedly lower than in normal lung tissue samples (n = 107; Fig 2B), particularly in samples from patients with lymph node metastasis and advanced clinical stages of NSCLC (Fig 2C and 2D, Table 1). Kaplan-Meier survival analysis also revealed that miR-3607-3p downregulation was associated with poor prognosis in patients with NSCLC (specimens from patients, n = 107; Fig 2E). Multivariate Cox regression analysis showed that the miRNA signatures were an independent prognostic factor for OS of NSCLC (Table 2).
qRT-PCR showed that serum expression levels of miR-3607-3p were significantly lower in NSCLC patients than in normal controls (Fig 2F, Table 3). Serum miR-3607-3p expression levels were negatively associated with lymph node metastasis and advanced clinical stages of NSCLC (Fig 2G and 2H, Table 3). The area under the curve (AUC) for plasma miR-3607-3p was 0.810, indicating that there was an obvious significance in NSCLC diagnosis by serum miR-3607-3p (Fig 2I).
Among the six NSCLC cell lines and normal human embryo lung fibroblast cell line (MRC-5), the H157 and H292 cell lines had the lowest expression of miR-3607-3p (Fig 3A) and the highest metastatic potential (S2 Fig); therefore, they were selected for the overexpression experiment. To further investigate the role of miR-3607-3p in the regulation of NSCLC cell proliferation, colony formation, invasion, and migration, H157 and H292 cells were transfected with an miR-3607-3p mimic and miR-3607-3p levels were examined using qRT-PCR. The efficiency of transfection was verified as a significant increase in miR-3607-3p expression in H157 and H292 cells by qRT-PCR (Fig 2B and S3A Fig). High exogenous expression of miR-3607-3p remarkably inhibited proliferation, colony formation, migration, and invasion of H157 and H292 cells (Fig 2C–2E and S3B–S3D Fig). Similarly, we observed that the exogenous high expression level of miR-3607-3p remarkably inhibited proliferation and invasion of normal human embryo lung fibroblast MRC-5 cell line (S4A–S4C Fig). To explore the possible mechanisms underlying the inhibitory effect on cell growth by overexpression of miR-3607-3p, a cell cycle analysis was performed. Upon upregulation of miR-3607-3p, the percentages of H157 and H292 cells in G0/G1 phase increased compared with the percentages measured in controls (Fig 3F and S3E Fig), indicating that overexpression of miR-3607-3p resulted in G1 phase arrest in NSCLC cells.
Next, we transfected NSCLC cells with inhibitors of miR-3607-3p to confirm the opposite results of mimic transfection (Fig 4A and S5A Fig). As expected, downregulation of miR-3607-3p using the inhibitors could enhance the malignant phenotype of A549 and H1299 cells in vitro, including cell growth (Fig 4B and S5B Fig) and colony formation (Fig 3C and S5C Fig), as well as cell migration and invasion (Fig 3D and S5D Fig). We further observed that forced expression of miR-3607-3p could decrease the number of cells in G1 phase and increase the numbers of cells in S phase (Fig 3E and S5E Fig). Hence, miR-3607-3p downregulation could be promoting cell proliferation.
To explore the mechanism through which miR-3607-3p regulates NSCLC cell progression, we searched for potential downstream targets of miR-3607-3p using several bioinformatics methods, including miRWalk, TargetScan, and miRDB (Fig 5A). Several candidate genes involved in proliferation, cycling and invasion-metastasis were identified by gene ontology (GO) terms and qRT-PCR. We found that TGF-β receptor 1 (TGFBR1) and cyclin E2 (CCNE2), two key proteins involved in TGF-β signaling and the cell cycle pathway, appeared to be potential targets of miR-3607-3p (Fig 5B). To verify whether TGFBR1 and CCNE2 are direct targets of miR-3607-3p, we transfected an miR-3607-3p mimic into cells and observed that this could markedly down-regulate the mRNA and protein levels of TGFBR1 and CCNE2 in H157 and H292 cells, respectively (Fig 5C and 5D). We also transfected NSCLC cells with inhibitors of miR-3607-3p to confirm the results of mimic transfection. As expected, downregulation of miR-3607-3p using inhibitors could enhance the TGFBR1 and CCNE2 mRNA and protein levels in A549 and H1299 cells (Fig 5E and 5F). We next applied the dual-luciferase reporter method to reveal the regulation of miR-3607-3p by TGFBR1 and CCNE2. The fragments containing the miR-3607-3p binding sequence or mutated sequence in the 3’UTR regions of TGFBR1 and CCNE2 were cloned into the pmiR-RB-REPORT vector luciferase reporter. These reporter constructs were co-transfected with miR-3607-3p mimic or miR-NC into H157 and H292 cells, and the luciferase activities were subsequently measured. The miR-3607-3p mimic significantly suppressed the luciferase activity of pmiR-RB-REPORT-TGFBR1 or CCNE2-3’UTR (Fig 5G and 5H), while miR-NC had no inhibitory effect on pmiR-RB-REPORT TGFBR1 or CCNE2-3’UTR. The miR-3607-3p inhibition of pmiR-RB-REPORT-TGFBR1 and CCNE2-3’UTR was sequence-specific because the luciferase activities of pmiR-RB-REPORT-TGFBR1 or CCNE2-mut did not decrease in the presence of miR-3607-3p. Taken together, these results suggest that miR-3607-3p can directly target the 3’-UTR of TGFBR1 and CCNE2.
A rescue experiment was performed to confirm that TGFBR1 or CCNE2 was the functional target of miR-3607-3p in H157 and H292 cells. The evidence was obtained from the observation that the TGFBR1 or CCNE2 mRNA and protein (endogenous) in the two cell lines were diminished by mimic transfection and recovered by transfection of both pEGFP-N1-TGFBR1 or CCNE2 expression constructs, respectively (Fig 6A–6C). The results showed that migration and invasion created by mimic transfection were reversed by transfection of both expression constructs (Fig 6D and 6E).
Finally, we evaluated the effects of miR-3607-3p on the growth and metastasis of NSCLC in nude mice. A549 cells were transfected with either a lentiviral expression vector to knock down miR-3607-3p or a negative control lentiviral vector. The downregulation of miR-3607-3p in the A549 cells following lentiviral infection was verified by qRT-PCR (Fig 7A). Then, we injected these A549 cells subcutaneously in BALB/c nude mouse to induce tumors. Beginning on day 7 after injection, the tumor lengths and widths were measured every 5 days for 40 days. The tumor growth curve revealed a significantly higher growth rate in the miR-3607-3p-downregulated group compared with the control group (Fig 7B). Subsequently, the tumors were dissected and the exact dimensions and weights were evaluated. Compared with the control group, the tumors in the miR-3607-3p-downregulated group were larger (Fig 7C and 7D).
Luciferase-labeled cells (106) were injected intravenously in the tail vein of mice and the animals were sacrificed 6 weeks later. Luciferase activity was used to evaluate the tumor burden in the lung and brain. The lung metastasis burden was significantly higher in the mice injected with miR-3607-3p-knockdown cells compared with the control group (Fig 7E and 7F). As expected, knockdown of miR-3607-3p significantly increased the brain metastasis through tail vein injection (Fig 7E and 7G). All these results obtained for the mouse models suggest that miR-3607-3p plays important roles in NSCLC growth and metastasis in vivo, particularly in lung and brain metastasis. To determine whether miR-3607-3p could inhibit the growth of NSCLC in nude mice, we established nude mouse tumorigenic and metastatic model using A549 cells, as above. After 7 days, miR-3607-3p agomir or miR agomir NC were directly injected into the subcutaneous tumor every 5 days for 35 days. The tumor volume was measured every 5 days until day 42. The tumor volume and weight of mice treated with miR-3607-3p agomir were significantly greater than those of mice treated with miR agomir NC (Fig 7H–7J). This result indicated that miR-3607-3p significantly inhibits the tumorigenicity of A549 cells in the nude mouse model.
In addition, the proliferative activities of the tumor cells were assessed by immunohistochemistry for Ki-67 in FFPE tissues of xenograft tumors. The Ki-67 staining was decreased in tumors from the miR-3607-3p agomir group (Fig 7K). Moreover, a distinct decrease in TGFBR1 or CCNE2 expression was observed in IHC slices of the miR-3607-3p agomir group compared with the expression observed for the miR agomir NC group (Fig 7K). TGFBR1 or CCNE2 expression was significantly higher in tumor tissues than in adjacent non-tumor tissues (n = 70; Fig 7L). miR-3607-3p expression was inversely correlated to the expression of TGFBR1 and CCNE2 in NSCLC specimens (n = 70; Fig 7M and 7N).
MicroRNAs have recently been demonstrated to contribute to carcinogenesis and the progression of various cancers and may provide new therapeutic strategies such as the use of biomarkers and therapeutic targets [19,20]. In this study, we found that miR-3607-3p, an intronic miRNA located at chromosomal position 5q14.3, is frequently downregulated in human NSCLC specimens. Our analyses suggest that low miR-3607-3p expression can be an important biomarker for discriminating between normal lung and tumor tissues. Correlations with clinicopathological parameters suggest that downregulation of miR-3607-3p is associated with tumor progression in NSCLC. Low miR-3607-3p expression was significantly associated with NSCLC of higher stages and with lymph node metastasis. These results are consistent with recent findings indicating that expression levels of tumor-suppressing miRNAs are often downregulated in NSCLC.
Many miRNAs have been associated with tumor-suppressive effects in NSCLDC and other types of cancer. For example, miR-195 expression has been shown to be lower in tumor tissues and was associated with clinical stages of NSCLC [13]. miR-720 acts as a tumor suppressor by inhibiting cell migration and invasion, and was found to be downregulated in primary breast carcinoma [20]. miRNAs have also been extensively investigated as prognostic factors [21]. Zhou et al. also demonstrated that high expression of miR-574-5p in serum was an independent factor of poor prognostic in patients with SCLC [22]. The present study suggests that miR-3607-3p is an independent prognostic factor for OS in NSCLC. In addition, our results showed that lower expression of miR-3607-3p might be associated with poor overall survival in NSCLC. Poor prognosis of patients with early-stage lung cancer is associated with lymph node metastasis and distant metastasis at the time of presentation [23]. Regarding the role of miR-3607 in other cancer types, Saini et al. [17] showed that the miR-3607-3p region is frequently deleted in prostate cancer, highlighting potential tumor-suppressive roles of this family. A recent study showed that CirclRAK3 sponges miR-3607-3p and facilitates breast cancer metastatic spread [24]. In hepatocellular carcinoma, low expression of miR-3607-3p predicts poor prognosis [25]. Collectively, these findings suggest that miR-3607-3p may be a novel biomarker for NSCLC prognosis and diagnosis. However, Lin et al. [18] reported that miR-3607-5p had opposite roles compared with miR-3607-3p. Since the sequences of the two miRNAs are different, it is possible that they possess different roles. Indeed, it is accepted that the mature miRNA -3p and -5p species originate from the opposite sides of the same miRNA hairpin and have complex post-transcriptional miR-3607 regulation. Therefore, both species would regulate a unique set of targets and possess very different biological functions. Hence, further study is necessary to elucidate this point. Indeed, miR-3607-3p has been reported to be significantly attenuated in prostate cancer [17]. On the other hand, miR-3607-5p was upregulated in lung cancer and miR-3607-5p overexpression promoted lung cancer A549 cells proliferation [18]. Those two studies support the results of the present study, but additional research is still necessary to examine this dual effects of the same miRNA.
Endogenous circulating miRNAs could be significant for the diagnosis, prognosis, and metastasis of cancer. Our results suggest that serum miR-3607-3p levels might be useful in delineating lung cancer stages because of decreasing expression of miR-3607-3p in higher-stage cancers. Indeed, the AUC value was 0.81 in NSCLC patients compared with normal controls. These findings indicate that miR-3607-3p levels could hold great diagnostic potential, with high sensitivity and specificity. In other words, serum miR-3607-3p level could be a good noninvasive biomarker for NSCLC diagnosis. Tumor-derived miRNAs were first described in plasma by Mitchell et al. [26]. Accumulating evidence suggests the potential of miRNAs in the early detection of several malignancies such as lung cancer [27], breast cancer [28], and gastric cancer [29]. In addition, serum miR-146a has been reported in papillary thyroid carcinoma [30], lung cancer [31], and gastric cancer [32], which is associated with diagnosis and can be used for early detection of tumors.
Differences in miRNA expression in tumor samples do not necessarily alter the function of tumor cells. Our in vitro results suggested that miR-3607-3p mimic suppressed the proliferation, colony formation, and migration abilities of NSCLC cells, while transfection of miR-3607-3p inhibitor had the opposite effects. Lower levels of miR-3607-3p not only enhanced cell viability but also promoted colony formation and cell migration. Importantly, data from the present study revealed that upregulation of miR-3607-3p reduced tumor growth and metastasis in vivo by delivery of the agomir into cancer cells and the tail vein in nude mice. Recent studies have indicated that miRNAs play a potential role as tumor suppressors in NSCLC and other types of cancers [33,34]. Joshi et al. demonstrated that miR-148a might act as a tumor suppressor and that the miRNA inhibited the migration and invasion of the A549 NSCLC cell line [5]. Yu et al. conducted an in vitro study of NSCLC cell lines and provided evidence that cell proliferation, migration and invasion could be evaluated by the transfection of miR-520a-3p [35]. Similarly, the roles of miR-675-5p in regulating tumor growth and metastasis in vivo have been demonstrated in previous studies [27]. In gastric cancer, forced expression of miR-26b led to inhibition of GC cell migration and invasion in vitro and lung metastasis formation in vivo [36]. Conversely, miR-151-5p could promote tumor growth and lung metastasis of SC-M1 cells in vivo through downregulation of p53 protein expression in NSCLC [37]. All these results support our findings that miR-3607-3p functions as a tumor suppressor gene. miR-3607-3p may be the basis of new approaches to cancer therapy via the mechanisms of its regulation of tumors.
Although the mechanism by which miRNAs alter gene expression remains controversial, most studies have suggested that miRNAs are primarily processed by the RNA-mediated interference machinery to trigger partial or complete target gene mRNA degradation [38]. Our bioinformatics analysis revealed that miR-3607-3p could bind to the 3′UTR of TGFBR1 and CCNE2, repressing their expression. TGFBR1 is an important element of the TGF-β/SMAD signaling pathway, which has emerged as a central mediator of cancer progression because of its ability to regulate cell growth, differentiation, and migration [39]. The TGF-β pathway controls a plethora of cellular responses and plays a crucial role in tumorigenesis as either tumor promoter or suppressor [40]. Regarding tumor promotion, the TGF-β pathway could stimulate cell invasion by inducing epithelial-mesenchymal transition (EMT) and ultimately promote metastasis in multiple tumors [41–43]. TGF-β binding to TGFBR2 leads to phosphorylation and activation of TGFBR1 by TGFBR2. With the help of SARA, Smad2/3 is phosphorylated by TGFBR1. Then, they form a heterotrimeric complex with Smad4 and translocate into the nucleus to regulate gene transcription [44]. Yang et al. demonstrated that miR-140-5p possesses the potency to suppress hepatocellular carcinoma growth and metastasis by regulating TGFBR1 and FGF9 [45]. Studies have also shown that repression of TGFBR1 inhibits the cell proliferation of lung cancer and cell migration and invasion of breast cancer [46,47]. Previous studies highlighted the role of TGFBR1 in NSCLC [48,49].
CCNE2 is a well-known cyclin involved in the progression of the cell cycle, specifically the G1/S transition. CCNE2 is also involved in the development of NSCLC [2,50,51]. In this study, we conducted FACS analysis to further confirm how miR-3607-3p acts as a negative regulator of the cell cycle and found that increasing the expression of this miRNA resulted in significant G0/G1 arrest and S phage reduction. Cyclin E proteins play critical roles in G1 phase and in the G1-S phase transition with cyclin-dependent kinase 2 (CDK2) [2]. Overexpression of CCNE1 and CCNE2 has been reported in many types of human cancers. Recent data clearly showed that restoration of miR-144-5p function in bladder cancer cells inhibited the expression of both CCNE1 and CCNE2 and significantly induced G1 arrest in bladder cancer cells [52]. These results suggest that miR-3607-3p might inhibit NSCLC growth and metastasis partly by targeting TGFBR1 and CCNE2. The Ki-67 staining intensities were decreased in the tumors from the miR-3607-3p-overexpressing group. Our results further showed stronger TGFBR1 and CCNE2 staining in xenograft tumors of the miR-3607-3p agomir group than in tumors of the NC group, indicating that the proliferation of tumor cells was increased by the down-expression of miR-3607-3p. The present study is not without limitations. TGFBR1 and CCNE2 were evaluated in the tumor tissue and normal adjacent tissue samples in the mouse model. Since the tumors cells are of human origin and the ‘normal adjacent tissue’ is of mouse origin, differing reactivity of the antibodies with human and mouse proteins could be concerning. Finally, miR-3607-5p should be tested along miR-3607-3p within the same sets of experiments.
In conclusion, our results strongly suggest that miR-3607-3p is downregulated in NSCLC tissues and miR-3607-3p might act as an independent predictor for overall survival in NSCLC. Our results also suggest that serum miR-3607-3p could be a novel and stable biomarker for NSCLC. Furthermore, we found that miR-3607-3p possesses the potency to suppress NSCLC growth and metastasis, and to induce cell cycle arrest by regulating TGFBR1 and CCNE2. Our findings suggest that miR-3607-3p functions as a tumor suppressor in NSCLC and holds promise as a prognostic biomarker and potential therapeutic target for NSCLC.
An organized chip array including 162 non-metastatic NSCLC tissue samples and non-neoplastic lung tissue samples was purchased from Outdo Biotech (HlugA180Su02 and HLug-Squ150Sur-02, Shanghai, China; http://www.superchip.com.cn/). A total of 107 paired frozen paraffin NSCLC tissue samples and matched adjacent non-cancerous tissue samples were obtained from the North China University of Science and Technology Affiliated People’s Hospital; the tissues were collected between 2009 and 2013. In addition, the serum samples from 80 patients with NSCLC and 40 healthy controls were obtained from the North China University of Science and Technology Affiliated People’s Hospital; serum samples were stored at -80°C. For studies using human data, the study was approved by ethics committee of the North China University of Science and Technology Affiliated People’s Hospital. (approval number: RH-2017-006) and written informed consent was obtained from all participants.
The human lung adenocarcinoma cell lines A549, NCI-H1299, NCI-H157, ANIP-973, NCI-H292, and PC-9 were obtained from the Cell Culture Center of Peking Union Medical College (Beijing, China). The human embryonic kidney (HEK) 293T cell line was obtained from the American Tissue Culture Collection (Manassas, VA). The NSCLC cell lines were cultured in RPMI-1640 medium. HEK 293 T cells were maintained in DMEM supplemented with 10% fetal bovine serum (Gibco BRL, Grand Island, NY) in a humidified atmosphere of 5% CO2 at 37°C. The human fetal lung fibroblast cell line (MRC-5) was cultured in Minimum Essential Medium (MEM) containing non-essential amino acids, Earle’s salts, and L-glutamine supplemented with 10% fetal bovine serum and 1% antibiotic-antimycotic solution (containing 100 U/mL penicillin, 100 μg/mL streptomycin, and 0.25 μg amphotericin), and was maintained in a humidified air atmosphere with 5% CO2 at 37°C.
In situ hybridization (ISH) was performed according to the manufacturer’s instructions (Roche Molecular Systems, Pleasanton, CA, USA). The miR-3607-3p and miR-3607-5p probes (5’-CATCAGAAAGCGTTTACAGT-3’ and 5’-GCAUGUGAUGAAGCAAAUCAGU-3’) were tagged with 3’ and 5’ digoxigenin (Redlandbio.biomart.cn, Guangzhou, China). U6 snRNA (5’-CACGAATTTGCGTGTCATCCTT-3’) and scrambled probes (5′-GTGTAACACGTCTATACGCCCA-3′) were used as positive and negative controls, respectively. The probe-target complex was detected using an antidigoxigenin-alkaline phosphates conjugate and nitro-blue tetrazolium and 5-bromo-4-chloro-3’-indolyphosphate as the chromogen. The signals were classified according to cytoplasmic miR-3607-3p intensity as: negative: negative or faint expression in most cells; low expression: low expression in most cells or moderate expression in <50% of the cells; and high expression: moderate to strong expression in most cells. At least five random fields were evaluated by assessors blinded to grouping and clinical features.
All endogenous mature miRNA mimics, inhibitors and agomirs were purchased from RiboBio (Guangzhou, China). Transfection was performed according to the manufacturer’s protocols. miRNA mimics (50 nmol/L), miRNA inhibitors (100 nmol/L), and miRNA negative controls (NC) (100 nmol/L) were transfected into the cells using Lipofectamine 2000 (Invitrogen, Carlsbad, USA) according to the manufacturer’s instructions. After 48 h of transfection, cells were used for further experiments.
pDonR223-TGFBR1 and pDonR223-CCNE2 plasmids carrying the human TGFBR1 and CCNE2 genes were purchased from Changsha Axybio Bio-Tech Co., Ltd (Changsha, China, S1–S5 Files). The complete coding sequences of human TGFBR1 and CCNE2 were amplified from the pDonR223-TGFBR1 and pDonR223-CCNE2 plasmids. TGFBR1 and CCNE2 products and pEGFP-N1 plasmid were digested with Xho I and Hind III, and the fragments were purified and ligated with T4 DNA ligase. The ligated products were transformed into TOP10 competent cells. The positive clones were named pEGFP-N1-TGFBR1 and pEGFP-N1-CCNE2.
To evaluate the expression of miR-3607-3p, TGFBR1, CCNE2, and other genes, total RNA was used for reverse transcription (RT) and quantitative polymerase chain reaction (qRT-PCR) was performed on a Step One Plus real-time system (AB Applied Biosystems, Carlsbad, CA). U6 and GAPDH were used as internal controls. All the primers used in this study are listed in S6 File.
Bioinformatics analysis was performed using the following software: miRWalk, miRanda, and TargetScan. The 3’-UTR of human TGFBR1 and CCNE2 was amplified from human genomic DNA and individually inserted into the pmiR-RB-REPORT (Ribobio, Guangzhou, China, S7-8 file) using the XhoI and NotI sites. Similarly, the fragment of TGFBR1 and CCNE2 3’-UTR mutant was inserted into the pmiR-RB-REPORT control vector at the same sites. For reporter assays, NSCLC cells were co-transfected with wild type reporter plasmid and miR-3607-3p mimics. Firefly and Renilla luciferase activities were measured in cell lysates using the Dual-Luciferase Reporter Assay system. Luciferase activity was measured 48 h post-transfection using the dual-glo luciferase reporter system according to the manufacturer’s instructions. Firefly luciferase units were normalized against Renilla luciferase units to control for transfection efficiency.
For cell proliferation assays, cells were seeded on a 96-well plate (5×103 per well) and cell proliferation was determined by MTS (3-(4,5-dimethylthiazol-2-yl)-5-(3–carboxymethoxyphenyl)-2-(4-sulfophenyl)-2H-tetrazolium) according to the manufacturer’s instructions. MTS solution was added (20 μl/well) to each well and incubated at 37°C for 2 h. The optical density of each sample was immediately measured using a microplate reader (BioRad, Hercules, CA, USA) at 570 nm.
Cells were transfected with miR-3607-3p mimic or miR mimic NC, miR-3607-3p inhibitor or miR inhibitor NC, as described above. After 24 h, the transfected cells were trypsinized, counted and re-plated at 1×103 cells/10-cm dish. After 10 days, colonies resulting from the surviving cells were fixed with 3.7% methanol, stained with 0.1% crystal violet and counted. Colonies containing at least 50 cells were scored. Each assay was performed in triplicate.
In vitro cell migration assays were performed as described previously using Trans-well chambers (8 μM pore size; Costar). Cells were allowed to grow to ~75–80% confluency and serum-starved for 24 h. After detachment with trypsin, cells were washed with PBS and resuspended in serum-free medium. Next, 100 μl of cell suspension (5×104 cells/mL) were added to the upper chamber. Complete medium was added to the bottom wells of the chambers. The cells that had not migrated after 24 h were removed from the upper face of the filters using cotton swabs, but the cells that had migrated were fixed with 5% glutaraldehyde solution to determine the number of migratory cells. The lower surfaces of the filters were stained with 0.25% Trypan Blue. Images of six random ×10 fields were captured from each membrane and the number of migratory cells was counted. The assay was performed in triplicate for each experimental condition and the mean values were used for analysis. Similar inserts coated with Matrigel were used to evaluate the cell invasive potential in the invasion assay.
Fluorescence-activated cell sorting (FACS) analysis was performed 48 h after transfection. The cells were harvested, washed with cold PBS, and fixed into 70% ethanol at -20°C for 24 h, stained with 50 μg/mL propidium iodide (PI) (4ABio, China), and analyzed using a FACS Calibur flow cytometer (BD Bioscience, MA). The results were analyzed using the ModFit software (BD Bioscience, USA). Assays were conducted three independent times.
For western blot analysis, RIPA buffer containing protease inhibitors and phosphatase inhibitors (Roche) was used to prepare whole-cell lysates. Equal amounts of proteins were separated by SDS-polyacrylamide gel electrophoresis (SDS-PAGE) and transferred to PVDF membranes (Millipore). After blocking with 5% bovine serum albumin (BSA), the membranes were probed with anti-TGFBR1, CCNE2, and anti-GAPDH (ab31013, ab32103, and ab8425; Abcam, Cambridge, UK), followed by incubation with a horseradish peroxidase-conjugated secondary antibody goat-anti-mouse IgG (1:2000) and goat-anti-rabbit IgG (1:3000). Proteins were visualized using Image Reader LAS-4000 (Fujifilm) and analyzed using the Multi Gauge V3.2 software.
Recombinant lentiviral vectors containing miR-3607-3p knockdown and irrelevant sequences were purchased from XIEBHC Biotechnology (Beijing, China, S9 File). In addition to the lentivirus expression vectors, there was a luciferase and puromycin reporter gene driven by the EF1α promoter to indicate the infection efficiency in a timely manner. To construct the lentiviral vectors, the precursor sequence for miR-3607-3p and the irrelevant sequence (negative control) were inserted into pHBLV-U6-MCS-EF1α-Luc-T2A-puromycin lentiviral vectors. The recombinant lentiviruses were packaged by co-transfection of HEK-293T cells with pSPAX2 and pMD2.G with the LipoFiter reagent. The supernatants with lentivirus particles were harvested at 48 and 72 h after transfection and filtered through 0.45-μm cellulose acetate filters (Millipore, USA). Recombinant lentiviruses were concentrated by ultracentrifugation. To establish stable cell lines, NSCLC cells were transduced with lentivirus with a MOI of approximately 5 in the presence of 5 μg/ml polybrene. The supernatant was removed after 24 h and replaced with fresh complete culture medium. Infection efficiency was confirmed by RT-PCR 96 h after infection. The cells were selected with 2 μg/ml puromycin for 2 weeks.
For studies using animal data, all experiments were performed according to the Regulations for the Administration of Affairs Concerning Experimental Animals of 1988, issued by the State Scientific and Technological Commission for China. And these experiments were approved by the Institutional Animal Care and Use Committee of the North China University of Science and Technology Affiliated People’s Hospital (approval number: RM-2017-035). The effects of miR-3607-3p on the tumorigenic and metastatic potential of NSCLC cells were analyzed in subcutaneous and systemic metastasis in vivo models via right hip subcutaneous tissue injection and tail vein injection, respectively. For the subcutaneous model, 4-6-week-old BALB/c nude mice were injected subcutaneously in the right hip with 1×106 transfected cells. For the experimental metastasis in vivo model, transfected cancer cells (1×106 in 100 μl of HBSS) were directly injected into the tail vein. Six weeks later, tumor colonies in subcutaneous tissue were observed by HE staining and histology examination. Bioluminescence images were collected to assess the growth and metastasis of implanted tumor cells. To quantify the in vivo bioluminescence signals, mice were anesthetized with isoflurane before in vivo imaging. A D-luciferin solution (in vivo imaging solutions, PerkinElmer, 150 mg/kg in PBS) was injected intravenously for both orthotopic and systemic xenografts. Bioluminescence images were acquired with the IVIS Spectrum imaging system (Perkin-Elmer Life Sciences, Waltham, MA, USA) 2–5 min after injection. The captured images were quantified using the Living Image Software package (Perkin Elmer/ Caliper Life Sciences) by measuring the photon flux (photons/s/cm2/steradian) within a region of interest (ROI) drawn around the bioluminescence signal.
The agomir and miRNA negative control were synthesized by Ribobio (Guangzhou, China) and used according to the manufacturer’s instructions. A 10-nmol miR-3607-3p agomir as well as the miRNA negative control in 0.1 ml of saline buffer were locally injected into NSCLC cell-forming tumor mass once every 5 days for 6 weeks. After treatment, the NSCLC cell-forming tumors were tested by immunohistochemistry. The tumor size was monitored by measuring the length (L) and width (W) with calipers every 5 days. The volumes were calculated using the formula (L × W2)/2. Mice were sacrificed by cervical dislocation on day 42. The tumors were excised and snap-frozen for protein and RNA extraction.
The sections were de-paraffinized and boiled in 10 mM citrate buffer (pH 6.0) for antigen retrieval. Endogenous peroxidase was blocked by 3% H2O2. Slides were blocked in serum, incubated with the indicated antibodies at 4°C overnight, incubated with anti-rabbit secondary antibody, and visualized with diaminobenzadine (Sigma). A negative control experiment was also performed. IHC staining images were captured at 200× under a microscope (Olympus).
All continuous data were expressed as means ± standard deviation. Error bars represented the standard errors of the means. Student’s t-test, χ2 test and repeated measures ANOVA were used to determine statistical significance, as appropriate. The log-rank test was used to analyze the effect of clinical variables and miRNAs on the overall survival (OS) of patients. Multivariate Cox regression models were used to assess factors associated with overall survival in NSCLC. Cox regression analysis was used to evaluate the independent prognostic value of the miR-3607-3p signature, with age, gender, T stage, histological type, N stage, clinical stage, and the miRNA signature used as covariates. Receiver operating characteristic (ROC) curves and the area under the ROC curve (AUC) were used to assess the feasibility of using serum miRNA as a diagnostic tool for detecting NSCLC. P < 0.05 was considered statistically significant. Statistical analyses were performed using SPSS 16.0 (IBM, Armonk, NY, USA).
This study was reviewed and approved by the Ethics Committee of North China University of Science and Technology Affiliated People’s Hospital.
|
10.1371/journal.pgen.1008335 | Telomere-binding proteins Taz1 and Rap1 regulate DSB repair and suppress gross chromosomal rearrangements in fission yeast | Genomic rearrangements (gross chromosomal rearrangements, GCRs) threatens genome integrity and cause cell death or tumor formation. At the terminus of linear chromosomes, a telomere-binding protein complex, called shelterin, ensures chromosome stability by preventing chromosome end-to-end fusions and regulating telomere length homeostasis. As such, shelterin-mediated telomere functions play a pivotal role in suppressing GCR formation. However, it remains unclear whether the shelterin proteins play any direct role in inhibiting GCR at non-telomeric regions. Here, we have established a GCR assay for the first time in fission yeast and measured GCR rates in various mutants. We found that fission yeast cells lacking shelterin components Taz1 or Rap1 (mammalian TRF1/2 or RAP1 homologues, respectively) showed higher GCR rates compared to wild-type, accumulating large chromosome deletions. Genetic dissection of Rap1 revealed that Rap1 contributes to inhibiting GCRs via two independent pathways. The N-terminal BRCT-domain promotes faithful DSB repair, as determined by I-SceI-mediated DSB-induction experiments; moreover, association with Poz1 mediated by the central Poz1-binding domain regulates telomerase accessibility to DSBs, leading to suppression of de novo telomere additions. Our data highlight unappreciated functions of the shelterin components Taz1 and Rap1 in maintaining genome stability, specifically by preventing non-telomeric GCRs.
| Tips of chromosomes, telomeres, are bound and protected by a telomere-binding protein complex called shelterin. Most previous studies focused on shelterin’s telomere-specific role, and its general role in genome maintenance has not been explored extensively. In this study, we first set up an assay measuring the spontaneous formation rate per cell division of gross chromosomal rearrangements (GCRs) in fission yeast. We found that the rate of GCRs is elevated in mutants defective for shelterin components Taz1 or Rap1. Detailed genetic experiments revealed unexpectedly that Taz1 and Rap1 have a novel role in repairing DNA double-strand breaks (DSBs) and suppressing GCRs at non-telomeric regions. Given that shelterin components are conserved between fission yeast and humans, future studies are warranted to test whether shelterin dysfunction leads to genome-wide GCRs, which are frequently observed in cancers.
| The integrity of chromosomal DNA can be compromised by mutations that vary in size, ranging from small perturbations, such as point mutations and short insertions/deletions, to large changes, such as deletions, duplications, inversions, and translocations of long chromosome segments. The latter are collectively called genomic rearrangements or gross chromosomal rearrangements (GCRs), which have profound implications in cancers as well as genetic diseases. Recent advances in DNA sequencing technology have enabled us to trace the history of GCRs in cancer cells, and it is now well-known that cancer development is accompanied by the frequent occurrence of GCRs [1]. Thus, elucidation of the molecular mechanism underlying GCR control is of critical importance in understanding the progression of cancer malignancy.
Previous studies have pointed to the requirement of chromosome maintenance mechanisms for suppression of GCRs, including DNA repair and telomere protection pathways [2, 3]. The telomere is a huge DNA-protein complex that is located at the termini of linear chromosomes. In humans, telomeric DNA comprises hexanucleotide TTAGGG repeats and consists of a double-stranded (ds) region and a single-stranded (ss) overhang. The telomeric dsDNA recruits TRF1-TRF2-Rap1, whereas the ss telomeric DNA recruits POT1-TPP1, and these two subcomplexes are bridged by TIN2 to form a complex known as shelterin (reviewed in [4]). This shelterin complex helps cells distinguish telomeres from DNA double-strand breaks (DSBs) that must be repaired. For instance, TRF2 depletion brings about the frequent occurrence of chromosome end-to-end fusions, which is due to deregulation of the non-homologous end joining (NHEJ) repair pathway at telomeres. Resultant dicentric chromosomes are unstable, leading to another round of chromosomal rearrangements (reviewed in [5]). It is thus evident that telomere protection by the shelterin complex is vital for repressing GCRs.
While the shelterin complex primarily serves to protect telomeric DNA, the telomere-associated DNA polymerase named telomerase is implicated in GCRs [6, 7]. On the one hand, telomerase is able to elongate the telomere repeat sequence using its RNA subunit as a template, thereby counteracting gradual telomere shortening at each round of DNA replication. At the same time, however, telomerase poses a potential threat to genome stability. In budding yeast, telomerase promotes GCRs through de novo addition of telomere repeats to DSB sites, resulting in terminal deletion of chromosomal DNA [7]. It has been reported that de novo telomere addition is suppressed through two mechanisms: activation of Pif1 helicase, which was proposed to remove telomerase from DSBs; and inhibition of Cdc13 accumulation by DNA damage signaling [8–10]. However, a previous study showed that fission yeast Pif1 is not a negative regulator of telomerase [11]. In human cells, recruitment of telomerase to telomeres and the activity of telomerase are regulated by the shelterin complex (reviewed in [12]). However, it is still unclear whether shelterin is also involved in the regulation of de novo telomere addition at non-telomeric sites.
Fission yeast, Schizosaccharomyces pombe, serves as a useful model to dissect the functions of shelterin, because this unicellular organism shares most of the shelterin components with humans. Fission yeast shelterin is composed of six proteins: Taz1, Rap1, Poz1, Tpz1, Pot1, and Ccq1. Among these, Taz1, Tpz1 and Pot1 are orthologs of human TRF1/2, TPP1, and POT1, respectively. Fission yeast Rap1 and human Rap1 are also homologous to each other, sharing several domains including a single BRCT domain at their N termini. Taz1 and Rap1 form a subcomplex that binds to telomeric ds DNA, while Poz1, Tpz1, Ccq1, and Pot1 form another subcomplex at the telomeric ss DNA. Similar to human shelterin, these two subcomplexes at telomeric ds and ss DNA are bridged by the physical interaction between Rap1 and Poz1 [13, 14].
To date, the shelterin components in fission yeast have been extensively investigated. Taz1, Rap1, and Poz1 negatively regulate telomerase activity and promote telomere heterochromatin formation [13, 15–17]. On the other hand, Tpz1 and Pot1 are essential for telomere protection, and thus telomere DNA is aggressively degraded after deletion of the tpz1+ or pot1+ gene [13, 18]. Ccq1 recruits telomerase to telomeres through direct binding to telomerase [19, 20]. Taz1 and Rap1 prevent telomere end fusions that would otherwise be caused by aberrant activation of the NHEJ repair pathway at telomeres [21, 22]. Taz1 and Rap1 also tether telomeres to the nuclear periphery via inner nuclear membrane (INM) protein Bqt4 in vegetative cell growth [23]. As such, the shelterin components perform distinct functions, even though they form a complex.
It is known that disruption of shelterin can trigger frequent GCRs through breakage of dicentric chromosomes formed by chromosome end-to-end fusion [5]. However, it is unclear whether the shelterin complex has an additional GCR-suppressive function apart from preventing such chromosome end-to-end fusions; this uncertainty can be ascribed to technical limitations in precisely measuring the occurrence rate of GCRs in mammalian cells. In budding yeast, an assay has been developed to measure the GCR rates, aptly termed the “GCR assay” [24, 25]. GCR rates are deduced from loss of two tandem counter-selective markers inserted in a non-essential chromosomal region. In this study, we adopted the GCR assay to fission yeast and examined whether the individual shelterin components as well as other telomere-binding proteins suppress GCRs in non-telomeric regions. We found that a fraction of the shelterin components, including Taz1 and Rap1, are required for GCR suppression. Deletion of DNA ligase IV, which is essential for NHEJ, did not rescue the increased GCR rates in taz1Δ and rap1Δ mutant cells, suggesting that Taz1 and Rap1 do not prevent GCR via suppressing NHEJ, unlike the Taz1- and Rap1-dependent protection of telomeres from fusion. Instead, derepression of telomerase is responsible for the increased GCR rates in taz1Δ and rap1Δ strains. Dissection of the Rap1 protein identified the N-terminal BRCT domain as an important domain for the GCR suppression. Moreover, when DSBs are site-specifically induced at a non-telomeric locus by I-SceI endonuclease, Taz1 and Rap1 are required for cellular survival and for inhibiting erroneous repair. We propose that Taz1 and Rap1 prevent GCRs by regulating telomerase activity and DSB repair, even in non-telomeric regions.
To measure GCR rates in fission yeast, we applied the assay system that was previously developed for budding yeast (Fig 1A) [24]. We constructed a DNA cassette containing two neighboring marker genes, ura4+ and TK (the latter encodes herpes virus thymidine kinase) in tandem. Cells expressing ura4+ and TK are sensitive to 5-fluoroorotic acid (5-FOA) and 5-fluoro-2’-deoxyuridine (FUdR), respectively. As expected, fission yeast cells with this marker cassette integrated at approximately 150 kb from the right telomere of chromosome I (the precise location is described in Materials and Methods) showed sensitivity to both of the drugs (5-FOA/FUdR) (S1A Fig). This strain is expected to become resistant to both drugs when the ura4+ and TK genes undergo simultaneous deletions and/or loss-of-function point mutations. However, such simultaneous point mutations seem highly unlikely because the probability of simultaneous point mutations occurring in two specific genes is thought to be quite low (~10−14/cell division, given that the spontaneous incidence of loss-of-function mutations for each gene, independently, is ~10−7, see Methods) [26]. Thus, as in the budding yeast GCR assay system, the vast majority of drug-resistant survivors in our assay should be derived from GCRs that result in simultaneous loss of the two marker genes. Because an essential gene closest to the marker cassette is sec16+, which is located about 16.8 kb centromeric from the cassette, and there is no essential gene telomeric to the cassette, our system can detect GCRs that take place within this ~16.8-kb region (Fig 1A). Hereafter, we will refer to this GCR target region as the “breakpoint region”. Because it lacks any sequence that shares apparent homology with other chromosome regions, our GCR assay is expected to detect GCRs that are mediated by no or little homology.
In the GCR assay, we counted the number of colonies on a plate with or without 5-FOA/FUdR and estimated GCR rates per cell division using fluctuation analysis [24]. In the case of wild-type cells, a GCR rate determined in our system was 2.6 × 10−9 per cell division (Fig 1B). This rate is actually far greater than the expected probability of dual independent point mutant survivors (10−14 per cell division), confirming that our system primarily detects GCRs. We then isolated 5-FOA/FUdR-resistant clones and performed DNA sequencing at the breakpoint region (See Materials and Methods). Based on the sequencing data, GCRs were classified into deletion and translocation types (Fig 1C, wild-type). In the deletion type, DSBs led to deletion of the chromosomal terminus containing the drug selection cassette. The sequencing analysis detected ectopic telomeric DNA repeats at the breakpoints, suggesting that de novo telomere addition healed the DSB (Fig 1D). Twelve out of fifteen wild-type-derived GCR clones examined here belonged to this type. In two other clones, breakpoints were fused with unique sequences from the left arm of chromosome I (opposite the right arm where the original marker cassette had been located prior to the rearrangement) in a head-to-tail orientation (same direction towards telomeres). Such fusions could be derived from either break-induced DNA replication or DNA recombination (Fig 1D, Translocation as diagramed in S1C Fig). Indeed, the breakpoint junctions consisted of 7 or 8 bp of microhomology in these survivors (Fig 1D). In both types of GCRs, the locations of the breakpoints seemed to be uniformly distributed in the breakpoint region, rather than clustering at a particular hotspot (Fig 1E). In the last of the 15 survivors, we established the loss of the marker cassette but could not determine the precise change in the breakpoint region sequence. As we expected, we did not obtain any clones with simultaneous point mutations in both ura4+ and TK, validating the usefulness of our assay system to specifically evaluate GCR rates.
Of note, the GCR rate measured in our fission yeast system is comparable to that in the original budding yeast system: 2.6 × 10−9 /cell division in a 16.8 kb-long breakpoint region in wild-type fission yeast (this study) vs. 2.27 × 10−9 in a 19.2 kb breakpoint region in wild-type budding yeast [27]. If we assume that GCRs occur randomly throughout the genome, then GCR rates would be expected to be proportional to the DNA length of the breakpoint region for the GCR assay. The GCR rates normalized to unit length are 1.5 × 10−10/cell division/kb in fission yeast and 1.2 × 10−10/cell division/kb in budding yeast. Thus, wild-type cells of the two yeast species showed comparable normalized GCR rates. In addition, GCRs in wild-type fission yeast cells are mostly associated with terminal deletions, whereas translocations are relatively rare, just as in budding yeast [24]. We thus asked whether the GCR suppression mechanism identified by the budding yeast GCR assay system also functions in fission yeast. Budding yeast strains lacking nuclease FEN-1 or Mre11 show a 914- and 628-fold increases in GCR rates, respectively [20]. We therefore tested the effects of loss of FEN-1 and Mre11 in rad2Δ and mre11Δ S. pombe cells and observed a ~100-fold increase in GCR rates (Fig 1B). In budding yeast, Pif1 helicase suppresses telomerase-mediated telomere elongation at native telomeres and DSBs through destabilizing annealing of telomerase RNA template and single-stranded telomere DNA substrates [28]. We examined the impact of inactivating pfh1, the fission yeast homologue of PIF1 on GCRs. pfh1-mt* is a mutant that lacks nuclear functions but retains the essential mitochondrial functions [29]. pfh1-mt* showed 32-fold higher GCR rates than wild type, similar to the results reported for the budding yeast corresponding mutant, pif1-m2 (S1D Fig) [7]. From these similarities observed in the two distinct yeast species, we surmise that the regulatory mechanism suppressing GCRs is evolutionarily conserved, underscoring the significance of studying the GCR mechanism in fission yeast.
Since most of the GCR survivors that we isolated were derived from de novo telomere addition, we investigated a possible involvement of the telomere proteins in GCR regulation. In fission yeast, Taz1 directly binds to both Rap1 and ds telomeric DNA, thereby recruiting Rap1 to telomeres (Fig 2A). We found that taz1Δ and rap1Δ cells showed greatly increased GCR rates, 1.1 × 10−7 and 0.85 × 10−7 /cell division (42-fold and 33-fold higher than wild-type cells), respectively (Fig 2B). A Rap1-I655R mutant, in which the recruitment of Rap1 to telomeres is diminished due to a compromised Taz1-Rap1 interaction [30], showed an increased GCR rate that was comparable to taz1Δ or rap1Δ (Fig 2B, 2.6 × 10−7 /cell division, a 97-fold increase over wild-type), suggesting that Taz1 represses GCRs primarily through the physical interaction with Rap1. Consistent with this notion, taz1+ and rap1+ were found to be epistatic: a taz1Δ rap1Δ double mutant (1.2 × 10−7 /cell division) showed similarly increased GCR rates compared to each single mutant taz1Δ or rap1Δ (Fig 2B). We determined the sequences of GCR breakpoints in taz1Δ and rap1Δ survivors in GCR assay. We found only deletion type GCRs in taz1Δ and rap1Δ survivors, although the fractions of deletion in these mutants are not significantly higher than in wild type (Fig 2C & 2D). These results imply that the two shelterin components Taz1 and Rap1 function in the same pathway to prevent GCRs.
We also measured GCR rates of wild-type, taz1Δ, and rap1Δ strains at 20°C. It is known that taz1Δ, but not wild-type or rap1Δ delay the cell cycle progression and lose viability due to chromosomal entanglement at this temperature [31]. We found that taz1Δ, but not wild-type and rap1Δ, showed more than one order of magnitude higher GCR rates at 20°C than at 32°C (S2 Fig). Given the correlation between the increased GCR rate and cold-sensitivity among the three strains, it is possible that the chromosome entanglement contributes to the high GCR rate with taz1Δ at 20°C. Future study is necessary for concluding the molecular link between these phenotypes.
In sharp contrast to taz1+ and rap1+ deletion, deletion of the poz1+ gene, which encodes another Rap1-interacting shelterin component, did not affect the GCR rate (3.9 × 10−9 /cell division, Fig 2B). Interestingly, taz1Δ poz1Δ and rap1Δ poz1Δ double mutants showed lower GCR rates than the taz1Δ and rap1Δ single mutants, demonstrating that Poz1 is required for the derepression of GCRs in taz1Δ and rap1Δ cells (Fig 2E). Consistently, the abrogation of Poz1-Tpz1 binding by a I501A/R505E mutation in Tpz1 [32], another shelterin component that directly binds to Poz1, similarly suppressed the increased GCR rates in taz1Δ and rap1Δ mutant backgrounds (Fig 2E). This result suggests that Poz1 recruitment promotes GCRs in taz1Δ and rap1Δ, given that Poz1-Tpz1 binding is essential for telomere localization and function of Poz1 [14, 33]. We noticed that poz1+ deletion and Tpz1-I501A/R505E mutation individually caused strong reduction in GCR rates in rap1Δ but not much in taz1Δ cells. These results suggest that the increased GCR rates in taz1Δ and rap1Δ are caused by different mechanisms. It was reported that the formation of Rap1-Poz1-Tpz1 trimer is a hierarchical process in vitro [34]. First, Poz1 and Tpz1 form a dimer. The Rap1-binding domain of Poz1 undergoes allosteric changes upon the Poz1-Tpz1 dimer formation, which greatly increases the affinity with Rap1, and induces the Rap1-Poz1-Tpz1 trimer formation. Therefore, it is possible that the Tpz-Poz1 dimer stably exists in rap1Δ but not in taz1Δ. Such unusual shelterin subcomplexes may contribute to the differential effects of poz1+ deletion in taz1Δ and rap1Δ, as revealed in Fig 2E. These results suggest that genetic interaction of Taz1, Rap1, and Poz1 regarding GCR suppression is complex, similarly to that as for telomere length regulation and cold sensitivity [20, 22]. We also examined Stn1, a non-shelterin protein that binds to telomeric ssDNA. Because Stn1 is essential for telomere protection, we investigated a temperature sensitive mutant stn1-1, which has slightly elongated telomeres at semi-permissive temperature 25°C [35]. We found a moderately increased GCR rate at that temperature (Fig 2B), 1.6 × 10−8 /cell division, a 6-fold increase over wild-type.
Because both Taz1 and Rap1 are multi-functional (see below), we set out to dissect which specific function(s) is related to the GCR inhibition. It is known that, in taz1Δ and rap1Δ cells, but not in poz1Δ cells, telomeres are prone to fuse to each other by NHEJ when cells are arrested at G1 phase [21, 36]. It is thus possible that a failure to suppress NHEJ in taz1Δ and rap1Δ could lead to formation of dicentric chromosomes, which would trigger DSBs which could result in the observed chromosome terminal deletions. In order to examine whether NHEJ is responsible for frequent GCRs in taz1Δ and rap1Δ mutant cells, we deleted the DNA ligase IV-encoding lig4+gene, which is essential for NHEJ in fission yeast. It was reported that a lack of lig4+ suppresses the frequent telomere fusions in taz1Δ and rap1Δ [21, 22]. We found that disruption of lig4+ in taz1Δ and rap1Δ did not significantly suppress the increased GCR rates observed with taz1Δ and rap1Δ (Fig 3A), suggesting that NHEJ is dispensable for the high incidence of GCRs in taz1Δ and rap1Δ cells.
Because fission yeast in exponentially growing phase shows very short G1 phase, and NHEJ is active only in G1 but not in S and G2 phase, it was possible that NHEJ was dispensable for GCRs due to a small fraction of cells staying in G1 phase. We therefore arrested cells in G1 phase through nitrogen starvation, and measured the GCR frequency. Briefly, cells exponentially growing in YES media were divided into two groups, which were incubated in EMM media with (N+) or without (N-) ammonium sulfate for 24 hr, respectively, and then transferred to YES media for growth overnight. Then equal numbers of N+ and N- cells were subjected to the GCR assay. We found that taz1Δ(N-) cells showed an approximately two-fold increase in GCR frequencies compared to taz1Δ(N+) cells (S3A Fig). Because taz1Δ(N-) cells were G1-arrested and/or lost viability [21] while taz1Δ(N+) cells actively proliferated in EMM media (with or without supplementing nitrogen), the total number of cell divisions was greater in taz1Δ(N+) cells than in taz1Δ(N-) cells. Therefore, the two-fold increase of GCR frequencies is most likely an underestimate of a larger GCR rate (which is normalized per cell division) in taz1Δ(N-) compared to that in taz1Δ(N+). Interestingly, GCR frequencies in taz1Δ(N-) were partially suppressed by ligase IV deletion, while those in taz1Δ(N+) were not. Taken together, NHEJ also contributes to the increase of GCR frequencies of taz1Δ cells in G1 phase. Because G1 cells are rare in cells with unperturbed cell cycles, this effect is negligible in exponentially growing cell populations. These results suggest that the break-fusion-bridge cycle via formation of telomere-fusion-mediated dicentric chromosomes plays a minor role, if any, in the increased GCR rate in cycling taz1Δ cells.
Both Taz1 and Rap1 are essential for heterochromatin formation at telomeres and their adjacent regions, subtelomeres [37]. To examine whether telomere heterochromatin structure is important for suppression of GCRs, we deleted the clr4+ and swi6+ genes, both of which encode essential factors for heterochromatin formation (S3B Fig). Deletion of swi6+ in the wild-type background led to a small increase in the GCR rate, suggesting a potential contribution of heterochromatin to the suppression of GCRs. We examined poz1-W209A mutation. The shelterin component Poz1 is required for telomere silencing, and the poz1-W209A mutation is known to specifically disrupt the telomere heterochromatin regulatory function, among others [38]. No increase in the GCR rate was observed in poz1-W209A strain (S3C Fig). In contrast, deletion of clr4+ in rap1Δ background showed small but significant decrease GCR rates, although the underlying mechanism is unclear. From these results, we conclude that heterochromatin does not play a significant role in suppressing GCR except in rap1Δ.
In fission yeast, Taz1 and Rap1, but not Poz1, tether telomeres to the INM via binding of Rap1 to INM protein Bqt4 in vegetative cell growth [23, 39]. It is thus possible that the telomere tethering to the INM contributes to GCR suppression through regulation of chromosome positioning within the nucleus. We found that bqt4Δ cells showed moderately increased GCR rates (4.3 × 10−8 /cell division, Fig 3B). The GCR rate was also significantly increased by deletion of bqt3+ (3.0 × 10−8 /cell division), whose protein product Bqt3 stabilizes Bqt4 [23]. It was reported that the Ku70/80 complex and two INM proteins Lem2 and Man1 also promote tethering of telomeres to the nuclear envelope, although Man1 plays a minor role [40, 41]. Deletion of pku70+ or lem2+, but not man1+, led to moderately higher GCR rates (2.9 × 10−8 and 4.2 × 10−8 /cell division, respectively) than the wild-type strain (Fig 3B). These results imply that tethering of telomeres to the nuclear envelope facilitates GCR suppression.
Bqt4 localizes to the INM through its C-terminus transmembrane domain, and its N-terminal half is necessary and sufficient for binding Rap1 [23]. While telomeres are dissociated from the nuclear envelope in bqt4Δ, expression of an artificial fusion protein between Rap1 and an N terminus-truncated Bqt4 (Rap1-GFP-Bqt4ΔN) in bqt4Δ resumed telomere clustering at the nuclear envelope [23]. With our GCR assay, we found that bqt4Δ cells expressing the Rap1-GFP-Bqt4ΔN fusion protein from bqt4 promoter showed only a slightly lower GCR rate than bqt4Δ cells expressing GFP-Bqt4ΔN, in which telomeres are not tethered to the INM (Fig 3B, 2-fold difference). Moreover, the rap1-5E mutant (consisting of S213E, T378E, S422E, S456E, S513E mutations), in which the interaction between Rap1 and Bqt4 is impaired [39], displayed a comparable GCR rate to wild-type cells (Fig 3B). We also found that simultaneous deletion of bqt4+ significantly increased GCR rates in taz1Δ and rap1Δ cells (Fig 3C). These results suggest that Rap1-Bqt4 binding plays a minor role in suppressing GCRs, and that Bqt4 regulates GCRs at least in part by a Taz1- and Rap1-independent mechanism. By the same token, this result suggests that Rap1 utilizes Bqt4-independent mechanisms for suppressing GCRs.
Taz1 and Rap1 suppress telomerase-mediated telomere DNA elongation [42, 43]. Given that all GCRs examined in taz1Δ and rap1Δ were terminal deletions involving de novo telomere additions at breakpoints, it was likely that deregulated telomerase reactions facilitated GCRs through enhanced de novo telomere addition in taz1Δ and rap1Δ cells. Inactivation of trt1+, the gene encoding the catalytic subunit of telomerase, leads to chromosome self-circularization [42], making the GCR assay results difficult to compare with other cases. We therefore explored the effect of a Pof8 disruption on GCR rates. Pof8 is involved in maturation of telomerase RNA, and deletion of pof8+ leads to telomere shortening without extensive chromosome circularization, in contrast to a trt1+ deletion [44–47]. We found that taz1Δ pof8Δ and rap1Δ pof8Δ cells, which also do not show chromosome circularization, showed GCR rates which were lower than taz1Δ and rap1Δ cells, and similar to wild type cells (Fig 4A). These results suggest that telomerase activity is essential for the high GCR rates in taz1Δ and rap1Δ. In contrast, the GCR rate of a rad2Δ pof8Δ strain was in between that of rad2Δ alone and wild type cells, suggesting that Taz1 and Rap1 specifically suppress telomerase-dependent GCRs.
We considered two possibilities for how telomerase activity affects GCRs in taz1Δ and rap1Δ: (1) increased telomerase accessibility directly facilitates de novo telomere addition at breakpoints, or (2) abnormally elongated native telomeres indirectly affect non-telomeric GCRs. To determine which is the case, we examined GCR rates using cells with circular chromosomes in the presence or absence of Trt1 [4]. It is known that circular chromosomes in trt1Δ do not contain telomere DNA sequences [48]. For this purpose, we introduced a Trt1-expressing plasmid into trt1Δ cells with circularized chromosomes. In this setting, the majority of the trt1Δ cells harboring the Trt1 plasmid maintained circular chromosomes I and II (S4 Fig). As for trt1Δ taz1Δ, chromosomal configuration depends on the order of gene deletions during the strain preparation. When trt1+ is deleted first, followed by taz1+ deletion, the strain contains circular chromosomes. In contrast, linear chromosomes are maintained when taz1+ is deleted first, followed by trt1+ deletion [42]. Below, we will describe experiments using trt1Δ taz1Δ maintaining circular chromosomes, except otherwise noted. We also confirmed that trt1Δ taz1Δ expressing ectopic Trt1 retains circular chromosomes (S4 Fig).
When we subjected circular chromosome-containing cells to the GCR assay, it was expected that circular chromosomes needed to undergo complicated changes, such as two independent DSBs at the both sides of the selection cassette, and healing of the two DSBs by telomere addition to produce linear chromosomes. Consistently, all of the various strains maintaining circular chromosomes (except Trt1-overproducing trt1Δ taz1Δ) showed GCR rates below the detection sensitivity of the assay (Fig 4B).
When trt1Δ, trt1Δ taz1Δ and trt1Δ rap1Δ (all containing circular chromosomes) were transformed with Trt1-expressing plasmids, trt1Δ taz1Δ showed a significant increase in GCR frequency, while trt1Δ and trt1Δ rap1Δ did not (Fig 4B). These results suggest two points: first, Taz1 prevents GCR formation independent of its specific DNA binding to telomere DNAs, since circular chromosomes lack all telomere DNAs [48]; second, Taz1 has additional roles, which are not shared by Rap1, in preventing GCR formation from circular chromosomes. When we deleted poz1+ in trt1Δ taz1Δ cells, followed by over-expression of Trt1, GCR rates were decreased, suggesting that Poz1 promotes GCRs in taz1Δ cells in the absence of telomere DNA (Fig 4B, compare lanes 4 and 8). rap1Δ trt1Δ cells showed similar GCR rates to wild type even after Trt1 re-expression. We confirmed that both trt1Δ taz1Δ poz1Δ cells and rap1Δ trt1Δ cells maintained circularization of chromosomes I and II before and after Trt1 re-expression (S4 Fig).
In contrast to circular chromosomes-containing trt1Δ taz1Δ, linear chromosome-maintaining trt1Δ taz1Δ (see above), showed significantly increased GCR rates compared to linear-chromosome-containing wild-type cells (Fig 4B). Ectopic Trt1-over-expression further increased the GCR rates to the level of taz1Δ cells.
To further dissect the precise mechanism of GCR repression by Rap1, we exploited previously reported sequential N-terminal Rap1 truncations, Rap1-A to G [31] (Fig 5A). Among these, we found that only the Rap1-G mutant showed an increased GCR rate. Because the Rap1-A to F mutant strains all retain the Poz1-binding domain (Rap1 457–512 amino acids) but Rap1-G does not, the results raised the possibility that Rap1-Poz1 binding is required for the GCR suppression. However, deletion of the Poz1-binding domain alone did not increase GCR rates (Rap1ΔP, Fig 5A and 5B). Further dissection of Rap1 revealed that simultaneous deletion of the BRCT domain at the N terminus as well as the Poz1-binding domain led to an increase in GCR rates that was comparable to that in rap1Δ cells (Rap1-AΔP, Fig 5A and 5B, and S4 Fig). The phenotype of the rap1-AΔP strain was similar to rap1Δ regarding GCR suppression; GCRs from the rap1-AΔP strain primarily showed terminal deletion (Fig 5C), and deletion of pof8+ canceled the increased GCR rates. These results indicate that the BRCT domain and the Poz1-binding domain redundantly suppress GCRs. Since the Rap1-A mutant, which lacks the BRCT domain, maintains normal telomere length, we suggest that the BRCT domain does not regulate telomerase action at native telomeres, while the Poz1-binding domain suppresses GCRs through inhibition of telomerase at both telomeres and non-telomeric DSBs. The Rap1 BRCT domain may be involved in a general DNA repair pathway, failure of which causes various consequences including erroneous telomere addition by telomerase at non-telomeric regions.
Given that GCRs are thought to arise from aberrant DSB repair, the telomerase-independent GCR repression mechanism could potentially include DSB processing. In order to examine this possibility, we constructed a conditional, site-specific DSB induction system (Fig 6A). The DNA sequence-specific endonuclease I-SceI was expressed under the control of a tetracycline-inducible promoter, and a single I-SceI cut site (I-SceIcs) was integrated at approximately 150 kb centromeric from the right telomere of chromosome I, exactly at the same locus as the marker cassette that was inserted in our GCR assay strains. Addition of anhydrotetracycline (ahTET) to the culture media leads to a DSB at the I-SceIcs. Indeed, two hours after ahTET addition, quantitative PCR amplification of genomic DNA using primers flanking the I-SceIcs decreased to 40–50% of control levels in wild-type, taz1Δ, rap1Δ, and poz1Δ backgrounds, demonstrating that DSBs were induced in these strains with similar efficiencies (Fig 6B). With this system, we examined how efficiently the wild-type and mutant cells could repair the DSB. We transiently induced DSB formation at I-SceIcs by culturing cells in liquid media containing ahTET for two hours. After that, ahTET was washed out and the cells were spread onto ahTET-free plate media. Cells that were unsuccessful in repairing the I-SceI DSB (or healing it, e.g. by de novo telomere addition) would not form colonies. We examined genomic DNA extracted from 10 colonies each from wild type, taz1Δ, rap1Δ strains and confirmed that none of them contained mutation in I-SceIcs, indicating that GCRs were not involved in generating survivors. Subsequent to the transient DSB induction in wild-type cells, the frequency of colony formation decreased to 62% of control (uncut) levels (Fig 6C). Strikingly, taz1Δ and rap1Δ cells showed further lower viabilities (37% and 33%, respectively). This result suggests that Taz1 and Rap1 promote DSB repair. taz1Δ pof8Δ cells showed similar survival with taz1Δ, consistent with the idea that the survivors occur not through telomerase-mediated GCRs, but through DSB repair, and suggesting that Taz1 promotes DSB repair independently of telomerase regulation. Interestingly, the BRCT domain-lacking rap1-A mutant showed significantly lower survival (42%) than wild-type, indicating that the Rap1 BRCT domain plays a significant role in DSB repair. We note, however, that the survival rate of the rap1-A strain was still slightly but significantly higher than rap1Δ, suggesting that Rap1 is involved in two pathways that promote DSB repair: one BRCT-domain-dependent and the other independent (Fig 6D). We examined whether Taz1 and Rap1 physically bind to the DSBs by chromatin immunoprecipitation (ChIP). No significant ChIP signal was detected for both Taz1 and Rap1 at the sites 1.5 and 5 kb apart from I-SceIcs at 2 and 4 hrs after DSB induction, while they localized at telomeres (S6B Fig). It is possible that Taz1 and Rap1 were only transiently recruited to DSBs, which made the ChIP detection difficult. Alternatively, Taz1 and Rap1 are indirectly involved in the DSB repair (see Discussion).
To examine if the impaired DSB repair in taz1Δ and rap1Δ cells leads to GCRs, we measured GCR frequencies caused by the I-SceI-induced DSB. We allowed the I-SceI endonuclease to be continuously active by culturing cells on plate media containing ahTET. Under this condition, faithful DSB repair would be detrimental to cell viability because it would regenerate I-SceIcs, leading to incessant cut and repair cycles. In contrast, I-SceIcs would become resistant to I-SceI cleavage when the I-SceIcs was lost via mutagenic DSB repair, including GCRs, indels, and point mutations. In the wild-type background, only 0.58% of cells survived and formed colonies, suggesting that GCRs and erroneous DNA repair is rare (Fig 6F). The survival was likely caused by de novo telomere addition, because trt1Δ showed significantly lower survival rates than wild type (Fig 6E). ccq1Δ also decreased the survival rate to a similar level shown by trt1Δ, suggesting that Ccq1 is required for de novo telomere addition at DSBs, as in the case of telomerase-mediated telomere elongation at native telomeres. In contrast, a higher fraction of cells survived in the taz1Δ and rap1Δ backgrounds. In all wild-type, taz1Δ, and rap1Δ survivors, in which the breakpoints were identified (n = 9, 9, and 7, respectively), the I-SceIcs was eliminated by terminal deletion associated with de novo telomere addition, suggesting that the observed viabilities reflect frequencies of terminal deletions in wild-type, taz1Δ and rap1Δ strains. Therefore, Taz1 and Rap1 not only promotes faithful DSB repair (Fig 6C), but also prevents erroneous DSB repair by suppressing terminal deletion coupled with de novo telomere addition. In wild-type cells, breakpoints of the deletions were close to (1~10 bp) I-SceIcs in all 9 clones in which the breakpoints were identified (Fig 6G). In contrast, breakpoints of 4 out of 9 clones in the taz1Δ survivors and 5 out of 7 clones in rap1Δ survivors were within this range, but the breakpoints in the other 5 taz1Δ survivor clones (one was ~300-bp centromeric, and four were 9~13-kb centromeric to the I-SceIcs) and two rap1Δ clones were located far (>10 bp) from the I-SceIcs. Therefore, we monitored DNA resection around DSBs indirectly through ChIP experiments of RPA. Localization of RPA subunit Rad11 was increased 2 hours after ahTET addition at 1.5 kb distant from I-SceIcs, while it was increased 4 hours at 5 and 13 kb both in wild type and taz1Δ strains (S6C Fig). These results suggest that Taz1 and Rap1 repress de novo telomere addition associated with DNA resection. The uncontrolled telomere addition may have contributed to the higher frequency of chromosome deletions found in taz1Δ and rap1Δ compared to wild-type clones (Fig 6E and 6F). As rap1-A and poz1Δ strains did not show increased survival, disruption of either Rap1 BRCT domain or Rap1-Poz1 pathway is not sufficient for inducing mutagenic DNA repair, and they function redundantly for suppressing telomere addition. We also examined distribution of breakpoints of telomere addition in rap1-A survivors. Since the breakpoints often located far from I-SceIcs (>10 bp, 3/6) similarly with taz1Δ and rap1Δ, the broader distribution of breakpoints shown in the shelterin mutants would be caused consistently by a loss of the Rap1 BRCT domain-dependent DSB repair pathway. Collectively, these results suggest that Taz1 represses GCRs by facilitating proper DSB repair and suppressing de novo telomere addition.
In a previous research, Taz1 was implicated in DSB repair because taz1Δ cells were sensitive to DNA damaging agents, such as methyl methane sulfonate (MMS) and bleomycin [31]. The sensitivity was augmented by simultaneous deletion of cds1+ [31]. We tested if a loss of the impaired DSB repair was responsible for the higher GCR rates in taz1Δ and rap1Δ. We found that the increased GCR rates and sensitivity to DNA-damaging reagents were not always correlated: rap1Δ was not sensitive to bleomycin (S6D Fig), consistent with a study showing that rap1Δ is not sensitive to MMS [36]. Moreover, taz1Δ cds1Δ and rap1Δ cds1Δ double-mutant strains showed decreased GCR rates relative to taz1Δ and rap1Δ (S6E Fig). These results indicate that the high GCR rates in taz1Δ and rap1Δ is not simply due to defective DSB repair.
Here we have described the hitherto unappreciated functions of the shelterin components Taz1 and Rap1 in the maintenance of genome integrity. Deletion of Taz1 and Rap1 led to increased spontaneous formation of GCRs, especially those involving chromosome terminus deletions associated with de novo telomere addition. Deregulation of telomerase accessibility appeared to be essential for frequent GCRs in taz1Δ and rap1Δ strains. Importantly, disruption of either the Rap1-Poz1 association (ΔP in Fig 5) or the BRCT domain of Rap1 did not induce high GCR rates (mutant A in Fig 5). By contrast, simultaneous deletion of the Rap1 BRCT domain and Poz1-binding domain led to higher GCR rates (Fig 5B). These results indicate that Rap1 suppresses GCRs through two independent pathways; one is BRCT domain-dependent, and the other is via a Rap1-Poz1 interaction that contributes to suppressing telomerase recruitment. These two pathways compensate for each other to suppress GCRs because inactivation of either one did not significantly affected the GCR rate, but simultaneous inactivation of both increased GCR rates in our system (Fig 5B).
How do shelterin components Taz1 and Rap1 repress GCRs at the breakpoint region? Although it has been reported that Taz1 can be recruited to telomere-like sequences outside telomeres, the breakpoint region of our GCR assay system does not have any telomeric DNA motif and a previous genome-wide chromatin immunoprecipitation analysis failed to detect Taz1 in the breakpoint region [49]. In addition, other genome-wide studies have indicated that telomeres are unlikely to reside stably in close proximity to the breakpoint region by examining 3D chromosome positioning in the nucleus [50, 51]. One possibility is that the higher GCR rates observed in the taz1Δ and rap1Δ strains were caused indirectly from the abnormal telomeres in these mutants. For example, aberrantly elongated telomeres (taz1Δ and rap1Δ) or gapped telomeres (taz1Δ) sequester substantial amounts of DSB repair factors [52], thereby compromising DSB repair efficiency outside telomeres. However, we showed that Taz1 still significantly suppressed GCRs in cells with no telomeric DNA (circular chromosomes) (Fig 4B). This result favors another hypothesis that DSBs directly recruit Taz1 and Rap1 in a telomeric DNA sequence-independent manner, rather than the indirect model. However, we have not detected any localization of Taz1 and Rap1 at DSBs in ChIP experiments (S6B Fig). It is possible that although Taz1-Rap1 associates with DSBs, the association is very limited either temporally or stoichiometrically, which made detection by the ChIP experiment difficult. In humans, it was reported that human TRF1 and TRF2 are recruited to DNA damage sites to promote homologous recombination-directed DSB repair (HDR). It is known that the association happens only transiently immediately after DSB induction [53–55]. In budding yeast, inner nuclear envelope protein Mps3 binds unrepaired DSBs, thereby spatially recruiting them close to the nuclear envelope [56].
We showed that Taz1 and Rap1 promote survival after a transient site-specific DSB induction, suggesting that they are involved in DSB repair. Notably, the experiment with constitutive DSB induction showed that GCR breakpoints in taz1Δ were in some cases far (> ~10 kb) from the original break site. These large deletions in taz1Δ can be explained by excessive resection of DSB ends or defective HDR. Firstly, Taz1 might suppress DNA end resection at non-telomeric DSB sites. This is not surprising given that Taz1 and Rap1 suppress extensive resection at telomeres [57]. Excessively resected DNA may provoke loss of the opposite strand, because it is known that 3’-end strands are degraded several hours after resection of the 5’ strand at a DSB in budding yeast [58]. It is possible that de novo telomere addition occurred at such new DSB sites distant from the original DSB site, which may account for the de novo telomere additions that are far from the original DSB site in taz1Δ cells. Alternatively, since the Taz1 homolog TRF1 promotes HDR at non-telomeric DSBs [52, 53], it is possible that Taz1 suppresses large deletions by promoting HDR. According to this scenario, in the taz1Δ strain, cells that engage in grossly defective HDR to repair DSBs are inviable, but an increased accessibility of telomerase to DSBs would promote formation of de novo telomere addition to the DSBs. In addition, the longer reaction time required by the inefficient HDR in taz1Δ, compared to taz1+, would also lead to extensive resection, resulting in large deletions. Similar to our results, budding yeast pif1 mutants, in which de novo telomere addition is highly promoted, showed large deletions concomitant with spontaneous and HO endonuclease-induced de novo telomere addition [52]. This also can be explained by inefficient HDR, because it was recently shown that Pif1 promotes HDR [58].
Previous studies with budding yeast and mammalian cells have proposed that telomere-binding proteins prevent GCRs solely by suppressing fatal inter-chromosomal fusions or de novo telomere additions. Our results, however, raise the possibility that components of the shelterin complex have a previously unappreciated mechanism for suppressing genome instability. Remarkably, our Rap1 truncation analysis demonstrated that the previously uncharacterized BRCT domain together with the Poz1-binding domain suppresses GCRs. Although we do not know at this moment the detailed molecular mechanism of GCR suppression by the BRCT domain, it does not depend on NHEJ because DNA ligase IV is dispensable for the increased GCR rate in rap1Δ. The BRCT domain-dependent mechanism would also be independent of telomerase regulation, given that the Rap1-A strain lacking the BRCT domain has normal telomere lengths [27]. Rather, the BRCT domain would be involved in DSB repair, because the Rap1-A strain showed reduced survival after transient DSB induction. Although the N-terminal BRCT domain is a conserved feature of Rap1 among other species, including budding yeast and humans, its function has been unclear, except that a mutation in the BRCT domain of budding yeast Rap1 affects its regulatory activity related to transcription [59]. It would be interesting to examine whether the BRCT domain of mammalian Rap1 is involved in suppression of genome instability or promotion of DSB repair. Further analysis of fission yeast shelterin components and their counterparts in mammals will reveal the detailed mechanism of the GCR suppression.
All of the experiments were performed using the S. pombe strains listed in S1 Table. Growth media, basic genetics, and construction of strains carrying deletion alleles or epitope-tagged proteins were described previously [60]. Point mutations were introduced using a QuikChange Multi Site-Directed Mutagenesis Kit (Agilent, 200514), or by manually using Dpn I and PCR. All plasmids constructed in this study were sequence-verified.
TK was cloned from FY2317 [61] and was flanked by a cytomegalovirus promoter and S. cerevisiae LEU2 terminator. ura4+ and TK were inserted between SPAC29B12.14c and SPAC1039.01 on chromosome I (at nt 5442736 of the chromosome I sequence described in Pombase: http://www.pombase.org/)
For immunoblot of Rap1 and Cdc2, polyclonal anti-Rap1 antibody [29] and Cdc2 p34 [PSTAIRE] antibody (sc-53, Santa Cruz Biotechnology, Inc.) were used.
For measurements of GCR rates, a previously described method using fluctuation analysis in budding yeast was applied to fission yeast, with some modifications [62]. Cells were streaked on YES agar plates and incubated at 32°C to form single colonies. Whole colonies were picked up by excising colonies with a sterile scalpel, suspended in YES liquid media, and incubated at 32°C until saturation. A portion of the saturated cells (at most 500 μl per plate) was plated on a YES agar plate supplemented with 1 mg/ml 5-FOA and 20 μg/ml FUdR. At the same time, 100 μl of a 105-fold dilution of the cell suspension was plated on another YES agar plate. Both plates were placed at 32°C, and colony numbers of selective and non-selective media were counted after 5- and 2-day incubations, respectively. The total number of GCRs in each liquid culture was estimated from the colony numbers using a following equation [63]:
m=(r/z‐0.693)/ln(r/z+0.367)
where m represents a number of GCRs formed de novo, not formed by duplication of pre-formed GCRs, in each total liquid culture, r represents the number of colonies on the selective media, and z represents the fraction of cells plated on the selective media.
This procedure was performed with at least 7 independent colonies, and median values of [(number of GCRs: m) / (total cell number)] were shown as the GCR rate. When the median value was zero, a tentative median value was calculated by assuming m = 1 in all colonies and shown in figures as an upper bound, as described previously [62]. Confidence intervals for the GCR rates were calculated as previously described [62], and shown in graphs as error bars. In order to determine statistical significance of differences in GCR rates, a Mann-Whitney test was performed.
To determine breakpoint sites in GCR survivors, breakpoint sequences were mapped to ~400 bp resolution by sequential PCR analysis, as previously described [25] (S7A Fig). For survivors with no loss of amplification by any primer sets, PCR products spanning ura4+ and TK (in GCR assay) or I-SceIcs (in the site-specific DSB assay) were sequenced to check whether they contained point mutations. To assess whether de novo telomere addition had occurred, PCR analysis was performed with a primer designed to anneal to a sequence centromeric from breakpoint and a primer including telomere sequence (M13R-T1: 5’-caggaaacagctatgacctgtaaccgtgtaaccgtaac-3’ or M13R-19: 5’-caggaaacagctatgaccctgtaaccccctgtaacc-3’; the underlined sequence was added to increase specificity of amplification after the 2nd cycle) using iProof DNA polymerase (Bio-Rad) (S7B Fig) [64]. Reaction mixtures were incubated at 98°C for 30 s, then cycled 30 times at 98°C for 10 s, 68°C for 20 s, and 72°C for 1 min. For survivors without de novo telomere addition, the breakpoint sequence was determined by ligation-mediated PCR, as described previously [65] (S7C Fig).
To examine how G1 arrest affects GCR frequency, we modified the protocol for measurement of GCR frequencies after treatment with DNA damaging agents [66]. Cells exponentially grown in YES liquid medium were divided equally, washed twice with EMM medium with/without ammonium chloride (nitrogen source), and incubated with the same medium for 24 hours. The incubated cells were washed once with YES medium and incubated with YES liquid medium overnight until saturation. Defined numbers of cells were plated on YES plates containing/lacking 5-FOA and FUdR. Resulting colony numbers were counted, and the ratio of surviving cells in 5-FOA and FUdR was calculated to obtain the respective GCR frequencies.
Site-specific DSB induction was performed essentially as described previously [67], with the following modifications. The tetracycline-inducible I-SceI integration plasmid containing LEU2 was integrated at the leu1-32 locus. A plasmid containing the I-SceI cleavage site as well as ura4+ and hygromycin B selectable markers was integrated at the same site as the ura4+-TK cassette in our GCR assay strain. DSB was induced by addition of ahTET (Sigma, 3 μM final).
qPCR was performed using a StepOnePlus real-time PCR system (Applied Biosystems). Sequences of PCR primer sets are listed in S2 Table.
ChIP assays were performed essentially as described previously [13], with the following modifications. The cell concentration was adjusted to 1.0 × 107 cells/mL just before addition of ahTET and a defined volume of the culture was collected at indicated time points for fixation. Immunoprecipitation was performed with anti-myc antibody (9B11, Cell Signaling) using Dynabeads M‐280 Sheep anti‐Mouse IgG (Invitrogen). DNA was purified and extracted from washed beads and input samples with Chelex 100 Resin (BioRad) as described previously [68], and analyzed by qPCR.
According to a previous report [26], fission yeast acquires 5-FOA resistance by spontaneous inactivating point mutation of either ura4+ or ura5+, and the rate of spontaneous mutation which confers 5-FOA resistance is 1.3 × 10−7 (/cell division). Among these mutations, ratio of mutation in ura5/ura4 is 1.85, so inactivating mutation rate of ura4+ can be calculated as 1.3 × 10−7 / (1.85+1) = 4.6 × 10−8. Given that the length of ORF of ura4+ and TK are ~800 and ~1100 bp, respectively, we estimated mutation rate of TK as 4.6 × 10−8 × (1100/800) = 6.3 × 10−8. If this estimate is true, the rates of point mutations that confer 5-FOA and FUDR resistance are both nearly or less than 10−7 per cell division.
|
10.1371/journal.pbio.1002478 | Indirect Fitness Benefits Enable the Spread of Host Genes Promoting Costly Transfer of Beneficial Plasmids | Bacterial genes that confer crucial phenotypes, such as antibiotic resistance, can spread horizontally by residing on mobile genetic elements (MGEs). Although many mobile genes provide strong benefits to their hosts, the fitness consequences of the process of transfer itself are less clear. In previous studies, transfer has been interpreted as a parasitic trait of the MGEs because of its costs to the host but also as a trait benefiting host populations through the sharing of a common gene pool. Here, we show that costly donation is an altruistic act when it spreads beneficial MGEs favoured when it increases the inclusive fitness of donor ability alleles. We show mathematically that donor ability can be selected when relatedness at the locus modulating transfer is sufficiently high between donor and recipients, ensuring high frequency of transfer between cells sharing donor alleles. We further experimentally demonstrate that either population structure or discrimination in transfer can increase relatedness to a level selecting for chromosomal transfer alleles. Both mechanisms are likely to occur in natural environments. The simple process of strong dilution can create sufficient population structure to select for donor ability. Another mechanism observed in natural isolates, discrimination in transfer, can emerge through coselection of transfer and discrimination alleles. Our work shows that horizontal gene transfer in bacteria can be promoted by bacterial hosts themselves and not only by MGEs. In the longer term, the success of cells bearing beneficial MGEs combined with biased transfer leads to an association between high donor ability, discrimination, and mobile beneficial genes. However, in conditions that do not select for altruism, host bacteria promoting transfer are outcompeted by hosts with lower transfer rate, an aspect that could be relevant in the fight against the spread of antibiotic resistance.
| In bacteria, genes can move between cells, sometimes with the donor host cell actively involved in the gene transfer mechanisms. This movement of genes is called horizontal gene transfer, and it increases the prevalence of mobile genes in bacterial populations. However, it is not clear if donor host cells benefit from gene spread, or are simply exploited by selfish genes. Here, we show with both modelling and experiments that for the donor host, investing in the transfer of beneficial genes—such as those conferring antibiotic resistance—can be understood as an altruistic behaviour. This behaviour is costly to the donor but beneficial to recipients and can be selected for if a sufficient proportion of recipient cells share the donors’ transfer allele. Preferential transfer from donors towards recipients that share this allele occurs when dispersal is limited or if discrimination mechanisms are present. Our work suggests that both processes are likely to be widespread in nature, promoting horizontal gene spread by host donor cells. As many antimicrobial resistance and virulence genes are mobile, our work further implies that the spread of harmful traits among human pathogens may be modulated by host bacteria in a direction that depends on the bacterial ability to transfer the traits specifically to their kind.
| MGEs such as plasmids or phages are defined by their ability to undergo horizontal gene transfer (HGT) between bacterial hosts [1], and are widespread in nature. Genes present on MGEs often affect their hosts’ fitness in a specific environment [2]. Particularly, many mobile genes increase virulence or antibiotic resistance and thus have harmful consequences on human health. Antibiotic resistance genes are enriched on plasmids [3], leading to their fast spread among bacterial species via horizontal transfer [4]. Genes coding for secreted proteins, often involved in virulence, are also enriched on MGEs promoting cooperative secretion [5,6]. In order to better combat the medical issues arising from horizontal transfer, we must understand the selective pressures acting on gene mobility. The population dynamics and evolution of transfer have mostly been studied by focusing on MGEs themselves [7]; however, transfer is influenced not only by MGE genes, but also by genes of the bacterial host chromosome. Both donor [8] and recipient cells [9,10] can regulate transfer, with different donor and recipient genetic backgrounds resulting in as much as eight orders of magnitude variance in the transfer rates for the same plasmid [11,12]. Thus, to fully understand the evolution of horizontal gene spread and the natural variation in transfer rates among hosts, we must consider the selective pressures acting on hosts.
On one side, horizontal transfer confers varied and often extreme costs onto the bacterial host. Phage mobility usually requires host cell lysis that leads to death, while plasmid transfer through conjugation renders host cells sensitive to male-specific phages [13] and decreases the host's growth rate and fitness [14,15]. Because of these costs, horizontal transfer has classically been considered as a selfish trait of parasitic MGEs, selected as it favours their spread [7]. Direct support for transfer being a purely costly trait to the host came from studies of plasmid–host coevolution, where host genes that decrease transfer were selected [16]. On the other side, it has also been suggested that HGT could benefit the host because of the transfer of accessory genes not directly involved in MGE maintenance and transfer. Indeed, MGEs are often thought to constitute a communal pool of genes [2], a flexible genome [17] that can be quickly shuffled by HGT in response to environmental changes, making host populations more robust [18]. In this view, HGT is beneficial to the host population because it allows cells to share beneficial traits and provides diversity at the population level. However, it is not clear that these proposed benefits are sufficient for HGT to be favoured by the host. Traits advantageous at the group level—here the maintenance of a communal pool of genes—are not necessarily selected for at the individual level, especially when individuals can benefit from others that invest in the trait while not paying the cost of investing themselves [19]. Indirect, population-wide benefits alone are not necessarily sufficient to explain the selection of host genes promoting costly transfer [20]. The ability to receive genes can clearly be directly selected for when these genes enhance individual fitness: for instance, CRISPR immunity against antibiotic resistance plasmids, a form of HGT resistance, was rapidly lost in the presence of antibiotics when receiving plasmids was beneficial to the host [21]. On the contrary, the ability to donate genes need not be selected, as the donor cell does not directly benefit from transferring genes to neighbouring recipients.
To quantitatively understand HGT, the selection acting on donor ability must be analysed in a social context, taking into account both the costs and benefits transfer bestows onto donor and recipient hosts. Here, we theoretically and experimentally analyse the evolution of host genes controlling plasmid transfer. We show that from the host side, transfer represents a form of altruism: actors pay a cost of investing in transmission and deliver a benefit to recipients of beneficial mobile elements. Altruistic donation of MGEs can be maintained when transfer is sufficiently biased towards cells sharing donation alleles, increasing the donor allele inclusive fitness. This bias can arise in structured populations or by an association between transfer and discrimination alleles. Fitness gains due to the transfer of mutualistic plasmids further select for genotypes where donor ability alleles, discrimination alleles, and mutualistic plasmids are associated.
We first perform a qualitative analysis to identify if and in which conditions a strain with high donor ability can be selected. We model the fitness of nonmobile host genes controlling donor ability for a given plasmid using a neighbour-modulated fitness approach that partitions fitness into the effects of an individual’s own genotype and those of social neighbours [22,23]. We consider a population of bacteria structured in an infinite number of patches [24] and model a simplified life cycle with nonoverlapping patch generations, in which the following processes occur successively [25,26]: founding, reproduction, transfer, selection, and dispersal (see S1 Text for details).
A cell i in patch j is characterized by three traits: plasmid carriage pij (pij = 1 for plasmid-bearing cells and 0 for plasmid-free cells), donor ability qij and recipient ability sij. Successful transfer is controlled by three factors: the probability of contact between plasmid-bearing and plasmid-free cells, the donor ability of plasmid-bearing cells, and the recipient ability of plasmid-free cells. We assume that plasmid and host traits are distributed independently in the starting population so that the cell's donor ability qij is independent from its initial plasmid content pij. Initially uninfected cells become infected with a probability proportional to the patch level frequency of plasmid-bearing cells modulated by their own recipient ability and by the average patch donor ability. A cell i in patch j will thus be modified by transfer with the probability (1 − pij) pj qjsij.
Plasmid presence has an effect ep on the host cell, and we can express the plasmid effect on host fitness as ep pij. The cost of donor ability is cq leading to an effect of transfer on host fitness that is proportional to donor ability, experienced only by cells bearing plasmids before transfer, and equal to −pij cq qij. Donor ability is costly independently of actual transfer efficiency, modelling the effect of expressing the transfer machinery (which happens even in the absence of successful transfer).
The fitness of an individual founding cell i in patch j, measured over the patch life cycle, is noted by Wij. With W0 being the basal host fitness, we obtain (see S1 Text):
Wij=W0+ep[pij+(1−pij)pjqjsij]−pijcqqij
(1)
To understand selection acting on donor ability q, we apply the Price equation [27,28] to Eq (1). We obtain the regression coefficient between fitness and donor ability, β(Wij,qij), that describes the effect of donor ability on fitness (Eq 2). We provide the derivation of Eq (2) and a detailed analysis in S1 Text.
The Ej[pj(1 − pj)] term describes the effect of patch composition on the efficiency of plasmid transfer: transfer events are more likely when both plasmid-bearing and plasmid-free cells are abundant within each patch. β(qjsij,qij) is a regression coefficient between individual donor ability qij and the product of individual recipient ability with patch-level donor ability. It corresponds to the relatedness between plasmid donor and recipient cells, noted by Rq, at the locus determining donor ability (see S1 Text for a detailed analysis): Rq is higher when donor cells preferentially encounter recipients that share their donation allele and when transfer is more successful towards those cells. Rq thus determines how much a donor cell transfers plasmids to individuals bearing the same donation allele because of population structure and specificity in transfer. Finally, p cq is the average cost of transfer for the donor genotype: high donor ability is costly to the proportion of cells that bear plasmids and express their transfer machinery.
An increase in donor ability is selected for when it is correlated with increase in fitness, namely when β(Wij,qij) > 0, which combined with Eq (2) leads to the following condition:
epEj[pj(1−pj)]Rq>pcq
(3)
Eq 3 is a form of Hamilton’s rule [29], which postulates that a cooperative allele is selected for when its indirect benefits, weighted by relatedness among actors and recipients, outweigh its direct cost, maximizing its inclusive fitness (fitness inclusive of alleles present in other individuals). Applied here to donor ability, the indirect benefits are the benefits of plasmids to the recipient cells after transfer ep Ej[pj(1 − pj)], and the direct cost is the cost of donor ability for cells bearing plasmids p cq. Rq is the relatedness at the donor ability locus among donor and recipient cells of plasmid transfer. High, positive Rq implies that most of transfer events from cells with high donor ability will be directed towards recipients sharing their donation allele. On the contrary, negative Rq means that transfer will be biased towards cells with a different allele than the one carried by the donor. Thus, a high donor ability allele can be selected even when individually costly, when transfer maximizes its inclusive fitness through plasmid effects on recipient cells. We note that relatedness in bacteria can vary across loci [30], as it can be modified in a locus-specific way by mutation [31] or HGT [5,6]. Thus, unlike relatedness arising from genealogical kinship in sexually reproducing organisms, it will not necessarily tend to be the same across the genome. To underline this and avoid any potential semantic confusion, we follow nomenclature defined already in [30] and consider that cells that specifically share alleles at the locus of interest (plasmid donation) are cells of the same kind but not necessarily kin.
Rq will be positive when donors preferentially transfer plasmids to recipients of their kind. Positive relatedness generally arises through the combination of two processes: limited dispersal and discrimination mechanisms [29,32]. Here, limited dispersal is due to patch structure: the correlation between qj and qij is governed by the initial repartition of genotypes among patches, with no migration before transfer occurs. Positive relatedness can arise from strong population bottlenecks leading to stochastic variations in founding cell frequencies among patches, followed by clonal reproduction [33]. Effective discrimination in transfer also leading to positive relatedness arises if sij and qij are positively correlated, with genotypes with high donor ability having higher recipient ability than average, or if donors have a way to direct transfer specifically to their kind (see S1 Text for discussion). Alternatively, negative relatedness can arise if sij and qij are negatively correlated, leading to preferential transfer to cells bearing a different donation allele.
We can distinguish two scenarios for the effect and selection of transfer depending on the plasmid effects on the host cell. In the first case, the transferred plasmid is mutualistic with its host (ep > 0), for instance conferring antibiotic resistance: transfer is therefore an altruistic behaviour [29] with a direct cost of performing transfer and indirect benefits through the plasmid benefits in recipient cells. Transfer is selected if Rq is positive and sufficiently high: Rq > p cq / (ep Ej[pj(1 − pj)]). In the second case, the transferred plasmid is parasitic (ep < 0): donor ability for parasitic plasmids, decreasing the fitness of recipient cells, is selected if Rq is negative and sufficiently low: Rq < p cq / (ep Ej[pj(1 − pj)]). This would be a case of spiteful behaviour [32,34]. Specific population structure or discrimination processes are required to produce negative relatedness, and spite is thus thought to be less common than altruism [34].
We focus here on the transfer of mutualistic MGEs and more specifically on antibiotic resistance plasmids that allow their hosts to grow when antibiotics are present. The main prediction arising from our model is that donor ability for these mutualistic plasmids is an altruistic trait, counterselected if transfer occurs indiscriminately towards any cell, but selected for when plasmid donors and recipients share donation alleles. We present the model graphically in Fig 1, focusing on the three relevant scenarios affecting relatedness: random interactions between individuals (Fig 1A), discrimination in transfer (Fig 1B), and structured populations (Fig 1C). We next test the model’s predictions with both simulations and experiments, performing competition assays between strains with varying donor ability in order to investigate quantitatively if and how much selection favours donor ability in biologically realistic settings.
Discrimination in transfer occurs if during the encounters between a donor and potential recipients the plasmids are transferred to cells of the donor’s kind more often than would be expected based on its frequency in the population. Discrimination of plasmid recipients could be based on differences in the initial recognition between cells or differences in plasmid establishment in recipient cells. To search for evidence of discrimination, we analyse two available datasets [11,12] that quantify plasmid conjugation rates among different pairs of natural isolates. Both studies measured conjugation rates for the multiresistant R1 plasmid, among 10 strains from the ECOR collection [11] or 9 other natural Escherichia coli strains [12]. In each dataset, we compute normalized donor ability for each pair of donor and recipient strains (see Materials and Methods), which corrects for basal differences in donor ability between strains. We find that transfer to self occurs at rates higher than average in 18 out of 19 cases (Fig 2A). Additionally, in 8 out of 19 cases, the highest rate of transfer is from a strain to itself. Overall, transfer to self is 7.3 times higher than average donor ability over all tested isolates (two tailed t test for difference from 0 for normalized donor ability to self, p = 0.0003). In a mixed population with many different strains, the high rates of transfer to self we describe here would translate into a biased transfer between cells sharing donation alleles. This apparent discrimination does not imply that the same locus is responsible for high donor ability and for discrimination, as multiple genes could be involved in discrimination. However, the signal we observe in Fig 2A suggests that alleles for high donor ability and for discrimination in transfer are linked in natural isolates sufficiently to lead to an effective discrimination at the donor ability locus.
We next experimentally investigate if discrimination may allow for the selection of host transfer genes. We use two widely studied E. coli strains, the K12 strain MG1655 [35] and the B strain REL606 [36], and the multiresistant R1-19 plasmid [37]. K12 and B strains bear different restriction-modification systems [9], potentially leading to discrimination in plasmid transfer [38,39]. We first measure the conjugation rate in a well-mixed environment between all four combinations of K12 and B as donor and recipient strains (Fig 2B, red bars). We find that K12 is generally a better donor, but also that K12 transfers R1-19 plasmid to itself at a 5-times higher rate than to B (Mann-Whitney Wilcoxon test, p = 0.003). Overall, the K12 strain is an example of good donor strain displaying discrimination for transfer, in comparison to the lower donor B strain. Moreover, R1-19 carriage leads to a 54% reduction in exponential growth rate for K12 strain, compared to a 1.6% reduction only for B strain (S1 Fig). To test if part of the costly effect R1-19 has on K12 is due to donor ability, we use a K12-derived strain with a deletion in the arcA gene, a gene known to affect transfer [40], and the repressed R1 plasmid, which transfers approximately 1,000-fold less than R1-19 [37,15]. The K12ΔarcA mutant transfers R1-19 plasmid at a strongly reduced rate to both itself and B (Fig 2B, grey bars), and R1-19 cost is reduced as well (6.4%, S1 Fig). Similarly, R1 plasmid imposes almost no cost to K12 growth (1.8%, S1 Fig). Both results suggest that most of R1-19 cost to K12 is due to its high transfer rate. We then test whether K12 discrimination in transfer can lead to biased transfer towards other K12 cells in a well-mixed population and subsequent selection of the better donor strain. We compete the K12 and B strains by mixing them equally in a well-mixed population, with a common proportion of cells from each strain initially bearing R1-19 plasmid. In the absence of antibiotic selection, the better donor K12 decreases in frequency (Fig 2C, dark blue), showing a lower basal fitness than B in those culture conditions. When antibiotic selection is applied at the end of the competition by plating the population on kanamycin (Kn)-containing medium, only Kn-resistant, plasmid-bearing cells grow. When all cells initially bear plasmids, selection does not favour the K12 strain (Fig 2C orange, 19% decrease in K12 frequency, two sided t test for difference from 0, p = 0.003). However, when only a fraction (2.5%) of both K12 and B cells initially bear R1-19 plasmid, providing opportunity for plasmid transfer, K12 is selected (19% increase in K12 frequency, two sided t test for difference from 0, p = 0.009). Finally, to confirm that this specific selection of donors is due to R1-19 transfer to K12 cells, we analyse the outcome of competition when the arcA gene is deleted from K12 and transfer is impaired. In the absence of antibiotic selection (Fig 2C, light blue), the arcA deletion does not affect K12 fitness when plasmids are absent or rare and increases K12 fitness when all cells bear plasmids (11% decrease for K12ΔarcA versus 25% decrease for K12, two-sided t test, p = 0.043), possibly because of the reduced plasmid cost for K12ΔarcA. With antibiotic selection, the specific selection of K12 when a fraction of cells bear plasmids disappears for K12ΔarcA (Fig 2C middle, yellow bars, two-sided t test, p = 2.10−5), demonstrating that K12 selection was due to plasmid transfer. Discrimination effectively biases antibiotic resistance transfer strongly enough so that the better donor K12 strain is selected for in the presence of antibiotics. Thus, when transferred plasmids are needed for growth, discrimination in transfer towards kind, at naturally appearing levels, can be sufficient to select for the better donor.
A second possible reason for transfer bias is bacterial growth in structured populations, where donors interact preferentially with their kind. Next, we examine whether, in the absence of discrimination, structured populations can provide a sufficient bias in transfer to select for good donors. To analyse the effect of biased transfer in structured populations, we use a synthetic system with fluorescently tagged plasmids in which we can identify plasmid transfer between two strains. We adapted the system from the one we designed for an earlier study on interaction between conjugation and cooperation [6]. A helper plasmid FHR, that is nonmobile and thus behaves like a chromosomal allele, governs the host cell donor ability for a mobile plasmid C, which confers chloramphenicol (Cm) resistance. We compete (Fig 3A) two strains differing in their donor ability: the good donor D+ strain bears FHR that transfers C plasmids, and the nondonor D− strain does not (S2A Fig). After a transfer phase (t0 to t1), populations are grown with or without Cm during the selection phase (t1 to t2). We compare a single, well-mixed population (m), where D+ and D− are mixed in equal proportions, to a structured metapopulation (s), consisting of two subpopulations that grow separately during the transfer phase, s1 and s2, founded respectively with a 10% and 90% proportion of D+ (leading to equal proportions of D+ and D− at the metapopulation level). In this setup, the changes in the good donor frequency can be followed both within and among populations to evaluate the effect of population structure on donor selection.
D+ strain frequency does not change significantly in m or s populations during the transfer phase (Fig 3B left). D+ frequency then increases at t2 only for the structured s population grown in the presence of Cm (Fig 3B right, 26% increase from t0, Mann-Whitney Wilcoxon test, p = 0.004). It decreases for m population with Cm (19% decrease, two-sided t test for difference from 0, p = 3.10−9) and stays constant in the absence of Cm. The dynamics generally follows our predictions: D+ selection requires both population structure and plasmid selection. However, the expected cost of D+ during the transfer phase is not present at the population level. We next investigate in more detail both this cost and the selection of the good donor strain.
By looking at the dynamics of individual subpopulations during the transfer phase, we observed that D+ increases in frequency when prevalent (S3A Fig). We confirmed with an independent experiment that D+ fitness in competition with D− linearly increases with D+ frequency (S3B Fig). This positive frequency-dependence for donor fitness could be due to lethal zygosis, a phenomenon known to damage recipients at high donor cell frequencies [41], which could be aggravated by the absence of entry exclusion in our strains [42]. In natural systems, entry exclusion may protect new transconjugants but would also make initial plasmid-bearing donors immune to lethal zygosis, probably leading to a similar frequency-dependence of fitness when most donor cells initially bear plasmids. In our system, frequency-dependence leads to no observable cost for D+ at the metapopulation level. At low frequencies, donor ability still has a cost, which is also observed as a decrease in the strain’s growth rate when growing in isolation (D+ versus D−, S2C Fig). Interestingly, donor cells grow significantly more slowly when they bear C plasmids, which is not the case for nondonor cells (D+C versus D−C, S2C Fig), suggesting that donor ability cost is enhanced by the presence of transferable plasmids in the cell.
During the selection phase, good donors are selected only in the structured s population and only in the presence of Cm, meaning that donor selection requires both population structure during the transfer phase and subsequent antibiotic selection. We see that, as predicted by our model, biased transfer due to population structure promotes indirect selection of the donor strain. To better understand the factors affecting D+ selection, we proceed to analyse the dynamics of C plasmids (Fig 3C). During the transfer phase, plasmid frequency changes depend on the proportion of cells able to transfer. In the s1 population where D+ cells are few, plasmid frequency declines slightly. It increases mostly in the s2 population enriched in D+ strain. Increases are due to transfer, as the increase in plasmids present in D− strain is due to plasmids that originate from D+ (as identified by fluorescence markers, see Materials and Methods and S4 Fig). We then follow the proportion of C plasmids that are present in D+ cells, as plasmid localization controls survival in the presence of antibiotics. During the transfer phase, the proportion of C plasmids present in D+ cells compared to D− cells decreases in the well-mixed m population (13% decrease, Mann-Whitney Wilcoxon test, p = 0.004) but increases in the structured s population (28% increase, Mann-Whitney Wilcoxon test, p = 0.004) (Fig 3D). In the well-mixed population, the decrease is probably due to the strong fitness cost D+ cells incur specifically when they bear C plasmids (S2 Fig). The same cost also explains the subsequent decrease in D+ strain frequency under Cm selection in both populations. The enrichment of C plasmids in D+ cells depends on the population structure of the s population: total plasmid transfer is more prevalent in the s2 subpopulation, effectively biasing transfer towards D+ at the metapopulation level. These results experimentally confirm our models prediction: in the absence of discrimination mechanisms, donor ability for antibiotic resistance plasmids can be selected when population structure ensures preferential transfer to cells sharing donation alleles.
So far, we have shown that both discrimination and population structuring can select for donor ability. However, we have always assumed and ensured that discrimination and population structure are present in the system. Here, we study how both phenomena can themselves emerge.
When analysing effects of population structure, we imposed starting proportions of both the strains and the plasmids in order to dissect the dynamics of transfer. We continue by studying the outcome of competitions arising in more natural population structures using simulations. In the absence of a mechanism for discrimination of recipients, our model suggests that population structure may influence the selection of donor ability in two opposing ways (Eq 3). Efficient transfer is favoured by the coexistence of plasmid-bearing and plasmid-free cells within patches, while biased transfer towards kind is favoured by high relatedness at the donation locus. The selection of transfer thus requires preferential interactions between cells sharing donor ability alleles but also sufficient cell mixing that would favour the encounter between plasmid-bearing and plasmid-free cells. To study the effects of a natural population structure and the possibility that both conditions are met simultaneously, we simulate strong population dilution, which leads to stochastic founder cell numbers and genotype frequencies [6,33]. We follow the frequency of the donor D+ strain in a simulated metapopulation initiated from a strongly diluted mix of equal proportions of D+ and D−. With growth parameters based on our experimental results, we vary both the dilution factor applied to founding populations and the proportion of plasmid-bearing cells (Fig 4). The results exhibit a clear pattern: the donor strain is selected under the combination of strong initial dilution and low initial plasmid frequency. D+ selection in the presence of antibiotics is controlled primarily by the enrichment of plasmids in D+ at t1 (S5A Fig), which occurs when there is biased transfer towards D+ during the transfer phase. As predicted by Eq 3, biased transfer requires high relatedness at the donation locus but also effective transfer, which will be affected by dilution and initial plasmid prevalence. First, diluting down to low founding cell numbers provides sufficient variation in D+ frequencies among populations to ensure high relatedness (S5B Fig). Second, the contribution of transfer to plasmid abundance declines with increasing dilution and increasing plasmid initial abundance (S5C Fig). Transfer promotes plasmid invasion primarily when plasmid-bearing cells are initially scarce, because in those conditions, the majority of plasmid-bearing cells actually arise from transfer. Increasing dilution leads to the frequent absence of one of the cell types from each population, which in turn decreases the number of possible encounters between plasmid-bearing and plasmid-free cells. Selection for donor ability is strongest at high dilution because of high relatedness; however, dilution simultaneously limits transfer, which decreases the strength of selection compared to the optimal population structure studied in Fig 3. Overall, our simulations suggest that the conditions for selection of donor ability can be met in natural environments through limited dispersal alone, despite the trade-off between relatedness and transfer efficiency that arises from population structure.
We now focus on the association between discrimination mechanisms and plasmid transfer, which was apparent in natural isolates data (Fig 2A) and ask how this association itself could emerge. The effect on donor ability (Fig 2B and Fig 2C) suggests that a genotype bearing both discrimination and high donor ability alleles could be favoured by selection. We consider as an example the case of a strain inactivated at a chromosomal restriction-modification locus: such a mutant appearing in a wild type population will transfer plasmids preferentially to clonemates bearing the same allele, as unmodified plasmids transferred from this mutant will be degraded in a wild-type recipient cell [38,39]. The mutant allele with no modification is denoted by M−, and the wild type allele by M+, D+, and D− stand for high and low donor ability, as before. We use simulations to follow the dynamics of modification and donation alleles, in populations of cells with no initial association between M− and D+. In a well-mixed population (analogous to the one studied in Fig 2C), the M− allele is selected for, but the D+ is not (Fig 5A, left). Discrimination in transfer by M− cells leads to a reduced total transfer to M+ cells, and direct selection of M− with comparatively higher recipient ability. D+ cells are outcompeted, as they do not receive more plasmids than D− cells in the absence of association to M− alleles. On the contrary, in a population where D+ cell frequency differs sufficiently among subpopulations, both M− and D+ alleles are selected for (Fig 5A, right): similarly to the dynamics presented in Fig 3, population structure biases transfer towards D+ cells, allowing for D+ selection in the presence of antibiotics. Linkage between M− and D+ alleles is also controlled by population structure (Fig 5B): with antibiotic selection, positive linkage appears when D+ cell frequency varies among subpopulations. This does not occur in the absence of antibiotics, suggesting that linkage is due to the specific selection of plasmid-bearing cells. With increasing D+ population structure, most plasmids end up in D+ M− cells (Fig 5C) due to the combined effect of higher recipient ability and biased transfer by D+ cells. We conclude that the association between discrimination and transfer alleles can emerge simply through selection of plasmid-bearing cells, when population structure ensures that cells with high donor ability are favoured.
Selective pressures acting on donor ability depend on the fitness effects of genes carried by the transmitted plasmids. Here, we focus on mutualistic antibiotic resistance plasmids, but parasitic and mutualistic plasmids can coexist in host populations, and hosts may not be able to evolve differential control of transfer based only on the accessory genes plasmids carry. Next, we consider the presence of parasitic plasmids N, which are similar to mutualistic plasmids C but do not confer benefits during the selection phase (see S2B Fig and Materials and Methods). In simulations, increase in the initial proportion of parasitic plasmids lead to a decrease in final frequency of the good donor strain (S6 Fig). We can conclude that the strength of selection for donor ability decreases when donors encounter a mix of mutualistic and parasitic plasmids, as they cannot distinguish between the two plasmid types. However, the benefits conferred by mutualistic plasmids could indirectly favour their association to the host.
To study the association between plasmids and hosts, we measure experimentally and with simulations the linkage disequilibrium between plasmids and the good donor allele in a metapopulation similar to the previous one (Fig 3) but now with a mix of C and N plasmids in equal proportions instead of only C plasmids, and with no initial association between plasmids and a specific strain (each plasmid is equally present in D+ and D- strains). After the transfer phase, both C and N plasmids become significantly linked to the good donor D+ strain in the structured population, but not in the well-mixed population (Fig 6A, left). Moreover, after Cm selection, linkage to D+ slightly increases for C plasmids but decreases to zero for N plasmids in the structured population and remains at zero in the well-mixed population (Fig 6A, right, plain lines). Finally, this pattern relies on the specific selection of C-bearing cells: when selecting with Kn (an antibiotic to which both plasmids confer resistance) instead of Cm, linkage decreases to zero for both plasmids (dashed lines). Our experiments show that linkage between plasmids and good donor cells arises only when two specific conditions are met: (1) the population is structured, and (2) plasmids are beneficial.
To better understand the factors controlling the association, we independently vary D+ donor ability and C plasmid beneficial effect on the host strain fitness in our simulations. The linkage that appears at t1 between both plasmids and D+ strain increases with D+ donor ability, independently of subsequent plasmid benefits (S7A Fig). This linkage arises because transfer is biased towards D+ cells at the metapopulation level (as seen in Fig 3). In our experiments, the observed linkage is stronger for N plasmids because they are preferentially transferred. At t2, however, C plasmid benefits modify the linkage patterns. With increasing C benefits, C linkage to D+ increases (Fig 6B), and N linkage to D+ decreases (S7B Fig), when donor ability is sufficiently high. The specific association between C plasmids and the donor strain thus arises from the benefits provided by C plasmids to the host: antibiotics promote the selection of cells bearing C plasmids, which are mainly good donor cells because of previous transfer. Overall, this mechanism selects for cells simultaneously bearing both D+ and C alleles. Linkage between the good donor strain and beneficial plasmids arises without directly enforcing any association between the two, due to the combination of two effects: population structure biasing transfer towards good donor cells and the plasmids benefiting the host.
Our work investigates the evolution of host genes controlling the transfer of mutualistic MGEs such as those conferring antibiotic resistance. We focus on genes modulating plasmid donation, a property that, unlike plasmid reception, does not directly benefit the host. Earlier interpretations have described MGEs as a communal pool of genes conferring benefits at the population level [2,18]. We demonstrate here that donor ability, when it is costly to the host, is not selected directly. However, we do not need to invoke population-level benefits to explain why the host may promote MGE transfer. Instead, we show that host donor ability alleles can be selected indirectly when transfer increases their inclusive fitness (Fig 1). We then investigate further this qualitative result by measuring selection direction and strength in simulations as well as experiments using both natural and synthetic microbes, in situations close to ones that could be observed in nature.
In the absence of discrimination, population structure is a simple mechanism ensuring that cells encounter preferentially neighbours of the same kind. Here, we demonstrate that in a synthetic biological system devoid of any mechanism for discrimination in transfer, population structure enables the selection of donor ability, biasing plasmid transfer prior to the selection of plasmid-bearing cells (Fig 3). Donor ability is not selected within well-mixed populations where donors do not interact preferentially with their kind, and good donors decline in frequency due to donor ability costs. However, donor ability is selected at a metapopulation scale, where population structure provides sufficient relatedness at the locus controlling donor ability. With simulations, we then show that populations with sufficient relatedness can arise simply through strong population dilution, despite the reduction in transfer due to fewer interactions between plasmid-bearing and plasmid-free cells (Fig 4).
Further, we experimentally demonstrate that differences in transfer rates between isolates, leading to effective discrimination in transfer, can also be sufficiently high to favour a strain with high donor ability (Fig 2). In natural isolates, we observe discrimination for the transfer of an antibiotic resistance plasmid. These results motivate future studies that would quantify the generality of discrimination by examining other plasmids and strains, as well as determine the underlying mechanisms. Discrimination can result from specific recognition during cell–cell contact [30] or even direct spread through the cytoplasm of clonemates in the case of bacterial chains [43]. Alternatively, discrimination can arise during the establishment of plasmids in recipient cells. In particular, plasmid transfer rate is greatly diminished when restriction-modification systems present in recipients differ from those in donor cells [9,38,39]. At a larger phylogenetic scale, a plasmid host range can be limited by its mechanisms of replication or transfer [44]. Even when plasmids are successfully transferred, they need not confer any fitness benefit, because genes beneficial in the initial donor may be suboptimal in a novel, unfamiliar host [45], favouring a donor strain over distant competitors in which the transferred accessory genes are not fully beneficial. Finally, discrimination may rely on quorum-sensing mechanisms regulating transfer [46], which can provide an indication of the local abundance of related cells.
Any of these mechanisms could lead to discrimination among transfer recipients, but they may not all be controlled by the same locus as donor ability. Discrimination by plasmid donors towards their kind necessitates genetic linkage between donation and discrimination alleles. The pattern observed for natural isolates (Fig 2A) suggests that a sufficient level of association does exist in nature, at least for the R1 plasmid. Moreover, this observation may be explained and maintained by the dynamics we describe in Fig 5, where linkage between discrimination and donor alleles emerges from their coselection in structured populations.
Biased transfer to kind can thus happen in host cells that differ from others at a single locus modulating donor ability in structured populations; the benefits of transfer then promote the emergence of discriminating genotypes through linkage with a second locus determining the specificity of transfer. Population structure plays a central role, allowing both the spread of donor alleles in the absence of discrimination mechanisms and the emergence of discrimination.
As the selective pressures we describe here are indirect, they may be too weak to have a significant effect on the evolution of transfer rates. To examine this, we calculate selection coefficients acting on the donor allele in our experiments and simulations. The strength of selection observed for the discriminating strain K12 in competition with B is high (s = 0.35, S8 Fig). As the degree of discrimination displayed by K12 is close to the average one measured across natural isolates (Fig 2A), this result suggests that selection of donor strains that transfer preferentially to their kind may occur widely in nature, even in unstructured populations. In the structured populations we studied, the strength of selection depends on the details of population structure: when relatedness at the donor ability locus is high but plasmids are present in each strain in equal abundance, donors are again efficiently selected for in our experiments (s = 0.10, S9 Fig). When both parameters are controlled purely by initial dilution, they behave in opposing ways and selection is lower (S10 Fig, s = 0.0025 in the optimal case). In natural populations, selection arising through population structure might thus be weaker than the one due to discrimination in transfer and vary depending on the details of host and plasmid population dynamics. Still, bacteria are characterized by large population sizes, leading to estimates of effective population sizes around 107 [47,48], which implies that mutations with selection coefficients larger than 10−7 can be selected for [49]. Thus, even in the presence of a trade-off between relatedness and transfer efficiency, selection acting on hosts can result in biologically significant changes in transfer rates.
In the long-term, continued selection for transfer requires that plasmids do not spread to fixation. As plasmid transfer itself increases plasmid prevalence in host populations, selection for donors will be progressively decreased with plasmid spread. However, many factors may contribute to maintaining plasmids at intermediate frequencies in bacterial populations. Accessory genes on plasmids are often beneficial in transient or local conditions [50,51] and could be repeatedly lost when they are not selected for. Plasmid-free segregants occur regularly and will rapidly invade populations when plasmids are costly. Other factors like the presence of bacteriophages can also lead to unstable dynamics, increasing plasmid loss [52]. Moreover, transfer is strongly regulated as a function of environmental conditions [8] and could be induced specifically in the conditions where plasmid-bearing cells are favoured. A striking case of such a scenario are mobile elements providing tetracycline resistance, whose transfer is induced by subinhibitory concentrations of tetracycline [53]: transfer occurs in conditions where mobile elements are likely to increase host fitness in the near future, as indicated by antibiotic gradients. Regulation of plasmid transfer will also modify the cost of transfer to the host. In our model, we assumed that plasmid-bearing cells experience a constitutive cost proportional to donor ability, leading to a higher cost to donor genotypes when plasmids are abundant. Transfer can be repressed when plasmid-free cells are likely to be rare [46], leading to the expression of transfer genes only when transfer efficiency is maximised. Finally, on a wider scale, cell migration between populations that experience different selection pressures for plasmid traits strongly increases the potential for horizontal transfer, as the immigration of plasmid-free cells in populations where plasmid traits are beneficial prevents plasmid fixation and allows sustained transfer [54].
Transferring plasmids increases the donor allele inclusive fitness because it enriches cells of the same kind with beneficial alleles. This phenomenon can be compared to the evolution of teaching in animals: teaching of adaptive information can be selected when teachers and pupils are related [55]. The difference between genetic information transfer in bacteria and cultural transmission is that beneficial genes are by default also transmitted vertically (together with the donor allele), making transfer ineffective if they are already prevalent in the population. Thus in order to be selected, horizontal transfer needs to improve transmission to kind compared to vertical transmission. Indeed, horizontal transfer is selected mostly when initially only few cells bear plasmids (Fig 4 and S5C Fig), as in these conditions it allows a more rapid and efficient spread than vertical transfer.
Interestingly, the phenomenon of lethal zygosis suggested by the positive frequency-dependence observed in our synthetic system (S3 Fig) [41,42] could act on the selection of donor ability in a complementary way, by selecting for donor genotypes when plasmids are prevalent. Transfer would then be a spiteful behaviour, in this case, not because of the indirect effects of transferred genes but due to the direct damage to recipient cells.
Bacteria frequently encounter parasitic MGE decreasing fitness. Eq 3 suggests that the transfer of parasitic elements could be selected if it can be preferentially directed towards cells of another kind. The spread of parasites has been suggested to be a typical case of spiteful behaviour, since the donors may be immune to the negative effects of their parasites [56]. Bacteriophages are a well-known example, where phage lysogeny ensures that most cells of the initial strain are protected from lysis and phages preferentially lyse the cells of a competing strain, at the same time ensuring phage spread [57]. Similar mechanisms are not yet known for plasmids. However, transfer to unrelated cells is well described in the case of the Ti plasmid of Agrobacterium tumefasciens, where the T-DNA is transferred to plant cells [58], and specific transfer and gene expression ensure that another species produces resources. Even with no specific targeting, suboptimal effects of transferred genes could render the plasmid harmful, damaging specifically unrelated recipients and effectively leading to spite.
We conclude that the inclusive fitness benefits conferred by transferred plasmids can lead to indirect selection for host donor ability. Plasmid transfer rates thus can be shaped not only by their direct effects on plasmid fitness [7] but also by their indirect effects on host fitness. The direction and strength of selection acting on donor ability will depend on the potential for plasmid transfer, its bias towards kind, the fitness effects of plasmids present in the host population, and the costs of transfer. Thus, all these factors might, at least in part, determine both the strikingly large variability of transfer rates observed among bacterial isolates [11,12] and the existence of high donor ability strains.
Our findings have consequences in the context of the fight against the spread of antibiotic resistance, as the indirect selection of donor strains could promote widespread dissemination of antibiotic resistance. Treatments that decrease the spread of MGEs have already been considered, like male-specific phages that inhibit plasmid transfer but also kill preferentially the cells that actively transfer plasmids [59]. Our work suggests that, the same as for other cooperative behaviours [60], bacteria resistant to such treatments may evolve, but relatively slowly [61], which should be taken into account when aiming to diminish horizontal transfer [62].
More generally, our results underline the active role hosts may play in the evolution of transfer rates and the necessity to take bacterial social interactions into account when studying plasmid transfer. Plasmids themselves often bear public good genes involved in host sociality and interaction with neighbouring cells [5,20], and plasmid transfer promotes host public good production by modifying relatedness in structured populations [6]. The indirect benefits of added public good production may in turn further favour the hosts that are investing in transfer.
Finally, we show that biased transfer in structured populations combined with selection of plasmid-bearing cells promotes association between hosts with high donor ability, discrimination mechanisms (Fig 5), and beneficial plasmids (Fig 6). Donor ability can be selected in the absence of initial linkage with discrimination alleles or mutualistic plasmids, but selection itself creates linkage at the population level. This dynamic will alleviate the cost of parasitic plasmids and lead to a prevalence of donor strains associated with mobile, transiently beneficial plasmids. In the long term, the phenomenon could promote mutualistic coevolution between beneficial plasmids and strains that transmit them at high rates to their kind, in a way analogous to the evolution of mutualism between species. The benefits generated by mutualism can create an association between mutualistic partners [63], while the association itself favours further mutualism [64,65]. Plasmids would be a special case of mutualism, with a complex and important role of horizontal transmission, a mechanism that is generally expected to inhibit mutualism [66] but here actually benefits both partners. Social selection promotes host investment in plasmid transfer, increasing plasmid fitness but simultaneously promoting host association to mutualist plasmids. This will likely lead to complex social selective pressures acting on plasmids themselves and shape the mobile gene pool.
To test for discrimination, the better plasmid donor was the Escherichia coli K12 strain MG1655red, which is MG1655 [35] marked with the td−Cherry gene. The worse donor was the E. coli B strain REL606 [36]. To measure conjugation rates, two spontaneous mutants resistant to rifampicin (RifR) for each strain MG1655 and REL606 were used as recipients. The plasmid used was the multiresistant R1-19 plasmid (that provides resistance to Cm, sulfonamides, ampicillin, Kn, streptomycin, and spectinomycin) [37]. The K12ΔarcA strain was MG1655red transduced with the Keio collection arcA deletion mutant [67].
To test for transfer selection in structured populations, we used two synthetic strains, D− and D+, and two associated plasmids, C and N. D− strain is E. coli K12 MG1655. D+ strain is a derivative of MG1655 marked with the td−Cherry gene and bearing the helper plasmid FHR. FHR is a variant of the F plasmid with low self-transfer and entry exclusion [6], which provides efficient mobilization of plasmids carrying F oriT sequence. N plasmids bear F oriT and an aph gene providing Kn resistance, while C plasmids additionally carry a cat gene providing Cm resistance. N and C plasmids express either YFP or GFP under control of the strong promoter PR. For selection experiments (Fig 3), D− strain initially bears C-GFP plasmid, and D+ strain initially bears C-YFP plasmid, in order to identify the origin of C plasmid (see S2A Fig). For linkage experiments (Fig 6), C-YFP and N-GFP plasmids were used respectively as C and N plasmids (see S2B Fig).
FHR and N-GFP plasmids and their construction are described in detail in [6] (where N-GFP was called the T+P− plasmid). N-YFP was constructed by amplification of YFP sequence with primers AGCGACTCGAGGATAAATATCTAACACCGTGCGTGTTGAC and AGCACAAGCTTTTCCCGGGTCATTATTTGTATAG, then ligation of N-GFP plasmid and the PCR product after digestion with XhoI and HindIII. To construct C plasmids, the cat gene was amplified from pKD3 plasmid [68] with primers TACTAAGACGTCAGGAACTTCATTTAAATGGCG and TACTAGCTCGAGAAGAGGTTCCAACTTTCACC. The PCR product was ligated into the corresponding (GFP or YFP) N plasmid after digestion with AatII and XhoI.
The D+ and MG1655red strains were constructed by integration of the pRNA1-tdCherry gene construction on pNDL32 plasmid obtained from Nathan Lord (Paulsson laboratory, Harvard Medical School). pNDL32 was transformed into MG1655 with selection on 100 μg/mL ampicillin, then streaked twice at 30°C on LB-agar (Luria-Bertani, BD Difco). Colonies were streaked overnight on LB-agar at 42°C, and plasmid loss was confirmed by checking that clones were ampicillin-sensitive. FHR was finally added to D+ strain by conjugation. To construct K12ΔarcA, the Keio collection arcA deletion mutant was used for P1 transduction of MG1655 red, then the kan resistance cassette was removed with pCP20 plasmid [68].
Spontaneous RifR mutants of MG1655red and REL606 were obtained by plating overnight cultures on LB-agar with Rif (Sigma-Aldrich) at 100 μg/mL.
Experiments were conducted under well-mixed conditions with 5 mL medium in 50 mL tubes (Sarstedt). Exponential growth rates (S1 Fig and S2C Fig) were measured in a Tecan Infinite M200 reader on 100 μL cultures with 50 μL mineral oil (Sigma) in 96-well plates, after 100-fold dilution from stationary phase cultures.
Our simulations mimic the experimental conditions of strain growth and plasmid transfer in the same way as described in our previous work [6]. Plasmid transfer follows a mass-action law: the number of transfer events is proportional to both donor and recipient cell densities in the local population. The probability coefficient is the transfer rate constant γ (mL.cell-1.h-1). Strains are characterized by their donor ability q that modulates effective transfer and leads to a proportional cost of donor ability for the donor cell cq. Similarly to our experiments, we model two steps: a transfer phase (from t0 to t1), then a selection phase, in conditions where the plasmid genes affect growth (from t1 to t2). The length of the transfer phase is set to 12 h after 100-fold initial dilution from carrying capacity, and growth for the selection phase is allowed for 36 h after a second 100-fold dilution. Equations governing changes in cell densities, presented below, are common to the two steps. Ntot is the total cell density.
|
10.1371/journal.ppat.1007405 | Polymicrobial sepsis influences NK-cell-mediated immunity by diminishing NK-cell-intrinsic receptor-mediated effector responses to viral ligands or infections | The sepsis-induced cytokine storm leads to severe lymphopenia and reduced effector capacity of remaining/surviving cells. This results in a prolonged state of immunoparalysis, that contributes to enhanced morbidity/mortality of sepsis survivors upon secondary infection. The impact of sepsis on several lymphoid subsets has been characterized, yet its impact on NK-cells remains underappreciated–despite their critical role in controlling infection(s). Here, we observed numerical loss of NK-cells in multiple tissues after cecal-ligation-and-puncture (CLP)-induced sepsis. To elucidate the sepsis-induced lesions in surviving NK-cells, transcriptional profiles were evaluated and indicated changes consistent with impaired effector functionality. A corresponding deficit in NK-cell capacity to produce effector molecules following secondary infection and/or cytokine stimulation (IL-12,IL-18) further suggested a sepsis-induced NK-cell intrinsic impairment. To specifically probe NK-cell receptor-mediated function, the activating Ly49H receptor, that recognizes the murine cytomegalovirus (MCMV) m157 protein, served as a model receptor. Although relative expression of Ly49H receptor did not change, the number of Ly49H+ NK-cells in CLP hosts was reduced leading to impaired in vivo cytotoxicity and the capacity of NK-cells (on per-cell basis) to perform Ly49H-mediated degranulation, killing, and effector molecule production in vitro was also severely reduced. Mechanistically, Ly49H adaptor protein (DAP12) activation and clustering, assessed by TIRF microscopy, was compromised. This was further associated with diminished AKT phosphorylation and capacity to flux calcium following receptor stimulation. Importantly, DAP12 overexpression in NK-cells restored Ly49H/D receptors-mediated effector functions in CLP hosts. Finally, as a consequence of sepsis-dependent numerical and functional lesions in Ly49H+ NK-cells, host capacity to control MCMV infection was significantly impaired. Importantly, IL-2 complex (IL-2c) therapy after CLP improved numbers but not a function of NK-cells leading to enhanced immunity to MCMV challenge. Thus, the sepsis-induced immunoparalysis state includes numerical and NK-cell-intrinsic functional impairments, an instructive notion for future studies aimed in restoring NK-cell immunity in sepsis survivors.
| Sepsis is an exaggerated host response to infection that can initially lead to significant morbidity/mortality and a long-lasting state of immunoparalysis in sepsis survivors. Sepsis-induced immunoparalysis functionally impairs numerous lymphocyte populations, including NK-cells. However, the scope and underlying mechanisms of NK-cell impairment and the consequences for NK-cell-mediated pathogen control remain underappreciated. NK-cells contribute to early host control of pathogens through a balance of activating and inhibitory receptors, and alterations in the number and capacity of NK-cells to exert receptor-mediated immunity can lead to dramatic impairment in host control of infection. The present study defines sepsis-induced numerical and cell-intrinsic functional impairments in NK-cell response to cytokine stimulation and receptor signaling that contribute to impaired host capacity to mount NK-cell-mediated effector responses and provide protection to bacterial and/or viral pathogens. Impairments in receptor signaling were due to reduced expression of adaptor protein DAP12. Importantly, the diminished ability of NK-cells from CLP hosts to provide anti-viral (MCMV) immunity is partially restored by IL-2 complex (IL-2c) therapy, which increased the number, but not function, of protective Ly49H+ NK-cells. Thus, these findings define sepsis-induced changes of the NK-cell compartment and provide insight into potential therapeutic interventions aimed at resolving sepsis-induced immunoparalysis in sepsis survivors.
| Sepsis is fatal to approximately 250,000 Americans every year and presents a significant economic burden (>$20 billion annually). The cytokine storm, which characterizes a septic event, is the result of a mismanaged infection and is composed of both pro- and anti-inflammatory cytokines [1]. However, the ~75% of patients that survive the cytokine storm can enter a state of chronic immunoparalysis associated with increased susceptibility to unrelated secondary infection, increased viral reactivation, and decreased 5-year survival compared with control cohorts [2–6]. Additionally, the sepsis-induced cytokine storm is associated with apoptosis of lymphocytes leading to severe and transient lymphopenia [7–11].
While the impact of sepsis on several lymphoid populations has been explored, including work from our labs examining the effect of sepsis on CD4 and CD8 T cell responses, the influence of sepsis on the NK-cell compartment remains understudied [9, 12–20]. The predominant focus of research to date has been on the contribution of NK-cells to sepsis severity, due to NK-dependent release of cytokines during the cytokine storm [21–30]. As such, NK-cells are largely considered detrimental in the context of sepsis, however, NK-cells are also important early mediators in the control of infection. Thus, sepsis-induced impairment in NK-cell function(s) may contribute to the increased host susceptibility to unrelated infection(s). Indeed, the enhanced susceptibility of sepsis patients to secondary infection(s) and viral re-infections indicate potential long-term impairment of NK-cells after sepsis. Additionally, there is some indication that sepsis leads to apoptosis and functional impairment of NK-cells [31–38]. However, robust characterization of these lesions, the underlying mechanisms of the sepsis-induced NK-cell dysfunction, and the direct consequences of sepsis-induced NK-cell impairment to host health remain to be elucidated.
NK-cells can be activated either by cytokine stimulation and/or receptor stimulation, and receptor expression and signaling is key to NK-cell receptor-mediated function. NK-cells recognize target cells through a balance of activating and inhibitory receptors [39–42], and when the balance is shifted in favor of activating receptors an immune synapse is formed to promote NK-cell effector function (e.g. cytokine release and cytolysis of the target cell). This process is contingent on multiple signaling events including activation of signaling cascades, incorporating events such as phosphorylation of AKT and calcium flux [42–44]. Consequently, even minor changes in these pathways can lead to cumulative downstream impairment [45]. Thus, sepsis-induced qualitative changes in NK-cell receptor-mediate function, compounding with the lymphopenic state, may lead to dramatic impairment in NK-cell-mediated control of pathogens.
Herein we describe sepsis-induced numerical loss and cell-intrinsic changes in NK-cells in response to cytokine and receptor signaling, associated with functional impairment and reduced pathogen control. While sepsis led to numerical loss, it did not alter the subset composition of surviving NK-cells. However, alterations in NK-cell-mediated cytotoxicity gene transcription and impaired receptor signaling were observed following sepsis induction. These cell-intrinsic changes were associated with a per-cell functional impairment in response to both cytokine stimulation and specific receptor stimulation. As a consequence, mice had reduced NK-cell-mediated pathogen control that could be improved by therapeutic administration of interleukin-2 complexes (IL-2c). This improved pathogen control was a result of an increased number, but not inherent function, of NK-cells following IL-2c therapy.
NK-cell-mediated protection against infection is contingent on the number of NK-cells capable of recognizing the presence of pathogen at the time of infection [46]. Sepsis leads to a loss of several lymphocyte populations, including NK-cells, in a variety of tissues [8, 47, 48]. To further examine the impact of sepsis on the NK-cell compartment, inbred C57Bl/6 (B6) mice underwent Sham (control) or cecal-ligation and puncture (CLP) surgery. At early stages of the sepsis-induced immunoparalysis state, day 2 post-CLP surgery (at the time when cytokine storm is diminished/absent [17, 49]), spleens and livers were harvested (Fig 1A). NK-cells were identified as NK1.1+ (activating receptor expressed early in NK-cell development) and CD3-, to exclude natural killer T lymphocytes (NKT) (Fig 1B) [50, 51]. Consistent with existing data, total lymphocyte and NK-cell numbers were dramatically diminished in both the spleen (Fig 1C and 1D) and liver (Fig 1E and 1F) from CLP-treated mice compared to Sham counterparts [8, 47]. However, no alteration in the frequency of NK-cells in either of these tissues was observed.
This decline in NK-cell numbers during the early immunoparalysis state suggests that CLP leads to apoptosis of NK-cells. Alternatively, as a result of initial sepsis-induced inflammation, NK-cells can follow inflammatory cues and be represented in substantially higher numbers at locations proximal to the initial insult. To address these possibilities NK-cell numbers were determined in proximal (e.g. mesenteric lymph node (mLN), peritoneal lavage (PL)) and/or distal sites (e.g. inguinal lymph nodes (iLN)) 24 hrs post-sepsis induction. A numerical loss of NK-cells was observed across all tissues analyzed, suggesting that redistribution, including to the abdominal cavity, does not account for overall decline in numbers (Fig 1G). To address whether apoptosis contributed to the loss of NK-cells, spleen-derived NK-cells were evaluated for expression of activated caspase (FLICA) and the loss of membrane integrity (propidium iodide [PI]) at 12, 24, and 48 hrs after sepsis induction. Importantly, a significant increase in the frequency of apoptotic (FLICA+PI+) NK-cells was observed in septic hosts 12 hrs after sepsis induction, which resolved by 24 and 48 hrs (Fig 1H) suggesting that apoptosis of NK-cells occurs early and is transient. To further explore the timing of numerical decline (apoptosis) of NK-cells after sepsis induction, splenocytes from Sham and CLP hosts at 48 hrs post-surgery were harvested and placed in culture for an additional 48 hrs (Fig 1J). Importantly, the number of NK-cells recovered over 48 hrs was indistinguishable between the two groups suggesting that potential ‘intrinsic’ differences in NK-cell compartment could not be attributed to increased frequency of dying NK-cells (Fig 1H–1J).
Multiple cell death pathways, including receptor- and mitochondria-mediated cell-death pathways, contribute to sepsis-induced apoptosis and lymphopenia [52, 53]. Yet fratricide has been shown to contribute to NK-cell loss following infection [54]. To address whether fratricide was also a factor in the NK-cell loss during sepsis, splenocytes were transferred into Thy1 disparate WT and perforin knockout (Prf-/-) mice (S1A Fig). Thy1.1+ (transferred) NK-cells were identified in the spleen 2 days after sepsis induction and the fold loss was calculated for both WT and Prf-/- mice. We did not observe a difference in fold loss of NK-cells between WT and Prf-/- mice (S1B–S1D Fig), suggesting fratricide did not contribute for the loss of NK-cells during sepsis.
While inbred mouse strains, such as B6, are valuable for in-depth analyses of NK-cells with well-defined activating and inhibitory receptor repertoires, the genetic homogeny within individual strains does not reflect the true genetic diversity observed in human population [55]. To compensate for this lack of genetic diversity we have previously utilized outbred Swiss Webster (SW) mice to evaluate immunologic responses in genetically heterogeneous populations [17, 19, 56, 57]. To determine the extent to which the loss of NK-cells following sepsis was recapitulated in genetically heterogeneous population, we performed Sham or CLP surgery on outbred SW mice (S2A Fig). Similar to data from inbred B6 mice, a decline in total NK-cell (here defined as NKp46+CD3-(S2B Fig)) numbers in both the spleen (S2C Fig) and liver (S2D Fig) was observed. These data suggest sepsis-induced numerical loss of NK-cells represents a global phenomenon not restricted to particular inbred strain of mice.
Sepsis could lead to numerical decline of NK-cells and changes in effector capability of NK-cells [31–38]. To address how sepsis alters the functional capacity of NK-cells we evaluated the NK-cell transcriptome after sepsis (via RNA-sequencing; RNA-seq.) coupled with gene-set enrichment analysis (GSEA) (Figs 2 and 3). GSEA uses enrichment scores to compare the enrichment of genes in a ranked list. For this analysis RNA was isolated and sequenced from sorted splenic NK1.1+CD3- NK-cells derived from Sham and CLP mice at days 1 or 2 post-surgery (Fig 2A). We detected 441 and 287 genes that are differentially expressed in NK-cells (fold change ≥ 1.5, p<0.05) between day 2 Sham and day 1 and 2 CLP NK-cells, respectively. Additionally, we noted 169 differentially expressed genes between day 1 and day 2 CLP NK-cells, demonstrating an evolving transcriptional profile as the cytokine storm develops and resolves. Fig 2B shows a heat map of genes with significantly different expression between NK-cells; gene changes are enumerated in Fig 2C. Principal component analysis of gene expression subsequently identified distinct clustering of sampled groups, indicating time-dependent changes in NK-cell gene expression (Fig 2D).
GSEA was performed comparing day 2 Sham and CLP samples to address functional changes that manifest at the beginning of the immunoparalysis phase (post-cytokine storm). We focused on enriched pathways that are closely associated with NK-cell effector functions. Interestingly, CLP samples demonstrated negative enrichment for genes associated with NK-cell-mediated cytotoxicity and calcium signaling, an important component of receptor signaling for target cell recognition (Fig 3A). This analysis revealed impairments in NK-cell receptor signaling may be associated with the impaired effector capacity of NK-cells following sepsis. Expression of genes in these pathways were evaluated to assess how sepsis alters NK-cell function. As expected we observed core enrichment of genes reduced in expression after sepsis (Fig 3B and 3C). We next evaluated the expression of the core enriched genes to define potential lesions. While we did not observe changes in NK-cell receptor expression among core enriched genes, many of the modulated genes were associated with receptor signaling (e.g. Tyrobp, Lat, Plcd1) or effector molecules (e.g. Ifng, Tnf, Prf1) (Fig 3D and 3E). Notably, expression changes observed in receptor signaling primarily encode proteins that participate in early activating receptor signaling cascades [58–60]. Thus, they are likely to impact a wide range of receptors even if expression of the receptor itself is unaltered. These changes represent a potential cell-intrinsic perturbation in receptor signaling and cumulatively suggest a potential sepsis-induced impairment in NK-cell effector functionality.
To assess to what extent NK-cells (analyzed on the population level) in the post-sepsis environment exhibit functional impairment, as suggested by the GSEA, mice were infected with virulent Listeria monocytogenes (L.m.) 2 days post-sepsis induction. While L.m. is not an infection commonly associated with sepsis clinically, it can be used to directly probe NK-cell capacity to produce the effector cytokine IFN-γ [61, 62]. Similar to data shown before, total numbers of NK-cells were diminished in spleens and livers of CLP hosts one day after L.m. infection (Fig 4A–4E). Importantly, the frequency of IFN-γ producing NK-cells directly ex vivo in response to secondary infection was also reduced in both tissues of CLP hosts (Fig 4F). The reduced number of NK-cells and the reduced frequency of IFN-γ producing NK-cells results in a dramatic reduction (14 and 17-fold in the spleen and liver, respectively) in the number of IFN-γ producing NK-cells (Fig 4G–4J). Thus, these data are consistent with GSEA analysis suggesting sepsis has the capacity to influence effector functionality of the remaining NK-cells in response to secondary bacterial infection.
However, sepsis induces perturbation in the dendritic cells (DCs) and other cellular compartments could also influence pathogen-induced cytokine secretion in vivo (ex. DC derived IL-12 [19]) necessary to facilitate IFN-γ production by NK-cells. To determine the extent to which intrinsic defects contributed to the impaired production of IFN-γ by NK-cells during L.m. infection, NK-cells obtained from Sham and CLP hosts 48 hours after surgery were stimulated with IL-12 and IL-18 directly ex vivo to bypass sepsis-induced changes in endogenous levels of stimulatory cytokines [19] (Fig 4K). Importantly, NK-cells from CLP hosts exhibited impairment in IFN-γ production following cytokine stimulation compared to Sham counterparts (Fig 4L and 4M), which corresponded with the reduced expression of the Il12rb1 gene identified in the RNA-seq (1.3-fold reduction in expression; p<0.05). Therefore, the data in Fig 4 collectively show sepsis reduces capacity of NK-cell compartment to produce effector cytokines in response to infection in vivo and/or cytokine stimulation directly ex vivo.
To precisely define the extent to which sepsis affects NK-cell functional capacity we switched to a model system in which we could probe a single NK-cell receptor for its capacity to mediate effector functions and protect against infection. Ly49H, an activating receptor expressed exclusively by subpopulations of NK-cells, has no endogenous ligand and recognizes the m157 immunoevasion protein of murine cytomegalovirus (MCMV) [63, 64]. Additionally, Ly49H+ NK-cells are critical in the control of MCMV and expression of the receptor confers resistance to MCMV by mouse strains (such as B6) [65–67]. Thus, high ligand specificity and importance in host immunity to infection make Ly49H+ NK-cells an ideal population to mechanistically examine the impact of sepsis on NK-cell receptor-mediated immunity.
To determine the extent to which sepsis affects the frequency Ly49H+ NK-cells and the level of Ly49H expression, NK-cells from the spleens and livers were evaluated 2 days after surgery (Fig 5A). Although sepsis did not change the frequency and/or relative expression of Ly49H on NK-cells (Fig 5B–5D), numerical loss of the Ly49H+ NK-cells was observed in both the spleens (Fig 5E) and livers (Fig 5F) suggesting similar susceptibility of Ly49H+ NK-cells to sepsis-induced apoptosis. Additionally, the maturation status of NK-cells determines their function [68, 69]. To address whether sepsis altered the maturation of surviving NK-cells, subset composition of Ly49H+ NK-cells was determined by CD27 and CD11b expression [68]. Additionally, KLRG1 and Ly6C were used as markers of NK-cell terminal maturation [70–72]. As predicted, numerical loss of all subsets was observed in the spleen and liver; however, no changes in the composition of NK subpopulations was detected (S3A–S3G Fig). Finally, inflammation can alter NK-cell expression of effector molecules [73]. Thus, we sought to address the extent to which sepsis altered the steady state expression of Granzyme B (GzmB) in NK-cells. Baseline GzmB expression of NK-cells was not impacted by sepsis (S3H–S3J Fig). Therefore, sepsis induces the loss of Ly49H+ NK-cells without changing the maturation status and/or levels of Ly49H receptor expression on NK-cells.
To define if sepsis influences the killing capacity of Ly49H+ NK-cells, wild type (WT) and m157-expressing splenocytes (m157-Tg) were used as targets in an in vivo cytotoxic assay [66, 74]. Two days post-sepsis induction target cells, CFSE-labelled m157-Tg (CFSElo) and WT (CFSEhi) splenocytes were injected in a 1:1 mixture i.v. into Sham- and CLP-treated mice. Additional groups of mice, in which the NK-cell compartment (αNK1.1 groups) was depleted prior to the sepsis induction, were included as necessary controls (Fig 5G). Results in Fig 5H and 5I clearly indicate impaired specific lysis of m157 target cells in the spleens of CLP hosts compared to Sham controls. Importantly, the specific lysis of m157 target cells strongly correlated with the number of Ly49H+ NK-cells present in the spleen (Fig 5J). Thus, these data suggest the sepsis-imposed numerical loss of Ly49H+ NK-cells resulted in impaired target cell killing in vivo.
While the reduced number of Ly49H+ NK-cells resulted in a deficit in CLP host capacity to kill target cells, the RNA-seq. data and reduced IFN-γ production upon cytokine stimulation also indicated NK-cells may exhibit sepsis-induced functional impairments on a per-cell basis. To assess this possibility, splenocytes from Sham and CLP mice were incubated I) in vitro with either m157-Tg or WT cells separately to assess degranulation, II) with the target cell mix of m157-Tg and WT splenocytes described above to assess killing, or III) plate bound stimulatory antibodies (or control IgG) to assess effector molecule production (Fig 6A). Of note: since sepsis alters the number but not the frequency of both total and Ly49H+ NK-cells, the capacity of Ly49H expressing NK-cells to perform effector function(s) (e.g. degranulation, target cell lysis and cytokine production) on per-cell basis was determined by plating equivalent numbers of either Sham or CLP splenocytes per well.
Relative to Sham, Ly49H+ NK-cells from CLP hosts exhibited less degranulation (as determined by analyzing CD107a expression) [66, 75] in response to Ly49H stimulation after 6 hrs incubation in vitro (Fig 6B). Impaired degranulation can subsequently impair the capability to kill target cells [75]. To determine the extent to which sepsis impairs the intrinsic capacity of Ly49H+ NK-cells to perform cytolysis splenocytes from Sham or CLP hosts were incubated with the m157-Tg: WT target mix (effector to target ratio 0.5: 1) for 18 hrs, after which m157 specific lysis was assessed. While the number of Ly49H+ NK-cells per well was equivalent (Fig 6C) there was significantly less specific lysis of m157 target cells by CLP splenocytes compared to Sham splenocytes (Fig 6D). Target cells were separated for the degranulation assay to avoid ‘cold target’ competition; however, the killing assay required the presence of both target cell populations in the same well. Finally, to examine if sepsis modulated the ability of subsets of NK-cells to produce effector cytokines (e.g. IFN-γ), Sham or CLP-derived splenocytes were stimulated with plate bound stimulatory antibodies to different activating NK-cell receptors (NK1.1, NKp46, Ly49D, or Ly49H) or control IgG antibodies [73], and we assessed IFN-γ production and upregulation of GzmB for cells expressing the receptor stimulated with its agonistic/specific Ab compared to IgG control. Importantly, a defect in the capacity to produce IFN-γ and/or upregulate GzmB in response to receptor stimulation was most prominent in Ly49H and Ly49D expressing NK-cells (Fig 6E–6H). This suggests that sepsis-induced receptor impairment is potentially a result of a defect(s) shared by Ly49H and Ly49D but not NK1.1 and NKp46, a notion that will be explored further. In summary, these data collectively suggest that sepsis changes the per-cell capacity of NK-cells to respond to precisely defined viral ligands or specific receptor stimulation.
Although sepsis did not alter Ly49H expression, the diminished capacity of Ly49H NK-cells to exert effector functions upon receptor ligation suggests potential lesions in receptor signaling. One of the genes indicated by RNA-seq. analysis as having reduced expression after sepsis was Tyrobp. This gene encodes DAP12, the adaptor protein for Ly49H (and Ly49D but not NKp46 or NK1.1) and is required for Ly49H expression and signaling [60, 66, 76, 77]. We thus hypothesized that reduced DAP12 expression contributed to impaired Ly49H signaling and subsequently impairing effector function on a per-cell basis. To address whether DAP12 expression was impacted by sepsis, NK-cells were enriched by negative selection (~85% NK-cell purity, of which ~70% were Ly49H+) to minimize signaling events that might occur as a result of antibody-receptor interaction. The sorted cells were then lysed and assessed for expression of DAP12 by immunoblotting using monoclonal αDAP12 antibody (Fig 7A). Dramatic reduction in the expression of DAP12 by NK-cells from CLP hosts was observed (Fig 7B and 7C). Thus, sepsis impairs expression of Ly49H/D adaptor protein DAP12 in NK-cells.
Because of the proximity of the lesion to the receptor it is likely the clustering of the receptor with its adaptor protein necessary for proper immune synapse formation may be impaired [78]. To address this possibility, NK-cells were again negatively sorted before adherence to an α-Ly49H mAb coated slide. Cells were allowed to interact with the antibody for 15 minutes before being fixed and stained with αDAP12. Slides were then examined by total internal reflection fluorescence (TIRF) microscopy, which allows for imaging of the plasma membrane at the interface with the glass slide. This allows evaluation of microcluster formation, as assessed by DAP12 density, at the site of receptor activation (Fig 7D–7F). Of important note, enriched NK-cells obtained after CLP surgery showed a statistically significant impairment in their ability to adhere to the slide. This is a receptor-mediated event, thus non-adherent cells that are potentially the most influenced by sepsis are, by necessity, excluded from TIRF analysis (Fig 7D). To further corroborate this, the reduced capacity of NK-cells obtained from CLP hosts to adhere was observed in response to various concentrations of αLy49H Abs (S4A and S4B Fig). Thus, the reduced expression of DAP12, coupled with reduced microcluster formation, contributes to the diminished Ly49H-mediated effector responses.
Impaired DAP12 expression should also alter subsequent signaling events, as even relatively minor changes in upstream signaling can lead to dramatic differences in functional outcome [45]. While there are many signaling events that follow Ly49H receptor stimulation, AKT phosphorylation and calcium flux in the NK-cells from CLP hosts were further examined because of their relative proximity to receptor stimulation and the negative enrichment of calcium signaling associated genes identified in the RNA-seq. analysis [79].
To determine how sepsis alters Ly49H receptor capacity to induce AKT phosphorylation splenocytes were harvested and labeled with α-Ly49H mAb. Addition of cross-linking beads was used to stimulate the cells for 2 hrs, while no bead addition was used for unstimulated controls (Fig 8A). To ensure the results were not affected by a sepsis-induced change in total AKT, the amount of AKT present in stimulated and unstimulated Sham and CLP samples was evaluated and determined to be equivalent (Fig 8B and 8D). Unstimulated Sham and CLP cells did not have detectable pAKT(pS473, activating phosphorylation of AKT [79]), compared to FMO (fluorescence-minus-one) controls. However, upon stimulation pAKT could be detected in Sham cells but not in CLP cells (Fig 8C and 8E). These data indicate that while sepsis does not alter AKT expression it does alter the capacity of Ly49H to induce activating phosphorylation of AKT following receptor stimulation.
To evaluate how sepsis affects calcium signaling in NK-cells following specific receptor stimulation, splenocytes were harvested and stained to identify NK-cells (NK1.1+CD3-) before being labeled with a calcium sensitive dye (Fluo-4-AM) to detect free intracellular calcium (fluorescence correlates with calcium concentration) (Fig 8A) [80]. A baseline reading was taken to determine calcium concentration in the absence of stimulation. Intriguingly, NK-cells from CLP hosts had higher internal calcium at baseline indicating a potential dysregulation in calcium sequestration (Fig 8F and 8G). It is important to note that in the context of calcium signaling during receptor stimulation, the most important factor is how much calcium fluxes in response to receptor stimulation rather than the total amount of calcium present in the cell [77, 80]. Therefore, baseline readings were followed by stimulation with the addition of a fluorescently labeled α-Ly49H mAb. This had the additional benefit of distinguishing stimulated (Ly49H+) from unstimulated (Ly49H-) NK-cells for an effective internal control. We found that while Sham Ly49H+ NK-cells fluxed calcium efficiently (relative to Ly49H- NK-cells), CLP Ly49H+ NK-cells did not (Fig 8F and 8G). This was further quantified using area under the curve (AUC) of Ly49H+ or Ly49H- NK-cell populations as a surrogate measurement of internal calcium flux during the stimulation period (Fig 8H). Finally, results were validated by stimulating with Ionomycin (Iono) to determine that the peak amount of calcium present in all cells was similar suggesting that differences were the result of different capacity to flux calcium rather than different total amounts of calcium (Fig 8F and 8G). Thus, these results suggest sepsis impairs the ability of NK-cells to respond to viral ligands due to reduced expression of the receptor’s adaptor protein leading to impairments in downstream signaling, including fluxing calcium and phosphorylating AKT, and reducing formation of the immunological synapse.
The previous results indicate that a loss of DAP12 is associated with the reduced functionality of Ly49H and Ly49D receptors. Therefore, to establish whether DAP12 loss is causal in impaired function of these receptors bone marrow chimeras (BM) were generated for control and DAP12 overexpression (DAP12-OE) vectors. Vectors were transfected into LSKs used for BM reconstitution and transfection was demarcated by expression of mCherry (Fig 9A and 9B). We next sought to address whether this rescued NK-cell DAP12-dependent receptor-mediated IFN-γ production of CLP hosts. Similar to previous results we observed reduced functionality of both Ly49H and Ly49D in cells from CLP Control hosts, relative to Sham counterparts. However, functionality was completely rescued by DAP12-OE in CLP hosts (Fig 9C and 9D). Thus, sepsis-induced loss of DAP12 expression is causal in impairment in DAP12-dependent NK-cell-receptor functionality.
The results thus far indicate proximal impairments in receptor signaling. To address whether downstream signaling impairments also exist, splenocytes were stimulated with phorbyl 12-myristate 13-acetate (PMA) and Iono to mimic DAG and calcium signaling (S5A Fig). PMA/Iono stimulation represents a very potent stimulus that bypasses receptor signaling. Yet even with this strong stimulation condition, both total (S5B and S5C Fig) and Ly49H+ (S5B and S5D Fig) NK-cells from CLP-treated mice demonstrated a significant reduction in IFN-γ production. These data suggest that in addition to the proximal impairment in DAP12 signaling, distal signaling events may also be impacted.
The numerical and cell-intrinsic impairments described thus far suggest pathogen control by NK-cells would be severely impaired in CLP hosts. To address this directly, MCMV was chosen as a model pathogen since Ly49H+ NK-cells are critical for controlling MCMV infection in B6 mice [65–67]. To test the role of NK-cells in control of MCMV in the post-sepsis environment, mice were treated with control IgG or α-NK1.1 mAb to deplete the NK-cell compartment 2 days prior to Sham or CLP surgery. Mice were subsequently infected with the Smith strain of MCMV (105 PFU i.p.) 2 days after surgery. Viral titers in the spleens and livers were analyzed 3 days after infection (5 days post-surgery; Fig 10A) [81]. Consistent with existing data, the contribution of NK-cells is critical in providing anti-MCMV immunity since NK-depleted control (Sham) groups of mice had 15–50 fold higher viral loads in spleens and livers compared to their IgG-treated counterparts (Fig 10B and 10C) [66]. As expected, sepsis diminished the ability of the host to respond to MCMV challenge, and NK sufficient (IgG-treated) CLP-treated mice had a significant increase in viral load in both organs examined compared to Sham controls. Importantly, the contribution of NK-cells in anti-viral immunity in vivo was significantly diminished in the post-sepsis environment since similar levels of infection were observed in NK deficient (α-NK1.1 mAb group) or sufficient (IgG group) CLP hosts (Fig 10B and 10C). These data highlight the requirement for NK-cells in controlling MCMV infection and pinpoint the dramatic effect sepsis has on the ability of NK-cells to exert their effector functions in vivo.
The impaired control of MCMV infection highlights a need to recover the NK-cell compartment (number and/or function) in the post-septic environment. Recent clinical trials have demonstrated promise in recovering lymphocyte numbers by administration of lymphoproliferative cytokines (e.g. IL-2/7/15) [82–87]. Therefore, the efficacy of IL-2/α-IL-2 mAb complexes (IL-2c) therapy in reversing sepsis induced impairment of NK-cell-mediated MCMV control was explored next. The α-IL-2 mAb S4B6 was used for complex formation because it does not result in the expansion of regulatory T cells (Treg) [88, 89]. Mice received either control IgG or IL-2c at 24 hrs post-surgery, the earliest time point post-sepsis-induction at which the frequency of apoptotic cells in the spleen is no longer elevated (Fig 1I). Ly49H+ NK-cells in the PBL were monitored prior to therapy administration, 3 days after therapy (D4 post-sepsis), and 6 days after therapy (D7 post-sepsis). Spleen and liver Ly49H+ NK-cells were determined 6 days after therapy (Fig 11A). We observed robust expansion of Ly49H+ NK-cells in the PBL 3 days after IL-2c treatment in both groups of mice (Fig 11B). The number of Ly49H+ was also elevated in spleen, liver and PBL of IL-2c treated CLP hosts, relative to IgG treated controls 6 days post-treatment (Fig 11C–11E). It is, however, noteworthy that the number of Ly49H+ NK-cells in the spleen of IL-2c treated CLP hosts is still not fully recovered, relative to Sham mice. Thus, IL-2c therapy leads to robust expansion of Ly49H+ NK-cells in multiple tissues that remains elevated up to 6 days after surgery.
The numerical increase in Ly49H+ NK-cells in CLP host treated with IL-2c was interesting, but it was unclear whether this therapeutic intervention with IL-2c merely resulted in the proliferation of a ‘broken’ population of cells. Promisingly, IL-2 has been described to both increase the number of NK-cells and enhance activating receptor function (such as Ly49H), in part by tuning calcium flux [43, 58, 73, 90, 91, 92]. To address whether IL-2c therapy improved the intrinsic functional capacity of NK-cells, Sham and CLP mice were again treated with control IgG or IL-2c at D1 post-surgery. NK-cell capacity to produce IFN-γ in response to Ly49H or Ly49D receptor stimulation was evaluated as well as expression of DAP12 by NK-cells 6 days after start of the therapy (Fig 12A). The capacity of NK-cells from IgG-treated CLP hosts to produce IFN-γ in response to receptor stimulation remained impaired, demonstrating longevity of the previously described lesion. Additionally, NK-cells, from IL-2c treated CLP hosts, to produce IFN-γ in response to receptor stimulation also remained impaired, indicating that IL-2c does not improve the intrinsic function of NK-cells (Fig 12B). To determine whether this functional impairment (despite the numerical restoration) was associated with altered expression of DAP12, splenic NK-cells were isolated from Sham and CLP hosts 6 days after either IgG or IL-2c therapy (7 days post-surgery). Importantly, the reduced DAP12 expression observed in NK-cells from IgG-treated CLP hosts was not markedly increased upon IL-2c treatment further suggesting a strong association between DAP12 expression and diminished receptor-mediated cytokine production (Fig 12C and 12D). Thus, these results indicate impaired NK-cell receptor-mediated function is maintained after CLP induction and treatments that increase NK-cell numbers do not necessarily improve the ‘per-cell capacity’ of NK-cells to function properly.
The results of Figs 11 and 12 indicate that IL-2c rescues the number but not the intrinsic function of Ly49H+ NK-cells. However, we wanted to determine whether this numerical increase alone was sufficient to improve pathogen control. To address this, the numbers of Ly49H+ NK-cells were evaluated in the PBL 6 days after IL-2c therapy, a time at which cells are undergoing contraction following the IL-2-induced proliferative burst (Fig 13A) [93]. Importantly, IL-2c therapy induced an increase in Ly49H+ NK-cell numbers in PBL of both Sham and CLP mice at the time of MCMV challenge (Smith; 105 PFU i.p) (Fig 13B and 13C). Viral titers were evaluated in the spleens and livers 3 days after infection (10 days post-surgery; Fig 13D and 13E). Viral titers in both the liver and spleens of Sham mice were low and indistinguishable indicating IL-2c therapy does not further improve control of MCMV in non-septic hosts (Fig 13D and 13E). IgG-treated CLP hosts had higher viral titers than their NK-sufficient Sham counterparts even 10 days post-surgery suggesting sepsis can induce long-term changes in NK-cells preventing them to function properly. Importantly, IL-2c-treated CLP mice showed statistically significant improvement in viral control (Fig 13D and 13E). Cumulatively, these data suggest that IL-2c therapy reduces sepsis-induced impairment of host NK-cell-mediated pathogen control potentially by improving the number but not the intrinsic function(s) of NK-cells.
Sepsis as systemic infection can lead to a cytokine storm causing tissue damage and/or death [1]. However, patient survival of the cytokine storm has gradually risen over the recent decades leading to a population of sepsis survivors [2, 3]. These sepsis survivors can, though, enter a state of chronic immunoparalysis characterized by increased susceptibility to pathogens and decreased long-term survival [4–6]. This immunoparalysis is defined by severe lymphopenia and reduced functionality of surviving lymphocytes [7–9, 31–41, 94]. Paradoxically, while NK-cells are beneficial during the early stages of infection, the contribution of NK-cells to the septic event, including cytokine release (e.g. IFN-γ), has largely framed these cells as detrimental in the context of sepsis [21–30]. Thus, their capacity to mediate early control of pathogens suggests sepsis-induced impairment of NK-cell function may also contribute to the immunoparalysis state of sepsis survivors. Indeed, their capacity to produce cytokine in response to infection and TLR stimulation is impaired, highlighting differential roles of NK-cells during the cytokine storm and sepsis-induced immunoparalysis [37, 38]. However, the scope of sepsis-induced lesions in NK-cells, the underlying cause of this functional loss, and its subsequent impact on host capacity to control infection had not been fully defined.
NK-cell response to target cells is dictated by a balance of activating and inhibitory receptor signaling [39–42]. Receptor expression and capacity to effectively signal will determine whether an immunologic synapse is formed with the target cell and effector function (e.g., cytokine production and cytolysis) is performed [78]. Thus, changes in the number of receptor expressing cells and/or their relative expression of the receptor or its capacity to signal can lead to changes in NK-cell capacity to mediate control of infections. Therefore, we investigated how sepsis impacts NK-cell receptor-mediated control of pathogens. Using Ly49H as a model receptor, we were able to effectively probe NK-cell-intrinsic impairments in receptor signaling, independent of receptor expression. This impairment, in conjunction with the sepsis-induced lymphopenic state led to a dramatic impairment in NK-cell-mediated pathogen control. Informatively, therapeutic administration of IL-2c led to the numerical increase in NK-cells and enhanced NK-cell-mediated pathogen control.
Our use of transcriptome analysis and GSEA provided a strong foundation for investigating how sepsis alters the function of NK-cells. The modulation in gene expression in NK-cells occurring within the first 2 days following sepsis induction was profound and served as the impetus for our interrogation of NK-cell effector function and Ly49H signaling. For example, the downregulation in genes associated with NK-cell cytotoxicity (e.g., Ifng, Tnf, and Prf1) correlated strongly with our data showing reduced frequencies of IFNγ+ NK-cells from septic hosts and reduced target cell killing. The reduced functionality connected with the revealed deficiencies in the expression of DAP12 and downstream signaling from the Ly49H receptor. Not surprisingly, the numerical and functional defects in NK-cells led to the marked reduction in NK-cell-specific protection against secondary pathogen challenge in the post-septic host. We chose MCMV as the model pathogen for testing the fitness of the NK-cell compartment, finding the impairment in NK-cell-mediated pathogen control lasted up to 10 days after the septic insult. It is unclear, however, how long this impairment in pathogen control lasts. The duration of the sepsis-induced immunoparalysis in preclinical and clinical setting remains difficult to reliably define and serves as a significant knowledge gap not only for NK-cells, but across other immune populations. This includes known impairments in dendritic cell (DC) cytokine production in the post-septic environment [19]. Loss of DC-derived cytokine may lead to NK-cell-extrinsic impairment in NK-cell function, similar to that observed in T cell function [19]. This combination of cell-intrinsic and extrinsic impairments is reminiscent of those observed in T cells [20].
The administration of exogenous cytokines can have a marked effect on the number and function of various immune cell populations in sepsis survivors, as demonstrated by the preclinical and clinical data touting the benefits of IL-2, IL-7, IL-15, and Flt3L (among others). Of all the common γc receptor binding cytokines, IL-2 is the only cytokine currently approved for medical use and most promising in terms of improving functional responses. IL-2 has been intensively studied and shown to lower the rate of opportunistic infections in AIDS patients [95] and reduce bacterial abscess after CLP surgery in mice [96]. IL-2, however, is highly toxic at clinically significant levels. IL-2 signaling via low-affinity receptors present in the endothelium promotes vascular leak syndrome (VLS) [97], leading to pulmonary edema, severe hypotension, and liver cell damage[98]. One way to avoid IL-2 toxicity (but maintain its potency) uses IL-2/α-IL-2 mAb complexes (IL-2c). IL-2 can signal via αβγc heterotrimers with high affinity, or βγc heterodimers with intermediate affinity. The action of IL-2 complexed to different α-IL-2 mAb, such as S4B6 IgG1 mAb, differs because these mAb can avoid interactions with low affinity receptors and bind a specific IL-2 receptor type. Thus, IL-2c (IL-2 bound by S4B6) expands effector cells (CD8 and CD4 T cells and NK-cells), and can lower pathogen burden in bacterial infections [99] without signs of VLS [100]. Thus, IL-2c therapy may serve as a short-term replacement in the absence of properly functioning DC (for example). This relates directly to the question of whether the supportive nature of the IL-2c therapy is sustained or wains with time. Understanding the duration of both the intrinsic/extrinsic impairment of NK-cells and therapeutic effect of IL-2c are important considerations that should be addressed in future experiments. While we observed that therapeutic IL-2c administration led to expansion of NK-cells and enhanced NK-cell-mediated pathogen control, it is clear that per-cell functional impairment still exists. This has direct relevance to recent assessment of therapeutic IL-7, a related cytokine, administration and indicates the need for additional therapeutic intervention beyond the numerical expansion that the study evaluated [82, 101].
Finally, a major consideration of the experiments performed and data presented is that, while intrinsic impairment in DAP12-dependent receptors, Ly49H and Ly49D, was observed it is not representative of all NK-cell receptors [55, 102]. Thus, it is important to address the impact of sepsis on other NK-cell receptors–for example, activating receptors such as NKG2D and NKG2C and inhibitory receptors such as NKG2A and PD-1 –in future experiments, in addition to NK1.1 and NKp46 assessed here. Evaluating other receptors, such as NKG2D, that can signal through DAP10, a related adaptor protein, may be of particular interest given that gene expression of DAP10 was not observed to be altered by sepsis [58]. Further determining how sepsis impacts the capacity of inhibitory receptors to limit activating receptors from signaling, by activating SHP-1 and SHP-2, may reveal additional nuances to impaired NK-cell signaling [103]. Sepsis increases the expression of PD-1 by NK-cells [104, 105]. This, and other similarly increased expression of inhibitory receptors, may contribute to shifting the receptor balance to an inhibitory state, in addition to the intrinsic impairment in activating receptor signaling, subsequently impairing NK-cell function. These data contribute to a growing body of literature on sepsis-induced lymphocyte intrinsic and extrinsic, which contribute to the immunoparalyzed state of sepsis survivors [20].
Experimental procedures using mice were approved by University of Iowa Animal Care and Use Committee under ACURF protocol numbers 7051102 and 6121915. The experiments performed followed Office of Laboratory Animal Welfare guidelines and PHS Policy on Humane Care and Use of Laboratory Animals. Cervical dislocation was used as euthanasia method of all experimental mice.
Inbred C57Bl/6 (B6; Thy1.2/1.2) and outbred Swiss Webster mice were purchased from the National Cancer Institute (Frederick, MD) and maintained in the animal facilities at the University of Iowa at the appropriate biosafety level. Perforin knockout mice (Prf-/-) were a generous gift from the Karandikar lab (Department of Pathology, University of Iowa). m157-transgenic (m157-Tg) mice (Thy1.2/Thy1.2) were obtained from the Tripathy lab at the Washington University (St. Louis) and were bred and maintained at the University of Iowa (Iowa City, IA). Recombinant virulent Listeria monocytogenes (L.m.; strain 1043S) was injected i.v. (104 CFU). For MCMV-Smith infection, mice were injected i.p. (105 PFU). MCMV viral titers were quantified using plaque assay on M2-10B4 cells, as previously described [81].
Bone marrow (BM) chimeric mice were generated as previously described [106]. Briefly, the Tyrobp gene was cloned into an overexpression plasmid (pMSCV-IRES-mCherry), gene insertion into plasmid was confirmed by sequencing. The Tyrobp-containing or control (ctrl) plasmid was then transfected into 293 cells before being infected with adenovirus to generate adenovirus constructs for BM progenitor transfection. Lin- (B220, Gr-1, Ter119, NK1.1, TCRγ/δ, CD11b, CD11c, CD4, CD8, CD3) BM was collected from B6.SJL (CD45.1) mice and transfected with either of the adenoviral constructs. (5000 mCherry+ LSK cells along with 0.2 million protector BM cells) transfected BM was then injected into each irradiated 6–8 wk old C57Bl/6 recipients (CD45.2). Mice received Uniprim diet for the first 2 wks following transplant before returning to normal chow. 8 wks after transplantation mice were used for experiments.
Peripheral blood (PBL) was collected by retro-orbital bleeding. Peritoneal lavage was gathered by injecting 1mL of cold RPMI into the peritoneal cavity. The abdomen was then gently massaged before removing the fluid from the peritoneal cavity. Single-cell suspensions from spleen, liver, and lymph nodes were generated after mashing tissue through 70 μm cell strainer without enzymatic digestion. Livers were subsequently suspended in a Percoll (35%) and RPMI (65%) gradient to isolate mononuclear cells.
Flow cytometry data were acquired on a FACSCanto (BD Biosciences, San Diego, CA) and analyzed with FlowJo software (Tree Star, Ashland, OR). To determine expression of cell surface proteins, mAb were incubated at 4°C for 20–30 min and cells were fixed using Cytofix/Cytoperm Solution (BD Biosciences) and, in some instances followed by mAb incubation to detect intracellular proteins. The following mAb clones were used: NK1.1 (PK136, eBioscience), CD3 (17A2, eBioscience), Ly49H (3D10, eBioscience), Ly49D (4E5, eBioscience), NKp46 (29A1.4, eBioscience), CD27 (LG.7F9, eBioscience), CD11b (M1/70, eBioscience), IFN-γ (XMG1.2; eBioscience), Granzyme B (MHGB04, Invitrogen), CD107a (1D4B, BD Pharmingen), AKT1 (55/PKBa/AKT, BD Pharmigen), pS473 (M89-61, BD Pharmigen).
Intracellular cytokine staining: For direct ex vivo, staining cells were incubated for 1 additional hour in the presence of Brefeldin A (BFA) before surface and intracellular IFN-γ staining. For cytokine staining following in vitro stimulation BFA was added during the last hour of stimulation. Intracellular signaling staining: For detection of intracellular AKT and pAKT (pS473) cells were methanol fixed and permeabilized according to BD protocol. Apoptosis was evaluated using Vybrant FAM Caspase-3/7 Assay Kit (Invitrogen) according to manufacturer’s protocol.
Splenocytes obtained from Sham or CLP hosts 48 hrs post-surgery were placed in culture for additional 48 hrs. At the end of each timepoint the number of live cells per well was enumerated and then assessed by flow cytometry to determine the frequency of live NK-cells. The % of input NK-cells was determined by the following equation: [(# of live NK-cells)at indicated timepoint] / [(# of live NK-cells)prior to culture].
Splenocytes were incubated at 37°C with 20 ng/mL each of rIL-12 and rIL-18 (R&D Systems) for 8 hrs. BFA was added during the last 4 hrs of stimulation.
Splenocytes (107/mL) from m157 and wild type littermate control mice were labeled with CarboxyFluorescein diacetate Succinimidyl Ester (CFSE; eBioscience) by incubating the cells at room temperature for 15 minutes with 1μM (CFSEhi) or 0.1 μM (CFSElo) CFSE. The labeled cells were incubated for 5 minutes with 1mL FCS on ice to remove any free CFSE, and washed three times with RPMI prior to adoptive transfer by i.v. injection, as we performed previously [107].
In vivo target cell killing was performed as previously described [66]. Briefly, target splenocytes from m157-Tg and littermate control mice were disparately labeled with CFSE (as described above). Target cells were mixed 1:1 before being injected (106 total cells) i.v. into NK1.1-depleted and control (NK sufficient) Sham and CLP hosts and analyzed 3 hrs after injection. The ratio of CFSEhi:CFSElo cells was determined by flow cytometry. m157 specific lysis was calculated using the previously described calculation: [(1-(Ratio(CFSElo: CFSEhi)sample/ Average(Ratio(CFSElo: CFSEhi)NK1.1-depleted))) x 100]. Sham and CLP groups used separate NK1.1-depleted controls for corresponding group. The specific lysis calculation adjusts the killing for the different environments present in Sham and CLP hosts.
In vitro target cell killing used the same above target cell mixture. However, target cells in the absence of effector cells served as the control. Thus, the above equation was modified to be: [(1-(Ratio(CFSElo: CFSEhi)sample/ Ratio(CFSElo: CFSEhi)target cells only))) x 100]. Splenocytes were mixed with target cells at a ratio of 1:1:1 (splenocytes: m157: littermate), i.e. 0.5:1 (effector: target). Cells were incubated for 18 hrs at 37°C. Cells were surface stained to determine the number of Ly49H+ NK-cells present at the end of stimulation.
The same target cell mix, described above (Methods: M157 Target cell killing assays), was used. Splenocytes were stimulated at multiple effector-to-target cell ratios (sample w/m157 targets) in the presence of fluorescently labeled α-CD107a mAb and monensin for 6 hrs at 37°C. Splenocytes in the absence of target cells served as controls (sample w/WT targets). CD107a expression served as a marker of degranulation and was assessed by flow cytometry in conjunction with surface staining of NK-cells. Degranulation in response to receptor stimulation was calculated as (% of CD107a+ Ly49H+ NK-cells)sample w/m157 targets−(% of CD107a+ Ly49H+ NK-cells)sample w/WT targets
Plate bound antibody stimulation was performed as described [73]. Briefly, flat-bottomed 96-well plates (Thermo-Fisher) were coated overnight (at 4°C) with 100μL of control IgG or fluorescently labeled anti-receptor antibody (diluted 1:100). Fluorescently-labeled antibody was used to identify cells that had internalized the receptor as a result of stimulation. Equivalent numbers of splenocytes were then stimulated for 8 hrs at 37°C in the presence of BfA to promote accumulation of cytokine. Cells were then surface stained, fixed/permeabilized, and stained intracellularly to evaluate cytokine production. Receptor specific response is calculated as % Stim[αreceptor antibody]- % Unstim[IgG] of receptor expressing cells.
Splenocytes were incubated at 37°C with 50μM phorbyl 12-myristate-13-acetate (PMA) and 500 μM Ionomycin (Iono) for 4 hrs. BFA was added during the last hour of stimulation.
Calcium flux assay was performed as described with some modification [108]. Briefly, 106 splenocytes were labeled with Fluo-4AM according to manufacturer’s instructions (F10489, Thermo-Fisher). Cells were then surface stained. A baseline reading was taken for 20 seconds. Cells were then stimulated by adding fluorescently labeled α-Ly49H mAb and reading was taken for 110 seconds. Cells were finally stimulated with Iono (500μM) and readings were taken for 50 seconds.
Immunoblotting was performed as described [108). Antibodies used for immunoblotting analysis were: monoclonal mouse/human α-DAP12 (D7G1X, Cell Signaling Technology) and mouse α-GAPDH (H86045M, Meridian Life Science). NK-cells were isolated by negative selection using Miltenyi Biotec NK-cell isolation kit (130-115-818) according to manufacturer’s instructions. 5x105 cells were used for each well; Sham = 3 mice/sample, CLP = 9 mice/sample. Cells were then lysed with 2X lysis buffer (20mM Tris pH 8.0, 2mM EDTA, 2 mM Na3VO4, 20mM DTT, 2% SDS, and 20% glycerol) at 95°C for 5 min. Lysates were sonicated to reduce viscosity and loaded on 10–20% Tris-HCl Protein Gel (3450033, Bio Rad Criterion). Separated proteins were transferred to PVDF membranes (Millipore) and blocked for 1 hr in 1:1 PBS:SEA Block (Thermo-Fisher) IRDye 800CW or IRDye 680-conjugated secondary antibodies were diluted in SEA Block and incubated with PVDF membranes for 30 min at room temperature. Membranes were imaged using Licor Odyssey Infrared detector.
Images were taken using Leica AM TIRF MC imaging system as described with the following modifications [108]. NK-cells were isolated by negative selection and placed on glass chamber slides (5x104 cells/chamber; LabTek II) precoated with 10μg/mL α-Ly49H mAb. Cells were stimulated for 15 minutes, fixed with 4% paraformaldehyde, and permeabilized with 0.25% Triton-X. Cells were blocked with SEA blocking buffer (Thermo-Fisher) for 1 hour and stained with 5 μL rabbit α-human/mouse DAP12 antibody (ab219765, Abcam) overnight at 4°C. Cells were washed and incubated with DyLight 488-conjugated donkey α-rabbit IgG (poly4064, BioLegend) secondary antibody for 2 hrs at room temperature. Cells were washed and fresh PBS was added to each well. Images were taken at room temperature using 100X oil submersion lens and Leica AM TIRF MC imaging system at the University of Iowa Central Microscopy Research Facility. Laser intensity and exposure parameters remained constant within each experiment. TIRF microscopy images were analyzed using ImageJ software. Membrane DAP12 was quantified by measuring mean pixel intensity in the longest axis of cells.
Cellular adhesion was performed as previously described with some modification [108, 109]. Briefly, flat-bottomed 96-well plates (Thermo-Fischer) were coated with 0–10μg of αLy49H (3D10). 5x106 splenocytes were incubated on the plate for 30 min. Non-adherent cells were removed by quickly inverting the plate to empty contents. Adherent cells were stained with αNK1.1-APC-Cy7 (PK136). Cells were washed twice with PBS before being imaged utilizing Licor Odyssey Infrared detector.
Mice were anesthetized with ketamine/xylazine (University of Iowa, Office of Animal Resources), the abdomen was shaved and disinfected with Betadine (Purdue Products), and a midline incision was made. The distal third of the cecum was ligated with Perma-Hand Silk (Ethicon), punctured once using a 25-gauge needle, and a small amount of fecal matter extruded. The cecum was returned to abdomen, the peritoneum was closed with 641G Perma-Hand Silk (Ethicon), and skin sealed using surgical Vetbond (3M). Following surgery, 1 mL PBS was administered s.c. to provide post-surgery fluid resuscitation. Bupivacaine (Hospira) was administered at the incision site, and flunixin meglumine (Phoenix) was administered for postoperative analgesia. This procedure created a septic state characterized by loss of appetite and body weight, ruffled hair, shivering, diarrhea, and/or periorbital exudates with 0–10% mortality rate, similar to our previous reports [16–19]. Sham mice underwent identical surgery excluding cecal ligation and puncture.
Total RNA was extracted from NK1.1+CD3- cells sorted 1-day post-CLP and 2 days post-Sham or CLP, two biological replicates were obtained for each group. RNA-seq. was performed as previously described and was processed by the University of Iowa Bioinformatics Division [110]. Gene expression is given as DESEQ2 values. The sequencing quality of RNA-seq. libraries was assessed by FastQC v0.10.1 (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/). RNA-seq. libraries were mapped to mouse genome using Tophat (v2.1.0) [111], and gene expressions were calculated with featureCounts, a read summarization program suitable for counting reads generated from RNA-seq. [112]. The reproducibility of RNA-seq. data was evaluated and visualized by PCA and correlation heatmap [113]. Pair-wise group comparative analyses were performed with DESeq2 with multiple test correction of fdr [114] to identify differentially expressed genes. Upregulated or downregulated genes in when comparing groups were identified by requiring a greater than 1.5-fold expression change and a false discovery rate (FDR) <0.05, as well as a non-zero DESEQ2 value. UCSC genes from the iGenome mouse mm9 assembly (https://support.illumina.com/sequencing/sequencing_software/igenome.html) were used for gene annotation. The RNA-seq. data are deposited at the GEO (accession number GSE#114739). Principal component analysis was performed using MATLAB R2017a software. Gene set enrichment and functional assignment were performed in DAVID bioinformatics resources and software from the Broad Institute as described [107, 110, 115]. Enrichment was evaluated for Day 2 CLP samples relative to Day 2 SHAM samples.
Complexes were made as previously described [89, 116]. Briefly, murine IL-2 (PeproTech) was incubated with S4B6 α-IL-2 mAb at a 2:1 molar ratio (1.5μg/mL IL-2: 50μg/mL S4B6) at 37°C for 15 min. 1.5μg of rat IgG or murine IL-2/IL-2c were injected.
Unless stated otherwise data were analyzed using Prism6 software (GraphPad) using two-tailed Student t-test (for 2 individual groups, if unequal variance Mann-Whitney U test was used), one-way ANOVA with Bonferroni post-hoc test (for >2 individual groups, if unequal variance Kruskal-Wallis with Dunn’s post-hoc test was used), two-way ANOVA (for multiparametric analysis of 2 or more individual groups, pairing was used for samples that came from the same animal) with a confidence interval of >95% to determine significance (*p ≤ 0.05). Data are presented as standard error of the mean.
|
10.1371/journal.pgen.1006785 | The bantam microRNA acts through Numb to exert cell growth control and feedback regulation of Notch in tumor-forming stem cells in the Drosophila brain | Notch (N) signaling is central to the self-renewal of neural stem cells (NSCs) and other tissue stem cells. Its deregulation compromises tissue homeostasis and contributes to tumorigenesis and other diseases. How N regulates stem cell behavior in health and disease is not well understood. Here we show that N regulates bantam (ban) microRNA to impact cell growth, a process key to NSC maintenance and particularly relied upon by tumor-forming cancer stem cells. Notch signaling directly regulates ban expression at the transcriptional level, and ban in turn feedback regulates N activity through negative regulation of the Notch inhibitor Numb. This feedback regulatory mechanism helps maintain the robustness of N signaling activity and NSC fate. Moreover, we show that a Numb-Myc axis mediates the effects of ban on nucleolar and cellular growth independently or downstream of N. Our results highlight intricate transcriptional as well as translational control mechanisms and feedback regulation in the N signaling network, with important implications for NSC biology and cancer biology.
| Stem cells are functional units in the development, maintenance, and regeneration of tissues in multicellular organisms. Defects in stem cell regulation can compromise tissue homeostasis and result in proliferative or degenerative diseases. Our understanding of the molecular and cellular mechanisms regulating the in vivo behavior of stem cells is still incomplete. The Drosophila central nervous system neural stem cells called neuroblasts have offered an excellent model system for uncovering key mechanisms and player involved in stem cell regulation. Previous genetic studies have uncovered the evolutionarily conserved Numb-N signaling pathway that regulates the self-renewal vs. differentiation choices of the cell fates of NSCs during their asymmetric division. Our understanding of how Numb-N signaling regulates NSC fate is still rudimentary. Recent studies have implicated the involvement of microRNAs in stem cell regulation in both mammalian and Drosophila systems. But the molecular mechanism and logic of miRNA action remain to be delineated. In this study we show that the bantam microRNA is a direct transcriptional target of the N signaling pathway, and that bantam feedback regulates N by negatively regulating the expression of Numb, an inhibitor of N. This feedback regulation of N helps maintain the robustness of NSC fate. We further show that bantam also impinges on a Numb-Myc axis of cell growth regulation, apparently in a N-independent manner. Together, our results highlight the importance of both transcriptional and translational control mechanisms in NSC regulation by the N signaling network. These findings have important implication for our understanding of the basic biology of NSCs and the therapeutic intervention of N-induced cancers.
| Balancing self-renewal with differentiation is a key property of all stem cells [1–3]. Tipping such balance can have detrimental consequences, resulting in lineage depletion or tumorigenesis. N signaling is critically required for lineage homeostasis of both Drosophila and mammalian NSCs [3–5]. In the Drosophila larval central brain, there are two different types of neuroblast (NB) lineages, the type I and type II NBs. N signaling appears to be dispensable for the homeostasis of type I NB lineages. In contrast, in the type II NB lineages, which differs from type I NB lineages by possessing transit-amplifying intermediate progenitors (IPs) and are hierarchically similar to mammalian NSCs, impaired N signaling leads to NB loss whereas N hyperactivation causes the dedifferentiation of IPs into cancer stem cell (CSC)-like tumor-initiating NBs [6–8]. Dedifferentiation has also been recognized as a key mechanism in tumorigenesis in mammals, highlighting the relevance of Drosophila type II NBs to the understanding of human cancer biology.
The mechanism by which N signaling maintains NB lineage homeostasis is not well defined. Cell growth regulation has recently been implicated as a key mechanism by which N maintains NSCs [7], and is particularly relied upon by CSC-like NSCs [7, 9]. Understanding how N signaling regulates the growth and maintenance of normal NSCs and CSC-like NSCs will therefore have important implications for NSC biology and cancer biology. MicroRNAs are non-coding mRNAs that regulate gene expression by base-pairing with target mRNAs to inhibit their translation or stability. The mode of microRNA action in regulating gene expression tends to be fine-tuning instead of on-or-off, making them excellent candidate players in the maintenance of the robustness of cell fates and tissue homeostasis. The microRNA pathway has emerged as a fundamental gene regulatory pathway with important roles in cell metabolism, proliferation, differentiation, and survival [10–15]. Although recent studies have implicated the involvement of microRNAs in stem cell regulation in various organisms, the molecular mechanisms and logic of microRNA action remain to be elucidated.
In this study we set out to examine the role of the bantam (ban) microRNA in the regulation of NB homeostasis in the Drosophila brain. We show that ban is a direct transcriptional target of the N signaling pathway, and that ban feedback regulates N through negative regulation of its target mRNA numb, which encodes an inhibitor of N. We show that this feedback regulation of N helps maintain the robustness of NB cell fate. Our results further show that ban also impinges on a Numb-Myc axis of cell growth regulation, apparently in a N-independent manner, thus revealing novel mechanisms in NSC regulation by the N signaling network. These findings have important implications for both the basic biology of NSCs and the therapeutic intervention of cancers caused by deregulated Numb-N signaling.
To identify new players in the N signaling network important for NSC and CSC-like growth, we tested the microRNA pathway. In the fly larval central brain, N signaling is normally required for the maintenance of type II but not type I NBs, and N hyperactivation results in the formation of CSC-like NB within the type II but not type I NB lineages [6–8]. We used clonal overexpression (OE) of N intracellular domain (N-intra), an activated form of N, to induce ectopic formation and overproliferation of type II NBs and ensuing tumorous brain growth (Fig 1A and 1B). Inactivation of Dicer-1 (Dcr-1), a key component of the miRNA pathway [16], effectively rescued N-intra induced ectopic NB formation and tumorous growth, supporting a critical role for miRNA in N-regulated NSC lineage homeostasis (Fig 1A and 1B). Given the role of ban miRNA in controlling tissue growth, cell proliferation, and survival [17, 18], we tested its involvement in NSC regulation by N. Loss of ban function as in banΔ1 null mutant had similar effect as dcr-1 mutation in rescuing N-intra induced ectopic NB formation (Fig 1A and 1B), implicating ban as a key miRNA influencing NB homeostasis. To further confirm these results, we used a transgene overexpressing ban-sponge (ban-sp), which could effectively interfere with ban function [19]. CSC-like ectopic NB proliferation induced by OE of N [7, 9] was partially blocked by ban-sp (Fig 1C and 1D). Conversely, ban OE enhanced the N OE effect (S1 Fig). Overexpression of Dpn, a putative effector of the N pathway in NBs, caused the formation of CSC-like NB within the type II but not type I NB lineages [20, 21]. NB overproliferation caused by Dpn-OE was also attenuated by ban-sp (Fig 1E and 1F). Intriguingly, in banΔ1 mutant type II NB clones without N OE, the parental NBs were preserved (Figs 2A and S2E), suggesting that ban is not essential for NB formation or maintenance under normal condition. The number of IPs, however, is reduced in banΔ1 mutant type II NB clones (S2F Fig). These results are consistent with a recent report [18], which showed that both type I and type II NBs are reduced in ban mutant brains. However, in clonal analysis, each ban mutant clone still contains a NB with appropriate marker expression, albeit with reduced cell size [18]. These results suggest that ban may act cell autonomously to regulate NB cell size, but its effect on NB number may be mediated by a non-autonomous mechanism. Together, these results suggest that ban is preferentially required for the formation and proliferation of CSC-like cells induced by N pathway hyperactivation.
We next tested whether cell growth regulation is a main mechanism by which ban mediates N effect on NSC/CSC regulation. Compared to WT NBs, banΔ1 mutant NBs in NB clones are smaller in size, consistent with a previous report [18]. This is true in type I and type II NB lineages (S2A–S2C Fig). banΔ1 mutant IPs in type II NB lineages are also smaller in size (S2D Fig). Conversely, overexpression of ban using a UAS-ban-D transgene increased the size of IPs, without obvious change of type II NB size (S2G–S2I Fig). These results support the notion that ban is involved in the growth control of NBs and IPs [18]. Previous studies demonstrated that nucleolar growth is a key aspect of cell growth in the dedifferentiation of IPs into ectopic NBs induced by N hyperactivation [7]. We found that the nucleolar sizes of both NBs and IPs in banΔ1 type II NB lineages were smaller than WT (Fig 2A–2C). Conversely, when ban was overexpressed, the nucleolar size of IPs was increased (Fig 2A–2C). The growth regulator Myc is a key mediator of N-regulated nucleolar growth in NB lineages [7]. Knockdown of dMyc effectively attenuated ban OE induced nucleolar growth (Fig 2A and 2C), whereas dMyc OE rescued the nucleolar growth defect in banΔ1 mutant (Fig 2A and 2C). As reported before [7], dMyc OE promotes nucleolar growth in IP but not NBs, whereas the depletion of dMyc leads to reduction of nucleolar size in both NBs and IPs. It appears that the nucleolar size in NBs has reached a limit making it hard for dMyc to further increase it. These results suggest that ban influences NB cell growth at least in part through Myc-mediated nucleolar growth, although we do not rule out the possibility that ban and dMyc may act in parallel to regulate nucleolar growth.
To better understand how ban regulates NB lineage homeostasis, we examined its expression and activity in type II NB lineages, which is the source of the ectopic NBs induced by N hyperactivity. Using a ban-lacZ transcriptional reporter, we found that ban expression is highly enriched in NBs (Fig 3A and 3C). This is true for both type I and type II NBs. In type II NB lineages, low level of ban expression was detected in IPs but not the differentiated neurons. Correlating with this expression pattern, ban activity as detected with a GFP sensor, the expression of which correlates inversely with ban activity [17], was high in NBs and adjacent IPs but low in differentiated neurons (Fig 3B and 3D). Given that ban expression and activity are highly enriched in NBs, we next tested whether this is under N regulation. We found that ban expression and activity were elevated by N OE (Fig 3A–3D, S3 Fig), or in α-adaptin (ada) mutant condition (S4A and S4B Fig), where N activity is high due to compromised turnover of N receptor on the cell surface [22]; conversely, ban expression and activity were diminished when N was knocked down by RNAi (Fig 3A–3D, S3 Fig). These results suggest that ban not only mediates the effect of N on NB lineage homeostasis but its expression and activity are under the control of N. qRT-PCR analysis showed that mature ban miRNA level in larval brain was increased by N-OE but decreased by N-RNAi (Fig 3E), similar to the response of known N pathway targets genes (S4C Fig). We next tested whether ban is a direct transcriptional target of N signaling. Through chromatin immunoprecipitation (ChIP) using a Su(H) antibody [22], we found that regions of ban locus including its promoter region containing putative Su(H) binding sites were preferentially pulled down in the ChIP assay. This is true in both larval brain (Fig 3F) or wing imaginal discs (S4D Fig). qRT-PCR analysis showed that ban miRNA level was increased by N-OE but decreased by N-RNAi in the wing discs (S4E Fig) as well. Collectively, these results support that ban is a direct transcriptional target of the N pathway.
The differential expression and activity of ban in progenitor cells and differentiated daughter cells, and the sharp boundary between cells with high and low ban expression and activity, raised the possibility that ban may regulate N activity to form a positive feedback loop, a mechanism commonly used to generate “all-or-none” switches during cell fate determination [23]. Using a Notch activity reporter, E(spl)mγ-GFP [24], we found that N activity in NBs was increased by ban OE, but decreased by ban-sp OE (Fig 3G and 3H). This is true for both type II and type I NBs. We chose type I NBs located at stereotypic positions in the posterior brain for analysis because they express the E(spl)mγ-GFP reporter at a higher level than other NBs, and their scattered distribution made it easier to do GFP signal quantification. We assume that what is learned from type I NBs on the regulation of N activity by ban may well be relevant to type II NBs, but given the differences in Notch function between these cell types, this is an assumption that requires further testing in the future. Our qRT-PCR analysis revealed that other known targets of N were also transcriptionally activated by ban OE but repressed by ban-sp OE (S4C Fig). These results indicate that ban participates in a regulatory network with N to help maintain NSC cell fate.
We next sought to identify the target of ban that participates in the regulatory network underlying NB fate determination. Although numb 3'-UTR has been reported to contain two predicted ban binding sites [18], the ban binding site(s) responsible for numb translational repression remains to be determined. Using the RNAHybrid program [25] for miRNA target prediction, we identified several candidate ban-binding sites in Drosophila numb mRNA 3´-UTR as well as CDS (S5A Fig). In translational reporter assays, the addition of numb 3´-UTR made the translation of the luciferase reporter specifically sensitive to the presence of ban but not let-7 miRNA (Fig 4A). Mutating the seed sequence in a best-predicted ban-binding site in numb 3´-UTR, which is distinct from the two predicted sites reported previously [18], abolished the sensitivity of the reporter to ban (Figs 4A and S5B). These results suggest that numb mRNA is a potential target of ban. To gather in vivo evidence of ban regulation of numb, we first examined the effect of ban LOF and GOF on endogenous Numb expression. Western blot and qRT-PCR analyses showed that in banΔ1 mutant or when ban-sp was ubiquitously expressed, levels of Numb protein (Fig 4B) and mRNA (S5C Fig) expression were significantly increased in the brain, whereas ban OE led to a moderate reduction of Numb protein and mRNA levels (Figs 4B and S5C).
We next examined the in vivo functional relationship between ban and numb in NB regulation. Knockdown of numb by RNAi in type II NB lineages resulted in enlarged nucleolar size in newly born IPs, and numb RNAi rescued the nucleolar size reduction caused by ban-sp OE (Fig 2D and 2E). Importantly, the nucleolar size increase caused by numb RNAi was Myc-dependent (Fig 2D and 2E; S6A–S6C Fig), as in ban OE case (Fig 2A and 2C). These results support the notion that Numb is a key target mediating the effect of ban on nucleolar growth.
We further examined the biochemical relationship between Numb and Myc underlying their functional interaction in ban-regulated nucleolar growth. In mammalian HEK293 cells, Numb and c-Myc exhibited physical interaction (Fig 4C). Overexpression of Numb led to reduced level of endogenous cMyc (Fig 4D and 4E), an effect abolished by treatment with the proteasome inhibitor MG132 (Fig 4F and 4G), suggesting that Numb affects Myc protein level through the ubiquitin-proteosome system (UPS). We have also examined Myc levels in Drosophila larval brain NBs with altered Numb activities. We found that dMyc level is increased when Numb is inhibited by RNAi, and decreased when Numb is overexpressed. This is true for both endogenous dMyc (S7A) or overexpressed dMyc (S7B). We have also tried to examine the effect of altered ban activities on dMyc expression. The immunostaining did not reveal consistent clear-cut results as seen in Numb manipulation case, probably because ban acts through Numb to indirectly affect dMyc expression, making its effect on dMyc protein level not as robust as Numb. However, when we examined the effect of altered ban activity on endogenous dMyc expression by western blot analysis of brain extracts, we saw increased dMyc protein level in ban GOF condition and reduced dMyc level in ban LOF condition (S7C Fig).
Numb has previously been shown to regulate transcription factor stability by stimulating E3 ligase activity [26]. We found that the effect of Numb in promoting Myc degradation was attenuated by knocking down Huwe1 in HEK293 cells (S7D and S7E Fig). In fly larval brain, Huwe1 RNAi resulted in increased nucleolar sizes of NBs and IPs in type II NB lineages in a Myc-dependent manner (S8A–S8C Fig). Moreover, Huwe1 functionally interacted with Numb and Myc to regulate type II NB maintenance, as indicated by the ability of Huwe1 RNAi to facilitate Myc in rescuing the type II NB loss caused by Numb OE (S8D and S8E Fig).
To examine the effect of ban on Numb protein expression specifically in NBs, we first examined Numb protein expression in the NBs in ban LOF and GOF FLP-out clones. Our results showed that Numb expression level change in the NBs displays similar trends as detected by the western blot analysis of brain tissues (S9A and S9B Fig), although the difference did not achieve statistical significance. This may be due to the sensitivity of the immunostaining method, the specificity of the antibody, or the relatively high basal expression of Numb in the otherwise wild type NBs. Consistent with the last scenario, when we examined Numb expression in N-V5 overexpression NBs that have lower basal level of Numb, the effect of ban-sp in elevating endogenous Numb expression became more significant (S10A and S10B Fig). This result further supports the notion that the translation of numb mRNA is regulated by ban in vivo.
Previous studies showed that OE of a phospho-mimetic form of Numb (Numb-TS4D) caused ectopic NB formation and tumorous brain growth, an effect likely reflecting a dominant-negative effect of Numb-TS4D in inhibiting endogenous Numb, as co-expression of Numb-WT completely rescued the Numb-TS4D effect [27]. We found that the Numb-TS4D effect was also rescued by ban-sp (Fig 4H and 4I), presumably due to elevation of the level of endogenous Numb by ban-sp that counteracted Numb-TS4D action. In contrast, co-overexpression of ban using UAS-ban-D did not change Numb-TS4D effect (Fig 4H and 4I), presumably because endogenous Numb activity has been sufficiently inhibited by Numb-TS4D such that its further translational repression by ban-D OE will not have additional phenotypic effect.
To further test for a critical role of Numb in mediating the effects of ban on N activity and CSC-like growth, we performed genetic epistasis experiments. First, we found that the effect of ban inhibition by ban-sp in attenuating N activity was mediated by Numb, as numb RNAi or removal of one copy of numb (Fig 5A and 5B; S11A and S11B Fig) blocked the ban-sp effect. In contrast, removal of one copy of pros or brat, two other genes that were recently identified as ban targets that regulate normal NB growth and proliferation [18], was without effect (S11A andS11C–S11E Fig). Consistently, numb RNAi or removing one copy of numb, but not pros or brat, effectively rescued the effect of ban-sp in attenuating ectopic NB formation and tumor-like growth induced by N hyperactivation (Fig 5C and 5D). In N-OE condition, numb RNAi or removing one copy of numb alone had no obvious effect on NB number (S12A and S12B Fig). Moreover, promotion of cell growth by Myc-OE, but not cell cycle progression by Cyclin E-OE, which failed to affect nucleolar growth (S12C and S12D Fig), suppressed the effect of ban-sp on ectopic NB formation and tumorous growth (Fig 5E). Consistent with Huwe1 being a negative regulator of dMyc, Huwe1 RNAi also suppressed the effect of ban-sp on ectopic NB number and tumorous growth (S12E and S12F Fig). Together, these results suggest that at least in N-induced CSC-like NB growth and proliferation, Numb is a key target that mediates the effect of ban, and cell growth conferred by the Numb-Myc axis is a key mechanism of NB homeostasis regulation by ban.
By revealing the involvement of the miRNA pathway, here we highlight the complexity of the N signaling network in normal NSCs and tumor-forming CSC-like NSCs. Previous studies implicated critical roles for both canonical and non-canonical N signaling pathways in NSCs and CSC-like NSCs, and revealed particular dependence of CSC-like NB growth on non-canonical N signaling, which involves PINK1, mTORC2, and mitochondrial quality control [9]. Our current study reveals a particular requirement for ban in CSC-like NBs induced by N hyperactivation. The CSC-like NB overproliferation induced by hyperactivation of N or N pathway component Dpn (Figs 1A, 1B, 1C–1F, 4I, 4J, 5C and 5D) can all be assumed to be of type II NB origin, since previous studies have clearly established that Notch signaling is essential for the development and/or maintenance of type II NBs, but dispensable for type I NBs, and that hyperactivation of Notch or its downstream effector Dpn induced ectopic CSC-like NB growth by altering the lineage homeostasis of the type II but not type I NBs [3–8, 20, 21]. It would be interesting to test whether, in addition to ban’s role in canonical N signaling, there exists a link between ban and non-canonical N signaling. Our data indicate that the ban-Numb signaling motif regulates NSC/CSC behavior through at least two mechanisms. On one hand, it regulates cell growth and particularly nucleolar growth, through Myc, a known regulator of cellular and nucleolar growth [28]. Consistently, we observed negative regulation of Myc protein level by Numb through Huwe1 and the UPS. c-Myc is an essential regulator of embryonic stem cell (ESC) self-renewal and cellular reprogramming [29], and Myc level and stability can be controlled in stem cells through targeted degradation by the UPS [30, 31], suggesting conserved mechanisms. A key function of the nucleolus is the biogenesis of ribosomes, the cellular machinery for mRNA translation, and previous studies in Drosophila have supported the critical role of nucleolar growth in NSC self-renewal and maintenance [7, 32]. On the other hand, the ban-Numb axis feedback regulates the activity of N by a double negative regulation, with the end result being positive feedback regulation. This feedback mechanism may help transform initial not so dramatic differences in N activity between NB and its daughter cell generated by the asymmetric segregation of Numb during NB division [33] into “all-or-none” decision of cell fates (Fig 6). Feed-forward regulatory loops, both coherent and incoherent, are frequently found in gene regulatory networks [23], and although ban miRNA is not conserved in mammals, miRNAs have been implicated in an incoherent feed-forward loop in the Numb/Notch signaling network in colon CSCs in mammals [34].
Given the role of ban in a positive feedback regulation of N and the potency of N hyperactivity in inducing tumorigenesis, one may wonder why ban overexpression is not sufficient to cause tumorigenesis. As in any biological systems, feedback regulation is meant to increase the robustness and maintain homeostasis of a pathway. Feedback alone, either negative or positive, should not override the main effect of the signaling pathway. Thus, in the NB system feedback regulation by ban is built on top of the available N signaling activity in a given cell and serving to maintain N activity. Because of ban’s “fine-tuning” rather than “on/off switching” of Numb expression, its effect on N activity during feedback regulation will also be “fine-tuning”, serving to maintain N activity in NB within a certain range. Overexpression of ban in a wild type background may not be sufficient to cause tumorigenesis because N activity is not be elevated to the level sufficient to induce brain tumor as in N-v5 overexpression condition. Consistent with this, the extent of Numb inhibition by ban is also modest, not reaching the threshold level of Numb inhibition needed to cause tumorigenesis. Consistent with the notion that feedback regulation by ban is built on top of the available N signaling activity in a given cell, and that there is dosage effect of N activity in tumorigenesis, overexpression of ban in N-v5 overexpression background further enhanced N-v5 induced tumorigenesis (S1 Fig). It is likely that ban or other miRNAs may participate in additional regulatory mechanisms in the N signaling network in Drosophila. Of particular interest, it would be interesting to test whether miRNAs may impinge on the asymmetric cell division machinery to influence the symmetric vs. asymmetric division pattern [35], a key mechanism employed by NSCs and transit-amplifying IPs to balance self-renewal with differentiation.
Our results emphasize the critical role of translational control mechanisms in NSCs and CSC-like NSCs. Compared to the heavily studied transcriptional control, our knowledge of the translational control of NSCs and CSCs is rather limited. As fundamental regulators of mRNA translation, miRNAs can interact with both positive and negative regulators of translation to influence gene expression [36, 37]. Thus, miRNA activity can be regulated context-dependently at both the transcriptional and translational levels, which may account for the opposite effect of N on ban activity in the fly brain and wing disc [38], although the ban genomic locus is bound by Su(H) in both tissues. Whether N regulates the transcription of ban or its activity as a translational repressor in the wing disc remains to be tested. With regard to the translation of numb mRNA, the conserved RNA-binding protein (RNA-BP) Musashi [39] has been shown to critically regulate the level of Numb protein in mammalian hematopoietic SCs and leukemia SCs [40, 41]. Further investigation into the potential interplay between miRNAs and RNA-BPs in the translational control of Numb in NBs and CSC-like NBs promises to reveal new mechanisms and logic in stem cell homeostasis regulation, with important implications for stem cell biology and cancer biology.
Fly culture and crosses were performed according to standard procedures and were raised at indicated temperatures. Drosophila stocks were obtained from the Bloomington Drosophila Stock Center, the Vienna Drosophila Resource Center (VDRC), or individual investigators in the Drosophila research community. Please see Supplementary Materials and Methods for details.
For immunostaining of Drosophila brains, late third instar larva were dissected, processed for immunohistochemistry, and imaged by confocal microscopy essentially as described [7]. Please see Supplementary Materials and Methods for details of the antibodies used for immunostaining.
The primary antibodies used for western blot analysis in HEK293T were: chicken anti-GFP (1:20,000; Abcam), mouse anti-c-myc (1:500; 9E10, Santa Cruz Biotechnologies), rabbit anti-c-Myc (1:2000; Y69, Abcam), rabbit anti-c-Myc (1:1000; Cell Signaling Technology), rabbit anti-β-actin (1:20,000; Millipore), mouse anti-Actin (1:20,000; AbD Serotec). Nuclear and cytosolic extracts were obtained using the NE-PER Nuclear and Cytoplasmic Extraction Kit (Thermo Scientific). For western blot analysis of ban LOF, GOF, and additional mutants in larval brains, protein extracts were prepared from late third instars of various genotypes, resolved on SDS-PAGE, transferred to Immobilon-P membrane (Millipore) and probed with the indicated antibodies. Please see Supplementary Materials and Methods for details of the other antibodies used for western blotting. Target protein versus loading control band intensities was measured from three independent blots with the Tina2.0 software (raytest Isotopenmessgeraete GmbH, Straubenhardt, Germany) or Image Studio Lite.
To generate MARCM clones, larva at 24 h after larval hatching (ALH) were heat-shocked for 1 hr at 38°C and further aged for 72–96 h at 25°C before dissection. For ban mutant MARCM clonal analysis, hsFLP, elav-Gal4; UAS-mCD8-GFP; FRT2A, tubP-Gal80/TM6b were crossed to FRT2A, banΔ1 /TM6b to examine ban LOF effects in normal NSCs, or crossed to Nact; FRT2A, banΔ1/TM6b to examine ban LOF effects in CSC-like NBs. For flip-out clonal analysis, w, hsFLP; Actin 5c>CD2>Gal4, UAS-GFP-NLS was crossed with the indicated UAS lines, and 24 h ALH larva were heat-shocked for 1 hr at 38°C and further aged for 72–96 h at 25°C before dissection. Occasionally, two clones may be adjacent to each other, especially in brain tumor backgrounds. These large “fused clones” can be distinguished from true “tumor clones” derived from single CSC-like NBs by the lack of GFP signal at the clone boundary in the former, and the distinct topologies in the organization of the stem cells and differentiated progenies in these two types of clones.
For NB cell size or nucleolar size quantification, measurements were performed as previously described [7]. In all Figures, unpaired Student’s t-tests were used for statistical analysis between two groups.
|
10.1371/journal.pgen.1003066 | Spastic Paraplegia Mutation N256S in the Neuronal Microtubule Motor KIF5A Disrupts Axonal Transport in a Drosophila HSP Model | Hereditary spastic paraplegias (HSPs) comprise a group of genetically heterogeneous neurodegenerative disorders characterized by spastic weakness of the lower extremities. We have generated a Drosophila model for HSP type 10 (SPG10), caused by mutations in KIF5A. KIF5A encodes the heavy chain of kinesin-1, a neuronal microtubule motor. Our results imply that SPG10 is not caused by haploinsufficiency but by the loss of endogenous kinesin-1 function due to a selective dominant-negative action of mutant KIF5A on kinesin-1 complexes. We have not found any evidence for an additional, more generalized toxicity of mutant Kinesin heavy chain (Khc) or the affected kinesin-1 complexes. Ectopic expression of Drosophila Khc carrying a human SPG10-associated mutation (N256S) is sufficient to disturb axonal transport and to induce motoneuron disease in Drosophila. Neurofilaments, which have been recently implicated in SPG10 disease manifestation, are absent in arthropods. Impairments in the transport of kinesin-1 cargos different from neurofilaments are thus sufficient to cause HSP–like pathological changes such as axonal swellings, altered structure and function of synapses, behavioral deficits, and increased mortality.
| Hereditary spastic paraplegias (HSPs) comprise a group of inherited neurological diseases. The main feature of HSP is progressive stiffness of the lower limbs due to a dysfunction of nerve cells. We study HSP type 10, which is caused by mutations in the neuronal motor protein KIF5A. HSP type 10 is inherited in an autosomal-dominant manner, which means that patients have a normal and a mutated copy of the KIF5A gene. KIF5A plays an important role in neuronal function: it transports cargos to the synapse that can be up to 1 m from the cell body. We use the fruit fly as a model to investigate the role of mutations in KIF5A. Our fly model replicates a central feature of HSP: muscles that are activated by nerve cells that have long cellular processes are more severely impaired. We now address why one mutated copy of KIF5A is sufficient to cause HSP. To date, it has been thought that patients might have HSP because they have insufficient functional KIF5A or because mutated KIF5A disturbs the function of normal KIF5A. We provide evidence for the latter possibility.
| Hereditary spastic paraplegia (HSP) is a group of genetically heterogeneous neurodegenerative disorders characterized by distal axonopathy that affects the longest axons in the corticospinal tract [1], [2]. To date, 48 HSP loci have been described. The three most common causes of HSP - accounting for more than 50% of all cases - are mutations in SPG3A (Atlastin), SPG4 (Spastin) and SPG31 (Reep1). Both Atlastin and Spastin mutations as well as mutations in 6 other identified SPG genes: (KIF5A, Nipa, Spatacsin, Spastizin, Spartin and Maspardin) have been implicated in disturbances of the intracellular transport. This suggests that perturbations in long-range, tubulin based transport might be a common pathological mechanism underlying different forms of HSP (for review see [3]).
SPG10 is inherited in an autosomal-dominant manner, with age of onset varying from childhood to the fourth decade of life [4]. The SPG10 gene KIF5A encodes the heavy chain of the neuronal microtubule motor kinesin-1 [5], [6]. The kinesin-1 family is the major anterograde motor complex.
To date, 21 different SPG10 mutations have been described [5], [6], 19 of which localize to the motor domain of KIF5A. Neither genomic deletions nor “truncating” mutations were identified as causes of autosomal-dominant SPG10, suggesting that SPG10 may be caused by a dominant-negative effect rather than by haploinsufficiency.
We generated a Drosophila model for SPG10 to validate the proposed dominant-negative interaction between mutant and wild-type Khc in the context of a living organism. Our results imply that SPG10 is not caused by haploinsufficiency but by the loss of endogenous kinesin-1 function. Thereby, KhcN262S acts as an antimorph, not as a neomorph.
In a previous in vitro study, four point mutations in the KIF5A gene that cause HSP in humans were analyzed [7]. One of these mutations, N256S, caused a reduction of motor velocity and displayed a dominant-negative effect on the function of wild-type KIF5A at physiologically relevant ratios of mutated and wild-type kinesin, but did not influence its microtubule binding affinity [7]. The N256S mutation was selected to generate the first in vivo Drosophila model for SPG10. Using this model we aimed at demonstrating that the mutated protein is stable in the context of an intact organism. Consistent with in vitro results [7] human genetic studies suggest [5], [6] that SPG10 is not caused by haploinsufficiency but by the dominant-negative interaction of mutated and wild-type kinesin. Thus, we wanted to prove that the dominant-negative action of mutated kinesin persists in vivo at physiologically relevant ratios, and is not abolished by cellular quality control mechanism, which might either prevent the hetero-dimerization of mutant and wild-type kinesin or which might selectively destabilize these dimers.
The KIF5AN256S mutation, which corresponds to the amino acid exchange N262S in Drosophila kinesin heavy chain (khc) (Figure 1A, red star), is located on loop 11. Loop 11 (Figure 1A, red line; Figure 1B, brown cylinder) connects the microtubule- (α-helix 4: Figure 1A, green box; Figure 1B, yellow helix) and the ATP-binding site (β-sheet 7: Figure 1A, blue arrow; Figure 1B, pink β-sheet 7) of Khc [8]. Five of the 21 described human mutations in KIF5A (Figure 1A, red dots) map to seven amino acids (EAKNINK) of loop 11 [9], highlighting the importance of this structure, which is essentially 100% conserved between nematodes, arthropods, and chordates (Figure 1A).
We used two complementary approaches to address the putative dominant negative action of KhcN262S in Drosophila. Both involve - unlike previous studies of loss of function alleles [10]–[16] - the ectopic expression of Khc in the wild-type background. Therefore, pathological alterations should only occur if KhcN262-Khcwt heterodimers are dysfunctional or if dimers containing KhcN262S are directly toxic. Ectopic expression of KhcN262S was induced either alone or in combination with Khcwt. The phenotypic severity of an antimorphic mutation will be decreased by increasing wild-type gene dosage. If KhcN262S acts as an antimorph, coexpression of Khcwt should ameliorate all observed defects.
Strong expression of KhcN262S in motoneurons (D42-Gal4) of wild-type larvae raised at 25°C or 29°C caused severe pathological symptoms, resulting in the death of larvae in the L2 or L3 stage, respectively. Initially, larvae display the characteristic tail-flipping phenotype originally described for khc null mutant larvae [13]. As the paralysis progresses, it ascends from posterior to anterior until larvae can only move their head (Figure 1C, red arrow; Video S1). Complete paralysis and death ensue. This “ascending” paralysis mirrors human pathological symptoms characterized predominantly by affliction of the lower limbs due to the particular vulnerability of the long descending spinal tracts. In Drosophila, axons innervating the posterior segments are considerably longer than those innervating anterior segments and seem to be - just as in humans - primarily affected.
In larvae coexpressing both KhcN262S and Khcwt (Khcwt+N262S), phenotypes are ameliorated, but still persist. This suggests that KhcN262S acts as an antimorph, not as a neomorph. At the age at which KhcN262S-expressing larvae are only able to move their head, Khcwt+N262S larvae can still locomote while displaying a tail-flipping phenotype (Figure 1C, green arrowhead; Video S2). This delayed paralysis of Khcwt+N262S larvae is progressive but does not cause larval lethality (Figure 1C). Khcwt overexpression alone does not cause any obvious defects at either 25°C (Figure 1C; Video S3) or 29°C.
When the flies were raised at 18°C, decreased expression levels of KhcN262S or Khcwt+N262S resulted in extremely short-lived flies or flies with a 25% reduction in life span (Figure 1D). Expression of Khcwt led to only a minimal but significant reduction in life span (Figure 1D, 1E).
Furthermore, the behavior of flies expressing KhcN262S at 18°C was strongly impaired. If forced to fly, animals could not sustain stable flight, but fell to the ground. At rest, they held their wings in an abnormal position (Figure 2A). Whereas control flies (Figure 2A, green arrowhead) held their wings parallel to their body axis, KhcN262S-expressing flies held their wings up (Figure 2A, red arrow). This behavior has been previously described in Parkinson disease-related Drosophila pink-1 mutants, in which the wing posture defect was attributed to apoptotic degeneration of the indirect flight muscles [17]. However, we did not observe signs of muscle loss in KhcN262S-expressing flies when thoracic indirect flight muscles were analyzed (Figure 2B), suggesting that the abnormal wing posture is secondary to impaired motoneuron function and not a result of muscle degeneration.
In the above-described models, KhcN262S is expressed throughout development. To test whether conditional expression of KhcN262S (elav-Gal4, tub-Gal80ts) after completion of development is sufficient to cause neurodegeneration, we induced expression 0 to 24 hours after eclosion. Thirteen days of overexpression at 29°C led to a 3–4 fold excess of ectopic Khc compared with endogenous Khc (Figure 2C). Importantly, equal protein levels of Khcwt and KhcN262S were detected, showing that N262S does not affect protein stability (Figure 2C). Sixteen days after initiation of conditional overexpression, flies overexpressing KhcN262S were essentially unable to climb a vertical plastic surface (Figure 2D), whereas flies overexpressing Khcwt showed no obvious locomotion defects.
In summary, KhcN262S acts as an antimorph, not as a neomorph. Either chronic or conditional expression of KhcN262S is sufficient to cause HSP-like pathological changes in Drosophila.
To further validate that KhcN262S interferes with the function of wild-type Khc in a dominant-negative manner, we compared the effects of ectopic expression of KhcN262S in a wild-type background to those observed in heterozygous and homozygous khc mutants. To this aim four to five day old larvae were selected. The expressions of KhcN262S lead - likely due to a decreased overall fitness of these larvae - to delays in larval growth as exemplified by the decreased size of 120 h old larvae (D42>w1118: 3.3 mm, D42>khcN262S: 2.4 mm, p<0.001, Student's T-Test, two-tailed). D42>khcN262S larvae showed also a trend to transit from the L2 to L3 stage being 15% smaller (data not shown). We thus decided to analyze the locomotion speed of larvae of the same size rather than of the same larval stage. This is reasonable as the length of a larva is the main biophysical parameter promoting or limiting the fast movement of a larva. The use of Animaltracer - a custom build algorithm - allowed us to determine both speed and length of larvae in a non-biased, automated manner. We next quantified the average speed of small (1–3 mm) and large (3–5 mm) larvae (Figure 3A). Locomotion of KhcN262S-overexpressing and khc deficient larvae was dramatically impaired compared with control and Khcwt-overexpressing larvae. Whereas locomotion of small Khcwt+N262S-expressing larvae was indistinguishable from that of controls, large larvae displayed a significant reduction in locomotion speed. No difference in the locomotion speed between wild-type larvae and larvae lacking one copy of khc were reported. These data support the hypothesis that SPG10 is caused by a dominant-negative action of mutated KIF5A that dramatically reduces endogenous kinesin-1 function. Consistent with the degenerative nature of HSP, locomotion defects caused by the ectopic expression of mutated Khc were generally more pronounced in larger larvae.
In the khc null mutant, the accumulation of the synaptic vesicle (SV) protein, cysteine string protein (CSP), has been reported [13]. We sought to address whether expression of KhcN262S and of Khcwt+N262S is sufficient to cause the accumulation of synaptic cargos in nerves. CSP has been reported to be a cargo of Khc [13], [18]. The expression of KhcN262S and of Khcwt+N262S led to the accumulation of cargos in axons (Figure 3B). Both the number and area fraction of nerves that were positive for cargo accumulations were highest in khc mutant larvae and in larvae overexpressing KhcN262S (Figure 3C). Axonal cargo accumulations were absent in Khcwt-expressing larvae. Similar results were obtained by staining for another SV protein: the Drosophila vesicular glutamate transporter (DVGlut) (Figure 3D, 3E) [19], [20]. Accumulations of cargo coincided with an increased intensity of the neuronal membrane marker anti-HRP (arrowheads in Figure 3B, 3D, 3F, 3H).
Next, we scored for nonselective disturbances of axonal transport. Thereby we could show that both the transport of dense core vesicles (Figure 3F, 3G; visualized by ANF-GFP), [21] and the transport of the active zone protein Bruchpilot (Brp) [22], [23] is disturbed in KhcN262S expressing larvae (Figure 3H, 3I). Both cargos are transported by the kinesin-3 family member unc-104 [18], [24]; indicating that both the fast axonal transport of kinesin-1 and kinesin-3 cargos is disturbed by the expression of mutated Khc.
We then aimed to validate that accumulations of cargo characterized by strong fluorescence for both anti-HRP and SV (Figure 3B, 3D, arrowhead) are axonal swellings. A swelling is defined by (1) the accumulation of cargo, (2) a local increase in anti-HRP staining intensity, and (3) a strong increase in axon diameter. Hurd and Saxton reported that at sites of axonal swellings, individual axons increase in diameter up to 10-fold within a micron [13]. We performed ultrastructural analysis to address whether axons were swollen upon KhcN262S overexpression (Figure 4A). An example of a typical cross-section of nerves of control larvae and those expressing KhcN262S is shown in the left panel of Figure 4A. The median axonal diameter was markedly increased in KhcN262S-expressing larvae (w1118: 0.25 µm, n = 320; KhcN262S: 0.36 µm, n = 301) (Figure 4B). Axons of control larvae contain microtubules (Figure 4A, cyan arrowhead), cargos of fast axonal transport such as mitochondria (Figure 4A, green arrowhead), and clear vesicles. As reported earlier for khc mutants [13], axonal swellings observed in KhcN262S-overexpressing larvae additionally contained large dark-staining organelles, including multivesicular bodies (Figure 4A, purple arrowhead), dark prelysosomal vacuoles (PLVs) (Figure 4A, red arrowhead), and autophagosomes (Figure 4A, dark blue arrowhead). To further validate the presence of lysosomal organelles in axonal swellings we expressed LAMP-GFP a marker for late endosomes and lysosomal compartments in motoneurons [25]. While no strong LAMP-GFP fluorescence was detected in control larvae, LAMP-GFP was strongly enriched in axonal swellings (Figure 4C, arrowhead) of KhcN262S-expressing larvae. The autosomal and autolysosomal marker ATG8-mrfp [26] also localized in axonal swellings (Figure 4D, arrowhead). The accumulation of PLVs can be triggered by impairments in the retrograde transport of small prelysosomal organelles, which then fuse and mature, giving rise to the PLVs observed in swellings [13]. Alternatively, stress-driven autophagy of the cytoplasm [27] might further contribute to the formation of PLVs [13]. The fact that the swellings are positive for ATG8 is consistent with the hypothesis that stress-driven autophagy of the cytoplasm might contribute to the formation of PLVs observed in electron microscopy [13].
We sought to use in vivo analysis of axonal transport to estimate disturbances in the delivery of cargo to synapses. The frequency at which organelles are delivered to synapses can be predicted by measuring cargo flux, i.e. the number of organelles that pass a defined cross-section of a nerve in a given time interval. Transport velocity allows estimation of how long it takes an organelle to reach its destination. A slow velocity might indirectly cause crowding in the axon. Upon a 50% reduction of cargo velocity, a 100% increase of cargo density is necessary to obtain the same flux.
Both genetic deletion of KIF5A [28] and ectopic expression of KIF5AN256S [29] reduced anterograde and retrograde flux of neurofilaments in cultured mouse cortical neurons. Although no effects of overexpression of KIF5AN256S on anterograde velocity were reported [29], deletion of KIF5A (KIF5A−/−) reduced both maximum and average velocities of neurofilament transport [28]. Neurofilaments were not depleted from distal axons upon overexpression of KIF5AN256S [29]; neither axonal swellings nor increased apoptosis was reported [29]. In cultured KIF5A−/− motoneurons [30], anterograde and retrograde transport velocities of mitochondria were reduced [30] compared with those of controls (KIF5A+/+). Effects on mitochondrial flux had not been investigated to date. We thus sought to determine the effects of deleting KIF5A on cargo flux. Anterograde (KIF5A+/+: 0.10±0.022 min−1; KIF5A−/−: 0.06±0.015 min−1, p = 0.045) but not retrograde flux of mitochondria (KIF5A+/+: 0.08±0.018 min−1; KIF5A−/−: 0.08±0.017 min−1, p = 0.93) was affected by loss of KIF5A. The number of stationary mitochondria detected within a 20 µm segment of the axon (KIF5A+/+: 3.11±0.75; KIF5A−/−: 2.83±0.429, p = 0.958) was not altered. Reductions in flux might be directly caused by reductions in transport velocity, or might be attributable to secondary defects. The 50% reduction of anterograde velocity that had been reported for KIF5A−/− motoneurons [30] fully explains the observed 44% reduction in anterograde flux that we observed.
Although loss of khc [12] resulted in the reduction of retrograde flux rates in Drosophila, no impairment was observed in the KIF5A−/− motoneuron culture model. Our measurements were performed in motoneurons isolated at day E12.5 and assessed at day 4 in vitro, an early developmental time point that corresponds to an early stage of pathological progression at which no retrograde depletion of cargo would occur. We suggest that secondary defects, e.g. impaired microtubule stability, the formation of axonal traffic jams, or the distal depletion of mitochondria, might contribute to reductions in retrograde flux observed in khc deficient larvae [12].
We were thus interested in performing in vivo imaging to address the effects of the expression of Khcwt+N262S on anterograde and retrograde cargo transport in motoneurons of the Drosophila in vivo model at a time point at which behavioral impairments can be observed. Data obtained in Khcwt+N262S-expressing larvae are therefore of particular importance. These larvae allowed us to quantify, for the first time, the effects of expression of mutated Khc at physiologically relevant ratios on axonal transport in the context of an intact nerve. In Khcwt+N262S-expressing larvae, stochastically more than 25% of the Khc motors are expected to be Khcwt-homodimers. We did not observe a significant change in the velocity of mitochondria in either direction in any of the investigated genotypes. We did, however, detect a strong reduction in both anterograde and retrograde flux in both KhcN262S and Khcwt+N262S-expressing larvae. Khcwt-expressing larvae did not show any changes (Figure 5A–5F; Videos S4, S5, S6, S7).
Ebbing and colleagues [7] reported that in vitro velocities of purified KIF5A constructs were reduced more than two-fold upon mixing wild-type and N256S-mutant kinesin at a stoichiometric ratio of 1∶1. The authors further assumed that kinesin cargo vesicles are moved by 5 to 8 motors [7]. Under these conditions, each organelle is expected to have a high probability of being attached to at least one mutant motor, leading to slower motility and shorter run lengths. The fact that we did not observe slower mitochondria suggests that the assumptions used to extrapolate single-molecule measurements to organelle transport in a cellular environment might be oversimplified. Alternatively, the experimental approach chosen to measure transport in vivo might be flawed. Bleaching is routinely used to quantify mitochondrial transport in Drosophila [12], [18]. To exclude an influence of the bleaching procedure on our results, we sought to compare flux and velocities obtained before and after bleaching. We are not aware of a study that experimentally validates that transport velocities are not affected by the bleaching procedure. Theoretically, slow mitochondria might not enter the bleached region during the analyzed time interval. Thus, they might be excluded from the analysis, resulting in a biased analysis due to an inappropriate selection of fast-moving mitochondria. By comparing the mitochondrial flux in bleached and in non-bleached nerve segments, we could show that bleaching has an effect on flux rates; a higher flux is observed when analysis is performed after bleaching (Figure 6A, 6B). This observation is best explained by the fact that bleaching allows for better visualization of moving mitochondria, which are less likely to be obscured after stationary mitochondria have been bleached. Both anterograde and retrograde flux is affected to the same degree by the method chosen. Thus, the ratio of retrograde to anterograde transport flux is not affected by the experimental procedure (Figure 6C). We observed no effect of bleaching on transport velocities (Figure 6D, 6E). To further confirm our results, we additionally performed a comparison of transport velocities obtained from two non-bleached control genotypes (D42>w1118; D42>Khcwt) and two non-bleached mutant genotypes (D42>KhcN262S; khc−/−). No significant reduction of anterograde or retrograde transport velocities was detected in any of the investigated phenotypes (Figure 6D, 6E). As no effect of bleaching on velocities could be observed, we suggest the use of bleaching when quantifying mitochondrial transport in Drosophila larvae.
We were next interested in further investigating the cause of the reduced mitochondria flux. The reduced flux might be directly caused by impairments in axonal transport or by depletion of mitochondria in the cell body or near synapses. No obvious reduction in mitochondrial abundance in motoneuron cell bodies was detected (Figure 7A). Thus, the observed reduction in anterograde flux is likely caused by impaired transport of mitochondria.
Although the size of mitochondria was not affected by expression of mutated Khc, a strong reduction in the number and density of mitochondria at neuromuscular junctions (NMJs) 6/7 in segment 2 was detected in KhcN262S-expressing larvae (Figure 7B–7E). A trend toward a reduced mitochondrial number (p = 0.07) and density (p = 0.08) in Khcwt+N262S-expressing larvae was observed (Figure 7C–7E). Thus, reductions in the retrograde flux might be the result of impaired retrograde axonal transport or of reduced abundance of mitochondria at the synapse, or a combination of both effects.
Next, we were interested in studying the structure and function of NMJs in more detail. Behavioral experiments in Khcwt+N262S-expressing larvae and data obtained in khc null mutants [13] predict strong defects at posterior segments, whereas anterior segments should be less affected. Quantification of both the NMJ area and the synapse number (postsynaptic glutamate receptor fields) revealed that this is indeed the case (Figure 8A–8E). Overexpression of KhcN262S or Khcwt+N262S led to a strong reduction in the area of NMJs 6/7 in segment A5 but not in segment A2 (Figure 8A–8C). Affected NMJs are furthermore characterized by inhomogeneity in anti-HRP staining (Figure 8A, arrowheads). No significant reduction in the number of synapses was detected in anterior segment A2 (Figure 8D). Overexpression of KhcN262S or Khcwt+N262S caused a significant reduction, however, in the number of synapses (Figure 8E) in posterior segment A5.
We next sought to address whether reduced axonal transport does limit the supply of NMJs with SVs and active zone proteins. To this aim, we used CSP and DV-Glut as markers for SVs and Brp as a marker for AZs. All three proteins are present in axonal swellings (Figure 3B, 3D, 3H). We could confirm that the abundance of both AZ (Figure 9A, 9B) and SV (Figure 9C–9F) proteins is reduced at the NMJ. SV proteins are inhomogenously distributed in KhcN262S-expressing larvae. While few boutons stain intensively for CSP and DV-Glut (Figure 9C, 9E arrowheads), other boutons display a weaker staining intensity (Figure 9C, 9E arrows). There is a strong correlation between the inhomogeneity observed in the staining for HRP and SV proteins. This inhomogenous distribution might resemble defects in the delivery of SV, in endo-/exocytosis, or in membrane trafficking.
Thus, we were interested in addressing functional impairments of the NMJ in detail. To this aim, we recorded postsynaptic potentials by using current clamp recordings at muscle 6 in segment A4. The evoked excitatory junction potentials (eEJPs) of KhcN262S-expressing larvae were drastically decreased in amplitude and had an increase in the half-width time (Figure 10A, 10B, 10C). The increased half-widths in KhcN262S- and Khcwt+N262S-expressing larvae might be caused by impairments in the synchronization of vesicle fusion with the arrival and spread of the action potential to all release sites [23]. Loss of Brp - an active zone protein that has been shown to be important for establishing close proximity between SV and release sites - leads, in like manner, to an increase in the half-width of evoked excitatory junctional currents [23].
In contrast to eEJPs, the amplitude of miniature excitatory junction potentials (mEJPs) in response to single, spontaneous vesicle fusion events was slightly - but not significantly – increased in both KhcN262S- and Khcwt+N262S-expressing larvae (Figure 10D, 10E). This might represent a postsynaptic compensation presynaptic defect. Indeed, although there was no significant difference in eEJPs between Khcwt+N262S-expressing larvae and controls, the former group revealed a significant reduction in quantal content (Figure 10F). As quantal content is a measure of the number of vesicles released per presynaptic action potential, it is better suited for characterizing presynaptic defects than eEJP size.
As HSP is a neurodegenerative disorder characterized by distal axonopathy, we were interested in whether we could observe any signs of synapse degeneration in our models. Both KhcN262S- and Khcwt+N262S-expressing larvae showed pathological alterations in neuronal membrane organization (Figure 8A; Figure 9A, 9C, 9E), as well as reduced abundance and altered distribution of SV proteins (Figure 9C, 9E; Figure 11A–11D). However, typically the complete absence of the SV protein synapsin from parts of the NMJ was not detected in KhcN262S-expressing larvae (Figure 9C, 9E; Figure 11C, 11D). These data are consistent with the reduction in abundance and inhomogeneous distribution of CSP and DV-Glut (Figure 9C, 9E) observed in KhcN262S-expressing larvae. Synaptic footprints [31], areas of the NMJ that are, after a retraction of the nerve, positive for postsynaptic marker proteins (Dlg or GluRIIC), but negative for presynaptic marker proteins, are commonly scored by using either synapsin [31]–[33] or Brp [34] as a presynaptic marker. The absence of a single presynaptic marker protein is a clear indication of pathological alterations at the NMJ. It is not sufficient, however, to prove that a nerve ending has retracted. We thus defined only the simultaneous absence of an SV marker (synapsin) and of the presynaptic membrane (HRP) as a retraction event. Using this more conservative assay, retraction was seldom detected. Dystrophic boutons characterized by a strong reduction in the intensity of SVs, in combination with an inhomogeneous HRP signal, were, however, frequently detected in animals expressing mutant Khc. To quantify the degree of neurodegenerative pathological alterations at the NMJs, we used a scoring system that assessed the frequency of retractions, the occurrence of dystrophic boutons, and the presence of minor pathological alterations such as weaker staining for SV (compare Figure S1A–S1F and Text S1). Using this scoring system, we detected a significant degree of neurodegenerative alterations in larvae expressing mutant Khc (Figure 11E).
A strong HRP staining at a subset of boutons at the NMJs of larvae expressing mutant Khc suggested the local accumulation of membrane rich organelles. As autolysosomal organelles were detected in axonal swellings, we sought to address whether the observed strong HRP signal might be indicative of increased autophagy at the NMJ. Indeed, ultrastructural analysis of larvae overexpressing Khcwt+N262S indicated that PLVs (red arrowhead), autophagosomes (dark blue arrowhead), and multivesicular bodies (purple arrowhead) are frequently present in mutant but not in control NMJs (Figure 12A and 12BI–12BIII).
Next, we questioned whether the observed degeneration represents a classic distal synaptopathy or whether it is preceded by the loss of the motoneuron cell body. To this aim, we analyzed motoneuron cell bodies in 4-day-old early-mid L3 larvae, the stage at which degeneration was observed at the NMJ. No substantial motoneuron loss was detected in the subset of motoneurons positive for eve in KhcN262S-expressing larvae when compared with the control group (Figure 12C). To monitor cell death in all motoneurons we used an antibody that allows visualizing the activation of the putative initiator caspase DRONC [35]. Using this antibody, which is commonly used to quantify dying cells in Drosophila (for review see [35]), we found no evidence for increased apoptosis of motoneurons in mid L3 KhcN262S-expressing larvae (Figure 12D).
The fact that we did not observe substantial motoneuron cell loss in mid-L3 KhcN262S-expressing larvae, a stage when degeneration of synapses was already pronounced, is consistent with the concept that HSP is primarily caused by synaptopathy and axonopathy, whereas motoneuron cell loss is neither causative for HSP nor an early feature of the pathological process.
Human genetics suggest that autosomal-dominant SPG10 is not caused by haploinsuffiency. The above presented Drosophila model further supports this hypothesis. While no difference in the locomotion speed between wild-type larvae and larvae lacking one copy of khc were observed, larvae ectopically expressing KhcN262S were severely impaired. The observed behavioral impairments are both qualitatively and quantitatively similar to impairments observed in khc null larvae.
These results are consistent with KhcN262S acting as an antimorph or a neomorph. Antimorphic mutations act in opposition to the normal gene function. Thus, they are also referred to as dominant negative mutations [36]. The phenotypic severity of an antimorphic mutation will be decreased by increasing wild-type gene dosage [36]. A neomorphic mutation leads to a change in the nature of the gene resulting in a dominant gain of function. This function, which is not produced to a relevant degree by the native gene, is often toxic. Increasing wild-type gene dosage will not reduce the phenotypic severity of a neomorphic mutation, as the newly gained function is by definition different from the normal gene function [36].
Animals ectopically expressing KhcN262S alone were more severely affected than animals expressing KhcN262S in combination with an additional copy of Khcwt. We thus conclude that KhcN262S is an antimorphic mutation: KhcN262S causes loss of kinesin-1 function via a dominant-negative mechanism.
We propose that all observed pathological changes are downstream of a loss of kinesin-1 function and cannot be attributed to more generalized toxicity of mutant Khc or kinesin-1 complexes containing mutant Khc. In human patients, 25% of the Khc motors are expected to be Khcwt-homodimers. Our results suggest that these should be capable of transporting synaptic cargo.
Two models have been proposed to explain how molecular motors coordinate cargo transport (for review, see [37]). The tug-of-war model assumes that the direction of movement is the result of the dynamic competition of opposing motors. The simultaneous action of two motors thus exerts a stretching force on the particle. Loss of motors responsible for transport in one direction should increase transport velocities in the opposite direction. While the direct opposing action of dynein and the Dictyostelium kinsesin-3 family member Unc-104 could be experimentally validated [38], most frequently, the loss of motors responsible for transport in one direction leads to transport disturbances in both directions [30], [39]. This is best explained by a model in which the activity of opposing motors is controlled by coordination complexes that are regulated such that only one set of motors is active at any given time. These two models are, however, not mutually exclusive. Both models offer distinct advantages for molecular motors. Ideally, coordinated action of opposing motors might allow for faster, more energy-efficient transport, but simultaneous binding of opposing motors has been proposed to decrease the probability of detachment from microtubules, resulting in increased processivity of movement [40]–[42]. Thus, it is most probable that in most cellular environments cargo is transported by the coordinated, simultaneous action of opposing motors. The exact balance between coordinated inactivation of opposing motors and their active role as a stabilizing “dragging force” might vary for different cargos, developmental time points, stages of disease progression, and distinct cell types. Thus, it is of no surprise that studies investigating mutation in molecular motors performed in different model systems lead to seemingly opposing results.
In Drosophila, mitochondrial transport velocities are affected neither by the deletion of khc [12] nor by overexpression of KhcN256S. These observations are in accordance with data obtained by measuring neurofilament transport in cultured mouse cortical neurons ectopically expressing KIF5AN256S [29]. Yet how can these results be reconciled with the observation that maximum and average velocities of mitochondrial [29] and neurofilament [28] transport were reduced in neurons isolated from KIF5A−/− mouse embryos? The loss of KIF5A might be partially compensated by KIF5B or KIF5C. KIF5B and KIF5C might be less effective, however, in transporting mitochondria, thus reducing both average and maximum transport velocity. While no compensatory up-regulation of KIF5B or KIF5C was detected in KIF5A−/− mice [30], indirect evidence nonetheless suggests that residual mitochondrial transport observed in KIF5A−/− motoneurons might be driven by KIF5B or KIF5C. First, both mitochondria and neurofilaments are actively transported in the anterograde direction [28], [30]. Thus, they must be bound to an anterograde motor. Second, KIF5C and KIF5B have been shown to be important for mitochondrial transport [43]. In KIF5B−/− cells, mitochondria are clustered around the nucleus rather than being appropriately dispersed throughout the cell. This defect can be rescued by ectopic expression of KIF5A, KIF5B, or KIF5C, highlighting the fact that any of these motors are capable of transporting mitochondria [43]. While the frequency of neurofilament transport is reduced by 75% in KIF5A−/− neurons, KIF5C or KIF5B overexpression is sufficient to partially rescue this defect [28].
We thus propose that cellular quality control mechanisms ensure that cargos are equipped with a minimal number of molecular motors. Thus, obvious defects such as the genetic deletion of KIF5A are detected and partially compensated by targeting similar motors to cargos awaiting initiation of transport. Transport driven by these alternative motors might be slower and less efficient [28], [30]. Still, a basal cargo flux can be obtained despite the complete absence of one molecular motor type [28], [30]. Mutations in KIF5A that impact neither the stability of the protein nor its ability to interact with regulatory complexes might not be detected by this cellular quality control system. Thus, cargos might be loaded with defective motors that cause problems after initiation of transport. These secondary problems might include a drop of transport flux [28], [29]. The phenotypic strength of defects might therefore depend on the expression level as well as the nature of the mutation. Thus, observed defects might be either less severe (KIF5AN256S [29]) or more severe (head-less KIF5A, [28]) than loss of KIF5A [28].
KhcN262S-expressing larvae exhibit the characteristics of classical distal degeneration. No human SPG10 autopsy reports have been published till date, but studies of SPG4 cases [1] suggest a “dying back” axonopathy as probable disease mechanism in HSP. Synapse and axon loss may therefore be a primary step in the pathophysiological manifestation of SPG10, while neuronal loss may occur only later.
Could protection of the cell bodies be used as a suitable treatment strategy for HSP? While little data is available on the treatment of HSP patients, treatment strategies for Amyotrophic lateral sclerosis (ALS), a clinically related disease, have been explored in more detail. The neurodegenerative disease ALS is characterized by the progressive loss of motoneuron in the brain and the spinal cord [44]. Pathological changes occur in ALS patients and animal models - similar to our observations in the Drosophila HSP model - first at the NMJ and are followed by axon and neuronal loss [45], [46].
Treatment strategies aiming at prevention of motoneuron loss lead only to limited success in ALS models and human patients [44]. Thus, treatment that aims at delaying motor impairments during the progression of ALS is currently favored [44]. We suggest that preservation of synapses and axons will be an important requirement for a successful therapeutic intervention of both HSP and ALS.
Impaired neurofilament transport has been implicated in the pathogenesis of SPG10 [29]. The human genome contains 3 genes encoding the heavy chains of conventional kinesin: the neuronally expressed genes KIF5A and KIF5C and the ubiquitously expressed gene KIF5B. Neurofilaments are transported by KIF5A and KIF5C [28], [29]. Impairments in neurofilament transport due to impaired KIF5A function might thus explain why only mutations in KIF5A, but not KIF5B, have been identified as a genetic cause for HSP.
The Drosophila genome, on the other hand, does not contain a neurofilament gene [47]. Electron microscopic studies concluded that neurofilaments are absent in all arthropods [48] and yet ectopic expression of mutant Khc leads to the formation of axonal swellings and HSP-like pathological changes in Drosophila.
Conversely, the hypothesis that impairments in neurofilament transport represent an important cellular cause of HSP has been further supported by reports that swellings in SPG4 patients and the SPG4 mouse model were positive for neurofilaments [49].
These swellings also contained, however, multiple other cargos, including mitochondria, the amyloid precursor protein, tubulin, and tau [49]. In our study, both kinesin-1 and kinesin-3 cargos [24], as well as lysosomal and autophagic organelles, accumulated in swellings. This suggests that eventually all kinds of cargo might become trapped in axonal swellings.
Future studies involving the use of in vivo imaging in mouse and Drosophila models will be needed to shed light on the formation of swellings. It is therefore important to identify cargos that accumulate first within a swelling. The cargo, which causes the formation of the swellings, should accumulate prior to other cargos.
A more detailed understanding of the temporal sequence of cargo accumulation and an increase in the understanding of impaired neurofilament transport will be instrumental to further decipher common pathological hallmarks of HSP. It is to be hoped that this will aid in the design of successful treatment strategies.
Site-directed mutagenesis was used to introduce the amino acid exchange N262S in a full-length wild-type Drosophila khc cDNA (SD02406) which was next inserted into a modified pUAST attB vector. Details are described in Text S2.
Flies were maintained at 25°C on standard fly medium seeded with yeast. For experiments, flies were raised at 18°C, 25°C, or 29°C. For fly strains used in this study, see Table S1. Transgenic stocks UAS-khcwt and UAS-khcN262S were created by BestGene using integrase mediated site-specific transgenesis at cytological position 86F (Fly strain BDSC 23648).
Proteins were extracted from whole fly heads using 65 mM Tris (pH 6.8), 5% (w/v) SDS, 1× Protease Inhibitor (Roche) buffer. Samples were separated on 7.5% SDS gels and transferred to nitrocellulose membranes. For antibodies used in this study, see Tables S2 and S3.
Third instar larvae were dissected in chilled Ca2+-free HL3 solution and fixed in 4% formaldehyde in PBS. For antibodies used in this study, see Tables S2 and S3. Larval preparations were mounted in Vectashield (Vector). Images were captured using a Zeiss LSM 710 confocal microscope with the following settings unless otherwise noted: Objective: 40× plan Apochromat, 1.3 N.A.; Voxel Size: 100 nm×100 nm×500 nm; pinhole: 1 AU, average: 2–4. Images used for illustration purposes were processed as follows: (1) A Gaussian filter (radius = 2) was applied to the raw data stack. Brightness and contrast were appropriately adjusted. The relevant slices of the modified stacks were maximum-projected. Projected images were scaled by 2, and gamma adjustment (gamma = 0.75) was applied. ImageJ 1.41o, 1.44p, 1.46r or 1.45s (US National Institutes of Health; http://rsb.info.nih.gov/ij/download.html) was used to process and analyze images. For quantification of Glutamate receptor fields, Delta 2D software (Decodon GmbH, Germany) was used.
In vivo imaging was essentially performed as previously described [50]–[52], using a Zeiss LSM 710 confocal microscope equipped with a 40× Plan Apochromat Objective (1.3 N.A.). For better visualization of moving mitochondria, all mitochondria in a 20 µm segment of the nerve were bleached. This allowed for easy visualization of moving particles passing through the bleached region. Next, confocal Z-stacks (Z-planes: 10; voxel size: 100 nm×100 nm×1500 nm, pinhole: 1.6 AU, average: 2) were recorded at maximal speed corresponding to 100 stacks per 426 seconds. For details on the generation of kymographs, see Text S1.
Third instar larvae were dissected, rinsed, and transferred into the recording chamber. A fixed-stage upright microscope (Model BX51WI with 40× water immersion lens; Olympus Optical, Tokyo, Japan) was used to visualize the nerve and the muscles. Intracellular current clamp recordings were performed in HL3 solution with 1 mM extracellular Ca2+ at 19°C. Evoked excitatory junction potentials (eEJPs) and spontaneous miniature excitatory junction potentials (mEJPs) were obtained from muscle 6 segment A4 with an Axoclamp 900A amplifier (Axon Instruments, Union City, CA), digitized (Digidata 1440A, Axon Instruments, Union City, CA), recorded at 10 kHz (pClamp 10, Axon Instruments, Union City, CA), and analyzed using AxoGraph X software. Sharp, bee-stinger-shaped glass microelectrodes filled with 3 M KCl and a resistance between 10 and 20 MΩ were used. Cells with resting potentials between −60 and −70 mV and input resistance >4 MΩ were selected for analysis.
For stimulation, the cut end of the segmental nerve was pulled into a fire-polished suction electrode (6–8 µm inner diameter), and brief (300 µs) depolarizing pulses were passed at 0.1 Hz (ISO-STIM 01D, NPI Electronics, Tamm, Germany, stimulus generator and isolation unit). The amplitude of the pulse was set to about 7V, which results in the stable recruitment of both innervating motoneurons. It corresponds to 1.5 times the amplitude needed to recruit both motoneurons innervating muscle 6. For each eEJP and mEJP average, 15 eEJPs and 120 s of mEJP recordings were used for subsequent analysis.
Larval fillets were fixed with 4% PFA (in PBS) for 10 min at room temperature followed by fixation in 2.5% glutaraldehyde (in PBS) overnight at 4°C. Postfixation was done with 1% osmium tetroxide in 100 mM phosphate buffer, pH 7.2, for 1 h on ice. Larval fillets were rinsed with water, treated with 1% aqueous uranyl acetate (UA) for 1 h at 4°C, dehydrated through a graded series of ethanol concentrations, and stored in liquid Epon overnight. Next, larval fillets were pinned on a dissection pad. Muscles 4 of segment 4 were dissected with sharp insect pins, embedded in Epon, and polymerized for 48 h at 60°C. Ultrathin sections were stained with UA and lead citrate and viewed in a Philips CM10 electron microscope.
Living L2 larvae were transferred to aluminum platelets with a 150 µm recess containing 1-hexadecene as an external nonpenetrating filler. The platelets were sandwiched with platelets that had no recess and cryofixed with a high-pressure freezer (Bal-Tec HPM 010, Balzers, Liechtenstein). Larvae were freed from external hexadecene under liquid nitrogen and then transferred to 2 ml microtubes with screw caps (Sarstedt, #72.694, Germany). As a freeze substitution medium, we used a 2% osmium tetroxide solution in anhydrous acetone, supplemented with 25 µl of 20% methanolic UA solution to give a final UA concentration of 0.5%. The freeze substitution was carried out in a Leica AFS-2 with the samples kept at −90°C for 27 h, −60°C for 6 h, and −40°C for 6 h. The temperature increase between steps was set to 10°C/h. At −40 h, glutaraldehyde from a 25% solution in water (EMS #16530, Electron Microscopy Sciences, Fort Washington, PA) was added to give a final concentration of 0.6% glutaraldehyde and 2% water. After 6 h at −40°C, the microtubes were placed on ice for another 1 h. Samples were then washed 3 times with acetone and infiltrated with Epon at room temperature in a series of increasing Epon concentrations in acetone (30%, 60%, 90%, 2× 100% Epon each for 1 h, with the second 100% Epon change continuing overnight on a rotating wheel). After embedding, the Epon samples were polymerized for 48 h at 60°C. Ultrathin sections were stained with UA and lead citrate and viewed in a Philips CM10 electron microscope. Micrographs were taken on EM-film (Maco ES 208, Hans O. Mahn GmbH & Co KG, Stapelfeld, Germany). Alternatively, sections were imaged by using a FEI-Tecnai Spirt, 120 kV electron microscope equipped with a Gatan USC 4000 camera.
Thoraxes were dissected by careful removal of heads, wings, limbs, and abdomen parts, prefixed at 4°C overnight (4% PFA, 3% glutaraldehyde, 0.1% sodium cacodylate), and postfixed with 1% osmium tetroxide for 3 h at 4°C. Next, thoraxes were washed with 30%, 50%, 70%, and 100% ethanol for 10 min each followed by washing with 100% acetone and 3∶1 acetone∶Epon for 1 h. Thoraxes were then immersed in 1∶1 acetone∶Epon and 100% Epon for 24 h. The Epon was polymerized for 48 h at 60°C. Semi-thin sections (2 µm) were prepared on a Reichert-Jung Supercut 2050 microtome with glass knifes. The semi-thin sections were stained with Toluidine blue solution (0.5% Toluidine blue O [C.I 52040, Roth] in 1% [w/v] disodium tetraborate buffer) for 1 minute and then washed under running water. The semi-thin sections were documented on a Zeiss Imager.Z1m microscope by using a 20× Zeiss Neofluar, 0.5 N.A objective. All washing, fixation, or staining procedures were performed at room temperature unless otherwise noted.
Flies were raised at 18°C. Offspring were collected on the day of eclosion and 15–20 male flies transferred to vials containing standard fly media. The flies were transferred to fresh fly media every 3 days. A Kaplan-Meier plot was used for depicting survival curves.
Flies were raised at 18°C. Emerged male flies were split into two batches of 50 flies within 24 h after eclosion. One batch of flies was raised at 18°C and the other at 29°C to induce the expression of either Khcwt or KhcN262S for 16 days. Motor function of 16-day-old flies was monitored by analyzing their ability to climb 6 cm at the wall of a vertical plastic tube within 15 s. Fifty flies from each genotype were analyzed. A successful trial was scored with 1 and a nonsuccessful trial with 0. Each fly was allowed to climb three times and the average climbing score per fly was calculated.
To monitor locomotion behavior, we placed individual larvae on a thin slice of apple juice agar. Locomotion was examined at 25°C at 70% humidity by using a DCM510 (ScopeTek, P.R. China) camera integrated in a custom-built stereomicroscope. Larval locomotion was recorded at a frame rate of 30 fps for 5 min. The videos were then converted into avi format by using a Prism Video Converter, v 1.61 (NCH Software Inc., Australia). Next, images were cropped and compressed by using VirtualDub 1.9.10 (http://www.virtualdub.org/).
To measure locomotion speed, we placed up to 200 larvae on a 15×15 cm agar plate and filmed for 10 min. Locomotion and size of the larvae were analyzed with the custom-built software Animaltracer. This software was developed by us on the basis of the MATHLAB software package Worm Tracker & Track Analyzer (Department of Molecular and Cellular Physiology at Stanford University) [53]. This algorithm can be divided into two parts, the larval tracker and the track analyzer. The larval tracker identifies and tracks individual larvae within a movie. The track analyzer analyzes the movies and returns the size and the velocity of single larvae. For comparative analysis of the different genotypes larvae within a certain size range were grouped. Average locomotion speed is calculated for this size group for every movie. Larvae that touched each other were automatically excluded from analysis. Larvae whose velocity was less than 10% of average velocity of the respective genotype and size group were likewise excluded from analysis. A minimum of six movies per genotype were analyzed. For all further statistical analysis, n was defined as the number of movies analyzed.
KIF5A+/− mice were obtained from MMRRC (Mutant Mouse Regional Resource Centers, University of California, Davis, USA). Mouse embryonic motoneuron culture, staining of mitochondria, and time-lapse imaging were performed as previously described [30]. Flux of mitochondria was measured in a 20 µm segment of motoneuron axons.
The number of mitochondria passing two defined cross-sections, both in the anterograde and the retrograde direction, were counted in a time interval of 30 minutes. The flux is the average of these two measurements. Mitochondria that moved less than 5 µm within the 20 µm segment were classified as stationary. 18 KIF5A+/+ and 23 KIF5A−/− axons from four independent experiments were analyzed.
All animal work in this study were approved by the German Government (Regierungspräsidium Tübingen) and the University of Tübingen.
Statistical tests were performed with the software PAST.exe (http://folk.uio.no/ohammer/past/index.html) unless otherwise noted. Sample errors are given as standard deviation (s.d.) and standard error of the mean (s.e.m). Data were first tested for normality by using the Shapiro-Wilk test (α = 0.05). Normally distributed data were analyzed either by student's t-test (two groups) or by a one-way analysis of variance followed by a Tukey-Kramer post-test for comparing multiple groups. Non-normal distributed data were analyzed by using either a Mann-Whitney test (two groups) or a Kruskal-Wallis H-test followed by a Dunn's test for comparisons between multiple groups. Differences in survival were determined by the Mantel-Cox test using Prism. The p values obtained from the Mantel-Cox test were corrected for the total number of comparisons made. Statistical tests for analyzing axonal transport in KIF5A-deficient motor neurons were performed with IBM SPSS Statistics, Version 20. The following alpha levels were used for all tests: * p<0.05; ** p<0.01; *** p<0.001.
|
10.1371/journal.pntd.0005947 | How does competition among wild type mosquitoes influence the performance of Aedes aegypti and dissemination of Wolbachia pipientis? | Wolbachia has been deployed in several countries to reduce transmission of dengue, Zika and chikungunya viruses. During releases, Wolbachia-infected females are likely to lay their eggs in local available breeding sites, which might already be colonized by local Aedes sp. mosquitoes. Therefore, there is an urgent need to estimate the deleterious effects of intra and interspecific larval competition on mosquito life history traits, especially on the duration of larval development time, larval mortality and adult size.
Three different mosquito populations were used: Ae. aegypti infected with Wolbachia (wMelBr strain), wild Ae. aegypti and wild Ae. albopictus. A total of 21 treatments explored intra and interspecific larval competition with varying larval densities, species proportions and food levels. Each treatment had eight replicates with two distinct food levels: 0.25 or 0.50 g of Chitosan and fallen avocado leaves. Overall, overcrowding reduced fitness correlates of the three populations. Ae. albopictus larvae presented lower larval mortality, shorter development time to adult and smaller wing sizes than Ae. aegypti. The presence of Wolbachia had a slight positive effect on larval biology, since infected individuals had higher survivorship than uninfected Ae. aegypti larvae.
In all treatments, Ae. albopictus outperformed both wild Ae. aegypti and the Wolbachia-infected group in larval competition, irrespective of larval density and the amount of food resources. The major force that can slow down Wolbachia invasion is the population density of wild mosquitoes. Given that Ae. aegypti currently dominates in Rio, in comparison with Ae. albopictus frequency, additional attention must be given to the population density of Ae. aegypti during releases to increase the likelihood of Wolbachia invasion.
| Several countries are seeking new vector control tools to reduce the transmission of arboviruses such as dengue, chikungunya and Zika. One of these innovative approaches relies on the release of Aedes aegypti mosquitoes infected with the endosymbiont Wolbachia, since this bacterium can block the aforementioned viruses and interrupt transmission. Countries in Latin America and Southeast Asia have a strong co-occurrence of Ae. aegypti and Ae. albopictus in their urban landscapes. Therefore, Wolbachia-infected Ae. aegypti mosquitoes are likely to lay their eggs in local breeding sites already colonized by wild uninfected conspecifics and/or Ae. albopictus larvae. We conducted experiments to study larval competition with varying larval densities, species proportions and food levels. Interestingly, Ae. albopictus proved to be a superior competitor under different scenarios: its larvae had superior survivorship, faster development rate and a higher performance index than Ae. aegypti (both infected and uninfected groups). The presence of Wolbachia increased larval survivorship of Ae. aegypti. Our data show that the population density of wild mosquitoes, especially interspecific competition, can become an additional force to reduce the pace of Wolbachia invasion in endemic regions.
| Infectious diseases caused by arboviruses are a growing global health concern. Among the disease vectors, mosquitoes from the genus Aedes and mostly important Ae. aegypti (Linnaeus, 1762) and Ae. albopictus (Skuse, 1894) have a prominent role in transmitting several arboviruses to humans. In the last 50 years, dengue virus (DENV) has shown a 30-fold increase in global incidence, with around 400 million new infections yearly [1–3]. In the last decade, chikungunya became pandemic after spreading from limited regions of Africa and Asia and arriving into the Americas. Two CHIKV genotypes were detected in Brazil: The Asian genotype has probably invaded the country through the Caribbean and the East-Central-South African (ECSA) genotype was first detected in the Bahia State [4,5]. Most recently, Zika virus (ZIKV) emerged in the Pacific and later in the Americas, causing a public health emergency due to its association with microcephaly in newborns [6–8].
The Ae. aegypti mosquito is more frequently observed in highly urbanized areas. It is extremely well adapted to live in close association with human dwellings. Females blood feed preferentially on humans and lay eggs on man-made containers often located on the surroundings of residences [9–12]. Conversely, Ae. albopictus is more frequently collected in wooded areas next to humans and tends to bite on a variety of vertebrates outdoors [13]. Both species overlap their spatial distribution in suburban areas, especially in those districts with high vegetation coverage [11,14–16]. Thus, eventually, Ae. aegypti and Ae. albopictus share the use of the same breeding sites, which triggers a series of ecological interactions due to the limited resources available.
Many studies have investigated the negative outcomes of competing environment on adult life history traits. Inadequate nutrition during the larval stage of mosquitoes can be associated with reduced wing size, shorter longevity and flight performance, higher susceptibility to arboviral infections and replication under laboratory conditions [17–21]. Density-dependent competition in larval stages causes mortality and thus reduced recruitment to the adult stage, showing that Ae. aegypti vectorial capacity is strongly dependent on the larval habitat quality.
Effective vector control activities are used as the primary approach to mitigate arbovirus transmission, especially in the absence of vaccines. Ae. aegypti control still relies massively on source reduction and on using chemicals such as insecticides. However, maintaining infestation level below a theoretical threshold to avoid outbreaks requires a constant and somehow utopic military discipline of field health agents over time [22]. Moreover, the overuse of insecticides favors the dissemination of alleles that confer resistance among wild Ae. aegypti populations, jeopardizing insecticide efficiency as tools for vector control [23,24]. Due to the low capacity of traditional control measures to reduce mosquito populations, new approaches to mitigate transmission must be tested.
One of the innovative approaches currently being tested is the deployment of the maternally inherited endosymbiont Wolbachia pipientis into wild mosquito populations [25]. The use of Wolbachia as a natural control agent is supported by findings showing that Ae. aegypti females infected with the wMel strain are able to block DENV, CHIKV and ZIKV [26–28]. Accordingly, Wolbachia releases may be used to mitigate arbovirus transmission by two different strategies: suppression of mosquito population by massive male-releases or the substitution of a highly susceptible population by one that blocks arbovirus transmission [29,30].
Wolbachia deployments are taking place in five countries, including Brazil (www.eliminatedengue.com). When Wolbachia-infected mosquitoes are released, females tend to behave as their wild counterparts, i.e., will blood feed on local householders and lay eggs in the available breeding sites. Considering that in Rio de Janeiro city there is a strong co-occurrence of Ae. aegypti and Ae. albopictus [11, 16], females carrying Wolbachia will eventually lay eggs in containers already colonized by wild Aedes sp. mosquitoes. Therefore, our main objectives were to estimate the deleterious effects of larval competition on mosquito life history traits, but also to determine to what extent larval competition of wMel-Ae. aegypti mosquitoes with wild Ae. aegypti and Ae. albopictus may jeopardize Wolbachia invasion.
We used three different mosquito populations: Ae. aegypti infected with Wolbachia (wMelBr strain), wild Ae. aegypti and wild Ae. albopictus.
The lineage of Ae. aegypti with wMel was imported from Australia to Brazil (IBAMA license 11BR005873/DF). Briefly, a backcrossing with 250 virgin females (wMel) and 200 wild males was conducted for nine consecutive generations, producing wMelBr [31]. After that period, the wMelBr colony was outcrossed every five generations with 10% wild males from a pool of four districts (Jurujuba, Tubiacanga, Urca and Vila Valqueire) with high nuclear genome homogeneity across Rio to refresh the genetic pool [32]. We used the F19 of wMelBr generation.
Wild Ae. aegypti and Ae. albopictus mosquitoes were obtained from four districts in Rio de Janeiro city (the same districts with which the wMelBr colony was outcrossed) by collecting eggs laid on the wooden paddle of ovitraps. A total of thirty ovitraps were installed uniformly in each area, of approximately 1 km2, to represent the genetic variation of the wild population. Wooden paddles were brought to Fiocruz, eggs were hatched and larvae were classified using taxonomic keys [33]. Larvae of each species were pooled and reared in dechlorinated water and fed with TetraMin (fish food), maintained in a climate controlled insectary, at 26 ± 1°C and 70 ± 10% relative humidity. Adult females were kept under a 12:12 hour light:dark cycle, ad libitum access to sugar solution (10%) and blood fed twice a week using anesthetized mice (CEUA L-0007/09). Eggs were stored under insectary conditions until the experiment. We used the F1 of wild Ae. aegypti and F2 of Ae. albopictus.
The larval competition experiments were performed in a semi-field setting, an open building located at the Army Institute of Biology in Rio de Janeiro, Brazil (22°53’34”S, 43°14’33”W), but with limited control access to unauthorized personnel. The experiment was subject to the influences of climate variation, such as humidity and air temperature, as well as rainfall. These conditions were continuously recorded by means of a weather station (Instrutemp, ITWH model 1080) installed on site.
The intraspecific larval competition of Ae. aegypti (with and without wMelBr) and interspecific competition with Ae. albopictus were investigated by monitoring the development of larvae at different densities, species proportions, and food levels in containers. Twenty-one treatments were set and used different proportions of wild Ae. aegypti: wMelBr-Ae. aegypti: Ae. albopictus (20:0:0, 40:0:0, 60:0:0, 0:20:0, 0:40:0, 0:60:0, 0:0:20; 0:0:40, 0:0:60, 20:0:20, 30:0:30, 20:0:40, 40:0:20, 0:20:20, 0:30:30, 0:20:40, 0:40:20, 20:20:0, 30:30:0, 20:40:0, 40:20:0). The densities evaluated herein were based on that of Braks et al. (2004) [19] and represent larval crowding in nature [12,34]. Larvae were placed as L1 in 400 ml of tap water into black plastic containers (9.5 cm in height, 8.5 cm base diameter). Each treatment had eight containers with two distinct food levels. The food consisted of 0.25 or 0.50 g of Chitosan (an analogue of insect chitin used to simulate the remains of arthropods), and fallen avocado leaves (extra source of natural nutrition commonly used in Aedes competition assays), in the same proportion, that were collected, washed, dried, broken into small pieces and weighed. Therefore, our experimental design consisted of 200 plastic containers (12 cm in diameter x 15 cm in height).
Each container was identified and received the appropriate quantity of Chitosan and leaf litter, with 400 ml of tap water, three days before the addition of larvae. Containers were covered with black tulle to prevent oviposition by wild mosquitoes. One hour after eggs were hatched, larvae were counted with the help of a stereo microscope and then added to their appropriate containers. Each container was monitored daily for the presence of pupae, which were collected and placed in small covered vials (6.5 cm height x 2.5 cm diameter) and kept until adult emergence. On the day of emergence, adults were killed with acetyl acetate and, after being sexed, one wing was removed. Wing length was defined as the distance from the axillary incision to the apical margin excluding the fringe [35]. The experiment ended when the last pupa became adult.
In treatments with the presence of wMelBr and wild Ae. aegypti simultaneously, all adults and dead pupae were screened for Wolbachia. Screening was performed using the Taqman multiplex Real Time—Polymerase Chain Reaction. Adult mosquitoes and dead pupae were individually screened on ViiA7 Real Time PCR machine (Life Technologies). The genomic DNA was extracted using a squash buffer (0.1 M NaCl; 10 mM Tris Base; 1 mM EDTA; pH 8.2) supplemented with 9 μg of Proteinase K per mosquito (Qiagen). After macerating the mosquitoes with a 2mm glass-bead on a Mini-beadbeater (Biospec Products), samples were placed on a thermocycler following the thermal cycle: 56° C for 5 minutes and 98° C for 15 minutes. Genomic DNA was diluted 1:10 in ultra-pure water and then used as the template for Wolbachia screening. We used the WD0513 gene that amplifies a fragment of 110 bp with the following primers: TM513-Forw: CAA ATT GCT CTT GTC CTG TGG and TM513-Rev: GGG TGT TAA GCA GAG TTA CGG and TM513-probe 5’-/FAM Cy5/ TGA AAT GGA AAA ATT GGC GAG GTG TAG G -–BHQ-1/-3’. In the same reaction, a ribosomal gene from Ae. aegypti that amplifies a fragment of 68 bp was analyzed with the primers: RPS17-Forw: RPS17-Forw: 5’- TCC GTG GTA TCT CCA TCA AGC T- 3’ and RPS17-Rev: 5’- CAC TTC CGG CAC GTA GTT GTC- 3’, and RPS17-probe: 5’-/FAM/CAG GAG GAG GAA CGT GAG CGC AG/3BHQ_1/-3’. Negative and positive controls of Ae. aegypti (with and without wMelBr) and Ae. albopictus were used in all reactions.
Reagents used in the qPCR were: 5 μL of TaqMan Universal PCR Master Mix (Thermo Fisher), 0.5mM of RPS17 primers, 0.6mM of TM513 primers, 0.1mM of RPS17 probe, 0.25mM of TM513 probe and 1μL of diluted DNA. Water was added to complete a final volume of 10 μL.
Three biological aspects were observed throughout the experiment: larval survivorship, developmental time and wing length. Survivorship was calculated, for each container, by the frequency of larvae that reached the adult stage. Developmental time per container was calculated as the average number of days from hatching until the emergence of the adult was observed in the plastic vial.
An important parameter in population ecology is the performance index λ’, related to the growth rate r’ by λ’ = exp(r´). We calculate λ’ using values of observed biological aspects, such as survival of immature, development time and adult size of cohorts of mosquitoes, for each replicate. An estimate of the performance index has been adapted by [36] from the equation established by [37] using r’ as a measurement of population growth. According to this index, the condition λ′ > 1.0 represents an increase in the population, whereas condition λ′ < 1.0 points to a population decrease. The λ’ index was calculated for each replicate as follows:
r′=ln(1N0∑xAxf(w¯x))D+∑xxAxf(w¯x)∑xAxf(w¯x),
where N0 is the initial number of females in a cohort, which we assumed to be 50% of the added larvae, since the sex ratio of the species studied here is generally 1:1 [38,39]; Ax is the number of adult females on day x; wx is the average size of the female wing on day x; fecundity of females is modeled by a function ƒ (wx) of the wing size, as proposed for Ae. aegypti [40] and Ae. albopictus [41]. No significant differences in fecundity have been found due to Wolbachia infection [31], thus we assumed the same relationship between mosquito size and fecundity for infected and uninfected mosquitoes. D is the time required (in days) for a newly hatched female to mate, blood feed and lay eggs. In our experiments, D is typically equal to the number of days that a female takes to reach the adult stage plus four days, the length of the first gonotrophic cycle [42].
The effects of competition conditions on the performance of Ae. aegypti infected with wMelBr were analyzed using a Generalized Linear Model (GLM). Development time, wing length, survival proportion and the performance index were analyzed each separately as outcomes using as explanatory variables: species, nutrients, competing numbers of wild Ae. aegypti, Ae. albopictus, and Ae. aegypti with wMelBr. For development time, performance index, and wing length we used a normal distribution and logarithmic link function. For survivorship we used logistic regression models with a binomial family/logit link function. For each of the outcomes we selected the model with lowest Akaike Information Criterion (AIC). P-values lower than 0.05 were considered significant. We used R 3.0.1 software for these analyzes.
The index values of λ' were used to make a model to simulate the impact of different levels of infestation of Ae. aegypti wild type and Ae. albopictus in the performance of Ae. aegypti with wMelBr in larval competition. Three nonlinear regressions were applied to each of the indices λ' for the three populations: wild Ae. aegypti, Ae. albopictus and Ae. aegypti with wMelBr, with the number (nx) of individuals in each cohort x (aeg for wild Ae. aegypti, albo for Ae. albopictus, wmel for Ae. aegypti with wMelBr) in competition, according to the following model: log (λ’) ~ naeg + nalbo + nwmel. These analyses allowed us to evaluate the effect on the performance index when increasing both interspecific and intraspecific competition. The values obtained in the regressions were used to simulate the interspecific competition among the three populations.
Once coefficients for interspecific competition were obtained we evaluated the intensity of interspecific competition that makes the growth rate negative, i.e., r´ = log (λ’) < 0. For instance, if population of Ae. aegypti with wMelBr suppresses the wild Ae. aegypti population, this permitted us to evaluate the frequency of Ae. albopictus that causes a severe interspecific competition that might compromise sustained growth of Ae. aegypti with wMel. In this case, we find value for r´wMel = log (λ’wMel) = βwmel + αwmel nalbo + γwmel nwmel < 0, where β, αwmel, and γwmel are coefficients obtained in the regression analysis.
The use of anesthetized mice to blood feed mosquitoes was authorized by Fiocruz Ethical Committee for Animal Use (CEUA L-0007/09), which follows the National guidelines for the scientific use of animals disposed on the Law 11.794/2008.
Under intraspecific competition, survival was inversely proportional to larval density in the three tested populations, as expected (Ae. albopictus: t = -27.2, P<0.05, Ae. aegypti: t = -28.3, P<0.05, Ae. aegypti with wMelBr: t = -26.3, P<0.05). Ae. albopictus presented higher tolerance for increasing competition than wild Ae. aegypti and Ae. aegypti with wMelBr. On the other hand, Ae. aegypti presented a significant decrease in survivorship when larval density per container doubled. This pattern was observed independently of Wolbachia presence (Fig 1, Table 1).
Under interspecific competition, the survival of Ae. albopictus and Ae. aegypti with wMelBr larvae was significantly higher than survival of wild Ae. aegypti, irrespective of the amount of food resources (Table 1). Nonetheless, competitive advantage of Ae. albopictus over wild Ae. aegypti seemed slightly more evident in the most stressful and competitive treatments. Ae. aegypti with wMelBr also survived less than Ae. albopictus, although its survival is marginally higher than that observed for wild Ae. aegypti. In some particular treatments, Ae. aegypti larvae infected with wMelBr presented better survival than wild Ae. aegypti.
Under intraspecific competition, overcrowding was directly related to the increase in developmental time (DT) (P<0.05). Wild Ae. aegypti and those infected with wMelBr have a longer DT starting at 40 larvae per container, while Ae. albopictus DT was notably affected only at a higher density, i.e., 60 larvae per container (Fig 2, Table 2). Wild Ae. aegypti presented the longest DT at high densities with an average duration of 57.17 and 47.99 days in low and high food resources, respectively. Under interspecific competition, Ae. aegypti with wMelBr had a similar DT to Ae. albopictus at high food resources, but was outcompeted when food resources were scarce. Interestingly, Ae. albopictus presented a shorter DT than wild Ae. aegypti (t = 4.33, P<0.05), but the presence of Wolbachia did not alter the DT in Ae. aegypti.
The three populations had a significant decrease in wing size due to overcrowding (Table 3, Fig 3). Ae. albopictus had a sharper decrease than wild Ae. aegypti and Ae. aegypti with wMelBr. The presence of Wolbachia did not seem to influence mosquito wing size. The amount of nutrients had a positive effect for wild Ae. aegypti and a negative effect for Ae. albopictus.
Overall, overcrowding had a significant effect on the performance of the three populations (Fig 4, Table 4). The value of λ' for wild Ae. aegypti and Ae. aegypti with wMelBr suffered a reduction from 1.2 to 1.0 when the larval density was doubled (treatment 2). On the other hand, the λ' for Ae. albopictus was only reduced to 1.0 when the larval density was tripled (treatment 3). Remarkably, all populations tested were able to maintain λ' above 1 under low densities, meaning that they could be successfully sustained in the wild. Under these experimental settings, Ae. albopictus showed superior performance to wild Ae. aegypti and the presence of Wolbachia did not seem to affect Ae. aegypti performance (Fig 4, Table 4).
We simulated interspecific competition among the three different populations applying the results from the nonlinear regression analyzes for each of the Ae. aegypti, Ae. albopictus, Ae. aegypti with wMel performance indices. As expected, increasing either intraspecific or interspecific competition makes the performance index smaller for the three populations. In Fig 5, when performance indices reach values below lines for λ = 1, the interspecific competition does not allow a positive growth rate. We generally observe that the values under which Ae. albopictus can sustain a positive growth rate are larger than values for both Ae. aegypti and Ae. aegypti with wMel populations.
We also studied the frequency of Ae. albopictus that could make interspecific competition intense enough to make Ae. aegypti/wMel performance index λwmel<1, i.e. impacting severely on the sustained growth assuming only larval competition (Fig 6). As the number of Ae. aegypti/wMel larvae increases (intraspecific competition), the frequency of Ae. albopictus that causes the performance index to reach an unsustainable level λwmel<1 decreases.
The deployment of Wolbachia to reduce dengue transmission is currently being undertaken in several regions of the world. During releases, Ae. aegypti mosquitoes infected with Wolbachia will lay eggs in breeding sites in which wild mosquitoes previously colonized, i.e., intra-specific competition with local Ae. aegypti and other native mosquitoes such as Ae. albopictus might be an important issue to determine the pace of Wolbachia invasion. Due to the co-occurrence of Ae. aegypti and Ae. albopictus in several countries of Southeast Asia and Latin America, we investigated how the intra- and interspecific competition with Ae. albopictus might undermine wMel invasion.
We explored three critical aspects of mosquito biology under different competition scenarios: larval survivorship, developmental time and wing length. Using these estimates, we calculated a performance index that is related to growth rate for wild Ae. aegypti, Ae. albopictus, and Ae. aegypti with wMelBr.
Overcrowding significantly reduced larval survivorship in the three populations tested, as expected [18,19,43–45]. Our data show that under intraspecific competition settings, larval survivorship decreased more intensely for Ae. aegypti than for Ae. albopictus. Also, the presence of Wolbachia did not affect this pattern (Fig 1). The effects of Ae. albopictus density on Ae. aegypti mortality and vice-versa has been evaluated elsewhere [19,34,46,47]. In summary, one of these populations gets more severely affected when the density of the other increases, and this increase in mortality might be seen in the larval or adult stage [48,49]. Ae. albopictus is frequently pointed as a better competitor than Ae. aegypti [19,34,36,46,50,51] as well as other species such as Ae. japonicus [52] and Culex pipiens [53]. Larvae survival under interspecific competition conditions may vary due to the difference in efficacy with which larvae exploit food resources [54]. Ae. albopictus showed a larval survivorship higher than that observed for both wild and infected Ae. aegypti. This finding strongly suggests Wolbachia has limited role in affecting larvae mortality under competitive scenarios.
The wMel strain has relatively mild effects on mosquito fitness [25,55], but interestingly, a superior survivorship of infected larvae was reported in a competitive environment when compared with uninfected larvae [39]. On the other hand, Wolbachia infection reduced the tolerance of Ae. aegypti larvae to starvation, probably due to an increasing rate of depletion of accumulated energy reserves [45]. Our data support an overall beneficial impact of Wolbachia infection on Ae. aegypti larval survivorship, since infected larvae present a superior survival rate than their uninfected counterparts (the exception being observed in starvation scenarios, in which Wolbachia reduced larval survivorship).
The development time from egg hatch to adult is a critical fitness aspect of mosquito biology under field conditions. Delayed larvae are more subject to external factors such as predation [56], water evaporation and breeding site treatment or removal. Our results show that nutrient levels caused longer development time for all three populations when food resources were scarce [19,48,57]. Ae. albopictus had a rapid development time when compared with Ae. aegypti [11] even in the more competitive treatments. The presence of Wolbachia did not accelerate Ae. aegypti developmental time. Differences in larvae development time were highest in treatments with interspecific competition with 0.25 g of litter. The exception was the Ae. albopictus/Ae. aegypti with wMel, in which larvae of both populations have distinct development time differences only at 0.50 g litter, with the former developing faster than the latter.
Overall, results regarding the influence of Wolbachia on larval development time are conflicting. The experimental design and settings established in our competitive assays were unable to detect any changes on Ae. aegypti development time due to Wolbachia presence. At intermediate (50 larvae/tray) and high densities (250 larvae/tray), wMel infection led to more rapid larval development for both males and females, with no effect under a less crowded and more stressful condition [28]. Opposing results to our data, a slight delay was observed in wMel infected larvae related to their uninfected counterparts [39]. Despite these findings in disagreement, few strains of Wolbachia are known to modify adult feeding behavior, and might interfere with larval foraging capability as well [26,56,58,59]. Potential explanations for the effects of Wolbachia on mosquito larval development time involve immune up-regulation or increased metabolism observed in the adults, which may also influence larval development rate [26,60]. Other aspects still need an evaluation to better understand the effect of Wolbachia on immature development time, such as the effects of the population genetic background, Wolbachia strain and experimental design.
Mosquito body size is ultimately a manifestation of larval habitat quality and can produce significant effects on an insect’s fitness and then alter mosquito vectorial capacity [12,61,62]. Physiological stress in juvenile stages produces negative effects that may pass into adulthood [63]. For instance, highly competitive environments produce mosquitoes with a small wing length, which are less likely to promote Wolbachia invasion since they should blood feed more often, possess shorter longevity and lower flight performance than bigger mosquitoes [18–21]. Hence, Ae. aegypti vectorial capacity is strongly dependent on the larval habitat quality [11,17–21]. Previous reports have shown an inversely proportional correlation between wing size and larval density in Ae. aegypti, as we observed [28,45]. Our results indicate reduction in mosquito size due to overcrowding in all three populations, which is highly expected [19, 49]. Ae. albopictus wing size was consistently smaller than Ae. aegypti in almost all treatments, with visible differences when competition was intra or interspecific. The interaction between nutrients and population produced unexpected results. Ae. albopictus body size decreased at 0.5 g litter when compared with the 0.25 g treatment. Interestingly, body size of Ae. aegypti with wMel was not affected by availability of food resources. Hence, from the perspective of Wolbachia deployment, the infection with wMel strain does not pose a significant disadvantage during competition against wild mosquitoes [39].
We used three population growth correlates, i.e. larval survivorship, time to adulthood and adult wing size to estimate a composite index of mosquito performance (λ') for each container [36,37]. Overall, larval density negatively affected the performance index λ' of the three populations, but remarkably only Ae. albopictus population growth was positive in all treatments. In fact, population growth of Ae. albopictus was significantly superior from the observed for wild Ae. aegypti, while the presence of Wolbachia provided no advantage to infected Ae. aegypti. Interspecific assays using Ae. aegypti and Ae. albopictus at different densities have shown a superior competitive ability of the the latter [19,34,36]. Despite being frequently described as a superior larval competitor to Ae. aegypti, these two species coexist in much of Brazil and in southeast US and Southeast Asia [16,64]. Part of the explanation for coexistence may rely on life-history trade-offs and abiotic factors [40,65–67] Coexistence between Ae. aegypti and Ae. albopictus may be possible due to dry and warm climates that would favor the former and mitigate effects of larval competition via differential mortality of Ae. albopictus [67]. This hypothesis was reinforced by Camara et al. (2016) [34], observing that intensity of competition at the larval stage may vary seasonally, with harsh effects on development time during warmer Summer. Abiotic factors may also contribute to habitat segregation since urbanized areas tend to be warmer than arborized surrounding areas [68]. Additionally, one force that can impact Ae. aegypti/wMelBr invasion is the asymmetric reproductive interference among mosquitoes, in which male Ae. albopictus can inseminate and thus sterilize Ae. aegypti females. The act of reducing the reproductive success of a different species by mating a female of an incompatible species is called satyrization [69–71]. Evidence of satyrization of Ae. aegypti females seems to be more likely than on Ae. albopictus females, although still low (less than 5%), biasing the asymmetric nature of cross matings in favor of the latter [72–74]. Therefore, although still not observed in Brazilian sites where Wolbachia deployment is ongoing, additional concern would be required if invasion is lagging.
The major force that can affect Wolbachia invasion is the population density of wild mosquitoes [75,76]. This concern is even more important if we consider that mosquitoes from other species can lay eggs in the same breeding sites of Ae. aegypti. Therefore, during Wolbachia deployment, infected mosquitoes will lay their eggs in breeding sites already colonized by local mosquitoes, such as uninfected Ae. aegypti and Ae. albopictus. Assuming Ae. albopictus is a better competitor and the presence of Wolbachia does not increase mosquito performance at the larval stage, the natural density of Ae. albopictus may become an additional obstacle to slow invasion. However, we observed a negative growth rate of Ae. aegypti/wMelBr only when Ae. albopictus frequency was high. In Rio de Janeiro, we selected four neighborhoods with different landscapes and performed adult mosquito collections with BG-Sentinel Traps installed at the peridomestic area of local householders on a weekly basis for 104 consecutive weeks [31]. We observed that the frequency of Ae. albopictus was lower than 5% in the four sites during the 104 weeks of trapping. In fact, during approximately four consecutive months, no Ae. albopictus mosquitoes were collected in any trap from any field site (Eliminate Dengue Program). Therefore, Ae. albopictus is more likely to slow down Wolbachia invasion, rather than to stop it. Density-dependent traits can promote strong effects on Wolbachia dynamics in Ae. aegypti field populations [77]. Therefore, an estimation of the population sizes of Ae. aegypti and other mosquito populations that can occasionally lay eggs in the same breeding sites, such as Ae. albopictus, Culex quinquefasciatus and Limmatus durhami, might provide important information on the Wolbachia invasion pattern in highly infested field sites [31,76–80].
|
10.1371/journal.pgen.1003874 | A Novel Highly Divergent Protein Family Identified from a Viviparous Insect by RNA-seq Analysis: A Potential Target for Tsetse Fly-Specific Abortifacients | In tsetse flies, nutrients for intrauterine larval development are synthesized by the modified accessory gland (milk gland) and provided in mother's milk during lactation. Interference with at least two milk proteins has been shown to extend larval development and reduce fecundity. The goal of this study was to perform a comprehensive characterization of tsetse milk proteins using lactation-specific transcriptome/milk proteome analyses and to define functional role(s) for the milk proteins during lactation. Differential analysis of RNA-seq data from lactating and dry (non-lactating) females revealed enrichment of transcripts coding for protein synthesis machinery, lipid metabolism and secretory proteins during lactation. Among the genes induced during lactation were those encoding the previously identified milk proteins (milk gland proteins 1–3, transferrin and acid sphingomyelinase 1) and seven new genes (mgp4–10). The genes encoding mgp2–10 are organized on a 40 kb syntenic block in the tsetse genome, have similar exon-intron arrangements, and share regions of amino acid sequence similarity. Expression of mgp2–10 is female-specific and high during milk secretion. While knockdown of a single mgp failed to reduce fecundity, simultaneous knockdown of multiple variants reduced milk protein levels and lowered fecundity. The genomic localization, gene structure similarities, and functional redundancy of MGP2–10 suggest that they constitute a novel highly divergent protein family. Our data indicates that MGP2–10 function both as the primary amino acid resource for the developing larva and in the maintenance of milk homeostasis, similar to the function of the mammalian casein family of milk proteins. This study underscores the dynamic nature of the lactation cycle and identifies a novel family of lactation-specific proteins, unique to Glossina sp., that are essential to larval development. The specificity of MGP2–10 to tsetse and their critical role during lactation suggests that these proteins may be an excellent target for tsetse-specific population control approaches.
| Tsetse flies are the sole vector for African trypanosomes, causative agents of sleeping sickness in humans and nagana in cattle. Transcriptome and proteome analyses were utilized to examine the underlying mechanisms of tsetse lactation that occur during each reproductive cycle. These analyses revealed a dramatic shift to the synthesis of milk proteins during lactation and a novel milk-specific protein family. All members of this family were co-localized, shared sequence similarity and were expressed at 40× basal levels during milk secretion. Suppression of gene from this lactation-associated family impaired progeny development by reducing milk protein content and altering milk homeostasis. These novel genes represent an excellent target for tsetse-specific reproductive-based control mechanisms. In addition, the characterization of tsetse milk production revealed multiple factors that are functionally analogous between tsetse and mammalian lactation.
| Tsetse reproductive biology is unusual among insects. Female tsetse give birth to a fully mature third instar larva (viviparity) after an extended intrauterine gestation. This reproductive strategy limits the capacity of tsetse mothers to only 8–10 offspring per lifetime [1]. To accommodate intrauterine larval development, the morphology and physiology of the female tsetse reproductive organs have undergone extensive modification. The reproductive tract has been expanded into a uterus to serve as a safe harbor for developing larvae. Ovarian development alternates between the right and left ovaries to produce a single oocyte during each gonotrophic cycle. The female accessory gland has been modified and expanded to provide milk that is secreted into the uterus and consumed by the developing larva [1]. The distinctive aspects of tsetse viviparity represent significant reproductive bottlenecks that could be exploited for population control. Furthermore, identification of factors specific to milk production could lead to development of novel tsetse-specific compounds that interfere with larval development and induce abortion (abortifacients) without impacting non-target insects.
The nutritional components of tsetse milk consist mainly of proteins and lipids emulsified in an aqueous base [2]. In total, 6–10 mg of nutrients (combined with 10 mg of water) are transferred to the larva in the milk during intrauterine development. Few studies have examined regulation of tsetse milk production, including an investigation of structural changes in the milk gland, radioisotope studies of nutrient movement within the mother during lactation, and direct examination of specific milk proteins [1], [3]–[9]. To date, six milk proteins have been characterized, including Transferrin [7], [10], a lipocalin (Milk Gland Protein1, MGP1 [6], [11]), two unknown milk proteins (MGP2–3; [12]), Acid Sphingomyelinase 1 (aSMase1; [9]) and Peptidoglycan Recognition Protein-LB (PGRP-LB, [13]). Furthermore, we recently showed that lipid metabolism is governed by the cooperative activity of insulin and juvenile hormone signaling pathways during the pregnancy cycle [14]. However, the full suite of proteins present in the milk and underlying mechanisms for their regulation during tsetse lactation and pregnancy have yet to be determined.
In this study, a satellite paper to our report on the whole genome sequence of the tsetse species Glossina morsitans morsitans [15], we used differential RNA-seq analyses to compare transcript abundance in females carrying an intrauterine larva (lactating) with females 24–48 hours after parturition (non-lactating or dry). The lactation period occurs during larvigenesis, while the dry period occurs over the course of oogenesis and embryogenesis [14]. In addition to transcriptome analysis, protein constituents of tsetse milk were identified through LC/MS/MS analyses of gut contents from nursing larvae. We describe the expression profile, the predicted structure based on an in silico approach, and the microsyntenic genomic organization of nine tsetse-specific milk proteins (MGP2–10) that we propose represent a highly divergent, novel protein family. siRNA-based knockdown analysis was employed to examine the functional roles of the MGP2–10 proteins during tsetse reproduction. Since MGP2–10 are tsetse-specific and have substantial influence over tsetse fecundity, we discuss their potential for exploitation in novel population reduction approaches. Lastly, we discuss our findings in light of milk secretions described from other lactating organisms.
To understand the major products of lactation and factors that may be responsible for regulating their expression, we analyzed two RNA-seq libraries. The first library represents lactating females carrying an early third instar intrauterine larva, while the second represents dry females collected approximately 48 hours post-parturition, at which time they had an early embryo developing in the uterus and lactation has yet to commence. In total, over 42 million reads from each of the two libraries were recovered (Table 1). Overall read quality was high for both sample sets based on FastQC analysis. Removal of contaminating tsetse symbiont (Wigglesworthia, Sodalis and Wolbachia) specific sequences and cleanup resulted in a 4% reduction in the total number of sequences identified in lactating flies and a 5.1% reduction for dry flies, respectively (Table 1).
De novo assembly of the two datasets by Abyss [16], [17] and Trinity [18] generated 42,935 contigs that were subsequently identified according to multiple search parameters (Table S2). There were 34,674 contigs at least 200 bp in length, with the longest contig measuring 24,573 bp in size (Fig. S1). Distribution of reads per contig was comparable between the two datasets with the exception that there was a greater number of highly expressed genes in lactating flies (Fig. S2). Comparative analyses of contigs revealed that most were more highly expressed in dry flies compared to their lactating counterparts. A total of 297 contigs (2.1%) with at least 50 mapped transcriptome reads showed elevated expression in lactating versus dry flies (Fig. 1a; Table S2). Only 1311 were expressed at statistically different levels between the two datasets, with 48 contigs (4.2%) more highly expressed in lactating flies (Fig. 1b; Table S3). Classification of the lactation-expressed contigs based on specific metabolic and structural functions revealed enrichment for lipid metabolism, transport and storage, protein synthesis, secreted proteins, and those of unknown function (Fig. 2a).
Examination of the most highly expressed contigs in lactating flies revealed known milk protein genes along with novel transcripts not previously associated with lactation. The known milk protein genes (mgp1, mgp2–3, tsf and asmase1) were expressed at least 10-fold higher in lactating versus dry flies (Fig. 3a–c; Table S2, S3). Of particular interest was the discovery of a group of seven new genes similar to the previously identified mgp2–3 genes that were upregulated during pregnancy (mgp4–10; Fig. 3b; Table S2). Transferrin and aSMase1, involved in iron transport and sphingolipid metabolism respectively, were the only other proteins that were highly expressed and showed increased transcript abundance in lactating flies (Fig. 3c; Table S2). Recently, the immunoregulatory Peptidoglycan Recognition Protein LB (PGRP-LB) was also detected in tsetse milk [13]. Based on this analysis, PGRP-LB expression patterns are different from those of other lactation associated-proteins, as its expression did not increase throughout pregnancy (Table S2). In addition to the genes described above, specific ribosomal RNAs were significantly elevated in lactating flies (Fig. 3d; Fig. S3), and may account for the overall increase in contigs coding for genes involved in protein synthesis (Table S2). Confirmation of transcript abundance during lactation was achieved by qPCR analysis of the mgp1–10, 28S, transferrin, pgrp-lb and asmase1 genes (Table 2; Text S1).
The majority of the contigs (94%) were more abundant in dry (non-lactating) compared to lactating flies (Fig. 1a; Table S2). Multiple gene families were highly expressed in dry flies (Fig. 2a, Table S2). Contigs encoding heat shock proteins and antioxidant enzymes were increased in dry flies, indicating that dry flies may be better suited than their lactating counterparts to respond to stress and environmental insult (Fig. S4). In particular, qPCR analysis validated the transcriptome data for Cu/Zn superoxide dismutase and catalase, which encode proteins that remove reactive oxygen species to prevent damage (Table 2). Lipid metabolism contigs were more abundant in dry flies with the exception of Brummer lipase, which was only two-fold higher than in lactating flies (Fig. S4; Table S2). Expression of tsetse yolk proteins, yolk protein 1–3 (yp1–3) was also higher in dry flies, reflecting the yolk protein synthesis that occurs between bouts of lactation (Fig. S4; Table S2). Contigs identified as trypsin showed decreased transcript abundance in lactating flies (Table S4). Given that the transcriptome analysis was from whole females, this finding likely correlates with lactating females' smaller bloodmeals that result from limitations on abdominal space imposed by developing intrauterine larva [5]. These results suggest that many processes in tsetse mothers are down regulated during lactation ( = higher in dry flies), when the female devotes energy and resources to synthesize milk-associated proteins to nourish the developing larva.
We conducted a secondary RNA-seq analysis after removing reads that mapped directly to the twelve most abundant lactation-specific genes (asmase1, mgp1–10 and transferrin). Removing these reads yielded only a 2.6% reduction in the dry fly dataset, but a drastic reduction of 47.2% was observed in the lactating fly dataset (Fig. 4a). This difference suggests that lactating flies invest over 47% of their total transcriptional activity toward producing the main protein constituents of tsetse milk (Fig. 4a). Each milk-specific gene accounted for 1.4–6.7% of the total read count in lactating flies, with the most reads mapping to mgp10 and transferrin (Fig. 4a). This removal resulted in a total of 2238 genes that were more highly expressed in lactating flies, but with only 151 that were significantly elevated relative to dry flies (Fig. 4b,c). Assignment by metabolic category resulted in a more balanced distribution of highly expressed contigs in lactating and dry flies (Fig. 5a). This second analysis revealed a few additional contigs whose expression increased during lactation; their expression was previously overshadowed by highly expressed milk-specific genes (Table S5; Table S6). These included dawdle (an activin signaling molecule), glyoxylate/hydroxypyruvate reductase (an enzyme that converts glycerate to hydroxypyruvate), choline-phosphate cytidylyltransferase (an enzyme in the Kennedy pathway that catalyzes choline phosphate to CDP-choline) and multiple ribosomal proteins (Fig. 5b, Tables S5, Table S6).
Using LC/MS/MS analyses on the gut contents of feeding larva, we identified 155 proteins that may be constituents of tsetse milk. Most of these proteins have a low exponentially modified protein abundance index (empai) value and are likely present in milk at low levels or may be products from the larval gut (Table S7). Most of the highly abundant proteins identified in tsetse milk were products of genes identified as highly expressed during lactation in our transcriptomics analysis, including MGP1–10, Transferrin and aSMase1 (Table 2). Previously, PGRP was documented in tsetse milk [13] and we confirmed the presence of this immune protein in the milk proteome (Table 2). In addition to the transcriptionally-abundant proteins, the milk proteome identified three other abundant proteins (empai >1.2; Table 2). These three proteins include a sterol binding protein (Niemann-Pick C-2g, NPC2G), Ubiquitin Associated and SH3 Domain Containing A (UBASH3A, a protein belonging to the T-cell ubiquitin ligand, TULA, family [19]), and a putative tsetse protein with unknown function (GmfB8). Transcript levels for NPC2G, GmfB8-like protein and UBASH3A were measured in the milk gland/fat body fraction during and after lactation and in the larval gut to confirm whether these are generated by the milk gland or if they are products of the gut (Fig. 6). Both npc2g and gmfb8-like protein expression were detected at high levels in the larval gut. Transcript level for UBASH3A was higher in the milk gland/fat body, showing an expression profile similar to PGRP-LB (Fig. 6), suggesting that this is likely a low abundance protein generated by the milk gland during lactation. These results provided further validation for our transcriptome-based identification of milk protein genes as actual secreted products in tsetse milk. In addition, we recovered potential milk proteins that are not under extensive transcriptional regulation during lactation.
BLASTx searches of the NCBI nucleotide collection failed to recover orthologous sequences to the MGP2–10 from any organism. Partial gene sequences encoding MGPs were identified from two other tsetse species, Glossina fuscipes fuscipes (MGP2, 5) and Glossina pallidipes (MGP3, 4), using RT-PCR with degenerate primers (Fig. S5, S6). Mining of sequence data from recent EST projects on the flesh fly, Sarcophaga crassipalpis [20], [21], revealed one sequence with marginal sequence similarity with the tsetse MGP2–10 (Fig. S5, S6). The average number of amino acids for MGP2–10 was 179.3 (range 170–191, Table 3) with a predicted molecular weight of 21.4 kD (range 20.4–22.4 kD; Table 3). The average isoelectric point for MGP2–10 was 6.2 (range 5.9–6.5; Table 3). The aliphatic index, or the relative volume of a protein occupied by aliphatic side chains (alanine, valine, isoleucine, and leucine) that is indicative of the stability of globular proteins [22], is moderate for MGP2–10, ranging from 64–86 (Table 2). There are no predicted glycosylation sites on MGP2–10, but there are at least four predicted phosphorylation sites for each MGP (Table 3). Amino acid alignments of MGP2–10 identified a conserved secretory peptide and three conserved regions with 68–100% and 12–100% nucleotide and amino acid similarity, respectively (Fig. 7a,b; Fig. S5). Phylogenetic analysis shows MGP2 and MGP4 as recently duplicated paralogs sharing 92% amino acid similarity (Fig. 8a; Fig. S6 a,b). Mapping of mgp2–10 coding sequences to genomic scaffolds revealed that these genes localize to a 40 kb microsyntenic region (Fig. 8a). The phylogeny for mgp2–10 splits these genes into two distinct groups, one consisting of mgp2,4,5,6,9,10 and the other of mgp3,7,8. When the phylogeny is mapped against the genome location of mgp2–10, mgp2,4,5,6,9,10 localized with a region surrounded by mgp3,7,8 (Fig. 8a). All mgp genes share a conserved exon-intron structure (Fig. 8a), despite showing varying levels of amino acid sequence similarity amongst them (Fig. 8b; Fig. S5). Our results indicate that MGP2–10 proteins are likely specific to Glossina, but it remains to be seen if evolutionarily-related sequences may exist in other closely-related viviparous genera (i.e. bat flies and sheep ked; data not available for these species). The sequence obtained from the flesh fly may represent either a class of proteins distinct from the tsetse MGP family, or could be highly divergent ancestral sequence to MGP2–10 genes found in tsetse.
Based on amino acid composition, the novel tsetse milk proteins provide all essential amino acids necessary for larval growth and development (Table S8). Protein structure predictions for the novel MGPs were generated by four individual programs. Structural predictions were ab initio as no homologous protein structures were available. The ab initio structure predictions from the four programs revealed that MGPs usually form multiple α-helices (6–10 per protein). However, no functional insights were provided by the I-TASSER program, since MGPs lack structural similarity to other characterized proteins (Fig. S7). Of particular interest, these proteins contain high percentage of hydrophobic amino acids (this study, [12]), including the hydrophobic secretory peptide that was identified in tsetse milk, indicating that this region is not always cleaved during secretion.
Examination of amino acid alignments identified several regions of moderate conservation across MGP2–10 from G. morsitans. To assess the relative selective pressures acting on these paralogous genes, we performed several computational analyses of nonsynonymous-to-synonymous substitution ratios (dN/dS) along the coding sequences for these genes. This type of analysis is usually conducted on orthologous genes in different species or on multiple alleles within a species, but utilization of this on paralogous genes could provide insight into regions critical to their function. When dN/dS substantially exceeds 1, evidence for positive selection ( = adaptive evolution) is inferred. In contrast, dN/dS = 1 implies neutral evolution, while dN/dS values closer to zero provide evidence for negative or purifying selection. Sequences were translated and multiple alignments were performed in ClustalX [23], followed by optimization in BioEdit [24] or MEGA 4/5 [25], [26]. We reverse-translated amino acid sequences to obtain codon alignments as input sequences for dN/dS analyses under both PARRIS [27] and FEL (Fixed Effects Likelihood; [28]) analyses in DataMonkey (www.DataMonkey.org; [29], [30]), a web-based implementation of the HyPhy algorithm [31]. PARRIS allows detection of positive selection across an entire coding sequence, while the FEL method is suitable for detecting positive or negative selection in a site-specific manner in small (10–15 sequences) datasets [28]. Under the PARRIS algorithm, we found no evidence for positive selection across the coding sequence of MGP2–10, suggesting that no residues in these proteins are targets of adaptive evolution. FEL analysis likewise showed no evidence for individual codons subject to positive selection. In contrast, while the preponderance of residues in MGP2–10 are apparently undergoing neutral evolution, FEL analysis indicates that the identified N-terminal secretory signal peptide is largely subject to purifying selection (Fig. 7a,b), suggesting that this region is indispensible for protein function or appropriate intra/intercellular transport. A minority of residues, largely dispersed throughout the C-terminal half of MGP2–10 are additionally subject to negative selection as evidenced by dN/dS ratios significantly less than 1 (p = 0.05, Fig. 7a,b). A role for amino acids under purifying selection outside of the secretory region is unknown. A majority of these conserved sites are proline (33.3%) and phenylalanine (26.6%) residues, suggesting these amino acids may be critical for MGP2–10 folding and/or function.
Previous examination of mammalian milk proteins revealed that discrete, specific sections of each gene are subject to neutral, negative or positive selection [32]. Using a similar MEGA-based analysis to specifically investigate MGP2–10 from G. morsitans, the secretory peptide and conserved region 1 appear to be largely under negative selection (Fig. 7b). The first variable region has a high dN/dS and is likely subject to neutral or positive selection (Fig. 7b), but additional MGP sequences need to be recovered from other Glossina sp. to more confidently determine site- or region-specific selective pressure across MGP coding sequences. Overall, these analyses indicate that only the secretory peptide and the first conserved region are likely subject to purifying selection, but additional analysis will be necessary once full-length MGP genes are recovered from other members of Glossina to establish regions of selection.
Using RT-PCR analysis to determine the tissue specificity of MGP expression, we found that expression of mgp2–10 is specific to the female fat body/milk gland (Fig. 9a). Temporal expression profiles obtained for these genes showed that mgp2–10 transcripts increase dramatically during larvigenesis and then rapidly decline within 24–48 h following parturition (Fig. 9b). This temporal and spatial expression profile is consistent with other characterized milk proteins, including mgp1 (Fig. 9a, [6]), asmase1 [9] and transferrin [7]. The temporal expression profiles for the MGP genes we identified from the two other tsetse species (gpmgp3,4 from G. pallidipes and gfmgp2,5 from G. fuscipes) were similar to those observed for mgp2–10 in G. morsitans. Transcript abundance was lower in teneral females and in females with developing intrauterine embryos, becoming progressively greater through larvigenesis (Fig. S8). In contrast, the MGP-like sequence discovered from the flesh fly, another brachyceran that exhibits larviposition but does not nourish the developing larva, was not expressed in this manner; we observed no differences in MGP-like gene expression between male and female flesh flies (nonpregnant vs. pregnant minus larval expression level; Fig. S8). Thus, even though a gene with moderate sequence similarity was identified in S. crassipalpis, its expression profile is incongruent with tsetse MGPs.
Injection of siRNAs targeting mgp5,7–9 significantly reduced corresponding target transcripts in lactating flies (Fig. 10a). Differences in the knockdown efficiency are likely due to the combined effects of technical variation and slight natural variation in the pregnancy cycle. Suppression of individual transcripts (even the highly expressed mgp7) had no effect on the number of pupae produced per female, the length of pregnancy, or the incidence of pupal emergence (Fig. 10b–d). This suggests functional redundancy among the multiple mgp paralogs, which are all expressed in a similar spatio-temporal manner during pregnancy. Simultaneous suppression of two MGPs (i.e. 5 and 7) reduced the number of pupae deposited per female by 10–15% and extended the duration of pregnancy by 2–3 d, but no difference was observed in the incidence of adult eclosion (Fig. 10b–d). When mgp5,7–9 were co-suppressed, fecundity was reduced by nearly 70% and in cases where mothers produced viable progeny, pregnancy was extended by 6–8 d (Fig. 10b–d). Together, these results suggest that the paralogous mgps share a critical function in tsetse reproduction.
Bradford assay of the milk protein content indicated that simultaneous knockdown of mgp5,7–9 reduced overall milk protein by nearly 22% (0.43±0.06 mg protein/5 µl milk) compared to siGFP treated controls (0.56±0.04 mg protein/5 µl milk). We hypothesized that these proteins may serve a function in maintaining milk lipid stability. To explore this possibility, we assessed the stability of milk emulsification after MGP knockdown utilizing an emulsion stabilization assay. Knockdown of mgp5,7–9 resulted in an increased rate of separation of the aqueous and lipid fractions of the milk by over two-fold (Fig. 10d,e). These findings suggest that the novel MGPs are not only an important amino acid/protein resource for the developing larva, but function to stabilize lipids within tsetse milk, allowing fat to remain homogenously distributed.
Human African Trypanosomiasis (HAT; sleeping sickness) is a fatal disease that affects millions of people in sub-Saharan Africa. A disease caused by closely related parasites in animals, known as African Animal Trypanosomiasis (AAT; nagana), devastates agricultural and livestock production systems as well as land use in Africa. There are no prophylactic drugs that are affordable and highly efficacious, and simple tools for diagnosis or mammalian vaccines for control of HAT or AAT are lacking. Tsetse (Glossina sp.) are vectors of African trypanosomes and reduction of this insect population is a highly effective control strategy. The success of population reduction based control efforts largely is due to the low reproductive capacity of tsetse, which is viviparous yields few progeny per female over their lifespan.
Our main goals in this study were to provide an in-depth characterization of the tsetse lactation process, which is essential for intrauterine larval development and to discover novel biochemical, molecular or physiological targets for tsetse population control through reproductive suppression. Through transcriptomic analysis we identified twelve major tsetse milk proteins (MGP1–10, aSMase1 and Transferrin). Expression of these genes is tightly regulated through the transition between dry and lactating periods in order to optimize resource allocation for milk production. We also analyzed milk collected from larval guts using a proteomics approach both to characterize its composition and to verify secretion and transfer from mother to offspring of the nine milk proteins in addition to Transferrin, aSMASe1 and MGP1. Among the proteins identified as lactation products, MGP2–10 represent a group of secreted proteins unique to tsetse. Using a knockdown approach we showed that tandem suppression of multiple mgp variants resulted in substantial delays in parturition and up to a 70% reduction in fecundity. Likely, this impaired fecundity results both from a lack of protein resources and impaired lipid stabilization in the milk emulsion.
Several of our major findings from this transcriptome analysis included evidence for a substantial shift during tsetse milk production toward secreted proteins, genes involved in lipid metabolism, transporter genes, and specific genes coding for protein synthesis machinery. Structural studies have previously shown that there are extensive arrays of endoplasmic reticulum (ER) that develop in actively secreting milk gland cells, and then degenerate during an involution period following parturition [33]. A similar pattern of increased ER production in the milk gland was documented in another tsetse fly species, Glossina austeni [34] and in Melophagus ovinus, a viviparous sheep ked [35]. The rate of protein synthesis is elevated in the milk gland during periods of lactation, in accordance with an increase in ER [1], [36]–[38]. Increase of milk gland associated ER to allow for production of lactation-associated proteins is likely the reason for high abundance of transcripts for genes involved in protein synthesis during tsetse milk production. The products of mgp1–10 likely constitutes over 95% of the protein content in tsetse milk (this study, [6], [12]), accounting for the increase in contigs for secreted proteins. Transferrin and aSMase1 account for the high read abundance of contigs associated with transporters and lipid metabolism, respectively.
The extreme elevation of asmase1, mgp1–10 and transferrin transcript levels during lactation indicate that the expression of these 12 genes which constitutes less than 0.0005% of the contigs from the de novo library, represents over 45% of the RNA-seq library. In contrast, in dry flies, these same gene transcripts represent less than 2.6% of the library. A heavy investment in specific genes during reproduction is not uncommon [39]–[43], but in most studies the effect was measured directly within a specific organ, rather than the entire organism. As an example, the lactating wallaby invests over 50% of transcript abundance in the mammary gland to the production of protein bound for transfer in the milk [39]. In addition, most milk production transcripts, specifically those directly incorporated into milk, show drastic changes throughout lactation in mice and bovine within the mammary gland [44]–[46]. For tsetse, heavy transcript investment according to total Illumina read levels is based on the entire female fly rather than the isolated milk gland organ. This high investment is not surprising since at least 4–5 mg (20–25% of the total mass of the mother) of proteins are secreted by the milk gland during the 4–5 day period of lactation, representing 40–50% of the nutritional content of the milk [2]. Thus, female tsetse may be uniquely adapted to generate milk with its entire resources devoted to the transcription of milk proteins at the expense of other biological processes during lactation.
Increases in transferrin, asmase1 and mgp1 expression during lactation were expected since these proteins are recognized as major constituents of tsetse milk [7]. The role for Transferrin in tsetse milk has yet to be determined [7], though speculation suggests transferrin may serve as a source of iron as well as for immune development/protection [7]. Regarding other milk proteins, knockdown of asmase1 in lactating females reduces fecundity and severely impacts progeny fitness [9]. Biochemical studies have revealed that secreted aSMase1 is inactive and conversion to the biologically active form, which allows sphingomyelin digestion, occurs upon encountering the acidic conditions of the larval gut [9]. As a lipocalin, MGP1 likely carries a critical unknown hydrophobic ligand in the milk [6], and has been documented to be critical for tsetse fecundity [6].
Lipids, specifically triacylglycerides, constitute the other major nutritional components in tsetse milk [2], [47]. Recent studies have shown that Brummer (Bmm) lipase- and adipokinetic hormone (AKH)-mediated lipolysis are both critical for mobilizing lipids during tsetse reproduction [47]. Our transcriptome data indicate that only a single lipase, bmm, is increased during lactation, while expression of most other lipid metabolism genes are suppressed or expressed at levels equivalent to those seen in dry flies. Such minimal transcript variation is perhaps not surprising in light of recent studies on insect lipolysis, which reveal that most control occurs at the post-translational level, either through insulin signaling or through other factors that interact with the surface of the lipid droplet [48]–[51].
An argument for post-translational regulation is further supported by our recent study showing that insulin and juvenile hormone signaling both act to coordinate lipid metabolism in tsetse mothers through transcriptional regulation of select lipolytic/lipogenic genes including midway and bmm, while other such genes associated the lipid metabolism show little variation [14]. Further, a reasonable explanation for the lack of an observed increase in lipolysis genes is that these genes typically increase prior to lactation (late embryogenesis/early larvigenesis [14]), while our lactating sample was collected during the peak of lactation occurring during the last stages of larvigenesis. In addition, perilipin1 and perilipin2 transcripts are elevated in dry versus lactating flies and perhaps these proteins, which interact with lipid droplets, are necessary to accommodate the drastic increases in fat body volume that occurs during the involution periods that separate lactation cycles. In general, our results regarding expression of lipid metabolism genes support our previous studies that bmm-mediated lipolysis plays a critical role in regulating lipid homeostasis during pregnancy [14], [47].
Removal of reads mapping to the 12 abundant milk protein genes during RNA-seq analysis allowed for the identification of three additional genes that were enriched and highly expressed during milk production. Dawdle is an activin signaling molecule that has been linked to synaptic growth at the neuromuscular junction [52] and immunity [53] in Drosophila. As a member of transforming growth factor beta (TGF β) superfamily of growth factors, activin may be signaling growth of a specific tissue, possibly the milk gland, during lactation. In addition to the role in Drosophila, activin is key in regulating growth of the mammary gland during lactation in multiple mammals and has a critical role in breast cancer [54]–[56]. The increased levels of glyoxylate/hydroxypyruvate reductase, grhpr, may provide additional substrates to maintain homeostasis of proline, the main nutrient source in tsetse hemolymph [57]–[59], as milk production requires a massive amino acid investment [1]. Finally, expression of choline-phosphate cytidylyltransferase, cct, has been linked to changes in lipid droplet size [60], and this enzyme may be playing a role either in the fat body during the rapid lipolysis associated with tsetse lactation [47], or in the generation of fat globules for incorporation into the milk. Alternatively or in combination, CCT could be critical for the allocation of choline and choline-derivatives into milk during lactation. The provision of choline is essential for proper organismal growth and development [61], [62].
Our transcriptome data revealed that the majority of genes are expressed at higher levels in dry versus lactating flies. This difference is likely due to the fact that transcript levels for most genes are reduced in lactating flies at the expense of generating lactation-specific proteins. In dry flies, transcript elevation for genes associated with digestive processes likely corresponds to the increased bloodmeal size in flies not harboring an intrauterine larva [1]. Elevated transcripts for genes coding for heat shock proteins suggest that dry flies may be better suited for stress tolerance than their lactating counterparts. In addition, proteins involved in oocyte development are elevated in dry flies, likely since oocyte development is nearly complete before lactation begins [1]. Thus, the transcript profile diversity in dry flies is more robust, featuring a more global/representative expression of genes, compared to the rather specific gene set expressed in lactating flies.
Recent studies focusing on MGP2 and MGP3 failed to verify their transfer to the nursing larvae since antisera were not available [12]. The proteomic analysis performed here confirms that these highly expressed genes synthesize milk proteins that are indeed transferred to the intrauterine larva. The proteomics data also confirm that Transferrin, MGP2–10, and aSMase1 are the primary protein components of the tsetse milk [6], [7], [9]. In addition our proteomic analysis also identified UBASH3A as a component of the tsetse milk. UBASH3A is a member of the TULA protein family and contains ubiquitin-associated (UBA) and Src-homology 3 (SH3) domains along with a histidine phosphatase domain [19], [63], [64]. A potent regulator of cellular function documented in most metazoan species [19], [64], UBASH3A is critical for regulation of T-cell proliferation and other aspects of the mammalian immune response, specifically for suppressing immune cell proliferation. The role of insect UBASH3A has not yet been determined but the presence of UBASH3A in tsetse milk suggests that it may play a role in modulating the immune system of the mother or progeny to allow intrauterine larval development. Along with potentially modulating mother-offspring immune relationship, tsetse's milk secretions also provide a route for the transmission of tsetse's microbial symbionts (Wigglesworthia and Sodalis, [65], [66]) and host immune responses may need to be regulated for symbiotic homeostasis. Our prior studies had shown that the presence of PGRP-LB in the milk is critical for symbiont transfer and overall offspring fitness [13] and the presence of UBASH3A may play a similar role in host-symbiont dynamics during the bacteria transfer within the milk. The ability to transfer symbionts to allow for maintenance of the microbiome in the offspring has been documented to be critical for tsetse immune maturation [67] and the development of the peritrophic matrix development [68]. Many other proteins were observed at lower levels; these low abundance proteins may be critical for larval development. Due to the recovery of milk from within the larval gut contents, we cannot rule out the possibility that these proteins could be products of the larval alimentary canal. Studies devoted to each low abundance peptide will be necessary to determine if it is a product of tsetse milk.
We identified seven new milk gland proteins, MGP4–10, that are similar to MGP2–3. MGP2–10 each contains a conserved secretory signal and multiple sites throughout three moderately-to-highly conserved regions with several residues under apparent strong purifying selection. Structural analysis of these MGPs failed to provide functional insights, but did reveal that these proteins are likely globular, consisting of multiple α-helices. Further study is necessary to conclusively determine the structures of these novel proteins. The coordinated high expression levels observed for mgp2–10 during lactation and reduced expression after parturition indicate that these proteins are under similar transcriptional regulation and that they may also serve as a source of proteins for larval nutrition [this study, 12]. Indeed, milk protein content was reduced by 20–25% when mgp5,7–9 transcripts were suppressed by 60–70%, suggesting that MGPs, based on total transcript abundance, likely account for 70–75% of the total protein content of tsetse milk. The MGP2–10 proteins also contain all amino acids, supporting the notion that they function as a complete protein resource for the developing larva. Furthermore, multiple phosphorylation sites associated with each protein suggests that MGP2–10 may also serve as a source of phosphate in tsetse milk. The lack of predicted glycosylation sites on MGP2–10 is not surprising since carbohydrate levels are extremely low in tsetse milk [2].
Previous studies have shown that low molecular weight proteins interact with lipids in tsetse milk [2]. This prompted us to investigate a potential role of MGPs for stabilization of milk-borne lipids. Here, we show that RNA interference of mgp7–9 results in acceleration of lipid separation from the aqueous phase of milk. This suggests that MGP7–9 (and likely the other MGPs) may represent the previously-documented, unidentified low molecular weight proteins associated with tsetse milk lipids [2]. MGP2–10 have a high proportion of hydrophobic amino acids [this study, 12], which may enable these proteins to interact with milk lipids. Thus, it appears that these newly-identified proteins are critical for maintenance of proper lipid/water dynamics in tsetse milk.
Similarities among MGP2–10 suggest that these proteins represent a highly divergent lactation-specific protein family from tsetse flies. These genes are localized to a single 40 kb chromosomal loci, have similar gene structures and their phylogeny correlates with their chromosomal organization indicating that mgp2–10 may have expanded by multiple gene duplication events from a common ancestor. It is possible that ancestral duplication events yielded two separate groups which may have been subsequently expanded as a result of unequal genetic crossing-over with the mgp2,4,5,6,9,10 being encoded on the antisense strand. Predicted three-dimensional structures between MGP2–10 is similar including multiple α-helices and a globular protein tertiary arrangement. mgp2–10 are under nearly identical transcriptional regulation showing increased expression during tsetse fly lactation and rapid decline during involution. These proteins also exhibit functional redundancy as a source of secreted amino acids in the milk and in sustaining lipid-protein homeostasis within the aqueous milk base. Although MGP2–10 have varying levels of amino acid similarities (18–91%), there are conserved regions they share outside of the secretory peptide. Specifically, 23 sites are under purifying selection (8 in the secretory peptide coding sequence and 15 dispersed throughout the remaining portions of the sequence), and these are likely critical to the functional role of MGP2–10 during tsetse lactation. Collectively, similarities between MGP2–10 indicate that these proteins constitute a novel family in tsetse similar to other highly divergent protein families, including caseins [69], [70], aquaporins/major intrinsic proteins [71], [72], odorant binding proteins [73], [74] and small heat shock proteins [75].
Our previous work demonstrated that several mechanisms underlying tsetse lactation parallel characteristics of mammalian lactation. First, both systems have highly specialized lactating cells that cycle through periods of high productivity during lactation to low activity following involution [76], [77]. Second, there are multiple, functionally analogous proteins involved in tsetse and mammalian lactation [78], [79]. These proteins include a lipocalin (MGP1 vs. β-lactoglobulin [6], [38], [45], [80]), an iron-transfer protein (Transferrin vs. Lactoferrin [7], [10]), SMase in milk or the gut contents of feeding progeny [9], [81]–[83] and various immunity proteins (PGRP and UBASH3A vs. multiple mammalian immunity proteins, [this study], [ 45], [78], [84,85]. Third, the lipid content transferred to the developing offspring is similar during lactation in both systems. Fourth and finally, microbial symbionts are transferred from the mother to the developing offspring in both tsetse and mammals [66], [86]–[88]. There are however a few noteworthy differences between tsetse and mammalian lactation, such as the abundance of calcium transport proteins in mammalian not found in tsetse milk [this study], [76], [79,89]. This difference is unsurprising, since insects do not require large amounts of calcium for their chitin-based exoskeleton. In addition, tsetse milk contains a lower carbohydrate content than mammalian milk [76], [90], indicating that tsetse flies rely solely on lipids and protein for growth and development, rather than a combination of sugar/lipids/protein as in the mammalian case. Such reduced reliance on sugar is also unsurprising as tsetse flies have little to no detectable levels of glucose within their bodies and use proline as their circulating hemolymph resource, rather than a glucose-based substrate such as trehalose [1], [57].
Mammalian genomes contain no orthologous sequences to the nine novel tsetse MGPs. However, MGPs might function analogously to caseins in mammalian milk. Caseins are the major amino acid and calcium source for the mammalian neonate [65], [70], [91]. While MGPs do not carry calcium, they do, like caseins, represent a major amino acid resource in the milk [39], [46], [69], [92]. The presence of multiple phosphorlyation sites in MGPs suggests that this novel protein family may also act in tsetse milk as a source of phosphate as do caseins in mammalian milk [69], [70]. Furthermore, caseins are amphipathic molecules that form micelles, which interact directly with lipids both in vivo and in vitro [69], [93]. According to our results, MGPs likewise interact with lipids to promote stability of lipid emulsions in the aqueous tsetse milk. To determine if MGP2–10 have amphipathic structural properties like caseins, direct protein structural studies, rather than protein modeling, will be necessary. In addition, expansion of the casein and MGP gene families has occurred for both mammals and tsetse within specialized regions of their genomes. This indicates that expansion of these protein families (MGPs and caseins) is advantageous for provisioning the necessary nutrients in both tsetse and mammalian milk, respectively [this study], [ 69,92]. Finally, members of the MGP and casein families show substantial divergence in sequence similarity [this study], [69,78], which is a characteristic of proteins that are mainly nutritional components of milk. Proteins involved in mechanics of lactation, i.e. milk fat globule formation or have an enzymatic function, are typically more conserved within and between organisms [78]. These similarities further support the idea that MGPs perform an analogous role to mammalian caseins in tsetse milk.
Few studies have examined the effects of casein knockdown/knockout in mammals. In mice, knockout lines have been developed for α-, β- and κ-casein [94]–[96], and in goats there are naturally occurring deficiencies in α-casein [97]. Knockdown phenotypes differ dramatically, depending on the casein variant targeted. The knockout mutant for β-casein in mice [95] and null αS1-casein in goats [97] have no or minimal apparent effects on milk production, potentially due to increased expression of other casein genes to compensate for the loss of β-casein or αS1-casein, respectively. Offspring receiving milk from α-casein-null mothers experience delayed growth and life-long body size reduction, but only transient effects on physical and behavioral development [96]. The most drastic change is noted in κ-casein null mice, which fail to lactate [94]. Similar to suppression of caseins, knockdown of individual tsetse MGPs had only minimal effects on tsetse fecundity; more drastic changes occurred upon silencing multiple transcripts. In addition, a reduction in tsetse MGPs accelerated separation of lipid emulsions. Caseins likely interact similarly with lipids in mammalian milk to promote lipid emulsifications. Indeed, in addition to their biological roles, caseins have also been industrialized as emulsifying agents [98], [99]. This feature highlights the ability of these proteins to stabilize lipids present in the milk, as noted in tsetse. Proteomic studies examining mammalian milk fat globules have identified caseins, indicating that these proteins are associated with milk lipids [100]–[102]. Specifically, casein modification alters lipid composition and protein components of the milk fat globule in goats [103]. The analogous functions of MGP2–10 and caseins suggest roles for these proteins as a source of amino acids, as stabilizers of milk homogeneity, and as carriers of polyatomic ions (i.e. phosphate groups). These roles must be fulfilled by a specific abundant protein or protein family to satisfy nutritional requirements of an immature organism during periods of lactation.
This study provides the first complete examination of the mechanisms underlying tsetse fly lactation. In general, our results show that the majority of genes have lower expression during lactation with the exception of those directly involved in milk production. The combination of transcriptomic and proteomic analyses reveals there are 12 major milk gland proteins, which comprise ∼47% of the transcriptome of lactating flies, along with multiple minor protein constituents of tsetse milk. We have provided an overview of the combined results of this study (Fig. 11). Furthermore, we discovered a novel, tsetse-specific protein family, MGP2–10, that is expressed highly during lactation. Interference with expression of these proteins reduces tsetse fly fecundity, suggesting that this family of MGP genes may provide a target for development of tsetse-specific abortifacients. This study has also revealed that many of the underlying functional aspects of tsetse fly lactation are analogous to those of other lactating organisms. This example of convergent evolution suggests that tsetse flies could be used as an invertebrate model system to investigate the complex molecular and physiological aspects associated with obligate lactation.
Colonies of G. morsitans morsitans were reared at Yale University and the Institute of Zoology at the Slovak Academy of Sciences (SAS). The other two species (G. pallidipes and G. fuscipes) were reared at SAS. Flies were maintained on blood meals provided through an artificial feeding system at 48 h intervals [104]. Two groups of females were used for transcriptome analysis: the first group carried a third instar larva (lactating) while the second group was examined 48 h post parturition (dry or non-lactating). Developing progeny were removed from each female to ensure transcript changes were representative of differences between the mothers. For sex specific transcript analysis, males and females were collected 16–18 d after adult emergence. Tissue samples were collected from pregnant females (16–18 d after adult emergence) carrying third instar larvae 24 h after blood feeding. Samples for temporal expression analyses were collected according to progeny development status based on previous studies [9], [47]. Flesh flies, S. crassipalpis, acquired from Ohio State University were reared according to standard procedures [105].
Total RNA was extracted from individual flies or dissected tissues using Trizol reagent (Invitrogen, Carlsbad, CA, USA), following the recommended protocol. RNA was treated twice with the TURBO DNA Free kit (Ambion, Austin, TX, USA) to remove DNA, alcohol precipitated to remove residual salt, and further cleaned using the RNeasy kit (Qiagen, Maryland, USA). Total RNA (2–3 µg) was pooled from 10 flies extracted individually for each treatment. Sample quality and concentration was determined using a Bioanalyzer 2100 (Agilent, Palo Alto, CA, USA). Library construction was performed using standard protocols for Illumina mRNA-Seq sequencing by the W. M. Keck Foundation Microarray Resource at the Yale School of Medicine. Each single-end library was sequenced on one lane of the Genome Analyzer II platform (Illumina, San Diego, CA, USA).
To determine Illumina read quality, FastQC analysis was performed on the transcriptomes generated from dry and lactating flies. Due to the prevalence of tsetse symbiont sequences in the reads, a specific quality control step was included to reduce bacterial sequence reads using the known whole genome sequence data from Wigglesworthia [106], Wolbachia (unpublished) and Sodalis [107] determined from the same host species G. morsitans. Following symbiont specific sequence removal, remaining sequences were trimmed in CLC Genomics (CLC Bio) to remove ambiguous nucleotides. Contig libraries were constructed using Abyss [16], [17] followed by a secondary assembly with Trinity [18]. Functional annotation was accomplished using the BLASTx algorithm through comparison with sequences included in the NCBI protein database [108] as well as the KOG [109] and GO databases [110]. Conserved protein domains were detected using rpsBLAST [111] searches against the CDD, Pfam and Smart databases [112]. Predicted protein translations were submitted to SignalP to identify potential secretion products by screening for secretion signal motifs [113]. Additionally, contigs were compared to several proteomes obtained from Flybase [114] (D. melanogaster) and Vectorbase [115] (An. gambiae). Each read from each library was compared by BLASTn to the assembled coding sequences (CDS) using a word size of 25, m8 output and low complexity filter turned off. CDS coverage and CDS number of read “hits” from each library were computed from the BLAST output file. A hit was only considered significant if it had 97% or better identity to its target and no more than one gap. The same read could be mapped up to three different CDS to the extent that their BLAST scores were identical. Expression levels were determined using CLC Genomics Workbench (CLC bio, Cambridge, MA). Reads were mapped to our de novo assembly with an algorithm allowing only two mismatches and a maximum of 10 hits per read. RPKM was used as a measure of gene expression [116]. The proportion of read counts for each contig in relation to the total read counts in each sample was determined in order to calculate P-value differences in proportions by a Z-test following Bonferroni correction [117]. Fold change was determined as the ratio of RPKM of lactating flies vs. RPKM of dry flies. In addition to the analysis of the complete Illumina libraries, a secondary analysis was conducted featuring Illuminia libraries filtered to eliminate milk-specific contigs to reduce bias by these highly abundant proteins [116], [117]. Data from this study are available in Sequence Read Archive (SRA075330).
Pulled glass capillary tubes were used to collect milk samples by negative pressure from the guts of feeding third instar larvae, which were microscopically dissected from the uterus of pregnant females. Samples were stored in 1× protease inhibitor cocktail (Sigma-Aldrich). Proteins were precipitated with 10% trichloroacetic acid (Fisher Scientific) at 4°C overnight, collected by centrifugation (11,000×g, 30 minutes, 4°C) and washed two times with ice-cold acetone. Protein pellets were briefly dried and dissolved in 10 µl of protein pellet buffer (8M urea, 3M thiourea, and 1% dithiothreitol). Trypsin digestion was performed at 37°C for 12–16 h following dilution with distilled H2O to a final volume of 100 µl. Samples were stored at −80°C until analysis. Peptides were separated with a Waters nanoAcquity UPLC system (75 µm×150 mm BEH C18 eluted at 500 nl/min at 35°C) with Buffer A (100% water, 0.1% formic acid) and Buffer B (100% CH2CN, 0.075% formic acid). A linear gradient was established with 5% Buffer B, increasing to 50% Buffer B at 50 minutes and finally to 85% Buffer B at 51 minutes. MS/MS was acquired with an AB Sciex 5600 Triple Time-of-Flight mass spectrometer using 1 microscan followed by four MS/MS acquisitions. Neutral loss scans were obtained for 98.0, 49.0 and 32.7 amu. Seven separate 1 µl injections at an estimated 0.351 µg/µl concentration for a total of 2.457 µg on the column were used for analysis.
Mascot algorithm was used to analyze uninterrupted MS/MS spectra [118]. The Mascot Distiller program used MS/MS spectra to generate Mascot compatible files by combining sequential MS/MS scans from profile data that have the same precursor ion. Charge states of +2 and +3 were preferentially located with a signal-to-noise ratio of 1.2 or greater. A list of protein sequences was created and used in the BLASTx search against Trinity-assembled library from the pregnancy-specific analysis and positive matches were identified by tBLASTx against the NCBI and Swiss-Prot databases. Mascot scores were based on MOlecular Weight SEarch (MOWSE) relying on multiple matches of more than one peptide to the same predicted protein [119], [120]. The MOWSE based ions score is equal to (−10)*(Log10P), where P is the absolute probability that a match is random. Matches were considered significant when the probability of a random match fell below 5% (E value<0.05). Therefore, Mascot scores greater than 68 were above the significance threshold when searching the newly assembled library. Proteins were considered to be successfully identified when two or more peptides matched the same predicted protein and the Mascot score exceeded the significance threshold. The exponentially modified protein abundance index (empai) was employed to estimate levels of protein species based on the number of species detected compared to the number of possible peptides for specific protein [121], .
Chromosomal organization of genes and full length mRNA sequences for mgp2–10 were obtained by mapping Illumina high-throughput reads against G. m. morsitans genomic scaffolds in the CLC Genomics software package. Nucleotide and predicted protein sequences were aligned using PROMALS3D [123] and Clustal [124] and formatted with BioEdit [24]. Flesh fly, Sarcophaga crassipalpis, sequences were obtained from a previous EST project [20], [21]. Sequences of mgp2–10 from other tsetse species (G. pallidipes and G. fuscipes) were obtained from female cDNA by RT-PCR followed by cloning into T-vector plasmid (Invitrogen) and sequenced at the DNA Analysis Facility at Yale University (New Haven, CT). Pairwise phylogenetic tree construction and bootstrap analysis (10000 replicates) were performed using the MEGA4/5 sequence analysis suite [25], [125]. dN/dS analyses were performed using the FEL (Fixed Effects Likelihood [28]) and PARRIS [27] algorithms available via DataMonkey [29], [30], which is a web-based implementation of the HyPhy phylogenetic analysis program [31]. Sequences were translated, aligned, reverse translated and the stop codons removed in accordance with the requirements for sequence input to DataMonkey. Under the FEL method, posterior probabilities cutoffs were set at 95, which is equivalent to a p-value of 0.05 for the site-specific detection of codons under positive or negative selection. Analysis of specific regions of the MGP2–10 coding regions was conducted using MEGA5 according to previous milk protein studies [32] and individual regions were based upon protein coding regions with high or low levels of amino acid homology.
For sex- and tissue-specific RT-PCR expression analyses, total RNA isolated from males and females and from dissected tissues was used as template for the Superscript III reverse transcriptase kit following the manufacturer's protocols (Invitrogen). Fat body and milk gland were analyzed as a combined samples since complete separation is nearly impossible due to the intricate association of these organs. PCR was performed with gene-specific primer pairs (Table S1) using the GoTaq DNA polymerase kit (Promega). The PCR amplification conditions were as follows: 95°C for 3 min, 35 cycles of 30 sec at 95°C, 52–56°C for 1 min, and 1 min at 70°C using a Bio-Rad DNA Engine Peltier Thermocycler (Hercules, CA).
For pregnancy-specific transcript abundance determination, qPCR analyses were performed using a CFX PCR detection system (Bio-Rad, Hercules). Data were analyzed with CFX manager software version 3.1 (Bio-Rad). Primer sequences used were the same as used in RT-PCR analyses (Table S1). Comparative Ct values for genes of interest were standardized by Ct values for the control gene (tubulin) relative to the average value for the control treatment or newly emerged flies, yielding the delta Ct value. All experiments were analyzed in triplicate and subject to ANOVA followed by Bonferroni correction and Dunnett's test.
Short interfering RNAs (siRNA) consisting of two Duplex sequences (Table S1) were designed using Integrated DNA Technologies online software (IDT). Control siRNAs were designed against Green Fluorescent Protein (GFP; Table S1). Each oligo, designed to target a single mgp gene, was also compared to the reference RNA library/G. morsitans genome [14]) and the Trinity contigs library from this study to verify target specificity. The oligos for each strand of the siRNA were combined, and the concentration was determined spectrophotometrically followed by adjustment to 800–850 ng/µl per siRNA. Each female fly was injected with 2 µl siRNA solution into the thorax 8–10 d after adult emergence. Five days post-injection, gene expression levels were determined by qPCR and normalized to tubulin transcripts. For combined knockdown studies, siMGP constructs were mixed to yield a sample concentration of at least 600 ng/µl for each siRNA targeting a specific MGP transcript. Fecundity following MGP knockdown was assessed as previously described [9]. Finally, milk protein content was determined by Bradford assay (Bio-Rad) after extraction from the larval gut contents as described above.
Emulsification assays were based on milk turbidity measurements. For this assay, milk was acquired from the guts of actively feeding larvae as before and diluted 10× prior to the assay. Samples were vortexed for 1 min at 10,000 rpm, and absorbance of the diluted emulsion was measured at 500 nm. Changes in absorbance were measured hourly for 10 h. Results were analyzed based on the slope of a regression, where ln (ABSt/ABS0) is plotted versus time based on the exponential model (ABSt = ABS0 e−kt). For this model, ABSt denotes absorbance at any time t, ABS0 is the initial absorbance, and k is the rate of absorbance decline in %/h.
To generate structural models for MGP2–10, four web-based de novo protein modeling programs were consulted. QUARK is a recently developed ab initio assembly program that will first break proteins into small sequences, following which full-length sequence models are assembled using Monte Carlo simulations [126]. The I-TASSER program first develops a three-dimensional model and subsequently predicts function based on structural similarity with functionally defined proteins [127]. Phyre2 is a widely used protein homology/analogy recognition engine that can rapidly predict the structure of 250 residue proteins [128]. Finally, SPARKS-X is a program that performs well in comparison to other programs [129]. Each program was run under the default configuration and the resultant predicted protein structures were visualized using Discovery Studio 3.1 (Accelrys).
|
10.1371/journal.ppat.1005965 | A Viral Immunity Chromosome in the Marine Picoeukaryote, Ostreococcus tauri | Micro-algae of the genus Ostreococcus and related species of the order Mamiellales are globally distributed in the photic zone of world's oceans where they contribute to fixation of atmospheric carbon and production of oxygen, besides providing a primary source of nutrition in the food web. Their tiny size, simple cells, ease of culture, compact genomes and susceptibility to the most abundant large DNA viruses in the sea render them attractive as models for integrative marine biology. In culture, spontaneous resistance to viruses occurs frequently. Here, we show that virus-producing resistant cell lines arise in many independent cell lines during lytic infections, but over two years, more and more of these lines stop producing viruses. We observed sweeping over-expression of all genes in more than half of chromosome 19 in resistant lines, and karyotypic analyses showed physical rearrangements of this chromosome. Chromosome 19 has an unusual genetic structure whose equivalent is found in all of the sequenced genomes in this ecologically important group of green algae.
| We propose that chromosome 19 of O. tauri is specialized in defence against viral attack, a constant threat for all planktonic life, and that the most likely cause of resistance is the over-expression of numerous predicted glycosyltransferase genes. O. tauri thus provides an amenable model for molecular analysis of genome evolution under environmental stress and for investigating glycan-mediated host-virus interactions, such as those seen in herpes, influenza, HIV, PBCV and mimivirus.
| Eukaryotic micro-algae of the genus Ostreococcus and related species of the order Mamiellales are globally distributed in the photic zone of the world's oceans where they contribute to fixation of atmospheric carbon and production of oxygen, besides providing a primary source of nutrition in the food web [1–3]. Their tiny size (1–3 μm), simple cells (one chloroplast, one mitochondrion), ease of laboratory culture and extremely small genomes, several of which have been completely sequenced [4–8], render them attractive as models for marine ecology [9–11], cell biology [12] and evolution in the green lineage [4,13]. Typical features of the highly streamlined genomes of the Mamiellales include a higher GC content (48–64% [7,14], 59% for O. tauri) than higher plants (41%, [15]) and two unusual “outlier” lower GC chromosomes that can carry higher proportions of transposons and genes predicted to originate from prokaryotes (see [16] for a recent review). In the type species O. tauri, these two chromosomes are chromosome 2 (the big outlier chromosome or BOC) and chromosome 19 (the small outlier or SOC), both of which greatly vary in size between strains isolated from the environment [17].
In marine environments, unicellular organisms account for the largest biomass, but they are outnumbered by about ten to one by viruses, which can lyse host cells and thereby contribute to biogeochemical cycles by promoting the turnover of plankton populations [18–20]. While large DNA viruses infecting algae have been known for many years (see [21,22] for reviews), including viruses of Micromonas pusilla, a member of the Mamiellales [23], viruses of Ostreococcus spp. were discovered more recently. Complete genomes of many viruses infecting Mamiellales (genus Prasinovirus) are now available [24–27]. Given the size of an Ostreococcus cell (about 1 μm) prasinoviruses are huge icosahedral particles, about one eighth of the host cell diameter between faces (110 nm), carrying a repertoire of about 250 genes. Resistance to prasinoviruses arises spontaneously in cultured algal lines [28] and their complex host–strain specificity patterns [29,30] witness their involvement in the planktonic 'arms race' that assures for rapid turnover and evolution in micro-algal populations (see [31] for a review). Viruses of algae in general and prasinoviruses in particular play an important role in controlling phytoplankton populations [32–36], the effectiveness of host defence responses thus determines the fate of algal populations.
Regrowth of eukaryotic microalgae following viral lysis with large dsDNA viruses has been observed in phylogenetically diverse lineages [37,38] and such virus-resistant lines can be stable in culture [28,38,39]. In the diplont marine bloom-forming haptophyte Emiliania huxleyii viral infection of diploid cells led to the selection of haploid cells which were resistant to infection [40]. Cell lines of the toxic bloom-forming dinoflagellate Heterocapsa circularisquama became resistant to viral infection by HcRNAV after co-culture with this single-stranded RNA virus. In this case, resistance appeared to be reversible and a variable proportion of resistant cells in culture harboured viruses and produced particles [41]. Some of these authors suggest that changes in the host cell surface might occur, or that non-infectious defective particles in culture might occupy available receptor sites on host cells, but no specific molecular mechanism has been demonstrated.
In O. tauri, Thomas et al. [28] observed two kinds of virus-resistant cell lines, either lines which were resistant but chronically infected, producing viruses at a low level in culture (“resistant producers” referred to as RP), or cell lines devoid of viruses and immune to re-infection (“resistant non-producers”, or RNP). Here we aimed to unravel the molecular mechanisms underlying viral resistance using clonal lines of O. tauri RCC4221 [4] and its virus, OtV5. We produced numerous independent OtV5-resistant lines to examine host and viral gene expression in detail with RNA-Seq technology and to test whether altered gene expression could be responsible for the resistance phenotypes.
The experimental schema in Fig 1 shows how independent clonal lines of O. tauri resistant to OtV5 were generated from a clonal starting population. The parent OtV5-susceptible O. tauri was subcultured into 46 parallel independent culture lines (one set of 30 and another set of 16). OtV5 was introduced into 38 of the O. tauri lines while eight were maintained as non-infected controls. All cultures with added OtV5 lysed, appearing clear of visible cells within 3–5 days. The cultures exposed to OtV5 showed regrowth of cells after approximately one week and, upon re-exposure to OtV5, were subsequently resistant to lysis. Periodic re-testing for OtV5 resistance showed viral resistance persisted over the almost two-year duration of the study while the control lines continued to be susceptible (8 susceptible control lines were maintained alongside the resistant lines shown in Fig 1, and at each of the 10 dates shown 8/8 lines remained susceptible to OtV5; at the end of the time course, all of the remaining 36 lines were found to be resistant). Resistant lines were periodically tested for production of infective OtV5 by taking cell-free culture medium from resistant lines, adding it to the susceptible parent and observing lysis. The majority of resistant lines were RP at the start of the experiment, however, the proportion of RP diminished over time so that RNP lines were more abundant at the end of the study (Fig 2).
Twenty-three O. tauri transcriptomes were sequenced from the experiment including 19 virus-resistant and four susceptible. To test if OtV5-resistance was linked to differential gene expression and to identify the genes involved, we compared four resistant to four susceptible transcriptomes from matching RNA library batches (S1 Table). At this early point in the study all but two lines were RP, so resistant lines used in the comparison were RP, however over the course of the study several lines of evidence indicated RP and RNP were probably due to similar molecular mechanisms (discussed below). Over 95% (7432 of the 7749) of O. tauri chromosomal genes analysed were transcribed to some extent (≥10 aligned reads in at least one sample) in both resistant and susceptible O. tauri lines suggesting that the majority of genes were expressed at the time of sampling.
A total of 170 O. tauri chromosomal genes were significantly differentially transcribed in resistant lines of which 103 were up-regulated and 67 down-regulated (Fig 3, and see S2 and S4 Tables for full gene descriptions), which represents only 2% of all expressed O. tauri genes. Most strikingly, 49 differentially transcribed genes, representing almost a third of differentially transcribed genes were concentrated on chromosome 19, the SOC (Fig 3A). These genes on chromosome 19 also had overall larger log2 fold change values indicating larger changes in expression levels on genes from this chromosome than the others (Fig 3B). Predicted glycosyltransferases (GTs) appear to be enriched on chromosome 19 compared to the other chromosomes, the majority of which were differentially transcribed in OtV5-resistant lines (Fig 3C). The only other chromosome encoding GTs regulated in virus resistant lines was chromosome 2, the BOC, where the GTs were down-regulated. Annotated genes on chromosome 19 are known to belong to few functional categories [5] relating to surface membrane proteins, the building of glycoconjugates, as well as methyltransferases (MTs) (Fig 4, S2 Table). The majority (16 of 22) of differentially transcribed carbohydrate transport and metabolism genes (Fig 3D) were located on chromosome 19. Furthermore, almost half of the differentially transcribed genes of unknown function (29 of 54) (Fig 3D) were also encoded by chromosome 19. Thus, genes involved in virus immunity in O. tauri are preferentially encoded by the SOC; these genes are strongly associated with carbohydrate modification and metabolism and their specific regulation was associated with viral resistance.
Other functional categories that were significantly differentially transcribed were related to translation, transcriptional regulation (chromatin remodelling, RNA modification, transcription factors), protein modification and turnover, amino acid transport and modification, and other transporters (Fig 3C). Amino acid biosynthesis genes involved in four different pathways were down-regulated in resistant lines, implicating a decrease in a broad range of amino acids and their downstream metabolites during viral immunity (S4 Table). Ribosomal subunits were also under-expressed suggesting a slowing of growth-related processes. Genes involved in expression regulation were likely involved in maintenance of the virus-resistant state. In particular, histone modification genes were all over-transcribed in the resistant lines pointing to a key role for chromatin restructuring in resistance. Proteasome-mediated protein degradation (ubiquitination) as well as post-translational modification (glutathione-associated and chaperones) were up-regulated and may be part of the viral immunity response or translational-level regulation. Transporters whose substrates were related to undefined small, generally inorganic molecules had mixed regulation with a few cases of transporters with related functions being regulated in opposing directions. For example, a transmembrane phosphate transporter (ostta06g00210) over-expressed while another calcium-dependent phosphate transporter (ostta17g00940) was under expressed (S4 Table). Altered transcription of transporters may modulate the available substrate pool required by OtV5 curbing viral replication. However, these genes show lower significance levels of differential expression and may be involved in processes that are not directly relevant for resistance.
Re-sequencing of the O. tauri parent line with single molecule PacBio technology was able to resolve a large inverted repeat region (LIRR) on chromosome 19 totalling ~180 kb (Fig 4) that was predicted to exist from previous work [4]. The assembled PacBio contig was 311,428 bp, in agreement with its expected size from gel mobility (Fig 5). Notably, all over-transcribed genes in the resistant lines from the SOC were located in this region (Fig 4). Differentially transcribed genes on other chromosomes (S4 Table) were in some cases adjacent to each other, but were not grouped together as on chromosome 19. Remarkably, transcription of blocks of genes in the LIRR was almost completely suppressed in controls indicating close to half the SOC was silenced in the OtV5-susceptible wild-type state. By contrast, in resistant lines, these genes were highly transcribed, having among the highest mean fragment counts of the differentially transcribed genes (S2 Table). The co-transcription of genes on the LIRR suggests they were under the control of the same regulatory factors.
Re-sequencing was also able to resolve a tandem repeat region (TRR) comprised of ~2,255 bp monomeric repeats sharing 97–99% identity (S1 Fig), which was under-transcribed in resistant lines but still showed some transcription with high sample–sample variance (Fig 4). Although small genes of unknown function (ostta19g00035 and ostta19g00240) occurred in the TRR, the sequences outside of these genes appeared transcribed, suggesting a functional role. The repeats were interrupted at two loci by unique sequences; (1) a 2,232 bp putative terminal repeat retrotransposon in miniature (TRIM) possessing long terminal repeats (LTRs) and (2) a 3,354 bp sequence of unclear origin (S1 Fig). The putative TRIM is likely derived from the complete 5,537 bp LTR retrotransposon, retrostreo2 on chromosome 8, with which it shares 99% identity in the LTRs. Extensive deletions in the intervening gene sequences suggests the putative TRIM is non-autonomous and is an example of LTR-transposon miniaturisation in O. tauri, as well as transposition between chromosomes. Tandem repeats interrupted by transposons are hallmarks of centromeric repeats [42,43]. However, the repeat sequence was not found on other chromosomes, as would be expected in centromere recognition sites, leaving the functional significance of the TRR unclear.
We conducted an exploratory analysis of all sequenced transcriptomes (S1 Table) to see if the chromosome 19 genes detected in the differential expression analysis were consistently regulated the same way in independent resistant lines that were not part of the comparative analysis. Hierarchical clustering of the long inverted repeat region (LIRR) gene transcription profiles grouped all susceptible samples separately from the resistant samples (S2A Fig). In particular, two blocks of genes (ostta19g00070–120 and ostta19g000560–610) were transcribed across all independent resistant lines, albeit with variation in transcription levels, but not in the susceptible controls, supporting the LIRR having a shared regulatory mechanism activated by OtV5 infection. By contrast, down-regulated genes were not consistently under-transcribed in all resistant lines (S2B Fig). This indicates over-expression of the LIRR genes was a strong determiner of resistance while under-expression of chromosome 19 genes, including the TRR, was not consistently correlated with resistance.
Several GT genes from chromosome 19 induced in the virus-resistant state were of putative foreign origin (S2 Table) and appeared to be clustered with genes encoding functionalities related to sugar metabolism and modification (Fig 4, left gene map). These genes include a rhamnan synthesis F/Wbx GT (ostta19g00070), family 92 GT (ostta19g00600) and the co-regulated sugar modification enzymes (ostta19g00110, NAD-dependent epimerase/dehydratase; ostta19g00610, CMP-N-acetylneuraminic acid hydrolase).
Ostta19g00070 comprises merged rhamnan synthesis F and Wbx GT domains. The former is allied with bacterial rhamnose-glucose polysaccharide F (RgpF), which in Streptococcus transfers rhamnose to the nascent rhamnan backbone, which is incorporated into surface lipopolysaccharide [44]. RgpF domain in O. tauri is currently the only occurrence in eukaryotes, the species distribution being otherwise restricted to bacteria (Pfam: PF05045). Similarly, Wbx GT domain is found in bacterial gene clusters involved in synthesis of O-antigen, which in Shigella comprises repeated monomers of L-rhamnose, D-galacturonic acid and N-acetylgalactosamine residues [45,46]. One putative gene cluster (ostta19g00110–140) comprises a predicted NAD-dependent epimerase/dehydratase, glycoprotein-N-acetylgalactosamine 3-beta-galactosyltransferase, integral membrane galactosyltransferase and triose-phosphate transporter. Strikingly, the 3-beta-galactosyltransferase was associated with sequences from metazoans, while the adjacent genes appear to have homologues in the green lineage. A second putative gene cluster was a CAZy (carbohydrate-active enzymes database) family 92 GT containing a triose phosphate transporter domain (ostta19g00600) adjacent to CMP-N-acetylneuraminic acid hydrolase (ostta19g00610) of metazoan origin. These two genes were putatively involved in the activation of sugars for the synthesis of cell surface sialic acid. In proximity to these clusters of sugar modifcation genes was a predicted FkbM methyltransferase (ostta19g00560), which suggests it has a role in glycan methylation. Two of the up-regulated GTs (ostta19g00130 and ostta19g00630), were predicted to be membrane-associated and could be part of the Golgi mannosyltransferase complex.
Two different GTs and a triose-phosphate transporter on chromosome 19 appear to be recently generated paralogous copies because they share up to 97% amino acid identity (S3 Fig). The paralogues curiously occur in distinct genetic contexts (Fig 4). For example, the previously mentioned 3-beta-galactosyltransferase (ostta19g00120) is adjacent to sugar modification genes and is over-transcribed in resistant lines while the putative paralogue (ostta19g00320) is adjacent to a conserved eukaryotic algae hypothetical protein and was under-transcribed (Fig 4, GT a). In this case, it suggests the same, or similar, base sugar substrate is utilised in both conditions, but alternatively modified in virus resistant lines. How these GTs and triose-phosphate transporters have been duplicated and shifted in this modular fashion is unclear, but the transposon-related genes on chromosome 19 are candidates for mediating these transfers.
The only two GTs that were differentially regulated in resistant lines that were not located on the SOC, were located on chromosome 2. These genes were down-regulated and included glycosyltransferase AER61 (ostta02g00040) and a membrane glycosyltransferase (ostta02g02060). However, other carbohydrate modification and metabolism genes located on chromosome 2 that were up-regulated in resistant lines included a somatomedin B domain-containing protein (ostta02g03940) and carbohydrate glycoside hydrolase (ostta02g04570). The former is involved in binding polysaccharides and the latter in the lysis of O-glycosidic bonds. This implicates a role also for the BOC in carbohydrate modification during viral resistance.
As almost all lines were RP at the time of RNA sequencing, we explored the transcriptomes in sequenced samples for evidence of OtV5 transcription (S1 Table). In susceptible controls, a low number of read counts were assigned to few OtV5 genes (S4 Fig); in particular, OtV5_154c and OtV5_159, which correspond to highly transcribed viral specific genes (S3 Table). As susceptible cells had not been exposed to OtV5, apparent transcription of viral genes were taken to be artefacts probably arising from mis-assigned sequencing sample indexes to highly expressed transcripts within the shared sequencing flow cell lane (all samples were multiplexed together). BLAST searches at neither nucleotide levels nor protein levels revealed similarities that would intimate horizontal transfer of these genes between host and virus. As expected, lines that were RNP had no OtV5 transcription and clustering with susceptible controls in their OtV5 transcription profiles (S4 Fig). The RP lines fell into three clusters corresponding to relatively high (five samples), moderate (five samples) and negligible OtV5 transcription (seven samples near-identical to controls) showing that for most RP lines OtV5 transcripts were nearly undetectable.
As some OtV5 genes appeared to be highly transcribed in certain RP lines, we sought to estimate the amount of virus transcription relative to that of the host. The ratio of the mean transcript counts of the most expressed viral gene (OtV5_154c: 5,878) over that of one of the most highly expressed host genes, the ribulose bisphosphate carboxylase small subunit (rbcS, ostta18g01880: 49,249), gives 0.12. As a comparison, the transcript abundance ratio of the corresponding genes during Paramecium bursaria chlorella Virus 1 (PBCV-1) infection of Chlorella was 37, far exceeding host transcription levels [47]. Since relative OtV5 transcription was lower than that of the host, even in RP lines with relatively higher OtV5 transcripts levels, this suggests viral activity was certainly much lower than expected during infection of susceptible O. tauri.
Nonetheless, the OtV5 genes that were transcribed to a significant level (mean normalised count > 9) covered over 40% of all OtV5 genes. Expressed OtV5 genes were distributed along the entire viral genome with the expression profile varying for each sample (S5 Fig). Although we could not compare the OtV5 expression in RP with control cells in a normal lytic cycle to confirm this, the OtV5 transcription profiles did not indicate specific genes linked to chronic infection, nor a particular stage in viral replication. Crucially, transcripts of capsid proteins were detected (S3 Table) consistent with OtV5 forming virions. This corroborates with results of the virus production assay, which detected infective lytic virions in the medium of RP lines (Fig 2). Apart from virion structure, transcribed OtV5 genes were associated with DNA replication, transcription, amino acid synthesis and carbohydrate modification and metabolism (S3 Table).
Expressed OtV5 carbohydrate metabolism genes included two enzymes involved in the biosynthesis of nucleotide sugars, OtV5_011, a predicted GDP-D-mannose 4,6-dehydratase, and OtV5_042, a putative dTDP-D-glucose 4,6-dehydratase. The homologue of this GDP-D-mannose 4,6-dehydratase in PBCV-1 has a dual functionality acting in both the synthesis of GDP-D-rhamnose and GDP-L-fucose, both of which are monosaccharides present in the capsid glycan but are rare in the host [48–50]. Adjacent to the GDP-D-mannose 4,6-dehydratase is a group 1 GT (OtV5_012c) that was also expressed. The spatial proximity of these genes suggests the glycosyldonor of the group 1 GT is the nucleotide sugar produced by the GDP-D-mannose 4,6-dehydratase. Three additional OtV5 GTs were transcribed including a predicted membrane-localised group 34 GT (OtV5_033), a group 2 diphosphosugar GT (OtV5_035) and a GT with no assigned CAZy group (OtV5_160). All OtV5 encoded GTs had their closest homologues in prasinoviruses indicating they have conserved virus-specific functions.
Given the massive transcriptional changes in chromosome 19 and the known plasticity in outlier chromosome size between O. tauri strains [17], we examined the karyotypes of all experimental lines by pulsed-field gel electrophoresis (PFGE). Resistant lines showed karyotype changes, most notably as a shift in the size of chromosome 19 (34 of 36 resistant lines tested) while no change was evident in the susceptible controls (Fig 5). The most common changes were an increase of ~20–490 kb (17 lines) or a decrease of ~40–140 kb (13 lines) in the size of chromosome 19. This was likely due in the former case to duplications within chromosome 19 and in the latter, to deletions within, or fission of, chromosome 19. Several resistant lines (R1, R22, R29 and R30) appeared to have lost chromosome 19 or its location was ambiguous. In R9a and R2, there were possible translocations of regions of chromosome 19 to chromosomes 6 and 8 (respectively) or an increase in the size of chromosome 19 with concurrent insertions-deletions in other chromosomes. Indeed, variations in the size of chromosomes other than chromosome 19 were noted in R11a, R3, R5, R22, R23 and R30. More complex changes were also evident, most notably in R30 where bands apparently larger than 1.1 Mb are present that hybridize with the chromosome 19 probe. Interestingly, seven resistant lines (R9a, R13a, R1, R2, R12 R18 and R26) showed densely hybridizing material in the wells, almost always coinciding with a large increase in the size of chromosome 19 (>430 kb). This could correspond to large circular forms that cannot migrate in the PFGE [51], suggesting the extensive changes to chromosome 19 could occur via circular intermediates. However, the presence of circular DNA in the nucleus such as episomes has neither been previously detected nor specifically investigated in O. tauri.
To investigate the chronic production of OtV5, electron microscopy was performed on two RP lines. This showed a minority of cells (<0.5%) was visibly infected and/or in the middle of lysis (Fig 6) demonstrating viral reproduction proceeded by a typical lytic cycle. Nonetheless, a population crash of the resistant cultures was never observed, with the only loss of lines occurred following antibiotic treatment, so the frequency of lysis was low. Since all lines were re-cloned upon acquiring resistance (Fig 1), the minority of susceptible cells present in RP lines were presumably resistant cells that had switched to a susceptible state.
Production of the OtV5 genome was also observed by PFGE as a ~200 kb band (Fig 5) in 15 of the 19 lines that were RP at that time point. In the four RP lines where the band was not apparent, viral genomes were likely present below the detection limit of DNA staining. The intensities of the viral bands were both greater and less than that of the host chromosomes suggesting, (1) a large variability in amount of viral production between RP lines and (2), given that the number of genomes in the sample is proportional to band intensity, the number of copies of the viral genome was not directly proportional to that of the host. The fact that the OtV5 genome copy number could be much lower than that of the host implies not every cell in the host population was carrying the virus genome. Higher OtV5 band intensities could correspond to viral genomes in virions as well as multiple viral genomes generated during replication in the cell.
Outlier chromosomes are found in all Mamiellales sequenced to date, and all SOCs encode genes with a similar set of predicted functions [7], although these genes have very little detectable sequence homology compared to the non-outlier chromosomes. Only 11% of genes on the SOCs of O. lucimarinus and O. tauri are orthologues, whereas the average proportion of orthologous genes is 86% on standard chromosomes [52]. We re-analysed the functional genes in all available Mamiellales genomes observing a consistent over-representation of genes involved in carbohydrate transport, synthesis and modification, particularly GTs (Table 1). Prasinoviruses are known to infect all clades of Ostreococcus, Micromonas, and Bathycoccus so far tested [29]. Furthermore, prasinovirus resistance has also been generated in Bathycoccus prasinos against BpV2 (Bathycoccus prasinos Virus 2) and Micromonas pusilla against MpV1 (Micromonas pusilla Virus1) in culture [28]. Given that differentially expressed genes on the SOC in O. tauri was strongly linked to a switch to a virus-immune state, we hypothesize the SOC has a similar role in virus resistance in these other genera of the Mamiellales.
Initially, we designed the experimental strategy to test whether resistant cells may arise spontaneously by mutation, and we planned to look for single nucleotide changes linked to resistance. However, the first visible signs of re-growth of resistant cells appeared about a week after inoculation in all of the cultures. Since cultures do not become visibly green until they reach at least 106 cells.ml-1, and assuming the cells have an optimum division rate under these conditions (1.4 divisions.day-1 in this growth chamber, M. Krasovec, personal communication), we find that about 1 in 1,000 cells may have become resistant at the time of or just after inoculation with the virus. This, however, may be an underestimation, since after a shock (such as after sub-culturing), there is usually a lag before re-growth. This frequency far exceeds the expected spontaneous mutation rate in O. tauri since comparison of the re-sequenced O. tauri genome between 2001 and 2009 revealed the fixation of eight single nucleotide substitutions and two deletions during the approximately 6,000 generations in the lab [4]. We thus preferred an alternative hypothesis, that resistance is induced by the biotic challenge of virus infection, for example by epigenetic modifications affecting gene expression patterns. Since the original culture was clonal, it seemed unlikely that a proportion of host cells was resistant at the moment of inoculation; although we cannot rule out that a reversible regulatory switch to a resistant state occurred in a small proportion of cells analogous to the minority of cells in RP lines that was lysed by OtV5. However, after a period of selection, resistance in the majority of cells appears to be stable since none of the resistant lines were lysed after re-infection over the course of the experiment.
In several loci, up-regulated genes on the LIRR related to carbohydrate metabolism and modification appeared to be spatially grouped. We speculate genes that are clustered together act on the same carbohydrate synthesis and glycosylation pathway. Notably, genes uniquely expressed in virus-resistant lines might dramatically alter the saccharide composition and glycosylation state of the cell. Several of the highly expressed GTs are associated with synthesis of surface glycans. Surface glycans are known to be important for host–virus interactions, where they mediate initial binding and recognition events of both host cells and pathogens at the cell surfaces [53]. In particular, rhamnan is a polymer important for bacterial host–virus interactions [54,55], and sulphated derivatives of rhamnans are known to have anti-viral activity in mammals [56]. The metazoan-derived CMP N-acetylneuraminic acid hydroxylase (ostta19g00610), in humans makes the influenza A virus receptor, N-acetylneuraminic acid, whose decorations affect receptor specificity [57]. Likewise, glycoprotein-N-acetylgalactosamine 3-beta-galactosyltransferase (ostta19g00120), known to function in the synthesis of extended mucin type O-linked glycans, has been associated with certain immune mediated diseases in humans [58]. Thus, expression of these genes may affect resistance through masking, or altering the O. tauri surface receptor used by OtV5 to inhibit viral adsorption. However, Thomas et al., [28] observed no statistical difference in OtV5 adsorption to RP, RNP and susceptible cells suggesting defective adsorption was not the mechanism for viral resistance. This may be due in part the very low proportion of O. tauri cells ‘competent’ to OtV5 adsorption at a given time point; even after inoculating susceptible O. tauri with high titres of OtV5, at most 20% of cells had visibly adsorbed virions despite the fact the majority of cells subsequently lyse [24]. One possible explanation is that the OtV5 receptor is only available during a defined point in host cell cycle, and as O. tauri cells are not perfectly synchronised in culture [59], adsorption is staggered over time making differences difficult to detect by a standard adsorption assay.
Alternatively, the over-expression of these specific host GTs may curtail OtV5 replication at another stage of the replication cycle, one possible candidate being virion assembly. In the model Chlorovirus, PBCV-1, the major capsid protein is glycosylated at least six sites and the glycans include rare monosaccharides whose biosynthesis genes are virus-encoded (see Piacente et al., [60] and references therein). As OtV5 similarly encodes numerous GTs and homologues to the PBCV-1 sugar metabolism enzymes, many of which are conserved in prasinoviruses (S3 Table), glycoconjugates are likely to be similarly essential to virion structure. Viral resistance may arise by the host switching carbohydrate metabolic pathways, thereby perturbing proper virion formation. This could be mediated, for instance, by feedback inhibition of viral sugar biosynthesis pathways by host generated products, directly altering the viral glycosyldonors/acceptors or simply altering the substrate pool. For example, virally produced GDP-D-rhamnose may be depleted by the host by redirecting it into extracellular rhamnan. Much remains unknown about the glycobiology of both prasinoviruses and Mamiellales and studies of their sugar composition, glycan structure and functional characterisation of GTs will help to pinpoint how the host–virus interaction is suppressed during resistance.
The link between resistance to prasinoviruses and alterations of glycosylation might extend to other Mamiellales, since there are numerous sugar modification genes clustered on the SOC in other members of the Mamiellales with sequenced genomes. They are extremely diverse, with different GT genes in each species, in keeping with the notion that they are important in viral specificity.
Thomas et al. [28] reported that chronically infected cells appeared to release viruses by budding, but after examination of many electron micrographs, we found that viral-induced lysis of host cells was occurring, which was further supported by the overall low level of OtV5 transcription in RP lines. RP cell lines are more likely to be unstably resistant, giving rise to a small proportion of revertant susceptible cells. Thus, apparent chronic virus production occurs by lysis of a minority of susceptible cells emerging in a majority of uninfected population at a frequency sufficient to propagate OtV5 over subsequent generations, although the serial cell transfers during the experiment tended towards eventual OtV5 extinction. This further suggests the two apparent kinds of resistance, RP and RNP, may reflect observing virus production at the level of a batch culture population and does not correspond to distinct mechanisms of immunity. Over time, revertant susceptible cells arising in cultures where OtV5 were absent could competitively exclude resistant strains, although this was not observed within the almost two-year duration of the study. In keeping with this observation RP cells have been shown to have growth rates similar to wild-type in culture, and differences in fitness could only be shown by competition between co-cultured strains [28].
It is tempting to speculate that the biotic stress induced, by an unknown mechanism, the activation of a transposon that mediated rearrangements in chromosome 19. Activation of LTR-retrotransposons during pathogen attack was observed in tobacco [61] and has since been observed following diverse kinds of biotic or abiotic stresses in diverse organisms (reviewed in [62–64]). This process may lie at the heart of genomic plasticity in response to environmental changes, since genome evolution might be facilitated by bursts of transposon activity [65]. In eukaryotic marine plankton, retrotransposons are probably the most abundantly expressed sequences [66]. The LTR of retrotransposons carry promoters that can affect the transcription of adjacent genes, although the effects of environmental conditions on LTR retrotransposons, which require transcriptional activity for their activation, are complex [62,67,68]. The putative TRIM on chromosome 19 (S1 Fig) may be activated in trans by the predicted complete version of this element, retrostreo2, on chromosome 8. However, since its reverse transcriptase (ostta08g00390) was down-regulated in resistant lines (S4 Table), it is not a likely candidate for causing changes in the SOC sequence. In addition, the tandem repeats surrounding the retroelement do not look like LTRs, they may be too long (2.3 kb, S1 Fig) and it seems unlikely that transcription of the TRR could influence the LIRR and the other distant and extensive set of resistance genes.
In Mamiellales, DNMT5 was proposed to function as the CpG maintenance methyltransferase [69] as DNMT1 and DNMT3 are lacking, the latter two enzymes being responsible for both de novo and maintenance CpG methylation in other organisms [70]. However, extensive CpG methylation occurs in nucleosome linkers and thus apparently functions to guide nucleosome positioning within the extremely compact nucleus. This raises the question of whether or not CpG methylation still plays a role in gene silencing in picoeukaryotes, albeit mediated by an alternative DNMT. A possible MT involved in epigenetic silencing is ostta19g00045, a predicted 5-cytosine methyltransferase located on chromosome 19 that was over-transcribed in the resistant lines (S2 Table). However, if indeed silencing of the LIRR of chromosome 19 in the susceptible control cells occurs by CpG methylation as in plants and animals [70], we would expect decreased activity of de novo CpG MT and increased activity of demethylases, which was not the case in this study. This suggests that silencing of chromosome 19 occurs via an alternative mechanism.
A more likely candidate is an up-regulated DDE superfamily endonuclease (Fig 4) shared between both outlier chromosomes (ostta19g00165 and ostta02g00275) that may mediate rearrangements. Type II transposons can carry silencer domains in the C-terminal region of their transposases [71] that not only repress their own expression, but can repress expression from adjacent genes, probably by histone modifications that change chromatin conformation. Expression of these transposons can be induced by certain transcription factors. Given that this gene lies at the heart of the over-transcribed region, it might explain both the transcriptional switch and the genome instabilities that we observe. However, it is also possible that an unknown mechanism up-regulating the inverted repeat region leads to de-repression of the predicted endonuclease, leading to the genomic instability observed. In O. lucimarinus both of the predicted orthologues of ostta19g00165 reside on the chromosome 2, but we have not yet examined the karyotypes of this species after viral infection. Three predicted transcription factors are up-regulated in resistant lines, ostta15g02170, containing a Myb-SANT domain of the REB1 family [72]; ostta01g00380, whose N-terminal regions shows weak similarity to RNA polymerase III subunit delta (BLAST to pfam12657, 1.39e-06) that is important in synthesis of some essential non-coding RNAs [73]; and ostta08g00865. Further experimental investigation would be required to determine whether they play a role in transcriptional activation of the SOC, since these conserved domains are important in numerous kinds of regulatory processes. We might thus speculate that one of these up-regulated transcription factors bind to the ostta19g00165 C-terminal domain after viral attack, inducing the adjacent glycosyltransferases, perhaps by changing chromatin configuration, and activating this transposon. Since the region is a large inverted repeat, an endonuclease target site might also be duplicated and physically distant, perhaps explaining the large size variations observed in this chromosome between independent resistant lines.
Given the presence of numerous repeated regions, and the reported occurrence of homologous recombination in O. tauri [12], a further alternative might be that the selective pressure for resistance to viruses promotes the growth of relatively rare cells in which genomic rearrangements have arisen by recombination, leading to changed genomic contexts and activation of the inverted repeat region. We have, however, never observed a change in the size of the O. tauri SOC in culture in our routine laboratory strain RCC4221 over a period of about 20 years at ~300 kb ([5,17] and this work). We were surprised to discover that extensive karyotypic changes had occurred in practically all of resistant lines entailing both increase and decrease in size. This suggests insertions/deletions of genetic material to the SOC after the acquisition of resistance to OtV5 is a consequence of activating resistance genes and not necessarily causative. Future work examining the SOC sequence in virus resistant O. tauri strains as well as in strains isolated from the environment and monitoring SOC changes during the critical period when resistance is established would inform which genetic changes occur and their relationship to viral resistance.
Several parallels with other genetic systems that involve biotic interactions between partners of symbionts or pathogens have come to light in our system. In the marine cyanobacterium Prochlorococcus, genes involved in resistance to viruses are clustered in hypervariable genomic islands [74] that facilitate rapid evolution of resistance and persistence of the host in some species [75]. As in Ostreococcus, in these regions most of these genes may be involved in cell surface interactions. Supernumerary chromosomes, which can vary in size and number between members of the same biological species, have been found in diverse fungal pathogens (reviewed in [76]). They often differ in base composition and structure with other autosomes of the same species, and they can carry genes encoding pathogenicity effectors, probably facilitating the rapid evolution of virulence. In Nectria haematococca, for example, pathogenicity effectors are clustered on a dispensable chromosome with an altered GC content that also carries transposons [77] and in Fusarium such chromosomes can be transferred horizontally between strains, broadening their host specificity [78]. In Ostreococcus, the tables are turned, since it is the host cell that shows genomic plasticity, whereas prasinovirus genomes show less nucleotidic variation between strains than those of their host genomes [26].
We speculate that the evolution of the small outlier chromosome as a viral immunity chromosome in the Mamiellales may have arisen following selection for small genome size, since genome compaction probably led to loss of genes involved in the silencing of viral attack by small RNAs, subsequently forcing the evolution of the defence mechanism typical to this group of algae. Small RNAs are central for epigenetic modifications and genome stability in eukaryotes in general [79,80], and in higher plants mechanisms of resistance to viral infection have been widely studied, where small RNAs have a central role in controlling viral infections (see [77] for a review). Although Ostreococcus is in the green lineage, most of the analysed genomes of algae in this group lack detectable homologues of the canonical genes Argonaut and Dicer thought to be important for their fabrication [81], possibly because it lost these genes during evolution of its extremely small genome that may result from “streamlining” in its marine environment [82]. Loss or gain of the biochemical machinery for fabrication of different kinds of small RNA has led to diversification in eukaryotic systems, and in some well-studied fungal systems loss of gene functions might provide clues about alternative pathways for dealing with viral assault [83].
To our knowledge, this is the first time that a chromosome specialized in defence against viruses has been described. Our results throw some light on the variability in the size of the SOC observed between different wild-type isolates of the O. tauri population in coastal north-western Mediterranean Sea [17]. These waters are indeed infested with prasinoviruses [84], and prasinoviruses are similarly abundant in the English Channel, where they infect Micromonas spp. [85]. While the precise mechanism for resistance remains to be elucidated by subsequent experimental investigations, we show that challenging cultures with a compatible virus reliably leads to extensive chromosome rearrangements, providing an opportunity to study natural genome evolution in controlled laboratory conditions in O. tauri. The observed rearrangements in the O. tauri SOC may be produced by an adaptive mechanism that quickly generates genetic variants able to evade viral lysis. Clustering of the genetic features observed in the SOC helps to explain several of its unusual characteristics. In particular, a faster rate of evolution than the other chromosomes, evidenced in the high level of species-specific genes. The conservation of the same functional gene categories, but not of gene homologues, would also be consistent with rapid “arms-race” evolution against the particular prasinoviruses that infect each species and implies the SOC as a whole has a conserved function in viral immunity.
The strains used in the experiment were O. tauri RCC4221 (Roscoff Culture Collection [86] and Ostreococcus tauri virus 5 (OtV5). Liquid L1 medium (NCMA, Bigelow Laboratory for Ocean Sciences, USA) was used for culturing throughout and prepared using autoclaved seawater from offshore Banyuls Bay (MOLA station: 42°27'11''N, 3°8'42''E) diluted 10% with MilliQ water and filtered prior to use through 0.22 μm filters. All cultures were maintained under a 12:12 hour light/dark regime in 100 μmol photon m-2 s-1 white light at 20°C. To grow O. tauri on a solid medium, molten agarose (1.5%) equilibrated to 60°C was added to cultures to give a final agarose concentration: 0.15% and immediately poured into 9 cm diameter petri dishes. Single O. tauri colonies were picked from the solid medium using a sterile pipette tip and placed into liquid L1. OtV5 plaques were produced by mixing serial dilutions of OtV5 lysate with O. tauri RCC4221 culture and growing on solid medium [24]. A stock OtV5 lysate used in all subsequent infections was produced by randomly selecting a plaque, inoculating it into an O. tauri RCC4221 culture, purifying the resulting lysate (centrifugation at 8,000 g for 20 min and filtration through 0.2 μm) and storing at 4°C.
To test viral susceptibility, stock OtV5 lysate was added to O. tauri lines growing in 48-well plates with the precursor susceptible O. tauri RCC4221 serving as a positive control. Lines were considered susceptible if visible lysis occurred within approximately one week. To test if O. tauri resistant lines were producing infective virus, an aliquot of the culture medium from actively growing cells was sampled, then centrifuged at 8,000 g for 20 min to remove O. tauri cells. The cell-free medium was added to susceptible O. tauri RCC4221 growing in 48-well plates. O. tauri lines were considered to be producing virus if visible lysis occurred in the susceptible line within approximately one week.
DNA was extracted from the initial 1 L culture of the O. tauri from which all subsequent lines were cloned (Fig 1) using a CTAB extraction method [34]. GATC Biotech performed whole genome sequencing using six single molecule real time (SMRT) cells on the PacBio RS II platform and de novo assembly with the Hierarchical Genome Assembly Process (HGAP, Pacific Biosciences). The final updated chromosome 19 sequence was obtained by scaffolding the HGAP-generated unitigs in Geneious [87,88]. Before total RNA extraction, O. tauri culture lines were subjected to antibiotic treatment to reduce the amount of bacteria in the cultures according to [89]. The antibiotic treated culture was used to inoculate a 200 mL culture for RNA extraction. Cultures were grown to exponential phase (11–32 million cells ml-1) and cells harvested by centrifugation at 8,000 g for 20 min in a swinging bucket rotor during the first 3 hours of the light cycle. Total RNA was extracted using the Direct-zol RNA kit (Zymo Research). Cell densities for each culture line are recorded in S1 Table. Selection for polyadenylated RNA, library preparation and sequencing was performed at GeT (Génome et Transcriptome, GENOTOUL, Toulouse, France). RNA libraries were sequenced on the Illumina Hi-Seq 2000 platform by multiplexing all samples on a single flowcell lane, which generating paired end reads of 101 bp in length. RNA sequence reads were checked for quality using FastQC [90].
In order to have a balanced experimental design and to control for batch effects, four OtV5-resistant and four control lines that were from matching RNA extraction batches were selected for differential gene expression analysis (S1 Table). Transcriptome read pairs (fragments) were aligned using TopHat2 [91] (alignment parameters: -i 10 -I 4000, -G) to the annotated genome sequence of O. tauri RCC4221 [4]. The counts of fragments aligning to each gene was determined using the htseq-count function of HTSeq [92] with parameters (-t CDS -s no -m intersection-nonempty). Exploratory and differential gene expression analyses were performed on fragment count tables using the R package DESeq2 [93]. Hierarchical clustering of the distance between transcriptome profiles after correction for batch effects separated resistant and susceptible lines in two groups (S5 Fig). For differential gene expression, RNA batch number was used as a secondary design variable to adjust for batch effects and the primary variable tested was virus resistance vs susceptibility using the DESeq function accepting genes as significantly differentially transcribed with adjusted p-value <0.1.
PFGE and in-gel hybridization was conducted as previously described [94]. Briefly, cell cultures were grown to mid-exponential phase, harvested by centrifugation at 8000 g for 20 min and cells were resuspended in TE buffer (10 mM Tris-HCl, 125mM EDTA, pH 8) at a density of 8.7 × 108 cells ml-1. Electrophoresis was performed in 0.8% agarose gels in 0.5× TBE buffer (44.5 mM Tris, 44.5 mM boric acid, 1 mM EDTA at pH 8) using the CHEF-DR III (Bio-Rad) system. For each sample, 2 mm of plug was loaded into the wells. Electrophoresis was run at 6 V cm-1 at 14°C with 120° pulse angle for 15 h with a switch time of 60 s and followed 9 h at a switch time of 90 s. After PFGE, DNA was chemically denatured and dehydrated in a vacuum dryer. PCR was used to amplify a 1681 bp region of the O. tauri gene ostta19g00640 (5052 bp), located on chromosome 19, with the primer pair 19e_Fw 5´-GCGATGCGGTGCTCTACC-3´ and 19e_Rv 5´CGTGGAGTTATCCCCGAACC-3´ (PCR conditions described in [17]. This gene was chosen as a probe for chromosome 19 because it was consistently transcribed between resistant and susceptible lines (S6 Fig), and thus was not deleted and it has a putative orthologue in the SOC of O. lucimarinus suggesting a conserved function. After gel purification of the PCR product (Wizard-Prep, kit, Machery-Nagel), the amplicon was randomly labelled with [α-32P]CTP (Perkin-Elmer) according to manufacturer’s instructions (Prime-a-Gene kit, Promega) for use as a DNA probe. Dried gels were equilibrated in hybrization buffer (6× SSC, 5× Denhardt's solution, 0.1% (w/v) sodium dodecyl sulphate, 10 μg ml-1 tRNA), radiolabelled probe was added, hybridized overnight at 65°C and the gel exposed to radiographic film.
Sequence data used in this study can be found in the GenBank data libraries. O. tauri RCC4221 chromosome sequences are under accession numbers CAID01000001.2 to CAID01000020.2 [4] except for the updated chromosome 19 sequence (this study), which is available from http://wwwphp.obs-banyuls.fr/publications/data/2/. Transcriptome data is available from PRJNA344946. The updated genome sequence and annotation of OtV5 [24] is under accession EU304328.2. Gene models and annotations for O. tauri RCC4221 are also available from the Online Resource for Community Annotation of Eukaryotes (ORCAE) under Ostreococcus tauri V2 [95, 96].
Transmission electron microscopy was performed as previously described in [24] with some modifications. Briefly, 200 mL cultures of O. tauri in exponential phase were fixed in 1% glutaraldehyde for 30 min and fixed cells were centrifuged for 30 min at 2500 g and the cell pellet quickly resuspended in molten (37°C) 1% low melting point agarose (agarose type II, Sigma) and allowed to set in a disposable micropipette (SMI, Emerville, CA, USA). The agarose cell plug was fixed in buffer of one volume of sodium cacodylate (0.4M) in two volumes of L1 medium containing 2.5% glutaraldehyde for 2 h, then washed in the same buffer without glutaraldehyde. Post fixation was performed in 1% OsO4 in sodium cacodylate (0.2M) for 1 h. After two 15 min washes in sodium cacodylate (0.2M), the agarose cell plug was cut into small pieces and dehydrated in ethanol and embedded in Epon 812 resin at 60°C for 48 h. Ultra-thin slices (80–90 nm) were placed on a 300 mesh grid and stained with uranyl acetate for 15 min, followed by lead citrate staining for 2 min then visualised with a Hitachi H 7500 transmission electron microscope.
|
10.1371/journal.pgen.1005824 | Specific Hsp100 Chaperones Determine the Fate of the First Enzyme of the Plastidial Isoprenoid Pathway for Either Refolding or Degradation by the Stromal Clp Protease in Arabidopsis | The lifespan and activity of proteins depend on protein quality control systems formed by chaperones and proteases that ensure correct protein folding and prevent the formation of toxic aggregates. We previously found that the Arabidopsis thaliana J-protein J20 delivers inactive (misfolded) forms of the plastidial enzyme deoxyxylulose 5-phosphate synthase (DXS) to the Hsp70 chaperone for either proper folding or degradation. Here we show that the fate of Hsp70-bound DXS depends on pathways involving specific Hsp100 chaperones. Analysis of individual mutants for the four Hsp100 chaperones present in Arabidopsis chloroplasts showed increased levels of DXS proteins (but not transcripts) only in those defective in ClpC1 or ClpB3. However, the accumulated enzyme was active in the clpc1 mutant but inactive in clpb3 plants. Genetic evidence indicated that ClpC chaperones might be required for the unfolding of J20-delivered DXS protein coupled to degradation by the Clp protease. By contrast, biochemical and genetic approaches confirmed that Hsp70 and ClpB3 chaperones interact to collaborate in the refolding and activation of DXS. We conclude that specific J-proteins and Hsp100 chaperones act together with Hsp70 to recognize and deliver DXS to either reactivation (via ClpB3) or removal (via ClpC1) depending on the physiological status of the plastid.
| In this paper we report a relatively simple mechanism by which plant chloroplasts deal with inactive forms of DXS, the main rate-determining enzyme for the production of plastidial isoprenoids relevant for photosynthesis and development. We provide evidence supporting that particular members of the Hsp100 chaperone family contribute to either refold or degrade inactive DXS proteins specifically recognized by the J-protein adaptor J20 and delivered to Hsp70 chaperones. Our results also unveil a J-protein-based mechanism for substrate delivery to the Clp complex, the main protease in the chloroplast stroma. Together, this work allows a better understanding of how chloroplasts get rid of damaged DXS (and potentially other proteins), which should contribute to take more informed decisions in future approaches aimed to manipulate the levels of plastidial metabolites of interest (including vitamins, biofuels, or drugs against cancer and malaria) in crop plants.
| Organelles like mitochondria and plastids play fundamental roles in all eukaryotic organisms. In particular, plastids were acquired by a symbiosis between photosynthetic cyanobacteria and eukaryotic cells. Today, plastids (like mitochondria) are intimately integrated into the metabolism of plant cells but they still remain as separate functional entities that regulate their own biochemistry by relatively independent mechanisms. An important part of this regulation relies on the effective control of plastidial enzyme activities. Most of the enzymes required for plastidial metabolism are encoded by nuclear genes, synthesized in precursor form in the cytosol, and transported into plastids using energy-dependent import machineries [1]. Following import, specific proteases cleave the transit peptides and complex networks of plastidial chaperones ensure proper folding, assembly, or suborganellar targeting of the mature proteins. Chaperones and proteases are also essential components of the protein quality control (PQC) system that promotes the stabilization, refolding, or degradation of mature proteins that lose their native conformation and activity after metabolic perturbations or environmental challenges such as excess light, temperature peaks, oxidative stress or nutrient starvation [2,3]. While plant plastids contain many groups of prokaryotic-like chaperones (such as Hsp70 and Hsp100) and proteases (including Clp, Lon, Deg, and FstH), their specific targets and PQC-related roles remain little studied [1–4].
Due to the presence of plastids, plants have biochemical pathways that are not found in other eukaryotic kingdoms. For example, isoprenoid precursors are produced by the methylerythritol 4-phosphate (MEP) pathway in bacteria and plant plastids, whereas animals and fungi synthesize these essential metabolites using a completely unrelated pathway which is also used by plants to produce cytosolic and mitochondrial isoprenoids [5,6]. MEP-derived isoprenoids include compounds essential for photosynthesis (such as carotenoids and the side chain of chlorophylls, tocopherols, plastoquinone and phylloquinones) and growth regulation (including the hormones gibberellins, cytokinins, strigolactones and abscisic acid). Many plastidial isoprenoids also have nutritional and economic relevance [6]. All MEP pathway enzymes are located in the plastid stroma [5,7]. While transcriptional regulation of genes encoding biosynthetic enzymes is known to exert a coarse control of the MEP pathway, fine-tuning of metabolic flux appears to rely on post-transcriptional or/and post-translational regulation of enzyme levels and activity [8–12]. This is most evident for deoxyxylulose 5-phosphate synthase (DXS), the homodimeric enzyme that catalyzes the first step of the pathway. Metabolic control analysis calculations confirmed that DXS is the enzyme with the highest flux control coefficient (i.e. the main rate-determining step) of the MEP pathway [13]. Consistent with this prime regulatory role, DXS activity is tightly regulated by several post-translational mechanisms [10–12]. In particular, DXS enzymatic activity is allosterically inhibited by MEP pathway products [14,15], which also repress DXS protein accumulation [8,14,16–18]. Mathematical modeling recently showed that the post-translational control of DXS protein abundance and enzyme activity is crucial for the adjustment of the MEP pathway flux to persistent changes in environmental conditions, such as substrate supply or product demand [18]. Despite the central relevance of this type of regulation, little is known about the molecular mechanisms behind it.
We previously showed that the Arabidopsis thaliana J-protein J20 interacts with inactive forms of DXS to deliver them to the Hsp70 chaperone for eventual activation (which involves folding or refolding) or degradation (which involves unfolding) [19]. However, the particular protease involved and the specific components of the two J20-dependent antagonistic pathways remained unknown. Here we show that DXS is primarily degraded by the Clp protease complex through a pathway involving J20 and Hsp100 chaperones of the ClpC type. We also demonstrate that Hsp70 can physically interact with ClpB3, another plastidial Hsp100 chaperone, to promote the activation of non-functional DXS enzymes.
The main protease families involved in the degradation of terminally damaged or surplus proteins in plastids are Clp, Lon, Deg, and FtsH, all of them of prokaryotic origin [3,4]. We and others have previously shown that Arabidopsis mutants with a decreased activity of the stromal Clp protease complex display an accumulation of several MEP pathway enzymes, including DXS [20–24]. However, whether other plastidial proteases involved in PQC networks could also contribute to DXS degradation in the stroma remains unexplored. Several functional Lon homologues are found in Arabidopsis, but only Lon1 [25] and Lon4 [26] have been localized to chloroplasts, where they are attached to the stromal side of thylakoids. The Deg gene family in Arabidopsis contains 16 members, with 5 of them experimentally confirmed to be localized in chloroplasts [27]. From these, the isoforms Deg1, Deg5 and Deg8 were found in the thylakoid lumen, whereas Deg2 and Deg7 were detected in the stromal side [28,29]. FtsH proteases are encoded by 12 genes in Arabidopsis, and 9 of them can be found in chloroplasts [30]. The four major chloroplast isomers (FtsH2, FtsH5, FtsH8 and FtsH1, in order of abundance) have been shown to reside in the thylakoid membrane with their catalytic domain facing the stromal side [31–33].
DXS protein levels were examined by immunoblot analysis in Arabidopsis wild-type (WT) plants and single mutants defective in plastidial Lon, Deg, or FtsH isomers (Fig 1 and S1 Table). As a control, we included the Clp protease mutant clpr1, which displays a reduction of other subunits of the Clp proteolytic core [34] but increased DXS protein levels [20]. As shown in Fig 1A, DXS protein levels in the analyzed mutants were similar to those in WT plants with only three exceptions. Lines defective in Lon1 and Deg7 showed a decreased accumulation of the protein compared to the WT, whereas DXS levels were only increased in the clpr1 mutant (Fig 1A). No changes in DXS transcript levels were detected in any of the mutant lines (Fig 1A). Since enzyme levels would be expected to be post-translationally increased (but not decreased) in the mutants impaired in DXS-degrading proteases, we conclude that the Clp complex is likely the primary protease for DXS removal. The contribution of Lon, Deg, or FtsH proteases, however, cannot be fully discarded as we only tested mutants for individual isoforms of those other proteases and it is possible that different isoforms may have redundant functions.
Clp proteases are found in almost all bacteria and endosymbiotic organelles (mitochondria and plastids). In bacteria (S1 Fig), they are formed by a barrel-like catalytic core of two heptameric rings of proteolytic subunits (ClpP) and a dynamically interacting hexameric ring of Hsp100 chaperones (ClpA and ClpX in Escherichia coli; ClpC, ClpX, and ClpE in Bacillus subtilis) that unfold substrates for translocation into the proteolytic chamber [35]. Additionally, interaction of Hsp100 members with adaptor proteins (such as ClpS and SspB in E. coli and ClpS, MecA, McsB, and others in B. subtilis) enhance or expand substrate specificity [35–37]. In plant plastids, the Clp protease is more complex [38,39] but the basic components are conserved (S1 Fig). It presents a protease core (formed by two heptameric rings of plastome encoded ClpP1 and nuclear-encoded ClpP3-P6 and ClpR1-R4 subunits) stabilized by two plant-specific subunits (ClpT1-T2). The Arabidopsis homologues of the bacterial ClpA and ClpC unfolding chaperones are ClpC1, ClpC2, and ClpD. A ClpS adaptor is also found in chloroplasts [40], where it might form a plant-specific binary adaptor complex with the ClpF protein [41]. The possibility of other pathways delivering proteins to the Clp protease, however, remains open.
As shown above (Fig 1A), DXS levels increase in mutants defective in Clp protease activity such as clpr1 [20]. If DXS is targeted to the Clp protease for degradation, we would also expect a post-translational upregulation of DXS enzyme levels in mutants impaired in the adaptors and chaperones that deliver the protein to the Clp catalytic core. A systematic analysis of such mutants (clps, clpc1, clpc2, clpd, clpt1, and clpt2) showed that only those defective in ClpC1 accumulated higher levels of DXS protein than WT plants (Fig 1B and S2 Fig). Quantification of DXS-encoding transcripts in the same mutant lines showed WT levels in all cases (Fig 1B), confirming that the observed accumulation of DXS polypeptides in ClpC1-defective lines was not a consequence of increased gene expression.
It has been proposed that the two Arabidopsis ClpC paralogs ClpC1 and ClpC2 perform similar if not identical functions in the chloroplast [42]. However, proteolytic assays with known Clp protease substrates only showed a greatly reduced degradation rate in clpc1 plants [42], which showed the strongest reduction in total ClpC content (Fig 1B and S2 Fig). Estimation of DXS degradation rates upon treating WT and mutant plants with the protein synthesis inhibitor cycloheximide also showed a slower proteolytic removal of DXS polypeptides in clpc1 mutants (Fig 2A). As expected, a defective Clp catalytic core in the clpr1 mutant led to similarly reduced DXS degradation rates (Fig 2A), again supporting our conclusion that DXS is a target for this proteolytic complex. To confirm whether DXS might be a ClpC1 substrate, tagged versions of the Arabidopsis proteins (DXS-GFP and ClpC1-MYC) were overproduced in Nicotiana benthamiana leaves by agroinfiltration and co-immunoprecipitation assays were next performed. As shown in Fig 2B, these assays confirmed that DXS and ClpC1 can indeed interact. Together, we conclude that DXS might be mainly unfolded by ClpC1 for degradation by the Clp proteolytic core.
Recent results have shown that client proteins of the stromal Clp protease are recognized and delivered to ClpC chaperones by ClpS and ClpF adaptors [40,41]. While DXS might actually be a target of ClpS in bacteria [43], a wild-type phenotype in terms of DXS protein levels was observed in Arabidopsis plants defective in the proposed chloroplast adaptors (Fig 1B) [40,41]. Although ClpC could possibly directly deliver client proteins to the Clp protease without the need of an adaptor, we reasoned that further substrate specificity should be achieved using an alternative ClpS/ClpF-independent adaptor system.
Our previous work showed that inactive forms of DXS are recognized by J20, a J-protein adaptor that delivers them to the Hsp70 chaperone [19]. Computational analysis of the Arabidopsis DXS monomer with the Aggrescan3D algorithm revealed the presence of several aggregation-prone clusters (S3 Fig). Consistent with the conclusion that DXS tends to aggregate and that J20 prevents its aggregation, GFP-tagged DXS proteins accumulate in plastidial speckles that are larger in j20 plants (S4 Fig) [19]. In addition, the endogenous DXS enzymes are less accessible to proteinase K cleavage in the j20 mutant (S4 Fig), again suggesting that DXS aggregation is increased in the absence of J20, likely because the delivery of aggregated (and hence inactive) DXS proteins to the Hsp70 chaperone is impaired. The main role of Hsp70 is actually to prevent the formation of toxic aggregates of damaged proteins and, together with Hsp100 chaperones, promote their solubilization [44–50]. However, Hsp70 chaperones also facilitate the transfer of irreparably damaged client proteins to proteolytic systems [49,51–53]. For example, cytosolic Hsp70 is involved in the degradation of Arabidopsis protein clients by the eukaryotic 26S proteasome [51].
Despite the absence of conserved domains for direct interactions between Hsp70 and ClpC-type Hsp100 proteins (S5 Fig) [36,45,46], co-immunoprecipitation experiments showed that both chaperones can be found together in the chloroplast envelope [54,55]. It is therefore possible that Hsp70 and ClpC might interact either directly (using unidentified chaperone binding motifs) or indirectly (via third partners) to participate in PQC events at the stromal side of the inner envelope membrane [1,42,56,57]. Because in Arabidopsis the two plastidial isoforms of Hsp70 (Hsp70.1 and Hsp70.2) and ClpC (ClpC1 and ClpC2) are also found in the stroma [42,58], we reasoned that Hsp70 and ClpC proteins might collaborate to deliver DXS to the Clp protease using J20 as an adaptor. Interestingly, overexpression of J20 in transgenic Arabidopsis plants leads to decreased DXS protein levels, whereas loss of J20 function causes a reduced degradation rate of the enzyme (Fig 2C) [19]. Since both the J20 adaptor and ClpC chaperones are involved in the control of DXS degradation, we next tested whether they might function in the same pathway. We followed a genetic strategy based on comparing the DXS accumulation phenotype of single mutants defective in either J20 or ClpC1 with that of double j20 clpc1 mutants (Fig 3). All three mutants accumulated higher levels of DXS proteins (but not transcripts) compared to WT plants. In particular, DXS levels increased ca. 2-fold in j20 plants and 4-fold in single clpc1 and double j20 clpc1 mutants (Fig 3A). The absence of an additive or synergistic phenotype in the double mutant supports the conclusion that J20 and ClpC1 actually function in the same pathway delivering DXS to degradation in Arabidopsis plastids. Such a ClpS/ClpF-independent pathway could potentially be functioning for other plastidial clients of J-proteins. However, the lack of bona-fide substrates for other plastidial J-proteins prevents to experimentally testing this possibility at the moment.
The results described above suggest that damaged DXS polypeptides might be recognized by J20 and then delivered to Hsp70 and ClpC chaperones for unfolding and degradation by the Clp proteolytic core. But unlike that observed for J20-defective mutants [19], DXS accumulates in a mostly active form in clpc1 mutants (Fig 3B). Thus, measurement of DXS activity in plant extracts showed increased total activity but unchanged specific activity (i.e. relative to protein levels) in clpc1 compared to WT controls (Fig 3B). Estimation of DXS activity in planta by quantification of the resistance to clomazone (CLM), a specific DXS inhibitor [19,59,60] further supported the presence of higher DXS activity levels (i.e. increased resistance to the inhibitor) in the clpc1 mutant, opposite to the increased sensitivity detected in the case of j20 plants (Fig 3C and S6 Fig). To reconcile these results, we propose that loss of J20 causes an accumulation of aggregated (i.e. enzymatically inactive) DXS because the protein cannot be normally reactivated (refolded) or discarded (degraded). By contrast, accumulation of active DXS enzyme when ClpC activity decreases might be due to the existence of a functional pathway to disaggregate and refold the excess protein that cannot be degraded in the clpc1 mutant. The participation of J20 in such a putative reactivation pathway is supported by the observation that loss of J20 function in degradation-impaired (ClpC1-defective) j20 clpc1 plants results in a higher proportion of inactive DXS enzyme (Fig 3B) and hence a reduced resistance to CLM (Fig 3C and S6 Fig) compared to the single clpc1 mutant.
Work in different systems has shown that Hsp70 can be assisted by Hsp100 chaperones of the ClpB type in the solubilization of toxic aggregates of damaged proteins [45–50,61–64]. ClpB3 is the only ClpB-type chaperone found in Arabidopsis plastids [44]. Unlike the rest of plastidial Hsp100 chaperones present in this plant (ClpC1, ClpC2, and ClpD), ClpB3 lacks the IGF motif (or ClpP-loop) required for interaction with proteolytic subunits of the Clp core but it harbors a domain responsible for the interaction with Hsp70 chaperones (S5 Fig) [36,45,46]. Interestingly, the levels of ClpB3 were increased in mutants defective in Clp protease subunits, including ClpC1 [21,22,24,40,65] (Figs 1B and 3A and S2 Fig), suggesting that ClpB3 might contribute to mitigate protein folding stress caused by a defective Clp protease activity. In agreement, impairment of both ClpB3 and Clp protease activity results in a seedling lethal phenotype [22]. Based on these data, we speculated that ClpB3 might also participate in the DXS reactivation pathway mediated by J20 and Hsp70 chaperones. To evaluate this possibility, we first analyzed DXS protein levels and activity in ClpB3-defective Arabidopsis plants (Fig 4). If ClpB3 promotes DXS protein disaggregation (and hence activation), it was expected that clpb3 mutants would show a transcription-independent accumulation of inactive forms of DXS, assuming that the degradation rate of J20-delivered proteins would remain constant. Indeed, clpb3 plants showed a WT rate of DXS degradation (Fig 2C) but an enhanced accumulation of DXS enzyme without changes in transcript levels (Fig 4A). Also as predicted by our model, the specific activity of the DXS protein found in the ClpB3-defective mutant was much lower than that measured in WT plants (Fig 4B). Loss of both ClpB3 and J20 activities in the double j20 clpb3 mutant resulted in an even higher accumulation (Fig 4A) of mostly inactive DXS protein (Fig 4B), presumably because the absence of J20 prevents the targeting of non-functional enzymes to ClpC for eventual degradation by the Clp protease. The dramatic phenotype displayed by single clpb3 and double j20 clpb3 mutant plants (Fig 4C) [66] prevented the reliable quantification of their CLM resistance. In any case, the available data suggests that when the proteolytic degradation of inactive (e.g. aggregated) forms of DXS delivered to the Clp protease by J20 via ClpC is impaired (e.g. in clpr1 and clpc1 mutants), an increase in ClpB3 levels promotes the disaggregation and activation of the enzyme, eventually resulting in higher levels of enzymatically active DXS. When J20 activity is missing, however, inactive DXS forms cannot be properly reactivated via ClpB3 (as deduced from the similar levels of DXS protein but lower proportion of active enzyme found in the double j20 clpc1 mutant compared to the single clpc1 line; Fig 3) or degraded via ClpC (as deduced from the increased levels of inactive DXS protein present in double j20 clpb3 plants compared to the single clpb3 mutant; Fig 4).
As described above, the mechanistic basis for the collaboration between J20, Hsp70, and ClpC chaperones is currently unknown. However, the presence of a Hsp70-binding motif in the amino acid sequence of ClpB3 (S5 Fig) suggests that plastidial Hsp70 isoforms might be able to directly interact with ClpB3 to synergistically activate damaged DXS proteins recognized by the J20 adaptor. In agreement with this possibility, the ClpB3 protein was efficiently immunoprecipitated from WT extracts using an anti-Hsp70 serum (Fig 4D). When a similar experiment was performed with the Arabidopsis hsp70.2 mutant, previously shown to contain lower amounts of plastidial Hsp70 proteins than the WT [58], the level of immunoprecipitated ClpB3 protein was concomitantly decreased (Fig 4D). These results confirm that plastidial Hsp70 isoforms can be found together with ClpB3 in Arabidopsis chloroplasts, providing a mechanistic frame for the observed collaboration between these two families of chaperones in the J20-mediated activation of DXS.
The results described above are consistent with a model involving the participation of ClpB3 and ClpC1 on opposite pathways resulting in either reactivation or degradation, respectively, of inactive DXS proteins recognized by the Hsp70 adaptor J20. Under normal growth conditions, the levels of ClpB3 transcripts and protein are lower than those of ClpC1 (S7 Fig) [42,67]. However, ClpB3 transcript levels have been shown to strongly increase upon exposure to high temperatures [66,68,69] whereas virtually no changes in RNA or protein levels have been detected for ClpC1 or ClpC2 in response to heat or other types of stress, including cold, drought, salt, and oxidative stress [66,70]. The ratio between plastidial ClpB3 and ClpC1 chaperones (and hence the potential capacity to reactivate damaged or/and aggregated DXS polypeptides) could therefore increase when plants are challenged with at least some types of stress (S7 Fig).
DXS-derived isoprenoids such as carotenoids and tocopherols protect plants against oxidative stress, whereas others (including chlorophylls and prenylated quinones) are essential for photosynthesis. Therefore, a decreased production of these isoprenoids (e.g. upon down-regulating DXS activity) is expected to trigger a stress response. We observed that a specific reduction in DXS activity in Arabidopsis WT plants germinated and grown in the presence of CLM caused an increased accumulation of ClpB3 but not ClpC chaperones compared to controls grown in the absence of inhibitor (Fig 5). A similar ClpB3 protein accumulation response was also observed in mutants with a defective MEP pathway (Fig 5). As previously observed [8,14,16–18], the pharmacological or genetic blockage of the pathway also resulted in increased accumulation of DXS protein. Most interestingly, the DXS and ClpB3 accumulation response was detected as soon as 5 hours after reducing the MEP pathway flux by treatment with specific inhibitors (Fig 5). We therefore conclude that stress situations (including those causing a decreased DXS activity and/or MEP pathway flux) could rapidly trigger an increased accumulation of ClpB3, but not ClpC chaperones, likely aimed to promote the reactivation pathway that would keep DXS enzymes in an enzymatically active condition. Furthermore, our data show that ClpB3 levels are more prone to change compared to those of ClpC proteins, suggesting that ClpB3 concentration might be a major factor regulating the fate of inactive DXS polypeptides recognized by J20 and delivered to Hsp70.
Based on the presented data, we propose a model for the regulation of DXS enzyme levels and activity by different types of plastidial chaperones (Fig 6). According to this model, J20 (a plastidial member of the J-domain protein family, also known as J-proteins or Hsp40 co-chaperones) acts as an adaptor providing substrate specificity [19]. In particular, J20 delivers inactive DXS proteins to Hsp70 chaperones that would next act together with particular Hsp100 proteins to either degrade (ClpC1) or reactivate (ClpB3) the enzyme (Fig 6A). J20 might recognize DXS polypeptides that remain unfolded after plastid import or become misfolded by ordinary perturbations and eventually aggregate (S3 and S4 Figs), a process that would render the protein more insoluble and enzymatically inactive. Under normal growth conditions, most DXS proteins remain soluble but some are indeed found associated to the insoluble fraction (Fig 6B). This might be due to the relative low levels of ClpB3 relative to ClpC1 (S7 Fig) [42,67]. In agreement, a further reduction in ClpB3 levels (e.g. in the clpb3 mutant) results in a higher proportion of DXS protein associated to the insoluble fraction (Fig 6B) and hence inactive (Fig 4). By contrast, an enhanced accumulation of ClpB3 takes place in stress situations (Fig 5 and S7 Fig) or when Clp protease function is impaired (Figs 1B and 3A and S2 Fig) [21,22,24,40,65], likely aimed to mitigate general protein folding stress. In the case of DXS, a reduced degradation rate in the clpc1 mutant (Fig 2A) results in increased levels of active (soluble) enzyme (Figs 3 and 6B) likely because a higher accumulation of ClpB3 prevents DXS aggregation. Similar to that proposed in other systems [47–50,61–64], ClpB3 directly interacts with Hsp70 to synergistically perform this role (Fig 4). The observed changes in DXS protein levels and solubility appear to be highly specific, as the next enzyme of the MEP pathway (Fig 5A), deoxyxylulose 5-phosphate reductoisomerase (DXR), was found to be essentially soluble in WT and Hsp100-defective mutants (Fig 6B) and to remain unchanged in J20-defective plants [19].
Interaction with CHIP, a co-chaperone that functions as an E3 ubiquitin ligase, converts Hsp70 from a protein-folding machine into a degradation factor that targets unfolded substrates for degradation by the eukaryotic 26S proteosome [51,71]. Based on genetic evidence (Fig 3) and published results that ClpC and Hsp70 chaperones can be found together in plastid complexes [54,55], we propose that Hsp70 and ClpC chaperones could somehow interact (either directly or by means of unidentified partners) to deliver client proteins like DXS to the Clp catalytic complex. In summary, our model (Fig 6A) proposes that collaboration of Hsp70 with Hsp100 chaperones might deliver inactive (misfolded or/and aggregated) forms of DXS (and potentially many other plastidial proteins recognized by specific J-proteins, the substrate adaptors for Hsp70) to either refolding (via ClpB3) or degradation (via ClpC chaperones). The seedling lethal phenotype of double mutants with no ClpB3 and Clp protease activity [22] illustrates the key relevance of these two seemingly antagonistic pathways for plant life. We speculate that taking a specific pathway (i.e. deciding whether to repair or degrade the protein) might depend on the relative abundance of these Hsp100 partners, particularly as a consequence of changes in ClpB3 levels. The main reason behind the existence of such sophisticated and expensive pathways for the regulation of DXS levels and activity is likely to be the major role demonstrated for this enzyme in the control of the MEP pathway flux [10,11,13]. Future work should next determine how the collaboration of different sets of plastidial chaperone types, and hence the fate of the client protein, is specifically regulated.
Arabidopsis thaliana mutant lines used here are indicated in S1 Table (all in the Columbia background). Sibling lines expressing 35S:DXS-GFP in WT and j20 backgrounds were previously generated [19]. Seeds were surface-sterilized and germinated on solid Murashige and Skoog (MS) medium supplemented with 1% sucrose. Plants were grown under long day conditions as described [19]. For cycloheximide experiments, seeds were germinated on top of a sterile disc of synthetic fabric (SefarNitex 03-100/44). At day 7, the disc with the seedlings was transferred to fresh medium supplemented with 100 μM cycloheximide and samples were collected at different times afterwards (up to 12h) for immunoblot analysis. Inhibition of protein synthesis with cycloheximide had no visual effect on treated seedlings at the times used for the experiment (S8 Fig). Treatments with MEP pathway inhibitors were performed by transferring discs with 7-day-old seedlings to fresh medium supplemented with 10 μM clomazone (CLM) or 100 μM fosmidomycin (FSM). CLM resistance was estimated by quantifying chlorophyll levels in the presence of increasing concentrations of the inhibitor as described [60].
For transient expression and co-immunoprecipitation assays, an Arabidopsis full-length cDNA encoding ClpC1 without the stop codon was PCR-amplified, cloned into the pDONOR207 vector (Invitrogen), and subcloned into the Gateway vector pGWB417 to be expressed under the 35S promoter with a C-terminal MYC epitope (35S:ClpC1-MYC construct). A 35S:DXS-GFP construct was available in the lab [19]. Transient expression of these constructs was carried out by agroinfiltration of Nicotiana benthamiana leaves using the Agrobacterium GV3101 strain. Samples for immunoprecipitation were collected after 3 days.
Protein extracts were obtained from whole plants and used for immunoprecipitation assays or/and immunoblot analysis as described [19]. For the separation of soluble and insoluble (with protein aggregates) fractions, native protein extracts were obtained in a buffer containing 100 mM Tris-HCl pH7.9, 10 mM MgCl2, 1% (v/v) glycerol, and 20 μl/ml protease inhibitor cocktail (Sigma). After centrifugation for 10 min at 10.000 xg, the supernatant was collected as the soluble fraction. The pellet was washed with fresh buffer and centrifuged again. The obtained pellet fraction was then resuspended in denaturing TKMES buffer [19] and centrifuged again to collect the supernatant as the insoluble fraction. In all cases, protein concentration was determined using the Bio-Rad protein assay. For immunoblot assays, antibodies raised against DXS and DXR [19], GFP (Life Technologies), MYC (Millipore), and chloroplast Hsp70, ClpC, and ClpB proteins (Agrisera) were diluted 1:500 for DXS, 1:7,000 for DXR, 1:1,000 for GFP and MYC, 1:6,000 for Hsp70, 1:2,000 for ClpC, and 1:3,000 for ClpB. The total amount of protein loaded per lane was calculated for each particular antibody to remain in the linear range (S9 Fig). Chemiluminescent signals were visualized using a LAS-4000 (Fujifilm) image analyzer and quantified with Quantity One (Bio-Rad). Student´s t test was used to assess statistical significance of quantified differences.
For protease accessibility assays, protein extracts from 10-day-old WT and j20 seedlings containing 30 μg of total protein were incubated for 5 min at 37°C with increasing concentrations of Proteinase K (Invitrogen). After stopping the reaction with SDS-PAGE loading buffer, extracts were used for immunoblot analysis.
DXS enzyme activity measurements were carried out as described [19,72]. Specific activity was calculated by dividing the total activity measured in extracts with the amount of DXS protein found in the corresponding sample.
RNA isolation, cDNA synthesis, and qPCR experiments were performed as described [19] using the APT1 (At1g27450) gene for normalization.
The Aggrescan3D algorithm [73] was used to analyze protein aggregation propensity. Predictions were performed in static mode using a distance of aggregation analysis of 10 Å. The Arabidopsis DXS structure was modelled using Swiss-Model [74] on top of the 2.40 Å resolution E. coli DXS structure with PDB code 2O1S. Residues 72 to 707 of the Arabidopsis DXS monomer, sharing a sequence identity of 41.08% with the E. coli protein, were structurally aligned and modelled. The interface of the generated homodimer was evaluated with PDBePISA (http://www.ebi.ac.uk/pdbe/pisa/) rendering an area of 8892 Å and a predicted dissociation ΔG for the dimer of 51.2 kcal/mol (close to those of the template E. coli crystal structure, which exhibits an interface of 7970 Å and a dissociation ΔG of 59.1 kcal/mol).
|
10.1371/journal.pgen.1000515 | A Genome-Wide Screen for Regulators of TORC1 in Response to Amino Acid Starvation Reveals a Conserved Npr2/3 Complex | TORC1 is a central regulator of cell growth in response to amino acid availability, yet little is known about how it is regulated. Here, we performed a reverse genetic screen in yeast for genes necessary to inactivate TORC1. The screen consisted of monitoring the expression of a TORC1 sensitive GFP-based transcriptional reporter in all yeast deletion strains using flow cytometry. We find that in response to amino acid starvation, but not to carbon starvation or rapamycin treatment, cells lacking NPR2 and NPR3 fail to fully (1) activate transcription factors Gln3/Gat1, (2) dephosphorylate TORC1 effector Npr1, and (3) repress ribosomal protein gene expression. Both mutants show proliferation defects only in media containing a low quality nitrogen source, such as proline or ammonia, whereas no defects are evident when cells are grown in the presence of glutamine or peptone mixture. Proliferation defects in npr2Δ and npr3Δ cells can be completely rescued by artificially inhibiting TORC1 by rapamycin, demonstrating that overactive TORC1 in both strains prevents their ability to adapt to an environment containing a low quality nitrogen source. A biochemical purification of each demonstrates that Npr2 and Npr3 form a heterodimer, and this interaction is evolutionarily conserved since the human homologs of NPR2 and NPR3 (NPRL2 and NPRL3, respectively) also co-immunoprecipitate. We conclude that, in yeast, the Npr2/3 complex mediates an amino acid starvation signal to TORC1.
| Before a eukaryotic cell commits to cell division it must be large enough so that both daughter cells would be of viable size. The control of cell size is largely mediated by nutritional input signals via an evolutionarily conserved protein complex termed TORC1. In particular, TORC1 has been shown to sense the level of amino acids and its activity correlates with the level of amino acids present in the media. Yet, it is largely unknown how TORC1 senses amino acids. Here we demonstrate that the evolutionarily conserved Npr2/3 complex mediates the amino acid scarcity signal to TORC1. Cells lacking NPR2 and NPR3 genes fail to inactivate TORC1 when amino acids are scarce. Overactive TORC1 prevents these cells from adapting to an amino acid scarce environment, and, as a result, these cells are unable to proliferate in media that is not rich in amino acids. Artificially inhibiting TORC1 with rapamycin can completely rescue these defects. These results provide insight into how cells sense amino acid deficiency. Moreover, as deletions of NPR2 have been implicated in tumor growth, these results offer a fertile ground to study the role overactive TORC1 might play in those cancers.
| The Target of Rapamycin (TOR) kinase has emerged as a central processing unit that can incorporate signals from diverse pathways regarding the availability of glucose, oxygen and amino acids [1]. The TOR kinase is a large scaffolding protein-kinase that is highly conserved in all eukaryotes. In recent years, TOR has been shown to nucleate two different multiprotein complexes. In the TOR Complex 1 (TORC1), TOR is associated with raptor and mLST8 (yeast Kog1 and Lst8, respectively), and its known mammalian phosphorylation targets are p70 S6K1 and 4E-BP1 [2]. In the phosphorylated state, both of these substrates are involved in promoting protein translation. In the TORC2, TOR partners with rictor (yeast Avo3), instead of raptor, and mLST8 [3],[4]. Mammalian TORC2 phosphorylates AKT protein at Ser473, but not TORC1 substrates [5]. Interestingly, the natural compound rapamycin only inhibits TORC1 when complexed with intracellular protein FKBP12. The inhibition of TORC1 by the treatment of cells with rapamycin leads to a nutrient starvation-like phenotype, in cells ranging from yeast to human [6],[7].
Befitting its ability to process numerous nutritional signals, TORC1 controls a diverse set of effector pathways. In a nutrient rich environment, TORC1 in yeast is responsible for promoting translation initiation, ribosome biogenesis, and anabolic pathways [8]–[10]. Upon nutrient limitation or rapamycin treatment, these processes are inhibited, and the cell initiates macroautophagy, activates enzymes necessary for catabolic processes, and ultimately enters the G0 phase of the cell cycle [1],[11]. In yeast, an active TORC1 controls these processes by preventing several transcription factors from entering the nucleus and activating their target genes. In rich media, transcription factors Gln3/Gat1 are prevented from activating enzymes necessary for nitrogen catabolism, Msn2/4 from inducing stress response, and Rtg1/3 from activating retrograde signaling [12]–[14]. TORC1 controls ribosomal protein (RP) gene expression by excluding corepressor Crf1 from entering the nucleus and inactivating RP gene transcription [15].
While there has been some progress in understanding how glucose, oxygen and energy status are signaled to TORC1 [1], the molecular mechanism by which an amino acid starvation leads to TORC1 inactivation is unclear and controversial [16]. An amino acid starvation, just like rapamycin treatment, leads to a rapid dephosphorylation of the mTORC1 effectors S6K1 and 4E-BP1 [2]. This inactivation was proposed to be dependent on the presence of tuberous sclerosis proteins TSC1 and TSC2, which form a heterodimer and connect the insulin/IGF-IRS-PI3K-Akt pathway to the mTORC1 kinase [17]. However, other labs have found TSC2 to be unnecessary for this inactivation [18],[19]. Further, organisms that appear to have an amino acid sensitive TORC1, such as S. cerevisiae and C. elegans, do not have homologs of TSC1 and TSC2. hVPS34, encoding a class III PI3-kinase, has been shown to be required for amino acid induced phosphorylation of S6K1 [19], and its kinase activity correlates with amino acid levels [20]. However, at this point it is unclear how hVPS34 regulates TORC1 [21]. If Vps34 was necessary to maintain active TORC1, then one would expect cells lacking the VPS34 gene to exhibit severe growth defect and mimic rapamycin treatment. Yet, yeast vps34Δ cells grow normally in rich media [22], despite the fact that Vps34 is the only PI3K found in yeast. Also, a recent report was unable to detect defective TORC1 signaling in Vps34 mutant D. melanogaster [23].
With the hypothesis that the genes necessary to inactivate TORC1 in response to amino acid starvation are distinct from the Vps34 pathway and that these genes are not essential for growth under a rich nutrient environment, we sought to screen the yeast deletion collection for genes necessary to inactivate TORC1 in response to amino acid starvation. We devised a GFP-based transcriptional reporter assay where TORC1 activity can be monitored by measuring intracellular GFP content by high throughput flow cytometry. The screen yielded two largely uncharacterized genes, NPR2 and NPR3, and here we demonstrate that they form a heterodimer that is responsible for inactivating TORC1 specifically in response to amino acid starvation.
Gene expression profiling has revealed that either rapamycin treatment or amino acid starvation in both yeast and human cells lead to a robust activation of genes involved in nutrient catabolism [7],[10]. In yeast, an active TORC1 complex binds and maintains an association between a cytosolic Ure2 and two transcription factors, Gln3 and Gat1 [24]. Upon a shift to a nitrogen free media, or rapamycin treatment, Ure2 releases both Gln3 and Gat1, which then move into the nucleus to activate the transcription of largely common target genes [25]. One of the genes activated this way is DAL80 (Figure 1A). For our genetic screen, we employed the expression of DAL80 as a readout for the activity status of TORC1. An active TORC1 suppresses the transcription of DAL80 mRNA, whereas inactivation of TORC1 leads to a transcriptional upregulation of DAL80.
To measure the Dal80 reporter in thousands of different deletion strains, we constructed a Dal80-based GFP reporter plasmid Dal80pr-GFP (Figure 1B). We could monitor the expression of this reporter by flow cytometry in a high-throughput fashion. In cells harboring this plasmid, GFP levels can be induced by either rapamycin treatment or amino acid starvation (Figure 1C). As DAL80 mRNA level stays constant following the start of induction (data not shown) and GFP has a long half-life, GFP accumulates linearly over time in the cell. As a result, this method enabled us to identify mutants that fail to express the reporter, as well as mutants that partially express or over-express the reporter.
We assembled all non-essential yeast deletion strains into fifty-five 96-well plates, such that each well contained a known deletion strain. Each well was transformed with Dal80pr-GFP and the hygromycin B-resistant transformants were outgrown from the non-transformed cells in liquid media containing hygromycin B. Even though we set out to perform our screen for mutants that fail to induce the Dal80pr-GFP reporter in response to amino acid starvation, we also screened the genome for mutants that (1) activate Dal80pr-GFP in the rich media even in the absence of rapamycin, and (2) have altered induction of Dal80pr-GFP in response to rapamycin treatment. In summary, we screened the Dal80pr-GFP reporter expression in ∼5100 strains under three different conditions (Figure 1D).
In this screen, we sought to find mutants that have active Gln3 and Gat1 in the rich YPD media. It is well known that a lack of Ure2, a cytosolic inhibitor of Gln3 and Gat1, leads to the activation of both transcription factors even in amino acid rich medium [12]. When we screened the genome in YPD medium, only ure2Δ cells show elevated levels of the reporter expression (Figure 2A). ace2Δ cells are large and clumpy, and as a result they exhibit a higher background signal [26]. These data show that among the non-essential deletion strains, only the ure2Δ mutant leads to Gln3 and Gat1 activation in the rich media.
To successfully identify the mutants that under-express and over-express the Dal80pr-GFP reporter in a single screen, we selected to perform the flow cytometry analysis after a 15-hour induction in a rapamycin containing rich YPD media. We chose the 15-hour timepoint because by then the intracellular GFP levels were about 10 fold above background, enabling us to identify mutants that either under-express or over-express the Dal80pr-GFP reporter. As can be seen from Figure 2B, deletion strains that over-express and under-express the reporter can be identified. We re-transformed 43 mutants that over-expressed our reporter two-fold or higher (mutants above the red line in Figure 2B), and performed an induction time course for each mutant. Of these 43 mutants, 36 exhibited GFP levels higher than a WT strain (Figure S1). Interestingly, in the re-analysis, the top three strains that over-express GFP are rtg1Δ, rtg2Δ and rtg3Δ cells. The Rtg1–3 proteins are part of the retrograde gene expression mechanism that leads to a transcriptional activation of the first three TCA cycle enzymes in respiratory-deficient cells [27]. These findings demonstrate that the presence of Rtg1–3 proteins suppresses the activation of Gat1 and Gln3 transcription factors in response to rapamycin treatment and suggest that other mutants identified in this screen might be involved in retrograde gene expression (such as prm10Δ cells with the 4th highest GFP expression).
We also reanalyzed the top 40 strains that under-expressed Dal80pr-GFP (strains below the green line in Figure 2B). The top three deletion strains that fail to express the reporter were as expected: fpr1Δ, gln3Δ and gat1Δ strains. Fpr1 encodes the yeast version of FKBP12 that binds rapamycin before inhibiting the TORC1 complex, and fpr1Δ cells are completely rapamycin resistant. These three strains had no detectable induction of GFP, while the rest of the strains that lie below the green line had above-background levels of GFP. Upon inspection of the list of strains that only partially express Dal80pr-GFP, we noticed that many of these deletion strains are missing genes necessary for efficient transcription or translation. As our screen relies on efficient transcription and translation of GFP, we devised a filter to identify deletion strains specific to the TORC1 pathway. We constructed a reporter plasmid similar to our Dal80pr-GFP plasmid, but used the Gal1 promoter in lieu of the Dal80 promoter to drive the expression of GFP (Gal1pr-GFP). In cells harboring this plasmid, GFP can be induced by shifting the cells from glucose-containing media to galactose-containing media. By separately transforming the Dal80pr-GFP and Gal1pr-GFP reporter plasmids into strains that showed partial activation of the Dal80pr-GFP reporter, and inducing the GFP with a corresponding inducer, we were able to separate the TORC1-specific mutants from the mutants that are unable to express GFP efficiently (Figure 2D). For instance, fpr1Δ strain fails to induce Dal80pr-GFP in response to rapamycin treatment, yet it can efficiently upregulate GFP expression from the Gal1pr-GFP reporter under galactose induction. From the top forty strains with defective inductions of Dal80pr-GFP, twenty are TORC1-specific (Figure 2D and Figure S2). The molecular basis for failure to activate Dal80pr-GFP can be attributed to most of these strains. The fpr1Δ, gln3Δ and gat1Δ deletion strains fail to induce the reporter for reasons discussed above. Of the remaining 17 mutants, six (sod1Δ, ymr073cΔ, yml122cΔ, rrd1Δ, ppm1Δ and spt8Δ strains) are partially rapamycin-resistant. We have recently shown that rapamycin resistance in these strains is due to the inability of the Fpr1:rapamycin complex to bind to TORC1 [28]. As a result, these strains fail to fully respond to rapamycin treatment. Another seven (ede1Δ, ppz1Δ, apn1Δ, yor322cΔ, snf7Δ, bck1Δ and slt2Δ strains) are rapamycin hyper-sensitive, and presumably rapamycin treatment leads to cell death in these cells. For instance, bck1Δ and slt2Δ cells fail to induce Gal1pr-GFP in response to galactose induction when rapamycin is present (data not shown).
These findings illustrate that our screen was able to capture all major players in the pathway. Further, demonstration that we were able to identify strains that over-express and under-express our reporter suggests that we could identify even minor modifiers of TORC1 in response to amino acid starvation.
To identify strains that fail to inactivate TORC1 in response to amino acid starvation, we exposed the cells harboring the Dal80pr-GFP reporter to media lacking all sources of nitrogen for four hours. Nitrogen-free media contains 2% glucose (as in rich YPD medium) and all salts, vitamins and trace elements that yeast cells need for cell proliferation. No nitrogen source was included as yeast can synthesize all 20 AAs de novo. This screen identified 17 strains that have a GFP content less than 40% of the WT. By imposing the Gal1pr-GFP filter again, we eliminated 13 strains due to their inability to activate Gal1pr-GFP in response to galactose induction (Figure 2D). As expected, the gln3Δ and gat1Δ strains specifically failed to induce the Dal80pr-GFP reporter. Lastly, npr2Δ and yhl023CΔ strains fail to induce the Dal80pr-GFP reporter, suggesting that proteins encoded by NPR2 and YHL023C are involved in signaling the absence of amino acids to TORC1. For reasons that will become clear in next sections, we renamed YHL023C locus NPR3. NPR stands for Nitrogen Permease Regulator, as Npr1 and Npr2 have been shown to regulate amino acid transport activity [29],[30]. Both NPR2 and NPR3 are largely uncharacterized genes. npr2Δ cells were isolated as unable to grow in the presence of low concentrations of amino acids [29], and NPR3 was shown to be necessary for sporulation [31]. Both genes are conserved among all eukaryotes, yet nothing is known about their function. A bioinformatic analysis is also unable to predict an activity/function for either protein. We characterized these two proteins further.
To verify that Npr2 and Npr3 are involved in activation of Gln3 and Gat1, we monitored the translocation of Gat1 into the nucleus when npr2Δ and npr3Δ cells are shifted from the rich YPD medium to the nitrogen-free medium. Consistent with data from the screen, genomically GFP-tagged Gat1 fails to translocate into the nucleus upon nitrogen starvation in npr2Δ and npr3Δ cells (Figure 3A). In fact, while about 90% of WT cells exhibited nuclear Gat1 only about 20% of npr2Δ and npr3Δ cells had any nuclear Gat1 (Figure 3B). Further, the npr2Δ npr3Δ double mutant behaved exactly as either single mutant, suggesting that Npr2 and Npr3 lie in the same pathway. These data confirm that Npr2 and Npr3 are involved in translocation of TORC1-responsive transcription factor Gat1, and that their absence is not suppressing the transcription or translation of our Dal80pr-GFP reporter.
If Npr2 and Npr3 are indeed upstream regulators of TORC1, then TORC1 in npr2Δ and npr3Δ cells should be sensitive to rapamycin and insensitive to nitrogen starvation. Consistent with the results from the screen, Dal80pr-GFP cannot be activated in npr2Δ and npr3Δ cells in response to a nitrogen starvation, yet the reporter can be readily activated by rapamycin in both mutants (Figure 3C). These data posit Npr2 and Npr3 upstream of TORC1.
To confirm that Npr2 and Npr3 mediate the amino acid starvation signal to TORC1, we examined the effect of nitrogen starvation in npr2Δ and npr3Δ cells on different TORC1 readouts. TORC1 has been shown to control the phosphorylation status of Npr1, a protein kinase that controls the post-Golgi sorting of amino acid permeases [32]. In rich media Npr1 is hyperphosphorylated, whereas nitrogen starvation or rapamycin treatment leads to a faster migrating dephosphorylated form [30],[33]. Consistent with our hypothesis, cells missing NPR2 or NPR3 are unable to dephosphorylate Npr1 upon nitrogen starvation (Figure 3D). Yet, inclusion of rapamycin in rich YPD media leads to a dephosphorylation of Npr1 in both mutants. To more clearly define the signal that both mutants fail to heed, we determined the Npr1 phosphorylation state in cells grown with either 10 mM proline or 10 mM glutamine as the sole nitrogen source. The wildtype cells grown in proline media exhibited dephosphorylated Npr1, whereas wildtype cells grown in glutamine media showed a hyperphosphorylated Npr1 that is indistinguishable from npr2Δ and npr3Δ cells. To our surprise, wildtype cells grown in the presence of high concentrations of ammonia never exhibited the same level of hyperphosphorylated Npr1 as npr2Δ and npr3Δ cells. Even at 50 mM ammonium sulfate, wildtype cells exhibited a faster migrating band of Npr1 than either mutant. This suggests that ammonia itself is not sensed by Npr2 and Npr3, but rather the signal about the levels of amino acids, such as glutamine, is mediated to TORC1. That is, npr2Δ and npr3Δ cells, irrespective of the nitrogen source, exhibit an Npr1 phosphorylation state similar to the phosphorylation state of wildtype cells grown in glutamine containing media. Therefore, the absence of NPR2 or NPR3 appears to mimic the presence of glutamine, even if a low quality nitrogen source is provided instead.
To further confirm the nature of the signal that npr2Δ and npr3Δ mediate to TORC1, we examined the effect of total carbon starvation in npr2Δ and npr3Δ cells on Npr1 phosphorylation. The carbon starvation leads to a rapid dephosphorylation of Npr1 in both mutants, suggesting that amino acid starvation and carbon starvation are distinct signaling inputs to TORC1. These data confirm that Npr2 and Npr3 mediate the signal about amino acid starvation to TORC1.
As TORC1 exerts an exquisite control over the transcription of ribosomal protein (RP) genes, we were interested whether amino acid starvation is able to downregulate RP gene expression in the two mutants. Ribosome biosynthesis is very energy expensive, and RP genes are responsible for ∼50% of the total Pol II transcription initiation events [34]. Hence, an overproduction of RP mRNA can be very costly for starving cells as it depletes valuable energy resources. To study the effect of nitrogen starvation on gene transcription, we performed a whole genome microarray experiment where we shifted WT, npr2Δ or npr3Δ cells from YPD media to N-free media for 30 minutes. For each gene, an induction ratio was calculated by dividing the expression of that gene after the shift in media by its expression in YPD media. Genes that are downregulated upon the shift in media have an induction ratio below 1, and vice versa. As can be seen in Figure 4A, mRNA species encoding large and small ribosomal protein subunits (RPL and RPS, respectively, highlighted in red) are downregulated less in npr2Δ cells than in WT cells. (Data from npr3Δ data is identical to npr2Δ data and is therefore not shown). In fact, this effect appears to scale linearly, such that on average, the RP mRNA levels are ∼2.5 times higher in npr2Δ and npr3Δ cells than in WT cells in N-free media (Figure 4B). No difference in expression was observed in rich YPD media. Independent semi-quantitative RT-PCR analysis on representative RPL and RPS genes confirmed the microarray results (Figure 4C). Based on the Npr1 phosphorylation data, we speculated that RP genes would exhibit a higher expression in npr2Δ and npr3Δ cells when grown in proline or ammonium sulfate as a nitrogen source, whereas no difference would be observed when cells are grown in the presence of glutamine. Indeed, both mutants exhibited a 2–3 fold higher expression of RP genes when grown in the presence of proline as a nitrogen source (Figure 4D). Congruent with the Npr1 phosphorylation data, no difference was observed between the two mutants and WT cells when they were grown in media containing glutamine or 2% peptone. npr2Δ and npr3Δ cells grown in 0 mM, 1 mM or 50 mM ammonium sulfate consistently exhibited a 2–3 fold higher level of RP gene expression than WT cells (Figure 4E). Again, it appears that npr2Δ and npr3Δ cells grown in a proline or ammonia containing media are incapable of inactivating TORC1. As cells grown on ammonia are known to activate retrograde signaling pathway by activating TORC1 responsive transcription factors Rtg1/3, we sought to determine whether npr2Δ and npr3Δ cells are capable of inducing the expression of the Rtg1/3 target gene CIT2 [35]. As expected, the activation of CIT2 mRNA was greatly diminished in the two mutant strains.
In conclusion, we examined four independent TORC1 effector pathways to confirm the effect of Npr2 and Npr3 on TORC1. Gat1 nuclear translocation, Npr1 phosphorylation, RP gene expression profiling and Rtg1/3 reporter expression analysis confirm that Npr2 and Npr3 are necessary to inactivate TORC1 by amino acid starvation. In particular, npr2Δ and npr3Δ cells fail to inactivate TORC1 when a non-preferred amino acid source is presented. As TORC1 is a key player in a cell's response to nutrient availability, we hypothesized that cells lacking NPR2 and NPR3 would fail to adapt to an environment containing non-glutamine based nitrogen source.
To address the in vivo function of Npr2 and Npr3, we first sought to determine the ability of cells lacking NPR2 or NPR3 to proliferate in an environment containing low amounts of preferred nitrogen. In our previous experiments, we noticed that npr2Δ and npr3Δ cells do not appear to have any growth defects in rich YPD media. Quantitatively measuring the doubling time during an exponential phase of proliferation confirmed this observation (Figure 5A). Next, we grew prototrophic npr2Δ and npr3Δ cells in N-free media supplemented with 2% peptone, 10 mM proline, 10 mM glutamine or increasing concentrations of ammonia. Consistent with our molecular TORC1 activity data, npr2Δ and npr3Δ cells grown in the presence of peptone and glutamine do not exhibit a proliferation defect. However, when the same cells are provided with proline or ammonia as a nitrogen source, both mutants proliferate slowly. These data suggest that the inability to inactivate TORC1 in both mutants leads to a failure to adapt to an environment with a scarce source of preferred amino acids. In all cases, the double mutant behaves exactly as the single mutants, further confirming that Npr2 and Npr3 lie in the same pathway upstream of TORC1. Consistent with the finding that Npr2 and Npr3 are responsible for inactivating TORC1 only in response to amino acid starvation, all three mutants behave exactly as WT cells when limited for glucose (Figure S3) or grown on a non-fermentable carbon source (data not shown).
We next sought to establish a role for Npr2 and Npr3 in the activation of amino acid transporters. TORC1 inactivation leads to a robust transcriptional activation of various amino acid permeases [10]. In our unpublished studies, we have found that threonine uptake is particularly sensitive to the presence of a preferred nitrogen source and we sought to examine the ability of npr2Δ and npr3Δ cells to import radiolabeled Thr. We grew cells in either proline or glutamine and measured their Thr uptake. The cells grown in glutamine exhibit a barely detectable uptake of Thr, with no difference between WT and the two mutants (Figure 5B). The cells grown in proline exhibit a robust increase in Thr uptake activity, but this activity is about 50% of WT levels in npr2Δ and npr3Δ cells. Therefore, npr2Δ and npr3Δ cells fail to fully activate threonine amino acid transporters under amino acid limitation.
A long-term characteristic of nitrogen-starved yeast is their ability to undergo sporulation, and this phenotype has been linked to TORC1 [36],[37]. In fact, a large-scale study has implicated Npr3 in sporulation [31]. Our study confirmed this report and here we show that npr2Δ cells also undergo sporulation much less efficiently than WT cells (Figure 5C).
Lastly, as the phenotypes associated with a lack of Npr2 and Npr3 are due to the failure to inactivate TORC1, we wondered whether artificially inhibiting TORC1 in npr2Δ and npr3Δ cells might rescue the proliferation defects observed in these cells under nitrogen limitation. When npr2Δ and npr3Δ cells are shifted from rich YPD media to nitrogen limitation media containing 0.5 mM ammonium sulfate, three unique phenotypes can be observed. First, in both mutants about 40% of cells undergo lysis within the first 10 hours (Figure 5D). Second, npr2Δ and npr3Δ cells resume logarithmic proliferation after about 12 hours, whereas WT cells do this in about 2 hours. Third, the maximum cell density of npr2Δ and npr3Δ cultures is about 50% of WT levels. Remarkably, all three phenotypes can be completely rescued by including rapamycin in the nitrogen limitation media. In other words, the phenotypes associated with a lack of NPR2 and NPR3 genes can be completely abolished by artificially inhibiting TORC1.
These data clearly demonstrate that Npr2 and Npr3 are responsible for mediating the signal about amino acid starvation to TORC1 and that TORC1 remains active in npr2Δ and npr3Δ cells when preferred amino acid source (e.g. glutamine) is absent. As a result, these cells fail to adapt and thrive in an environment where nitrogen source is scarce or non-preferred. Our data also suggest that both proteins are specific for amino acid starvation, with no relevance in mediating the signal about carbon limitation. Lastly, the genetic data demonstrate that both proteins act in the same pathway upstream of TORC1.
To better understand the role of Npr2 and Npr3 in cells, we purified both proteins by tandem affinity purification (TAP) expressed under the endogenous promoters [38]. Both proteins are relatively scarce in cells, especially Npr2. When we ran the two purifications in adjacent lanes we noticed a single unique band where Npr2 runs in the Npr3-TAP lane, and a single unique band where Npr3 runs in the Npr2-TAP lane (Figure S4). This suggested to us that Npr2 and Npr3 might be forming a heterodimer and we sought to verify this interaction by co-immunoprecipitation. By employing a strain with a genomically incorporated HA-tag at the C-terminus of Npr3 and a myc-tag at the C-terminus of Npr2, we find that when Npr3 is immunoprecipitated with anti-HA antibodies these immunoprecipitates contain Npr2-myc (Figure 6A). This interaction is not medium-dependent, as it occurs in the medium made of either YPD or N-free media. There is a decrease in the amount of Npr2 pulled down, but this appears to be due to degradation of Npr3. The finding that Npr2 and Npr3 form a complex is consistent with the observation that npr2Δ, npr3Δ and npr2Δ npr3Δ cells have identical phenotypes with respect to amino acid starvation.
A homology search of Npr2 and Npr3 shows that both are conserved eukaryotic-specific proteins (Figure 6B and Figure S5), and their homologs are uncharacterized. The yeast Npr2 shares 30% identity with human NPRL2 (Npr2 Like), and the yeast Npr3 shares 19% identity with human C16orf35 (Accession number Q12980; termed NPRL3 here). If the human homologs are orthologous to yeast Npr2 and Npr3, then NPRL2 and NPRL3 should also form a complex in human cells. To verify this, we co-transfected HA-NPRL2 with three N-terminally myc-tagged proteins: XMis12 (an unrelated Xenopus protein), γ-tubulin (an unrelated human protein) and NPRL3. Upon immunoprecipitation with anti-myc antibodies, only NPRL3 immunoprecipitates contained HA-NPRL2, demonstrating that NPRL2 and NPRL3 also form a complex. The conserved nature of this complex suggests that NPRL2/3 might also signal to TORC1, although further studies are needed to verify this hypothesis.
In this study we aimed to systematically screen the yeast genome for genes necessary in sensing amino acid starvation. To achieve this goal, we first established a robust flow cytometry-based assay to monitor the expression of a TOR-sensitive transcriptional reporter. We believe a GFP-based readout has three advantages over a luciferase-based readout: (1) by employing GFP, there is no need to lyse cells; (2) a fluorescence reading can be obtained for each individual cell, thereby abolishing the need to normalize the signal for the number of cells; and (3) this assay offers a possibility of more complicated analyses with sub-populations of cells, such as by co-induction of two different fluorescent proteins. Compared to assays where the plate could be scanned, however, performing flow cytometry in a high throughput manner is still time-consuming. It takes about 45 minutes to process each 96 well plate, with the majority of time spent flushing the fluidics system between samples.
Using this assay, we found an evolutionarily conserved Npr2/3 complex that is responsible for the inactivation of TORC1 in response to amino acid limitation. In response to amino acid withdrawal, cells missing NPR2 and NPR3 fail to completely inactivate such TORC1 readouts as Npr1 phosphorylation, Gat1 inhibition and ribosome biosynthesis. Correspondingly, these cells are unable to adjust to a growth environment where the nitrogen source is limiting or of poor quality. Our data appear to indicate that the Npr2/3 complex specifically mediates a signal about the low quality/quantity of amino acids since phenotypes associated with NPR2/3 deletion are evident only when cells are grown on nitrogen sources that are sub-optimal for TORC1 activity. For instance, when cells are grown with peptone or glutamine as nitrogen sources, TORC1 is fully active and under those conditions npr2Δ and npr3Δ cells do not exhibit proliferation defects. Under conditions where a lower activity of TORC1 is required, such as in the presence of proline or ammonia, Npr2/3 are responsible for lowering this activity. In the simplest model, the Npr2/3 complex senses the levels of glutamine and inactivates TORC1 when those levels are low. Further studies must determine whether the Npr2/3 complex is itself responsible for sensing intracellular amino acid pools or whether this complex simply mediates the signal. Interestingly, our results indicate that neither protein plays a role in carbon starvation, since in both mutants TORC1 is inhibited when glucose is removed from the media and both mutants proliferate normally under glucose-limiting conditions (Figure 6D).
Despite the high degree of conservation, our best efforts to assign an activity or function to either Npr2 or Npr3 were unsuccessful, as neither protein exhibits characteristic domain structures that could predict their cellular function. Our efforts to co-immunoprecipitate TORC1 and Npr2/3 were inconclusive (data not shown). Also, the Tor1/Kog1 complex was unaltered in npr2Δ and npr3Δ cells (data not shown). Further complicating analysis of both proteins is their relatively low abundance in the cell. In fact, in our hands human NPRL2 is detectable in only a few cell lines (data not shown). This has hindered our attempts to purify additional members of the complex, as well as visualize their sub-cellular location. Clearly, further biochemical and genetic approaches are needed to shed light on the mechanistic aspects of their function.
Recently, two groups characterized a role Rag GTPases play in activating TORC1 by amino acids in human tissue culture cells and D. melanogaster [39],[40]. These studies demonstrated that Rag GTPases are part of the extended TORC1 complex and that cells expressing constitutively active RagA/B fail to inactivate TORC1 when amino acids are withdrawn. Interestingly, Rag A and RagB are similar to each other and they appear to be orthologus to yeast Gtr1. If yeast Gtr1 was responsible for activating TORC1, then deletion of GTR1 should lead to lethality or at least mimic a rapamycin-like slow proliferation phenotype. Yet, gtr1Δ cells appear normal and don't express the TORC1-sensitive Dal80pr-GFP reporter in rich media (data not shown). This discrepancy could be explained by either a substitution of Gtr1 by another GTPase in yeast, or by a divergence of the amino acid signaling pathway to TORC1 between metazoans and yeast. By a similar token, yeast Vps34 did not emerge from our screen as a regulator of TORC1 (see Introduction). Further studies must examine the role of Rag GTPases in lower eukaryotes and whether the Npr2/3 complex is functionally conserved in higher eukaryotes. Keeping with the conserved nature of TORC1 and its regulation by amino acids, it would be exciting if a unifying mechanism emerges.
Except for mTOR itself, many of the upstream and downstream components of the mTOR pathway are known to be either proto-oncogenes or tumor suppressors [1]. Interestingly, early reports have implicated NPRL2 as a tumor suppressor gene. NPRL2 is located on the human chromosome 3p21.3 homozygous deletion region [41], a region that is deleted in various human cancers [42]. Further, reintroduction of NPRL2 into these cells inhibits cell proliferation both in cell lines and in human lung cancer mouse models [43],[44]. Human NPRL2-null cells and yeast npr2Δ cells both have been shown to exhibit resistance to cisplatin [44], suggesting functional conservation and raising the possibility that NPRL2 might act as a repressor of mTORC1. It remains to be determined whether a mutation in NPRL2 is truly the driver mutation that gives rise to neoplastic growth via overactive mTORC1, but a high mTORC1 activity has certainly been ascribed to invasive tumors [45]. Further, as cancer cells with hyperactive mTORC1, either due to loss of PTEN or TSC1/2, have been shown to be particularly sensitive to rapamycin treatment [46],[47], we speculate that cancer cells with mutated NPRL2 might also be particularly sensitive to rapamycin. Also, since both Npr2 and Npr3 are necessary for inactivation of TORC1 in yeast, it would be interesting to examine whether a deletion of human NPRL3 also correlates with tumorigenesis.
In conclusion, here we report the discovery of a novel, conserved Npr2/3 complex that specifically inactivates TORC1 in response to amino acid limitation in yeast. Inactivation of TORC1 is important for adaptation to an environment with scarce amino acids, since diminished TORC1 activity leads to activation of amino acid permeases and catabolic enzymes, repression of ribosomal protein gene expression, and induction of macroautophagy. Cells lacking NPR2 or NPR3 are unable to inactivate TORC1, and they fail to thrive in nitrogen challenged environment as a result. These results demonstrate a unique way by which the cells respond to nutrient limitation, and this complex provides a fertile ground to study the regulation of this important pathway.
The genetic screen was conducted with diploid yeast deletion collection [48]. A strain with genomically integrated GFP at the C-terminus of Gat1 has been described [49]. For Gat1-GFP studies, this strain served as the wildtype. From that strain, NPR2 was deleted using the kanMX cassette and NPR3 was deleted using the hygromycin B-phosphotransferase cassette from pAG26 plasmid. The double mutant strain was created by deleting NPR3 from Gat1-GFP npr2Δ strain. Npr1, Npr2 and Npr3 were epitope tagged at the C-terminus with PCR products amplified from pFA6a-3HA-kanMX6, pFA6a-13myc-His3MX6 and pFA6a-3HA-kanMX6, respectively [50]. YPD media consists of 1% yeast extract, 2% peptone and 2% glucose. N-free media consists of 0.17% Yeast Nitrogen Base without amino acids and ammonium sulfate (Sigma) and 2% dextrose. Carbon-free media consists of 1% yeast extract and 2% peptone. All yeast experiments were carried out at 30°C. The GFP expression-reporter vectors, with added hygromycin B-phosphotransferase and GFP genes, are based on pRS426, a multicopy 2µ-based plasmid. Six hundred basepairs of Dal80 or Gal1 promoter was fused in-frame with GFP to obtain the final Dal80pr-GFP or Gal1pr-GFP, respectively. NPRL2 and γ-tubulin were amplified from ProQuest cDNA library (Invitrogen). Full length NPRL2 (380 AA) was cloned into pCS-HA vector. NPRL3 was amplified from Mammalian Gene Collection clone MGC2816. This isoform is missing the first 179 amino acids from the amino terminus and is therefore 390 amino acids long. Both γ-tubulin and NPRL3 were cloned into pCS-myc vector. Xmis12-myc was a kind gift from Aaron Straight (Stanford).
All non-essential yeast deletion strains were distributed over fifty-five 96-well plates such that each strain was in a designated well. The strains were grown in 150 µL of YPD to saturation at 30°C on orbital shakers. 450 µL of fresh YPD media was added and the cells were allowed to enter a logarithmic growth phase. After 3 hours, the cells were pelleted by centrifugation of the plate at 750 g and the media was poured out without disturbing the cell pellet. To each well, 250 µL of the transformation mixture (36% polyethylene glycol, 0.1 M lithium acetate, 100 µg of salmon sperm DNA, and 5 µg of plasmid DNA) was added. The cell pellet and transformation mixture were mixed by shaking. The plates were incubated at 30°C for one hour, followed by a heat shock at 42°C for 45 minutes. The cells were then pelleted by centrifuging the plate at 750 g, the transformation mixture was poured out, a fresh 300 µL of YPD media was added, and the cells were allowed to express the hygromycin B-phosphotransferase protein by shaking the plates at 30°C. After 1.5 hours of growth, 300 µL of YPD with 2× hygromycin B was added. The final concentration of hygromycin B in 600 µL of YPD was 300 µg/mL. The cells were grown with shaking for 2 days, after which 5 µL of cells were transferred to a new 600 µL of YPD with hygromycin B. Cultures were grown for 24 hours and stored at 4°C until use. About 98% of wells showed growth after reinoculation. Usually, about 10 plates were transformed in parallel.
15 µL of transformed cells were allowed to recover for 6 hours in 450 µL of YPD at 30°C with shaking. The plasmid selection was maintained by including hygromycin B in the medium. After 6 hours, 10 µL of cells were transferred to 190 µL of PBS with cycloheximide (to prevent non-specific induction of GFP). These plates were used for “YPD Screen”. Another 20 µL of recovered culture was transferred to an YPD culture containing 20 ng/mL of rapamycin and hygromycin B. These plates were grown for 15 hours with shaking at 30°C. After 15 hours, 10 µL of the rapamycin culture was transferred to 190 µL of PBS with cycloheximide. These plates were used for “RAPA Screen”. The remaining 430 µL of recovered cells were pelleted, washed with PBS once, and inoculated in 450 µL of N-free media. These 96-well plates were incubated at 30°C with shaking for 4 hours, after which 10 µL of cells was transferred to 190 µL of PBS with cycloheximide. These plates were used for “N-free Screen”. All flow cytometry readings were conducted on the Guava PCA-96 AFP system (Guava Technologies, Hayward, CA). For each well, 500 cells were analyzed and the average GFP fluorescence was noted. It typically took 40 minutes for the flow cytometer to process one 96-well plate.
The WT, npr2Δ, npr3Δ and npr2Δnpr3Δ cells, all with incorporated Gat1-GFP, were grown in the YPD medium in early logarithmic growth phase. For the N-free media analysis, the strains were pelleted, washed quickly with N-free media and inoculated in fresh N-free media for 30 minutes. For the YPD analysis, the strains grown in YPD were pelleted, washed with PBS once and inoculated in fresh PBS. The GFP signal was imaged directly by fluorescence microscopy.
Twenty five milliliters of YPD-grown cells (OD600 = 1) were pelleted by centrifugation, washed once with water and inoculated into 25 mL of indicated media for 30 minutes. The cells were pelleted by centrifugation and frozen. One milliliter of lysis buffer (PBS/1 mM EDTA/1% NP-40) with protease inhibitors (2 µg/mL aprotinin, 2 µg/mL leupeptin, 1 mM PMSF, 1 mM benzamidine) and phosphatase inhibitors (10 mM sodium fluoride, 2 mM sodium vanadate, 5 mM sodium pyrophosphate, 10 mM β-glycerophosphate) was used to resuspend the frozen cell pellet. The cell mixture was transferred to 15 mL Falcon tube containing 400 µL of glass beads (500 µm in diameter, Sigma). The tubes were vortexed at 4°C for 10 minutes. Supernatants were collected after centrifugation at 750 g for 5 min at 4°C, followed by another centrifugation at 10,000 g for 10 minutes at 4°C. The total protein lysate was normalized by Bradford assay, and 50 µg of protein was loaded onto 8% Bis-Tris gel (Invitrogen). The resolved lysate was transferred to a nitrocellulose membrane and probed with anti-HA antibodies. HRP-labeled secondary antibodies were detected by ECL detection kit (Amersham).
Indicated prototrophic haploid cells were grown for 4 hours in media containing 2% glucose and 0.17% Yeast Nitrogen Base w/o amino acids and ammonium sulfate, supplemented with either 5 mM proline or 5 mM glutamine. One tenth OD600 units of cells (corresponding to ∼2 million cells) was washed with media containing 2% glucose and 0.67% Yeast Nitrogen Base w/o amino acids. Cells were inoculated in 7.5 mL of this media and 2.5 µCi of [3H]-Threonine (Amersham) was added. At indicated timepoints, 1 mL of cells was vacuum filtered onto nitrocellulose filters (Whatman glass microfiber filters, diameter 25 mm) and washed twice with 5 mL of PBS. Retained radioactivity was determined by liquid scintillation method. After the last timepoint, OD600 was determined and it never differed more than 5% between cultures. Blank sample contained 1 mL of vacuum filtered and washed cells without radioactivity.
Indicated diploid cells from the deletion collection were grown in YPD plates for 24 hours. They were then inoculated into 1% potassium acetate solution, and the percentage of cells that had undergone sporulation was determined by microscope.
Indicated prototrophic haploid cells were grown in 150 mL of YPD to OD600 = 0.5. For N-free samples, YPD cultures were pelleted, washed once with N-free media, and inoculated into 150 mL of N-free media for 30 minutes. Cultures were quickly spun down and cells stored at −80°C until use. mRNA isolation, cDNA synthesis and labeling, and array processing was done as previously described [51]. For ribosomal protein gene expression analysis, all RPL and RPS genes were included in the analysis.
Total RNA was prepared by the same protocol as for the microarray experiment. Twenty micrograms of total RNA was treated with 1 U of DNase (Invitrogen) for 15 min at room temperature. Following heat denaturation at 65°C for 10 min, RT using SuperScript II was performed as recommended by the manufacturer (Invitrogen). One twentieth of the reaction was used for PCR. Thermal cycling was carried out for 20 cycles at 94°C for 20 s, 58°C for 20 s and 68°C for 45 s. For transcripts not visible at 20 cycles, 25 cycles of PCR was performed. The primer sequences were designed to distinguish A and B versions of RP mRNA.
Cell lysate was prepared as described for Npr1 phosphorylation detection. Five milligrams of total protein in 1 mL was used for immunoprecipitation experiments. Five micrograms of rabbit anti-HA antibody was added and immunocomplexes were allowed to form with rotation for 2 hours at 4°C. Forty microliters of a 50%-slurry of protein A-agarose was then added and incubated another 2 hours. Agarose beads were washed with lysis buffer (without protease inhibitors) four times and the beads were boiled in LDS Sample Buffer (Invitrogen) for 2 minutes. Immunoprecipitates were resolved by SDS-PAGE, transferred to a nitrocellulose membrane and probed with indicated antibodies.
HeLa cells were grown in DMEM, supplemented with 10% fetal bovine serum. At 60% confluency, 10-cm plates of cells were transfected with indicated constructs using FuGene transfection reagent (Roche). After 48 hours cells were washed with PBS and lysed in 1 mL of lysis buffer (PBS/1 mM EDTA/40 mM HEPES/1% NP-40) with protease inhibitors. Lysates were first cleared with a spin at 10,000 g for 5 min and then precleared with one-hour incubation with agarose G beads. One milligram of total cell lysate was used for immunoprecipitation with one microgram of mouse anti-myc antibody as described above for yeast.
|
10.1371/journal.pcbi.1002382 | Paradoxical Evidence Integration in Rapid Decision Processes | Decisions about noisy stimuli require evidence integration over time. Traditionally, evidence integration and decision making are described as a one-stage process: a decision is made when evidence for the presence of a stimulus crosses a threshold. Here, we show that one-stage models cannot explain psychophysical experiments on feature fusion, where two visual stimuli are presented in rapid succession. Paradoxically, the second stimulus biases decisions more strongly than the first one, contrary to predictions of one-stage models and intuition. We present a two-stage model where sensory information is integrated and buffered before it is fed into a drift diffusion process. The model is tested in a series of psychophysical experiments and explains both accuracy and reaction time distributions.
| In models of decision making, evidence is accumulated until it crosses a threshold. The amount of evidence is directly related to the strength of the sensory input for the decision alternatives. Such one-stage models predict that if two stimulus alternatives are presented in succession, the stimulus alternative presented first dominates the decision, as the accumulated evidence will reach the threshold for this alternative first. Here, we show that for short stimulus durations decision making is not dominated by the first, but by the second stimulus. This result cannot be explained by classical one-stage decision models. We present a two-stage model where sensory input is first integrated before its outcome is fed into a classical decision process.
| Decision making is of crucial interest in many disciplines such as psychology [1], [2], neuroscience [3]–[5], economics [6], [7], and machine learning [8]. Binary decision theories relate to situations where an observer (or machine) is confronted with one of two possible noisy stimuli ‘A’ and ‘B’. A decision has to be made whether ‘A’ or ‘B’ is present. For example, human readers have to decide whether a handwritten character is an or a ; a trader has to decide whether to sell or to keep; a monkey has to decide whether dots on a screen are moving to the left or to the right [9]. While engineering and economical decision theories focus on how to compute optimal decisions [6], [7], [10], psychology and neuroscience investigate the actual decision making process in humans and animals [9], [11]–[14].
Decision making is usually assumed to be a one-stage process where evidence integration and decision making are identical (but see [15], [16]). In a standard accumulator model each bit of evidence is integrated and a decision is reached once the accumulated evidence for one of the two response alternatives crosses a threshold [13], [14], [17]–[33]. If the evidence itself is noisy, then the accumulation of evidence for each of the two stimulus alternatives leads to a diffusion-like process. For example, in the well-known random motion paradigm [9], moving dots appear at random moments in time, so that evidence for leftward or rightward moments arrives probabilistically and the accumulator is expected to evolve along a stochastic path that can be approximated by a drift-diffusion process. This is in good accordance with experimental studies where neurons in the macaque lateral intraparietal cortex (LIP) increase firing rates along a noisy trajectory up to the moment of decision [9], [32]–[34]. Since evidence is very noisy in this case, and arrives slowly over time, the decision process is rather slow [9]. Most experimental [9], [11], [12] and theoretical work on decision making [5]–[8] focuses on paradigms where noisy stimuli are presented for long durations, e.g. until a response is elicited (for exceptions see [31], [32]).
In other paradigms, where stimuli are less noisy, decisions can be extremely fast. For example, humans only need a fraction of a second to recognize objects such as animals in a picture [35]. This astonishing speed is also evident in sports such as table tennis or soccer requiring rapid reactions to moving balls. In these examples, the brain has to decide rapidly upon visual information available for only a hundred milliseconds or less. Note that even in these scenarios where stimuli are of high contrast (“low noise”), the responses of the observers can still be “noisy”.
Here, we first show psychophysically that one-stage models of the noisy accumulator or drift-diffusion type cannot explain the results of feature fusion experiments where two stimulus alternatives are presented in rapid succession for durations in the range of 20–160 ms. Second, we propose, instead, a two-stage model, where evidence integration is separated from a noisy drift-diffusion decision making process. Our results reveal additional aspects of the dynamics of decision making that are hidden in standard experimental paradigms where only one stimulus alternative is presented per trial.
In our psychophysical experiments we worked with visual stimuli comprising two vertical bars with a small horizontal offset either to the left or to the right (Vernier stimulus, Figure 1). The contrast of the bars and the horizontal offset was chosen such that, after flashing the stimulus for 10 ms, human observers can reliably identify (accuracy above 90 percent correct) whether the lower vertical bar is offset to the left or right with respect to the upper vertical bar.
Next, we presented a sequence of two Vernier stimuli. As a reference we performed an experiment where the presentation of a first Vernier stimulus ‘A’ for 10 ms was separated from the presentation of a second Vernier stimulus ‘B’ by an interstimulus interval (ISI; blank screen) of variable duration. If the ISI was 50 ms, observers easily distinguished the two stimuli and could report, with an accuracy of above 90 percent correct, whether the first or the second Vernier stimulus was offset to the right. If the interstimulus interval was shorter, the accuracy dropped (Figure 1). The high precision of the subjects in spite of the short stimulus duration suggests that – in contrast to e.g. traditional random dot stimuli – the stimulus is highly informative with relatively little stimulus noise.
If the ISI is 0 ms, i.e. the two verniers are presented in immediate succession, feature fusion occurs [36]. Observers perceive only one single vernier with a smaller offset because the vernier offsets integrate and partially cancel each other out (Figure 2A; [37], [38]). Our feature fusion experiments with Vernier stimuli are analogous to classic feature fusion experiments with color. For example, observers perceive one single yellow disk when a red disk is rapidly followed by a green disk [39].
In experiment one, vernier stimulus ‘A’, offset either to the left or right, was immediately followed by a second vernier stimulus ‘B’ with opposite offset direction (right or left, respectively). The durations and of both verniers were equal, i.e. , but varied from 10 to 80 ms, each. Vernier stimulus ‘B’ dominates the percept the stronger the longer both vernier stimuli ‘A’ and ‘B’ are presented (Figure 2D). For example, when the two vernier stimuli are presented for 20 ms each, observers report a percept corresponding to stimulus ‘B’ in 60% of the trials, while ‘A’ is reported in only 40% of the trials. When the two stimuli are presented for 40 ms each, observers report a percept corresponding to stimulus ‘B’ in 67% of the trials, while ‘A’ reported in only 33% of the trials.
We wondered whether the dominance of the second stimulus could be explained by classical noisy accumulator models, also called Drift-Diffusion models. In the standard, one-stage Drift-Diffusion Model [20], [22], [23], [27], evidence for ‘A’ or ‘B’ translates directly into the drift rate (upward for ‘A’, downward for ‘B’) of a decision variable (Figures 2B, C). As usually, we added noise to the drift process leading to a random walk of the trajectory. The noise accounts for both noisiness of the evidence itself (an important aspect in the moving-dot paradigm [9], [33], [34], [40]) and internal noise in the brain. After presentation of both stimuli, the drift goes back to zero. A decision is made when hits the upper (for ‘A’) or lower bound (for ‘B’).
In this one-stage model, dominance of stimulus ‘A’ is the stronger the longer the presentation times of ‘A’ and ‘B’, and respectively. This is in striking contrast to the experimental results. We found that the qualitative nature of the results is independent of the specific choice of parameters of the one-stage drift diffusion model: for all tested parameters, the dominance of the second stimulus decreased with increasing duration (whereas the dominance of the second stimulus increased in the experiments). Whereas, for certain, fixed stimulus durations, we could achieve dominance of the second stimulus with specifically optimized parameters, we could never achieve dominance of the second stimulus for the entire range of stimulus durations with one set of parameters.
We explored whether minor modifications of the one-stage drift-diffusion model can explain the dominance of the second vernier. For example, we replaced the noisy accumulator by a noisy leaky accumulator. However, this did not change the results qualitatively. We then tested a very basic two stage model. During stimulus presentation, the stimulus served as the drift in a noisy leaky integrator model. After stimulus termination, the leak was artificially set to zero and the integration continued as a free, unbiased noisy diffusion process. In other words, the result of the leaky evidence integration served as initial condition for the leak-free diffusion process. While qualitatively such a drift-diffusion model explains the dominance results well (Supporting Figure S3B), we suggest an alternative model, which accounts very well for both the dominance and the reaction time distributions.
In this two-stage model, the evidence integration enters the second stage as a drift rate rather than as a bias in the initial condition. (a) During stage one, evidence integration is leaky and dominated by the intrinsic noise of the stimulus. The variable of noisy evidence integration is . (b) Stage two starts after a fixed time after stimulus onset and ends when a second variable hits the upper or lower decision threshold. (c) The variable of the leaky integrator of stage one sets the drift in the (leak-free) drift-diffusion model of stage two.
The combination of (b) and (c) implies that, for long stimuli, stage two is a drift-diffusion model with time-dependent drift set by the momentary value of the integration variable of stage one. In case that the total duration of the stimulus is shorter than the time needed to reach the decision threshold in stage two, the value of the leaky integrator of stage one at the end of the stimulus is written into a buffer and this buffered value serves, during the remaining time, as the (constant) drift for the diffusion process in stage two until a decision is reached. In the limit that stimuli are shorter than , stage two has therefore a constant drift. In the limit that stimuli are presented for times much longer than (so that is negligibly short compared to the stimulation time), our two-stage model becomes equivalent to a standard one-stage drift-diffusion model with a time-dependent drift that is given by the low-pass filtered version of the input signal. However, for very short stimuli, the prediction of our two-stage model is remarkably different from that of a standard one-stage model – and these ultra-short stimuli are at the center of our study.
The results on stimulus dominance during the feature fusion paradigm with two short Verniers can indeed be explained by the two-stage model (Figures 2E, F). Since our stimuli are comparatively strong (over 90 percent accuracy for stimli presented separately), we consider the limit where the evidence integration in stage one is noise-free. Hence, in the first integration stage, evidence for stimulus ‘A’ and ‘B’ is simply accumulated in a noiseless forgetful (leaky) integrator (see also [30]). The time scale of forgetting is related to the time over which an ideal observer expects stimuli to remain constant (see Materials and Methods). The second phase, the decision stage starts at a fixed time and consists of a standard drift-diffusion model without leak (Figure 2F, bottom panel). For a sequence of two short stimuli, the stimulation ends before so that at the termination of the second stimulus (), the output of the evidence integration is written into a buffer and fed later from the buffer as a constant drift rate into stage two. The two-stage model captures the dominance of the second vernier very well (Figure 2G).
The critical test for models of decision making is to account for reaction time distributions rather than accuracy [20]. We therefore wondered whether the two-stage model captures the reaction time distributions in the fusion experiments. In experiment two, stimulus ‘A’ (the first vernier stimulus) was presented for a duration , immediately followed by stimulus ‘B’ (a vernier with opposite offset) of duration with (Figure 3A). Parameters of the two-stage model were adapted individually for each observer and kept fixed across all stimulus conditions. The dominance of the first vernier stimulus increased when increased (Figure 3B). Reaction times for strongly biased situations (e.g. where the first vernier stimulus is much longer than the second one or vice versa) are faster (75% of decisions made before 560 ms) than those in conditions with dominance around 50% (75% of decisions made before 610 ms) leading to an inverted-U-shaped curve of the reaction time quantiles (Figure 3C). The same pattern is observed when responses for the first and second vernier stimulus are analyzed separately (Figure 3D).
Median response times varied strongly across the 13 observers (Figure 4A). We separated the observers into a group of fast responders (median reaction time <500 ms) and one of slow responders (median reaction time >500 ms). While the reaction times of both groups show an inverted U-shape function, the qualitative picture is different between slow and fast responders. If the first vernier stimulus is presented for a short time only, fast responders are particularly fast whereas slow responders are particularly slow. The two-stage model qualitatively reproduces this behavior (Figures 4B,C, Supporting Figure S1).
For each stimulus condition, the outcome of the leaky integration in the first stage serves as a drift of the leakfree drift-diffusion model during the second stage of the two-stage model. For short stimuli, like the ones considered so far, where the stimulus ends before the integration, the result of stage one is written into a buffer and used as a constant value of the drift in the decision stage. In other words, the evidence at stimulus termination serves as drift value, rather than as an initial condition of stage two.
As an alternative, we have also analyzed a drift-diffusion model where the drift was taken as a free parameter, optimized for each stimulus condition independently so as to optimally predict the distribution of reaction times. The drift predicted from this model (which has more degrees of freedom) is statistically not different (as determined by a two way repeated measures analysis of variance) from our two-stage model where the drift is not a free parameter but the result of stage one. This finding suggests that the simple preprocessing by leaky integration correctly determines the drift rate (Figure 4D).
Further above we had reported that a qualitative fit of the dominance was possible by a noisy leaky integrator, if the leak was set to zero at the end of the stimulus. In such a model, the result of the leaky evidence integration serves as the initial value for a free diffusion process. The results of Figure 4D, however, indicate that the result of the leaky integration in the first stage should be used as the drift, and not as an initital condition for the diffusion in stage two. The results of Figure 4D can therefore be considered as a strong argument in favor of the two-stage model. In the following, we consider other aspects of the two-stage model.
If the writing into the buffer is triggered at stimulus termination, as assumed in the two-stage model, the question arises why the switch from ‘B’ to the background, but not that from stimulus ‘A’ to ‘B’, triggers the transition from stage one to stage two in the two-stage model. We suggest that the large change from a vernier stimulus to background is “interpreted” as stimulus termination because there is a strong neural off-transient for a change from ‘A’ to a blank screen, whereas there are no on- and off-transients for a change from ‘A’ to ‘B’, respectively [41]. This is well in accordance with a Bayesian approach (see Supporting Figure S2) suggesting that feature integration should terminate when it becomes unlikely that the momentary stimulus is a continuation of the previous stimulus. The readout in the two-stage model should therefore start when a novelty value of the momentary stimulus crosses a predetermined threshold (cf. Supporting Text S1).
We tested this prediction by the psychophysical experiment in Figure 1, where the first vernier stimulus was followed by a blank background (interstimulus interval; ISI) before the second vernier was presented. With an ISI of 20 ms, the two vernier stimuli, presented for 10 ms each, became individually discriminable. Observers could tell whether the first stimulus was offset to the left or to the right by motion cues [41], [42]. However, for a sequence of ‘A’ immediately followed by ‘B’ with , verniers are not individually visible even though the total duration is 40 ms as in the sequence with the 20 ms ISI. This suggests that in the condition with the 20 ms ISI the termination signal of the first vernier stimulus stopped evidence integration and wrote the result into a buffer, for later use in stage two, whereas evidence was integrated across the two vernier stimuli in the experiment without the blank, before the final result was written into a buffer.
In our experiments with ultra-short stimuli, the time where the read-out from the buffer starts, occurs after stimulus termination (and is included in the non-decisional time . We also tested a model where the decision process was triggered at stimulus termination, i.e., at the same moment when the result of evidence integration is written into the buffer (i.e. ). Such a model predicts that reaction times increase with total stimulus duration (data not shown), which disagrees with our observation that, for a given level of dominance, the mean reaction times remain largely constant for total stimulus durations of 20 ms, 40 ms, and 80 ms (Supporting Figure S3 D).
For most of the stimuli considered so far, the total stimulus duration was below 40 ms. In this case, the two stages of the model are sequential and do not overlap. However, for longer stimuli, evidence integration of stage one is not finished at the moment of when the diffusive decision process in stage two is started.
Indeed, a model with fixed drift in stage two works well for stimuli up to a total duration of 80 ms, but breaks down at 160 ms (data not shown). However, our two stage model assumes that as soon as stimuli extend beyond , the momentary value of the evidence integration stage is written into the buffer and immediately used as drift in the diffusion process of stage two. The drift is updated continuously so that the diffusion process becomes time-varying. The fact that a constant drift in stage two fails when the stimulus extends over 160 ms indicates that the parameter of our model is much shorter than 160 ms. We tested this by fitting for individual subjects such that the mean square error in the dominance was minimized across all stimulus durations, including the 160 ms conditions. The optimal values for were indeed smaller than 160 ms (, , ).
In the model, we explored the situation that the first stimulus becomes much longer than . Obviously, if the first stimulus is made very long, our two-stage model then predicts that the first stimulus dominates.
Most models of decision making do not account for the timing of stimuli. Likewise, most experiments use long stimulus durations in the range of several hundreds of milliseconds to seconds [9], [31], [32] or constant stimuli [20], [22], [25]. However, decision making has to occur in many situations in less than 100 ms, for example, when driving a car or playing soccer. Here, we have shown that rapid decision processes show very different characteristics than decision processes on longer time scales. For example for short stimulus durations, later presented stimuli dominate over stimuli presented earlier. We propose that these processes are also present in longer lasting decision processes, but are hidden and barely measureable.
To study the dynamics of rapid decision processes, we used a feature fusion paradigm. This paradigm relies on the well known effect of visual integration masking [43], [44] and follows partly Bloch's law [45]. The results of our experiments are in agreement with earlier results on feature fusion [38] and backward masking experiments [43], but do not agree with the traditional one-stage models of decision making in which evidence is integrated until a decision boundary is reached. The results of our experiment rather support a two-stage model in which evidence integration is separate from the actual decision process. This model is fundamentally different from classical drift-diffusion models [20], [22], [23], [27], race models [18], [19], [25], [32], attractor models [10], [30], [31], one-stage models with pre-processing [28], and “parallel” two-stage models [21], [46]. All these models predict the first stimulus to dominate when in contrast to the fusion results.
In our model, we assumed several components which are worth discussion each. First, evidence integration in stage one must be leaky. It is the leak that explains why, when the first and second vernier stimulus are of the same duration (), the second vernier stimulus dominates (experiment one). The leak in our model arises naturally from a Bayesian approach and can be traced back to the fact that stimuli are expected to change in natural environments. Similar to our Bayesian novelty detection approach (cf. Supporting Text S1), the leaky evidence integration can also be derived in the framework of Kalman filters [47]–[50].
Second, the accumulated evidence must transferred at an appropriate moment and written into a temporal buffer. Such a buffer is necessary since decisions often occur a considerable time after the stimulus has disappeared. We suggest that the precise moment of transfer is set by a novelty score monitored during evidence integration (see Supporting Text S1). Such a novelty signal and subsequent buffering explains why the two vernier stimuli are perceived individually, if the stimuli are separated by a blank screen (ISI), but fused into a single percept in the absence of the blank. In this sense, feature fusion can be interpreted as a failure to detect the onset of a new stimulus because the new evidence is not sufficiently different to raise a ‘novelty signal’. In contrast, the switch from stimulus to background creates a sufficiently strong transient to stop the feature integration process (Supporting Figure S2).
Third, the noisy decision process is triggered at a fixed time after stimulus onset. If the decision process were triggered at , reaction times would increase with stimulus duration. This is, however, not the case (Supporting Figure S3D). From the fact that our model assumes a fixed start time of the second stage, it necessarily follows that we have to distinguish two different situations: If the total stimulus duration is shorter than , we need to bridge the time between the end of the stimulus and decision by storing the intermediate result of evidence integration into a buffer. This value is then used in stage two as a fixed mean drift rate. If the total stimulus duration is longer than , the result of stage one is used online as a time-dependent drift for all times until the end of the stimulus (at which point it is again ‘frozen’ and transferred into the drift-buffer.).
Our two stage model is similar to previous two stage models in which sensory processing, e.g. motion processing or contrast detection, precedes a decision making stage (e.g. [15], [16], [46]). In our model, the sensory integration stage is leaky to account for the dominance of the second vernier.
Our two-stage model comprises a leaky integration stage followed by a drift-diffusion stage. The question arises whether or not a one-stage model with leak in the drift-diffusion process can explain the results. However, this is not the case because in such a model always the first stimulus dominates because the leak pushes the decision variable towards the starting point and not across it (Supporting Figure S3).
Another way to integrate the leak into a one-stage model is to directly transform the input by a leaky integrator (like our stage one) and to use the outcome of the leaky integrator as a time-variant drift in stage-two (, ). However, using stage one only for pre-processing will not change the pattern of results [28]. In such models, the decision variable also moves towards the decision bound for stimulus ‘A’ before dropping back to chance level. Therefore, these models also show a dominance of the first stimulus.
The novel features of the two-stage models are observable well only for stimuli in the range of up to about 100 ms. This duration is in line with the duration of visual integration found in other studies [51]–[53]. One of the paradoxical aspects of our model is that the second stage starts at a fixed time . Obviously, if the duration of a stimulus extends beyond , then the stage of evidence integration and that of stochastic decision making (stage two) will overlap and the separation into two distinct phases disappears (see Supporting Text S1). Therefore it is not surprising that for longer stimulus durations standard one-stage models work well [27], [29], [54].
In our model, a deterministic filter (leaky integrator) is applied in stage one to a step-like input, representing a noiseless stimulus. This is the limiting case where the stimulus is considered to be of high contrast. In a more realistic scenario the stimulus itself is noisy. The stochasticity of stimuli leads, after stage one, to a noisy result of evidence integration, which is written into the buffer and then used as drift for stage two. This noisy result is modeled by the variance of the drift constant of stage two. It is therefore tempting to relate the stochasticity of drift constants to sensory or physical noise. The stochasticity of stage two may be related to internal noise in the brain [32], [55]. What is the advantage of adding a separate noisy decision process? It is well known that human observers can manipulate the speed-accuracy trade-off according to instruction or reward scheme by a change in strategy corresponding to a shift of the initial condition, , or the decision thresholds in the drift-diffusion process [13], [54].
Neurons in the superior colliculus [56], the LIP [9], [32], [33], the pre-motor cortex [57], [58], and the dorsoventral lateral prefrontal cortex [11], [12], [59] were shown to be involved in decision making. The firing rate of these neurons increases as long as stimuli are displayed. This ramping activity may relate either to evidence accumulation (“stage one”) or to decision making (“stage two”). Future experiments with feature fusion stimuli may be used to decide between these two alternatives.
In summary, it is often (intuitively) assumed that visual input directly translates into decisions. A stimulus presented first should drive decisions stronger and faster than a later stimulus (first in, first out). This is obviously correct when the two stimuli are long, because a decision may be reached even before the second stimulus can influence decision masking. In this case, we can assume that evidence integration and decision making are the same. However, for short stimuli this is not the case. Evidence integration and decision making can only be disentangled, when the two stimulus alternatives are presented within one trial (feature fusion) but not when only one stimulus is presented per trial, as it is usually. The distinction between evidence integration and decision making is described well by our two-stage model, where rapid stimuli are integrated and buffered before the decision process starts.
All participants signed informed written consent. The study was approved by the Commission cantonale (VD) d'éthique de la recherche sur l'être humain (Lausanne, Switzerland) and conducted according to the principles expressed in the Declaration of Helsinki.
A total of 24 observers (8 female, aged 21–32 years) signed informed written consent. Participants had normal or corrected-to-normal visual acuity as measured by the Freiburg visual acuity test [60]. All but two observers (the first and second author) were naive to the purpose of the study. Naive observers were paid students from local universities.
Stimuli were presented on a Tektronix 608 X-Y display or a HP 1332A X-Y display. Both X-Y displays were equipped with a P11 phosphor and controlled by a PC via a fast 16 bit DA converter. Stimuli were presented at , a 1 MHz dot rate, a 500 Hz refresh rate, and a dot pitch of . Viewing distance was 2 m. The room was dimly illuminated by a background light () to prevent adaptation to scotopic vision. Stimulus contrast was close to 1.0. In each experiment, the conditions have been presented randomly interleaved to reduce the influence of hysteresis, learning, or fatigue in the averaged data.
The vernier stimuli were composed of two vertical segments. Each segment was 10′ (arc min) long, 0.5′ wide, separated by a vertical gap of 1′. A small horizontal offset was inserted between the upper and the lower segments (Figure 2A). Horizontal offset sizes ranged from 30″ to 40″ (arc sec). Offsets were chosen individually to be at least twice the offset size of the offset discrimination threshold for a single vernier stimulus of 20 ms duration as determined using the adaptive PEST procedure [61]. A sequence of two vernier stimuli with opposite offset directions was presented foveally in rapid succession. The offset direction of the first vernier (stimulus ‘A’) was chosen randomly in each trial (left or right). The second vernier (stimulus ‘B’) had an offset direction opposite to that of the first vernier. If, for example, the first vernier stimulus was offset to the left, the second vernier was offset to the right, and vice versa. Observers perceived only one fused vernier and were asked to report the position of the lower segment with respect to that of the upper segment by pressing one of two push buttons. Observers were instructed to respond as rapidly as possible, but also as accurately as possible. No feedback about performance was given. Naive observers did not know that a sequence of two vernier stimuli was presented.
We computed dominance, defined as the proportion of trials on which the response matched the offset direction of the first vernier stimulus. Thus, values above 50% indicate dominance of the first vernier (stimulus ‘A’); values below 50% indicate dominance of the second vernier (stimulus ‘B’). 50% vernier dominance is the point of subjective equality, i.e. first and second vernier stimulus equally contribute to performance.
First vernier stimulus (‘A’) and second vernier (‘B’) were presented in immediate succession (Figure 2). Both vernier stimuli had either the same duration or the duration of one of the verniers was four times longer than the other. The total duration of the first and second vernier was 20 ms, 40 ms, 80 ms, or 160 ms. All conditions were presented in a random order. Every condition has been repeated 400 times per observer.
As Experiment 1, except for that the duration of the first vernier was varied in 12 steps between 0 ms and 40 ms. The total duration + always summed up to a total of 40 ms (Figure 3A). Every condition has been repeated 400 times per observer.
In Figure 1, an ISI was inserted between the first and second vernier stimulus. Observers were informed about the experimental design and asked to indicate whether the first or second vernier stimulus was offset to the right.
Reaction times below 300 ms or above 1200 ms were excluded from analysis to reduce the impact of motor errors and unattended trials (less than 3% of the trials).
We model the stimuli by a time-varying input signal , which is +1 during the presentation of stimulus ‘A’, −1 for stimulus ‘B’ and 0 otherwise. In the evidence accumulation stage of the two-stage model, the stimulus is subjected to leaky integration: Since our stimuli have high contrast, the evidence integration is modeled as a noise-free process.
For times larger than the integrated evidence is fed as the drift into the noisy drift-diffusion model at stage two. We distinguish two different cases. a) Stimuli are shorter than . At the termination of stimulus ‘B’ () the integrated evidence is stored and written into a buffer. Later, for the buffered value is used as the mean drift rate with a fixed scaling factor for the decision stage, which encompasses a standard drift-diffusion model. b) Stimuli are longer than . In this case the momentary evidence is used as the mean drift for . Again, at the end of the stimulus, the last value of the evidence is buffered and used as drift henceforth.
During stage two, in every trial, a decision variable is initialized at and evolves according to the Langevin equation where is the drift rate and is a Wiener process, which introduces noise to the decision process. A decision is made when the decision variable reaches one of two decision boundaries (decision ‘A’) or (decision ‘B’). The associated reaction time is the sum of a non-decisional time (which accounts for sensory delay , and the evidence integration and buffering times as well as motor delays) and the time when the decision variable reaches the boundary. We used the Ratcliff extension [20] of a standard drift diffusion model, in which the non-decisional time , the initial condition and the drift rate vary stochastically from trial to trial. The non-decisional time is drawn from uniform distributions with mean and width . The initial condition is drawn from uniform distributions with mean and width . The drift rate is drawn from a Gaussian distribution with mean – the output of the first stage – and standard deviation . and represent noise in the evidence accumulation.
As a reference, we used a one-stage model, which encompasses a standard drift diffusion model, in which the drift rate depends on time and is given by the input signal: . In this model, the drift becomes zero after the end of the stimulus. We also simulated leaky variants of this one-stage model, for details see Supporting Text S1.
In the first step, the parameters , , , , , and of the decision stage were fitted to the experimentally obtained cumulative reaction time distributions by minimizing the product of the p-values of the Kolmogorov-Smirnov statistic for each stimulus condition [62], [63]. Responses to stimuli ‘A’ and ‘B’ and different stimulus conditions were fitted simultaneously using the fast-dm software of Voss & Voss [64]. For both experiments, fits were done individually for each observer. In the experiment of Figures 2, all parameters except the mean drift rate and the drift variability were the same in all stimulus conditions. The drift was calculated from stage one. Drift variability was a function of stimulus duration. In the experiment of Figures 3 and 4, only the mean drift rate was varied across conditions and calculated from stage one. In order to obtain the parameters and of the evidence integration in stage one we ran a simulation experiment with free drift rates as in Figure 4D. The obtained mean drift rates were then used to fit the time constant and the scaling factor , again separately for each observer. This fit was done using the fit-routine of MATLAB. Finally, to extract the optimal values for , we first used the data of experiment 1 with stimulus durations and fitted the parameters of both stages with the described procedure. Then, we performed a line scan of all values of and identified the value that minimized the mean square error of the measured dominance, now including the long duration of 160 ms.
Parameters are different for each observer, i.e. , , , , , , , and for stage one and .
|
10.1371/journal.pcbi.1004104 | A Biologically Plausible Computational Theory for Value Integration and Action Selection in Decisions with Competing Alternatives | Decision making is a vital component of human and animal behavior that involves selecting between alternative options and generating actions to implement the choices. Although decisions can be as simple as choosing a goal and then pursuing it, humans and animals usually have to make decisions in dynamic environments where the value and the availability of an option change unpredictably with time and previous actions. A predator chasing multiple prey exemplifies how goals can dynamically change and compete during ongoing actions. Classical psychological theories posit that decision making takes place within frontal areas and is a separate process from perception and action. However, recent findings argue for additional mechanisms and suggest the decisions between actions often emerge through a continuous competition within the same brain regions that plan and guide action execution. According to these findings, the sensorimotor system generates concurrent action-plans for competing goals and uses online information to bias the competition until a single goal is pursued. This information is diverse, relating to both the dynamic value of the goal and the cost of acting, creating a challenging problem in integrating information across these diverse variables in real time. We introduce a computational framework for dynamically integrating value information from disparate sources in decision tasks with competing actions. We evaluated the framework in a series of oculomotor and reaching decision tasks and found that it captures many features of choice/motor behavior, as well as its neural underpinnings that previously have eluded a common explanation.
| In high-pressure situations, such as driving on a highway or flying a plane, people have limited time to select between competing options while acting. Each option is usually accompanied with reward benefits (e.g., avoid traffic) and action costs (e.g., fuel consumption) that characterize the value of the option. The value and the availability of an option can change dynamically even during ongoing actions which compounds the decision-making challenge. How the brain dynamically integrates value information from disparate sources and selects between competing options is still poorly understood. In the current study, we present a neurodynamical framework to show how a distributed brain network can solve the problem of value integration and action selection in decisions with competing alternatives. It combines dynamic neural field theory with stochastic optimal control theory, and includes circuitry for perception, expected reward, effort cost and decision-making. It provides a principled way to explain both the neural and the behavioral findings from a series of visuomotor decision tasks in human and animal studies. For instance, the model shows how the competitive interactions between populations of neurons within and between sensorimotor regions can result in “spatial-averaging” movements, and how decision-variables influence neural activity and choice behavior.
| From very simple decisions, such as selecting what to wear or choosing a place for dinner, to more complex decisions, such as trading in the stock market, we usually have to select among competing options. Selecting between alternatives requires assigning and integrating values along a multitude of dimensions with different currencies, like the energetic cost of movement and monetary reward. Solving this problem requires integrating disparate value dimensions into a single variable that characterizes the “attractiveness” of each option. In dynamic decisions, in which the environment changes over time, this multi-dimensional integration must be updated across time. Despite the significant progress that has been made in understanding the mechanisms underlying dynamic decisions, little is known on how the brain integrates information online and while acting to select the best option at any moment.
A classic psychological theory, known as “goods-based decision making”, posits that decision making is a distinct cognitive function from perception and action and that it entails assigning values to the available goods [1–5]. According to this theory, multiple decision determinants are integrated into a subjective economic value at the time of choice. The subjective values are independently computed for each alternative option and compared within the space of goods, independent of the sensorimotor contingencies of choice. Once a decision is made, the action planning begins. This view is in accordance with evidence suggesting the convergence of subjective value in the orbitofrontal cortex (OFC) and ventromedial prefrontal cortex (vmPFC), where the best alternative is selected (for a review in “goods-based” theory see [4]). Despite the intuitive appeal of this theory, it is limited by the serial order assumption. Although many economic decisions can be as simple as choosing a goal and pursuing it, like choosing between renting or buying a house, humans evolved to survive in hostile and dynamic environments, where goal availability and value can change with time and previous actions, entangling goal decisions with action selection. Consider a hypothetical hunting scenario, in which a predator is faced with multiple alternative valuable goods (i.e., prey). Once the chase begins, both the relative value of the goods and the cost of the actions to pursue these goods will change continuously (e.g., a new prey may appear or a current prey may escape from the field), and what is currently the best option may not be the best or even available in the near future. In such situations, the goals dynamically compete during movement, and it is not possible to clearly separate goal decision-making from action selection.
Recent findings argue against a purely “goods-based” theory for decisions between actions, suggesting that decisions are made through a continuous competition between partially prepared action-plans. According to this “action-based” theory, when the brain is faced with multiple goals, it initiates concurrent and partially prepared action-plans that compete for selection and uses value information accumulated during ongoing actions to bias this competition, until a single goal is pursued [6–11]. This theory received support from neurophysiological [8, 12–15] and behavioral [9, 10, 16–21] studies, and it is in accord with the continuous flow model of perception, which suggests that response preparation can begin even before the goal is fully identified and a decision is made [22–24]. However, it is vague as to how action costs are dynamically integrated with good values and other types of information.
In the current study, we propose a distributed neurodynamical framework that models the neural basis of decision-making between actions. It allows dynamic integration of value information from disparate sources, and provides a framework that is rich enough to explain a broad range of phenomena that have previously eluded common explanation. It builds on successful models in dynamic neural field theory [25] and stochastic optimal control theory [26] and includes circuitry for perception, expected reward, selection bias, decision-making and effort cost. We show that the complex problem of action selection in the presence of multiple competing goals can be decomposed into a weighted mixture of individual control policies, each of which produces optimal action-plans (i.e., sequences of actions) to pursue particular goals. The key novelty is a relative desirability computation that dynamically integrates value information to a single variable, which reflects how “desirable” it is to follow a policy (i.e., move in a particular direction), and acts as a weighting factor on each individual policy. Because desirability is state- and time- dependent, the weighted mixture of policies automatically produces a range of behavior, from winner-take-all to weighted averaging. Another important characteristic of this framework is that it is able to learn sensorimotor associations and adapt the choice behavior to changes in decision values of the goals. Unlike classic neurodynamical models that used hard-wired associations between sensory inputs and motor outputs, we allow for plasticity in the connections between specific dynamic neural fields (DNFs) and, using reinforcement learning mechanisms, we show how action-selection is influenced by trained sensorimotor associations or changing reward contingencies.
By integrating dynamic neural field theory with control theory, we developed a biologically plausible framework that explains many aspects of human and animal behavior and its neural underpinnings in decisions between actions. It provides insights to a variety of findings in neurophysiological and behavioral studies, such as the competitive interactions between populations of neurons within [27, 28] and between [29] brain regions that result frequently in spatial averaging movements [10, 18, 30] and the effects of decisions variables on neuronal activity [31, 32]. We also make novel predictions concerning how changes in reward contingencies or introduction of new rules (i.e., assigning behavioral relevance to arbitrary stimuli) influence the network plasticity and the choice behavior.
The basic architecture of the framework is a set of dynamic neural fields (DNFs) that capture the neural processes underlying cue perception, motor plan formation, valuation of goods (e.g., expected reward/punishment, social reward, selection bias, cognitive bias) and valuation of actions (e.g., effort cost, precision required), Fig. 1. Each DNF simulates the dynamic evolution of firing rate activity within a neural population. It is based on the concept of population coding, in which each neuron has a response tuning curve over some set of inputs, such as the location of a good or the end-point of a planned movement, and the responses of a neuronal ensemble represent the values of the inputs. The functional properties of each DNF are determined by the lateral interactions within the field and the connections with the other fields in the architecture. Some of these connections are static and predefined, whereas others are dynamic and change during the task. The projections between the fields are topologically organized, that is, each neuron in one field drives activation of the corresponding neuron (coding for the same direction) in the fields to which it projects.
Let’s consider the hunting scenario, in which a predator is faced with multiple goods (i.e., prey) located at different distances and directions from the current state xt. The architectural organization of the present framework to model this type of problem is shown in Fig. 1. The “spatial sensory input field” encodes the angular spatial representation of the alternative goods in an egocentric reference frame. The “goods value field” encodes the goods values of pursuing each prey animal irrespective of sensorimotor contingencies (e.g., effort). The “context cue field” represents information related to the contextual requirements of the task. The outputs of these three fields send excitatory projections to the “motor plan formation field” in a topological manner. Each neuron in the motor plan formation field is linked with a motor control schema that generates both a direction-specific optimal policy πi, which is a mapping between states and best-actions, and an action cost function Vπi that computes the expected control cost to move in the direction ϕi from any state (see Methods section and S1 text for more details). It is important to note that a policy is not a particular sequence of actions—it is rather a function that calculates the best action-plan ui (i.e., sequences of actions/motor commands) to take from the current state xt to move in the direction ϕi for tend time-steps (i.e., πi(xt) = ui = [ut, ut+1, …, utend]).
The neurons in the motor plan formation field with activation levels above a threshold, γ, trigger the associated motor schemas. Once a motor schema is active, it generates a policy π(xt) towards the preferred direction of the corresponding neuron. The associated action cost is encoded by the action cost field. The “action cost field” in turn inhibits the motor plan formation field via topological projections. Therefore, the role of the motor plan formation field is two fold: i) trigger the motor schemes to generate policies and ii) integrate information associated with actions, goals and contextual requirements into a single value that characterizes the “attractiveness” of each of the available policies. The output of the motor plan formation field encodes what we call the relative desirability of each policy and the activity of the field is used to weigh the influence of each policy in the final action. As soon as the chase begins, the action costs and the estimates of the values related to the goals will change continuously—e.g., a new prey may appear at the field modulating the relative reward of the alternatives. The advantage of our theory is that it integrates value information from disparate sources dynamically while the action unfolds. Hence, the relative desirability of each policy is state- and time- dependent and the weighted mixture of policies produces a range of behavior from winner-take-all selection of a policy to averaging of several policies.
The general framework described above can be translated into more concrete and testable scenarios, such as visuomotor decision tasks with competing goals and/or effectors for comparison with experimental data. We show how the proposed computational model can be extended to involve motor plan formation DNFs for various effectors which interact competitively to implement effector as well as spatial decision-making.
One of the most fundamental questions in decision neuroscience is how the brain selects between alternative options. Classic serial theories, such as the “goods-based” model, suggest that decision-making is an independent cognitive process from the sensorimotor processes that implement the resulting choice [1]. According to this view, decisions are made in higher cognitive centers by integrating all the decision values of an option into a subjective value and selecting the best alternative option. Once a decision is made, the sensorimotor system implements the selected choice [2–4]. Clinical and neurophysiological studies have shown that a neural representation of subjective values exists in certain brain areas, most notably in the orbitofrontal cortex (OFC) and ventromedial prefrontal cortex (vmPFC) [3, 45].
Although the “goods-based” theory can explain a variety of decisions, such as choosing a place to have dinner tonight, it is limited by the serial-order assumption. This raises the question of how we decide between competing actions in dynamic environments, where the value and the availability of an option can change unpredictably. One possibility that has been suggested by a number of researchers is that decisions between actions are made through a continuous competition between concurrent action-plans related to the alternative goals [8, 11, 13, 46]. According to this “action-based” theory, the competition is biased by decision variables that may come from higher cognitive regions, such as the frontal cortex, but the sensorimotor system is the one that decides which action to select [8, 15]. This view has received apparent support from series of neurophysiological studies, which found neural correlates of decision variables within sensorimotor brain regions (for review, see [7, 13, 15]). Additionally, recent studies in non-human primates have shown that reversible pharmacological inactivation of cortical and sub-cortical regions, which have been associated with action planning or multisensory integration, such as lateral intraparietal area (LIP) [47, 48], superior colliculus (SC) [49], and dorsal pulvinar [50], cause decision biases towards targets in the intact visual field. These findings suggest that sensorimotor and midbrain regions may be causally involved in the process of decision-making and action-selection.
While these experimental studies have contributed significantly to understanding the mechanisms underpinning decisions between multiple actions, they are framed using theories that are insufficient to describe, much less predict, existing findings. Several computational frameworks exist that explain how information from disparate sources, such as goods values, action costs, prior knowledge, and perceptual information, are integrated dynamically to evaluate and compare the available options [51–53]. However none of these generate continuous actions which change the desirability and availability of options as they unfold. Instead, they generate discrete choices and not population codes that can be used to guide action in continuous parameter spaces, and they do not capture the interactions between networks of brain regions in any realistic way (see below). In the current study, we propose a neurodynamical framework that models the neural mechanisms of value integration and action-selection in decisions with competing options. It is comprised of a series of dynamic neural fields (DNFs) that simulate the neural processes underlying motor plan formation, expected reward and effort cost. Each neuron in the motor plan formation DNF is linked with a stochastic optimal control schema that generates policies towards the preferred direction of the neuron. A policy is a function that maps the current state into optimal sequences of actions. The key novelty of the framework is that information related to goals, actions and contextual requirements is dynamically integrated by the motor plan formation DNF neurons and that this information changes and is re-evaluated as the action is performed. The current activity of each of these DNF neurons encodes what we call “relative desirability” of the alternative policies, because it reflects how “desirable” it is to follow a particular policy (i.e., move in a particular direction) with respect to the alternative options at a current state, and weights the influence of each individual policy in the final action-plan.
The present framework is related to classic “sequential sampling models”, such as the drift diffusion model (DDM) [54] and the leaky competing accumulator (LCA) [55], which suggest that selection between competing options involves the gradual accumulation of sensory evidence until a threshold is reached. Although the framework also employs an “accumulator” mechanism to dynamically integrate value information and decide between options, and an action threshold to determine when to trigger motor schemas, it is quite different from these classic models. The classic models assume that populations of neurons in particular brain areas accumulate the sensory evidence and other brain areas compare the accumulated evidence to make decisions. For instance, studies in oculomotor decision-making suggest that LIP neurons accumulate the evidence associated with the alternative options and the accumulated activity of these neurons is compared by “decision brain areas” to select an option [13, 56]. Unlike the classic models, in the present framework the decisions are made through a dynamic transition from “weighted averaging” of individual policies to “winner-take-all”—i.e., a set of policies associated with an available goal dominates the rest of the alternatives—within the same population of neurons in the motor plan formation field. It does not assign populations of neurons to individual options, rather the alternative options emerge within a distributed population of neurons based on the current sensory input and prior knowledge. Once any neuron reaches the action threshold its associated motor schema is activated, but the decision process continues and other motor schemas can become active during the performance of an action. Because of these characteristics, the present framework can handle not only binary choices, but also decisions with multiple competing alternatives, does not require a “decision” threshold (although it does use a pre-defined “movement initiation” threshold), and can model decisions in which subjects cannot wait to accumulate evidence before selecting an action; rather they have to make a decision while acting.
Additionally, the present framework shares many features with other systems-level computational frameworks that have been previously proposed to model decisions between multiple actions [57, 58]. However, these approaches do not incorporate the idea of dynamically integrating value information from disparate sources (with the exception of Cisek’s (2006) model [58], which demonstrated how other regions, including prefrontal cortex, influence the competition), they do not model action selection tasks with competing effectors, and they do not model the eye/hand movement trajectories generated to acquire the choices. By combining dynamic neural fields with stochastic optimal control systems the present framework explains a broad range of findings from experimental studies in both humans and animals, such as the influence of decision variables on the neuronal activity in parietal and premotor cortex areas, the effect of action competition on both motor and decision behavior, and the influence of effector competition on the neuronal activity in cortical areas that plan eye and hand movements.
The model presented here bears some similarity to decision-making models in the hierarchical reinforcement learning (HRL) framework [59]. HRL suggests that decision-making takes place at various levels of abstraction, with higher levels determining overall goals and lower levels determining sub-sequences of actions to achieve them. HRL does capture the dynamic aspect of decision-making shown by our model in that it re-evaluates all goals and selects the best current goal at each time point, but there are three major differences. The first is that HRL chooses a single policy and typically pursues it until all actions in a sequence are performed, whereas our model uses a weighted mixture of policies and has no notion of a fixed sequence of actions. The second is that HRL uses a softmax rule to choose a goal, while our model uses both the “goods-” and “action-based” signals. Third, while an HRL framework could be used in place of our model’s motor plan formation DNFs to select an effector and target, our model goes beyond that by interfacing this system with optimal motor control models to control the arm and eyes in real time.
Besides the predictions that have been already validated by experimental studies, the present framework makes novel predictions that have not yet been confirmed by any studies to the best of our knowledge. For instance, we predict that when the brain is faced with multiple equally rewarded goals and also has to select between competing effectors to implement the choice, it first resolves the effector competition (i.e., decides which effector to use) before selecting which goal to pursue. This effect is related to the inhibitory interactions within and between the fields that plan the movements. The motor plan formation fields are characterized by local excitatory and global one-to-many inhibitory connections. However, the interactions between the motor plan formation fields of the competitive effectors are characterized by global inhibitory interactions in which each neuron in one motor plan formation field inhibits all neurons in the other motor plan formation field, resulting in greater net inhibition between rather than within fields. It is because of this architecture that the effector competition is resolved prior to target competition. Although this prediction has not been tested in experimental studies, a recent fMRI study in humans showed that when people have to select whether to use the left or the right hand for reaching to a single target, the effector selection precedes the planning of the reaching movement in the dorsal parietofrontal cortex [37]. We should point out that this prediction is made when both targets have the same expected value and requires the same effort. By varying the value of the targets or the action cost, it is possible that the framework will first select the most “desirable” target, and then it will choose the best effector to implement the choice.
The present framework is able to learn appropriate behavior for tasks in which actions result in different levels of expected reward based on contextual cues or goal-based reward contingencies. This is due to reinforcement learning on connection weights between the goods value field and context cue field and the motor plan formation DNF, and simulates the process by which cortico-basal ganglia networks map contextual cues and reward expectancy onto actions [60]. An unexpected result of our simulations was that although the model was trained on the cued effector task with targets in multiple locations, it could only learn the task when it was presented with a single target in one of the possible locations on each trial, rather than multiple targets at once. This was due to interference between the sensorimotor association and reward contingency learning processes. As the model was rewarded for reaching to a particular target with the correct effector, it began to associate that target with reward in addition to the effector. While various parameter settings which increased the noise levels in the network would promote exploration and thus spatial generalization of the effector cue, the learning proceeded optimally when a single target was presented in one of many possible locations on each trial. This effect has been extensively documented in many studies and is known as “dual-task interference” [61]—when conflicting and interfering stream of information must be processed simultaneously in dual tasks (e.g., talking on the phone while driving), the performance of the tasks is deteriorated substantially.
The computational framework presented in the current study is a systems-level framework aimed to qualitatively model and predict response patterns of neuronal activities in ensembles of neurons, as well as decision and motor behavior in action selection tasks with competing alternatives. It is not intended to serve as a rigorous anatomical model and because of this we avoid making any strict association between the components of the framework (i.e., individual DNFs and control schemes) and particular cortical and subcortical regions. However, it captures many features of neuronal activity recorded from different cortical areas such as the parietal reach region (PRR), area 5, lateral intraparietal area (LIP), premotor cortex, prefrontal cortex (PFC) and orbitofrontal cortex (OFC) in non-human primates that perform reaching and saccadic decision tasks with competing options.
This model can be conceived as a coarse-grained sketch of the fronto-parietal cortical network involved in decisions between competing actions. The “spatial sensory input field” encodes the spatial location of the targets in an egocentric reference frame and mimics the organization of the posterior parietal cortex (PPC). The “context cue field” represents information related to the task context (i.e., which effector to use to acquire the targets). Several neurophysiological studies have reported context-dependent neurons in the lateral PFC (LPFC) in non-human primates [62–64]. These neurons respond differently to the same stimulus when it requires different responses depending on the task context, whereas they are not sensitive to the color or pattern of the cue. The “goods value field” integrates the good values of the alternative options and represents how “desirable” it is to select a policy towards a particular direction without taking into account the sensorimotor contingencies of the choices (i.e., the action cost to implement the policy). The goods value field can be equated to ventromedial PFC (vmPFC) and OFC, which according to neurophysiological and clinical studies, have an important role in computation and integration of good values [3, 4, 65]. The action cost is encoded by the “action cost field”. Although it is not clear how action costs are encoded and computed in the brain, recent findings suggest that the anterior cingulate cortex (ACC) is involved in encoding action costs [11, 66, 67]. However, other studies have shown that ACC neurons also encode good values, such as the payoff of a choice and the probability that a choice will yield a particular outcome [68].
The “motor plan formation field” employs a neuronal population code over 181 potential movement directions and is responsible for planning the individual policies towards these directions. Hence, it could be equated with parts of the premotor cortex and parietal cortex, especially the parietal reach region (PRR) and dorsal area 5, for reaches and lateral intraparietal area (LIP) for saccades, which are involved in planning of hand and eye movements, respectively. One of the novelties of the proposed computational framework is that the motor plan formation field dynamically integrates all the decision variables into a single variable named relative desirability, which describes the contribution of each individual policy to the motor decision. Because of this property, the simulated neuronal activity in this field is modulated by decision variables, such as expected reward, probability outcome and action cost. This is consistent with series of neurophysiological studies, which show that the activity of neurons in LIP and premotor dorsal area (PMd) is modulated by the probability that a particular response will result in reward and the relative reward between competing targets, respectively [12, 32].
To reduce the complexity of the framework, we included only some of the main brain areas that are involved in visuomotor tasks and omitted other relevant cortical regions, such as the primary motor cortex (M1), the somatosensory cortex, the supplementary motor areas, as well as subcortical regions such as the basal ganglia. However, important motor and cognitive processes in action-selection, such as the execution of actions and the learning of sensorimotor associations and reward contingencies are implemented using techniques that mimic neural processes of brain areas that are not included in the framework.
We present a computational framework for dynamically integrating value information from disparate sources in decisions with competing actions. It is based on the concept that decisions between actions are not made in the medial frontal cortex through an abstract representation of values, but instead they are made within the sensorimotor cortex through a continuous competition between potential actions. By combining dynamic neural field theory with stochastic optimal control theory, we provide a principled way to understand how this competition takes place in the cerebral cortex for a variety of visuomotor decision tasks. The framework makes a series of predictions regarding the cell activity in different cortical areas and the choice/motor behavior, suggesting new avenues of research for elucidating the neurobiology of decisions between competing actions.
This section analytically describes the computational framework developed in this study to model the behavioral and neural mechanism underlying decisions between multiple potential actions.
DNF models simulate the activity of a network of neurons over a continuous space with a fixed connectivity pattern of local excitation and surrounding inhibition. Instead of some anatomically defined space, DNF models are defined over the space that is spanned by the parameters of the tasks. They are based on the concept of neuronal population coding, in which the values of task parameters, such as the location of a stimulus or movement parameters, are determined by the distribution of the neuronal activity within a population of neurons. Activation in DNFs is distributed continuously over the space of the encoded parameter and evolves continuously through time under the influence of external inputs, local excitation and lateral inhibition interactions as described by Equation (3):
τ u ˙ x , t = - u x , t + h + S x , t + ∫ w x - x ' f u x ' , t d x ' (3)
where u(x, t) is the local activity of the neuronal population at position x and time t, and u˙(x,t) is the rate of change of the activation field over time scaled by a time constant τ. In the absence of any external input S(x, t), the field converges over time to the resting level h from the current level of activation. The interactions between the neurons in the field are defined through the kernel function w(x − x′), which consists of local excitatory and global inhibitory components, Equation (4):
w ( x - x ' ) = c e x c e - ( x - x ' ) 2 2 σ e x c 2 - c i n h e - ( x - x ' ) 2 2 σ i n h 2 (4)
where cexc, cinh, σexc, σinh describe the amplitude and the width of the excitatory and the inhibitory parts of the kernel function, respectively. In this study we used a narrow Gaussian kernel for the excitatory interactions and a broader Gaussian kernel for the inhibitory interactions, such that neurons with similar tuning curves co-excite one another, whereas neurons with dissimilar tuning curves inhibit one another. The only fields with competitive interactions in this study were the reach and saccade motor plan formation DNFs; the other fields had cexc and cinh set to zero and therefore combined and encoded their inputs without performing further computation.
The kernel function is convolved with a sigmoidal transformation of the field activity f[u(x′, t)], such that only neurons with an activity level that exceeds a threshold participate in the intrafield interactions, Equation (5):
f ( u ( x ) ) = 1 1 + e - β u ( x ) (5)
where β controls the steepness of the sigmoid function.
Besides the context cue field, the fields consist of 181 neurons and their spatial dimension spans the circular space between 0° and 180°. The context cue field consists of 100 neurons, in which half of them respond to a saccade cue and half of them respond to a reach cue. The following sections describe how we integrate dynamic neural field theory with stochastic optimal control theory to develop a computational framework that can explain both neural and behavioral mechanisms underlying a wide variety of visuomotor decision tasks.
Stochastic optimal control theory has proven quite successful at modeling goal-directed movements such as reaching [26], grasping [69] and saccades [70]. It involves solving for a policy π that maps states onto actions ut = π(xt) by minimizing a loss function penalizing costly actions (i.e., effort) and deviations (i.e., accuracy) from the goal. Despite the growing popularity of stochastic optimal control models, the preponderance of them are limited only to single goals. However, real environments present people and animals at any moment with multiple competing options and demands for actions. It is still unclear how to define control policies in the presence of competing options.
In the current study we decompose the complex problem of action selection with competing alternatives into a mixture of optimal control systems that generate policies π′ s to move the effector towards specific directions. The core component of the present framework is the “motor plan formation” DNF that integrates value information from disparate sources and plans the movements to acquire the targets. Each neuron in this DNF is linked with a stochastic optimal control system. When the activity of this neuron exceeds a threshold γ at the current state xt, the controller suggests an optimal policy π* that results in a sequence of actions (ut = π*(xt) = [ut, ut+1, ut+T]) to drive the effector from the current state towards the preferred direction of the neuron from a period of time T. Note that the policy π is related to the preferred direction of the neuron and not to the location of the target. This is the main difference between the optimal control module used in the present framework and classic optimal control studies. However, the mathematical formulation of the optimal control theory requires defining an “end” (i.e., goal) state. In the present framework any “active” controller i generates a sequence of actions to move in the preferred direction of the neuron ϕi for distance r—where r is the distance between the current location of the effector and the location of the stimulus in the field encoded by that neuron. For instance, let’s consider a scenario in which a target is located at a distance r from the current hand location. The control schema with a preferred direction ϕi will suggest an optimal policy π i *, which is given by the minimization of the loss function in Equation (6):
J i ( x t , π i ) = ( x T i - S p i ) T Q T i ( x T i - S p i ) + ∑ t = 1 T i - 1 π i ( x t ) T R π i ( x t ) (6)
where πi(xt) is the policy for time instances t = [t1, t2, …, Ti] to move the effector towards the ϕi direction; Ti is the time-to-arrive at the position pi; pi is goal-position of the effector, i.e., the position that the effector is planned to arrive at the end of the movement and is given as: pi = [rcos(ϕi), rsin(ϕi)]; xTi is the state vector at the end of the movement; S matrix picks out the actual position of the effector and goal-position pi at the end of the movement from the state vector. Finally, QTi and R define the precision- and the control- dependent cost, respectively (see S1 text for more details).
The first term of the loss function in Equation (6) determines the current goal of the controller, which is related to the neuron with preferred direction ϕi i.e., to move the effector at a distance r from the current location, towards the preferred direction ϕi. The second term is the motor command cost (i.e., the action cost) that penalizes the effort required to move the effector towards this direction, for Ti time-steps.
For a DNF of 181 neurons, each of them with a preferred direction between 0 and 180 deg., we construct a simple loss function using an indicator variable ν(xt). This variable encodes the overall relative desirability of each policy with respect to the alternative options—in other words, it categorizes the state space into regions, where following one of the policies is the best option. We can write the loss function as a ν-weighted mixture of individual loss functions Ji’s, Equation (7):
J = ∑ i = 0 N ν i ( x t ) J i ( x t , π i ) (7)
where N is the number of neurons in the DNF (i.e., 181) and νj(xt) is the indicator variable associated with the controller j. When there is no uncertainty as to which policy to follow at a given time—i.e., only the activity of a single neuron exceeds the threshold γ—the ν-weighted loss function in Equation (7) is equivalent to Equation (6) with νi(xt) = 1 for the best current direction, and νj ≠ i(xt) = 0 for the rest of the alternative options. However, when more than one neuron is active, there is uncertainty about which policy to follow at each time and state. In this case, the framework follows a weighted average of the individual policies πmix to move the effector from the current state to a new one, Equation (8):
π m i x ( x t ) = ∑ i = 1 M ν i ( x t ) argmin π i J i ( x t , π i ) = ∑ i = 1 M ν i ( x t ) π i * ( x t ) (8)
where M is the number of neurons that are currently active, π i * ( x t ) is the optimal policy to move in the preferred direction of the ith-neuron from the current state xt, and νi is the relative desirability of the optimal policy π i * that determines the contribution of this policy to the weighted mixture of policies. For notational simplicity, we omit the * sign, and from now on πi(xt) will indicate the optimal policy related to neuron i.
To handle contingencies such as moving targets, perturbations, and the effects of noise, the framework employs a widely used strategy in optimal control theory known as “receding horizon” [71]. According to this strategy, the framework implements only the initial portion of the sequence of actions generated by πmix(xt) for a short period of time k (k = 10 in this study) and then recomputes the individual optimal policies πi(xt+k) from time t + k to t + k + Ti and remixes them. This strategy continues until the effector arrives at the selected target.
We focus next on computing the weights ν′ s by combining information from disparate sources. Recall that each policy πi is associated with a cost Vπi(xt), which represents the action cost—i.e., cost that is expected to accumulate while moving from the current state xt in the direction ϕi under the policy πi. The cost to implement each individual policy is represented by the “action cost” field, such that the higher the activity of the neurons in this field, the higher the action cost to move in the preferred direction of the neurons. Therefore, the output activity of this field ucost is projected to the “motor plan formation” field through one-to-one inhibitory connections in order to suppress the activity of neurons with “costly” preferred directions at a given time and state.
However, in a natural environment the alternative options are usually attached with different values that should be incorporated in the decision process. The present framework uses the “goods value” field to encode the good values (e.g., reward, outcome probability) associated with the available options. In the current study we assume that the goods value field represents the expected reward of the available goals, although it is straightforward to extend the field to encode other goods-related features. The expected reward is represented in a two-dimensional neural field in which each neuron is selective for a particular goal position in an allocentric reference frame, Ureward. For each effector (eye and hand in these simulations), its position is used to convert activity in this two-dimensional field into a one-dimensional neural field encoding an effector-centered representation of the goods value of each goal. The output activity of each effector-centered field, ureward, projects to the corresponding motor plan formation field with one-to-one excitatory connections. It thus excites neurons in each motor plan formation field that drive the effector towards locations with high expected rewards. One of the novelties of the present framework is that the weights of the connections between the goods value field and the motor plan formation field are plastic and are modified using a simple reinforcement learning rule.
Several of the tasks simulated in the current study require the use of a particular effector depending on a visual cue. The model includes a “context cue field” with neurons which respond noisily to the presence of one of the cues in the task. We used a context cue field with 100 neurons, half of which responded to the reach cue and the rest responded to the saccade cue. The output of this field, ucue, projects to each of the motor plan formation fields with one-to-all excitatory connections. Once the weights of these connections have been learned (see “Sensorimotor association learning in effector choice tasks” section), the field excites all neurons in the motor plan formation field corresponding to the cued effector.
Finally, each motor plan formation DNF receives visual input encoding the direction of each goal in effector-centered coordinates. This is provided by the “stimulus input” DNFs whose neurons had Gaussian tuning curves and preferred directions from 0 to 180 degrees. The output of this field, uvis, projected to the corresponding motor plan formation DNF via one-to-one excitatory connections.
The input to the motor plan formation DNF for each effector Smotor(f), is a sum of the outputs of the fields encoding the visual stimulus, cues, estimated cost, and expected reward, corrupted by additive noise ξ which follows a Gaussian distribution:
S m o t o r ( f ) = η v i s u v i s ( f ) + η c u e W c u e ( f ) u c u e - η c o s t u c o s t ( f ) + η r e w a r d u r e w a r d ( f ) + ξ (9)
The parameters ηvis, ηcue, ηcost, and ηreward scale the influence of the input stimulus, cue, cost, and expected reward inputs, respectively. While some studies attempt to find values for these parameters that capture the tradeoff subjects make between cost and reward [72, 73], we set them empirically in order to allow the model to successfully perform the task (see S1 Table, S2 Table, S3 Table and S4 Table in the supporting information for the values of the model parameters used in the current study).
For simulations of tasks using multiple effectors, each effector had its own copy of the cue weights and motor plan formation, cost, and effector-centered goods value fields. Competition between effectors was implemented via massive all-to-all inhibitory connections between their motor plan formation fields:
S m o t o r ( f ) = η v i s u v i s ( f ) + η c u e W c u e ( f ) u c u e + η c o s t u c o s t ( f ) + η r e w a r d u r e w a r d ( f ) - η e f f e c t o r ∑ g ≠ f ∑ u m o t o r ( g ) + ξ (10)
where ηeffector scales the inhibitory influence of the motor plan formation DNFs on each other.
|
10.1371/journal.ppat.1003989 | Modified Vaccinia Virus Ankara Triggers Type I IFN Production in Murine Conventional Dendritic Cells via a cGAS/STING-Mediated Cytosolic DNA-Sensing Pathway | Modified vaccinia virus Ankara (MVA) is an attenuated poxvirus that has been engineered as a vaccine against infectious agents and cancers. Our goal is to understand how MVA modulates innate immunity in dendritic cells (DCs), which can provide insights to vaccine design. In this study, using murine bone marrow-derived dendritic cells, we assessed type I interferon (IFN) gene induction and protein secretion in response to MVA infection. We report that MVA infection elicits the production of type I IFN in murine conventional dendritic cells (cDCs), but not in plasmacytoid dendritic cells (pDCs). Transcription factors IRF3 (IFN regulatory factor 3) and IRF7, and the positive feedback loop mediated by IFNAR1 (IFN alpha/beta receptor 1), are required for the induction. MVA induction of type I IFN is fully dependent on STING (stimulator of IFN genes) and the newly discovered cytosolic DNA sensor cGAS (cyclic guanosine monophosphate-adenosine monophosphate synthase). MVA infection of cDCs triggers phosphorylation of TBK1 (Tank-binding kinase 1) and IRF3, which is abolished in the absence of cGAS and STING. Furthermore, intravenous delivery of MVA induces type I IFN in wild-type mice, but not in mice lacking STING or IRF3. Treatment of cDCs with inhibitors of endosomal and lysosomal acidification or the lysosomal enzyme Cathepsin B attenuated MVA-induced type I IFN production, indicating that lysosomal enzymatic processing of virions is important for MVA sensing. Taken together, our results demonstrate a critical role of the cGAS/STING-mediated cytosolic DNA-sensing pathway for type I IFN induction in cDCs by MVA. We present evidence that vaccinia virulence factors E3 and N1 inhibit the activation of IRF3 and the induction of IFNB gene in MVA-infected cDCs.
| Modified vaccinia virus Ankara (MVA) is an attenuated vaccinia strain with large deletions of the parental genome that render it non-replicative in mammalian cells. MVA is a safe and effective vaccine against both smallpox and monkeypox. MVA has been investigated as a vaccine vector for infectious diseases and cancers. Dendritic cells (DCs) play important roles in innate and adaptive immunity. A better understanding of how MVA is detected by innate immune sensors in DCs would guide the development of more effective MVA-based vaccines. We report our findings that MVA infection induces the production of type I interferon (IFN) in conventional dendritic cells via a cytosolic DNA-sensing pathway mediated by the newly discovered DNA sensor cGAS, its adaptor STING, and transcription factors IRF3 and IRF7. By contrast, wild-type vaccinia virus fails to activate this pathway. Furthermore, we show that vaccinia virulence factors E3 and N1 play inhibitory roles in the cytosolic DNA-sensing pathway.
| Poxviruses are large cytoplasmic DNA viruses that cause human and veterinary diseases. Variola virus (the causative agent of smallpox) and monkeypox virus are important human pathogens [1]–[3]. Modified vaccinia virus Ankara (MVA) is an attenuated vaccinia virus that was developed through serial passaging in chicken embryonic fibroblasts. MVA has a 31-kb deletion of the parental vaccinia genome and was used successfully as a vaccine during the WHO-sponsored smallpox eradication campaign [4]–[6]. MVA has been investigated intensively as a vaccine vector against HIV, tuberculosis and malaria, as well as cancers [7]–[12].
Dendritic cells are the sentinels of the immune system. They can be mainly classified into two subtypes: conventional dendritic cells (cDCs) and plasmacytoid dendritic cells (pDCs). cDCs are professional antigen-presenting cells that can be activated via Toll-like receptors (TLRs), RIG-I-like receptors, and cytosolic DNA-sensing pathways [13], [14]. pDCs are potent type I interferon (IFN) producing cells that sense viral infections via TLR7, TLR8, and TLR9, and their adaptor MyD88 [15]. Delineating the innate immune responses of dendritic cells to MVA infection could guide vaccine design using MVA-based vectors.
We reported previously that wild-type vaccinia (WT VAC) infection of epidermal cDCs fails to induce the production of type I IFN and attenuates innate immune responses to lipopolysaccharide (LPS) or poly(I∶C) [16]. Infection of human or murine pDCs with live WT VAC also fails to induce type I IFN production, whereas infection with heat-inactivated vaccinia (Heat-VAC, by incubating at 55°C for 1 h) induces TLR7/MyD88-dependent type I IFN production [17], [18]. These results indicate that WT VAC produces inhibitor(s) to block poxviral sensing in cDCs and pDCs.
MVA has deletions or truncations of several intracellular immunomodulatory genes including K1L, N1L, and A52R, which have been implicated in regulating innate immune responses, especially the NF-κB signaling pathway [19]–[24]. Vaccinia N1 is a 14-kDa cytosolic protein that contributes to virulence in murine infection models [25], [26]. In addition to its role in inhibiting the NF-κB pathway, N1 also attenuates IRF3 activation [21]. On the other hand, MVA retains the E3L gene encoding a bifunctional Z-DNA/dsRNA binding protein, a key vaccinia virulence factor [27]–[35]. It has been shown that MVA infection of human monocyte-derived dendritic cells causes DC activation [36]. Waibler et al. [37] reported that MVA infection of murine Flt3L-DC triggered a TLR-independent type I IFN response. In addition, MVA infection of human macrophages triggers type I IFN and pro-inflammatory cytokines and chemokines via a TLR2/TLR6/MyD88 and MDA5/MAVS-dependent pathways [38].
In this paper, we report that MVA infection of murine pDCs fails to induce type I IFN induction. By contrast, MVA infection of cDCs triggers type I IFN production. This induction in cDCs is dependent on STING (stimulator of interferon genes; an adaptor for the cytosolic DNA-sensing pathway), cGAS (cyclic GMP-AMP synthase; a recently discovered cytosolic DNA sensor), transcription factors IRF3 and IRF7, and the IFN receptor IFNAR1. We find that IFN production in MVA-infected cDCs is independent of RNA-sensing pathways mediated by MDA5, MAVS, TLR3, or TRIF, and is modestly affected by the absence of the TLR9/MyD88 endosomal DNA-sensing pathway. We present evidence that vaccinia E3 and N1 proteins play inhibitory roles in the cGAS/STING/IRF3-dependent cytosolic DNA-sensing pathway. Type I IFN serves as an important link between innate and adaptive immunity. Therefore, the identification of the cytosolic DNA-sensing pathway mediated by cGAS/STING/IRF3 for type I IFN induction by MVA in cDCs has implications for MVA-based vaccine design to improve its immunogenicity and efficacy.
Waibler et al. [37] reported that MVA induces TLR-independent type I IFN responses in murine bone marrow-derived dendritic cells (BMDCs). Consistent with their report, we also observed that MVA infection of GM-CSF-cultured BMDCs or Flt3L-cultured BMDCs induces type I IFN secretion (data not shown). GM-CSF (Granulocyte/macrophage colony stimulating factor) is important for the development of CD11c+B220−PDCA1− cDCs. Flt3L (fms-like tyrosine kinase-3 ligand) is critical for the commitment and differentiation of hematopoietic progenitors to both pDC and cDC populations [39], [40]. To determine which DC subtype is responsible for the production of type I IFN in response to MVA infection, we enriched pDCs and cDCs to 98% purity from Flt3L-cultured BMDCs (Flt3L-DCs) using FACS. 2×105 pDCs and 1×106 cDCs were stimulated with CpG or infected with either WT VAC or MVA at a multiplicity of infection (MOI) of 10. Supernatants were collected at 22 h post infection. The levels of IFN-α and IFN-β were determined by ELISA. We found that MVA infection induced the secretion of IFN-α/β from cDCs, but not from pDCs (Figure 1A). Treatment of cells with the TLR9 agonist CpG induced IFN-α and β production from pDCs, but only IFN-β from cDCs (Figure 1A). WT VAC infection failed to induce type I IFN production in either pDCs or cDCs (Figure 1A).
To assess the time course of induction of type I IFN secretion by MVA-infected cDCs, we performed kinetic analysis using GM-CSF-cultured BMDCs (cDCs), which demonstrated that IFN-α and IFN-β proteins were detected by ELISA at 8 h post-infection with MVA and continued to accumulate up to 24 h post-infection (Figure 1B).
To test whether MVA induces type I IFN production in cDCs in a dose-dependent manner, we infected cDCs with MVA at increasing MOIs from 0.25 to 10. We found that MVA induced IFN-α and IFN-β production even at a low MOI of 0.25. IFN production was increased with higher doses of MVA, which reached the highest level at a MOI of 5 and 10 (Figure 1C). We used MOI of 10 for MVA infection in the rest of the in vitro infection experiments reported in this paper.
To test whether WT VAC or MVA infection of cDCs affects type I IFN gene expression, we performed quantitative real-time PCR analysis of RNA isolated from GM-CSF-cultured cDCs infected with WT VAC or MVA at 6 h post-infection. Mock-infection controls were also included. We observed that MVA infection of cDCs increased IFNA4 and IFNB mRNA levels by 6-fold and 105-fold, respectively, when compared with untreated cells. By contrast, infection with WT VAC increased IFNA4 and IFNB mRNA levels by 2-fold and 6-fold, respectively (Figure 1D). These results indicate that MVA is a stronger inducer of IFNA4 and IFNB gene expression than WT VAC (p<0.001).
We hypothesize that the differences we observed in IFN gene expression between MVA and WT VAC might be related to their abilities to activate transcription factor IRF3. IRF3 is a cytoplasmic protein expressed constitutively in many cell types. Upon virus infection, IRF3 is phosphorylated at multiple serine and threonine residues near the C-terminus. Phosphorylated IRF3 then translocates to the nucleus and activates IFN gene transcription [41]. We performed Western blot analysis of MVA- or WT VAC-infected cDCs, and found that MVA infection triggered IRF3 phosphorylation (which peaks at 4 h post infection), whereas WT VAC infection fails to do so. We also observed that MVA infection resulted in much higher levels of phosphorylation of TBK1 than WT VAC (Figure 1E), indicating that WT VAC might encode inhibitor(s) that interfere with phosphorylation of TBK1 and IRF3.
Similar to IRF3, the transcription factor IRF7 is another key regulator of type I IFN induction and is critical for host defense against virus infections [42], [43]. Using cDCs generated from IRF3−/− and age-matched WT control mice, we found that MVA-induced IFN-α/β secretion was abolished in IRF3 deficient cDCs (Figure 2A). IRF7−/− cells fail to produce IFN-α in response to MVA infection. IFN-β induction was reduced by 57% in MVA-infected IRF7−/− cells (Figure 2B). To assess whether the type I IFN positive feedback loop mediated by IFNAR1 is required for the induction of IFN, we infected IFNAR1−/− cDCs and WT controls with MVA at a MOI of 10. We found that IFN-α induction by MVA was abolished in IFNAR1−/− cells, whereas IFN-β induction by MVA was reduced by 45% in IFNAR1−/− cells compared with WT controls (Figure 2C). These results indicate that: (i) IRF3 is the critical transcription factor for MVA-induced type I IFN production, and (ii) IRF7 and IFNAR1 play roles in amplifying type I IFN signaling induced by MVA infection.
To test whether TLR7, TLR9 and MyD88 are involved in MVA induction of type I IFN in murine cDCs, we generated cDCs from TLR7−/−, TLR9−/−, and MyD88−/− mice or their age-matched WT controls. Cells were either treated with TLR9 agonist CpG or infected with MVA at a MOI of 10. Control experiments show that induction of IFN-β by CpG was abolished in MyD88−/− and TLR9−/− murine cDCs, but was not affected in TLR7−/− cDCs (Figure 3A, 3B and 3C). By contrast, MVA-induced production of IFN-α and IFN-β was reduced by 32% and 43% in MyD88−/− cDCs, and by 23% and 37% in TLR9−/− cDCs, respectively, but was not affected in TLR7−/− cDCs (Figure 3A, 3B and 3C). These results indicate that the endosomal TLR9/MyD88 pathway plays a minor role in MVA-sensing in cDCs.
We assessed the contributions of other known nucleic acid-sensing pathways to the induction of type I IFN by MVA. MAVS (mitochondrial antiviral signaling protein) is an adaptor for the cytosolic RNA sensors RIG-I and MDA5 [44]. TRIF (TIR-domain-containing adapter-inducing interferon-β) is an adaptor for the endosomal TLR3 that senses extracellular dsRNA [45], [46] and an adaptor for cytosolic RNA sensors [47]. We found that TLR3, TRIF, MDA5, and MAVS deficient cells secreted similar amounts of type I IFN compared to cDCs from age-matched WT control mice in response to MVA infection (Figure S1A, B, C and D). Taken together, these results indicate MVA-induced type I IFN production by infected cDCs does not require previously known RNA-sensing mechanisms.
STING is an endoplasmic reticulum-associated protein essential for type I IFN induction in response to intracellular DNA or DNA pathogens including bacteria and DNA viruses [48]–[53]. STING is also important for the development of CD8+ T cell responses after vaccination with vaccinia virus expressing ovalbumin [50]. To test whether STING is required for type I IFN induction in cDCs by MVA, we generated cDCs from the N-ethyl-N-nitrosourea (ENU)-induced Goldenticket (Gt) mutant mice (StingGt/Gt) harboring a single nucleotide variant of Sting resulting in a functionally null allele [54]. cDCs from age-matched WT mice were used as a control. Cells were either infected with MVA at a MOI of 10 or treated with lipopolysaccharide (LPS). MVA induction of IFN-α/β was abolished in StingGt/Gt cells, whereas LPS-induced IFN-α/β production was not affected (Figure 4A). Induction of IFNA4 mRNA by MVA was reduced from 6-fold in WT cells to 2-fold in StingGt/Gt cells, whereas induction of IFNB mRNA by MVA was reduced from 133 fold in WT cells to 14 fold in StingGt/Gt cells (Figure 4B). Western blot analysis demonstrated that MVA-induced IRF3 phosphorylation peaked at 4 and 6 h post infection in WT cDCs and was absent in StingGt/Gt cDCs (Figure 4C). Together, these results demonstrate that STING is essential for MVA-induced type I IFN production and IRF3 phosphorylation in cDCs.
The kinetics of type I IFN induction by MVA have been reported by Waibler et al. [55]. C57B/6 mice were infected with 1×107 pfu of MVA through intravenous (i.v.) inoculation and serum was collected at 6, 12, 18, and 24 h. They found that MVA induced highest IFN-α production at 6 h post infection. We performed i.v. inoculation of purified MVA (2×107 pfu) via tail vein injection of Sting+/+, StingGt/Gt, and IRF3−/− mice. Serum was collected at 6 h post-infection. We found that MVA infection of Sting+/+ mice induces IFN-α and IFN-β production to the levels of 798 pg/ml and 1017 pg/ml, respectively, which was abolished in StingGt/Gt, and IRF3−/− mice (Figure 4D). These results indicate that MVA-induced type I IFN production in vivo is also dependent on STING and transcription factor IRF3.
The STING/IRF3 pathway can be activated by cyclic GMP-AMP (cGAMP), a mammalian second messenger produced by cyclic GMP-AMP synthase (cGAS) in response to transfected DNA or DNA virus infection [56], [57]. Moreover, heat-inactivated vaccinia infection of THP1 (human acute monocytic leukemia cells) triggers the production of cGAMP. cGAS-deficient mice are more susceptible to herpes simplex virus 1 (HSV-1) infection than WT mice [58]. We generated cDCs from cGAS-deficient mice and WT controls. We found that MVA-induced IFN-α/β production was abolished in cGAS−/− cells (Figure 5A). Induction of IFNA4 and IFNB mRNA by MVA was also diminished in cGAS−/− cells compared with WT cells (Figure 5B). Western blot analysis demonstrated that MVA-induced phosphorylation of TBK1 and IRF3 was absent in cGAS−/− cells (Figure 5C). These results established that cGAS is the critical cytosolic DNA sensor for MVA.
Using quantitative real-time PCR analysis, we found that WT vaccinia and MVA DNA genomic copies were increased by 392-fold and 111-fold, respectively, at 24 h post-infection of BSC40 cells, whereas DNA copies of the WT vaccinia and MVA were increased by 86-fold and 46-fold, respectively, at 24 h post infection of BMDCs (Figure S2). This result raised the question of whether viral DNA replication is required for type I IFN induction by MVA in cDCs. Phosphonoacetate (PAA) inhibits vaccinia DNA polymerase [59], [60]. Using real-time PCR analysis, we found that treatment of cDCs with PAA prevented MVA DNA replication (Figure 6A). Western blot analysis demonstrated that PAA treatment did not affect MVA-induced IRF3 phosphorylation (Figure 6B), indicating that viral DNA replication is not required for MVA-induced activation of STING/IRF3 pathway in cDCs.
To test whether lysosomal processing of virions might contribute to host sensing of MVA, we took both pharmacological and genetic approaches. Chloroquine and bafilomycin A1 block endosomal/lysosomal acidification and thereby may prevent virion processing in the late endosomes and lysosomes [61]. CA-074-Me is a specific inhibitor of cathepsin B, a lysosomal cysteine protease [62]. We infected cDCs with MVA at a MOI of 10. To avoid drug effects on viral entry, cells were treated at 1 h post infection with chloroquine at a final concentration of 50 µM, or bafilomycin A1 at 100 nM, or CA-074-Me at 10 µM. Supernatants were collected at 22 h post infection and measured for the concentrations of IFN-α/β by ELISA. We observed inhibition of IFN-α/β production in the presence of chloroquine, bafilomycin A1, and CA-074-Me (Figure 7A). These results suggest that MVA-mediated type I IFN induction requires endosomal/lysosomal processing of vaccinia virions. To test the role of cathepsin B in host sensing of MVA, we took advantage of the cathepsin B knock-out (KO) mice generated by targeted deletion of the ctsb gene [63]. We infected cDCs from cathepsin B−/− (CathB−/−) mice and WT controls with MVA at a MOI of 10. Supernatants were collected at 22 h post infection. We found that IFN-α and IFN-β levels in the supernatant were reduced by 76% and 60%, respectively, in CathB−/− cDCs compared to WT cells, indicating that lysosomal enzyme cathepsin B activity contributes to MVA sensing in cDCs (Figure 7B). It is possible that other cathepsins might also participate in the processing of virions in the endosomal/lysosomal compartments because type I IFN induction by MVA is not abolished in cathepsin B−/− cDCs.
E3 is a key virulence factor that attenuates various innate immune responses, including type I IFN induction. MVA retains the E3L gene. Western blot analysis showed that E3 protein was produced in WT VAC and MVA-infected BMDCs, but not in MVAΔE3L-infected cells (Figure S3). To test whether E3 plays an inhibitory role in MVA sensing in cDCs, we compared the induction of type I IFN gene expression between MVA and MVAΔE3L-infected cDCs. We found that infection with MVAΔE3L induced higher levels of IFNA4 and IFNB mRNAs than MVA (Figure 8A) (p<0.001). This induction was abolished in cells lacking transcription factor IRF3 (Figure 8B). Furthermore, Western blot analysis demonstrated that MVAΔE3L infection induced higher level of phospho-IRF3 than MVA at both 4 and 8 h post infection (Figure 8C). These results suggest that E3 dampens innate immune sensing of MVA and that removing E3 from MVA results in enhanced activation of type I IFN gene expression.
Vaccinia virulence factor N1L gene is highly conserved among orthopoxviruses. The MVA orthologue of N1L contains a frameshift mutation resulting in a protein 4 amino acid (aa) shortened compared with N1 from WT VAC, but with a completely different C-terminal 23 aa segment [64], [65]. To test whether N1 alteration contributes to MVA-induction of the IFN pathway, we generated recombinant MVA that expresses the vaccinia N1L gene under its own promoter. Using Western blot analysis, we demonstrated that, unlike MVA, the recombinant virus MVA-N1L expresses vaccinia N1, which could be detected by a polyclonal antibody that recognizes the C-terminus of N1 (Figure S4). To investigate the role of N1 in modulating the IFN pathway, we infected cDCs with either MVA or MVA-N1L at a MOI of 10. Cells were harvested at 6 h post-infection and prepared for real-time PCR analysis for type I IFN mRNA expression. We found that the levels of IFNA4 and IFNB mRNAs in MVA-N1L-infected cDCs were 62% and 77% lower, respectively, compared with MVA-infected cells (Figure 9A) (p<0.001). Western blot analysis demonstrated that MVA-N1L infection triggered reduced levels of activation of p-TBK1 and p-IRF3 at 4 and 8 h post-infection when compared with MVA (Figure 9B). ELISA analysis of supernatants collected at 22 h post infection showed that MVA-N1-induced IFN-α and IFN-β levels were 39% and 30% lower, respectively, compared with those induced by MVA (Figure 9C) (p<0.01). Taken together, these data indicate that vaccinia N1 inhibits the cytosolic DNA-sensing pathway, which results in reduced type I IFN gene expression and secretion from cDCs.
Our study provides insights into how MVA infection is sensed by murine cDCs, a subset of DCs that are important for innate and adaptive antiviral immunity. We found that MVA induces type I IFN production in murine cDCs but not in pDCs, whereas WT vaccinia infection fails to induce type I IFN in either cDCs or pDCs. Using cDCs from knockout mice, we demonstrated that MVA-induced type I IFN induction is dependent on the newly discovered cytosolic DNA sensor cGAS and its adaptor, STING. The transcription factors IRF3 and IRF7, and the type I IFN feedback loop mediated by IFNAR1 are critical for the induction. By contrast, the TLR9/MyD88 pathway plays a relatively minor role in the sensing of MVA in cDCs. Treatment with inhibitors of either endosomal/lysosomal acidification or lysosomal enzyme activity attenuated MVA-induced type I IFN production, indicating that lysosomal enzymatic processing of vaccinia virions is important for MVA sensing in cDCs. Finally, we showed that vaccinia virulence factors E3 and N1 play inhibitory roles in IFN-β induction.
Based on these results, we hypothesize that following viral entry, possibly through macropinocytosis, some of the MVA virions are processed in the late endosomal and lysosomal compartments. As a result, viral DNAs are released into the cytoplasm and then sensed by the cytosolic DNA sensor cGAS, which leads to production of cGAMP [56], [58], [66]–[69]. cGAMP then acts as a ligand for STING to trigger the recruitment and activation of TBK1, which results in the phosphorylation of IRF3, as well as the subsequent induction of type I IFN gene expression [57], [70]–[72]. We surmise that viral DNA is the primary stimulator for type I IFN induction in response to MVA infection. We conclude that neither the endosomal dsRNA-sensing pathway (mediated by TLR3/TRIF) nor the cytosolic dsRNA-sensing pathway (mediated by MDA5/MAVS) is involved in MVA-induced type I IFN production.
STING has been identified as a critical adaptor for the cytosolic DNA sensor(s) [49]–[53]. It also acts as a direct sensor for cyclic dinucleotides, a type of bacterial second messenger [73]. STING is important for host defense against DNA viruses [50], [74]. Besides cGAS, several other cytosolic DNA sensors utilizing STING as an adaptor, including IFI16 and DDX41, have been identified [75], [76]. We found that shRNA knock-down of DDX41 in macrophage cell line failed to reduce MVA-induced phosphorylation of TBK1 and IRF3 (data not shown). We were unable to detect any difference of MVA-induced type I IFN production in IFI16-deficient cDCs compared with WT cells (data not shown).
Transcription factors IRF3 and IRF7 play important roles in the induction of type I IFN genes and the activation of antiviral viral immunity [77]. Our results indicate that IRF3 is required for the induction of IFN-α/β secretion by MVA-infected cDCs, whereas IRF7 also contributes to the initial induction of IFN-β production and the subsequent amplification of the type I IFN signal and induction of IFN-α production. Daffis et al. reported that West Nile Virus (WNV)-induced IFNB gene expression and that IFN-β production in cDCs is independent of IRF3 and IRF7 [78]. This could be due to different upstream sensing mechanisms for MVA and WNV in cDCs.
Vaccinia E3, a virulence factor, inhibits multiple signaling pathways that are important in host antiviral responses. E3 is a 190-aa protein that is composed of two distinct domains, the N-terminal Z-DNA binding domain (ZBD) and the C-terminal dsRNA binding domain (dsRBD). MVA retains the E3L gene, which is expressed in MVA-infected cells. MVAΔE3L infection of cDCs triggers higher gene expression of type I IFN than MVA, but no significant differences in type I IFN secretion was observed between MVA and MVAΔE3L-infected cDCs (Dai and Deng, unpublished). Others have shown that MVAΔE3L induces a higher level of apoptosis than MVA in HeLa cells and MEFs [79]–[81]. Consistent with published results, we found that MVAΔE3L infection of cDCs induces more apoptosis of infected cells than MVA.
How vaccinia E3 inhibits the cytosolic DNA-sensing pathway is currently unclear. It might prevent the activation of cGAS via its ZBD. Alternatively, it may block the activation of its downstream effector(s). We have previously shown that E3 ZBD plays an inhibitory role in TLR9/MyD88/IRF7-mediated myxoma virus-sensing in pDCs [17], whereas E3 dsRBD attenuated the cytosolic dsRNA-sensing pathway mediated by MAVS/IRF3 [35]. It has been recently shown that ZBD of E3 antagonizes PKR in primary MEF [82]. The present study illustrates another inhibitory function of E3 ZBD in the cytosolic DNA-sensing pathway. In fact, E3LΔ83N, a mutant vaccinia virus lacking the ZBD, is attenuated in intranasal and intracranial infection models [27], [31], further supporting the in vivo function of the ZBD domain of E3.
Vaccinia N1, another virulence factor, is altered in MVA due to a frameshift mutation, which results in a shortened polypeptide with a complete different 23-aa at the C-terminus. To test whether N1 inhibits the STING-dependent cytosolic DNA-sensing pathway, we generated a recombinant MVA that expresses the vaccinia N1L gene under its natural promoter. We found that infection with MVA-N1L virus induces lower levels of IFNA4 and IFNB mRNAs, as well as lower levels of p-TBK1 and p-IRF3. N1 has been shown to be able to interact with TBK1 [21]. We conclude that vaccinia N1 is an inhibitor of the STING-dependent type I IFN pathway, possibly through its interaction with TBK1.
Delaloye et al. [38] reported that MVA infection of human THP-1 cells induced type I IFN that was dependent on the cytosolic dsRNA-sensing pathway mediated by MDA5 and IPS-1/MAVS using shRNA knockdown. However, the cytosolic DNA-sensing pathway was not assessed in that paper. Using BMDCs from MDA5 or MAVS KO mice, we found that the cytosolic dsRNA- sensing pathway mediated by MDA5/MAVS was not required for type I IFN induction by MVA in these cells. This might be due to the intrinsic differences between the human and murine system and different experimental approaches. Future studies are needed to elucidate the induction of type I IFN in human DCs, which has implications for the use of MVA in vaccinations.
Waibler et al. [37] reported that MVA infection of bone marrow-derived Flt3-L cultured DCs or bone marrow-derived GM-CSF cultured DCs induced type I IFN induction, which was primarily triggered by non-TLR sensors and was independent of viral replication. At the time of the investigation, cytosolic DNA sensors had not been discovered. In our present study, we took advantage of FACS sorting to generate highly purified pDC and cDCs populations that allow us to demonstrate that MVA-induction of type I IFN mainly occurs in cDCs, but not in pDCs. Using STING-deficient BMDCs, we were able to show that STING plays a critical role in sensing MVA infection in cDCs.
Our present study and previous reports on host innate immune sensing of MVA and myxoma virus reveal some key differences between the two poxviruses. Whereas MVA is a strong inducer of type I IFN in cDCs via the activation of the cytosolic DNA-sensing pathway mediated by cGAS/STING/IRF3, it fails to induce IFN in pDCs. By contrast, myxoma infection of pDCs potently induces type I IFN through the endosomal DNA-sensing pathway mediated by TLR9/MyD88/IRF7 [17], [18], whereas myxoma infection of cDCs fails to activate IRF3 or to induce type I IFN (Dai and Deng, unpublished). Although MVA has been studied intensively as a vaccine vector, the studies have not been conducted with myxoma virus backbone and should be done in the future.
In summary, we present data demonstrating the critical role of the cGAS/STING pathway in mediating MVA-induced type I IFN production in murine cDCs. We postulate that upon viral entry, some of the MVA virions are processed in the endosomal/lysosomal compartment, and viral DNAs are detected by the cytosolic DNA sensor(s), leading to the assembly of the STING complex and activation of transcription factors IRF3 and IRF7. We provide evidence that vaccinia E3 and N1 play inhibitory roles in this pathway. Our future studies will focus on the mechanisms by which E3 and N1 attenuate the cGAS/STING-mediated cytosolic DNA-sensing pathway, as well as the role of cGAS and STING in host defense against vaccinia infection and MVA-induced adaptive immunity in vivo. The results of these studies will provide insights into improved poxvirus-based vaccines for clinical applications for both infectious diseases and cancers.
Mice were maintained in the animal facility at the Sloan Kettering Institute. All procedures were performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institute of Health.
The WR strain of vaccinia virus was propagated and virus titers were determined on BSC40 (African green monkey kidney cells) monolayers at 37°C. MVA and MVAΔE3L viruses were kindly provided by Gerd Sutter (University of Munich), and propagated in BHK-21 (baby hamster kidney) cells. MVA-N1L recombinant virus was generated by using a two-step, red-mediated recombination system [83]. The MVA orthologue of N1L containing a truncated and mutated C terminus of 71 nucleotides (due to a frameshift) was reconstituted by orthotopic insertion of 83 nucleotides of VACV N1L, which resulted in a full-length N1L gene. MVA-N1L was propagated on chicken embryonic fibroblasts (CEFs). All of the viruses were purified through a 36% sucrose cushion. BSC40 cells were maintained in Dulbecco's modified Eagle's medium (DMEM) supplemented with 5% fetal bovine serum (FBS). BHK-21 and RK13 cells were cultured in DMEM containing 10% FBS, 0.1 mM nonessential amino acids, and 50 µg/ml gentamycin. All cells were grown at 37°C in a 5% CO2 incubator.
Female C57B/6 mice between 6 and 10 weeks of age were purchased from the Jackson Laboratory and were used for the preparation of bone marrow-derived dendritic cells. These mice were maintained in the animal facility at the Sloan Kettering Institute. All procedures were performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institute of Health. The protocol was approved by the Committee on the Ethics of Animal Experiments of Sloan-Kettering Cancer Institute. cGAS−/−, MDA5−/−, MAVS−/−, IRF3−/−, IRF7−/−, TLR3−/−, MyD88−/−, TLR9−/−, TLR7−/−, TRIF−/− (TRIFLPS2/LPS2), Cathepsin B−/−, and STINGGt/Gt mice were generated in the laboratories of Drs. Zhijian Chen (University of Texas Southwestern Medical Center), Marco Colona (Washington University), Tadatsugu Taniguchi (University of Tokyo), Richard Flavell (Yale University), Shizuro Akira (Osaka University), Bruce Beutler (Scripps Research Institute), Christoph Peters (University of Freiburg), and Russell Vance (University of California, Berkeley). IFNAR1−/− mice were provided by Dr. Eric Pamer (Sloan Kettering Institute); the mice were purchased from B&K Universal and were backcrossed with C57BL/6 mice for more than five generations.
The bone marrow cells from the tibia and femur of mice were collected by first removing muscles from the bones, and then flushing the cells out using 0.5 cc U-100 insulin syringes (Becton Dickinson) with RPMI with 10% FCS. After centrifugation, cells were re-suspended in ACK Lysing Buffer (Lonza) for red blood cells lysis by incubating the cells on ice for 1–3 min. Cells were then collected, re-suspended in fresh medium, and filtered through a 40-µm cell strainer (BD Biosciences). The number of cells was counted. For the generation of Flt3L-BMDCs, the bone marrow cells (5 million cells in each well of 6-well plates) were cultured in complete medium (CM) in the presence of Flt3L (100 ng/ml; R & D systems) for 7–9 days. Cells were fed every 2–3 days by replacing 50% of the old medium with fresh medium. FACS was used to obtain highly purified pDCs and cDCs. pDCs were gated as CD11C+B220+PDCA-1+ and cDCs were gated as CD11C+B220−PDCA-1− as described [17]. For the generation of GM-CSF-BMDCs, the bone marrow cells (5 million cells in each 10 cm cell culture dish) were cultured in CM in the presence of GM-CSF (30 ng/ml, produced by the Monoclonal Antibody Core facility at the Sloan Kettering Institute) for 10–12 days. CM is RPMI 1640 medium supplemented with 10% heat-inactivated fetal bovine serum (FBS), 100 U/ml penicillin, 100 µg/ml streptomycin, 0.1 mM essential and nonessential amino acids, 2 mM L-glutamine, 1 mM sodium pyruvate, and 10 mM HEPES buffer. Cells were fed every 2 days by replacing 50% of the old medium with fresh medium and re-plated every 3–4 days to remove adherent cells. Only non-adherent cells were used for experiments.
RNA was extracted from whole-cell lystates with a RNeasy Mini kit (Qiagen) and was reverse transcribed with a First Strand cDNA synthesis kit (Fermentas). Quantitative real-time PCR was performed in triplicate with SYBR Green PCR Mater Mix and Applied Biosystems 7500 Real-time PCR Instrument (Life Technologies) using gene-specific primers. Relative expression was normalized to the levels of glyceraldehyde-3-phosphate dehydrogenase (GADPH).
Cells were infected with various viruses at a MOI of 10 for 1 h or mock infected. The inoculum was removed and the cells were washed with PBS twice and incubated with fresh medium. Supernatants were collected at various times post infection. Cytokine levels were measured by using enzyme-linked immunosorbent essay (ELISA) kits for IFN-α/β (PBL Biomedical Laboratories) and CCL5 (R & D systems).
BMDCs (1×106) were infected with MVA at a MOI of 10. At various times post-infection, the medium was removed and cells were collected. Whole-cell lysates were prepared at 1, 2, 4, 6, and 9 h post-infection. Equal amounts of proteins were subjected to sodium dodecyl sulfate-polyacrylamide gel electrophoresis and the polypeptides were transferred to a nitrocellulose membrane. Phosphorylation of IRF3 was determined using a rabbit polyclonal antibody specific for phosphoserine-396 of IRF3 (Cell Signaling). The level of IRF3 was determined by using a rabbit polyclonal antibody against IRF3. Anti-phospho-TBK1, anti-TBK1 and anti-STING antibodies were purchased from Cell Signaling. Vaccinia E3 protein level was determined by using anti-E3 monoclonal antibody (MAb 3015B2) kindly provided by Dr. Stuart N. Isaacs (University of Pennsylvania) [84]. Vaccinia N1 protein level was assessed by using mouse monoclonal anti-N1 antibody (7E5) kindly provided by Dr. Michael Way (Cancer Research UK). Vaccinia H3 protein level was assessed by using rabbit polyclonal anti-H3 antibody (a kind gift from Bernard Moss, NIH). Anti-glyceraldehyde-3-phosphate dehydrogenase (GADPH) or anti-β-actin antibodies were used as loading controls.
Viral DNA was purified with the DNeasy Mini Kit (Qiagen) according to the manufacture's protocol. Real-time PCR was performed with the Applied Biosystems 7500 Real-Time PCR Instrument (Life Technologies). The primers and TaqMan probe used in the quantitative PCR assay are specific for the vaccinia ribonucleotide reductase l4L gene. The sequences and PCR condition were described by Liu et.al. [85]. A standard curve was established from cloned DNA fragment of the vaccinia l4L gene. Corresponding CT (cycle threshold) values obtained by the real-time PCR methods were plotted on the standard curve to calculate the viral DNA copy number.
The commercial sources for reagents were as follows: CpG oligodeoxynucleotide ODN2216 (Invivogen); chloroquine, bafilomycin A1 and CA-074-Me (Sigma-Aldrich); Flt3L (R & D systems); anti-mouse CD11c APC (BD Pharmingen), anti-mouse B220 APC-Cy7 and anti-mouse PDCA-1 PE (Milteny Biotec); and Phosphonoacetate (Sigma).
Student's two-tailed t-test was used for each pairwise comparison. The p values deemed significant are indicated in the figures as follows: *, p<0.05; **, p<0.01; ***, p<0.001.
|
10.1371/journal.pntd.0007467 | On lifestyle trends, health and mosquitoes: Formulating welfare levels for control of the Asian tiger mosquito in Greece | The expansion of urban ecosystems and climate change, both outcomes of massive lifestyle changes, contribute to a series of side effects such as environmental deterioration, spread of diseases, increased greenhouse gas emissions and introduction of invasive species. In the case of the Athens metropolitan area, an invasive mosquito species—the Asian tiger mosquito (Aedes albopictus)–has spread widely in the last decade. This spread is favoured within urban environments and is also affected by changing climatic trends. The Asian tiger mosquito is accompanied by risks of mosquito-borne diseases, greater nuisance levels, and increased expenses incurring for its confrontation. The main aims of this paper are (i) to estimate the various costs associated with the control of this invasive species, as well as its health and nuisance impacts, (ii) to evaluate the level of citizens’ well-being from averting these impacts and (iii) to record citizens’ and experts’ perceptions regarding alternative control measures. Evidence shows that experts tend to place a high value on mosquito control when associated with serious health risks, while citizens are more sensitive and concerned about the environmental impacts of control methods. The synthesis of results produced by the current study could act as a preliminary guide for the estimation of societal welfare from the confrontation of similar problems in the context of a complex ecosystem.
| This paper is based on several years’ collaboration among researchers from various disciplines, key health policy makers and stakeholders in an attempt to evaluate the economic dimensions related to the presence of the Asian Tiger Mosquito (Aedes albopictus) and the challenges of tackling mosquito-borne disease outbreaks in Greece and Southern Europe. Similar studies have been conducted and continue to be published in Europe and the USA examining the socioeconomic benefit from the implementation of relevant control and prevention strategies. These studies conclude that there are significant benefits related both to the reduction of nuisance levels and the reduction of the health risks posed by various mosquito species. In our case, the application of an updated economic analysis on the effectiveness of relevant public control and prevention programs provides essential information for public health decision-making, bearing in mind the significant restructuring of the public sector and the fiscal crisis apparent in the European South.
| Recent reports highlight the impacts and risks to human and natural systems linked to global warming of 1.5°C compared to temperatures in the pre-industrial period [1]. The implications of rising temperatures for human health around the globe include changes in disease vector survival and pathogen development, and the emerging new sanitary and environmental risks are directly related to various socioeconomic impacts. Recent studies indicate that intense urbanization favours the spread of vector-borne diseases, which may also flourish due to the higher density of both people and animals (both domestic and peridomestic ones), as well as due to various environmental and socioeconomic modifications [2–4]. In addition, the globalization of trade and travel has facilitated the spread and establishment of invasive alien species (IAS). Insects predominate among non-native terrestrial invertebrates in Europe: of 1,522 established species, 1,306 (86%) are insects [5]. The IAS inadvertently introduced into Europe include several invasive mosquito species (IMS), which have found environmental and climatic conditions favourable for the establishment of permanent populations. These IMS are recognised as responsible for the emergence or reappearance of mosquito-borne diseases such as chikungunya, dengue and West Nile virus (WNV).
One IMS of major public health concern in Southern Europe is Aedes albopictus, the Asian tiger mosquito, which arrived in Europe in Albania in 1979 and then Italy in the early 1990s, through the trade in used tires. Ae. albopictus is already established in large areas of Greece and Southern Europe [6–8] and studies indicate that its rate of expansion in Greece is quite rapid [9–11]. Ae. albopictus has already been responsible for transmitting both dengue and chikungunya viruses in continental Europe, including over 200 laboratory-confirmed cases of the latter in Italy (Region of Emilia Romagna) in 2007 [12,13] and local dengue transmission in Croatia and France [14,15].
The IMS problem may affect the economy and society in various ways, through impacts on human and animal health, as well as on various services and activities. These impacts generate certain economic costs related to control strategies, public health measures, treatment of illness, productivity losses, information and awareness campaigns, and losses in tourism and other sectors. Economic impacts can be direct or indirect. Direct economic impacts are usually expressed as the net increase in public health spending as a result of the appearance of IMS and include, among other things, control-and-surveillance programs, private expenditures and direct medical costs. Direct impacts are the most clearly defined impacts as they can be explicitly expressed in monetary values. On the other hand, indirect impacts include the costs associated with new research and management services (in both the public and private sectors of the economy), as well as the effects of IMS on tourism, etc [16–19].
Thus, the gradual establishment of higher IMS populations in Greece has been accompanied by greater risks of mosquito-borne diseases, increased costs of implementing prevention measures, higher nuisance levels and side-effects on tourism and other economic sectors. The aim of this paper is thus to present the main categories of costs related to the aforementioned problem, to evaluate the potential benefit of enhanced prevention measures and to examine citizens’ and experts’ opinions concerning the various socioeconomic aspects of the problem. In this framework, the present study offers a chance to consider the evaluation and selection of strategies for similar socio-ecological problems, by various interest groups, under the prism of different institutional approaches in an ecosystemic context.
Prevention and control costs and data on health impacts were collected and analyzed in collaboration with the National Public Health Organization (formerly known as the Hellenic Centre for Disease Control and Prevention (HCDCP)), public health agencies and private companies specializing in mosquito control activities. In a previous work, a Cost-of-Illness study was carried out to estimate medical costs and productivity losses, from the West Nile Virus (2010) [20] while recent estimates are presented here concerning medical costs incurred by imported cases of dengue, chikungunya and Zika virus in Greece for the years 2013–2017. Citizens’ willingness to pay (WTP) for improved mosquito control programs was also based on an earlier study, which employed a contingent valuation method (CVM), specifically the discrete choice method [21].
Two new surveys were conducted for the present paper. These provide a deeper exploration of the socioeconomic impacts and benefits of implementing improved prevention and control strategies. The first was a nationwide web-based survey aiming to record citizens’ opinions and attitudes, and the second was a small-scale survey of experts involved in mosquito control activities in Greece. Fig 1 presents all the methods implemented and how they contribute to the overall estimation of the identified costs and benefits associated with the problem of IMS. It should be pointed out that costs and benefits are somehow interrelated, so that the elimination of the socioeconomic costs entails a positive consequence on the benefit side resulting from the control of IMS.
The annual public control and prevention costs examined in this study consist mainly of: (a) annual mosquito vector control activities, (b) contingency costs incurred in response to the WNV epidemic by the responsible national agency, the National Public Health Organization (NPHO), and (c) costs of the additional screening of blood donations that is imposed because of the risk of transmission of WNV through blood transfusion. Market prices for vector control activities were provided directly by regional and municipal authorities and private companies. The annual costs incurred by the NPHO from 2010 to 2013 were extracted from official reports and databases. The cost of additional blood safety testing was provided by the Hellenic National Blood Centre.
Health impact costs were assessed in two ways. First of all, we estimated the medical costs for all imported cases of dengue, chikungunya and Zika virus in Greece for the period 2013–2017 (Table 1). This calculation was based on anonymized data on the duration of hospitalization of reported cases, including intensive care treatment, provided through the official records of NPHO. It should be noted that the reported cases consisted mainly of infected travelers from chikungunya, dengue and Zika endemic countries who presented symptoms of these diseases upon their return to Greece. These estimates of medical costs are similar to those in the recent literature regarding the imported chikungunya cases in Italy (based on 2015 data), and the 2005–2006 chikungunya epidemic in La Reunion [22, 23].
On the other hand, in order to evaluate a proxy estimate of the health burden of mosquito species in public health we also used the WNV costs induced by other mosquito species. Specifically, we present in Table 2 the medical costs of the 2010 WNV outbreak in Central Macedonia, Greece and their associated public health prevention and control strategies’ costs [24,25]. In this epidemic, which occured mainly in this Region, a total of 260 cases were hospitalized during the first year. In the following three years of transmission the number of hospitalized cases fell to 30, 18 and 22 (Table 2). It should be underlined that enhanced surveillance and control measures, were implemented during the first year of the outbreak and particularly during the peak months of transmission (from June to October).
It might seem that costs associated with WNV (which is spread mainly by Culex mosquitoes) may be a rather poor approximation to health impacts of the Asian tiger mosquito. However, even though the costs for the WNV outbreak cannot be directly attributed to IMS, they represent an up-to-date indicator of regular public expenses incurred against the spread of mosquito-borne diseases in Greece. This indicator is directly comparable to other relevant estimates in South Europe [26], thus enabling better adjustment for country-level effects (variation) in the cost-of-illness assessment. In addition, it is also interesting to note that the annual regional surveillance program for arbovirsuses in the Emilia Romagna Region of Italy [27] is designed for the common surveillance of various vector borne diseases such as WNV, chikungunya and dengue, highlighting thus the importance of applying integrated approaches against all mosquito-borne diseases.
The estimation of health impact costs, (Tables 1 and 2) comprising medical costs and productivity losses, were based on the cost-of-illness (COI) analysis [28] in which the burden of a disease on society is estimated in financial terms using both direct and indirect measures. Direct costs consist mainly of medical care, both inpatient and outpatient, and are estimated using market prices. According to the National DRG (Diagnosis Related Groups) Indicators published in the 3054/18-11-2012 Official Government Gazette of the Hellenic Parliament, the average daily hospital care cost in Greek public hospitals is approximately 207€/day; this was multiplied by the total inpatient care days.
In addition, indirect costs represent the loss of productivity due to morbidity. These costs were estimated only for earnings lost during the reported days of sickness among people older than 18 years of age; the value of a lost working day was then multiplied by the total number of sick-days. The cost of a lost working day for people in the 18 to 65 years age range was calculated according to the per capita net income equivalent for the reference years (2011–2013) [29], divided by 220 working days. For people aged 65 years and over, the cost of a lost working day was calculated from the country’s median hourly earnings [30] for 2010, adjusted for inflation by the Consumer Price Index and then multiplied by 8 working hours. Due to lack of data on the age of patients, productivity losses for the imported cases of chikungunya, dengue and Zika virus, were calculated based from the median hourly earnings [30] for 2014, adjusted for inflation by the Consumer Price Index for each indicative year and then multiplied by 8 working hours.
We attempted to elicit household preferences for controlling IMS through a choice experiment approach [21]. Specifically, this stated preference method was implemented in order to examine household preferences regarding various attributes of mosquito control programs in relation to mosquito impacts (i.e. in order to assess the influence of these attributes in choosing a program). The initial selection of attributes was based on feedback from experts and on previous relevant studies. These attributes and their levels were then reduced to only those that were found to have a clear relationship with mosquito control programs, and this relationship was articulated in operational terms easily comprehended by citizens.
It should be noted that the control of mosquitoes is mainly carried out through annual activities which include monitoring and surveillance of the mosquito larvae population, implementation of larvicidal, adulticidal and surface residual ground treatments, and application of larvicidal and small scale adulticidal treatments by aerial spraying. On the other hand, controlling the Asian tiger mosquito calls for a more complex management plan and coordinated actions which have only recently been designed by the LIFE CONOPS research initiative (http://www.conops.gr/management-plan-for-aedes-albopictus-in-greece/?lang=en). The actions in this plan include (among others): standardized quantitative monitoring by special ovitraps, recording of mosquito population density data, involvement of the local population in the control campaign in private areas, residual door-to-door control interventions and use of larvicides in the road drains of public areas throughout the whole breeding season. The control methods and management plans according to the type of mosquitoes were described to all respondents at the beginning of the interview [21].
When selecting the attributes, two main categories of benefits that may be derived from improved mosquito control programs were identified: less nuisance and reduced risks to health. Another distinction was drawn between benefits from controlling native mosquitoes (principally of the Culex and Anopheles genera) and benefits from controlling invasive mosquitoes (the Asian tiger mosquito). For this purpose, two health risk attributes were used: (a) one related to the health risks that are mainly associated with native mosquitoes, such as WNV, and (b) another one related to the health risks due only to the Asian tiger mosquito (such as chikungunya fever). The nuisance attributes were likewise separated into: (a) nuisance during the day-time, which is a problem caused mainly by the Asian tiger mosquito, an “aggressive day-time biting mosquito” [31], and (b) nuisance at night, mainly associated with the native mosquito species. A cost attribute was included in order to elicit welfare effects, as determined by individuals’ preferences between alternative mosquito control programs.
Interviews were conducted from mid-June 2015 to the end of October 2015 in several districts within the Athens Metropolitan Area selected in order to represent the socioeconomic diversity of the city and the different degrees of exposure to the mosquito problem (either the Asian tiger mosquito or the native species). The survey was administered in face-to-face interviews by three trained interviewers and the average duration of each interview was 15 minutes. Although it was not possible to draw a strictly random sample from a sampling frame, the sample achieved a high degree of representativeness concerning geographical location and it was stratified based on location, sex and age (according to the 2011 Census). A total of 495 completed interviews were collected.
The socioeconomic evaluation of the mosquito control strategies was enhanced by conducting a survey of experts on the issue of the overall mosquito problem. This survey was designed to evaluate the socioeconomic impacts of the mosquito control plans by interviewing key stakeholders, public policy makers, medical practitioners, public health experts and regional administrators. The questions were formulated in order to evaluate the results of the preceding studies (especially the choice experiment) and provide qualitative evaluation of specific policy-related decisions (ecosystem services, adequacy of control programs, etc.). The questionnaire was distributed to a pool of 100 experts all over the country, selected on the basis of their experience and involvement in the design and implementation of mosquito control strategies. The survey was conducted through telephone interviews from May 2016 to May 2017 in collaboration with a member staff of the Ministry of Health and a total of 59 responses were collected.
Apart from the survey in Athens, another questionnaire was also designed for a nationwide web-based survey aimed at eliciting citizens’ opinions regarding certain socioeconomic aspects of the mosquito problem. Its particular focus was to examine and then to validate at the national level a set of parameters related to: a) private prevention costs for IMS, and b) individual preferences among various mosquito control programs. The questionnaire was distributed through a popular meteorological data website (www.meteo.gr) with a high daily number of visitors [32]. For the purpose of our survey, a special banner appeared on the home page, from which visitors followed a link to the web survey. The banner appeared randomly to visitors, but a selection bias could arise due to (i) the non-representative nature of the internet population, and (ii) self-selection of participants (the `volunteer effect') which was possibly related to their interest in mosquito control. The first set of questions focused on the respondents’ knowledge of the Asian tiger mosquito. Subsequent questions concerned: (a) the current perceived level of nuisance during the day and separately at night, both rated using a 5-point Likert scale (nuisance impacts from the Asian tiger mosquito and from other species were estimated by attributing the nuisance during the morning and late afternoon hours to the former and the nuisance during the evening and night hours to native mosquito species), (b) the period (months/year) with significant mosquito nuisance, (c) the monthly household expenditure for private prevention measures, and (d) the main reasons for taking individual prevention measures (i.e. they had to choose between health risk reduction and nuisance reduction). The survey took place in September and October 2016 with a total of 1,220 responses from all over the country.
According to national data published online on the governmental Greek Transparency Program Initiative (http://diavgeia.gov.gr), the average annual public mosquito control costs in the Athens Metropolitan area range from approximately 800,000 € to 1,330,000 € per year. This represents an average annual cost of about 0.6 € to 0.9 € per household. These programs consist mainly of adulticide and larvicide activities, mainly with the use of specific chemical larvicides currently available or undergoing the revision process in the EU [33], such as Diflubenzuron; these are designed for the control of Culex and Anopheles species and therefore target the elimination of their associated diseases (such as WNV). In other words, the implementation of these programs is not specifically tailored to the control of the Asian tiger mosquito and the prevention of chikungunya and dengue fever, even though Diflubenzuron also has high efficacy rates for the Aedes species [34]. It should be noted that some surveillance activities for the Ae. albopictus are currently implemented in most parts of Greece including the Athens metropolitan area [35], however, data are insufficient for calculating the cost of these as yet.
In contrast to Greece, some other European countries are implementing programs specifically aimed at the control of the Asian tiger mosquito, such as the "Italian Plan of the Emilia-Romagna Regional Health Authority for the fight against the Asian tiger mosquito and the prevention of Chikungunya and Dengue fever" [36]. The current Greek management plan differs from these chiefly in that because it is not focused on combating the Asian tiger mosquito, larvicide activities mainly take place in public spaces without considering specific urban (residential) areas with high breeding activity of Aedes albopictus. The lack of a regional or national plan aimed specifically at controlling the Asian tiger mosquito makes the measurement of the efficacy of larvicide activities against Aedes albopictus difficult.
What is more, due to high domestic breeding patterns of the Aedes species, information and communication activities can have a very high impact on the control of mosquito populations. According to recent estimates [36], the annual total expenditure for information activities in Emilia-Romagna during the years 2009–2011 ranged from 150,000 €/year to 0.6 mil €, significantly lower rate than the costs for regular anti-larval treatments which ranged from 3.6 to 4.4 million €/year. As previously noted, the annual integrated surveillance plan for arboviral diseases in Emilia Romagna is designed for the common surveillance of various vector-borne diseases such as WNV, chikungunya and dengue, while recent studies also emphasize the effectiveness of community participation also concerning the elimination of Culex species [37].
The overall public control and prevention costs associated with the WNV epidemic have already been presented in Table 2. Higher costs in the first year of application of measures are justified as a contingent response to the expansion of the outbreak. However, it appears that costs fell significantly during the following years. This could be also interpreted as a result of the epidemic being partially controlled; however, there are inadequate data within this study to support this argument.
According to the results Table 1, the average health cost for an imported case of Dengue was estimated to be 1,170 € for chikungunya, 2,774 € for dengue and almost 3,500 € for Zika virus. Even though the overall socioeconomic costs in the case of epidemic outbreaks for these diseases cannot be estimated with high precision (due to the limited number of disease cases), it is possible that in the scenario of future epidemics, disease complications could outweigh the present costs of treating the diagnosed imported cases.
The total cost of illness (COI) in the first year of the WNV outbreak (2010) was estimated at about 900,000 € [20]. This includes the cost of hospitalization for 260 recorded WNV cases, 25 of whom needed further hospitalization in intensive care at an extra cost of about 160,000 €. The total COI in the following year was estimated to be nearly 120,000 € for the hospitalization of 30 cases (two of whom required treatment in intensive care units). Subsequently, 18 cases were recorded and treated in 2012 with only one case requiring intensive care. The total COI for this year amounted to 71,000 €. Finally, in 2013, 22 cases were diagnosed (two in intensive care) and the COI was correspondingly slightly higher at 77,000 € (Table 2).
Even though it is very difficult to provide precise estimates of the total costs and the total social benefits of mosquito control programs, the results of our previous study [21] permit us to conclude that the benefits of mosquito control in terms of reduced nuisance and reduced health risks are likely to exceed the associated implementation costs.
Under our most conservative scenario (i.e. a medium prevention scenario, effective only against the native mosquito species), the estimated aggregate benefits from improved control programs can reach up to 11.2 million €/year, thus corresponding to a net benefit of 7.40 €/household/year (see Table 3). These results provide an order of magnitude estimate of the economic feasibility of improved mosquito control programs in the study area (Athens Metropolitan area). Specifically, the benefit-cost ratio of any program which is expected to achieve the selected target levels at a cost less than 13 times the cost of the current mosquito control program (800,000 €/year) will be greater than one (i.e. would be economically profitable). This cost could further increase (up to 31.3 €/household/year) if a high prevention scenario, effective against all mosquito species, were implemented. On the other hand, the expected added value of taking measures not only against native but also against the Asian tiger mosquito was found to be substantial, representing on average an additional benefit of about 15€/household/year [21]. As shown in Table 3, this benefit can be attributed mainly to the high health risks posed by the introduction of new invasive species into the study area.
In the survey of experts, 48% of the respondents considered the financial budget allocated to control programs to be adequate for confronting the problem while 34% suggested that an increase in public spending would be necessary. In addition, experts judge that the current control programs achieve balance between cost and effectiveness in their design and implementation. With respect to the potential negative impact of prevention measures on relevant ecosystem services, 65% of the experts stated that there are no (significant) negative impacts from these measures. Regarding the means of obtaining extra funds for supporting mosquito management, experts indicated that: (a) a redistribution of public resources would be necessary, (b) a reallocation of funds within national and regional budgets could improve the financing of mosquito control programs, and that (c) a financial contribution by citizens is equally important for the confrontation of the problem.
It should be noted that the Asian tiger mosquito can exploit water containers in private apartments for its breeding. Therefore, according to the experts, private prevention activities could contribute significantly to the reduction of the problem at a much lower cost, especially if supported by public information activities which, as shown in the case of Emilia Romagna in Italy, could be more cost-effective. Lastly, regarding the prioritization of the objectives of future control programs (Table 4), experts stated that the health impacts should be considered as the primary objective of these programs. Specifically, they consider the health threats of native and invasive mosquito-borne diseases as almost equally important, whereas they treat nuisance from mosquito species as a less important impact factor.
In the web survey of private citizens, 83% of respondents stated that the current prevention and control measures are insufficient or inadequate for dealing with the mosquito problems and therefore there is a need for further measures to be taken. The average private prevention costs of the sample were approximately 16€ per month in the period when mosquitoes are active, which amount to about 100 €/year. There was significant regional variation in these estimates, ranging from below 80 € (e.g. Region Thessaly and Region of North Aegean) to over 125 € (e.g. Region of Eastern Macedonia and Thrace, and Region of Central Greece). This variation may be an indirect indicator of the magnitude of the mosquito problem, which is strongly associated with the nuisance conditions in each area. It should be also noted that this revealed behavior concerning prevention costs can be used as a lower-bound proxy of individuals’ potential benefits from improved control measures in each region.
The results from this survey concerning the preferences of individuals for the diverse mosquito control programs are shown in Table 5. Concerning the main targets of these measures (Table 5), health impacts were considered to be more important than nuisance impacts, confirming the findings of previous surveys in Greece [20,21].
Furthermore, as in the other two studies, health risks from invasive species were considered to be a serious threat. Therefore, both group, (experts and citizens), appear to rate the health risks higher compared to the nuisance and cost factors of mosquito control programs. Finally, an important finding of this survey was that citizens seem to be aware of the environmental consequences of mosquito control measures. In particular, about 74% of respondents stated their disagreement with measures that may potentially affect the physical environment and the ecosystems.
The present paper aims at an overview of the socioeconomic aspects of the problem of invasive mosquitoes as recorded by the main interest groups in society (citizens and experts). It provides substantive indicators regarding the citizens’ perceived benefit derived from the implementation of improved mosquito control programs, as well as experts’ evaluation of the socioeconomic effectiveness of current and future programs for controlling the problem of invasive mosquitoes. In contrast to other studies [17,38,39] findings from both perspectives show a higher priority for the prevention and reduction of health risks as opposed to nuisance control.
Furthermore, based on the results of the survey conducted in Athens, citizens are willing to pay a considerable premium for effective protection against the spread of unfamiliar diseases, thus implying a risk-averting behavior against invasive mosquito threats. In other words, citizens are willing to pay today for improved control programs that will be able to eliminate potential future impacts and risks. The fact that climate change trends may worsen the mosquito problem and increase the risks of the transmission of new diseases (such as Zika virus) is likely to provide ever increasing potential individual and social benefits from implementing more efficient mosquito control management plans in the coming years [40].
The cost estimates extracted from the current study allow for a comparison with recent similar estimates in Southern Europe. According to our findings the costs of public mosquito control programs range from approximately 0.6 € to 0.9 € per household in Athens, while the public costs of informed arbovirus plans in the Region of Emilia Romagna in Italy reach almost 1.2 € per household [36]. Using the Purchasing Power Parity Index [41] this figure translates to an equivalent of 1.04 €/year per household in Athens, indicating that a small per capita increase in public costs could justify the design and implementation of more targeted programs in Greece, in terms of perceived citizens’ benefits levels as already presented in the current analysis.
What is more, a recent study estimated that the implementation of public intervention strategies against the spread of Aedes related arboviruses in Italy since 2007 may have saved up to 13.5 million € indicating the cost-effectiveness of these interventions both from an economic and a health perspective [23]. It should be noted that our analysis indicated an annual benefit of up to 11 million € from the implementation of optimal mosquito control programs in the Athens area intended to achieve public health targets similar to those of Italy.
With regard to health impact costs, the medical costs for an average cost of illness of an imported chikungunya virus case in Italy reaches approximately 3,500 € [23], while the average in Greece, based on our limited sample of imported disease cases estimated to be about 1,100 €. The average cost of illness for all three types of imported diseases in Greece (chikungunya, dengue, Zika virus) was found to be approximately 2,500 €. According to another recent study in La Reunion [22], the mean cost of illness per inpatient case of chikungunya reached approximately 2,000 €. Estimates of the cost of illness of recent WNV epidemics indicate a cost of about 3,500 € per case in Greece, while in Italy cost data are only available for the mean cost of illness for a WNV case with neuroinvasive complications (WNND), which reaches approximately 15,000 € per case [26].
The above estimates offer an important range of socioeconomic figures relevant to mosquito-related diseases in Southern Europe which could act as significant indicators for evaluating the societal benefits of integrated public control programs against the spread of arbovirus diseases under turbulent climatic and societal conditions.
The establishment of invasive species is usually associated with increased economic costs. For example, a study in the USA [42] estimated environmental damages and losses of almost $120 billion per year. According to a European Commission Impact Assessment [43], Invasive Alien Species are estimated to have cost the EU at least €12 billion/year over the past 20 years and the damage costs continue to increase. It is predicted that, due to the trends in climate change, the invasive mosquito problem will intensify in the immediate future [44]. Therefore, the evaluation of the socioeconomic costs of invasive mosquitoes is a vital but highly challenging task made even more complex by changing climatic conditions, as well as by globalization and urbanization trends that may call for the adoption of multi-disciplinary and more holistic approaches in order to evaluate the effectiveness of the expenses incurred in improving public health and social welfare [45].
|
10.1371/journal.pcbi.1005895 | Functional interrogation of Plasmodium genus metabolism identifies species- and stage-specific differences in nutrient essentiality and drug targeting | Several antimalarial drugs exist, but differences between life cycle stages among malaria species pose challenges for developing more effective therapies. To understand the diversity among stages and species, we reconstructed genome-scale metabolic models (GeMMs) of metabolism for five life cycle stages and five species of Plasmodium spanning the blood, transmission, and mosquito stages. The stage-specific models of Plasmodium falciparum uncovered stage-dependent changes in central carbon metabolism and predicted potential targets that could affect several life cycle stages. The species-specific models further highlight differences between experimental animal models and the human-infecting species. Comparisons between human- and rodent-infecting species revealed differences in thiamine (vitamin B1), choline, and pantothenate (vitamin B5) metabolism. Thus, we show that genome-scale analysis of multiple stages and species of Plasmodium can prioritize potential drug targets that could be both anti-malarials and transmission blocking agents, in addition to guiding translation from non-human experimental disease models.
| Malaria kills nearly one-half million people a year and over 1 billion people are at risk of becoming infected by the parasite. Plasmodial infections are difficult to treat for a myriad of reasons, but the ability of the organism to remain latent in hosts and the complex life cycles greatly contributed to the difficulty in treat malaria. Genome-scale metabolic models (GeMMs) enable hierarchical integration of disparate data types into a framework amenable to computational simulations enabling deeper mechanistic insights from high-throughput data measurements. In this study, GeMMs of multiple Plasmodium species are used to study metabolic similarities and differences across the Plasmodium genus. In silico gene-knock out simulations across species and stages uncovered functional metabolic differences between human- and rodent-infecting species as well as across the parasite’s life-cycle stages. These findings may help identify drug regimens that are more effective in targeting human-infecting species across multiple stages of the organism.
| Malaria is a worldwide problem of clinical significance causing an estimated 483,000 deaths, with a disproportionate percentage occurring in children less than 5 years of age, according to the World Health Organization[1]. Additionally, 1.2 billion people are at high risk of contracting the infection[1]. Plasmodium is a challenging organism to understand and treat, since it has a complex life cycle[2] and can remain latent within hosts. Indeed, current antimalarials target the symptomatic Plasmodium life cycle stages, while, allowing ample time for transmission before symptoms are seen. The use of experimental model organisms, such as mice, has provided a wealth of knowledge about the various life cycle stage in the Plasmodium genus; however many differences between rodent-, primate-, and human-infecting species remain incompletely understood. Thus, to identify effective means to eradicate malaria, there is a need to understand its biological capabilities as it relates to drug targeting in different stages of its life cycle and also across the different species.
Among potential drug targets, metabolic genes are of particular interest, since many anabolic and catabolic processes are critical for cellular growth and survival. Furthermore, methods have been developed to identify vulnerabilities in human pathogens by accurately predicting essential metabolic genes in genome-scale metabolic network reconstructions [3–14]. Here, we present detailed genome-scale metabolic network reconstructions of five life cycle stages of Plasmodium falciparum (P. falciparum). We used these stage-specific genome scale metabolic models GeMMs to characterize functional metabolic features of each stage as well as to predict essential targets whose inhibition would interfere with malaria growth across asexual, sexual (transmission) and mosquito stages. Moreover, we reconstructed GeMMs for four additional Plasmodium species that infect rodents, non-human and human primates (including P. vivax, berghei, cynomolgi, and knowlesi). We used these Plasmodium GeMMs to investigate cross-species similarities and differences with particular focus on characterizing functional metabolic differences between rodent- and human-infecting species. Results provide a means to rank order and stratify established and new malaria treatment targets in addition to providing key insights into differences between rodent versus primate specific infections and implications for the interpretation of experimental animal models.
A manually curated and quality-controlled[15] metabolic network reconstruction of P. falciparum (Fig 1, Step 1), iAM-Pf480, was constructed to interrogate the parasitic metabolic capabilities throughout the life cycle stages of malaria. iAM-Pf480 (Fig 2A) was built using the genome annotation of P. falciparum (Plasmodb.org, v24), the Malaria Parasite Metabolic Pathway (MPMP) Database (http://mpmp.huji.ac.il/), and specific biochemical and genetic characterization studies from 332 primary and review literature reference articles (Table A in S1 Tables). The metabolic network of the P. falciparum accounts for 1083 reactions, 617 unique metabolites and 480 genes localized to their respective intracellular compartments and organelles, including the cytoplasm, mitochondrion, the plastid-like apicoplast, endoplasmic reticulum, Golgi apparatus, and lysosome. Gene-protein-reaction (GPR) associations could be defined for 480 genes and 68% of all enzymatic reactions (Fig 2A).
In order to validate (Fig 1, Step 2) iAM-Pf480 predictions, we tested if iAM-Pf480 could correctly predict gene essentiality (Fig 2B). To accomplish this, we compiled a list of experimentally confirmed gene knockouts (n = 21, Table B in S1 Tables) and phenotypes resulting from targeted inhibition of enzymatic activities with drugs (n = 59, Table C in S1 Tables) in P. falciparum[16]. Under standard growth conditions, iAM-Pf480 correctly predicted 95% and 71% for the single gene knockouts and drug inhibition phenotypes, respectively (Fig 2B, Table B and C in S1 Tables). We also compared iAM-Pf480 gene essentiality predictions to iTH366[17], iPfa[18] and iPfal17[19], using our set of experimentally validated targets (Table B and C in S1 Tables), which revealed that iAM-PF480 accounts for a larger scope of genomic content, a larger biochemical complement, and functionally outperformed previously published P. falciparum models (see supplementary material, Fig A and Table L in S1 Text, Table B and C in S1 Tables).
iAM-Pf480 flux predictions were validated against available rapid stable-isotope labeling data to assess the metabolic flux changes in glycolysis of wild-type (WT) versus apicoplast pyruvate dehydrogenase knockout (PDHapi-KO) P. falciparum parasites[20]. Glucose and hypoxanthine uptake rates[21,22] were used to constrain the model (Table D and E in S1 Tables). Model-predicted glycolytic flux rates for both WT and PDHapi-KO showed good correlation (Pearson correlation coefficients 0.77 and 0.78 for WT and PDHapi-KO, respectively) with published experimentally measured fluxomic data[20] (Fig 2C and 2D).
Validation through fluxomic, single gene knockout, and drug targeting enzymatic assays provided confidence in the content and predictive capabilities of the metabolic model, setting the stage for further investigation into the consequences and capabilities of the parasites metabolic architecture (Fig 1, Step 3).
Only approximately 1% of the asexual parasites develop into male and female gametocytes in response to yet unknown cues[23]. However, most current therapies are targeted against the blood stages, which result in clinical infections[2]. Thus, there is a pressing need to investigate potential targets that are critical for both the asexual and gametocyte stages to suppress both malarial transmission and active infection[23]; identification of such targets require understanding the stage-specific metabolic capabilities of P. falciparum. Towards this end, stage-specific models of metabolism throughout malaria’s life cycle were constrained using multiple data types (Fig 1, step 3i). Stage-specific growth rates[24], glucose and lactate secretion rates[25], as well as stage-specific transcriptomic data[26] were used to constrain iAM-Pf480 producing five distinct stage-specific models; trophozoite (T), schizont, early gametocyte (GII), late gametocyte (GV) and ookinete (ook) (Fig 3A) (see Table F in S1 Tables and Methods for details). In the early gametocyte stage (GII), biomass precursors production was permitted; however in the late, metabolically quiescent[23,25] mature gametocyte stage (GV), ATP generation was optimized[27] without associated net biomass accumulation.
Generic, canonical reaction groupings into pathways do not inform functional states. However, GeMM-based simulations can be used to calculate groups of reactions with highly correlated metabolic fluxes under a set of condition(s), i.e., correlated reaction sets (co-sets), that in turn can yield insight into metabolic capabilities[28,29] and also reduce network size into functionally correlated modules of reactions. Artificial Center Hit and Run[30] sampling was used to determine the steady state flux distributions of the stage-specific models, and these were used to compute co-sets across different conditions (Table G in S1 Tables) (see Methods). 612 reactions were involved in co-sets with 3 or more reactions spanning pathways related to anabolic and catabolic processes for amino acids, fatty acids, and biomass production (Table G in S1 Tables). A quantitative assessment of flux magnitude as well as co-set size across the different stages are visualized as Voronoi plots (Fig 4A). Comparison across the different stages reflects that the biomass constraint provides the greatest influence on the size and modularity of the metabolic network, thus the late gametocyte, with the most relaxed biomass constraint, had the most modularity where high modularity reflects networks with co-sets that are similar in size, whereas low modularity refers to networks that have one or two co-sets that are very large in comparison to the rest of the co-sets in the network (Fig 4).
Across the different stages, there were considerable changes in central carbon metabolism, based on the co-sets (Fig 4). The metabolic simulations predicted splitting of glycolysis into independent upper and lower branches to divert biomass (nucleic acids, lipids, glycosylated proteins) required for proliferating parasite stages, whereas the late gametocyte stage is non-proliferative, accordingly, both the upper and lower glycolytic branches were in one co-set. Contrary to what was believed about the pentose phosphate pathway (PPP) deploying the oxidative arm only during the early stages of the parasite intra-erythrocytic development cycle (IDC) and the non-oxidative arm in the later stages of the IDC [31], simulations with stage-specific models showed that: 1) both the oxidative and non-oxidative arms were active with low fluxes in the early stages (trophozoite, schizont and GII), 2) the non-oxidative arm operated in the backward direction to produce glycolytic intermediates in the early stages while the oxidative arm produced NADP+ and ribose-5-phosphate (r5p), and 3) only the non-oxidative branch of PPP was active in the GV stage in the backward direction and was correlated with inositol metabolism (Fig 4) underscoring the importance of inositol metabolism across the transmissible stage of malaria[32] 4) similar to the GV stage, only the non-oxidative arm was active in the ookinete stage, albeit in the forward direction using fructose-6-phosphate (f6p) and glyceraldehyde-3-phosphate (g3p) supplied by glycolysis, thus maximizing the production of r5p (Fig 4B).
We constructed a comprehensive map of druggable targets for P. falciparum (Fig 5) using our curated list of experimentally validated targets (Table B and C in S1 Tables). This was used to compare to predictions made by the stage-specific P. falciparum. Selecting stage-specific targets spanning the parasite’s life cycle could promote the design of strategies for potential multi-stage targets or combination of existing drugs. The color scheme of highlighted reactions denotes model prediction classification across all stages. The red group in Fig 5 highlights reactions sensitive in the proliferative stage as well as the late gametocyte (GV) stage; this is of particular importance since the late gametocyte is of high clinical interest to target and represent a large percentage of the parasitic load that is not targeted by some of the more commonly used treatment drugs.
Reactions associated with genes from experimentally validated single gene targets (Fig 2B) are highlighted in the colored rectangles in Fig 5. There are several gene deletion associated reactions for which drugs have not been developed; these highlight potential targets for new drug development. We also note that the yellow group identifies reactions that were missed by the models, and highlights areas for future model refinement. By overlaying stage-specific model predictions on top of the experimental (single-stage) validated drug targets, the network map aids in the prioritization of drug target characterization.
The use of experimental model organisms, such as mice, has yielded a wealth of knowledge about the Plasmodium genus. However, there has been relatively limited investigation into species-specific differences in Plasmodium metabolism[33–35]. While there is 94% (422/448) homology among the metabolic genes of the different species (Table H in S1 Tables), it is unclear how they differ in their metabolic capabilities. Further, differences have been observed between rodent- and human-malaria infecting species to certain drug inhibitors[36] but no mechanistic explanation was attributed to these differences. Beginning with the iAM-Pf480 GeMM we systematically studied the functional metabolic differences between 5 different Plasmodium species (Fig 1, Step 3ii): Plasmodium falciparum 3D7 (Pfal), Plasmodium vivax (Pviv) Sal-1 (iAM-Pv461), Plasmodium berghei (Pber) ANKA (iAM-Pb448), Plasmodium cynomolgi (Pcyn) strain B (iAM-Pc455), and Plasmodium knowlesi (Pkno) strain H (iAM-Pk459). The core metabolic content reflects the intersection of the genes, reactions, and metabolites of all five reconstructions, whereas the pan metabolic content is the union of these entities. The core Plasmodial metabolic content is comprised of 1064 reactions and 422 orthologous genes (Fig 6A) and the pan metabolic capabilities include 1083 reactions corresponding to 448 orthologous genes (Table H in S1 Tables), reflecting a considerable level of conservation. However there are multiple functional differences cross the metabolic GeMMs (Fig 6B, Table H and I in S1 Tables). The differences in metabolic reaction content across the five reconstructed species predominantly involve co-factor metabolism (4 reactions), phospholipid metabolism (4 reactions), and purine/pyrimidine metabolism (3 reactions) (Fig 6B).
We performed in silico single gene deletion analysis for the set of 448 orthologous genes shared among the five species (Table H in S1 Tables). The deletion of 15% (67/448) of these orthologous genes caused a 100% reduction in growth across all species (Table H in S1 Tables). These genes spanned several metabolic subsystems with the majority involved in isoprenoid biosynthesis, phospholipid metabolism, as well as purine and pyrimidine metabolism. Interestingly, 19 genes out of this set have already been targeted by drug inhibitors (Table C in S1 Tables) while the remaining 48 orthologous genes represent overlooked novel druggable vulnerabilities in malaria.
14% (61/448) of the orthologs differed in their essentiality across the 5 Plasmodium species. Reactions with the most striking differences in essentiality across malaria species were between the rodent and non-rodent species, namely: thiamine pyrophosphokinase (TPK), and choline kinase (Table H in S1 Tables). Both genes were predicted to be essential in P.berghei only, while their deletion had no effect on growth in any of the non-rodent species.
Thiamine analogs[36] and choline kinase (CK) inhibitors[37] have been tested as antimalarial therapeutics both in plasmodium species that infect human (in vitro) and rodents (in vivo); however, it is not clear from these studies whether an equally potent antimalarial effect is observed in both species. Our multi-species reconstructions revealed three key thiamine (Vitamin B1) biosynthesis enzymes: phosphomethylpyrimidine kinase (PMPK), hydroxyphosphomethylpyrimidine kinase (HMPK1), and thiamine-phosphate pyrophosphorylase (TMPPP) that are absent in rodent malaria, but present in non-rodent malarial species (Fig 6C). Consequently, TPK was predicted to be essential for rodent malaria, but non-essential in non-rodent species, which can replenish thiamine pyrophosphate through the HMPK1-PMPK1-TMPPP pathway (Fig 6C).
In silico deletion of choline kinase (CK), the first enzyme in the Kennedy pathway (CDP-choline pathway) for synthesis of phosphatidylcholine (PC) (Fig 6D), inhibited growth of the rodent species while causing only marginal reduction in growth (3%) in the primate and human species. The model simulations revealed the difference in essentiality of CK was the result of the lack of phosphoethanolamine N-methyltransferase (PMT) in P. berghei, which in turn rendered it incapable of de novo PC synthesis from ethanolamine (Fig 6D). Thus, efforts to perturb PC for malaria treatments will require different strategies in human species than in non-human-infecting species since decreased potency is expected when perturbing choline metabolism in the non-rodent relative to rodent-infecting species.
A comparative analysis of stage-specific models of human and rodent species (see Methods and supplementary material for details), based on co-sets and in silico single gene deletion experiments, showed that pantothenate metabolism was not essential for growth in any of the life cycle stages of P. berghei. In contrast, pantothenate metabolism was essential for growth during the asexual and early gametocyte stages of P. falciparum (Fig 6E, Table K in S1 Tables). In line with recent evidence[38], our stage- and species-specific models predicted that the pantothenate transporter activity was essential in human malaria, but was mostly dispensable in rodent parasites. Pantothenate is a precursor of the enzyme cofactor coenzyme A (CoA) and the capability of de novo synthesis of CoA distinguishes P. falciparum asexual forms from its sexual counterparts as well as from the rodent and avian malaria parasites, thus challenging the assumption that rodent and human malaria parasites utilize similar nutrient acquisition strategies[39].
Therapeutic drugs that target multiple stages of the parasite, including sexual and asexual stages, will facilitate eradication of malaria. Developing effective medications will require understanding basic biological mechanisms, particularly the limitations of experimental animal models that are used as surrogates for understanding human Plasmodium pathogenesis. Systems analysis are thus needed to interpret and integrate multiple, large disparate datasets to unravel the complex life cycles of these pathogens.
In this study we use genome-scale metabolic modeling to interrogate malaria stage- and species-specific metabolic capabilities. Through the integration of high-throughput data, careful manual curation, and model prediction validation, we reconstructed detailed stage-specific models that span five distinct stages of the life cycle of P. falciparum. Since GeMMs allow condition-specific analyses, we were able to simulate the effects of reaction inhibition across different stages of the parasite’s life cycle and to identify drugs that are effective across more than one stage. Moreover, we detected stage-dependent metabolic redirection of flux in central carbon metabolism of the parasite that stage-matched proliferation requirements.
The overall genome organization and content across Plasmodium species is highly conserved, with about 4000 conserved syntenic genes located within the central core regions of the 14 chromosomes[40]. Subsequently, it is frequently assumed that findings from the animal models will directly translate to the human-infecting species, particularly in areas of essential, core metabolism, given the high degree of homology across the malaria species (94%). However, our GeMMs for multiple Plasmodium species highlight important metabolic differences. The GeMM simulations provide a mechanistic explanation for why P. berghei would be more sensitive to thiamine analogs as well as to drugs interfering with PC metabolism[36] and not in human infecting parasites. Additionally, differences in pantothenate metabolism were revealed in stage-specific analysis of P. falciparum and P. berghei, further highlighting potential differences between the metabolism of human- versus rodent-infecting species. This finding suggests a potential use of an auxotrophic mutant of P. falciparum defective in the de novo biosynthesis of pantothenate for vaccination in analogy to Mycobacterium tuberculosis[41].
These multi-species GeMMs have enabled us to make informed predictions about specific differences between rodent and non-rodent metabolic capabilities, underscoring the fact that the metabolic architecture and nutritional requirement of a rodent malaria species does not necessarily predict that of a human malaria species[39]. Furthermore, in addition to using the models to make predictions and gain systems-level insights to malaria metabolism in relation to drug targeting, these models can be used as a foundational structure upon which additional high-throughput data can be analyzed and predictive simulations can be conducted, thus leading to improved understanding, testable hypotheses and increased knowledge[42].
The methods employed for the reconstruction, simulation, and analyses presented in this manuscript are briefly summarized below, with further details regarding the procedures, protocols, calculations, and quality control measures provided in the supplementary material. All models are available as part of the supplementary material and will be deposited in the BiGG[43] database.
The genome sequence and genome annotations for P. falciparum were downloaded from Plasmodb.org (release 26). A list of P. falciparum metabolic pathways was built based on current genome annotation of P. falciparum (Plasmodb.org), the Malaria Parasite Metabolic Pathway (MPMP) Database (http://mpmp.huji.ac.il/), and malaria-specific biochemical characterization studies (Table A in S1 Tables). The stoichiometric matrix was constructed with mass and charge balanced reactions in the standard fashion and flux balance analysis was used to assess network characteristics and perform simulations[44]. Linear programming calculations were performed using Gurobi (Gurobi Optimization, Inc., Houston, TX) and MATLAB (The MathWorks Inc., Natick, MA) with the COBRA Toolbox[45,46].
We tested iAM-Pf480-predicted flux rates against kinetic flux data (rapid stable-isotope labeling) of glycolysis in wild-type (WT) and pyruvate dehydrogenase (PDH) deficient P. falciparum parasites cultured in vitro[20]. The generic iAM-Pf480 model was allowed to uptake metabolites available in standard in vitro growth conditions (RPMI 1640, 25 mm HEPES, 2 mm l-glutamine supplemented with 50 μm hypoxanthine and 10% A+ human serum[20]) (Table D in S1 Tables). Uptake rates for glucose and hypoxanthine were obtained from literature[21,22].
For validation of in silico single gene deletion essentiality predictions, we compiled a curated list of experimentally validated gene knock-outs (n = 21, Table B in S1 Tables) and phenotypes resulting from targeted inhibition of enzymatic activities with drugs (n = 59, Table C in S1 Tables) in P. falciparum based on our recently published list[16] of targeted chemical compounds in MPMP.
An experimentally measured growth rate (lower bound of 0.045 mmol/gDW/h corresponding to approximately 15 hours[24] was imposed on the biomass function). Lactate secretion for asexual stages (93% of uptake glucose) was applied to the iAM-Pf480 model simulating in vitro growth conditions (see fluxomics data section). For the gametocyte stages, we developed two models. The first model (GII) simulated early gametocyte stage II which is metabolically active and hence, the objective function was set to maximize the production of biomass precursors. The constraint on the lower bound of the biomass function was relaxed to 0 mmol/gDW/h since it’s expected that the early gametocyte stages will exhibit a lower growth rate compared to the asexual stages. Lactate secretion was set to a minimum of 80% of glucose uptake rate[25]. The second model (GV) represents mature, metabolically quiescent gametocyte stages. The objective function in the GV model was to set to optimize ATP production[23] and while no flux was allowed in the biomass function (lower bound = 0 and upper bound = 1e-9). Uptake of N-acetyl glucosamine (GlcNAc) was allowed in both gametocyte models since GlcNAc induces gametocytogensis[47]. For the ookinete model, the glucose uptake was constrained to 10% of the asexual stages glucose uptake rate since the mosquito gut is a glucose-rare environment[25].
P. falciparum 3D7 life cycle stage-specific RNA-Seq data was downloaded from SRA archive (SRP009370)[26]. SRA files were converted to fastq files using the sra-toolkit[48]. Tophat2[49] was used for the alignment (—library-type fr-unstranded) libraries. PICARD (http://broadinstitute.github.io/picard/) and samtools[50] were used for processing the aligned reads and HTSeq[51] was used to produce read counts (—stranded = no). The normalized read counts were then used to further constrain the stage-specific models (Fig B in S1 Text).
Validation in part was performed from growth rate predictions of overall biomass production rates in each life cycle stage of P. falciparum (Table F in S1 Text). The predictions showed an overall qualitative agreement with the experimentally observed growth phenotypes of the parasite during the asexual and sexual stages. Specifically, our models successfully predicted significant increase of the growth rate (FDR < 0.05) by 1.8 and 20 fold in the trophozoite relative to the early and late gametocyte stages, respectively (Table F in S1 Text). The ookinete model was the only malaria stage-specific model that was able to grow in absence of glucose (although the reduction in growth was 93%), which is in line with the glucose-rare medium in the mosquito gut where this stage develops[25].
Stage-specific model predictions were compared against differential gene expression (DEG) following a previously published workflow[52], outlined in Fig B in S1 Text. Briefly, differential gene expression analysis was carried out between every two stages and the lists of significantly differentially expressed genes (FDR < 0.05 and (> 75th or < 25th percentile of the log2 fold change in expression)) were later used for evaluation of stage—specific models’ predictions. The network flux states were sampled and significantly different reactions (FDR < 0.05 and (> 75th or < 25th percentile of the log2 fold change in reaction fluxes)) were identified following removal of loop reactions. The corresponding genes were selected using gene-protein-reaction relationships and were compared against the list of significantly differentially expressed genes.
Correlated reaction sets (co-sets) were calculated using the sampled steady state solution points for the iAM-Pf480 stage-specific models (COBRA toolbox[46] ‘identifyCorrelSets’ with a correlation cutoff threshold of 0.95). Only co-sets containing 3 or more reactions were labeled, since these co-sets generally represent transport of individual metabolites and not biochemical pathways per se. Sampled reaction fluxes in the pentose phosphate pathway were compared across the different stages and differential flux activity was acknowledged if the flux distributions were significantly different following multiple hypothesis correction, as previously described[52]. The modularity of the co-sets was assessed using the ratio of the mean size of co-sets divided by the maximum size of the co-sets for each stage. Voronoi plots were generated using TreeMap (v. 3.8.3) using the co-sets annotation (Table G in S1 Tables) and sampled flux distribution of each reaction in the corresponding co-set.
Genome-scale metabolic models were reconstructed for five Plasmodium species (P. falciparum ‘Pfal’, P. knowlesi ‘Pkno’, P. vivax ‘Pviv’, P. cynomolgi ‘Pcyn’, and P. berghei ‘Pber’). The details of the procedure for building Plasmodium multi-species genome-scale metabolic models are outlined in (Fig C in S1 Text).
|
10.1371/journal.pcbi.1003570 | Phylodynamic Inference for Structured Epidemiological Models | Coalescent theory is routinely used to estimate past population dynamics and demographic parameters from genealogies. While early work in coalescent theory only considered simple demographic models, advances in theory have allowed for increasingly complex demographic scenarios to be considered. The success of this approach has lead to coalescent-based inference methods being applied to populations with rapidly changing population dynamics, including pathogens like RNA viruses. However, fitting epidemiological models to genealogies via coalescent models remains a challenging task, because pathogen populations often exhibit complex, nonlinear dynamics and are structured by multiple factors. Moreover, it often becomes necessary to consider stochastic variation in population dynamics when fitting such complex models to real data. Using recently developed structured coalescent models that accommodate complex population dynamics and population structure, we develop a statistical framework for fitting stochastic epidemiological models to genealogies. By combining particle filtering methods with Bayesian Markov chain Monte Carlo methods, we are able to fit a wide class of stochastic, nonlinear epidemiological models with different forms of population structure to genealogies. We demonstrate our framework using two structured epidemiological models: a model with disease progression between multiple stages of infection and a two-population model reflecting spatial structure. We apply the multi-stage model to HIV genealogies and show that the proposed method can be used to estimate the stage-specific transmission rates and prevalence of HIV. Finally, using the two-population model we explore how much information about population structure is contained in genealogies and what sample sizes are necessary to reliably infer parameters like migration rates.
| Mathematical models play an important role in our understanding of what processes drive the complex population dynamics of infectious pathogens. Yet developing statistical methods for fitting models to epidemiological data is difficult. Epidemiological data is often noisy, incomplete, aggregated across different scales and generally provides only a partial picture of the underlying disease dynamics. Using nontraditional sources of data, like molecular sequences of pathogens, can provide additional information about epidemiological dynamics. But current “phylodynamic” inference methods for fitting models to genealogies reconstructed from sequence data have a number of major limitations. We present a statistical framework that builds upon earlier work to address two of these limitations: population structure and stochasticity. By incorporating population structure, our framework can be applied in cases where the host population is divided into different subpopulations, such as by spatial isolation. Our framework also takes into consideration stochastic noise and can therefore capture the inherent variability of epidemiological dynamics. These advances allow for a much wider class of epidemiological models to be fit to genealogies in order to estimate key epidemiological parameters and to reconstruct past disease dynamics.
| Genealogies can provide valuable information about the demographic history of a population because the demography of a population can dramatically shape the structure of a genealogy [1], [2]. For example, fluctuations in population size will shift the distribution of branching events, or coalescent times, over a genealogy relative to what would be expected for a population with a constant size [3]. Other aspects of a population's demographic history can also leave behind distinctive genealogical patterns. For example, the structuring of a population into different subpopulations can influence the topology of genealogies, which is often seen as clustering among individuals sampled from the same subpopulation [4]. These observations have led to great interest in statistical methods for inferring demographic trends and parameters from genealogies and given rise to the new field of phylodynamic inference [2], [5]–[8].
Most statistical methods for reconstructing the demographic history of a population from genealogies have been motivated by coalescent theory, which provides a probabilistic framework for relating the demographic history of a population to a genealogy of individuals sampled from that population [9], [10]. Critically, coalescent models provide a way to compute the probability of a given genealogy under a given demographic model. It is therefore possible to estimate parameters of a demographic model, such as population size, from a genealogy using likelihood-based inference methods. Extensions of this basic idea have been used to estimate changes in population size over time, for example by the Bayesian skyline methods available in the BEAST phylogenetic software package [11], [12]. Coalescent theory has also been extended to consider different forms of population structure, giving rise to structured coalescent models [13], [14]. Statistical methods that allow fitting of structured coalescent models to genealogies have the ability to estimate parameters relating to population structure, including migration rates between populations [7], [15].
Recent developments in phylodynamics have focused on developing models and statistical methods for more complex demographic scenarios, which have been largely motivated by the application of coalescent methods to pathogens like RNA viruses with rapidly changing population sizes. For example, coalescent models have been developed for populations where birth (i.e. transmission) rates vary over time [16], [17]. Importantly, the framework of Volz et al. [16] also considers the coalescent process in populations where transmission rates change over time in a nonlinear manner, as is often the case for epidemiological models like the well-known Susceptible-Infected-Recovered (SIR) model [18]. Coalescent models have also been developed for common epidemiological scenarios with population structure that alters the rate of coalescence in the population [19], but these models are limited to populations at equilibrium. Finally, Volz [20] presented a framework that brings together both complex population dynamics and population structure. This approach has great appeal as it generalizes coalescent models to allow both birth and migration rates to change over time as a function of the underlying population dynamics, which may be nonlinear and far from equilibrium.
Although recent advances with structured coalescent models have enabled the analysis of more complex epidemiological models, the statistical challenge remains of efficiently fitting stochastic population dynamic models to genealogies. These models can be extremely high-dimensional due to a large number of latent state variables for which we have no direct observations. In Rasmussen et al. [21], a particle filtering approach was used to marginalize out these latent variables by forward simulating population dynamic trajectories from the epidemiological model and then averaging over these trajectories to a compute a marginal likelihood. For unstructured models, adapting particle filtering methods to coalescent-based inference is relatively straightforward as the likelihood of a genealogy is simply a function of the simulated population dynamic trajectories. However, for structured models the likelihood also depends on the internal states of lineages in the genealogy, which may change over time as lineages move between populations [20]. The probable state of a lineage can only be calculated retrospectively conditional on the population's demographic history and the state of the lineage at the time of sampling. As we show below, these backward-time dependencies prevent the direct application of forward-time particle filtering methods to structured models.
We therefore present a new statistical approach for fitting stochastic population dynamics models to genealogies using the structured coalescent approach presented in Volz [20] using a modified particle filtering algorithm. This modified algorithm allows for efficient particle filtering under structured coalescent models where the probability that a lineage is in a certain population may depend on both the past dynamics of the population as well as future sampling of lineages. Using this algorithm, we can fit stochastic, nonlinear epidemiological models with essentially any form of population structure to genealogies as long as the model is Markovian. Because population structure arises naturally in many epidemiological models, we define population structure in a very broad sense and consider any model where the population of infected hosts is structured into different nonequivalent states and therefore lineages in different infected hosts do not necessarily have an equal probability of coalescing. This includes models with spatial structure, multiple stages of infection and models of vector-borne and other multi-host pathogens.
The paper has the following structure. First, we present the forward-time epidemiological models that we use as examples throughout the paper. Next, we review the framework first developed in Volz [20] for how coalescent models can be derived for a corresponding forward-time population dynamic model. We then describe how we can fit structured epidemiological models to genealogies given the corresponding structured coalescent model. The statistical method we describe combines MCMC methods with our particle filtering algorithm, and is a variation of the particle MCMC algorithm of Andrieu et al. [22]. Using simulated genealogies, we show that this algorithm can accurately reconstruct population dynamics in structured populations and obtain reliable estimates of epidemiological parameters such as transmission rates. We then apply our approach to the HIV epidemic in Detroit, Michigan in order to estimate stage-specific transmission rates and infer how prevalence and incidence have changed over the course of the epidemic. Finally, we explore under what conditions parameters relating to population structure can be inferred from genealogies and how factors such as sample size affect uncertainty in our estimates.
In this paper, we use epidemiological models to demonstrate how mechanistic population dynamic models can be fit to genealogies. More specifically, we will consider the type of Susceptible-Infected-Recovered (SIR) models widely used to study the transmission dynamics of infectious diseases [18], [23]. In SIR-type models, the host population is divided into different compartments depending on the host's state (e.g. susceptible or infected). For generality, we let be the vector that holds the number of hosts in each compartment at time , for example for the standard SIR model. For stochastic models, the state variables in are treated as random variables. We consider an epidemiological model to be structured if there is more than one class of infected host.
In this section, we consider formulating structured coalescent models for the type of structured epidemiological models just presented. As shown in Volz [20], thinking about population dynamic models as simple birth-death processes can be useful when deriving coalescent models that correspond to a given forward-time model. If we randomly sample individuals from a population and trace their ancestry back in time, then coalescent events in the genealogy will correspond to birth events in the population when both the parent and child lineages are ancestral to sampled individuals. While deaths may affect the overall population size, deaths can be ignored along lineages ancestral to sampled individuals because we know that a lineage could not have died out at an earlier time if it persisted to be sampled at some later time. For a structured population, we also must consider individuals transitioning between different subpopulations through migration events that occur independently of birth events, although for the type of models we will consider here a lineage can also transition between populations by being born into a different population than its parent.
The same birth-death-migration framework can be applied to pathogens if we assume that each infected host corresponds to a single individual in the pathogen population. In this case, births in the pathogen population occur at transmission events between hosts. Deaths in the population will correspond to recovery or mortality of infected hosts. If each infected host is represented by a single pathogen lineage, coalescent events in the genealogy will correspond to transmission events if both the infected host and the infector are sampled or give rise to descendent infections that are sampled. For structured epidemiological models, we also must consider a pathogen lineage transitioning among populations, or compartments in SIR-type models, independent of transmission events. For example, in the three-stage model, pathogen lineages can transition between different stages of infection. Here, we will refer to all transitions between states that occur independently of transmission as migration for generality. This allows many epidemiological models with some form of population structure to be thought of as a birth-death-migration process.
To formalize the birth process, we adopt the notation of Volz [20] and let be a matrix that specifies the birth rate of new lineages in the population at time , where , meaning that can be a function of the epidemiological parameters and the population state variables . Lineages may be in any one of states. The rate at which lineages currently in state give birth to lineages in state is given by the element . The rate at which migration, or transitions between states independent of birth events, occurs is given by another matrix . The rate at which lineages currently in state migrate to state is given by the element . We treat birth and migration as distinct processes because, as we will see, they affect the coalescent process in different ways since coalescent events can only occur at birth events but migration events can affect the probability of a particular lineage coalescing with another lineage. The total number of lineages in each state is given by a vector , such that gives the total number of individuals in the population in state at time . From here in, we drop the time indices and just refer to the matrices and or the vector , but emphasize that the rates in and and the population sizes in can be time-dependent.
We illustrate the and matrix notation by decomposing the three-stage and two-population SIR models presented above into their component birth and migration processes. For the three-stage model, we have(4)(5)In the matrix, births occur through transmission of the pathogen from any of the three stages of infection to susceptible individuals. Because all new infections begin in the early stage, only the leftmost column of the matrix has nonzero elements. The nonzero elements in the matrix correspond to migration between stages through disease progression from early to chronic and from chronic to AIDS.
For the two-population model, we have(6)(7)Because transmission events can move the pathogen within and between the two populations in either direction, all entries in the matrix are nonzero. The matrix has all zero entries because there is no migration between populations independent of transmission.
Before moving on, we note that for an infectious pathogen our coalescent models make the implicit assumption that coalescent events in the genealogy correspond to transmission events between hosts. In essence then, we are ignoring the within-host coalescent process and assuming that all infected hosts are represented by a single lineage. This implies that lineages immediately coalesce once in the same infected host, which may not be true for certain pathogens where multiple lineages can persist within a host for long periods of time. Nevertheless, in general our assumption that each infected host is represented by a single pathogen lineage will be valid as long as super-infection is rare and there is a strong bottleneck in the pathogen population at transmission events so that it is unlikely that more than one lineage is transmitted between hosts.
To fit a structured coalescent model to a genealogy, we need to compute the likelihood of the coalescent model given the genealogy. To compute this likelihood, we can partition the genealogy into any number of discrete time intervals. We label the time partitioned genealogy , where is the time of the first event in the genealogy and is the final event time going forwards in time (usually the terminal-most sampling event). Time points are chosen to correspond to the times at which events in the genealogy occur such as coalescent and sampling events. We can then further subdivide the genealogy into smaller intervals that correspond to the time steps used to simulate from the epidemiological model so that at any time point we have the state variables corresponding to that time. With the time partitioned genealogy , we can compute the likelihood over each interval in the genealogy, , and then take the product over all intervals to compute the total likelihood of the model given .
Computing the likelihood over a time interval requires us to first compute the probabilities that the lineages present in the genealogy did or did not coalesce within that time interval. The probability of a coalescent event in turn depends on the expected rate of coalescence under the model. This expected rate can be computed for a coalescent model with any arbitrary population structure using the formalism summarized above for the rates of birth in . As shown in Volz [20], the rate of coalescence for two lineages and is(8)where, for example, is the probability that lineage is in state . How these lineage state probabilities are computed is explained below. We can make intuitive sense of the coalescent rate in (8) by noting that is the total rate at which lineages in state give birth to lineages in state in the population and that is the probability that lineages and are the two lineages involved in a particular birth event. However, since we do not know the true states of and we must sum over all possible combinations of states for these two lineages.
The total rate of coalescence for all lineages present in the genealogy over an interval of time is then(9)
Given the rates of coalescence, we can then compute the likelihood over a time interval under the coalescent model. If the time interval does not end in a coalescent event, we have(10)Alternatively, if the interval does end in a coalescent event between two lineages and , we have(11)
As alluded to above, computing the coalescent rates requires us to compute the probability of each lineage in the genealogy being in each possible state. At the time of sampling, we may know the state of a lineage from information gathered from the infected host from which the sample was obtained. Alternatively, if we do not know the state of the host at the time of sampling exactly, we can assign prior probabilities to the lineage being in each state under a multinomial distribution. Either way, given the initial state or state probabilities at the time of sampling, we need to be able to compute the probability of the lineage being in each state at any point in the past.
Going backwards in time, the lineages transition between states at the rates given in the and matrices, which in turn depends on the population states and the parameters . Given these transition rates, we have a continuous time Markov process on a discrete state space along each branch. We can therefore use master equations to track how the lineage state probabilities change going backwards through time. In other words, we can write down differential equations for how the probability mass assigned to each state flows between states as we move into the past. As shown in Volz [20], the general form that these master equations take for any lineage and state is(12)where ; that is is the expected number of lineages in state in the genealogy at a given point in time. Further details on how the lineage state probabilities are computed and get updated at coalescent events are given in Text S1. For convenience, we introduce the notation to denote the lineage state probabilities for all lineages in the genealogy at time and to denote the complete mapping of lineage state probabilities onto the genealogy over the entire time partitioned genealogy .
The goal of phylodynamic inference for the type of models presented above will generally be to infer the parameters of interest from the genealogy along with the latent population state variables, such as the number of infected or susceptible hosts over time. In a Bayesian context then, we would like to infer the joint posterior density of the model parameters and the latent state variables . Up to a normalizing constant, this posterior density is given by(13)From (13), we see that this joint density can be factored into three parts: the coalescent likelihood which we outlined how to compute above; the prior density on the population state variables as defined by the epidemiological process model; and the prior density on the parameters . Although we may be able to compute each component individually and thereby the posterior probability of a given set of parameters and population states , the posterior density is not analytically tractable in general and we must resort to sampling from the posterior using MCMC methods.
However, it may be difficult or impossible to sample from complex, high-dimensional densities such as using standard MCMC methods. We could, for example, use a Gibbs sampler to iteratively sample from the conditional posterior densities of and any component of , but this strategy can be extremely inefficient owing to strong correlations among the parameters and the state variables, leading to slow MCMC mixing [28]. In Rasmussen et al. [21], a particle MCMC approach known as the particle marginal Metropolis-Hastings (PMMH) algorithm was therefore used to sample from the joint posterior density of and . The main motivation behind using the PMMH algorithm is that we can jointly update and together [22]. Each MCMC iteration, we first propose new parameter values and then run a particle filtering algorithm to get a numerical approximation of the posterior density of the latent state variables , which we refer to as . Particle filtering, also known as sequential Monte Carlo, provides a computational means of approximating high dimensional densities by providing samples (i.e the particles) distributed according to the desired density, and are often used in the context of nonlinear and non-Gaussian state space models [29]–[31]. We review how particle filters can be used to fit epidemiological models to genealogies in Text S1.
After running the particle filtering step in the PMMH algorithm, we can then sample a particle from to get a proposal for the latent state variables that is adapted to the parameters in . We can also use the particle filter to compute the marginal likelihood of by marginalizing out the state variables. Because we jointly accept and based on the marginal likelihood, we do not have to independently update , leading to a much more efficient MCMC sampler. Despite marginalizing out the latent state variables, the remarkable feature of the PMMH algorithm is it provides an exact (i.e. unbiased) approximation to the density of interest, . The PMMH algorithm is summarized in pseudo-code below.
We simulated mock genealogies under each model to test the performance of the PMMH algorithm before applying the method to real data. Mock genealogies were obtained by first forward simulating from the population dynamic model while tracking all infected hosts in the population and the parent-offspring relationships at transmission events. From the forward simulations, we could then trace the lineages of infected individuals backwards through time to obtain the true genealogy for a fraction of sampled lineages. All population dynamic simulations were performed using the tau-leaping algorithm so that the epidemiological dynamics included demographic noise [32].
The three-stage model was parameterized to reflect the natural history of HIV because we planned to apply our method to real HIV genealogies (see Table 1). We set the disease progression and AIDS death rate to values that give an average time between infection and death of about 10 years, consistent with observed patterns. The incidence scaling parameter was set to zero so that in the simulations there was a linear scaling between incidence and prevalence. The epidemic simulations were seeded with one early-stage infection at time zero and run for 37 years to reflect the timespan of the HIV epidemic in the U.S. To obtain mock genealogies from the complete infection trees, we sampled 200 individuals in the last six years of the epidemic to reflect the fact that most HIV sequences have been sampled in the recent past. For all parameters, we chose to use uniform priors over a wide range of biologically plausible values so that the choice of prior would have minimal influence on our estimates.
For the two-population model, we added seasonality to the model by seasonally forcing the base transmission rate using a sinusoidal forcing function, where(17)The strength of seasonality was the same in both populations but we allowed to differ between the two populations to get asynchronous dynamics between populations. The values of all fixed parameters in the model are also shown in Table 1. For the genealogies, 120 infected hosts were randomly sampled over time with sampling effort proportional to disease prevalence in each population. For the two-population model, we fixed the initial conditions for the number of susceptible and infected hosts in each population.
For the simulation experiments, we wished to compare estimates obtained by fitting stochastic models using the PMMH algorithm against estimates obtained by fitting deterministic models. To fit deterministic models, we used a Metropolis-Hastings sampler where, whenever new parameters are proposed, the likelihood of the genealogy under the new parameters is computed by conditioning on a deterministic trajectory of the state variables simulated from the model using these new parameters.
We applied our method to a set of HIV-1 partial pol sequences collected from men who have sex with men (MSM) in the metropolitan area of Detroit, Michigan. The dataset contained 437 HIV-1 subtype B sequences which were originally collected for drug resistance testing between 2004 and 2011. More information about this dataset can be found in Volz et al. [26]. Data were anonymized by staff at the Michigan Department of Community Health before being provided to investigators. Because this research falls under the original mandate for HIV surveillance and was de-identified, it was classified as human subjects research but was exempt from further IRB review.
We reconstructed time-scaled genealogies from the HIV sequences in BEAST using a relaxed molecular clock [33]. All sequences identified as likely recombinants were removed from the alignment prior to the analysis. Tips in the genealogy corresponding to sampled infected individuals were assigned prior probabilities of being in each infection stage based on the time since infection estimated from CD4 cell counts and genetic diversity within the host [34].
From the HIV genealogies, we estimated the transmission rates , and as well as the incidence scaling parameter . All other parameters were fixed at the values given in Table 1. Rather than estimate initial conditions, the time of the initial introduction of HIV into Detroit was estimated, at which point the epidemic was seeded with one early-stage infection in a completely susceptible population. All priors on the parameters were uniform. For the time of initial introduction the prior was truncated at 1973 as a lower bound and the root time of each tree as an upper bound. To ensure our phylodynamic estimates of HIV incidence were reasonable, we compared our estimates against incidence back-calculated from Michigan Department of Community Health surveillance data using the method of Yan et al. [35].
For all results shown in this paper, the PMMH algorithm was run for at least 100,000 iterations or until the MCMC fully converged. For the Metropolis-Hastings step, we chose a multivariate normal proposal density for , which can take into account the correlations among different parameters by optimizing the covariance parameters that specify the density.
For the particle filter, we found that using a small number of particles () was sufficient. Running the particle filter with a small number of particles tends to increase the error, or variance, in the marginal likelihood estimates. However, this error will not affect inference as long as the marginal likelihood estimates are not systematically biased because the error in the estimates will get averaged out in the encompassing MCMC algorithm. Nevertheless, with too few particles we run the risk of the MCMC getting stuck at erroneously high values of the likelihood. Our choice of was therefore a compromise between minimizing the error in the marginal likelihood estimates and the time taken to run the particle filter. Resampling within the particle filter was done by multinomial sampling with replacement. Resampling times were chosen to minimize the variance in the marginal likelihood estimates and were usually placed around coalescent events, as most of the variation in particle weights arises at coalescent times.
The PMMH algorithm was implemented in the software package PHYLter and Java source code is freely available at http://code.google.com/p/phylter/. Running the PMMH algorithm for 100,000 iterations using the simulated HIV genealogies took approximately 10 hours (0.36 s per iteration) on a 3.4 GHz Intel i7 processor without any parallelization across cores. The most computationally intensive component of the algorithm is computing the lineage state probabilities, which involves numerically solving the master equations for each lineage in the genealogy and has a time complexity of , where is the number of possible lineage states. On the other hand, run times scale linearly with the number of particles and lineages in the genealogy. Thus, the efficiency of the algorithm is mainly limited by the number of states in the model.
Before applying the PMMH algorithm to genealogies reconstructed from real data, we ran extensive simulations to ensure that we could accurately recover epidemiological parameters and population dynamics from mock genealogies. We simulated 100 stochastic realizations of an epidemic from the three-stage model, keeping track of the underlying infection tree so that we could obtain the true genealogy for a fraction of sampled lineages. From the simulated epidemic dynamics, we can see that demographic stochasticity generates considerable variation in when the epidemic begins and peaks (Figure S1). Even with this variability, we accurately inferred stage-specific prevalence and transmission rates from the mock genealogies using the PMMH algorithm (Figure 1). The 95% credible intervals generally contained the true prevalence for all three stages of infection (Figure 1A). We were also able to estimate the stage-specific transmission rates associated with each stage of infection (Figure 1B–D), even though there were strong correlations among the different transmission rates as seen in the pairwise joint posterior densities (Figure 1E–G). Overall, out of all 100 simulations, the 95% credible intervals contained all three transmission rates 94 times, while the posterior coverage was greater than 95% for each parameter individually. In contrast, when we fit deterministic models to the same set of genealogies, the credible intervals contained the true parameters only 79% of the time. The PMMH algorithm therefore appears to give reliable estimates of parameters and epidemiological dynamics and outperforms deterministic methods when stochasticity plays a role in the epidemic dynamics.
Given that we were able to reliably estimate transmission parameters and prevalence in our simulation study, we next applied the method to HIV genealogies reconstructed from sequences collected in Detroit, Michigan. A critical question in HIV epidemiology is to what extent transmission during the early stages of infection contributes to overall HIV incidence. Transmission during early infection may influence the effectiveness of interventions based on antiretroviral treatment in limiting the epidemic [36], [37]. If most new cases of HIV result from recently infected individuals, then prevention strategies that rely on treating diagnosed individuals, who are likely in later stages of infection, will directly prevent few transmissions. Thus, the transmission rate from early HIV infections (EHI) is a key parameter of great interest, although difficult to measure directly from traditional surveillance data. Phylogenetic studies of HIV have used the high degree of clustering and short branch times within these clusters to argue for a high EHI transmission rate [4], [38]. However, clustering alone cannot be taken as definitive evidence for high EHI transmission as similar patterns can arise simply from epidemic transmission dynamics [26]. In this section, we demonstrate that our inference framework can be used to estimate the EHI transmission rate and the number of new HIV infections attributable to EHI from HIV genealogies using models that explicitly consider HIV's transmission dynamics, as well as the stochastic nature of the epidemic dynamics.
Time-scaled genealogies were reconstructed using BEAST from HIV-1 partial pol sequences isolated from men who have sex with men (MSM) in the metropolitan area of Detroit. A representative genealogy randomly sampled from the BEAST posterior is shown in Figure S2. We then fit our three-stage SIR model to 10 genealogies sampled from the BEAST posterior to take into account uncertainty in the genealogy. From these genealogies, we estimated the transmission rate for each stage, including the EHI transmission rate, along with the stage-specific dynamics of prevalence and the incidence (i.e number of new cases) attributable to each stage over the course of the epidemic.
Parameters estimated from the representative HIV genealogy are shown in Figure 2 and estimates from all 10 genealogies are given in Table 2. We estimated that transmission rates are higher during the early and AIDS stages than during the chronic stage, as expected from previous studies [39]–[41]. The transmission rate from EHI is about 20 times higher than during the chronic stage and about five times higher than during the AIDS stage (Figure 2A–C). We also found evidence for a nonlinear dependence of incidence on prevalence, quantified through the incidence scaling parameter . Although estimated values of are small, the posterior density is clearly centered away from zero, indicating that incidence scales nonlinearly with prevalence (Figure 2D). Overall, parameter estimates were largely consistent across genealogies, although there was considerable variation in the time of initial introduction of HIV into Detroit estimated from different trees. This is likely attributable to the large amount of variation in the root times inferred for different trees, as we inferred earlier times of introduction from trees with earlier root times.
Stage-specific HIV prevalence inferred from the genealogies shows a predictable transition from most infections being in the early stage at the beginning of the epidemic to most infections being in the chronic or AIDS stages later in the epidemic (Figure 3A). This is expected given the longer duration of the chronic and AIDS stages. In general, our phylodynamic estimates of the epidemic dynamics closely track HIV incidence imputed from surveillance data from the beginning of the epidemic through the peak (Figure 3B). While our phylodynamic estimates do not capture the fluctuations in incidence that occur after 1990, there was nothing in our model that would allow us to reproduce this pattern, which likely results from complex changes in HIV treatment and behavioral changes [34]. Although there was also considerable variability in the population dynamics inferred from different genealogies, this variation occurs primarily during the early stages of the epidemic (Figure 3C). Again, this appears to be associated with uncertainty in the root times of trees; dynamics inferred from trees with earlier root times show an earlier rise and peak in incidence. After the epidemic peaks, the incidence estimated from different trees seems to converge on similar values.
Estimates of incidence attributable to each stage show that EHI contributed to most new infections at the beginning of epidemic when EHI prevalence was high (Figure 3B). After the epidemic peak, infections arising from EHI remains high proportional to EHI prevalence, consistent with the higher transmission rate we estimated for EHI. In the late 2000's, we estimated that between 40 to 50% of all new infections arise from EHI, indicating that early stage infections still play a major role in driving HIV transmission. These large estimates for number of new infections arising from EHI are consistent with the phylodynamic estimates of Volz et al. [34], who fit a more complex but deterministic epidemiological model to the same set of HIV sequences.
While our results for the three-stage model suggest that the PMMH algorithm works effectively and can be used to estimate key epidemiological parameters like HIV transmission rates, we were also interested in how much information genealogies contain about the structure of populations in general. To explore this question, we used the two-population model presented in (3), for which we can tune the strength of population structure by altering the mixing rate between populations. Mock genealogies were simulated under three values of : low (0.01), medium (0.05) and high (0.2). At , for example, about one in every one hundred transmission events occurs between populations. For all three values of , we were able to accurately infer the epidemiological parameters of interest and the population dynamics from the simulated genealogies (Figure 4 & Figure S3). While we can easily estimate under all three demographic scenarios, the posterior densities become skewed towards increasingly high values of as mixing increases between the populations (Figure 4A–C). This indicates that it may be very difficult to obtain precise estimates of or other parameters pertaining to population structure when populations are only weakly structured.
We can visually explore how much information a genealogy contains about population structure and pathogen movement by comparing the true lineage states to the computed lineage state probabilities. In Figure 5A–C, the true state of each lineage over time is mapped onto the genealogies. For ease of viewing, we only display a representative subtree of each genealogy. As expected, under low mixing lineages change states very slowly leading to a high degree of clustering among lineages sampled from the same population, whereas under high mixing lineages move rapidly between states and there is little clustering. We can then compare the true lineage states with the state probabilities computed under the median posterior values of the estimated parameters (Figure 5D–F). When is low, the state of the lineages at the time of sampling is highly informative about the state of the lineage going into the past. However, when we increase to 0.05, the state of the sampled lineages is less informative about the past states and we can see that the lineage state probabilities fluctuate seasonally according to the asynchronous dynamics between populations. When is high, the lineages move between states so rapidly that there is high uncertainty in the lineage states over the entire tree. This loss of information regarding the lineage states is readily observed by considering how the entropy, or uncertainty, in the lineage states changes going backwards in time (Figure 5G–H).
Visualizing the flow of information along the lineages in the trees shows how uncertainty in parameters like depends on how rapidly information about the lineage states decays. When is low, lineages remain in the same state long enough that once a coalescent event is reached, information about the probable state of the lineages is still present. In this case, the probable states of the coalescing lineages provides additional information about the transmission event with respect to whether the transmission event occurred within or between populations. By combining information from coalescent events across the entire tree, we can then estimate the rates at which transmission occurs within and between populations. However, if all information about the past lineage states is lost before lineages coalesce, the observed coalescent events will no longer be informative about whether transmission occurred within or between populations and therefore parameters like will be difficult to precisely estimate.
The preceding observations about uncertainty in lineage states suggest that it may be possible to estimate more precisely if we increase the number of sampled lineages. Increasing the sampling fraction will also increase the coalescent rate among lineages, thereby increasing the probability of lineages coalescing before all information about their probable state is lost. To test this idea, we simulated genealogies under the same three values of but varied the sample size. With a sample size of 100, the same as used above, we see that the likelihood is peaked around the true value of when mixing is low but the likelihood profile is fairly flat when mixing is high (Figure 6A–C). Increasing the sample size to 500 resulted in more curved likelihood profiles but the likelihood remains relatively flat with high mixing (Figure 6D–F). Doubling the sample size again to 1,000, the likelihood profiles show significant curvature for all values of (Figure 6G–I). This suggests that while the sample size does play a significant role in determining whether parameters like can be precisely estimated from genealogies, extremely large sample sizes may be required to estimate parameters pertaining to population structure when the population is only weakly structured.
The approach outlined in this paper allows for structured, stochastic epidemiological and other population dynamic models to be fit to genealogies in order to jointly infer past population dynamics and model parameters. We believe this to be an important step forward in the field of phylodynamics because many populations are structured in ways that could bias estimates of demographic parameters when using coalescent-based methods if population structure is not properly taken into account. Furthermore, unlike earlier methods for fitting structured coalescent models to genealogies (e.g. [7], [15]), our framework can accommodate non-equilibrium and nonlinear population dynamics and allows birth and migration rates to vary over time. We can also include stochasticity in our models when fitting them to data obtained from real populations, which may behave very differently than what would be expected under deterministic models. We can therefore fit the type of mechanistic population dynamic models typically used by epidemiologists and ecologists, which often include population structure, to genealogies.
As we have shown, fitting stochastic population dynamic models to genealogies through a structured coalescent model poses some challenges to statistical inference not normally dealt with in the statistical literature on fitting generic state space models to observational data. Under our structured coalescent models, the probability of a genealogy depends conditionally on both the population state variables as well as the states of individual lineages over time. However, going backwards in time, the probability that a lineage is in a certain state can strongly depend on the state that the lineage was sampled in at some future point in time. Particle filtering methods, which are widely used to fit state space models to other sources of data, can perform very poorly under these circumstances because the state of the system, in this case the lineage states, can depend strongly on the future states of the system. One strategy we initially tried was therefore to use a Gibbs sampling approach to iteratively sample from the conditional posterior densities of the population state and lineage state variables in independent steps to avoid the problem of having both forward and backward time dependencies in the model. Unfortunately, we found that such a Gibbs sampling strategy can be very inefficient and suffer from extremely poor MCMC mixing when there are strong correlations among the parameters and the lineage states. For example, in our two-population model, the mixing parameter controls how rapidly lineages move between states and is thus highly correlated with the lineage states. If we update conditional on our current lineage states, the proposed value of will need to be very close to the current value in order for the proposal to have high enough probability to be accepted conditional on the current lineage states. We therefore explore a potentially very large parameter space taking only small steps at a time.
Given these issues, we decided to use a modified version of the PMMH algorithm originally proposed by Andrieu et al. [22]. In this approach, we simply propose new parameter values each MCMC iteration and then run the particle filter to numerically integrate over the population state variables. To make the particle filtering algorithm as efficient as possible within each MCMC step, we allow for resampling by first weighting the particles according to the expected lineage state probabilities. Once we have run the particle filter forwards in time, we can then compute the true lineage state probabilities backwards in time and apply an additional round of importance sampling to correct for any bias introduced by using the expected lineage state probabilities. With the true lineage state probabilities of each particle, we can compute the coalescent likelihood of the genealogy while summing over all possible lineage states. We can therefore integrate over both the unobserved population state variables and the lineage state variables when computing the marginal likelihood of the parameter proposal. We thus have an efficient MCMC algorithm for sampling from the posterior density of the parameters without having to design independent proposals for the population states or the lineage states. The PMMH sampler therefore has a major practical advantage over other MCMC approaches that can be easily quantified. For the models considered in this paper, the PMMH algorithm typically converged in less than 100,000 iterations whereas for the Gibbs sampler we could run millions of MCMC iterations and still not converge. The efficiency of this approach will hopefully make it possible to also consider phylogenetic uncertainty in the future by sampling genealogies in addition to epidemiological parameters in the MCMC algorithm.
Whether or not the type of coalescent models considered here are appropriate for a particular pathogen is another important issue. The coalescent models assume that each infected host corresponds to a single pathogen lineage. If this were indeed always the case then coalescent events in the genealogy would always correspond to transmission events in the population. In reality, coalescent events will not occur instantaneously at transmission events but at some time before the actual transmission event because there will be a waiting time between when a lineage is transmitted and when it coalesces with another sampled lineage in the host. How closely the actual transmission event corresponds in time with the coalescent event will likely depend on the within-host dynamics of the pathogen [42]. For chronic viral infections like HIV where multiple lineages can persist within a given host for months or years, this may result in a large discrepancy in the timing of transmission and coalescent events. Nevertheless, a simulation study using a realistic distribution of within-host coalescent times for HIV found that the difference in timing between coalescent and transmission events was not sufficient to bias estimates of epidemiological parameters [34]. This may be due to the fact that a large fraction of HIV transmissions are due to recently infected individuals, in which case the within-host coalescent event cannot have occurred very long before the actual transmission event. A more principled approach to pursue in the future may be to impute the actual times of transmission conditional on the time of the coalescent events using information about within-host population dynamics. For example, additional information about pathogen population sizes over the course of a typical infection could provide an informative prior on waiting times between transmission events and coalescent events within hosts.
Another possible violation of the coalescent model occurs if sampled individuals have descendants that are themselves sampled, which can occur when samples are collected serially over time. The coalescent model implicitly assumes that when a new lineage is sampled, that lineage is sampled from a different host than any other lineage already in the genealogy. However, if a lineage is sampled from a host that has other sampled descendant lineages in the genealogy, then this results in a coalescent event in the tree that does not correspond to a transmission event in the population. A similar problem would arise if we unwittingly sampled more than one lineage from a single infected host. However this is likely to occur only if sampling is dense relative to prevalence over time. For example, if sampling is dense at the beginning and the end of an epidemic, then with a high probability hosts sampled at the beginning of the epidemic will likely have sampled descendants at the end of the epidemic. We acknowledge that the coalescent models used in this paper cannot adequately handle these types of situations, although for the HIV analysis it is unlikely that this is a serious problem seeing as all sequences were sampled in the recent past when prevalence was high. In cases where this is likely to be a serious problem, it may be worth developing metapopulation coalescent models, such as those introduced by Dearlove and Wilson [43], that allow hosts to be infected by more than a single lineage.
As our application to HIV showed, the PMMH algorithm allowed us to infer key epidemiological parameters like stage-specific transmission rates directly from genealogies. However, in the case of HIV, individuals stay in the same stage of infection for long periods of time relative to the timescale of the epidemic. The stage of infection of sampled individuals is therefore highly informative about the state of the lineage going into the past. Our experience with HIV may therefore not be representative of our general ability to infer parameters pertaining to pathogen transmission or movement in structured populations. In fact, our simple two-population SIR model revealed certain conditions under which it may be inherently difficult to estimate parameters relating to population structure. When lineages move between states rapidly due to transmission or migration any particular lineage is likely to have changed states multiple times before a coalescent event is reached, leading to high uncertainty about the state of lineage over the majority of the genealogy. This is somewhat analogous to the problem of site saturation in phylogenetic inference, where multiple transitions at a particular site along branches can render that site phylogenetically uninformative [44]. In the case of rapid transition rates among population states, observing the state of lineages at the time of sampling offers little or no information about the structure of the population because all information about the state of the lineage is quickly lost. Under these circumstances, it will be difficult to precisely estimate migration rates or other parameters relating to population structure from genealogies as we saw from the likelihood profiles of the mixing parameter in the two-population model, although it may be possible with many samples or a large sample fraction. This echoes earlier work on inference with structured coalescent models, where researchers have found it difficult to estimate migration rates from genealogies even without the complication of complex population dynamics [7], [45].
Although it may not always be possible to precisely estimate parameters relating to population structure from genealogies, we can imagine several cases in which the ability to fit mechanistic epidemiological models to genealogies that include population structure may be extremely useful. For example, our methods could be used to fit spatially structured models to genealogies of samples collected in different locations and could potentially complement recently developed phylogeographic methods that consider spatial structure but do not generally take into account local population dynamics at any particular location [46], [47]. For instance, incorporating both spatial and temporal dynamics could be important when the structure of a population is not static but changes over time due to changes in migration rates, which themselves may vary due to non-stationary population dynamics across locations. Our approach can also be applied in cases where sampling effort is distributed unevenly among populations so that the assumption of random sampling in unstructured coalescent models has obviously been violated. In this case, structured coalescent models can be used to control for non-random sampling as long as sampling is random within the subpopulations defined in the coalescent model. Finally, our methods can be applied to multi-host or vectored pathogens where lineages can move among different host or vector species. As shown in Rasmussen et al. [48] for the case of dengue, including the dynamics of both the host and vector populations in coalescent models may be necessary in order for population dynamics inferred from genealogies of vector-borne pathogens to be accurate.
We end by noting that the methods presented here can be used to fit epidemiological models to genealogies as well as other sources of data simultaneously. For example, we previously showed how unstructured epidemiological models can be fit to a genealogy and a time series of case reports simultaneously and it would be straightforward to extend the methods presented here to include time series or other observational data [21]. This could be especially helpful when certain parameters or aspects of the dynamics are difficult to infer from one data source but for which an alternative data source could be highly informative. For example, case report data may be aggregated over different subpopulations obscuring some of the heterogeneity present in the population but could be revealed by also considering information present in a genealogy. Consolidating data sources in this way will likely play an important role in epidemiological modeling in the future, especially as molecular sequence data become increasingly available and phylodynamic methods become integrated into modern epidemiology.
|
10.1371/journal.pntd.0002541 | Severe South American Ocular Toxoplasmosis Is Associated with Decreased Ifn-γ/Il-17a and Increased Il-6/Il-13 Intraocular Levels | In a cross sectional study, 19 French and 23 Colombian cases of confirmed active ocular toxoplasmosis (OT) were evaluated. The objective was to compare clinical, parasitological and immunological responses and relate them to the infecting strains. A complete ocular examination was performed in each patient. The infecting strain was characterized by genotyping when intraocular Toxoplasma DNA was detectable, as well as by peptide-specific serotyping for each patient. To characterize the immune response, we assessed Toxoplasma protein recognition patterns by intraocular antibodies and the intraocular profile of cytokines, chemokines and growth factors. Significant differences were found for size of active lesions, unilateral macular involvement, unilateral visual impairment, vitreous inflammation, synechiae, and vasculitis, with higher values observed throughout for Colombian patients. Multilocus PCR-DNA sequence genotyping was only successful in three Colombian patients revealing one type I and two atypical strains. The Colombian OT patients possessed heterogeneous atypical serotypes whereas the French were uniformly reactive to type II strain peptides. The protein patterns recognized by intraocular antibodies and the cytokine patterns were strikingly different between the two populations. Intraocular IFN-γ and IL-17 expression was lower, while higher levels of IL-13 and IL-6 were detected in aqueous humor of Colombian patients. Our results are consistent with the hypothesis that South American strains may cause more severe OT due to an inhibition of the protective effect of IFN-γ.
| Ocular toxoplasmosis (OT), due to protozoan parasite Toxoplasma gondii, is a potential complication of both acquired and congenital infection, leading to visual impairment in numerous countries and being responsible for 30 to 50% of uveitis cases in immunocompetent individuals. In this study we confirmed the presence of more severe ocular toxoplasmosis in a tropical setting of Colombia, when compared to France. The main hypothesis for these clinical differences is based on the idea that severe disease in humans may result from poor host adaptation to neotropical zoonotic strains of T. gondii Indeed, our results are consistent with the hypothesis that South American strains may cause more severe OT due to an inhibition of the intraocular protective immune response.
| Infection with the protozoan parasite Toxoplasma gondii is a leading cause of visual impairment in numerous countries, being responsible for 30 to 50% of uveitis cases in immunocompetent individuals [1]. Ocular toxoplasmosis (OT) is a potential complication of both acquired and congenital toxoplasmosis [2]. The incidence of ocular toxoplasmosis has been estimated in Colombia (Quindio region) to be of three new episodes by 100 000 inhabitants by year [3], while in British-born patients it has been estimated to be 0.4 cases per 100,000 population per year and the lifetime risk of disease to be 18 cases per 100,000 population [4].
In a Colombian study, 5.5% of the population in the province of Quindío exhibited retinochoroidal scars resulting from a postnatally acquired infection, with 20% of this group presenting reduced visual capacity. [3], [5]. In a retrospective study on uveitis conducted in 693 Colombian patients, 417 of whom had a definitive diagnosis, toxoplasmosis was the most frequent cause with 276 cases (39.8%) followed by idiopathic uveitis and toxocariasis [6].
Some differences between South American and European clinical case series were observed in terms of congenital transmission rates, probability of symptoms in congenital OT [7], [8], severity of ocular inflammation [9] and intraocular specific antibody levels [10]. However, no comparative clinical and biological studies have been performed yet in patients from both continents with laboratory-confirmed OT.
The population structure of T. gondii in North America and Europe includes three highly prevalent clonal lineages, Types I (haplogroup 1, Clade A), II (Haplogroup 2, Clade D), and III (haplogroup 3, Clade, C) which differ greatly in virulence in the mouse model. The vast majority of human and animal infections are caused by the relatively avirulent Type II strains. In contrast, heterogeneous atypical genotypes of T. gondii are associated with severe infections in humans in South America. They belong to various haplogroups: 4, 5, 8 10 and 15, Clade F [11], [12][13]. The high genetic diversity of Toxoplasma strains in the tropical zone of the Americas may partly explain why congenital toxoplasmosis is more symptomatic in South America than Europe, as was demonstrated in cohorts of congenitally infected children from different continents [8], [14], [15]. A comparative prospective cohort study of congenitally infected children in Brazil and Europe found that Brazilian children displayed eye lesions that were larger, more numerous, and more likely to affect the central part of the retina responsible for acute vision [7]. Anecdotal clinical cases were also reported in the literature, such as a severe atypical bilateral retinochoroiditis in a Brazilian patient, caused by a highly divergent, non-archetypal T. gondii strain [16].
Given the markedly different population structure of T. gondii in Europe and South America, it is relevant to study the implications of this diversity on human pathogenesis [17]. Therefore, we conducted a multicenter case series study in order to compare the different clinical and immunological characteristics between Colombian and French patients, collecting the same data and performing the same laboratory assays in patients with biologically confirmed OT. The findings were related to Toxoplasma strain genotyping and peptide-based strain serotyping in our patients.
We collected data from consecutive patients who consulted at the Departments of Ophthalmology at Strasbourg University Hospital (France) and Quindío University Health Center (Armenia, Colombia) between August 2008 and August 2010. Both departments were tertiary-level centers able to perform anterior chamber paracentesis. For both patient populations, a complete ocular examination was conducted, including best-corrected Snellen visual acuity, slit-lamp biomicroscopy, tonometry, and indirect ophthalmoscopy. The clinical diagnosis of OT was based on criteria previously described by G. Holland [6], [18]. Screened patients with clinically suspected OT and seropositive for anti-Toxoplasma immunoglobulin G (IgG) antibodies were subsequently submitted to biological investigations to assess the local presence of Toxoplasma DNA and/or the intraocular antibody synthesis [19] to confirm OT.
Ethics Committee/Institutional Review Board (IRB) approval were obtained from Hôpitaux Universitaires de Strasbourg (PHRC 2007/3964) and Quindio University (ACT 14, 2008/23-06). Written informed consent was obtained from all subjects.
We analyzed the clinical characteristics of 19 French and 23 Colombian patients with active uveitis and biologically confirmed OT. Patients who were immunocompromised, suffered from other ocular infections, or received local or systemic anti-Toxoplasma treatment for active uveitis, were excluded. An assessment of the inflammation level and anatomic classification of uveitis was carried out according to the criteria proposed by the International Uveitis Study Group (IUSG) [20]. The size of the retinochoroidal lesions was measured in disc-diameters (dd).
Paired samples of aqueous humor and serum were obtained from each subject at the time of clinical diagnosis for laboratory analysis. The Colombian samples were stored locally at −80°C and then shipped together on dry ice to Strasbourg for laboratory analysis. Aqueous humor samples (100–150 µL) were collected through anterior chamber paracentesis and stored, along with serum samples, at −80°C until analysis. The diagnosis of OT was first confirmed by real-time PCR detection of Toxoplasma DNA [21]. Positive PCR results were quantified using a standard curve with serial 10-fold dilutions from a calibrated suspension of T. gondii RH-Strain DNA. For PCR negative patients, immunoblot (IB) was performed in order to detect intraocular synthesis of Toxoplasma-specific antibodies (LDBIO Diagnosis, Lyon, France). If both PCR and IB were unconclusive, a modified Goldmann-Witmer test was used to prove intraocular specificantibody synthesis [22].
The Bio-Plex Human 27-Plex Cytokine Panel assay (Bio-Rad, Marne-la-Coquette, France) was used according to the manufacurer's recommendations to measure cytokine and chemokine levels in aqueous humor. The assay plate layout consisted in a standard series in duplicate (1 to 32 000 pg/mL), four blank wells and 20 µL duplicates of AqH samples, diluted to 50 µL with BioPlex Human serum diluent [23]. A set of Toxoplasma seropositive cataract patients were used as control, 9 Colombian and 10 French. Data were analyzed with Bio-Plex Manager TM software V1.1.
DNA extraction for genotyping analysis was performed directly on ocular fluid samples and indirectly on infected cell cultures for six reference strains. GT1, PTG, and CTG strains were selected as reference Types I, II, and III strains, respectively. TgCtCo02, TgCtCo05, and TgCtCo07 strains were selected as reference Colombian strains [24], [25]. T. gondii DNA samples were subjected to genotyping analysis with 15 microsatellite markers in a multiplex PCR assay, as described elsewhere [26].
Serotyping of Toxoplasma infections was performed using 5 polymorphic synthetic peptides derived from the T. gondii dense granule proteins (GRA), GRA6 and GRA7. This test detects the presence of strain specific antibodies raised against Type II or non-Type II GRA6/7 alleles in patients infected with Type II or non Type II (NE-II) parasites respectively, as previously described [14], [27]. Briefly, the ELISA results presented are an optical density (OD) index obtained by dividing the OD value at 405 nm for each of the 5 serotyping peptides by the mean of the OD readings for the 2 control peptides. Threshold values are determined by averaging the normalized OD ratio from 100 seronegative French samples and adding 2 standard deviations, above which normalized values are considered positive. Obtained results are divided in four populations depending on their reactivity to the 5 peptides: I/III, ATYP, no reactivity (NR), and II [28]. I/III, ATYP and NR are considered as NE-II [14]. Sera from pregnant women, tested Toxoplasma seropositive in our laboratories, were used to assess the Toxoplasma serotype in a larger population from each country, 45 serum samples from Colombia and 100 from France.
Mann-Whitney test followed by Dunn's Multiple Comparison test was applied for comparison of clinical and laboratory characteristics for French and Colombian patients with confirmed active ocular toxoplasmosis (P values<0.05 were considered statistically significant; Stata software, College Station (Tx) USA). Fisher's exact test was used to compare diagnostic performances of IB and PCR as well as the serotype prevalence. Wilcoxon matched-pairs signed rank test was performed to compare IB patterns. Mann-Whitney test was used to compare intraocular parasite loads (P values<0.05 were considered statistically significant. Kruskal-Wallis test followed by Dunn's Multiple Comparison test were applied for comparison of cytokine and chemokine levels in aqueous humor between control and OT populations in both countries (P values<0.05 were considered statistically significant) (GraphPad Prism, La Jolla, CA, USA).
The clinical findings for OT patients are summarized in Tables 1 and S1. Statistically significant differences between groups were found for eight parameters, being higher in Colombian patients in all cases: i) time between consultation and anterior chamber paracentesis (p = 0.02); ii) size of active lesions (p = 0.04); iii) unilateral macular involvement (p = 0.001); iv) unilateral visual impairment (p = 0.04); v) vitreous inflammation (p = 0.00001); vi) percentage of patients with synechiae (p = 0.04); vii) vasculitis (p = 0.04) and viii) bilateral involvement (p = 0.04). In addition, there was a trend towards higher values for the Colombian patients regarding the number of lesions, number of recurrences, and intraocular pressure (IOP), although these differences were not statistically significant. We conducted a stratified analysis in order to exclude the influence of time before anterior chamber paracentesis as a possible cause of the differences in clinical findings. We compared early (<20 days after symptom onset) and late consultations (>20 days after symptom onset). As shown in Table 2 and supplementary figure 1, most significant clinical differences between the populations were also visible when comparing only the early-consultant groups.
In Colombians, aqueous humor samples revealed the presence of T. gondii DNA in 11 out of 23 samples (47.8%). In French patients, T. gondii DNA could be detected in aqueous humor samples of 7 out of 19 patients (36.8%). This difference was not statistically significant. In contrast, parasite loads in aqueous humor were significantly higher in Colombian patients, 4.53 parasites ± 2 per 100 µL versus 0.35±0.13 parasites per 100 µL (p = 0.0006) (Figure 1). Aqueous humor samples from all French patients and 14 Colombian patients had an insufficient amount of T. gondii DNA for genotyping analysis. Only 9 Colombian ocular fluid samples were submitted for multilocus PCR-DNA sequence genotyping analysis. Six had unsuccessful PCR amplification for all 15 tested markers due to low T. gondii DNA concentration. The genotype of one clinical sample (case COL-#6) was closely related to a Type I strain, but harboring unique alleles at three MS loci, M102, N83 and AA, using 15 amplified markers (Table 3). Of note, the genotype of a reference Colombian isolate (TgCtCo07) collected from a cat in 2005 was also Type I-like, suggesting that Type I-like strains may not be uncommon in animals and humans in Colombia. The genotypes of the other two clinical samples (cases COL-#26 and COL-#38) could not be fully determined, with only four and five successfully amplified markers, respectively. However, the results of the amplified markers showed that both genotypes were different from the Type II or III strains, which are common in North America and Europe. They present a majority of Type I alleles (case COL-#26), like TgCtCo07 but distinct at the N61 marker, and a combination of Type I, III, and atypical alleles (case COL-#38), like TgCtCo02 and TgCtCo05, but again distinct at the N60 and N82 genetic markers.
IB detected local antibody production in 19/23 Colombian (82.6%) and 13/19 French (68.4%) patients (not significant). However, a significant difference was observed in number of bands and their recognition pattern of Toxoplasma proteins (p<0.0001) (Figure 2). Specific proteins were recognized in 3.3% to 63.3% of Colombian patients and 3.8% to 53.8% of French patients. Colombian patients recognized most frequently a 62 kDa protein, observed in 63.3% of patients. In French patients, the most frequently detected protein was at 34.2 kDa, found in 53.8% of patients.
As the amount of aqueous humor was insufficient for Toxoplasma strain typing using an ELISA peptide-based assay, we decided to serotype these patients using their sera. Ten OT patients from each center were assessed, all from the early consultation group. Among the Colombian patients, no Type II serotype was detected. We found 4 I/III, one atypical and 5 non reactive (NR) serotypes (Table 4). In contrast, all tested French OT patients showed Type II serotypes except one patient with an atypical serotype. These patterns were significantly different between the two groups (p<0.0001). The two cases COL#26 and COL#38, found as suspected Type I and Type I/III by genotyping, were serotyped as NR and type I/III, respectively (Table 4).
To test if certain T. gondii strains are associated with OT, we determined the overall distribution of serotypes in infected non-OT control populations from both countries. Among the 45 Colombian control patients, only 6 subjects (13.3%) had a type II whereas 39 (86.6%) had NE-II serotypes, which were subdivided in 6 NR, 29 type I/III and 4 atypical serotypes. Of 100 French control patients, we found 64 (64%) type II, and 36 (36%) with NE-II; 10 NR, 2 type I/III and 24 atypical serotypes. No statistically significant differences were observed between the control and OT groups in Colombian patients, however we found a significant difference (P = 0.02) between the French control and OT populations, with respect to the proportion of the two types, II and NE-II.
Cytokines patterns in aqueous humor of OT patients were compared to cataract controls (Figure 3 and Table S2 in Text S1). Several immune mediators were augmented in French, as well as in Colombian patients. In French patients, the Th1 type cytokines IFN-γ, IL-2 and IL-15 were expressed in all patients. This Th1 immune response was associated to a Th17 response with increased IL-17 production. Additionally, we observed a large proinflammatory response with increased levels of IL-6, IL-1β, IL-8, MIP-1β, MCP-1 and G-CSF. These patients also possessed a corresponding anti-inflammatory response was based on the presence of IL-4, IL-10, and IL-1RA. In contrast, Colombian patients had lower expression of major proinflammatory immune modulators, including IFN-γ, IL-15, IL-17, IL-2, IL-10, MIP-1β, GM-CSF and G-CSF, with the exception of elevated TNF-α and IL-6 levels. These patients also had elevated levels of the counterregulating Th2-type cytokine IL-13.
Previously published studies found differences between South American and European clinical case series on adult patients in terms of frequency of serological markers in OT [8], probability of symptoms in congenital infection [7], as well as inflammation levels and IOP [9]. However, these were mostly retrospective evaluations of multiple studies. Their main limitation is their inclusion of patients with “suspected” OT, rather than biologically confirmed cases. While the ocular signs of toxoplasmic retinochoroiditis are highly suggestive of this disease, they may be mimicked by other infections [22], while in some cases, the symptoms may be atypical [19], [29]. Therefore, we strengthened our evaluation by inclusion of biologically confirmed OT cases only, as well as by comparing the same bio-clinical data from two different populations of OT patients, located in South America and Europe in a cross sectional study. Among the 17 criteria analyzed in the two populations, the following were significantly higher in Colombian patients: macular involvement, vitreous inflammation, strabismus, bilateral involvement and synechiae. Our findings confirm and expand the data from the retrospective study of Dodds et al. from patients with biologically unconfirmed OT which found elevated IOP, increased presence of synechiae, AC cells, flare, and vitreous humor haze [9]. In our study, one key difference between the two patient populations was the date of consultation, as Colombian patients consulted later than the French. However, when our analysis was stratified regarding this aspect, the observed clinical differences remained significant.
The main hypothesis for these clinical differences is based on the idea that severe disease in humans may result from poor host adaptation to neotropical zoonotic strains of T. gondii [11]. Our study accumulated some clues supporting this hypothesis.
Central strain-specific parasite virulence factors in human infections were revealed in the last years [30]. Their role in the presence of more virulent parasite genotypes in South America [11], [12] is not yet thoroughly studied. Theses strains are rarely found in Europe [31] where Type II genotypes predominate, including in OT patients [32]. In the three Colombian OT patients where we could detect Toxoplasma DNA, we found one Type I and two atypical strains. The fact that no patient of the French group had a sufficient ocular parasite load for genotyping clearly shows the difference in ocular virulence. Additionally, we noticed that intraocular antibodies responses showed major differences in Toxoplasma antigen recognition by an immunoblotting assay. Even if this could be partly due to better detection of Toxoplasma Type I antigens used in this assay by Colombian patients, other, host immune specific factors are certainly crucial such as local antibodies, whose exact role and function should be explored.
Our serotyping assay confirmed that Colombian and French patients recognize different strain-specific epitopes. Colombian OT patients recognized a heterogeneous pattern of strain specific peptides, but none of them were from type II strains. The French OT patients recognized only Type II strain specific peptides, confirming the reliability of this test in a geographic region with predominant type II strains infections [33]. The corresponding control populations presented the same serological pattern for Colombia, but a slightly different pattern for France, where some sera were non reactive to Type II antigens. The difference may due to the unequal sample sizes, so this point needs further investigation using more samples and equilibrated OT and control population. However, these data indicate that type II and non-type II strains are able to cause ocular pathology, but with a markedly different clinical picture. Concerning the Colombian strains, current serotyping techniques might be not sensitive enough to distinguish the highly variable strains.
When we looked at the patients' local immunological reaction, we observed clearly different cytokine signatures. In French patients, the host-parasite relationship seems to be equilibrated between protection and inflammation. The protective effect of IFN-γ is balanced by anti-inflammatory cytokines such as IL-2 and IL-10. The role of IL-17 is controversial. We have previously observed an early pathologic and parasite promoting role for IL-17 in French patients and in an animal model infected by a Type II Toxoplasma strain [34]. In the intraocular ocular environment, IL-17 would attract neutrophils [35] and, accompanied by IL-15 and MIP-1β/CCL4, activates and attracts NK cells [36] and monocytes [37]. All these innate immune cells might cause retinal inflammation, but then permit to control Toxoplasma proliferation [38], [39]. As our recent findings implicate IL-27 and the Treg subset in counterbalancing deleterious inflammatory Th17 type responses [34], the corresponding mediators deserve to be examined more closely in future studies.
In contrast, in the clinically more severe Colombian cases, IFN-γ and other major immunomodulators such as IL-17 were barely detectable, while IL-6 and IL-13 were enhanced. Virulent strains encode virulence factors able to modulate multiple immune host cell signaling pathways through polymorphic effectors secreted into the host cells such as ROP16 and GRA15 [38], [40]. The presence of Toxoplasma effector proteins from virulent strains could explain the down-regulation of ocular IFN-γ, leading to higher ocular parasite loads in Colombian patients. The IL-17 down-regulation remains to be explained, but decreased levels of IL-17 and other immune modulators, including proangiogenic factors, could lead to a defect in the migration of leukocytes to the eyes and be another explanation for impaired control of parasites in the context of virulent South American infections. IL-6 could also antagonize the anti-microbial properties of IFN-γ by sustained activation of STAT3, a potent inhibitor of IL-12 and IFN-γ [41]. Down-regulation of IFN-γ and its anti-Toxoplasma activity was also observed for IL-13 in human fibroblasts [42]. It is important to note here that Type I strains express a ROP16 allele associated with prolonged activation of STAT3 and STAT6 signaling, which may in part contribute to the increased IL-13 levels, whereas Type II strains activate this pathway only transiently, allowing the establishment of an inflammatory reaction [43]. This may constitute the fundamental basis for the differential cytokine response observed in our study.
The theory of local T cell exhaustion may be also of interest in the settings of Colombian patients. Immune exhaustion is characterized by the modification of the CD8+ functions by reducing their polyfunctionality and their efficacy [44]. Indeed, high Toxoplasma loads associated with a decreased IFN-γ and IL-15 production and enhancement of TNF-α could be one aspect of this loss of CD8+ T cell polyfunctionality. In contrast, in French patients, elevated IL-15 is critical for homeostasis of memory CD8 T cells, and may lead to a better control of parasite proliferation and subsequent parasite latency in the retina.
Taken together, our results indicate that virulent strains observed in South America may suppress host-protective pathways, opening the way to multiplication and cytolytic activity of the parasite in retinal tissues including blood vessels. The presence of TNF-α in most of these patients could also contribute by enhancing an ongoing immunopathological retinal process [45]. In contrast, in French patients, the cytokinic environment may lead to the encystation of the parasite in the retinal tissues, leading to subsequent recurrences.
Of course, for ethical reasons, we were only able to take one time-point. Our results represent thus a snapshot of a developing immune response. Additionally, a multifactorial origin of the observed clinical and biological differences could not be excluded. In our study, the source of contamination may have been drinking water collected from surface water sources (i.e., rivers, lakes) [46], [47], [48], [49]. The more common macular involvement in Colombian patients is often associated with congenital toxoplasmosis [6], [15], [50], [51]. Even if we studied adult populations, we cannot exclude a congenital origin of infection in some Colombian patients. Moreover, acute toxoplasmosis was only diagnosed in 2 Colombian and 1 French case. The remaining population was considered to exhibit chronic toxoplasmosis. Finally, individual susceptibility was previously related to variations in various genes encoding immune response players, such as IFN-γ, IL-1α, IL-10, TLR-9 or ABCA4, COL2A1, and P2X7-R [52], [53], [54], [55]. These genetically susceptible patients are possibly less able to cope with a more virulent strain. Further investigations with larger cohorts including an evaluation of their immunological response and their individual susceptibility to Toxoplasma are needed to address these topics.
|
10.1371/journal.pgen.1000449 | Microarray Profiling of Phage-Display Selections for Rapid Mapping of Transcription Factor–DNA Interactions | Modern computational methods are revealing putative transcription-factor (TF) binding sites at an extraordinary rate. However, the major challenge in studying transcriptional networks is to map these regulatory element predictions to the protein transcription factors that bind them. We have developed a microarray-based profiling of phage-display selection (MaPS) strategy that allows rapid and global survey of an organism's proteome for sequence-specific interactions with such putative DNA regulatory elements. Application to a variety of known yeast TF binding sites successfully identified the cognate TF from the background of a complex whole-proteome library. These factors contain DNA-binding domains from diverse families, including Myb, TEA, MADS box, and C2H2 zinc-finger. Using MaPS, we identified Dot6 as a trans-active partner of the long-predicted orphan yeast element Polymerase A & C (PAC). MaPS technology should enable rapid and proteome-scale study of bi-molecular interactions within transcriptional networks.
| Specific interactions between protein transcription factors (TFs) and their DNA recognition sites are central to the regulation of gene expression. Inter-species conservation of these TF binding sites (TFBS), and their statistical enrichment in sets of co-expressed genes, facilitates their large-scale prediction through computational sequence analysis. A major challenge in characterizing these putative TFBS is the identification of the proteins that bind them. We have developed a new approach to this problem by expressing random genomically encoded protein fragments as fusions to the capsid of bacteriophage T7. We select this diverse phage-display “library” for binding surface-immobilized instances of the TFBS in the form of short double-stranded DNA. This in vitro selection strategy leads to the enrichment of phage whose capsid-fusion peptides interact with the specific DNA sequence. Because each phage carries the DNA encoding the peptide fusion, the identity of the enriched phage can be determined through population-level PCR amplification of DNA inserts and their hybridization to DNA microarrays. Here, we show that this technology efficiently reveals the identity of proteins that bind known and novel predicted regulatory elements. Its application to a predicted yeast element (PAC) reveals Dot6 as one of its interaction partners, both in vitro and within the yeast nucleus.
| The arrival of complete genomes and microarray technology has fueled a revolution in computational predictions of transcriptional regulatory elements, both through inter-species comparative genomics [1],[2] and mapping sequence to gene expression [3]. Application of these approaches to well-studies systems such as Saccharomyces cerevisiae has revealed the majority of previously-known TF binding sites, in addition to many novel predictions with strong evidence of function. As the list of high-confidence cis-regulatory element predictions grows, a more rapid and efficient approach is needed for the identification of proteins that bind these elements and connect them to the transcriptional regulatory network. Current biochemical and genetic methods of transcription factor identification are laborious and time-consuming. DNA affinity chromatography [4] requires chromatographic experience and biochemical skill, and typically entails several rounds of purification, requiring a significant investment of both time and input protein (due to losses) to isolate a single transcription factor [5]–[9]. Yeast one-hybrid and two-hybrid screens [10] to discover protein-DNA and protein-protein interactions are time-consuming and susceptible to both false positives and false negatives requiring extensive follow-up, especially when transcription factors are the potential interactants [11],[12]. Protein-binding microarrays [13],[14] are dependent upon choosing the right proteins for analysis and the ability to purify a functional epitope-tagged form of those proteins for use as a protein-binding microarray probe.
Phage display has been previously used to study protein-DNA interactions, but this work has focused mainly on the binding of specific zinc fingers to associated DNA nucleotide triplets [15]–[18]. Only a limited number of studies have used phage display libraries to enrich for a natural nucleic acid binding protein by selection against a specific nucleic acid target sequence [19]–[21].
We have developed a technology for identifying proteins that specifically bind predicted transcriptional regulatory elements (Figure 1). Our approach, called MaPS (for Microarray profiling of Phage-display Selections), selects a diverse (∼108) phage-display library of genomically encoded peptides for binding to surface-immobilized double-stranded DNA containing a DNA motif sequence of interest. After enrichment for a specific DNA-protein interaction, the bound phage are amplified, and can be used for more rounds of selection in order to further enrich the library for specific interactors. Typically, after the appropriate number of rounds of selection, the inserts from the enriched phage can be sequenced individually to identify the interacting proteins. However, in typical selections, the high level of background requires many rounds of phage display enrichment, followed by sequencing of sufficient number of plaques to develop a consensus sequence [22]–[24]. This leads to the selection of phage with only the highest binding affinities at the expense of lower-affinity but biologically relevant interactions. To bypass these limitations, we have developed a simple strategy that effectively ‘sequences’ the entire population of selected phage through PCR-amplification of inserts, labeling and hybridization to a microarray containing all the open reading frames (ORFs) encoded in the genome.
We have chosen a T7 phage system to display peptides between 300–1000 (with a mode of 500) amino acids in length. The T7 phage system offers multiple advantages over other display vehicles, such as Lambda or filamentous phage. The lytic T7 bacteriophage does not have to be exported through the bacterial inner membrane, placing fewer restrictions on the proteins that may be expressed than the more common phage display vector M13 [25],[26]. Other advantages of T7 include extreme robustness to environmental conditions, high capsid-fusion valency (up to 415 per phage), and rapid replication rate.
One of the strongest computationally predicted cis-regulatory elements in yeast is the Polymerase A & C (PAC) motif which was initially identified as a conserved sequence found in the upstream region of RNA polymerase I & III subunit genes [27]. Computational analysis of expression data found PAC, in conjunction with the Ribosomal RNA Processing Element (RRPE), to be highly enriched in the upstream regions of a cluster of genes enriched for RNA polymerase I & III transcription, RNA splicing, translation initiation, and other RNA metabolism functions [28]. These sequences are well-conserved among related yeast species [1],[29],[30], and their presence is highly predictive of the expression pattern of their downstream genes [31]. The transcription factor Stb3 was recently identified as the trans factor that binds RRPE [32], but to date no PAC-binding protein has been identified, despite numerous attempts [32]–[34].
Here we show that our MaPS technology allows for rapid and proteome-scale survey of sequence-specific protein-DNA interactions. We show that across a variety of test cases, corresponding to known TF-binding sites, MaPS identifies the cognate TF regulators. Moreover, in the most challenging application, we used MaPS to discover the transcription factor that specifically interacts with the PAC element.
In proof-of-principle experiments, the DNA-binding domains (DBD) and complete ORFs of RAP1 and MCM1 were cloned into T7 phage and tested for enrichment from the background of a phage display peptide library encoding restriction-digested yeast genomic DNA fragments (see Materials and Methods). Rap1 and Mcm1 were chosen as well-characterized transcription factors whose target genes [35],[36], recognition sequences [5],[37], specific dissociation constants [38],[39], DNA binding domains [40],[41], and crystal structures [42],[43] had been previously determined.
A library consisting of T7 phage displaying Rap1 and Mcm1 was mixed on an equal-titer basis with a T7 phage library containing fragments from a yeast genomic DNA complete restriction digest. This library was selected against biotinylated double-stranded oligonucleotides (bRAP1 and bMCM1) consisting of native sequences from upstream of open reading frames RPS21B (YJL136C) and MSG5 (YNL053W) and centered around the pairs of Rap1 and Mcm1 binding sites, respectively. Two rounds of selection were performed against three different amounts of target bDNA in buffers of three different salt concentrations, and the results of the selection determined by parallel PCR-amplification of phage inserts from liquid culture obtained from the second round of selection (Figure 2).
A band corresponding to the RAP1 DBD clone was present as a result of selection against the oligonucleotide containing Rap1 binding sites (1 pmol) under all three salt conditions. This band was not visible in both the starting library (input) and selections with no oligonucleotide present (0 pmol). The intensity of the band, implying level of enrichment, was dependent upon the amount of oligonucleotide added; the intensity of the band was reduced when only 0.1 pmol of oligonucleotide was used. The band was absent under all three salt conditions when the oligonucleotide selected against contained Mcm1 binding sites instead of Rap1 (data not shown), indicating that the enrichment was sequence-specific, rather than merely due to the presence of double-stranded DNA. No band corresponding to the entire RAP1 ORF clone was visible in any of the lanes in which a RAP1 DBD band was present, implying that the complete ORF failed to enrich by selection. This may be due to the under-representation of this clone in the library, due to a bias against large inserts proportional to the size of the translated product [26].
The intensity of the RAP1 DBD clone bands across the three salt conditions, for a given amount of target oligonucleotide (1 pmol), rose from a minimum in the low-salt condition to a maximum in the intermediate condition, before falling again at the highest salt concentration. This is consistent with greater competition at salt concentrations below physiological levels (100–200 mM NaCl), from phage displaying peptides with nonspecific DNA binding affinity [44]. At salt concentrations above physiological conditions-and approaching that used to elute transcription factors in DNA affinity chromatography [6]-sequence specific affinity is reduced, resulting in washing away of phage bearing sequence-specific interactors.
No band corresponding to either the MCM1 DNA binding domain clone or to the MCM1 ORF clone appeared from the selection against the oligonucleotide containing Mcm1 binding sites (data not shown). The bands corresponding to these two clones were visible in the selection against the Rap1 target oligonucleotide (Figure 2), but they appeared at a constant intensity regardless of the amount of DNA present at a given salt concentration, and even at a similar intensity in the starting library (input). We believe that the failure of these doped MCM1 clones to enrich by DNA affinity selection was due to the close proximity of the Mcm1 DNA-binding domain to the capsid protein. The RAP1 DBD clone encodes an additional 50 amino acids before the start of the actual binding domain, which may provide flexibility to the domain relative to the phage capsid. On the other hand, only 16 amino acids are encoded between the cloning junction and the MADS box homology region that mediates Mcm1 binding. This is likely insufficient to allow the Mcm1 DNA-binding domain to move into optimum DNA-binding configuration relative to the capsid, as supported by evidence presented later.
A diverse phage display library was constructed using yeast genomic DNA partially digested by restriction enzymes that recognize 4 bp restriction sites and leave blunt ends. The library was based on genomic DNA fragments to avoid the bias against low-abundance transcripts in cDNA libraries and the considerable time and labor necessary for constructing a complete ORF library, while at the same time achieving a sub-genic resolution capable of isolating critical domains. The genomic DNA was partially digested with the selected restriction enzymes to produce the greatest possible number of fragments directly ligatable into the T7 genome with no further enzymatic manipulation. The T7 phage genome was altered to accommodate these fragments, adding a 9-glycine linker and a variable 0–2 base pair frame-shift between the capsid protein and the insert site.
The genomic DNA phage display library was affinity-selected against the oligonucleotide bRAP1, a PCR product containing the sequence from the pair of Rap1 binding sites to the start of RPS21B (bRAP1–322), and a PCR product containing the entire upstream region of RPS21B (bRAP1–734). The library was selected for three rounds under 100–200 mM NaCl conditions, and PCR monitoring of the liquid lysate demonstrated the enrichment of discrete clones. PCR products were labeled and co-hybridized with a genomic reference. The mean percentile rank values for each ORF were calculated for one selection against bRAP1 and two each against bRAP1–322 and bRAP1–734 (Figure 3), with a high mean rank resulting from consistently high enrichment against all three targets. The ORF with the highest mean rank, and the only ORF >98th percentile, was RAP1 (YNL216W). The clone corresponding to this insert was also sequenced, and shown to contain a 1-kb fragment of RAP1 that included the DNA-binding domain, ligated in frame with T7 gene 10.
The genomic DNA phage display library was also selected against the oligonucleotide bMCM1 and a PCR product containing the entire upstream region of MSG5 (bMCM1–401). These targets contained a region with two Mcm1 binding sites flanking a Tec1 binding site [33]. As with the Rap1 binding site, after three rounds of selection, the phage inserts were amplified and profiled through microarray hybridization. The mean percentile rank of the two selections revealed TEC1 as the most highly ranked ORF and MCM1 as the 4th most highly ranked ORF. When the clones corresponding to these inserts were sequenced, they were shown to contain in-frame ligations of the entire MCM1 ORF, and a 1.3-kb fragment of TEC1 that included the DNA-binding TEA/ATTS domain. Moreover, the MCM1 clone included the 87 bp immediately upstream of the ATG, resulting in an additional 29 amino acids, plus the 9-glycine linker, between the T7 capsid and the MADS box that mediates DNA binding. This success implies that the direct clone had failed to enrich in the earlier test because the protein was too close to the phage capsid for proper orientation/activity. Most importantly, given that Mcm1 binds DNA as a dimer in vivo, its enrichment here clearly demonstrates that dimerization is not an absolute obstacle in the application of MaPS, and that relatively weak, yet specific, protein-DNA interactions can be discovered.
As an additional proof-of-principle, the phage display library was selected against the oligonucleotide bRPN4, derived from upstream sequence of PRE7 (YBL041W) centered around the Rpn4 binding site, and including a second Rpn4 site in tandem with the native. Three rounds of selection repeatedly resulted in a single clone bearing a 1.5-kb insert (Data not shown). Microarray analysis was not performed because of the absence of other clones and the consistency of the result. Sequence analysis confirmed that the insert contained the 3′ third of RPN4, including the region encoding the zinc-finger DNA-binding domain.
After the proof-of-principle validations presented above, we asked whether MaPS was able to discover the novel transcription factor associated with the computationally predicted cis-regulatory element PAC. To this end, the phage display library was affinity-selected against bPAC/RRPE, an oligonucleotide centered around the PAC and RRPE sites upstream of RPC82 (YPR190C), bPAC-320, a PCR product of the entire upstream region of RPC82, and bPAC4, a concatemer of four predicted genomic PAC sequences with high computational motif scores, computed using the probabilistic profile captured by the PAC Position Weight Matrix [31]. Three rounds of selection resulted in the enrichment of multiple bands (data not shown). PCR products from two selections against bPAC-320 and one each against bPAC/RRPE and bPAC4 were labeled and co-hybridized to microarrays with a genomic reference, and the mean percentile ranks calculated. The highest percentile rank belonged to YMR130W, an uncharacterized gene with a predicted hydrolase domain, but the second-most highly ranked ORF was DOT6 (YER088C). We focused on Dot6 as the most likely candidate PAC-binding protein based on previous evidence for a role in transcriptional regulation and because YMR130W also ranked relatively highly in selections for Rap1 (9 percentile) and Mcm1/Tec1 (11 percentile). Isolation and sequencing of the corresponding clone confirmed the in-frame ligation of an 800-bp fragment of DOT6, including the DNA binding domain.
Dot6 is a protein with a predicted myb DNA binding domain whose binding site could not be identified by chromatin IP [33]. Over-expression of Dot6 reduces silencing at rDNA loci [45], a side effect consistent with the induction of transcription of RNA polymerase I. Both Stb3 and Dot6 have been characterized as binding components of the Rpd3 histone deacetylase complex [46],[47], which has been shown by chromatin IP to bind genes with PAC and RRPE elements in their upstream regions [48].
To establish that Dot6 is indeed a PAC-binding protein, we performed gel shift assays using recombinant Dot6 and oligonucleotides bearing PAC elements. The DNA binding domain of Dot6 was cloned into the pGEX vector and purified by GST tag from E. coli, and the purified protein was tested for binding to bPAC4 probe (Figure 4). The probe produced additional shifted bands in the presence of protein purified from the Dot6-expressing strain, but not from the strain containing the empty vector, suggesting that the shift was Dot6-specific. This interaction is sequence specific, being successfully competed by ∼200-fold excess of unlabeled competitor (PAC4), but requiring a ∼107-fold excess of competitor with point mutations in each of the PAC elements (XPAC4).
We have conducted several studies to assess whether Dot6 contributes to regulation of expression in yeast of genes containing PAC motifs. In a series of genetic studies to be reported elsewhere (Lippman and Broach, unpublished observations), we have shown that deletion of DOT6 has little effect on expression of genes with PAC sites in their promoters during steady-state growth in rich media. Accordingly, Dot6 is not required for expression of such genes under normal growth conditions. However, we observed that Dot6 in conjunction with its paralog Tod6 are required for efficient repression of such genes during nutrient starvation or upon inactivation of the major nutrient responsive signaling pathways, mediated by PKA or TORC1. These studies suggest that Dot6 and Tod6 are redundant repressors of transcription of PAC-containing genes and that they are inactivated by nutrient induced signaling to enhance expression of PAC-site containing genes upon nutrient stimulation.
In a second series of experiments, we used chromatin immunoprecipitation to examine whether Dot6 binds in vivo to promoters of genes that contain PAC motifs. Given our genetic observations described above, we would anticipate that Dot6 would likely be bound to PAC-containing promoters only under conditions of attenuated nutrient signaling. Accordingly, we assayed for Dot6 DNA binding in cells subjected to carbon starvation, focusing on three genes, YDL063c, MRD1 and YTM1, whose promoters contain 5, 4, and 3 PAC sites, respectively. Transcription of these three genes is substantially repressed upon inhibition of PKA signaling in a DOT6 strain but this repression is attenuated by at least 3 fold in dot6Δ cells (Lippman and Broach, unpublished data). We measured the in vivo association of Dot6 with these promoters, as well as with the promoter of a control gene, GAP1, that lacks any PAC sites, by determining the relative amount of promoter DNA immunoprecipitated from a strain expressing a TAP-tagged version of Dot6. These values were then normalized to the relative amount of ACT1 promoter DNA immunoprecipitated in the same experiment. As shown in Figure 5, promoters for the three genes containing PAC-sites were enriched 12–30 fold over the ACT1 promoter following immunoprecipitation from the strain expressing TAP tagged Dot6. Significantly less DNA from these promoters was immunoprecipitated from a strain expressing the untagged Dot6 and the small amount precipitated was not enriched relative to the ACT1 promoter. Finally, the control promoter GAP1 was not enriched in the immunoprecipitate from the tagged strain. These results are consistent with the conclusion that Dot6 specifically associates with promoters containing PAC sites in vivo.
As both the number of predictions of biologically significant nucleic acid sequences and the number of laboratories making these predictions increases, a rapid and accessible method is needed by which investigators can quickly identify their cognate interacting proteins. The most commonly used approach for identifying nucleic acid-interacting proteins, DNA affinity chromatography, while effective at isolating transcription factors, has considerable up-front costs in material and biochemical expertise not available to most laboratories. Our MaPS technology utilizes phage display, in vitro affinity selection, and microarray display in order to rapidly survey the proteome for sequence-specific interactions with a DNA sequence of interest. Another major advantage of MaPS is the ability to simultaneously discover multiple transcription factors that interact with a region of DNA hundreds of base pairs in length. This capability should allow rapid high throughput characterization of the large fraction of non-coding DNA that is under selection for regulatory control. In addition, MaPS allows a laboratory to move quickly and easily from cis-regulatory motif prediction to identification of the interacting trans-factor without the need for specialized equipment or skills. Where possible, we have made use of commercially available systems and the most common molecular biological techniques in order to maximize the accessibility of the technology.
As with other technologies that are based on molecular libraries of complex genomes, successful identification of transcription factors, using MaPS, relies on high-coverage representation of the coding portion of the genome. We have shown that the use of genomic fragment libraries is a feasible solution in an organism with high gene density (S. cerevisiae). However, our success-rate of 70% for identifying known TF-DNA interactions, at least partly, reflects incomplete coverage of the yeast proteome in our phage-display library. The utilization of well-curated ORF collections or normalized cDNA libraries should partly address this challenge in the case of more complex metazoan genomes with much lower gene density. Another challenge is the competition for specific enrichment of TF-DNA interactions by phage whose capsid fusions interact with the solid phase or DNA non-specifically. In addition, some peptide fusions may provide phage with higher reproductive fitness relative to the rest of the library. The exponential amplification of these super-fit phage after every round of selection may also interfere with sequence-specific enrichment of TFs. Experiments presented here show evidence for enrichment of such false-positives. For example, although the highest and the fourth highest ORF signals belonged to the known cognate TFs in the Mcm1/Tec1 selection, the second (GDS1) and third (MET8) were likely false positives. The gene GDS1 encodes a mitochondrion localized protein of unknown function and MET8 encodes a bifunctional dehydrogenase and ferrochelatase involved in seroheme biosynthesis. As would be expected for non-specific enrichment, GDS1 was also ranked high in the selections against the PAC element (3rd highest) and Rap1 binding sites (112th highest).
The isolation of Dot6 as a PAC element binding protein illustrates the power and unique advantages of our whole-proteome in vitro approach. Transcription reporter experiments [49] and association with the Rpd3 histone deacetylase complex [46],[48] imply that Dot6 acts as a repressor of PAC-regulated genes, whose repression is removed under favorable growth conditions. Since both yeast one-hybrid and chromatin IP are normally performed during log-phase growth, the in vivo conditions made it unlikely that the Dot6-PAC element interaction could have been detected.
Successive rounds of selection provide us with increased discriminatory power in a manner similar to multiple columns in DNA affinity chromatography. Because the phage are regenerated before each selection round, it is free of material losses typically seen in affinity chromatography. Despite the requirement for multiple rounds, MaPS has rapid turnaround. It was possible to conduct two rounds of selection per day, such that a full PCR readout from three rounds of selection was available by the end of the second day. Because results from the second round were adequate for use on the microarray, it was possible to have fully processed data returned by the third or fourth day. This system is also easily amenable to automation with robotic liquid dispensers and automatic plate washers performing the liquid handling.
Strain Y3648 expressing Dot6-TAP was obtained from Open Biosystems (Huntsville, AL) and an isogenic untagged strain, BY4741, was obtained from Research Genetics.
T7 phage displaying Rap1 or Mcm1 were prepared by PCR amplification and cloning of the yeast genes into the T7 genome. Sequences encoding the DNA-binding domains and complete open reading frames of RAP1 (amino acids 289–618 and 1–878) and MCM1 (amino acids 1–135 and 1–286) were amplified from S. cerevisiae genomic DNA using primers that added EcoRI sites to the ends of the PCR products. After EcoRI digestion (which truncated the MCM1 ORF to amino acids 1–187 at an internal EcoRI site) the PCR products were mixed on an equimolar basis and cloned into the EcoRI site of the T7Select 10-3b phage vector. Approximately 0.17 µg of ligation product was packaged with an aliquot of T7 packaging extract (Novagen) and amplified in liquid culture as described [50]. Packaging reaction yield and amplified titers were determined by plaque assay as described [50].
A simple library of T7 phage containing short fragments (∼100 bp) of S. cerevisiae genomic DNA was prepared as a background from which to select Rap1- and Mcm1-expressing phage. Genomic DNA was completely digested with Tsp509I, which leaves 5′ overhangs compatible with EcoRI. The Tsp509I fragments were cloned into the EcoRI site of 10-3b phage vector, packaged, and amplified in the same way and in parallel with the Rap1 and Mcm1 PCR products above.
A short sequence encoding a nine-glycine linker and 0 bp, +1 bp, or +2 bp frameshift was introduced into T7Select 10-3 between the capsid protein gene and the SmaI site. Synthetic oligos encoding a sequence of 9 glycines, followed by 0–2 extra bases, were annealed to their complements to create double-stranded oligos with BamHI and SmaI half-sites at their ends, cloned into the corresponding sites of T7Select 10-3b DNA, and packaged. Samples of individual plaques were resuspended and heated for 10 min at 65°C in 100 µL TE buffer, followed by PCR with primers T7SelectUP and T7SelectDOWN, which flank the T7Select multiple cloning site. Plaques producing PCR products of the correct size were amplified in liquid culture, and successful incorporation of the inserts was confirmed by bidirectional sequencing. Cultures of the correct sequence were stored as T7 strains G9.0, G9.1, and G9.2.
A complex T7 phage display library of the yeast proteome was created from partial restriction digest fragments of S. cerevisiae gDNA. Genomic DNA was partially digested with AluI, BstUI, HaeIII, HpyCH4V, or RsaI to produce fragments distributed around a mode of 1.5 KB, and the fragments were size-selected by gel purification to a range of 1–3 KB. Fragments from digestion using each of the restriction enzymes were cloned into the SmaI site of T7 G9.0, T7 G9.1, and T7 G9.2 DNA in separate reactions. 1–4 µg of DNA from each ligation reaction was packaged using T7 packaging extracts, and the number of phage produced by each packaging reaction was estimated by plaque assays. The presence and size of inserts was determined by PCR of random plaques with primers T7SelectUP and T7SelectDOWN. Individual packaging reactions were amplified at a multiplicity of infection of ∼10−5 in log-phase Escherichia coli BLT5615 culture, made to 0.5 M NaCl after lysis, and clarified by centrifugation. Amplified lysates were pooled to provide the same number of copies of each independent clone from each combination of restriction digest and vector DNA, and the pooled lysate was aliquoted and stored at −80°C. The final library was estimated to contain 6.1×107 independent clones, of which 75% were recombinant, at ∼6500 copies/mL.
Target double-stranded oligonucleotides containing putative cis-regulatory motifs had native sequences from upstream of chosen genes. These sequences were chosen based on the strength of the motif's position weight matrix score and the level of correlation of the associated gene's expression profile to that of the cluster from which the motif was derived [31]. Short (≤70 bp) biotinylated double-stranded target oligonucleotides were prepared by annealing a 5′-biotinylated oligonucleotide to its complementary oligonucleotide in an equimolar ratio. Long (≥200 bp) biotinylated double-stranded DNA targets were prepared by PCR of S. cerevisiae ORF upstream regions using one biotinylated and one unbiotinylated primer.
Biotinylated target DNA (5 pmol) was bound to the wells of a StreptaWell High Bind strip (Roche) in 200 µL Binding Buffer (30–300 mM NaCl, 20 mM Tris pH 7.5, 2 mM KCl, 1 mM EDTA, 0.15 mg/mL purified BSA (NEB)) and salmon sperm DNA (10 µg, Invitrogen) for 30 min. T7 phage library (3×1010 pfu, ∼500 copies/clone) was added to the well and phage-DNA binding permitted for 60 min. Wells were washed five times with 300 µL of Wash Buffer (30–300 mM NaCl, 20 mM Tris pH 7.5, 2 mM KCl, 1 mM EDTA, 0.1% Tween-100) to remove unbound phage. Bound phage were eluted by incubating 30 min in 300 µL of Elution Buffer (1 M NaCl, 20 mM Tris-Cl pH 7.5, 2 mM KCl, 1 mM EDTA). Eluted phage (150 µL) were amplified in BLT5615 log phase culture (5 mL) until lysis (∼2 hr), and the lysate was clarified by centrifugation. Clarified lysate was used as the phage input to the next round, and changes in the phage population due to selection were tracked by PCR of the lysate using the T7SelectUP and T7SelectDOWN primers.
DNA was prepared by direct labeling of T7 liquid culture PCR products using Cy3-dUTP (GE Healthcare), Klenow fragment (NEB) and the T7SelectUP and T7SelectDOWN primers, or of yeast genomic DNA using Cy5-dUTP, Klenow fragment and random hexamer mix (GE Healthcare). Labeled probes were purified using the CyScribe GFX Purification Kit (GE Healthcare), aliquots of Cy5-labeled genomic DNA were mixed as a reference with each Cy3-labeled T7 product and concentrated.
Yeast whole-genome spotted ORF microarrays (Microarray Centre, Toronto, ON) were pre-hybridized as described [51], and hybridized as recommended by the vendor [51],[52]. Microarrays were scanned on an Agilent 2565 Microarray Scanner, and the TIFF files processed using GenePix 5.
Median feature and background intensities in both the Cy3 sample channel and Cy5 reference channel for every spot were analyzed by custom Perl scripts. Background-corrected intensities were calculated in each channel for those spots that met minimum signal requirements, with the intensities expressed as a fraction of the total signal intensity in each channel. The final intensity value for each ORF was calculated by averaging the ratio of sample channel to reference channel intensities across the replicate spots for each ORF on each array.
Dot6 was isolated by recombinant expression and GST affinity purification. Sequence encoding the Myb homology domain of Dot6 (amino acids 22–278) was amplified from S. cerevisiae genomic DNA using primers that added SmaI and NotI sites to the ends of the PCR product. The PCR product was cloned into the corresponding sites of expression vector pGEX4T-3, which provides an N-terminal GST tag, and transformed into E. coli strain BL21. Overnight cultures grown in LB broth+50 µg/mL ampicillin were diluted 1∶100 in fresh media, grown to an OD600 of 0.5, induced with 1 mM isopropyl-β-D-thiogalactoside, and incubated an additional 1.5 hr. Cells were pelleted by centrifugation and lysed with BugBuster Protein Extraction Reagent (Novagen) with 12.5 µg/mL DNaseI (Roche), 200 µg/mL lysozyme (Roche), and complete protease inhibitor (Roche). GST-Dot6 was purified from the soluble fraction remaining after centrifugation using Microspin GST Purification Columns (Amersham) according to the manufacturer's protocols, eluted with 10 mM glutathione in 50 mM Tris-HCl (pH 8.0), and stored at −20°C.
Purified protein was incubated for 1 hr at room temperature with 25 ng/µL salmon sperm DNA, 0.5 nM target bDNA, and 0–500 µM unbiotinylated competitor DNA in 150 mM NaCl Binding Buffer. The binding reactions were electrophoresed on a 5% polyacrylamide gel, transferred to nylon membrane, and analyzed using a Lightshift Chemiluminescent EMSA Kit (Pierce) according to the manufacturer's protocols.
We inoculated a 400 ml culture of SC+2% glucose to a density of OD600 = 0.12 and grew the cells to OD600 = 0.4 at 30°C. Cells were harvested by vacuum filtration and transferred to an equal volume of prewarmed SC media containing no glucose and incubated at 30°C for 80 minutes. Cells were fixed by addition of formaldehyde to a final concentration of 1% and incubated for 20 min at room temperature followed by incubation for 5 min with 0.25 M glycine. Cells were harvested by centrifugation, washed with ice cold PBS buffer, frozen in liquid nitrogen, and stored in −80°C. We resuspended cells from frozen pellets in pre-spheroplasting buffer (100 mM Tris pH 9.0, 10 mM DTT added freshly), incubated the suspension at 10 min at room temperature, harvested cells and resuspended them in spheroplasting buffer (50 mM KH2PO4/K2HPO4 pH 7.5, 1.0 M sorbitol, 10 mM DTT added fresh) containing 0.25 mg/ml zymolyase 100T (Seikagaku Corp, Japan). Cells were incubated at 30°C until converted to greater than 95% spheroplasts (ca. 30 min) and then disrupted by vortexing with an equal volume of glass beads. Lysates were sonicated using W-220 Ultrasonics Sonicator at power setting of 2.5 for 8 cycles of 10 sec each. The TAP-tagged protein and associated chromatin were immunoprecipitated using IgG-Sepharose beads (Amersham) overnight at 4°C. The chromatin cross-links were reversed by incubation at 65°C for 6 hrs and precipitated DNA was purified using QIAgen PCR Cleanup Kit (Valencia, CA). Quantitative PCR analysis was conducted using an Applied Biosystems 7900 instrument. Primers for each gene were designed to be less than 100 bp and to encompass all the PAC motifs present in the promoter of the individual gene. The enrichment of occupancy at a gene's promoter was calculated as the ratio of the fraction of input DNA present in the immunoprecipitate relative to the fraction input of ACT1 promoter DNA present in the same immunoprecipitate.
|
10.1371/journal.pntd.0006779 | A systematic review of antimicrobial resistance in Salmonella enterica serovar Typhi, the etiological agent of typhoid | The temporal and spatial change in trends of antimicrobial resistance (AMR) in typhoid have not been systematically studied, and such information will be critical for defining intervention, as well as planning sustainable prevention strategies.
To identify the phenotypic trends in AMR, 13,833 individual S. Typhi isolates, reported from 1973 to 2018 in 62 publications, were analysed to determine the AMR preponderance over time. Separate analyses of molecular resistance determinants present in over 4,000 isolates reported in 61 publications were also conducted. Multi-drug resistant (MDR) typhoid is in decline in Asia in a setting of high fluoroquinolone resistance while it is on the increase in Africa. Mutations in QRDRs in gyrA (S83F, D87N) and parC (S80I) are the most common mechanisms responsible for fluoroquinolone resistance. Cephalosporin resistant S. Typhi, dubbed extensively drug-resistant (XDR) is a real threat and underscores the urgency in deploying the Vi-conjugate vaccines.
From these observations, it appears that AMR in S. Typhi will continue to emerge leading to treatment failure, changes in antimicrobial policy and further resistance developing in S. Typhi isolates and other Gram-negative bacteria in endemic regions. The deployment of typhoid conjugate vaccines to control the disease in endemic regions may be the best defence.
| Typhoid is an invasive bacterial disease causing 26 million illness episodes globally, each year particularly in South Asia and sub-Saharan Africa afflicting children and poorer sections of society disproportionally. AMR is increasingly recognized among S. Typhi lineages spreading from South Asia to Africa, with resistance to first line antibiotics (co-trimoxazole, ampicillin and chloramphenicol), and fluoroquinolones and, of concern, cephalosporins which contribute to treatment failure. AMR in typhoid is not uniform globally and has evolved at different rates in various endemic regions. These trends have not been systematically analysed previously and the objectives of this study included reviewing the phenotypic and genetic determinants of AMR globally over time. The significance of this study revolves around identifying the different trends and mechanisms of AMR and planning interventional strategies accordingly, particularly in light of the Vi-conjugate vaccine candidate which recently received SAGE recommendation and WHO pre-qualification.
| Enteric fever is a systemic infection, caused by the Gram-negative bacteria Salmonella enterica subspecies enterica serovars Typhi and Paratyphi A, that continues to be a significant cause of morbidity and mortality in endemic regions. Annually, it is estimated that over 26 million people are culture positive for S. Typhi/ Paratyphi[1], and a significant proportion of isolates are resistant to multiple antimicrobials[2]. South and South-East Asia, continue to be critical hubs for enteric fever, dominated by the H58 haplotype of S. Typhi in many regions. Fluoroquinolone resistance is widely prevalent across Asia, in part because of the widespread use of this class of antimicrobials.
Terminology to describe AMR in typhoid can be confusing, with the term MDR S. Typhi, historically used to describe combined resistance to chloramphenicol, co-trimoxazole (trimethoprim-sulfamethoxazole) and ampicillin. These antibiotics are frequently termed first-line antimicrobials in the literature as these were amongst the first to be recommended for typhoid treatment by the WHO[3]. MDR S. Typhi is now generally on the decline in South and South-East Asia, potentially because these drugs are no longer in common use, in view of the previous widespread resistance to these agents[4–7]. Empiric antimicrobial use for treating suspected typhoid fever in this region is now predominantly with third-generation cephalosporins including ceftriaxone and cefixime or azithromycin, since fluoroquinolone resistance is so common.
In contrast to the situation in Asia, MDR typhoid appears to be on the increase in parts of Africa. Several regions have reported typhoid outbreaks in the last decade and these have been associated with MDR phenotypes. H58 S. Typhi disease is moving through areas of East and Southern Africa, while, non-H58 haplotypes are implicated in the Western and Northern regions, illustrating the heterogeneous nature of the disease on the continent[8,9].
The historical trend of antibiotic sensitivity and resistance in S. Typhi has not been systematically analysed and reported. Understanding this trend is important and may provide clues for sustaining treatment regimens in endemic areas as well as modelling the potential impact of typhoid vaccines in reducing AMR. This study uses a global genotypic and phenotypic approach to summarise such trends.
The objectives of this review were two-fold: to systematically delineate the historical trend of expressed phenotypic resistance to first-line antimicrobials, nalidixic acid, ciprofloxacin and cephalosporins as well as to describe the molecular mechanisms of AMR in typhoid. The search strategies for both objectives are described in Fig 1. Exclusion criteria such as time of publication, study design and language were not applied in the search builder in order to ensure complete data collection.
An isolate was considered resistant to an antimicrobial if it was reported as “resistant”, “intermediately susceptible”, “intermediately resistant” or “non-susceptible” based on minimum inhibitory concentration (MIC) values or diameters of zones of inhibition via disc diffusion using customary interpretive criteria such as the Clinical & Laboratory Standards Institute (CLSI) or the European Committee on Antimicrobial Susceptibility Testing (EUCAST) standards. For consistency, studies prior to the year 2000, that reported sensitivities of at least the first-line antimicrobials were included while studies conducted after the year 2000, which did not report antimicrobial sensitivities of either chloramphenicol, co-trimoxazole, ampicillin/amoxicillin, nalidixic acid, ciprofloxacin or at least one cephalosporin were excluded. Studies that reported antibiograms collectively and had not stratified these into intervals shorter than 5 years were also excluded.
Isolates identified from reports were then stratified based on year of isolation, geographic location and resistance phenotypes. Stratified isolates that were resistant to each antimicrobial were then expressed as a proportion of all the isolates reported. The trends of antimicrobial resistance were then expressed in 5-year intervals as represented in Table 1. This process was then repeated on isolates collected from Asia and Africa separately.
For the second objective, studies reporting molecular mechanisms of AMR of isolates either collectively or individually were included. These were only stratified based on country of isolation and type of mechanism reported as methods used to study these mechanisms were heterogeneous over the years and techniques employed have also changed drastically thus making temporal comparisons challenging.
Data from individual studies were extracted under the following parameters: (i) study identifier: first author, year of publication, year of study commencement, duration of study, country, study design and sampling population (hospital-based/ community and travel-associated/endemic or outbreak); (ii) methodology: sample size, site of isolation and antimicrobial susceptibility testing, interpretive criteria. For the studies included to study molecular determinants the technique of molecular detection was also recorded. (iii) results: numbers of S. Typhi isolates, frequency of MDR, nalidixic acid resistant, fluoroquinolone resistant and cephalosporin resistant strains. In addition we also collected data o the molecular mechanisms of MDR, fluoroquinolone and cephalosporin resistance in the form of AMR determining genes, resistance plasmids and AMR conferring SNPs. Study-specific data extraction was done twice–overall all for objective 1 and objective 2 separately.
Inclusion criteria were used to establish study validity. Risk of bias (RoB) was assessed using two tools (S3 Table). The first classifies studies based low-, moderate- or high- risk of bias and is known as the Quality In Prognosis Studies tool (QUIPS)[10]. The second is known as the Joanna Briggs Institute (JBI) tool[11] and reports RoB dichotomously. The JBI was adapted for use in this study similar to the adaptations used by Tadesse et al [12] We performed these RoB analysis separately on studies selected to meet the first and second objective. The isolates derived from these studies were then used for the frequency analysis. Parameters assessed for bias across the two tools included 1) Population description, i.e. whether community or hospital setting, 2) Study design, sample size and sampling techniques 3) Use of appropriate performance standards and quality control in microbiologic techniques such as bacteriologic culture and antimicrobial sensitivity and 4) the statistical analysis used for reporting summary measures.
In order to estimate frequencies of antimicrobial resistance in S. Typhi we set key criteria for such an analysis (See Methods). We initially focused on phenotypic data collected through classical antimicrobial susceptibility testing. Here, sixty-two studies (S1 Table) satisfied the inclusion criteria from which a cumulative number of 148 year-stratified summaries of antimicrobial resistant S. Typhi isolates were obtained. For example, Rahman et al[13] reported the isolates of their study in a year-stratified manner for 13 years, therefore providing 13 serial year-stratified summaries. Of our accepted 148 year-stratified summaries, 37 were undertaken prior to the year 2000 and more than 80% were retrospective in study design. The year-stratified summaries obtained from each report were then pooled into the following temporal intervals; pre-1991, 1991–1995, 1996–2000, 2001–2005, 2006–2010 and 2011–2015 and expressed as a proportion of resistant isolates for each antimicrobial (Table 1). In addition to RoB estimation for each included study that suggested most studies were in the spectrum of medium to low risk of bias, the confidence intervals estimated for each year-stratified summary further suggested that majority of data points in each temporal period contributed significantly to the overall trends and is illustrated in S1 Fig.
Of the 13,833 isolates obtained from the various reports, 63.2% were isolated from South Asia, 12.8% were from South-East Asia, 15% were from the continent of Africa mostly represented by countries in the East and South-West regions. The spatial distribution of isolates from endemic settings is illustrated in S2 Fig. Isolates that were cultured from travellers returning from endemic regions made up the remainder of the isolates included in this analysis. The number of isolates within each time interval rose steadily until 2001–2005, a period that accounted for the most isolates (4,725 isolates), the subsequent time intervals saw a decline in published data.
Nalidixic acid, ciprofloxacin and cephalosporin trends were only analysed from the late 1990’s as these drugs were not routinely tested as part of antimicrobial sensitivity studies prior to this period, although preliminary reports of ciprofloxacin resistance surfaced as early as 1992[14]. Fig 2A summarises the global AMR trends, which indicate a trend of decline in MDR and an increasing level of fluoroquinolone resistance.
The temporal distribution of isolates obtained from Asia and Africa, when analysed independently, revealed very different trends as shown in Fig 2B and 2C respectively. The proportion of MDR S. Typhi in Asia saw declining trends, accounting for less than 20% of isolates obtained between 2011 and 2015, whereas resistance to nalidixic acid and fluoroquinolones continued to increase during this period (from 20% in 2001–2005 to 65% in 2011–2015), prompting a change to the use of third-generation cephalosporins in the treatment of enteric fever. Third-generation cephalosporin resistance rose from 1.5% in the 2006–2010 to 4% in the 2011–2015 time interval. Azithromycin is now often used for the treatment of enteric fever, but the number of reports on the susceptibility did not meet the inclusion criteria for this systematic review and were too few to be presented in this study. However, there are sporadic reports of phenotypic resistance[15–17].
In Africa the scenario is very different, where MDR typhoid is still common, with over 90% resistance in some regions. Interestingly, fluoroquinolone and third-generation cephalosporin resistance are still low (< 1%).
To meet the second objective of this review 4,226 isolates spanning 61 studies (S2 Table) were included for the analysis of molecular mechanisms. Most studies (66%) incorporated the polymerase chain reaction (PCR) method to study the molecular determinants of antimicrobial resistance. However, four studies[4,8,9,18] reported whole genome sequence analysis of 2,118 isolates and between them provided valuable insights into the development of resistance in S. Typhi at a molecular level. In keeping with the phenotypic trends of AMR, the molecular findings of isolates between Africa and Asia were contrasting.
Genetic signatures associated with fluoroquinolone resistance were very distinct amongst isolates studied in Asia (Fig 3B). Single nucleotide polymorphisms (SNPs) in gyrA, gyrB, parC and parE, which include the quinolone resistance determining region (QRDR) in the S. Typhi genome, as well as fluoroquinolone resistance conferring plasmids containing qnrB2, qnrB4 and qnrS1 genes were reported. From these data it is apparent that fluoroquinolone resistance in S. Typhi is frequently linked to mutations with gyrA. A frequent position for SNPs in gyrA is codon 83 with the S83F being most common occurring in 1189 isolates. S80I was the most common SNP in the parC gene, detected in 260 isolates, together with a concordant SNP in S83F. The S83Y mutation was detected in 209 isolates, while 57 isolates harboured the mutation gyrA D87N, further underpinning the importance of gyrA–associated SNPs, likely in response to antimicrobial selection pressure. Isolates harbouring combinations of three SNPs in gyrA, at codons 83 and 87 as well as mutations at codon 80 in parC are associated with a high level of ciprofloxacin resistance and designated as ‘triple mutants’. These triple mutations were mostly commonly identified in S. Typhi isolates from South Asia[4,9], often in distinct sub-groups within the main H58 clonal population[4]. SNPs in parE and gyrB were also observed but to a much lower extent (3 and 7 isolates respectively). The qnrB2, qnrB4 and qnrS1 resistance determinants have been found in S. Typhi but they are still rare, being identified in 21 S. Typhi isolates from Asia. These are usually encoded on plasmids. We can anticipate that such isolates may become more common in the future.
The relatively recent trend to a decline in MDR S. Typhi in Asia has been accompanied by a decrease in the proportion of isolates carrying IncHI1 plasmids, which often harbour the resistance genes responsible for MDR typhoid. Such resistance genes are clustered on composite transposons and include catA, sul1, sul2, dfrA, blaTEM-1, strA, strB, tetA, tetB, tetC and tetD. These MDR-associated genes can also be found integrated on the chromosome of H58 S. Typhi in isolates from countries including India and Bangladesh[9]. The arrangement of these genes and transposons both in plasmids and embedded in the bacterial chromosome are illustrated in Fig 3A. Other plasmids identified in S. Typhi included R27-like, B7-like and those falling into IncH and IncN, but these are currently relatively uncommon.
The scenario in Africa was very different with MDR being widely prevalent, conferred in part by determinants encoded on IncHI1 plasmids. The H58 clade of S. Typhi is associated with much of the typhoid occurring in the last decade in East and Southern Africa, although other haplotypes do occur. The situation is somewhat different in Western Africa, where H58 is still uncommon and AMR typhoid is spread via non-H58 clades[8] with both IncH1 and IncY plasmids being present in the circulating population. Again, as elsewhere in Africa, genetic signatures of fluoroquinolone resistance were present in only a few of the analysed isolates. SNPs in gyrA, gyrB, parC and parE were detected in 36 isolates with the S83F SNP in gyrA being the most common. Plasmids encoding the qnrB2, qnrB4 and qnrS1 determinants have also been reported. Other plasmid-types identified in Africa are illustrated in Fig 3A.
Extended spectrum β lactamase (ESBL) producing S. Typhi isolates, which confer resistance to third-generation cephalosporins have been reported in India and Pakistan[19,20]. The Indian isolates carried IncX3 and IncA plasmids which encoded blaSHV-12 and blaCMY-2 determinants[19], as well as blaTEM-1B and blaDHA-1 probably on an IncN plasmid[21]. More recently, a case report of S. Typhi encoding ESBL (blaCTX-M15) on a IncY type plasmid has been reported in the Democratic Republic of Congo[22]. Other CTX-M producing isolates have been reported from Southern India[23], Nigeria[24], Japan[25] as well as from travellers returning from Guatemala[26] and Iraq[27]. A recent publication reported blaCTX-M15 producing S. Typhi isolates from Pakistan that were cephalosporin resistant in addition to MDR and fluoroquinolone resistant and have been labelled as XDR (extensively drug resistant). All the XDR isolates had a composite transposon as described above and an additional IncY plasmid containing blaCTX-M15 and qnrS genes[18].
There were also reports of azithromycin resistance mediated via the ereA from an isolate from Algeria, as well as via msrD and msrA from an Indonesian isolate[9].
The paucity of reliable point of care diagnostics for typhoid fever compels clinicians in the field to initiate presumptive antimicrobial therapy, often based on clinical judgment. In endemic settings, typhoid features high on the list of potential causes of undifferentiated febrile illness, and antimicrobial therapy is routinely started empirically with antimicrobials that are thought to be appropriate for local clades of S. Typhi. The data presented in this systematic review suggest that such antimicrobial use for the treatment of presumptive enteric fever is likely influencing the patterns of AMR in S. Typhi.
Two independent published reports entail global antimicrobial consumption trends. The first report assessed antimicrobial consumption between 2000 and 2010 and suggested that global antibiotic consumption increased by 36% based on national pharmaceutical sales. Most notably in 2010, India and China were the world’s first and second largest consumers of antibiotics respectively[28]. In India and similarly in other LMIC settings, the three classes of antimicrobials that were most consumed were beta-lactams, macrolides, and fluoroquinolones. The second report published in 2018 gauged antimicrobial consumption between 2000 and 2015 in defined daily doses (DDDs) as well in DDDs/1000 inhabitants/day and suggested that antimicrobial consumption has increased by 65% between 2000 and 2015 globally and interestingly showed that India and Pakistan were the two of the top three largest antimicrobial consumers. This rise was attributed to the to changing prescribing practices favouring cephalosporins for enteric fever among other virulent infections involving the respiratory tract, genito-urinary tract, skin and soft tissue in the setting of rising AMR for other antimicrobials including narrow spectrum beta-lactams and fluoroquinolones[29].
For the treatment of enteric fever, the first-line antimicrobials (chloramphenicol, co-trimoxazole and ampicillin) were recommended between 1948 and the early 1990s.[14] Unfortunately, the widespread use of these drugs facilitated the emergence of resistance to chloramphenicol and subsequently to ampicillin and co-trimoxazole, leading to MDR typhoid[14]. MDR typhoid became established in parts of Asia in the 1990’s and the phenotype was mainly conferred through the acquisition of horizontally acquired plasmids[14] harbouring transposons and integrons encoding resistance-determining genes. The most commonly implicated plasmids found in S. Typhi at this time were of the IncHI1 type[14,30,31]. Bayesian analysis suggests that this plasmid was first acquired by H58 and some other haplotypes of S. Typhi in Asia around the early 1990s[9]. With the establishment of widespread MDR typhoid, the use of chloramphenicol, ampicillin and co-trimoxazole became obsolete in this region. However, this analysis indicates that the subsequent circulation of these plasmids within S. Typhi in Asia markedly decreased over time, highlighting the adaptability of S. Typhi to changing antibiotic pressure.[4,32] The Global AMR trends in Fig 2 are driven very strongly by the trends in Asia and this is likely to be related to the magnitude of burden in South and South-East Asian countries, the relatively recent endemicity of the disease in Africa as well as a potential reporting bias related to under-reporting by African regions. Although this review did not attempt to estimate the true burden of antimicrobial resistant typhoid the contribution of Asian isolates to the over-all global trends is conspicuous in Table 1, Figs 2 and 3. Further, Fig 3B illustrates the striking contribution of isolates from Indian studies in determining Asian trends. The highest proportion of isolates analysed in this study came from India (30.4%), Bangladesh (30.1%) and Vietnam (14.4%). Among the African countries Nigeria (2.6%), Kenya (0.7%) and Ghana (0.54%) accounted for the largest proportion of total study isolates. The spatial distribution of isolates is depicted in S2 Fig. The trend observed in Africa is very different and may partially reflect the more recent introduction of S. Typhi isolates into the continent[9]. Transposon-mediated MDR typhoid associated with composite transposons either on plasmids or in the chromosome is increasingly reported, driven by both H58 and non-H58 clades[8]. Although IncHI1 plasmids are still the most commonly identified, other incompatibility (defined as the inability of two related plasmids to be stably transmitted together[33]) group plasmids such as IncY, IncN and IncFIIK (pKPN3) have also been identified in S. Typhi in Africa[8].
Following the emergence of MDR typhoid, fluoroquinolones were adopted as the treatment of choice for typhoid by the late 1990’s. The fluoroquinolone class of antimicrobials were highly effective, could be orally administered, had minimal side effects and had rapid rates of bacteraemic clearance times although potential adverse effects on the growing epiphysis of long bones was viewed with suspicion and initially restricted in children[14]. Nevertheless, ciprofloxacin and ofloxacin became favoured alternatives to the former first-line antimicrobials and consequently fluoroquinolone resistance began to develop. The antimicrobial pressure associated with fluoroquinolone usage likely facilitated the acquisition of alternative modes of antimicrobial evasion by S. Typhi. The spread of fluoroquinolone resistance was accelerated by the emergence of the H58 clade, which dominated circulating S. Typhi populations by the late 1990s, with an apparent increased fitness advantage and enhanced transmission success[9,34]. Unlike resistance to first-line antimicrobials, resistance to fluoroquinolones was mediated via the accumulation of non-synonymous SNPs in the genome inducing conformational changes in DNA gyrase and topoisomerase IV, the main sites of fluoroquinolone action. The genes in which SNPs occur include gyrA, parC, parE and gyrB, with gyrA SNPs correlating strongly with treatment failure[4]. Unfortunately, the standard method of gauging antimicrobial sensitivity, i.e. disc diffusion, suggested that S. Typhi was still relatively sensitive to ciprofloxacin despite ongoing treatment failure and relapse. A WHO report comprising of multi-centric antimicrobial surveillance data of typhoid isolates across India between 2008–2010 suggested that nalidixic acid sensitivity was a good indicator of fluoroquinolone sensitivity but there was a disparate correlation with nalidixic acid resistance and ciprofloxacin resistance[35] which is exemplified by the dissimilar nalidixic acid and ciprofloxacin trend lines in Fig 2A and 2B. These observations concluded that nalidixic acid break points on disc diffusion correlated more accurately with ciprofloxacin sensitivity, prompting a revision in the CLSI recommended break points in 2012[36]. An Indian study compared breakpoints for ciprofloxacin using the CLSI guidelines before and after the 2012 revision and also with the EUCAST guidelines and found that only 3% of isolates were sensitive using the revised guidelines vs 95% of isolates that were sensitive using the older guidelines[36]. The sensitivities of isolates reported using EUCAST breakpoints were comparable to the revised CLSI breakpoints[36]. The nalidixic acid and ciprofloxacin trend lines in Fig 2A and 2B which seem to converge may in reality be attributed to revisions in the CLSI guidelines for ciprofloxacin breakpoints. Fluoroquinolone-resistant S. Typhi isolates are currently widespread in Asia with over 60% of isolates in this review demonstrating resistance. In Fig 3B the proportion of isolates harbouring fluoroquinolone resistance conferring SNPs from South-Asia appears to be less that the proportions of South-East Asia and this is mainly due to two factors; the total number of isolates obtained during data extraction were dominated by reports from South Asia but a substantial proportion of these isolates were obtained prior to the era of widespread fluoroquinolone resistance, secondly it is difficult to do any temporal analysis as these results will be subject to high rates of bias due to the different methods employed in studying genetic determinants which have varied over time. For instance, the PCR, pulse field gel electrophoresis and conjugation transfer to E. Coli techniques employed do not always look for all MDR, fluoroquinolone and cephalosporin determinants of resistance, where as this is possible with whole genome sequencing resulting in broader information of AMR determinants. In Africa, 90% of isolates are still susceptible to fluoroquinolones with some reports of gyrA SNPs recently emerging[37][12]. Accumulating mutations in the QRDR cause S. Typhi to gradually increase the MIC values of ciprofloxacin. Ciprofloxacin-susceptible strains (MIC—0.06 μg ml) are known to acquire a gyrA S83F single mutation with a subsequent increase in MIC values (0.12–0.5 μg ml) and additional gyrA and parC mutations continue to cause an increase in MICs up to 4 μg ml[38].
More recently, third-generation cephalosporins and azithromycin have become the preferred treatment choices for typhoid in the face of MDR and fluoroquinolone resistance, owing to the broad spectrum of activity and the option of oral or intravenous administration. Nevertheless, widespread third-generation cephalosporin resistant typhoid is now on the horizon in South Asia with sporadic reports of treatment failure from India[19,21] and Pakistan[18,20,39] and the XDR typhoid outbreak in populous parts of the Sindh province in Pakistan[18]. In South Asia, cephalosporins such as ceftriaxone and cefixime are currently the mainstay of treatment for enteric fever, and are often started empirically, likely driving resistance in typhoid and other Gram-negative bacteria.
Confirmed typhoid and paratyphoid infections make up only a minority of the total proportion of all Gram-negative infections in endemic regions[40,41]. However, empirical antimicrobial treatment with cephalosporins for presumptive enteric fever confers an antimicrobial pressure, which encompasses all Gram-negative bacterial populations. It is thus plausible that the impact of empiric therapy for typhoid is of far greater importance in driving AMR than just as described in this study in S. Typhi. The mechanisms of resistance adopted by S. Typhi are similar to those among other Gram-negative bacteria [42] and the most contemporary concern stems from the emergence of extended spectrum β lactamases (ESBLs) produced by various Gram-negative species, which has originated as a result of the widespread cephalosporin use. Preventive approaches warrant a collective approach in tackling Gram-negative resistance as the molecular determinants of resistance are transferrable between Gram-negative organisms and thus reducing the use of cephalosporins for typhoid is likely to have an indirect effect on the other Gram-negative organisms [42].
A 2014 publication suggested that cephalosporins were the most commonly used antimicrobial in India and China, followed by broad-spectrum penicillins, fluoroquinolones and macrolides[28]. This trend might still hold true in 2017 which highlights the mounting antimicrobial pressure exerted by the use of cephalosporins culminating in the production of ESBLs by Gram-negatives, including S. Typhi.[19–21] These issues underscore the importance of controlling the spread of typhoid through the deployment of vaccines and prudent antimicrobial use in the short-term.
Single drug therapy (monotherapy) has been common practice in the treatment of typhoid, and monotherapy with former first-line antimicrobials may be a reasonable option in Asia. A single report from Nepal suggests that monotherapy with co-trimoxazole results in complete remission of typhoid fever caused by H58 which was fluoroquinolone-resistant but not MDR[43]. However, a more astute approach in Asia might involve combination therapy with a first-line antimicrobial and perhaps azithromycin. This approach for the treatment of enteric fever in Asia could potentially facilitate the conservation of cephalosporins. The decrease in MDR highlighted in this review following the reduction in use of first-line antibiotics (amoxicillin, chloramphenicol and co-trimoxazole) shows that cycling of these antibiotics for control of typhoid might be an option, where close monitoring of susceptibility is feasible. However, uncoordinated use of these agents would likely lead to a rapid re-emergence of MDR and it is difficult to see how such a programme could be undertaken globally. Immunization could theoretically reduce the number of circulating MDR, fluoroquinolone- and cephalosporin-resistant strains and, furthermore, decrease the incidence of undifferentiated febrile illness thereby reducing the need for empirical antimicrobial therapy.
This study has limitations in that the interpretive criteria employed by majority of studies was the CLSI guidelines which was improved periodically particularly with regard to ciprofloxacin breakpoints in 2012. It is hard to ascertain how quickly individual laboratories made the transition after each revision. Finally, it is also unlikely that true trends of Asian and African isolates are not represented in its entirety, which is mainly due to the lack of published data. Regions from West and Central Africa as well as regions from South-East Asia were under-represented. It was also difficult to account for methodological variations in studying molecular determinants of AMR over time with the rapid evolution of molecular techniques.
S. Typhi rapidly acquires resistance to the antimicrobials that are being used in the community, but can also lose resistance once these drugs are withdrawn. From these observations, it seems likely that antimicrobial resistance will emerge in areas endemic for typhoid, leading to treatment failure, changes in antimicrobial policy and further resistance developing in S. Typhi isolates and other Gram negative bacteria. Therefore, deployment of typhoid conjugate vaccines to control the disease may be the best defence against antimicrobial resistance in S. Typhi.
|
Subsets and Splits